EACL 2006
Third ACL-SIGSEM Workshop on Prepositions Proceedings of the Workshop
Workshop Chairs: Boban Arsenijevi´c Timothy Baldwin Beata Trawi´nski
3 April 2006 Trento, Italy
The conference, the workshop and the tutorials are sponsored by:
Celct c/o BIC, Via dei Solteri, 38 38100 Trento, Italy http://www.celct.it
Xerox Research Centre Europe 6 Chemin de Maupertuis 38240 Meylan, France http://www.xrce.xerox.com
CELI s.r.l. Corso Moncalieri, 21 10131 Torino, Italy http://www.celi.it
Thales 45 rue de Villiers 92526 Neuilly-sur-Seine Cedex, France http://www.thalesgroup.com
EACL-2006 is supported by Trentino S.p.a.
and Metalsistem Group
© April 2006, Association for Computational Linguistics Order copies of ACL proceedings from: Priscilla Rasmussen, Association for Computational Linguistics (ACL), 3 Landmark Center, East Stroudsburg, PA 18301 USA Phone +1-570-476-8006 Fax +1-570-476-0860 E-mail:
[email protected] On-line order form: http://www.aclweb.org/
Introduction
This volume contains the papers accepted for presentation at the Third ACL-SIGSEM Workshop on Prepositions, hosted in conjunction with the 11th Conference of the European Chapter of the Association for Computational Linguistics on April 3rd, 2006, in Trento, Italy. This meeting is supported by the ACL Special Interest Group on Semantics (SIGSEM, http://mcs. open.ac.uk/pp2464/sigsem/), which aims to promote research in all aspects of computational semantics. Two successful workshops endorsed by ACL-SIGSEM devoted to the topic of prepositions were held in Toulouse, France in September 2003, and Colchester, UK in April 2005. Prepositions have received a considerable amount of attention in recent years, due to their importance in computational tasks. For instance, in NLP, PP attachment ambiguities have attracted a lot of attention, and different machine learning techniques have been employed with varying degrees of success. Researchers from various perspectives have also looked at spatial or temporal aspects of prepositions, and their cross-linguistic differences, monolingual and cross-linguistic contrasts or the role of prepositions in syntactic alternations. Moreover, in languages like English and German, phrasal verbs have also been the subject of considerable effort, ranging from techniques for their automatic extraction from corpora, to methods for the determination of their semantics. In other languages, like Romance languages or Hindi, the focus has been either on the incorporation of the preposition or its inclusion in the prepositional phrase. All these configurations are of much interest semantically as well as syntactically. In the call for papers we solicited papers working on aspects of prepositions, such as: Descriptions: prepositions in lexical resources (WordNet, FrameNet), productive versus collocation uses, multilingual descriptions (mismatches, incorporation, divergences), prepositions and thematic roles. Applications: dealing with prepositions in applications e.g. for machine translation, information extraction or natural language generation. Representation of prepositions: prepositions in knowledge bases, cognitive or logic-based formalisms for the description of the semantics of prepositions (in isolation, and in composition/confrontation with the verb and the NP), compositional semantics; implications for AI and KR. Prepositions in reasoning procedures: how different kinds of preposition provide distinct challenges to a reasoning system and how they can be handled. Cognitive dimensions of prepositions: how different kinds of prepositions acquired/interpreted/represented, in terms of human and/or computational processing.
are
Of the 17 papers submitted, the program committee selected 13 papers for presentation at the workshop and inclusion in these proceedings, representative of the state of the art in this subject today. Each full-length submission was peer-reviewed by three members of the international program committee, whose job was made difficult by the high quality level of submissions. The accepted papers include proposals for the extraction of prepositional phrases from corpora, creation of a preposition ontology, disambiguation of verb particle sequences, translation of prepositions, interpretation of prepositions for
iii
practical applications, such as question answering systems, quantitative analyses of prepositions and prepositional phrases, as well as analyses within formal grammatical frameworks such as HPSG, LFG and OT. The papers deal not only with English, but also with languages such as German, Swedish, Polish, Finnish and Bengali. We would like to thank all the authors who submitted papers, the members of the program committee for the time and effort they expended in reviewing the papers, and the panelists for supporting and stimulating this event. Our thanks go also to the executive committee of the ACL-SIGSEM sub-group on prepositions, the organizers of the main conference and Maarten de Rijke and Caroline Sporleder, the EACL 2006 workshop chairs.
Boban Arsenijevi´c, Timothy Baldwin, Beata Trawi´nski
iv
Organizers: Boban Arsenijevi´c (University of Leiden, Netherlands) Timothy Baldwin (University of Melbourne, Australia) Beata Trawi´nski (University of T¨ubingen, Germany)
Program Committee: Doug Arnold (University of Essex, UK) Boban Arsenijevi´c (University of Leiden, Netherlands) Timothy Baldwin (University of Melbourne, Australia) John Beavers (Stanford University, USA) Bob Borsley (University of Essex, UK) Nicoletta Calzolari (Istituto di Linguistica Computazionale, Italy) Ann Copestake (University of Cambridge, UK) Markus Egg (University of Groningen, Netherlands) Christiane Fellbaum (Princeton University, USA) Anette Frank (DFKI, Germany) Julia Hockenmaier (University of Pennsylvania, USA) Tracy Holloway King (PARC, USA) Valia Kordoni (Saarland University, Germany) Ken Litkowski (CL Research, USA) Alda Mari (CNRS / ENST Infres, France) Paola Merlo (University of Geneva, Switzerland) Gertjan van Noord (University of Groningen, Netherlands) Stephen Pulman (University of Oxford, UK) Patrick Saint-Dizier (IRIT, France) Beata Trawi´nski (University of T¨ubingen, Germany) Jesse Tseng (Loria, France) Hans Uszkoreit (Saarland University and DFKI, Germany) Aline Villavicencio (Federal University of Rio Grande do Sul, Brazil) Martin Volk (Stockholms Universitet, Sweden) Joost Zwarts (Utrecht University, Netherlands)
v
Table of Contents
Spatial Prepositions in Context: The Semantics of ‘near’ in the Presence of Distractor Objects Fintan J. Costello and John D. Kelleher . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Polish Equivalents of Spatial ‘at’ Iwona Kna´s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 A Quantitative Approach to Preposition-Pronoun Contraction in Polish Beata Trawi´nski . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 Marked Adpositions Sander Lestrade . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 Semantic Interpretation of Prepositions for NLP Applications Sven Hartrumpf, Hermann Helbig and Rainer Osswald . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 Coverage and Inheritance in The Preposition Project Kenneth C. Litkowski and Orin Hargraves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 An Ontology-based View on Prepositional Senses Tine Lassen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 A Conceptual Analysis of the Notion of Instrumentality via a Multilingual Analysis Asanee Kawtrakul, Mukda Suktarachan, Bali Ranaivo-Malancon, Pek Kuan, Achla Raina, Sudeshna Sarkar, Alda Mari, Sina Zarriess, Elixabete Murguia and Patrick Saint-Dizier. . . . . . . . . . . . . . . . . . . . .51 German Particle Verbs and Pleonastic Prepositions Ines Rehbein . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 Automatic Identification of English Verb Particle Constructions using Linguistic Features Su Nam Kim and Timothy Baldwin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 On the Prepositions which Introduce an Adjunct of Duration Frank Van Eynde . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73 How Bad is the Problem of PP-Attachment? A Comparison of English, German and Swedish Martin Volk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 Handling of Prepositions in English to Bengali Machine Translation Sudip Kumar Naskar and Sivaji Bandyopadhyayn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
vii
Workshop Program 10 D ECEMBER , 2005 08:55-09:00
Opening Remarks
09:00-09:30
Spatial Prepositions in Context: The Semantics of ‘near’ in the Presence of Distractor Objects Fintan J. Costello and John D. Kelleher
09:30-10:00
Polish Equivalents of Spatial ‘at’ Iwona Kna´s
10:00-10:20
A Quantitative Approach to Preposition-Pronoun Contraction in Polish Beata Trawi´nski
10:20-10:40
Marked Adpositions Sander Lestrade
10:40-11:00
Coffee Break
11:00-11:30
Semantic Interpretation of Prepositions for NLP Applications Sven Hartrumpf, Hermann Helbig and Rainer Osswald
11:30-12:00
Coverage and Inheritance in The Preposition Project Kenneth C. Litkowski and Orin Hargraves
12:00-12:20
An Ontology-based View on Prepositional Senses Tine Lassen
12:20-12:40
A Conceptual Analysis of the Notion of Instrumentality via a Multilingual Analysis Asanee Kawtrakul, Mukda Suktarachan, Bali Ranaivo-Malancon, Pek Kuan, Achla Raina, Sudeshna Sarkar, Alda Mari, Sina Zarriess, Elixabete Murguia and Patrick SaintDizier
12:40-14:00
Lunch
14:00-15:00
Discussion
15:00-15:30
German Particle Verbs and Pleonastic Prepositions Ines Rehbein
ix
10 D ECEMBER , 2005 (continued) 15:30-16:00
Automatic Identification of English Verb Particle Constructions using Linguistic Features Su Nam Kim and Timothy Baldwin
16:00-16:20
Coffee Break
16:20-16:50
On the Prepositions which Introduce an Adjunct of Duration Frank Van Eynde
16:50-17:20
How Bad is the Problem of PP-Attachment? A Comparison of English, German and Swedish Martin Volk
17:20-17:40
Handling of Prepositions in English to Bengali Machine Translation Sudip Kumar Naskar and Sivaji Bandyopadhyayn
17:40-18:10
Closing Remarks and Business Meeting
x
Spatial Prepositions in Context: The Semantics of near in the Presence of Distractor Objects John D. Kelleher Fintan J. Costello, School of Computing, School of Computer Science and Informatics, Dublin Institute of Technology, University College Dublin, Dublin, Ireland. Dublin, Ireland.
[email protected] [email protected]
Abstract
some aspects of the grounding of language in nonlanguage. The conception of space underlying spatial locatives is fundamentally relativistic: the location of one object is specified relative to another whose location is usually assumed by the speaker to be known by the hearer. Moreover, underpinning this relativistic notion of space is the concept of proximity. Consequently, the notion of proximity is an important concept at the core of human spatial cognition. Proximal spatial relationships are often described using topological prepositions, e.g. at, on, near, etc. Terminology In this paper we use the term target (T) to refer to the head of a locative expression (the object which is being located by that expression) and the term landmark (L) to refer to the object in the prepositional phrase in that expression (relative to which the head’s location is described), see Example (1).
The paper examines how people’s judgements of proximity between two objects are influenced by the presence of a third object. In an experiment participants were presented with images containing three shapes in different relative positions, and asked to rate the acceptability of a locative expression such as ‘the circle is near the triangle’ as descriptions of those images. The results showed an interaction between the relative positions of objects and the linguistic roles that those objects play in the locative expression: proximity was a decreasing function of the distance between the head object in the expression and the prepositional clause object, and an increasing function the distance between the head and the third, distractor object. This finding leads us to a new account for the semantics of spatial prepositions such as near.
1
Example 1. [The man]T near [the table]L .
Introduction
In this paper, we present an empirical study of the cognitive representations underpinning the uses of proximal descriptions in locative spatial expressions. A spatial locative expression consists of a locative prepositional phrase together with whatever the phrase modifies (noun, clause, etc.). In their simplest form, a locative expression consists of a prepositional phrase modifying a noun phrase, for example the man near the desk. People often use spatial locatives to denote objects in a visual scene. Understanding such references involves coordination between a perceptual event and a linguistic utterance. Consequently, the study of spatial locatives affords the opportunity to examine
We will use the term distractor to describe any object in the visual context that is neither the landmark nor the target. Contributions The paper reports on a psycholinguistic experiment that examines proximity. Previous psycholinguistic work on proximal relations, (Logan and Sadler, 1996), has not examined the effects other objects in the scene (i.e., distractors) may have on the spatial relationship between a landmark and a target. The experiment described in this paper compares peoples’ judgements of proximity between target and landmark objects when they are presented alone and when there are presented along with other distractor objects. Based on the results of this experiment we
Proceedings of the Third ACL-SIGSEM Workshop on Prepositions, pages 1–8, c Trento, Italy, April 2006. 2006 Association for Computational Linguistics
1
propose a new model for the semantics of spatial prepositions such as near. Overview In §2 we review previous work. In §3 we describe the experiment. In §4 we present the results of the experiment and our analysis. The paper finishes with conclusions, §5.
2
Related Work
In this section we review previous psycholinguistic experiments that examined proximal spatial relations. We then present example spatial contexts, that the previous experiments did not examine, which motivate the hypothesis tested in this paper: the location of other objects in a scene can interfere with the acceptability of a proximal description being used to describe the spatial relationship between a landmark and a target.
1.74
1.90
2.84
3.16
2.34
1.81
2.13
2.61
3.84
4.66
4.97
4.90
3.56
3.26
4.06
5.56
7.55
7.97
7.29
4.80
3.91
4.47
5.91
8.52
O
7.90
6.13
4.46
3.47
4.81
6.94
7.56
7.31
5.59
3.63
3.25
4.03
4.50
4.78
4.41
3.47
3.10
1.84
2.23
2.03
3.06
2.53
2.13
2.00
Figure 1: 7-by-7 cell grid with mean goodness ratings for the relation near as a function of the position occupied by X. Spatial reasoning is a complex activity that involves at least two levels of representation and reasoning: a geometric level where metric, topological, and projective properties are handled, (Herskovits, 1986); and a functional level where the normal function of an entity affects the spatial relationships attributed to it in context (for example, the meaning of ‘near’ for a bomb is quite different from the meaning of ‘near’ for other objects of the same size; (Vandeloise, 1991; Coventry, 1998; Garrod et al., 1999)). There has been a lot of experimental work done on spatial reasoning and language: (Carlson-
2
Radvansky and Irwin, 1993; Carlson-Radvansky and Irwin, 1994; Hayward and Tarr, 1995; Gapp, 1995; Logan and Sadler, 1996; CarlsonRadvansky and Logan, 1997; Coventry, 1998; Garrod et al., 1999; Regier and Carlson, 2001; Kelleher and Costello, 2005). Of these only (Logan and Sadler, 1996) examined topological prepositions in a context where functional factors were excluded. The term spatial template denotes the representation of the regions of acceptability associated with a preposition. It is centred on the landmark and identifies for each point in its space the acceptability of the spatial relationship between the landmark and the target appearing at that point being described by the preposition (Logan and Sadler, 1996). The concept of a spatial template emerged from psycholinguistic experiments reported in (Logan and Sadler, 1996). These experiments examined various spatial prepositions. In these experiments, a human subject was shown sentences, each with a picture of a spatial configuration. Every sentence was of the form “The X is [relation] the O”. The accompanying picture contained an O in the center of an invisible 7-by-7 cell grid, and an X in one of the 48 surrounding positions. The subject then had to rate how well the sentence described the picture, on a scale from 1(bad) to 9(good). Figure 1 gives the mean goodness rating for the relation “near to” as a function of the position occupied by the X, as reported in (Logan and Sadler, 1996). If we plot the mean goodness rating for “near” against the distance between target X and landmark O, we get the graph in Figure 2.
Figure 2: Mean goodness rating vs. distance between X and O. Both the figure and the graph make it clear that the ratings diminish as we increase the distance between X and O. At the same time, we can observe that even at the extremes of the grid the ratings were still above 1 (the minimum rating). In-
deed, in the four corners of the grid, the points most distant from the landmark, the mean ratings nearly average twice the minimum rating. However in certain contexts other factors, apart from the distance between the landmark and the target, affect the applicability of a proximal relation as a description of the target’s position relative to the landmark. For example, consider the two scenes (side-view) given in Figure 3. In the scene on the left-hand side, we can use the description “the blue box is near the black box” to describe object (a). However, consider now the scene on the right-hand side. In this context, the description “the blue box is near the black box” seems inappropriate as an expression describing (a). The placing of object (c) beside (b) would appear to interfere with the appropriateness of using a proximal relation to locate (a) relative to (b), even though the absolute distance between (a) and (b) has not changed.
Figure 3: Proximity and distance In summary, there is empirical evidence that indicates that as the distance between the landmark and the target increases the applicability of a proximal description decreases. Furthermore, there is anecdotal evidence that the location of other distractor objects in context may interfere with applicability of a proximal description between a target and landmark object. The experiment presented in this paper is designed to empirically test the affect of distractor objects on proximity judgements.
