relating syntax and semantics: the syntactico-semantic lexicon of the

R E L A T I N G S Y N T A X A N D SEMANTICS: THE s Y N T A C T I C O - S E M A N T I C LEXICON OF THE SYSTEM V I E - L A N G Ingeborg

Steinacker,

Ernst B u c h b e r g e r

D e p a r t m e n t of Medical C y b e r n e t i c s U n i v e r s i t y of Vienna, Austria

ABSTRACT

Generic concepts are organized in a h i e r a r c h y of superand subconcepts in which a s u b c o n c e p t inherits all a t t r i b u t e s of its s u p e r c o n c e p t s .

This paper describes the structure and e v a l u a t i o n of the s y n t a c t i c o - s e m a n t i c lexicon (SSL) of the German Natural Language Understanding System VIE