3
Experiment
This work examines the impact of distractor objects on subjects’ judgment of proximity between the target and the landmark objects. To do this, we examine the changes in participants judgements of the appropriateness of the topological preposition near being used to describe a spatial configuration of the target and landmark objects when a distractor object was present and when it was removed. Topological prepositions (e.g., at, on, in, near) are often used to describe proximal spatial relationships. However, the semantics of a given topological preposition also reflects functional (Garrod et al., 1999), directional (Logan and Sadler, 1996)
3
and topological factors.1 Consequently, it was important to control for these factors during the design of the experiment. Functional factors were controlled for by using simple shapes in the stimuli. The preposition near was used to control the impact of directional factors. Previous psycholinguistic work indicated that near was not affected by any directional preferences. Finally, the influence of topological factors was controlled for by ensuring that the landmark and target maintained a consistent topological relationship (the objects never touched, overlapped or were contained in other objects). We approached our experiment with expectation that people’s proximity judgments between a target and a landmark will be a decreasing function of the distance between those two objects: the smaller the distance between a landmark and a target object, the higher the proximity rating people will give for those two objects. We expect that the presence of a distractor object will also influence proximity judgments, and examine two different hypotheses about how that influence will work: a target-centered hypothesis and landmark-centered hypothesis. In the target-centered hypothesis, people’s judgments of proximity between a target and a landmark will be a decreasing function of distance between those two objects, but an increasing function of distance between the target and the distractor object. Under this hypothesis, if the distractor object is near the target object, this will interfere with and lower people’s judgments of proximity between the target and the landmark. In the landmark-centered hypothesis, by contrast, people’s judgments of proximity between a target and a landmark will be a decreasing function of distance between those objects, but an increasing function of distance between the landmark and the distractor object. Under this hypothesis, if the distractor object is near the landmark, it will interfere with and lower people’s judgments of proximity between target and landmark. We test these hypotheses by varying target-distractor distance in our materials, but maintaining landmarkdistractor distance constant. If the target-centered hypothesis is correct, then people’s judgments of proximity should vary with target-distractor distance. If the landmark-centered hypothesis is correct, target-distractor distance should not influence 1 See (Cohn et al., 1997) for a description different topological relationships.
people’s judgments of proximity. 3.1
Material and Subjects
All images used in this experiment contained a central landmark and a target. In most of the images there was also another object, which we will refer to as the distractor. All of these objects were coloured shapes, a circle, triangle or square. However, none of the images contained two objects that were the same shape or the same colour.
1
2
3
4
5
a b
6 g
c
L f
d e
Figure 4: Relative locations of landmark (L) target positions (1..6) and distractor positions (a..g) in images used in the experiment. The landmark was always placed in the middle of a seven by seven grid (row four, column four). There were 48 images in total, divided into 8 groups of 6 images each. Each image in a group contained the target object placed in one of 6 different cells on the grid, numbered from 1 to 6 (see Figure 4). As Figure 4 shows, we number these target positions according to their nearness to the landmark. Each group, then, contains images with targets at positions 1, 2, 3, 4, 5 and 6. Groups are organised according to the presence and position of a distractor object. Figure 4 shows the 7 different positions used for the distractor object, labelled a,b,c,d,e,f and g. In each of these positions the distractor is equidistant from the landmark. In group a the distractor is directly above the landmark, in group b the distractor is rotated 45 degrees clockwise from the vertical, in group c it is directly to the right of the landmark, in d is rotated 135 degrees clockwise from the vertical, and so
4
on. Notice that some of these distractor positions (b,d, and f ) are not aligned with the grid. This realignment is necessary to ensure that the distractor object is always the same distance from the landmark. Each of these groups of images used in the experiment corresponds to one of these 7 distractor positions, with a distractor object occurring at that position for every image in that group. In addition, there is an eight group (which we label as group x), in which no distractor object occurs. Previous studies of how people judge proximity have typically examined judgments where the target is above, below, to the left or right of the landmark. The results of these studies showed that these distinctions are relatively unimportant, and the gradient of proximity observed tends to be symmetrical around the landmark. For this reason, in our study we ignore these factors and present landmark, target and distractor randomly rotated (so that some participants in our experiment will see the image with target at position 1 and distractor at position a in a rotated form where position 1 is below the landmark and position a is to the right of the landmark, but others will see the same relative positions at different rotations). In each image all objects present were placed exactly at the center of the cell representing their position. During the experiment, each image was displayed with a sentence of the form The is near the . The blanks were filled with a description of the target and landmark respectively. The sentence was presented under the image. 12 participants took part in this experiment. 3.2
Procedure
There were 48 trials, constructed from the following variables: 8 distractor conditions * 6 target positions. To avoid sequence effects the landmark, target and distractor colour and shape were randomly modified for each trial and the distractor condition and target location were randomly selected for each trial. Each trial was randomly reflected across the horizontal, vertical, or diagonal axes. Trials were presented in a different random order to each participant. Participants were instructed that they would be shown sentence-picture pairs and were be asked to rate the acceptability of the sentence as a description of the picture using a 10-point scale, with zero denoting not acceptable at all; four or five denoting moderately acceptable; and nine perfectly ac-
sults obtained from other groups. In particular, we compare the results from this group with those obtained from groups c, d and e: the three groups in which the distractor object is furthest from the set of target positions used (as Figure 4 shows, distractor positions c, d, and e are all on the opposite side of the landmark from the set of target positions). We focus on comparison with groups c, d, and e because results for the other groups are complicated by the fact that people’s proximity judgments are influenced by the closeness of a distractor object to the target (as we will see later). Figure 5: Experiment instructions. 9 8 7
proximity rating
ceptable. Figure 5 presents the instructions given to each participant before the experiment. Trials were self-paced, and the experiments lasted about 25-30 minutes. Figure 6 illustrates how the trials were presented.
6
group x group c group d group e
5 4 3 2
4
Results and Discussion
1 0
There are two questions we want to ask in our examination of people’s proximity judgments in the presence of distractor objects. First, does the presence of a distractor make any noticable difference in people’s judgements of proximity? Second, if the presence of a distractor does influence proximity judgements, is this influence target-centered (based on the distance between the target object and the distractor) or landmark-centered (based on the distance between the landmark and the distractor). We address the first question (does the distractor object have an influence on proximity judgments) by comparing the results obtained for images in group x (in which there was no distractor) with re-
Figure 6: Sample trial from the experiment.
5
1
2
3
4
5
6
target location
Figure 7: mean proximity rating for target locations for group x (no distractor) and groups c, d, and e (distractors present behind landmark) Figure 7 shows the average proximity rating given by participants for the 6 targets 1 to 6 for group x (in which there was no distractor object) and for groups c, d, and e (in which distractors occurred on the opposite side of the landmark from the target). Clearly, all three sets of distractor responses are very similar to each other, and are all noticably different from the no-distractor response. This difference was shown to be statistically significant in a by-subjects analysis comparing subjects’ responses for groups c,d and e with their responses for group x. This comparison showed that subjects produced significantly lower proximity ratings for group c than group x (Wilcoxon signed-rank test W + = 55.50, W − = 10.50, N = 11, p PRED in < OBJ > PART − FORM ein PCASE DIR OBL DIR PSEM + PRED PRO OBJ CASE acc PRED in < OBJ > PCASE LOC PSEM + ADJ " # PRED Erde OBJ SPEC der CASE dat
Figure 8: sickert [ PP in der Erde ]DAT ein ’soaks (through) the soil’
4.2.3 Lexical Entries and Grammar Rules In the f-structure in Figure 7 the pleonastic PP is subcategorized by the particle verb. Figure 9 shows the corresponding lexical entry for the verb. To prevent a locative PP in the dative from filling in the object position of the verb argument the lexical entry specifies that the object has to be assigned accusative case. einsickern
V
(↑ PRED) = ’einsickern’
VP →
(↑ OBL DIR:PART-FORM) = ein (↑ OBL DIR:OBJ:CASE)
ein- historically is derived from the preposition in and regarding its semantic features is comparable to the other two-way prepositions where particle and preposition have the same lexical form. The attributes PSEM and PCASE are added to the representation of the verb particles in Berman and Frank (1996). They are derived from the attribute set for prepositions, indicating the analogy in the semantics of particle and preposition. PSEM always has the value ’+’ for particle verbs formed by spatial prepositions, because they always have a semantic content. The attribute PCASE expresses the directionality in the semantics of the verb particle ( (↑ PCASE) = DIR). The predicate of the particle licences an object and behaves like a directional preposition. However, the object position is not lexically filled and therefore is assigned the predicate value ’PRO’. We also want to model the behaviour of the particle verb governing a locative PP in the dative (Figure 8). The lexical entry of the particle verb (Figure 9) explicitly requires accusative case assignment and prevents the locative dative PP from filling in the object position of the verb argument. The locative dative PP is attached to the adjunct set in the grammar rule shown in Figure 11. 3 V PP *
= acc (PP
Figure 9: Lexical entry for einsickern ’to soak’ However, as shown in example (4) the pleonastic PP can be omitted. In this case the argument OBL DIR subcategorized by the particle verb is provided by the particle ein- whose lexical entry is given in Figure 10. ein PART (↑ PRED)
= ’in’
(↑ PART-FORM)
= ein
(↑ PCASE) (↑ PSEM)
= DIR =+
(↑ OBJ PRED )
= PRO
Figure 10: Lexical entry for the particle ein In contrast to (Berman and Frank, 1996), in our representation the particle is assigned the PRED value ’in’ in the lexicon. The cause for the divergence between the lexical form of the particle and its PRED value is due to the fact that the particle
63
PART
↑=↓
↓ ( ↑ ADJ)
(↓ OBJ CASE) 6= acc
(↑ OBL DIR) = ↓)
(↑ OBL DIR) = ↓.
Figure 11: Grammar Rule specifying restrictions on particle verbs with pleonastic PPs The first PP in the grammar rule models the behaviour of a particle verb combining with one or more locative PPs in the dative. The constraint (↓ OBJ CASE) 6= acc ensures that this part of the rule will not be applied to a pleonastic PP with accusative case assignment.4 The second PP in the grammar rule captures a pleonastic PP in the accusative. The restriction that this PP has to be in the accusative is specified in the lexical entry for the particle verb (Figure 10). The last part of the rule expresses that the verb particle PART is also mapped to the OBL DIR ar3
For expository purposes we use a simple VP rather than a topological analysis. 4 The Kleene * notation indicates zero or more occurences of PP.
gument of the complex verb and so is able to saturate the argument structure of the verb. The formalisation in Figure 8 and 9 is consistent with the analysis that the particle has an implicit reference object which is identical to the object of a pleonastic PP in the accusative, but not to the object of a dative PP. The formalisation gives an adequate description of the behaviour of particle verbs in Group C, but it does not suppress the licencing of a pleonastic accusative PP for verbs in Group B which combine with locative PPs in the dative only. This problem is solved through the specification of a constraint (=c) in the lexical entries for all particle verbs in Group B (Figure 12). vorfahren V (↑ PRED) = ’vorfahren’
References
(↑ OBL DIR:PART-FORM) = vor (↑ OBL DIR:OBJ:CASE)
= acc
(↑ OBL DIR:OBJ:PRED)
=c PRO
(Group B). Here the verb particle saturates the directional OBL DIR argument required by the verb. Group C verbs allow both accusative and dative PPs. Only particle verbs governing PPs in the accusative are pleonastic, but the PP either modifies or adds new information to the inherent argument structure of the particle verb and therefore is not suppressed by the verb particle. Our formalisation describes the behaviour of particle verbs concerning their ability to licence pleonastic PPs. The semantic criteria restricting the behaviour of the particle verbs are embedded into the LFG representation and enable us to model the semantic differences on a syntactic level.
Figure 12: Lexical entry for vorfahren ’to drive up’ (Group B) The constraint checks that the predicate of the object in the OBL DIR f-structure is instantiated with the value ’PRO’. For all cases where the predicate is lexically realised, the constraint fails and thus the interpretation of pleonastic accusative PPs in the OBL DIR position for Group B verbs is suppressed.
5 Conclusions The aim of this paper is to explain the behaviour of German particle verbs formed by two-way prepositions and their ability to combine with pleonastic PPs. A classification of particle verbs based on semantic criteria was given, illustrating the restrictions imposed on their behaviour. It was shown that particle verbs occurring only with accusative PPs (Group A) always have a directional reading including the intrusion of the theme referent into a region specified by the relatum. Particle verbs which can not combine with an accusative PP (Group B) either have a static, nondirectional reading or describe a directed movement where the referent already may be present in the region specified by the relatum. Syntactically this results in the fact that the accusative PP is able to saturate the argument OBL DIR subcategorized by the particle verbs in Group A. The dative PP functions as an adjunct
64
Judith Berman and Anette Frank. 1996. Deutsche und franz¨osische Syntax im Formalismus der LFG. Max Niemeyer Verlag, T¨ubingen. Joan Bresnan. Blackwell.
2000.
Lexical-Functional Syntax.
Miriam Butt, Tracy Holloway King, Mar´ıa-Eugenia Nino and Fr´ed´erique Segond. 1999. A Grammar Writer’s Cookbook. CSLI Publications, Stanford, California. Mary Dalrymple. 2001. Syntax and Semantics. Lexical Functional Grammar, volume 34. Academic Press, San Diego, California. Junji Okamoto. 2002. Particle-Bound Directions in German Particle Verb Constructions. Projektbericht V: Typological Investigation of Languages and Cultures of the East and West. (Part II). Susan Olsen. 1998. Semantische und konzeptuelle Aspekte der Partikelverbbildung mit ein-. Stauffenburg, T¨ubingen. James Witt. 1998. Kompositionalit¨at und Regularit¨at, In: Olsen, Susan (ed). Semantische und konzeptuelle Aspekte der Partikelverbbildung mit ein-. Stauffenburg, T¨ubingen. Dieter Wunderlich. 1983. On the Compositionality of German Prefix Verbs. In: R. B¨auerle, Ch. Schwarze and A. von Stechow (eds.) Meaning, Use and Interpretation of Language. de Gruyter, Berlin.
Automatic Identification of English Verb Particle Constructions using Linguistic Features Su Nam Kim and Timothy Baldwin Department of Computer Science and Software Engineering University of Melbourne, Victoria 3010 Australia {snkim,tim}@csse.unimelb.edu.au
Abstract This paper presents a method for identifying token instances of verb particle constructions (VPCs) automatically, based on the output of the RASP parser. The proposed method pools together instances of VPCs and verb-PPs from the parser output and uses the sentential context of each such instance to differentiate VPCs from verb-PPs. We show our technique to perform at an F-score of 97.4% at identifying VPCs in Wall Street Journal and Brown Corpus data taken from the Penn Treebank.
1
Introduction
Multiword expressions (hereafter MWEs) are lexical items that can be decomposed into multiple simplex words and display lexical, syntactic and/or semantic idiosyncracies (Sag et al., 2002; Calzolari et al., 2002). In the case of English, MWEs are conventionally categorised syntacticosemantically into classes such as compound nominals (e.g. New York, apple juice, GM car), verb particle constructions (e.g. hand in, battle on), non-decomposable idioms (e.g. a piece of cake, kick the bucket) and light-verb constructions (e.g. make a mistake). MWE research has focussed largely on their implications in language understanding, fluency and robustness (Pearce, 2001; Sag et al., 2002; Copestake and Lascarides, 1997; Bannard et al., 2003; McCarthy et al., 2003; Widdows and Dorow, 2005). In this paper, our goal is to identify individual token instances of English verb particle constructions (VPCs hereafter) in running text. For the purposes of this paper, we follow Baldwin (2005) in adopting the simplifying assumption that VPCs: (a) consist of a head verb and a unique prepositional particle (e.g. hand in, walk off); and (b) are either transitive (e.g. hand in, put on) or intransitive (e.g. battle on). A defining characteristic of transitive VPCs is that they can gen-
erally occur with either joined (e.g. He put on the sweater) or split (e.g. He put the sweater on) word order. In the case that the object is pronominal, however, the VPC must occur in split word order (c.f. *He handed in it) (Huddleston and Pullum, 2002; Villavicencio, 2003). The semantics of the VPC can either derive transparently from the semantics of the head verb and particle (e.g. walk off ) or be significantly removed from the semantics of the head verb and/or particle (e.g. look up); analogously, the selectional preferences of VPCs can mirror those of their head verbs or alternatively diverge markedly. The syntax of the VPC can also coincide with that of the head verb (e.g. walk off ) or alternatively diverge (e.g. lift off ). In the following, we review relevant past research on VPCs, focusing on the extraction/identification of VPCs and the prediction of the compositionality/productivity of VPCs. There is a modest body of research on the identification and extraction of VPCs. Note that in the case of VPC identification we seek to detect individual VPC token instances in corpus data, whereas in the case of VPC extraction we seek to arrive at an inventory of VPC types/lexical items based on analysis of token instances in corpus data. Li et al. (2003) identify English VPCs (or “phrasal verbs” in their parlance) using handcoded regular expressions. Baldwin and Villavicencio (2002) extract a simple list of VPCs from corpus data, while Baldwin (2005) extracts VPCs with valence information under the umbrella of deep lexical acquisition.1 The method of Baldwin (2005) is aimed at VPC extraction and takes into account only the syntactic features of verbs. In this paper, our interest is in VPC identification, and we make use of deeper semantic information. In Fraser (1976) and Villavicencio (2006) it is argued that the semantic properties of verbs can determine the likelihood of their occurrence with 1 The learning of lexical items in a form that can be fed directly into a deep grammar or other richly-annotated lexical resource
Proceedings of the Third ACL-SIGSEM Workshop on Prepositions, pages 65–72, c Trento, Italy, April 2006. 2006 Association for Computational Linguistics
65
particles. Bannard et al. (2003) and McCarthy et al. (2003) investigate methods for estimating the compositionality of VPCs based largely on distributional similarity of the head verb and VPC. O’Hara and Wiebe (2003) propose a method for disambiguating the verb sense of verb-PPs. While our interest is in VPC identification—a fundamentally syntactic task—we draw on the shallow semantic processing employed in these methods in modelling the semantics of VPCs relative to their base verbs. The contribution of this paper is to combine syntactic and semantic features in the task of VPC identification. The basic intuition behind the proposed method is that the selectional preferences of VPCs over predefined argument positions,2 should provide insight into whether a verb and preposition in a given sentential context combine to form a VPC (e.g. Kim handed in the paper) or alternatively constitute a verb-PP (e.g. Kim walked in the room). That is, we seek to identify individual preposition token instances as intransitive prepositions (i.e. prepositional particles) or transitive particles based on analysis of the governing verb. The remainder of the paper is structured as follows. Section 2 outlines the linguistic features of verbs and their co-occuring nouns. Section 3 provides a detailed description of our technique. Section 4 describes the data properties and the identification method. Section 5 contains detailed evaluation of the proposed method. Section 6 discusses the effectiveness of our approach. Finally, Section 7 summarizes the paper and outlines future work.
(1) put = place EX :
ARGS :
Linguistic Features
2 Focusing exclusively on the subject and object argument positions. 3 All sense definitions are derived from WordNet 2.1.
66
verb-PP
(2) put on = wear EX : Put on the sweater . ARGS :
sweater OBJ = garment, clothing
ANALYSIS :
verb particle construction
While put on is generally used in the context of something, it usually occurs with clothingtype nouns such as sweater and coat, whereas the simplex put has less sharply defined selectional restrictions and can occur with any noun. In terms of the word senses of the head nouns of the object NPs, the VPC put on will tend to co-occur with objects which have the semantics of clothes or garment. On the other hand, the simplex verb put in isolation tends to be used with objects with the semantics of object and prepositional phrases containing NPs with the semantics of place. Also, as observed above, the valence of a VPC can differ from that of the head verb. (3) and (4) illustrate two different senses of take off with intransitive and transitive syntax, respectively. Note that take cannot occur as a simplex intransitive verb. wearing
(3) take off = lift off The airplane takes off.
ARGS :
When verbs co-occur with particles to form VPCs, their meaning can be significantly different from the semantics of the head verb in isolation. According to Baldwin et al. (2003), divergences in VPC and head verb semantics are often reflected in differing selectional preferences, as manifested in patterns of noun co-occurrence. In one example cited in the paper, the cosine similarity between cut and cut out, based on word co-occurrence vectors, was found to be greater than that between cut and cut off, mirroring the intuitive compositionality of these VPCs. (1) and (2) illustrate the difference in the selectional preferences of the verb put in isolation as compared with the VPC put on.3
book OBJ = book, publication, object
ANALYSIS :
EX :
2
Put the book on the table.
airplaneSUBJ = airplane, aeroplane
ANALYSIS :
verb particle construction
(4) take off = remove EX :
They take off the cape .
ARGS :
theySUBJ = person, individual capeOBJ = garment, clothing verb particle construction
ANALYSIS :
Note that in (3), take off = lift off co-occurs with a subject of the class airplane, aeroplane. In (4), on the other hand, take off = remove and the corresponding object noun is of class garment or clothing. From the above, we can see that head nouns in the subject and object argument positions can be used to distinguish VPCs from simplex verbs with prepositional phrases (i.e. verb-PPs).
3
Approach
Our goal is to distinguish VPCs from verb-PPs in corpus data, i.e. to take individual inputs such as Kim handed the paper in today and tag each as either a VPC or a verb-PP. Our basic approach is to parse each sentence with RASP (Briscoe and Carroll, 2002) to obtain a first-gloss estimate of the VPC and verb-PP token instances, and also identify the head nouns of the arguments of each VPC and simplex verb. For the head noun of each subject and object, as identified by RASP, we use WordNet 2.1 (Fellbaum, 1998) to obtain the word sense. Finally we build a supervised classifier using TiMBL 5.1 (Daelemans et al., 2004).
raw text corpus
RASP parser
Preprocessing
Verbs Particles
v+p with Semantics e.g. take_off := [.. put_on := [.. look_after := [..
Subjects Objects
WordNet
Word Senses
TiMBL Classifier
3.1 Method Compared to the method proposed by Baldwin (2005), our approach (a) tackles the task of VPC identification rather than VPC extraction, and (b) uses both syntactic and semantic features, employing the WordNet 2.1 senses of the subject and/or object(s) of the verb. In the sentence He put the coat on the table, e.g., to distinguish the VPC put on from the verb put occurring with the prepositional phrase on the table, we identify the senses of the head nouns of the subject and object(s) of the verb put (i.e. he and coat, respectively). First, we parse all sentences in the given corpus using RASP, and identify verbs and prepositions in the RASP output. This is a simple process of checking the POS tags in the most-probable parse, and for both particles (tagged RP) and transitive prepositions (tagged II) reading off the governing verb from the dependency tuple output (see Section 3.2 for details). We also retrieved the head nouns of the subject and object(s) of each head verb directly from the dependency tuples. Using WordNet 2.1, we then obtain the word sense of the head nouns. The VPCs or verb-PPs are represented with corresponding information as given below: P (type|v, p, ws SUBJ , ws DOBJ , ws IOBJ ) where type denotes either a VPC or verb-PP, v is the head verb, p is the preposition, and ws * is the word sense of the subject, direct object or indirect object. Once all the data was gathered, we separated it into test and training data. We then used TiMBL 5.1 to learn a classifier from the training data, which was then run and evaluated over the test data. See Section 5 for full details of the results. Figure 1 depicts the complete process used to distinguish VPCs from verb-PPs.
67
Figure 1: System Architecture 3.2 On the use of RASP, WordNet and TiMBL RASP is used to identify the syntactic structure of each sentence, including the head nouns of arguments and first-gloss determination of whether a given preposition is incorporated in a VPC or verb-PP. The RASP output contains dependency tuples derived from the most probable parse, each of which includes a label identifying the nature of the dependency (e.g. SUBJ, DOBJ), the head word of the modifying constituent, and the head of the modified constituent. In addition, each word is tagged with a POS tag from which it is possible to determine the valence of any prepositions. McCarthy et al. (2003) evaluate the precision of RASP at identifying VPCs to be 87.6% and the recall to be 49.4%. However the paper does not evaluate the parser’s ability to distinguish sentences containing VPCs and sentences with verb-PPs. To better understand the baseline performance of RASP, we counted the number of false-positive examples tagged with RP and false-negative examples tagged with II, relative to gold-standard data. See Section 5 for details. We use WordNet to obtain the first-sense word sense of the head nouns of subject and object phrases, according to the default word sense ranking provided within WordNet. McCarthy et al. (2004) found that 54% of word tokens are used with their first (or default) sense. With the performance of current word sense disambiguation (WSD) systems hovering around 60-70%, a simple first-sense WSD system has room for improvement, but is sufficient for our immediate purposes
in this paper. To evaluate our approach, we built a supervised classifier using the TiMBL 5.1 memorybased learner and training data extracted from the Brown and WSJ corpora.
4
Data Collection
We evaluated out method by running RASP over Brown Corpus and Wall Street Journal, as contained in the Penn Treebank (Marcus et al., 1993).
Group A Group B Group C Group D
Group A RP tagged data
Group B Group C
Group D
RP & II tagged data
II tagged data
FNR — — 10.15% 3.4%
Agreement 95.24% 99.61% 93.27% 99.20%
Table 1: False positive rate (FPR), false negative rate (FNR) and inter-annotator agreement across the four groups of token instances
4.1 Data Classification The data we consider is sentences containing prepositions tagged as either RP or II. Based on the output of RASP, we divide the data into four groups:
FPR 4.08% 3.96% — —
Group A Group B Group C Total
f ≥1 VPC V-PP 5,223 0 1,312 0 0 995 6,535 995
f ≥5 VPC V-PP 3,787 0 1,108 0 0 217 4,895 217
Table 2: The number of VPC and verb-PP token instances occurring in groups A, B and C at varying frequency cut-offs
Group A contains the verb–preposition token instances tagged tagged exclusively as VPCs (i.e. the preposition is never tagged as II in combination with the given head verb). Group B contains the verb–preposition token instances identified as VPCs by RASP where there were also instances of that same combination identified as verb-PPs. Group C contains the verb–preposition token instances identified as verb-PPs by RASP where there were also instances of that same combination identified as VPCs. Finally, group D contains the verb-preposition combinations which were tagged exclusively as verb-PPs by RASP. We focus particularly on disambiguating verb– preposition token instances falling into groups B and C, where RASP has identified an ambiguity for that particular combination. We do not further classify token instances in group D, on the grounds that (a) for high-frequency verb–preposition combinations, RASP was unable to find a single instance warranting a VPC analysis, suggesting it had high confidence in its ability to correctly identify instances of this lexical type, and (b) for lowfrequency verb–preposition combinations where the confidence of there definitively no being a VPC usage is low, the token sample is too small to disambiguate effectively and the overall impact would be negligible even if we tried. We do, however, return to considered data in group D in computing the precision and recall of RASP. Naturally, the output of RASP parser is not error-free, i.e. VPCs may be parsed as verb-PPs
68
and vice versa. In particular, other than the reported results of McCarthy et al. (2003) targeting VPCs vs. all other analyses, we had no a priori sense of RASP’s ability to distinguish VPCs and verb-PPs. Therefore, we manually checked the false-positive and false-negative rates in all four groups and obtained the performance of parser with respect to VPCs. The verb-PPs in group A and B are false-positives while the VPCs in group C and D are false-negatives (we consider the VPCs to be positive examples). To calculate the number of incorrect examples, two human annotators independently checked each verb–preposition instance. Table 1 details the rate of false-positives and false-negative examples in each data group, as well as the inter-annotator agreement (calculated over the entire group). 4.2 Collection We combined together the 6,535 (putative) VPCs and 995 (putative) verb-PPs from groups A, B and C, as identified by RASP over the corpus data. Table 2 shows the number of VPCs in groups A and B and the number of verb-PPs in group C. The first number is the number of examples occuring at least once and the second number that of examples occurring five or more times. From the sentences containing VPCs and verbPPs, we retrieved a total of 8,165 nouns, including
Type common noun personal pronoun demonstrative pronoun proper noun who which No sense (what)
Groups A&B 7,116 629 127 156 94 32 11
Group C 1,239 79 1 18 6 0 0
Table 3: Breakdown of subject and object head nouns in group A&B, and group C pronouns (e.g. I, he, she), proper nouns (e.g. CITI, Canada, Ford) and demonstrative pronouns (e.g. one, some, this), which occurred as the head noun of a subject or object of a VPC in group A or B. We similarly retrieved 1,343 nouns for verb-PPs in group C. Table 3 shows the distribution of different noun types in these two sets. We found that about 10% of the nouns are pronouns (personal or demonstrative), proper nouns or WH words. For pronouns, we manually resolved the antecedent and took this as the head noun. When which is used as a relative pronoun, we identified if it was coindexed with an argument position of a VPC or verb-PP, and if so, manually identified the antecedent, as illustrated in (5). (5)
EX :
Tom likes the books which he sold off.
ARGS :
heSUBJ = person whichOBJ = book
With what, on the other hand, we were generally not able to identify an antecedent, in which case the argument position was left without a word sense (we come back to this in Section 6). (6)
Tom didn’t look up what to do. What went on?
We also replaced all proper nouns with corresponding common noun hypernyms based on manual disambiguation, as the coverage of proper nouns in WordNet is (intentionally) poor. The following are examples of proper nouns and their common noun hypernyms: Proper noun CITI Canada Ford Smith
Common noun hypernym bank country company human
Sense 1 apple edible fruit(1st) produce, green goods, ... food(3rd) ... fruit(2nd) reproductive structure ... pome, false fruit fruit reproductive structure Sense 1 orange citrus, citrus fruit, citrous fruit edible fruit(2nd) produce, green goods, ... food(4th) ... fruit(3rd) ..
Figure 2: Senses of apple and orange When we retrieved the first word sense of nouns from WordNet, we selected the first sense and the associated hypernyms (up to) three levels up the WordNet hierarchy. This is intended as a crude form of smoothing for closely-related word senses which occur in the same basic region of the WordNet hierarchy. As an illustration of this process, in Figure 2, apple and orange are used as edible fruit, fruit or food, and the semantic overlap is picked up on by the fact that edible fruit is a hypernym of both apple and orange. On the other hand, food is the fourth hypernym for orange so it is ignored by our method. However, because we use the four senses, the common senses of nouns are extracted properly. This approach works reasonably well for retrieving common word senses of nouns which are in the immediate vicinity of each other in the WordNet hierarchy, as was the case with apple and orange. In terms of feature representation, we generate an individual instance for each noun sense generated based on the above method, and in the case that we have multiple arguments for a given VPC or verb-PP (e.g. both a subject and a direct object), we generate an individual instance for the cross product of all sense combinations between the arguments. We use 80% of the data for training and 20% for testing. The following is the total number of training instances, before and after performing hypernym expansion:
Group A Group B Group C
69
Training Instances Before expansion After expansion 5,223 24,602 1,312 4,158 995 5,985
Group B BA BC BAC
Frequency of VPCs (f≥1 ) (f≥5 ) (f≥1 & f≥1 ) (f≥5 & f≥5 ) (f≥1 & f≥1 ) (f≥5 & f≥1 ) (f≥1 & f≥1 & f≥1 ) (f≥5 & f≥5 & f≥1 )
Size test:272 train:1,040 test:1,327 train:4,163 test:498 train:1,809 test:1,598 train:5,932
Data RASP B BA BC BAC
Table 4: Data set sizes at different frequency cutoffs
5
Freq f≥1 f≥1 f≥5 f≥1 f≥1 f≥5 f≥5 f≥1 f≥1 f≥5 f≥1 f≥1 f≥1 f≥1 f≥5 f≥5 f≥1
Evaluation
We selected 20% of the test data from different combinations of the four groups and over the two frequency thresholds, leading to a total of 8 test data sets. The first data set contains examples from group B only, the second set is from groups B and A, the third set is from groups B and C, and the fourth set is from groups B, A and C. Additionally, each data set is divided into: (1) f ≥ 1, i.e. verb–preposition combinations occurring at least once, and (2) f ≥ 5, i.e. verb–preposition combinations occurring at least five times (hereafter, f ≥ 1 is labelled f≥1 and f ≥ 5 is labelled f≥5 ). In the group C data, there are 217 verb-PPs with f≥5 , which is slightly more than 20% of the data so we use verb-PPs with f≥1 for experiments instead of verb-PP with f≥5 . The first and second data sets do not contain negative examples while the third and fourth data sets contain both positive and negative examples. As a result, the precision for the first two data sets is 1.0. Table 5 shows the precision, recall and F-score of our method over each data set, relative to the identification of VPCs only. A,B,C are groups and f# is the frequency of examples. Table 6 compares the performance of VPC identification and verb-PP identification. Table 7 indicates the result using four word senses (i.e. with hypernym expansion) and only one word sense (i.e. the first sense only).
6 Discussion The performance of RASP as shown in Tables 5 and 6 is based on human judgement. Note that we only consider the ability of the parser to distinguish sentences with prepositions as either VPCs or verb-PPs (i.e. we judge the parse to be correct if the preposition is classified correctly, irrespective of whether there are other errors in the output).
70
P .959 1.0 1.0 1.0 1.0 .809 .836 .962 .964
R .955 .819 .919 .959 .962 .845 .922 .962 .983
F .957 .901 .957 .979 .980 .827 .877 .962 .974
Table 5: Results for VPC identification only (P = precision, R = recall, F = F-score) Data RASP BC BAC
Freq f≥1 f≥1 f≥1 f≥5 f≥1 f≥1 f≥1 f≥5 f≥1
Type P+V P+V P+V P+V P+V
P .933 .8068 .8653 .8660 .9272
R – .8033 .8529 .8660 .8836
F – .8051 .8591 .8660 .9054
Table 6: Results for VPC (=V) and verb-PP (=P) identification (P = precision, R = recall, F = Fscore) Also, we ignore the ambiguity between particles and adverbs, which is the principal reason for our evaluation being much higher than that reported by McCarthy et al. (2003). In Table 5, the precision (P) and recall (R) for VPCs are computed as follows: =
Data Correctly Tagged as VPCs Data Retrieved as VPCs
R =
Data Correctly Tagged as VPCs All VPCs in Data Set
P
The performance of RASP in Table 6 shows how well it distinguishes between VPCs and verbPPs for ambiguous verb–preposition combinations. Since Table 6 shows the comparative performance of our method between VPCs and verbPPs, the performance of RASP with examples which are misrecognized as each other should be the guideline. Note, the baseline RASP accuracy, based on assigning the majority class to instances in each of groups A, B and C, is 83.04%. In Table 5, the performance over highfrequency data identified from groups B, A and C is the highest (F-score = .974). In general, we would expect the data set containing the high frequency and both positive and negative examples
f≥1
P
f≥5
V
f≥5
P
# 4WS 1WS 4WS 1WS 4WS 1WS 4WS 1WS
P .962 .958 .769 .800 .964 .950 .889 .813
R .962 .969 .769 .743 .983 .973 .783 .614
F .962 .963 .769 .770 .974 .962 .832 .749
100
100
80
80
60
60
Types
Type V
Error Rate Reduction (%)
Freq f≥1
40
40
20
20
0
Table 7: Results with hypernym expansion (4WS) and only the first sense (1WS), in terms of precision (P), recall (R) and F-score (F) to give us the best performance at VPC identification. We achieved a slightly better result than the 95.8%-97.5% performance reported by Li et al. (2003). However, considering that Li et al. (2003) need considerable time and human labour to generate hand-coded rules, our method has advantages in terms of both raw performance and labour efficiency. Combining the results for Table 5 and Table 6, we see that our method performs better for VPC identification than verb-PP identification. Since we do not take into account the data from group D with our method, the performance of verb-PP identification is low compared to that for RASP, which in turn leads to a decrement in the overall performance. Since we ignored the data from group D containing unambiguous verb-PPs, the number of positive training instances for verb-PP identification was relatively small. As for the different number of word senses in Table 7, we conclude that the more word senses the better the performance, particularly for higher-frequency data items. In order to get a clearer sense of the impact of selectional preferences on the results, we investigated the relative performance over VPCs of varying semantic compositionality, based on 117 VPCs (f≥1 ) attested in the data set of McCarthy et al. (2003). According to our hypothesis from above, we would expect VPCs with low compositionality to have markedly different selectional preferences to the corresponding simplex verb, and VPCs with high compositionality to have similar selectional preferences to the simplex verb. In terms of the performance of our method, therefore, we would expect the degree of compositionality to be inversely proportional to the system performance. We test this hypothesis in Figure 3, where we calculate the error rate reduction (in F-score)
71
0
1
2
3
4
5
6
Compositionality
7
8
9
10
0
Figure 3: Error rate reduction for VPCs of varying compositionality for the proposed method relative to the majorityclass baseline, at various degrees of compositionality. McCarthy et al. (2003) provides compositionality judgements from three human judges, which we take the average of and bin into 11 categories (with 0 = non-compositional and 10 = fully compositional). In Figure 3, we plot both the error rate reduction in each bin (both the raw numbers and a smoothed curve), and also the number of attested VPC types found in each bin. From the graph, we see our hypothesis born out that, with perfect performance over non-compositional VPCs and near-baseline performance over fully compositional VPCs. Combining this result with the overall results from above, we conclude that our method is highly successful at distinguishing non-compositional VPCs from verb-PPs, and further that there is a direct correlation between the degree of compositionality and the similarity of the selectional preferences of VPCs and their verb counterparts. Several factors are considered to have influenced performance. Some data instances are missing head nouns which would assist us in determining the semantics of the verb–preposition combination. Particular examples of this are imperative and abbreviated sentences: (7)
a. Come in. b. (How is your cold?) Broiled out.
Another confounding factor is the lack of word sense data, particularly in WH questions: (8)
a. What do I hand in? b. You can add up anything .
7
Conclusion
In this paper, we have proposed a method for identifying VPCs automatically from raw corpus data. We first used the RASP parser to identify VPC and verb-PP candidates. Then, we used analysis of the head nouns of the arguments of the head verbs to model selectional preferences, and in doing so, distinguish between VPCs and verb-PPs. Using TiMBL 5.1, we built a classifier which achieved an F-score of 97.4% at identifying frequent VPC examples. We also investigated the comparative performance of RASP at VPC identification. The principal drawback of our method is that it relies on the performance of RASP and we assume a pronoun resolution oracle to access the word senses of pronouns. Since the performance of such systems is improving, however, we consider our approach to be a promising, stable method of identifying VPCs. Acknowledgements This material is based upon work supported in part by the Australian Research Council under Discovery Grant No. DP0663879 and NTT Communication Science Laboratories, Nippon Telegraph and Telephone Corporation. We would like to thank the three anonymous reviewers for their valuable input on this research.
References Timothy Baldwin and Aline Villavicencio. 2002. Extracting the unextractable: A case study on verb-particles. In Proc. of the 6th Conference on Natural Language Learning (CoNLL-2002), pages 98–104, Taipei, Taiwan. Timothy Baldwin, Colin Bannard, Takaaki Tanaka, and Dominic Widdows. 2003. An empirical model of multiword expression decomposability. In Proc. of the ACL-2003 Workshop on Multiword Expressions: Analysis, Acquisition and Treatment, pages 89–96, Sapporo, Japan. Timothy Baldwin. 2005. The deep lexical acquisition of English verb-particle constructions. Computer Speech and Language, Special Issue on Multiword Expressions, 19(4):398–414. Colin Bannard, Timothy Baldwin, and Alex Lascarides. 2003. A statistical approach to the semantics of verbparticles. In Proc. of the ACL-2003 Workshop on Multiword Expressions: Analysis, Acquisition and Treatment, pages 65–72, Sapporo, Japan. Ted Briscoe and John Carroll. 2002. Robust accurate statistical annotation of general text. In Proc. of the 3rd International Conference on Language Resources and Evaluation (LREC 2002), pages 1499–1504, Las Palmas, Canary Islands. Nicoletta Calzolari, Charles Fillmore, Ralph Grishman, Nancy Ide, Alessandro Lenci, Catherine MacLeod, and Antonio Zampolli. 2002. Towards best practice for multiword expressions in computational lexicons. In Proc. of the 3rd International Conference on Language Resources and Evaluation (LREC 2002), pages 1934–40, Las Palmas, Canary Islands.
72
Ann Copestake and Alex Lascarides. 1997. Integrating symbolic and statistical representations: The lexicon pragmatics interface. In Proc. of the 35th Annual Meeting of the ACL and 8th Conference of the EACL (ACL-EACL’97), pages 136–43, Madrid, Spain. Walter Daelemans, Jakub Zavrel, Ko van der Sloot, and Antal van den Bosch. 2004. TiMBL: Tilburg Memory Based Learner, version 5.1, Reference Guide. ILK Technical Report 04-02. Christiane Fellbaum, editor. 1998. WordNet: An Electronic Lexical Database. MIT Press, Cambridge, USA. B. Fraser. 1976. The Verb-Particle Combination in English. The Hague: Mouton. Rodney Huddleston and Geoffrey K. Pullum. 2002. The Cambridge Grammar of the English Language. Cambridge University Press, Cambridge, UK. Wei Li, Xiuhong Zhang, Cheng Niu, Yuankai Jiang, and Rohini K. Srihari. 2003. An expert lexicon approach to identifying English phrasal verbs. In Proc. of the 41st Annual Meeting of the ACL, pages 513–20, Sapporo, Japan. Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: the Penn treebank. Computational Linguistics, 19(2):313–30. Diana McCarthy, Bill Keller, and John Carroll. 2003. Detecting a continuum of compositionality in phrasal verbs. In Proc. of the ACL-2003 Workshop on Multiword Expressions: Analysis, Acquisition and Treatment, Sapporo, Japan. Diana McCarthy, Rob Koeling, Julie Weeds, and John Carroll. 2004. Finding predominant senses in untagged text. In Proc. of the 42nd Annual Meeting of the ACL, pages 280–7, Barcelona, Spain. Tom O’Hara and Janyce Wiebe. 2003. Preposition semantic classification via Treebank and FrameNet. In Proc. of the 7th Conference on Natural Language Learning (CoNLL2003), pages 79–86, Edmonton, Canada. Darren Pearce. 2001. Synonymy in collocation extraction. In Proceedings of the NAACL 2001 Workshop on WordNet and Other Lexical Resources: Applications, Extensions and Customizations, Pittsburgh, USA. Ivan A. Sag, Timothy Baldwin, Francis Bond, Ann Copestake, and Dan Flickinger. 2002. Multiword expressions: A pain in the neck for NLP. In Proc. of the 3rd International Conference on Intelligent Text Processing and Computational Linguistics (CICLing-2002), pages 1–15, Mexico City, Mexico. Aline Villavicencio. 2003. Verb-particle constructions and lexical resources. In Proc. of the ACL-2003 Workshop on Multiword Expressions: Analysis, Acquisition and Treatment, pages 57–64, Sapporo, Japan. Aline Villavicencio. 2006. Verb-particle constructions in the world wide web. In Patrick Saint-Dizier, editor, Computational Linguistics Dimensions of Syntax and Semantics of Prepositions. Springer, Dordrecht, Netherlands. Dominic Widdows and Beate Dorow. 2005. Automatic extraction of idioms using graph analysis and asymmetric lexicosyntactic patterns. In Proc. of the ACL-SIGLEX 2005 Workshop on Deep Lexical Acquisition, pages 48– 56, Ann Arbor, USA.
On the prepositions which introduce an adjunct of duration
Frank Van Eynde
Abstract
state of Silvia’s living in Paris held, but it can also denote a period which started three years ago and which includes the time of utterance (Kamp and Reyle, 1993, 567). The relevance of this distinction is clear from the fact that there are languages which use different prepositions for both interpretations. Italian, for instance, employs the preposition per in the translation of (1), (3) and the first interpretation of (4), whereas it employs da in the translation of the second interpretation of (4).
This paper deals with the prepositions which introduce an adjunct of duration, such as the English for and in. On the basis of both crosslingual and monolingual evidence these adjuncts are argued to be ambiguous between a floating and an anchored interpretation. To capture the distinction in formal terms I employ the framework of H EAD - DRIVEN P HRASE S TRUCTURE G RAMMAR, enriched with a number of devices which are familiar from D ISCOURSE R EPRESENTA TION T HEORY . The resulting analysis is demonstrated to be relevant for machine translation, natural language generation and natural language understanding.
(5) Maria suon`o il pianoforte per un’ora. (6) Laura star`a per due mesi nell’Ohio. (7) Silvia ha abitato per tre anni a Parigi. (8) Silvia abita a Parigi da tre anni. For ease of reference I will call the adjuncts in (1), (3), (4a), (5), (6) and (7) floating: they denote a stretch of time whose position on the time line is not defined. The adjuncts in (4b) and (8), by contrast, will be called anchored, since their position on the time line is fixed: their right boundary is supplied by the time of utterance. As illustrated in (9-10), the right boundary can also be supplied by a temporal adjunct, such as a PP[a] or a subordinate S[quando].
1 A typology of PP adjuncts of duration In many languages the adjuncts of duration take different forms depending on the aspectual class of the VP which they modify. In English, for instance, they are introduced by for if the VP denotes a state or a process and by in if the VP denotes an accomplishment. (1) Maria played the piano for an hour. (2) Anna wrote that letter in half an hour.
(9) A quel punto Silvia abitava da tre at that point Silvia lived for three anni a Parigi. years in Paris
Orthogonal to this distinction, there is another one, which can be made explicit by comparing (1) and (3) with (4).
‘By that time Silvia had lived in Paris for three years.’
(3) Laura will stay in Ohio for two months. (4) Silvia has lived in Paris for three years.
(10) Laura sar`a nell’ Ohio da due mesi, Laura will-be in Ohio for two months, quando verr`a raggiunta da Ivo. when she-will-be joined by Ivo
The adjuncts in (1) and (3) unambiguously specify the duration of Maria’s activity of playing the piano and of Laura’s stay in Ohio. The adjunct in (4), however, triggers an ambiguity: it can denote any three-year period in the past in which the
‘Laura will have been in Ohio for two months when Ivo will join her.’
Proceedings of the Third ACL-SIGSEM Workshop on Prepositions, pages 73–80, c Trento, Italy, April 2006. 2006 Association for Computational Linguistics
73
(17) Spesso suonavo il flauto per un’ ora. often I-played the flute for an hour ‘I often played the flute for an hour.’
The distinction between floating and anchored adjuncts is also relevant for the PP[in] adjuncts. To show this let us compare (2) and (11) with (12).
(18) Lea abita sempre a Roma da tre anni. Lea lives always in Rome for three years ‘Lea has lived in Rome for three years.’
(11) Pablo makes such a drawing in less than five minutes. (12) Leo will tune your piano in an hour.
The PP[per] in (17) is in the scope of the quantifying spesso ‘often’: there are several one hour periods of my playing the flute. By contrast the PP[da] in (18) outscopes the quantifying sempre ‘always’, yielding an interpretation in which Lea’s living in Rome is said to go on uninterruptedly for a period of three years. The same contrast can be observed in sentences with VP negation.
In (2) and (11) the PP[in] adjuncts are unambiguously floating, but (12) is ambiguous: it can either mean that it will take Leo one hour to tune your piano or that he will start the activity of tuning your piano in an hour from now. In the first interpretation, the adjunct is floating, as in (2) and (11), but in the second one it is anchored: the beginning of the hour which will pass before Leo starts tuning the piano is supplied by the time of utterance. The relevance of the distinction is, again, brought out by the Italian equivalents. While the floating PP adjuncts are introduced by in, as in the translation of (2), (11) and (12a), the anchored ones are introduced by fra, as in the tanslation of (12b).1
(19) Non suon`o il flauto per un’ ora. not played the flute for an hour ‘(S)he did not play the flute for an hour.’ (20) Non suona il pianoforte da un’ ora. not plays the piano for an hour ‘(S)he has not been playing the piano for an hour now.’
(13) Anna ha scritto quella lettera in mezz’ora.
The floating PP[per] in (19) is in the scope of the negation, yielding an interpretation which can be paraphrased as ‘it is not the case that (s)he played the flute for an hour’. The anchored PP[da] in (20), by contrast, outscopes the negation, yielding an interpretation which can be paraphrased as ‘for an hour it has not been the case that (s)he plays the piano’. To capture the semantic properties of the four types of durational adjuncts we need a framework for the analysis and representation of temporal expressions. As a starting point, I will use the HPSG framework, as defined in (Pollard and Sag, 1994) and (Ginzburg and Sag, 2000). This suffices to spell out what the four types have in common (section 2), but in order to also model what differentiates them (section 4) we will need some extensions to the standard HPSG ontology and notation (section 3).
(14) Pablo fa un disegno come quello in meno di cinque minuti. (15) Leo accorder`a il tuo pianoforte in un’ora. (16) Leo accorder`a il tuo pianoforte fra un’ora. The following table provides a summary of the data discussed so far.
EN IT EN IT
floating PP[for] (1,3,4a) PP[per] (5,6,7) PP[in] (2,11,12a) PP[in] (13,14,15)
anchored PP[for] (4b) PP[da] (8,9,10) PP[in] (12b) PP[fra/tra] (16)
The distinction between floating and anchored adjuncts is relevant for Machine Translation and for Natural Language Generation, since it conditions the choice of the preposition. At the same time, it is also relevant for Natural Language Understanding, since it bears on the issue of scope. More specifically, while the floating adjuncts can be in the scope of a VP quantifier, the anchored ones cannot.
2 What the durational adjuncts have in common
1
Instead of fra one also finds tra. The choice is mainly conditioned by phonological factors. To avoid alliteration speakers tend to prefer fra when (one of) the surrounding words start with t(r), and tra when (one of) the surrounding words start with f(r).
74
Since the adjuncts of duration are modifiers rather than arguments, they are not selected by their head sister. Instead, it is the head which is selected by the adjunct. Phrased in terms of the HPSG notation, a PP adjunct has a SELECT feature whose
value spells out the syntactic and semantic properties of its head sister.2
HEAD prep
(21)
SELECT
COMPS
synsem
(23)
(22)
sem-object rel t-unit-rel day-rel
prep HEAD SELECT CAT HEAD verb COMPS HEAD noun CAT COMPS COMPS MARK indefinite CONTENT 23
In words, the prepositions which introduce a durational adjunct take an NP complement and project a PP which modifies a VP. Besides these properties, which they share with many other types of PP adjuncts, there is the more specific requirement that the NP complement must denote an amount of time. This is modeled in terms of its MARK ( ING ) and its CONTENT values. Starting with the latter and employing the semantic ontology of (Ginzburg and Sag, 2000), in which the CONTENT value of a nominal is an object of type scope-object, the relevant constraint can be defined as follows: 2
For reasons which are given in (Van Eynde, 2005), I do not employ separate selection features for the modifiers and the specifiers. The SELECT attribute, hence, generalizes over the MOD ( IFIED ) and SPEC ( IFIED ) attributes of (Pollard and Sag, 1994). 3 I ignore the distinction between local and non-local properties. CAT is, hence, short for LOCAL CATEGORY and CON TENT for LOCAL CONTENT .
scope-object INDEX index RESTR t-unit-rel INST
In words, the index of the complement must be the argument of a predicate of type t-unit-rel. This is one of the intermediate types in the hierarchy of relations. Its subtypes are the predicates which express temporal units.
Since the SELECT feature is part of the HEAD value, it is shared between the PP and the preposition. From this it follows that prepositions which introduce an adjunct can impose constraints on the SYNSEM value of the phrase which the adjunct modifies. Exploiting this possibility we can capture the syntactic properties of the prepositions which introduce a durational adjunct in terms of the following AVM.3
75
year-rel
loc-rel
...
...
The defining property of temporal units is that they succeed one another without interruption. A day, for instance, is a temporal unit, since it is immediately followed by another day, but a Friday is not, since it is not immediately followed by another Friday. The relevance of this distinction is illustrated by the fact that for a day and in ten minutes can be adjuncts of duration, whereas for Friday and in April cannot. Whether a PP[for/in] can be used as an adjunct of duration is not only determined by the semantic class of the noun, but also by the prenominals: for every day and in that month, for instance, cannot be used as adjuncts of duration. This is captured by the constraint that the NP must be indefinite, rather than universal or determinate. Evidence for making this threefold distinction and for modeling it in terms of the MARK ING values is provided in (Van Eynde, 2005).4 A crucial factor in the semantic analysis of the durational adjuncts is their contribution to the meaning of the VP: the amount of time which is denoted by their NP daughter must somehow be related to the semantic properties of the VPs which they modify. To spell this out we first need a format for the semantic analysis of verbal objects. 4
If the NP is determinate, as in for the last five years, for the whole morning and da lunedi ‘since Monday’, it does not denote an amount of time, but an interval or an instant. Such PPs have meanings which resemble those of the durational adjuncts, but their contribution to the semantics of the VP is nonetheless different.
3 Times and temporal objects
projections, including the stative ones. The index is invariably the first argument of the relation which the verb denotes, as in greet-rel (i, x, y), and is linked to the V-time by means of the locrel relation. The function of this relation is to link the denotation of the V(P) to the time at which it holds. It is comparable to the overlap relation, familiar from Discourse Representation Theory: t. Since the temporal objects belong to a subtype of scope-obj, it follows that their indices are of the same type as those of the nominal objects. Given the ontology of (Ginzburg and Sag, 2000), this implies that they contain features for person, number and gender.5
To model the semantic properties of verbal projections I extend the semantic ontology of (Ginzburg and Sag, 2000) with times and temporal objects. sem-object time interval
scope-object instant
temp-obj
...
The temporal objects belong to a subtype of scope-object and, hence, have an index and a set of restrictions on that index. Besides, they have a TIMES attribute, which takes a list of times as its value.
temp-obj
(24)
TIMES
list time
(27)
NUMBER GENDER
BEGIN
END
EXTENT
instant instant scope-obj
Since the value of the EXTENT feature is of type scope-obj it can be identified with the amount of time which is expressed in an adjunct of duration. Of the various times which figure in the list of a temporal object, the rightmost one has a special role, since it is the one which is linked to the index of the verb. For ease of reference I will call it the V-time.
temp-obj
(26)
INDEX index loc-rel RESTR INST TIME
TIMES
(29) To make mistakes is/*are/*am human. Since the form is requires a subject with a third person singular index, it follows that the nonfinite VPs in (28) and (29) have a third person singular index, and since phrases share their index with their head daughter, this implies in turn that the verbs forging and make have a third person singular index.6 To avoid misunderstanding, it is worth stressing that this does not mean that they require a third person singular subject, but rather that they
list time
time
person number gender
(28) Forging banknotes is/*are/*am not easy.
The presence of these features in the CONTENT values of verbs may, at first, seem awkward, since they model properties which are typical of N(P)s. A form like greets, for instance, requires its NP subject to have a third person singular index, but does not have a third person singular index of its own, as argued in (Pollard and Sag, 1994, 82). Looking closer, though, the assignment of these features to verbs does have a number of advantages. One is that it accounts for the agreement in clauses with a verbal subject.
interval
index
PERSON
The objects of type time denote temporal entities and come in two kinds: instants and intervals. This distinction does not concern any inherent properties of the temporal entities, but rather their mode of individuation. The objects of type interval have a beginning, an end and a duration. (25)
5 The values of these features concern the mode of indivduation of a nominal’s referent and should not be confused with properties of the referent itself. A person, for instance, can be individuated by means of a second person pronoun, but this does not mean that (s)he has the property of being a second person. 6 That forging and make are verbs is clear from the fact that they take NP complements; if they were nouns, they would take PP[of ] complements.
The verb’s index ( ) is comparable to a Davidsonian event variable, but has a slightly different role. It is, for instance, not only assigned to verbs and VPs which denote an event, but to all verbal
76
themselves are third person singular. This distinction is especially relevant for the finite verbs, as illustrated by (30) and (31).
ABILITY feature to the objects of type number, adopting a proposal of (Van Eynde, 2005).
(32)
(30) That he/she snores is/*are/*am annoying. (31) That I/they snore is/*are/*am annoying.
number COUNTABILITY
countability
Its values are:
Also here the subjects are required to have a third person singular index, and since they are clauses which are headed by a finite verb, it follows that the finite verbs have a third person singular index. Moreover, this index is different from the one of their subject. Snore in (31), for instance, has a third person singular index, but requires its subject to have an index which is plural or nonthird person. In sum, one advantage of the assignment of a third person singular index to verbs is that it accounts in a straightforward manner for the agreement data in (28-31). Another advantage is that the indices provide a way to capture the distinction between the aspectual classes (Aktionsarten). To see this, let us first revisit the role of the indices in nominal objects. As argued in (Pollard and Sag, 1994), the indices are not only useful to model agreement between a finite verb and its subject, or between an anaphoric pronoun and its antecedent, but also between a determiner and its nominal head. The demonstrative these, for instance, requires a nominal with a plural index, whereas this requires a nominal with a singular index. A similar constraint holds for the combination of a quantifying determiner and its head. While every and a require their nominal head to be singular and count, much requires it to be singular and mass: every/a/*much table vs. much/*every/*a traffic. Despite the obvious similarity with the constraints for the demonstrative determiners, they cannot be modeled in terms of the indices of (Pollard and Sag, 1994), since their indices do not contain any information about the mass/count distinction. A natural move, therefore, is to redefine the indices in such a way that this distinction can be integrated. Independent evidence for this move is provided by the fact that the mass/count distinction concerns the mode of individuation of the referent(s) of the nominal, rather than an inherent property of the referent(s), see footnote 5. Another piece of evidence is the fact that the mass/count distinction closely interacts with the NUMBER distinction: most of the relevant constraints simultaneously concern a number and a mass/count value. To model this I add a COUNT-
countability bounded
unbounded
In terms of this dichotomy the count nouns have bounded indices, whereas the mass nouns have unbounded indices. Nouns which are used either way, such as glass, have the underspecified value in their lexical entry; this can be resolved by the addition of a determiner, as in a glass or much glass. Returning now to the verbs, it automatically follows from the presence of an index in their CON TENT values that they also have a COUNTABIL ITY feature. This is an advantage, since it provides us with the means to spell out the similarities between the count/mass distinction for nominals and the Aktionsart distinction for verbal projections. The states and processes, for instance, share the property of the mass nouns that their denotation is unbounded, whereas the accomplishments and the achievements share the property of the count nouns that their denotation is bounded (Bach, 1986). Exploiting the potential of this extended role of the indices I introduce a distinction between two types of temporal objects. The bounded ones have an index of type bounded and are subsumed by the following constraint: (33)
bd-temp-obj INDEX index NUM C bounded in-rel RESTR INST TIME TIMES list time time
In words, the index of a bounded temporal object is temporally included in the V-time. Since inclusion is a special type of overlap, this is a more constrained version of (26). It corresponds to DRT’s ‘e t’. The unbounded temporal objects obviously have an index of type unbounded, but the relation
77
piano went on for at least an hour. The generic loc-rel is not sufficient for this purpose, since it only requires overlap: it would make (1) true if the playing went on for five minutes. For the floating PP[for] and PP[per] adjuncts there is nothing which need be added to (34). Their anchored counterparts, however, are subsumed by one further constraint.8
of this index to the corresponding time is not subject to any further constraints; it is subsumed by the generic loc-rel. With the introduction of times, temporal objects and the boundedness distinction we have paved the way for a more detailed analysis of the various types of durational adjuncts.
4 What differentiates the four types of durational adjuncts
unbd-temp-obj INDEX index incl-rel H S C RESTR INST TIME TIMES ..., EXTENT COMPS
CONTENT
H S C is short for HEAD SELECT abbreviation is used in (36) and (38).
CONX C - IND UTT- TIME
instant
sem-object
The restricton to unbounded temporal objects accounts for the fact that these adjuncts combine with states and processes, but not with accomplishments or achievements. Notice, though, that this restriction does not exclude the combination with VPs whose CONTENT value is the underspecified temp(oral)-obj(ect). This is important, since few V(P)s are inherently bounded or unbounded. It is usually by the addition of an adjunct that the underspecification gets resolved. That the adjunct specifies the duration of the Vtime is illustrated by the examples of the first section. In (1), for instance, it is the time of playing the piano which is said to take an hour, and in (3) it is the time of Laura’s stay in Ohio which is said to have a length of two months. The relation between this time and the index of the V(P) is required to be the one of inclusion (s t). This accounts for the fact that (1) is only true if the playing of the 7
TIMES
temp-rel TIME TIME ..., END
In words, the interval whose duration is specified has a right boundary ( ) which is related to the time of utterance. This relation can be the one of identity, as in (5b) and (8), or it can be mediated by a temporal adjunct. In (9), for instance, the right boundary is specified by the PP a quel punto, which precedes the time of utterance, and in (10) it is specified by the clause quando verra` raggiunta da Ivo, which follows the time of utterance. To capture this variation I use the relation temp-rel. This stands for any binary relation between times.9
RESTR
SS ... C
The PP[for/per/da] adjuncts select a VP which denotes an unbounded temporal object and specify the duration of the V-time.7
The PP[for/per/da] adjuncts
(34)
I first discuss the adjuncts which combine with an unbounded temporal object, and then the adjuncts which combine with a bounded temporal object. In the last paragraph I return to the issue of scope. 4.1
(35)
rel loc-rel in-rel
temp-rel
incl-rel
m-rel
...
f-rel
As demonstrated in (Allen, 1984), the number of distinct binary relations between times is limited. He distinguishes seven basic relations: equal (=), before ( ), during (d), meets (m), overlaps (o), starts (s) and finishes (f). Each of these relations has an inverse: the one of before, for instance, is after ( ). This yields fourteen possible relations, but since equality is indistinguishable from its inverse, the number of distinct relations is 13. Of these 13 relations, only three are 8
is short CONX C - IND CONTEXT CONTEXTUAL - INDICES and SS SYNSEM CATEGORY HEAD SELECT CONTENT. 9
CONTENT. The same
... C
for for
In this respect, temp-rel and its subtypes contrast with loc-rel and its subtypes, which are relations between an index and a time.
78
(39) The bomb will explode in two minutes.
exemplified by (8-10), but most of the remaining ones are excluded by the constraint in (35) that the related times must be instants. This automatically excludes the relations in which at least one of the times must be an interval, such as overlap, during, start, finish and their respective inverses. 4.2
Since the beginning of an interval necessarily precedes its end, the V-time ( ) must follow the time of utterance. This accounts for the fact that the English PP[in] can have the anchored interpretation in a clause with a future auxiliary, such as (12) and (39), or in a clause with a futurate present tense, such as we are leaving in a minute, but not in a clause with a past tense verb, such as (2), or in a clause with a non-futurate present tense, such as (11).
The PP[in/fra/tra] adjuncts of duration
The floating PP[in] adjuncts select a VP which denotes a bounded temporal object and specify the duration of the V-time.
(36)
H S C
bd-temp-obj ... ,
TIMES
COMPS
EXTENT
CONTENT
4.3
Having spelled out the properties of the anchored adjuncts we can now account for the fact that they cannot be outscoped by a VP quantifier. What makes this impossibe is the fact that the interval whose duration they specify is linked to the time of utterance. The link can be more or less direct, but it does not allow for the intrusion of other intervening intervals. The floating adjuncts, by contrast, apply to intervals which are not linked to the time of utterance and, therefore, allow the intrusion of intervening times, as in (17), where spesso ‘often’ outscopes per un’ ora ‘for an hour’. Of course, the fact that the floating adjuncts can be outscoped by a VP quantifier does not imply that they must be outscoped whenever there is such a quantifier. To show this let us have a look at (40).
Since only intervals can have duration this constraint accounts for the fact that these adjuncts are not compatible with VPs which denote instantaneous events, as in: (37) ? The bomb exploded in two minutes. In contrast to their floating counterparts, the anchored PP[in] and PP[tra/fra] adjuncts do not specify the duration of the V-time, but rather of the interval which elapses between the time of utterance and the beginning of the V-time. In terms of Allen’s ontology, this can be characterized as an instance of m(eets)-rel: m(x, y) is true if and only if x immediately precedes y.10
(38)
R
H S C
SS CAT
COMPS
m-rel
TIME TIME
T BG , EX
CONTENT
CONX C - IND UTT- TIME
(40) We will train two hours a day for at least six months. While the adjunct two hours a day specifies the duration and the frequency of the V-time, i.e. the time of the individual training sessions, the PP[for] adjunct specifies the duration of the period in which the daily training sessions will take place.11 It, hence, outscopes the VP quantifier. This use of the adjunct is not covered by the analysis in section 4.1, since the latter only deals with those adjuncts which specify the duration of the V-time. To deal with the adjunct in (40) we would have to extend the hierarchy of temporal objects with a special subtype for the quantified temporal objects and add a constraint which captures the
instant
The leftmost interval is the one whose duration is specified. The rightmost time can be an instant or an interval. In (16) it is most likely an interval, since the tuning of a piano is bound to take some time, but it can also be an instant, as in the most plausible interpretation of (39). 10
R is short for RESTR ( ICTIONS ), T GINNING and EX for EXTENT .
Scope
11
The floating nature of the PP[for] adjunct is clear from the absence of a specification (implicit or explicit) of its right boundary and from the fact that its Italian equivalent is per almeno sei mesi rather than da almeno sei mesi.
for TIMES, BG for BE -
79
properties of the durational adjuncts which combine with such objects. Spelling this out is left for future work.
5 Conclusion The adjuncts of duration require an analysis in terms of two mutually independent distinctions. One concerns the aspectual class of the modified VP and is widely acknowledged as relevant. The other concerns the distinction between floating and anchored interpretations and is often ignored; its relevance, though, is clear from both crosslingual and monolingual data. For the analysis of the four types of durational adjuncts I have employed an extended version of HPSG. The extensions mainly concern the addition of times and temporal objects to the semantic ontology and the notation. The resulting analysis captures both the similarities and the differences between the four types of adjuncts, and provides an account for the fact that the floating adjuncts can be outscoped by a VP quantifier, whereas the anchored ones cannot.
References J. F. Allen. 1984. Towards a general theory of action and time. Artificial Intelligence, 23:123–154. E. Bach. 1986. The algebra of events. Linguistics and Philosophy, 9:5–16. J. Ginzburg and I. Sag. 2000. Interrogative Investigations. CSLI, Stanford. H. Kamp and U. Reyle. 1993. From Discourse to Logic. Kluwer Academic Publishers, Dordrecht. C. Pollard and I. Sag. 1994. Head-driven Phrase Structure Grammar. CSLI Publications and University of Chicago Press, Stanford/Chicago. F. Van Eynde. 2005. NP-internal agreement and the structure of the noun phrase. Journal of Linguistics, 42:1–47.
80
How bad is the problem of PP-attachment? A comparison of English, German and Swedish Martin Volk Stockholm University Department of Linguistics 106 91 Stockholm, Sweden
[email protected]
Abstract
languages (such as Dutch (Vandeghinste, 2002) or Swedish (Aasa, 2004)) have followed. In the PP attachment research for other languages there is often a comparison of the disambiguation accuracy with the English results. But are the results really comparable across languages? Are we starting from the same baseline when working on PP attachment in structurally similar languages like English, German and Swedish? Is the problem of PP attachment equally bad (equally frequent and of equal balance) for these three languages? These are the questions we will discuss in this paper. In order to find answers to these questions we have taken a closer look at the training and test data used in various experiments. And we have queried the most important treebanks for the three languages under investigation.
The correct attachment of prepositional phrases (PPs) is a central disambiguation problem in parsing natural languages. This paper compares the baseline situation in English, German and Swedish based on manual PP attachments in various treebanks for these languages. We argue that cross-language comparisons of the disambiguation results in previous research is impossible because of the different selection procedures when building the training and test sets. We perform uniform treebank queries and show that English has the highest noun attachment rate followed by Swedish and German. We also show that the high rate in English is dominated by the preposition of. From our study we derive a list of criteria for profiling data sets for PP attachment experiments.
1
2
Introduction
Any computer system for natural language processing has to struggle with the problem of ambiguities. If the system is meant to extract precise information from a text, these ambiguities must be resolved. One of the most frequent ambiguities arises from the attachment of prepositional phrases (PPs). Simply stated, a PP that follows a noun (in English, German or Swedish) can be attached to the noun or to the verb. In the last decade various methods for the resolution of PP attachment ambiguities have been proposed. The seminal paper by (Hindle and Rooth, 1993) started a sequence of studies for English. We investigated similar methods for German (Volk, 2001; Volk, 2002). Recently other
Background
(Hindle and Rooth, 1993) did not have access to a large treebank. Therefore they proposed an unsupervised method for resolving PP attachment ambiguities. And they evaluated their method against 880 English triples verb-noun-preposition (V-N-P) which they had extracted from randomly selected, ambiguously located PPs in a corpus. For example, the sentence ”Timex had requested duty-free treatment for many types of watches” results in the V-N-P triple (request, treatment, for). These triples were manually annotated by both authors with either noun or verb attachment based on the complete sentence context. Interestingly, 586 of these triples (67%) were judged as noun attachments and only 33% as verb attachments. And (Hindle and Rooth, 1993) reported on 80% attachment accuracy, an improvement of 13% over the baseline (i.e. guessing noun attachment in all
Proceedings of the Third ACL-SIGSEM Workshop on Prepositions, pages 81–88, c Trento, Italy, April 2006. 2006 Association for Computational Linguistics
81
cases). A year later (Ratnaparkhi et al., 1994) published a supervised approach to the PP attachment problem. They had extracted quadruples V-N-P-N1 (plus the accompanying attachment decision) from both an IBM computer manuals treebank (about 9000 tuples) and from the Wall Street Journal (WSJ) section of the Penn treebank (about 24,000 tuples). The latter tuple set has been reused by subsequent research, so let us focus on this one.2 (Ratnaparkhi et al., 1994) used 20,801 tuples for training and 3097 tuples for evaluation. They reported on 81.6% correct attachments. But have they solved the same problem as (Hindle and Rooth, 1993)? What was the initial bias towards noun attachment in their data? It turns out that their training set (the 20,801 tuples) contains only 52% noun attachments, while their test set (the 3097 tuples) contains 59% noun attachments. The difference in noun attachments between these two sets is striking, but (Ratnaparkhi et al., 1994) do not discuss this (and we also do not have an explanation for this). But it makes obvious that (Ratnaparkhi et al., 1994) were tackling a problem different from (Hindle and Rooth, 1993) given the fact that their baseline was at 59% guessing noun attachment (rather than 67% in the Hindle and Rooth experiments).3 Of course, the baseline is not a direct indicator of the difficulty of the disambiguation task. We may construct (artificial) cases with low baselines and a simple distribution of PP attachment tendencies. For example, we may construct the case that a language has 100 different prepositions, where 50 prepositions always introduce noun attachments, and the other 50 prepositions always require verb attachments. If we also assume that both groups occur with the same frequency, we have a 50% baseline but still a trivial disambiguation task. But in reality the baseline puts the disambiguation result into perspective. If, for instance, the baseline is 60% and the disambiguation result is 80% correct attachments, then we will claim that our disambiguation procedure is useful. Whereas 1
The V-N-P-N quadruples also contain the head noun of the NP within the PP. 2 The Ratnaparkhi training and test sets were later distributed together with a development set of 4039 V-N-P-N tuples. 3 It should be noted that important subsequent research, e.g. by (Collins and Brooks, 1995; Stetina and Nagao, 1997), used the Ratnaparkhi data sets and thus allowed for good comparability.
82
if we have a baseline of 80% and the disambiguation result is 75%, then the procedure can be discarded. So what are the baselines reported for other languages? And is it possible to use the same extraction mechanisms for V-N-P-N tuples in order to come to comparable baselines? We did an in-depth study on German PP attachment (Volk, 2001). We compiled our own treebank by annotating 3000 sentences from the weekly computer journal ComputerZeitung. We had first annotated a larger number of subsequent sentences with Part-of-Speech tags, and based on these PoS tags, we selected 3000 sentences that contained at least one full verb plus the sequence of a noun followed by a preposition. After annotating the 3000 sentences with complete syntax trees we used a Prolog program to extract V-N-P-N tuples with the accompanying attachment decisions. This lead to 4562 tuples out of which 61% were marked as noun attachments. We used the same procedure to extract tuples from the first 10,000 sentences of the NEGRA treebank. This resulted in 6064 tuples with 56% noun attachment (for a detailed overview see (Volk, 2001) p. 86). Again we observe a substantial difference in the baseline. When our student J¨orgen Aasa worked on replicating our German experiments for Swedish, he used a Swedish treebank from the 1980s for the extraction of test data. He extracted V-N-P-N tuples from SynTag, a treebank with 5100 newspaper sentences built by (J¨arborg, 1986). And Aasa was able to extract 2893 tuples out of which 73.8% were marked as noun attachments (Aasa, 2004) (p. 25). This was a surprisingly high figure, and we wondered whether this indicated a tendency in Swedish to avoid the PP in the ambiguous position unless it was to be attached to the noun. But again the extraction process was done with a special purpose extraction program whose correctness was hard to verify.
3
Querying Treebanks with TIGER-Search
We therefore decided to check the attachment tendencies of PPs in various treebanks for the three languages in question with the same tool and with queries that are as uniform as possible. For English we used the WSJ section of the Penn Treebank, for German we used our own ComputerZeitung treebank (3000 sentences), the
NEGRA treebank (10,000 sentences) and the recently released version of the TIGER treebank (50,000 sentences). For Swedish we used the SynTag treebank mentioned above and one section of the Talbanken treebank (6100 sentences). All these treebanks consist of constituent structure trees, and they are in representation formats which allow them to be loaded into TIGER-Search. This enables us to query them all in similar manners and to get a fairer comparison of the attachment tendencies. TIGER-Search is a powerful treebank query tool developed at the University of Stuttgart (K¨onig and Lezius, 2002). Its query language allows for feature-value descriptions of syntax graphs. It is similar in expressiveness to tgrep (Rohde, 2005) but it comes with graphical output and highlighting of the syntax trees plus some nice statistics functions. Our experiments for determining attachment tendencies proceed along the following lines. For each treebank we first query for all sequences of a noun immediately followed by a PP (henceforth noun+PP sequences). The dot being the precedence operator, we use the query: [pos="NN"] . [cat="PP"] This query will match twice in the tree in figure 1. It gives us the frequency of all ambiguously located PP. We disregard the fact that in certain clause positions a PP in such a sequence cannot be verb-attached and is thus not ambiguous. For example, an English noun+PP sequence in subject position is not ambiguous with respect to PP attachment since the PP cannot attach to the verb. Similar restrictions apply to German and Swedish. In order to determine how many of these sequences are annotated as noun attachments, we query for noun phrases that contain both a noun and an immediately following PP. This query will look like: #np_mum:[cat="NP"] > #np_child:[cat="NP"] & #np_mum > #pp:[cat="PP"] & #np_child >* #noun:[pos="NN"] & #noun . #pp All strings starting with # are variables and the > symbol is the dominance operator. So, this query says: Search for an NP (and call it np mum) that immediately dominates another NP (np child) AND that immediately dominates a PP, AND the
83
np child dominates a noun which is immediately followed by the PP. This query presupposes that a PP which is attached to a noun is actually annotated with the structure (NP (NP (... N)) (PP)) which is true for the Penn treebank (compare to the tree in figure 1). But the German treebanks represent this type of attachment rather as (NP (... N) (PP)) which means that the query needs to be adapted accordingly.4 Such queries give us the frequency of all noun+PP sequences and the frequency of all such sequences with noun attachments. These frequencies allow us to calculate the noun attachment rate (NAR) in our treebanks. N AR =
f req(noun + P P, noun attachm) f req(noun + P P )
We assume that all PPs in noun+PP sequences which are not attached to a noun are attached to a verb. This means we ignore the very few cases of such PPs that might be attached to adjectives (as for instance the second PP in ”due for revision in 1990”). Different annotation schemes require modifications to these basic queries, and different noun classes (regular nouns, proper names, deverbal nouns etc.) allow for a more detailed investigation. We now present the results for each language in turn. 3.1
Results for English
We used sections 0 to 12 of the WSJ part of the Penn Treebank (Marcus et al., 1993) with a total of 24,618 sentences for our experiments. Our start query reveals that an ambiguously located PP (i.e. a noun+PP sequence) occurs in 13,191 (54%) of these sentences, and it occurs a total of 20,858 times (a rate of 0.84 occurrences per sentences with respect to all sentences in the treebank). Searching for noun attachments with the second query described in section 3 we learn that 15,273 noun+PP sequences are annotated as noun attachments. And we catch another 547 noun attachments if we query for noun phrases that contain two PPs in sequence.5 In these cases the second PP is also attached to a noun, although not 4 There are a few occurrences of this latter structure in the Penn Treebank which should probably count as annotation errors. 5 See (Merlo et al., 1997) for a discussion of these cases and an approach in automatically disambiguating them.
Figure 1: Noun phrase tree from the Penn Treebank to the noun immediately preceding it (as for example in the tree in figure 1). With some similar queries we located another 110 cases of noun attachments (most of which are probably annotation errors if the annotation guidelines are applied strictly). This means that we found a total of 15,930 cases of noun attachment which corresponds to a noun attachment rate of 76.4% (by comparison to the 20,858 occurrences). This is a surprisingly high number. Neither (Hindle and Rooth, 1993) with 67% nor (Ratnaparkhi et al., 1994) with 59% noun attachment were anywhere close to this figure. What have we done differently? One aspect is that we only queried for singular nouns (NN) in the Penn Treebank where plural nouns (NNS) and proper names (NNP and NNPS) have separate PoS tags. Using analogous queries for plural nouns we found that they exhibit a NAR of 71.7%. Whereas the queries for proper names (singular and plural names taken together) account for a NAR of 54.5%. Another reason for the discrepancy in the NAR between Ratnaparkhi’s data and our calculations certainly comes from the fact that we queried for all sequences noun+PP as possibly ambiguous whereas they looked only at such sequences within verb phrases. But since we will do the same in both German and Swedish, this is still worthwhile. 3.2 Results for German The three German treebanks which we investigate are all annotated in more or less the same manner, i.e. according to the NEGRA guidelines which were slightly refined for the TIGER project. This enabled us to use the same set of queries for all
84
size noun+PP seq occur rate noun attachm NAR
CZ 3000 4355 1.4 2743 63.0%
NEGRA 10,000 6,938 0.7 4102 59.1%
TIGER 50,000 39,634 0.8 23,969 60.5%
Table 1: Results for the German treebanks
three of them. Since the German guidelines distinguish between node labels for coordinated phrases (e.g. CNP and CPP) and non-coordinated phrases (e.g. NP and PP), these distinctions needed to be taken into account. Table 1 summarizes the results. Our own ComputerZeitung treebank (CZ) has a much higher occurrence rate of ambiguously located PPs because the sentences were preselected for this phenomenon. The general NEGRA and TIGER treebanks have an occurrence rate that is similar to English (0.8). The NAR varies between 59.1% for the NEGRA treebank and 63.0% for the CZ treebank for regular nouns. The German annotation also distinguishes between regular nouns and proper names. The proper names show a much lower noun attachment rate than the regular nouns. The NAR in the CZ treebank is 22%, in the NEGRA treebank it is 20%, and in the TIGER treebank it is only 17%. Here we suspect that the difference between the CZ and the other treebanks is based on the different text types. The computer journal CZ contains more person names with affiliation (e.g. Stan Sugarman von der Firma Telemedia) and more company names with location (e.g. Aviso aus Finn-
land) than a regular newspaper (that was used in the NEGRA and TIGER corpora). As mentioned above, our previous experiments in (Volk, 2001) were based on sets of extracted tuples from both the CZ and NEGRA treebanks. Our extracted data set from the CZ treebank had a noun attachment rate of 61%, and the one from the NEGRA treebank had a noun attachment rate of 56%. So why are our new results based on TIGERSearch queries two to three percents higher? The main reason is that our old data sets included proper names (with their low noun attachment rate). But our extraction procedure comprised also a number of other idiosyncracies. In an attempt to harvest as many interesting V-N-P-N tuples as possible from our treebanks we exploited coordinated phrases and pronominal PPs. Some examples: 1. If the PP was preceded by a coordinated noun phrase, we created as many tuples as there were head nouns in the coordination. For example, the phrase ”den Austausch und die gemeinsame Nutzung von Daten . . . erm¨oglichen” leads to the tuples (erm¨oglichen, Austausch, von, Daten) and (erm¨oglichen, Nutzung, von, Daten) both with the decision ’noun attachment’. 2. If the PP was introduced by coordinated prepositions (e.g. Die Argumente f¨ur oder gegen den Netzwerkcomputer), we created as many tuples as there were prepositions. 3. If the verb group consists of coordinated verbs (e.g. Infos f¨ur Online-Dienste aufbereiten und gestalten), we created as many tuples as there were verbs. 4. We regarded pronominal adverbs (darin, dazu, hier¨uber, etc.) and reciprocal pronouns (miteinander, untereinander, voneinander, etc.) as equivalent to PPs and created tuples when such pronominals appeared immediately after a noun. See (Volk, 2003) for a more detailed discussion of these pronouns. 3.3 Results for Swedish Currently there is no large-scale Swedish treebank available. But there are some smaller treebanks from the 80s which have recently been converted
85
to TIGER-XML so that they can also be queried with TIGER-Search. SynTag (J¨arborg, 1986) is a treebank consisting of around 5100 sentences. Its conversion to TIGER-XML is documented in (Hagstr¨om, 2004). The treebank focuses on predicate-argument structures and some grammatical functions such as subject, head and adverbials. It is thus different from the constituent structures that we find in the Penn treebank or the German treebanks. We had to adapt our queries accordingly. Since prepositional phrases are not marked as such, we need to query for constituents (marked as subject, as adverbial or simply as argument) that start with a preposition. This results in a noun attachment rate of 73% (which is very close to the rate reported by (Aasa, 2004)). Again this does not include proper names which have a NAR of 44% in SynTag. Let us compare these results to the second Swedish treebank, Talbanken (first described by (Telemann, 1974)). Talbanken was a remarkable achievement in the 80s as it comes with two written language parts (with a total of more than 10,000 sentences from student essays and from newspapers) and two spoken language parts (with another 10,000 trees from interviews and conversations). We concentrated on the 6100 trees from the written part taken from newspaper texts. The occurrence rate in Talbanken is 0.76 (4658 noun+PP sequences in 6100 sentences), which is similar to the rates observed for English and German. The occurrence rate in SynTag is higher 0.93 (4737 noun+PP sequences in 5114 sentences). Talbanken (in its converted form) is annotated with constituent structure labels (NP, PP, VP etc.) and also distinguishes coordinated phrases (CNP, CPP, CVP etc.). The queries for determining the noun attachment rate can thus be similar to the queries over the German treebanks. In addition, Talbanken comes with a rich set of grammatical features as edge labels (e.g. there are different labels for logical subject, dummy subject and other subject). We found that the NAR for regular nouns in Talbanken is 60.5%. Talbanken distinguishes between regular nouns, deverbal nouns (often with the derivation suffix -ing: tj¨anstg¨oring, utbildning, o¨ vning) and deadjectival nouns (mostly with the derivation suffix -het: skyldighet, snabbhet, verksamhet). Not surprisingly, these special nouns have higher NARs than the regular nouns. The
deadjectival nouns have a NAR of 69.5%, and the deverbal nouns even have a NAR of 77%. Taken together (i.e. regarding all regular, deadjectival and deverbal nouns) this results in a NAR of 64%. Thus, the NARs which we obtain from the two Swedish treebanks (SynTag 73% and Talbanken 64%) differ drastically. It is unclear what this difference depends on. The text genre (newspapers) is the same in both cases. We have noticed that SynTag contains a number of annotation errors, but we don’t see that these errors favor noun attachment of PPs in a systematic way. One aspect might be the annotation decision in Talbanken to annotate PPs in light verb constructions. These are disturbing cases where the PP is a child node of the sentence node S (which means that it is interpreted as a verb attachment) with the edge label OA (objektadverbial). Nivre (2005, personal communication) pointed out that ”OA is what some theoreticians would call a ’prepositional object’ or a ’PP complement’, i.e. a complement of the verb that semantically is close to an object but which is realized as a prepositional phrase.” In our judgement many of those cases should be noun attachments (and thus be a child of an NP). For example, we looked at f¨oruts¨attning f¨or (= prerequisite for) which occurs 14 times, out of which 2 are annotated as OO (Other object) + OA, 11 are annotated as noun attachments, and 1 is erroneously annotated. If we compare that to betydelse f¨or (= significance for) which occurs 16 times out of which 13 are annotated as OO+OA and 3 are annotated as noun attachments, we wonder. First, it is obvious that there are inconsistencies in the treebank. We cannot see any reason why the 2 cases of f¨oruts¨attning f¨or are annotated differently than the other 11 cases. The verbs do not justify these discrepancies. For example, we have skapa (= to create) with the verb attachments and f¨orsvinna (= to disappear) with the noun attachment cases. And we find ge (= to give) on both sides. Second, we find it hard to follow the argument that the tendency for betydelse f¨or is stronger for the OO+OA than for f¨oruts¨attning f¨or. It might be based on the fact that betydelse f¨or is often used with the verb ha (= to have) and thus may count as a light verb construction with a verb group consisting of both ha plus betydelse and the f¨or-PP
86
being interpreted as an object of this complex verb group. Third, unfortunately not all cases of PPs annotated as objektadverbial can be regarded as noun attachments. But after having looked at some 70 occurrences of such PPs immediately following a noun, we estimate that around 30% should be noun attachments. Concluding our observations on Swedish let us mention that the very few cases of proper names in Talbanken have a NAR of 24%.
4
Comparison of the results
For English we have computed a NAR of 76.4% based on the Penn Treebank, for German we found NARs between 59% and 63% based on three treebanks, and for Swedish we determined a puzzling difference between 73% NAR in SynTag and 64% NAR in Talbanken. So, why is the tendency of a PP to attach to a preceding noun stronger in English than in Swedish which in turn shows a stronger tendency than German? For English the answer is very clear. The strong NAR is solely based on the dominance of the preposition of. In our section of the Penn Treebank we found 20,858 noun+PP sequences. Out of these, 8412 (40% !!) were PPs with the preposition of. And 99% of all of-PPs are noun attachments. So, the preposition of dominates the English NAR to the point that it should be treated separately.6 The Ratnaparkhi data sets (described above in section 2) contain 30% tuples with the preposition of in the test set and 27% of-tuples in the training set. The higher percentage of of-tuples in the test set may partially explain the higher NAR of 59% (vs. 52% in the training set). The dominance of of-tuples may also explain the relatively high NAR for proper names in English (54.5%) in comparison to 17% - 22% in German and similar figures for the Swedish Talbanken corpus. The Penn Treebank represents names that contain a PP (e.g. District of Columbia, American Association of Individual Investors) with a regular phrase structure. It turns out that 861 (35%) of the 2449 sequences ’proper name followed by PP’ are based on of-PPs. The dominance becomes even more obvious if we consider that the following 6 This is actually what has been done in some research on English PP attachment disambiguation. (Ratnaparkhi, 1998) first assumes noun attachment for all of-PPs and then applies his disambiguation methods to all remaining PPs.
prepositions on the frequency ranks are in (with only 485 occurrences) and for (246 occurrences). The dominance of the preposition of is so strong in English that we will get a totally different picture of attachment preferences if we omit of-PPs. The Ratnaparkhi training set without of-tuples is left with a NAR of 35% (!) and the test set has a NAR of 42%. In other words, English has a clear tendency of attaching PPs to verbs if we ignore the dominating of-PPs. Neither German nor Swedish has such a dominating preposition. There are, of course, prepositions in both languages that exhibit a clear tendency towards noun attachment or verb attachment. But they are not as frequent as the preposition of in English. For example, clear temporal prepositions like German seit (= since) are much more likely as verb attachments. Closest to the English of is the Swedish preposition av which has a NAR of 88% in the Talbanken corpus. But its overall frequency does not dominate the Swedish ranking. The most frequent preposition in ambiguous positions is i (frequency: 651 and NAR: 53%) followed by av (frequency: 564; NAR: 88%) and f¨or (frequency: 460; NAR: 42%).
5
Conclusion
The most important conclusion to be drawn from the above experiments and observations is the importance of profiling the data sets when working and reporting on PP attachment experiments. The profile should certainly answer the following questions: 1. What types of nouns where used when the tuples were extracted? (regular nouns, proper names, deverbal nouns, etc.) 2. Are there prepositions which dominate in frequency and attachment rate (like the English preposition of)? If so, how does the data set look like without these dominating prepositions? 3. What types of prepositions where regarded? (regular prepositions, contracted prepositions (e.g. in German am, im, zur), derived prepositions (e.g. English prepositions derived from gerund verb forms following, including, pending) etc.)
87
4. Is the extraction procedure restricted to noun+PP sequences in the verb phrase, or does it consider all such sequences? 5. What is the noun attachment rate in the data set? In order to find dominating prepositions we suggest a data profiling that includes the frequency and NARs of all prepositions in the data set. This will also give an overall picture of the number of prepositions involved. Our experiments have also shown the advantages of large treebanks for comparative linguistic studies. Such treebanks are even more valuable if they come in the same representation schema (e.g. TIGER-XML) so that they can be queried with the same tools. TIGER-Search has proven to be a suitable treebank query tool for our experiments although its statistics function broke down on some frequency counts we tried on large treebanks. For example, it was not possible to get a list of all prepositions with occurrence frequencies from a 50,000 sentence treebank. Another item on our TIGER-Search wish list is a batch mode so that we could run a set of queries and obtain a list of frequencies. Currently we have to trigger each query manually and copy the frequency results manually to an Excel file. Other than that, TIGER-Search is a wonderful tool which allows for quick sanity checks of the queries with the help of the highlighted tree structure displays in its GUI. We have compared noun attachment rates in English, German and Swedish over treebanks from various sources and with various annotation schemes. Of course, the results would be even better comparable if the treebanks were built on the same translated texts, i.e. on parallel corpora. Currently, there are no large parallel treebanks available. But our group works on such a parallel treebank for English, German and Swedish. Design decisions and first results were reported in (Volk and Samuelsson, 2004) and (Samuelsson and Volk, 2005). We believe that such parallel treebanks will allow a more focused and more detailed comparison of phenomena across languages.
6
Acknowledgements
We would like to thank J¨orgen Aasa for discussions on PP attachment in Swedish, and Joakim
Nivre, Johan Hall, Jens Nilsson at V¨axj¨o University for making the Swedish Talbanken treebank available. We also thank the anonymous reviewers for their discerning comments.
References
Ulf Telemann. 1974. Manual F¨or Grammatisk Beskrivning Av Talad Och Skriven Svenska. Inst. f¨or nordiska spr˚ak, Lund. Vincent Vandeghinste. 2002. Resolving PP attachment ambiguities using the WWW (abstract). In Computational Linguistics in the Netherlands, Groningen.
J¨orgen Aasa. 2004. Unsupervised resolution of PP attachment ambiguities in Swedish. Master’s thesis, Stockholm University. Combined C/D level thesis.
Martin Volk and Yvonne Samuelsson. 2004. Bootstrapping parallel treebanks. In Proc. of Workshop on Linguistically Interpreted Corpora (LINC) at COLING, Geneva.
Michael Collins and James Brooks. 1995. Prepositional phrase attachment through a backed-off model. In Proc. of the Third Workshop on Very Large Corpora.
Martin Volk. 2001. The automatic resolution of prepositional phrase attachment ambiguities in German. Habilitationsschrift, University of Zurich.
Bo Hagstr¨om. 2004. A TIGER-XML version of SynTag. Master’s thesis, Stockhom University.
Martin Volk. 2002. Combining unsupervised and supervised methods for PP attachment disambiguation. In Proc. of COLING-2002, Taipeh.
D. Hindle and M. Rooth. 1993. Structural ambiguity and lexical relations. Computational Linguistics, 19(1):103–120. Jerker J¨arborg. 1986. SynTag Dokumentation. Manual f¨or SynTaggning. Technical report, Department of Swedish, G¨oteborg University. Esther K¨onig and Wolfgang Lezius. 2002. The TIGER language - a description language for syntax graphs. Part 1: User’s guidelines. Technical report. Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn treebank. Computational Linguistics, 19(2):313–330. P. Merlo, M.W. Crocker, and C. Berthouzoz. 1997. Attaching multiple prepositional phrases: generalized backed-off estimation. In Proceedings of the Second Conference on Empirical Methods in Natural Language Processing. Brown University, RI. A. Ratnaparkhi, J. Reynar, and S. Roukos. 1994. A maximum entropy model for prepositional phrase attachment. In Proceedings of the ARPA Workshop on Human Language Technology, Plainsboro, NJ, March. Adwait Ratnaparkhi. 1998. Statistical models for unsupervised prepositional phrase attachment. In Proceedings of COLING-ACL-98, Montreal. Douglas L. T. Rohde, 2005. TGrep2 User Manual. MIT. Available from http://tedlab.mit.edu/ ∼dr/Tgrep2/. Yvonne Samuelsson and Martin Volk. 2005. Presentation and representation of parallel treebanks. In Proc. of the Treebank-Workshop at Nodalida, Joensuu, May. J. Stetina and M. Nagao. 1997. Corpus-based PP attachment ambiguity resolution with a semantic dictionary. In J. Zhou and K. Church, editors, Proc. of the 5th Workshop on Very Large Corpora, pages 66–80, Beijing and Hongkong.
88
Martin Volk. 2003. German prepositions and their kin. a survey with respect to the resolution of PP attachment ambiguities. In Proc. of ACL-SIGSEM Workshop: The Linguistic Dimensions of Prepositions and their Use in Computational Linguistics Formalisms and Applications, pages 77–88, Toulouse, France, September. IRIT.
Handling of Prepositions in English to Bengali Machine Translation Sudip Kumar Naskar Dept. of Comp. Sc. & Engg., Jadavpur University, Kolkata, India
[email protected]
Sivaji Bandyopadhyay Dept. of Comp. Sc. & Engg., Jadavpur University, Kolkata, India
[email protected]
Abstract The present study focuses on the lexical meanings of prepositions rather than on the thematic meanings because it is intended for use in an English-Bengali machine translation (MT) system, where the meaning of a lexical unit must be preserved in the target language, even though it may take a different syntactic form in the source and target languages. Bengali is the fifth language in the world in terms of the number of native speakers and is an important language in India. There is no concept of preposition in Bengali. English prepositions are translated to Bengali by attaching appropriate inflections to the head noun of the prepositional phrase (PP), i.e., the object of the preposition. The choice of the inflection depends on the spelling pattern of the translated Bengali head noun. Further postpositional words may also appear in the Bengali translation for some prepositions. The choice of the appropriate postpositional word depends on the WordNet synset information of the head noun. Idiomatic or metaphoric PPs are translated into Bengali by looking into a bilingual example base. The analysis presented here is general and applicable for translation from English to many other Indo-Aryan languages that handle prepositions using inflections and postpositions.
1
Introduction
Prepositions have been studied from a variety of perspectives. Both linguistic and computational
(monolingual and cross-lingual) aspects of prepositions have been contemplated by several researchers. Jackendoff (1977), Emonds (1985), Rauh (1993) and Pullum and Huddleston (2002) have investigated the syntactic characteristics of preposition. Cognitive theorists have examined the polysemous nature of prepositions and explored the conceptual relationships of the polysemy, proposing the graphical mental images (Lakoff and Johnson, 1980; Brugman, 1981, 1988; Herskovits, 1986; Langacker, 1987; Tyler and Evans, 2003). Fauconnier (1994) and Visetti and Cadiot (2002) have canvassed the pragmatic aspects of prepositions. A practical study of the usage of prepositions was carried out for the purpose of teaching English as a second language (Wahlen, 1995; Lindstromberg, 1997; Yates, 1999). The deictic properties of spatial prepositions have been studied by Hill (1982), while the geographical information provided by them was an interest of computational research (Xu and Badler, 2000; Tezuka et al., 2001). In the fields of natural language processing, the problem of PP attachment has been a topic for research for quite a long time, and in recent years, the problem was explored with a neural network-based approach (Sopena et al., 1998) and with a syntax-based trainable approach (Yeh and Vilain, 1998). Although past research has revealed various aspects of prepositions, there is not much semantic research of prepositions available for computational use, which requires a vigorous formalization of representing the semantics. A recent semantic study of prepositions for computational use is found in (Voss, 2002), with a focus on spatial prepositions. Spatial prepositions are divided into three categories according to which one of the two thematic meanings between place and path they acquire when they are in argument, adjunct and nonsubcategorized positions of particular types of
Proceedings of the Third ACL-SIGSEM Workshop on Prepositions, pages 89–94, c Trento, Italy, April 2006. 2006 Association for Computational Linguistics
89
verbs. The semantics of spatial prepositions dealt with in (Voss, 2002) is not lexical but thematic. There are some prepositions (e.g., over, with), which have many senses as preposition. By making use of the semantic features of the Complements (reference object) and Heads (verb, verb phrase, noun or noun phrase governing a preposition or a PP), the meaning of the polysemous prepositions can be computationally disambiguated. The different meanings of over call for different semantic features in its heads and complements [Alam, 04]. Prepositional systems across languages vary to a considerably degree, and this cross-linguistic diversity increases as we move from core, physical senses of prepositions into the metaphoric extensions of prepositional meaning (metaphor or rather, idiomaticity is one of the main realms of usage with prepositions) (Brala, 2000). The present study focuses on the lexical meanings of prepositions rather than on the thematic meanings because it is intended for use in an English-Bengali machine translation (MT) system, where the meaning of a sentence, a phrase or a lexical entry of the source language must be preserved in the target language, even though it may take a different syntactic form in the source and target languages. Bengali is the fifth language in the world in terms of the number of native speakers and is an important language in India. It is the official language of neighboring Bangladesh. There is no concept of preposition in Bengali. English prepositions are translated to Bengali by attaching appropriate inflections to the head noun of the PP, i.e., the object of the preposition. The choice of the inflection depends on the spelling pattern of the translated Bengali head noun. Further postpositional words may also appear in the Bengali translation for some prepositions. The choice of the appropriate postpositional word depends on the WordNet (Fellbaum, 1998) synset information of the head noun. Idiomatic or metaphoric PPs are translated into Bengali by looking into a bilingual example base. A brief overview of the English-Bengali MT System is presented in Section 2. Different types of English prepositions and their identification in the MT system are described in Section 3. Inflections and postpositions in Bengali are outlined in Section 4. Translation of English prepositions to inflections and postpositions in Bengali are detailed in Section 5. The conclusion is drawn in Section 6.
90
2
A Brief Overview of the EnglishBengali MT System
The handling of English prepositions during translation to Bengali has been studied with respect to an English-Bengali MT system (Naskar and Bandyopadhyay, 2005) being developed. In order to translate from English to Bengali, the first thing we do is lexical analysis of the English sentence using the WordNet, to gather the lexical features of the morphemes. During morphological analysis, the root words / terms (including idioms and named entities), along with associated grammatical information and semantic categories are extracted. A shallow parser identifies the constituent phrases of the source language sentence and tags them to encode all relevant information that might be needed to translate these phrases and perhaps resolve ambiguities in other phrases. Then these phrases are translated individually to the target language (Bengali) using Bengali synthesis rules. The noun phrases and PPs are translated using Example bases of syntactic transfer rules. Verb phrase translation scheme is rule based and uses Morphological Paradigm Suffix Tables. Finally, those target language phrases are arranged using some heuristics, based on the word ordering rules of Bengali, to form the target language representation of the source language sentence.
3
Prepositions in English
A preposition is a word placed before a “noun” to show in what relation the noun stands with regard to the other noun and verb words in the same sentence. The noun that follows a preposition, i.e., the reference object is in the accusative case and is governed by the preposition. Prepositions can also be defined as words that begin prepositional phrases (PP). A PP is a group of words containing a preposition, an object of the preposition, and any modifiers of the object. Syntactically, prepositions can be arranged into three classes – simple prepositions (e.g., at, by, for, from etc.), compound prepositions and phrase prepositions. A compound preposition is made up of a set of words which starts with and acts like a preposition (e.g., in spite of, in favor of, on behalf of etc.). A phrase preposition is a simple preposition preceded by a word from another category, such as an adverb, adjective, or conjunction (e.g., instead of, prior to, because of, according to etc.). Frequently prepositions follow the verbs together forming phrasal verbs and remain sepa-
rate. A word that looks like a preposition but is actually part of a phrasal verb is often called a particle. E.g. “Four men held up the bank.” Here held up is a verb [“to rob”]. Therefore, up is not a preposition, and bank is not the object of a preposition. Instead, bank is a direct object of the verb held up. A particle may not always appear immediately after the verb with which it makes up a phrasal verb (e.g., Four men held the bank up.). An idiomatic (metaphoric) PP starts with a preposition, but its meaning cannot be ascertained from the meaning of its components. Examples of idiomatic PPs are: at times, by hook or crook etc. All these syntactical characteristics are used to identify prepositions in the English-Bengali MT system. Moreover, the inventory of prepositions in English is a close set. So, identification of prepositions is not much of a problem in English. A simple list serves the purpose. The prepositions, compound prepositions, phrase prepositions and idiomatic PPs are identified during morphological analysis. Some of the phrasal verbs (when the phrasal verb appears as a whole) are identified during the morphological analysis phase and some during parsing (when the particle does not accompany the verb). However, there are some words that act as prepositions and fall into other POS categories as well. For example, the word before can be used as an adverb (e.g., I could not come before), preposition (e.g., He came before me) or a conjunction (e.g., He came before I came). Similarly, the word round can be used as an adjective (e.g., Rugby is not played with a round ball), noun (e.g., Rafter was knocked out of the tournament in the third round), adverb (e.g., They have moved all the furniture round), preposition (e.g., The earth revolves round the sun) and verb (e.g., His eyes rounded with anger). But depending on the POS of the neighboring words/terms, the parser easily identifies the correct POS of the word in the particular context. A preposition is usually placed in front of (is “pre-positioned” before) its object, but sometimes however may follow it (e.g., What are you looking at?). The preposition is often placed at the end when the reference object is an interrogative pronoun (e.g., Where are you coming from?) or a relative pronoun (e.g., My grandfather was a collector of coins, which we used to fight over). In such cases, the system finds out that the preposition is not a particle and is not followed by a noun either, so it must be a
91
stranded preposition. It searches the pronoun (relative or interrogative) that appears at its left and relates the stranded preposition to the pronoun. Thus during translation, the following conversion takes place. (1) Where are you coming from? ÅÆ From where are you coming? (2) My grandfather was a collector of coins, which we used to fight over. ÅÆ My grandfather was a collector of coins, over which we used to fight. But if the pronoun is missing, then the system has to find out the elliptical pronoun first. (3) I am grateful to the man I have spoken to. Æ I am grateful to the man [whom] I have spoken to. Æ I am grateful to the man to [whom] I have spoken. Prepositions represent several relations with the nouns governed by them. Spatial and temporal prepositions (which indicate a place or time relation) have received a relatively in-depth study for a number of languages. The semantics of other types of prepositions describing manner, instrument, amount or accompaniment largely remain unexplored. In case of an MT system, when a preposition has different representations in the target language for different relations indicated by it, identification of the relation is necessary. The WordNet synset information of the head noun of the PP, i.e., the object of the preposition serves to identify the relation.
4
Inflections and Postpositions in Bengali
In Bengali, there is no concept of preposition. English prepositions are handled in Bengali using inflections (vibhaktis) to the reference objects and/or post-positional words after them. Inflections get attached to the reference objects. An inflection has no existence of its own in the language, and it does not have any meaning as well. There are only a few inflections in Bengali: Φ (null), (-e) (-y) (-ye) (-te)
-å# , -Ì^ , -åÌ^ , -åTö , å#åTö (-ete), -åEõ (-ke), -åÌ[ý (-re), -å#åÌ[ý (-ere), -Ì[ý (-r) and -å#Ì[ý (-er) (an inflection is repre-
sented as a word with a leading ‘-’ in this paper). The placeholder indicated by a dashed circle represents a consonant or a conjunct. For examinflection is attached to the word ple, if
ject for any of these 3 English spatial and temporal prepositions. The choice depends on the spelling of the translated reference object. The rule is: if the last letter of the Bengali representation of the reference object is a consonant, ‘ ’ (-e) or
They have meanings of their own and are used independently like other words. A post-positional word is positioned after an inflected noun (the reference object). Some examples of the postpositional words in (colloquial) Bengali are:
ter of the Bengali word is a matra (vowel modifier) and if the matra is ‘ ’ (-a), any of ‘ ’
-å#
å# [ýçLçÌ[ý (bazar [market]) the inflected word is [ýçLçãÌ[ý (bazar-e [market-to]). On the other hand, å#åTö (-ete) is added to it (e.g., at/in marketÆ [ýçLçãÌ[ý [bazar-e / bazar-ete]), else if the last letpost-positional words are independent words.
×VãÌ^
, åUãEõ (theke [from]), LXî (jonno [for]), Eõçä»K÷ (kachhe [near]), aç]ãX (samne [in (diye [by])
front of]) etc.
5
Translating English prepositions to Bengali
When an English PP is translated into Bengali, the following transformation takes place: (preposition) (reference object) ÅÆ (reference object) [(inflection)] [(postpositional-word)]. The correspondence between English prepositions and Bengali postpositions (inflections and postpositional words) is not direct. As far as the selection of the appropriate target language representation of a preposition is concerned the reference object plays a major role in determining the correct preposition sense. Deciding whether the preposition is used in a spatial sense, as opposed to a temporal or other senses, is determined by the semantics of the head noun of the reference object. A noun phrase (NP) denoting a place gives rise to a spatial PP. Similarly, an object referring to a time entity produces a temporal expression. These relationships can be established by looking at the WordNet synset information of the head noun of the PP. 5.1
Translating English prepositions using Inflections in Bengali
The translation of the three English prepositions 'in', 'on', and 'at' involves identifying the possible inflection to be attached to the head noun of the PP. No postpositional words are placed after the head noun for these prepositions. The three prepositions 'in', 'on', and 'at' (which are both spatial and temporal in nature) can be translated into the Bengali inflections '- ' (-e), '- ’ (-te),
å#
åTö
-å#åTö (-ete) and 'Ì^'
(-y). Any of these 4 Bengali inflections can be placed after the reference ob-
92
#ç
åTö
Ì^
(-te), or ' ' (-y) can be added to the Bengali ref-
aµùîçãTö / aµùîçÌ^ [sandhya-te / sandhya-y]), otherwise 'åTö’ (-te) is added to it (e.g., at homeÆ [ýç×QÍöãTö [badi-te]). erence word (e.g., in eveningÆ
When translating the temporal expressions, if ‘on’ is followed by a day (like Sunday, Monday etc.) or by a date in English, null inflection is added. To translate this type of PPs, we take the help of an example base, which contains bilingual translation examples. Here are some translation examples from the example base (TLR – target language representation of the reference object). (1)
at
/
in
(place)
Ì å# åÌ^ / åTö ) [
(TLR) - ( / ye / te )]
ÅÆ
- ( e /
Ì Ì[ý /
(2) of (NP) ÅÆ (TLR) - (
å#Ì[ý / åÌ^Ì[ý
) [ - ( r / er / yer
)] 5.2
Translating English prepositions using Inflections and Postpositions in Bengali
Most of the English prepositions are translated to Bengali as inflections and postpositions to the noun word representing the reference object. To translate this type of PPs, we take the help of an example base, which contains bilingual translation examples. Here are some translation examples from the example base (TLR – target language representation of the reference object). (1)
before
(artifact)
ÅÆ
Ì Ì[ý å#Ì[ý åÌ^Ì[ý aç]ãX
(TLR) - ( / [ / ) ( r / er / yer ) samne ] (2) before (!artifact) ÅÆ
Ì Ì[ý å#Ì[ý / åÌ^Ì[ý %çãG [
(TLR) - ( / ) ( r / er / yer ) age ]
-
(3) round (place / physical object) ÅÆ (TLR) - ( / /
‘artifact’, whereas ‘evening’ and ‘me’ (which represents a person) are not. Thus ‘with’ is translated to Bengali as in sentence (1), and
Ì Ì[ý å#Ì[ý
åÌ^Ì[ý ) »JôçÌ[ý×VãEõ [
Ì[ý aç]ãX takes the meaning - Ì( Ì[ý / å#Ì[ý / åÌ^Ì[ý ) %çãG in
- ( r / er / yer
) chardike ] (4) after (time) ÅÆ (TLR) -
Ì( Ì[ý / å#Ì[ý / åÌ^Ì[ý ) YãÌ[ý÷ [
- ( r / er / yer ) pare ] (5) since (place / physical object / time) ÅÆ (TLR) [theke]
åUãEõ
The choice of inflection depends on the spelling of the translated reference object as said before. If the translated reference object ends with is added to it; else if ends with a a vowel,
åÌ^Ì[ý consonant, å#Ì[ý (er)is added to it; otherwise (it ends with a matra) Ì[ý (r)is appended with it. The
postpositional word is placed after the inflected reference object in Bengali. The choice of the postpositional word depends on the semantic information about the reference objects as collected from the WordNet. In cases with one postpositional word, there is no need to know the semantic features of the reference objects. For example, ‘since’, as a preposition, is always translated as (theke) in Bengali, irrespective of the reference object. Again in some cases, this semantic information about the reference object does not suffice to translate the preposition properly. Consider the following examples that include the preposition before in two different senses.
sentence (2) and (3). As there is no ambiguity in the meaning of compound prepositions and phrase prepositions, a simple listing of them (along with their Bengali representations) suffices to translate them. We have prepared a list that contains the phrase prepositions and compound prepositions in English along with their Bengali translations. English in spite of away from
Bengali [satteo]
aãüøC - Ì( Ì[ý / å#Ì[ý / åÌ^Ì[ý ) åUãEõ VÇãÌ[ý [ - ( r / er / yer ) theke dure ]
owing to
- Ì( Ì[ý / å#Ì[ý / åÌ^Ì[ý ) EõçÌ[ýãS
apart from
[ - ( r / er / yer ) karane ] [ chhadao ]
Instead of
K÷çQÍöçC - ( Ì[ý / å#Ì[ý / åÌ^Ì[ý ) Y×Ì[ý[ýãTöÛ [ - ( r / er / yer ) paribarte ]
along with
åUãEõ
(1) He stood before the door. ÅÆ (se [he] darja-r samne [the door before] dandalo [stood]) (2) He reached before evening. ÅÆ
åa VÌ[ýLçÌ[ý aç]ãX VñçQÍöç_
åa
aµùîçÌ[ý
- ( Ì[ý / å#Ì[ý / åÌ^Ì[ý ) açãU [ - ( r / er / yer ) sathhe ]
5.3
Translation of English Idiomatic PPs
The meaning of an idiomatic PP cannot be derived from the meanings of its components. The simplest way to tackle them is to maintain a listing of them. A list or a direct Example Base is used which contains idioms, which start with prepositions, along with their Bengali translations. Such an idiom is treated like any other PP during the word-reordering phase. Here are some examples of them:
a]ãÌ^ a]ãÌ^
(1) at times ÅÆ (samaye samaye) (2) by hook or crook ÅÆ (jebhabei hok)
%çãG
åYgì»K÷ç_
(se [he] sondhya-r age [evening before] pouchhalo [reached]) (3) He reached before John. ÅÆ (se [he] jan-er age [John before] pouchhalo [reached])
å^\öçã[ý+ åc÷çEõ
(3) to a fault ÅÆ (matratirikto)
åa LãXÌ[ý %çãG åYgì»K÷ç_
6
From the WordNet, the system acquires the semantic information that ‘door’ is a hyponym of
93
]çyç×Tö×Ì[ýNþ
Conclusion
In the present study, the handling of English prepositions in Bengali has been studied with reference to a machine translation system from English to Bengali. English prepositions are han-
dled in Bengali using inflections and / or using post-positional words. In machine translation, sense disambiguation of preposition is necessary when the target language has different representations for the same preposition. In Bengali, the choice of the appropriate inflection depends on the spelling of the reference object. The choice of the postpositional word depends on the semantic information about the reference object obtained from the WordNet.
Acknowledgements Our thanks go to Council of Scientific and Industrial Research, Human Resource Development Group, New Delhi, India for supporting Sudip Kumar Naskar under Senior Research Fellowship Award (9/96(402) 2003-EMR-I).
References Alam, Yukiko Sasaki. 2004. Decision Trees for Sense Disambiguation of Prepositions: Case of Over. In HLT/NAACL-04. Brala, Marija M. 2000. Understanding and trans-
lating (spatial) prepositions: An exercise in cognitive semantics for lexicographic purposes. Brugman, Claudia. 1988. The story of over:
Polysemy, semantics and the structure of the lexicon. New York: Garland Press. [1981. The
Langacker, Ronald. 1987. Foundations of cognitive grammar, vol. 1. Stanford, CA: Stanford University Press. Lindstromberg, Seth. 1997. English prepositions explained. Amsterdam: John Benjamins. Naskar, Sudip Kr. and Bandyopadhyay. Sivaji. 2005. A Phrasal EBMT System for Translating English to Bangla. In MT Summit X. Pullum, Geoffrey and Rodney Huddleston. 2002. Prepositions and prepositional phrases. In Huddleston and Pullum (eds.), 597-661. Rauh, Gisa. 1993. On the grammar of lexical and nonlexical prepositions in English. In Zelinskiy-Wibbelt (eds.), 99-150. Sopena, Joseph M., Agusti LLoberas and Joan L. Moliner. 1998. A connectionist approach to prepositional phrase attachment for real world texts. In COLING-ACL ’98, 1233-1237. Tezuka, Taro, Ryong Lee, Yahiko Kambayashi and Hiroki Takakura. 2001. Web-based inference rules for processing conceptual geographical relationships. Proceedings of Web Information Systems Engineering, 14-21. Tyler A. and Evans V. 2003. Reconsidering prepo-
sitional polysemy networks: the case of over*. In B. Nerlich, Z. Todd, V. Herman, & D. D. Clarke (Eds.), Polysemy: Flexible patterns of meanings in mind and language, pp. 95-160. Berlin: Mouton de Gruyter.
story of over. Berkely, CA: UC-Berkely MA thesis.]
Visetti, Yves-Marie and Pierre Cadiot. 2002. Insta-
Emonds, Joseph. 1985. A unified theory of syntactic categories. Dordrecht: Foris.
genbaum, Susanne and Dennis Kurzon (eds.), 9-39.
bility and the theory of semantic forms Starting from the case of prepositions. In Fei-
Fellbaum, Christiane D. ed. 1998. WordNet – An Electronic Lexical Database, MIT Press, Cambridge, MA.
Voss, Clare. 2002. Interlingua-based machine translation of spatial expressions. University of Maryland: Ph.D. Dissertation.
Fauconnier, Gilles. 1994. Mental spaces. Cambridge: Cambridge University Press.
Wahlen, Gloria. 1995. Prepositions illustrated. Michigan: The University of Michigan Press.
Herskovits, Annette. 1986. Language and spatial
Xu, Yilun Dianna and Norman Badler. 2000. Algorithms for generating motion trajectories described by prepositions. Proceedings of Computer Animation 2000, 30-35.
cognition An interdisciplinary study of the prepositions in English. Cambridge: Cambridge University Press. Hill, Cliffort 1982. Up/down, front/back, left/right. A contrastive study of Hausa and English. In Weissenborn and Klein, 13-42. Jackendoff, Ray. 1977. The architecture of the language. Cambridge, MA: MIT Press. Lakoff, George and Mark Johnson. 1980. Metaphors we live by. Chicago: University of Chicago Press.
94
Yates, Jean. 1999. The ins and outs of prepositions A guidebook for ESL students. New York: Barron’s. Yeh, Alexander S. and Marc B. Vilain. 1998. Some properties of preposition and subordinate conjunction attachments. In COLING-ACL ’98, 14361442.
Author Index Baldwin, Timothy, 65 Bandyopadhyayn, Sivaji, 89 Costello, Fintan J., 1 Hargraves, Orin, 37 Hartrumpf, Sven, 29 Helbig, Hermann, 29 Kawtrakul, Asanee, 51 Kelleher, John D., 1 Kim, Su Nam, 65 Kna´s, Iwona, 9 Kuan, Pek, 51 Lassen, Tine, 45 Lestrade, Sander, 23 Litkowski, Kenneth C., 37 Mari, Alda, 51 Murguia, Elixabete, 51 Naskar, Sudip Kumar, 89 Osswald, Rainer, 29 Raina, Achla, 51 Ranaivo-Malancon, Bali, 51 Rehbein, Ines, 57 Saint-Dizier, Patrick, 51 Sarkar, Sudeshna, 51 Suktarachan, Mukda, 51 Trawi´nski, Beata, 17 Van Eynde, Frank, 73 Volk, Martin, 81 Zarriess, Sina, 51
95
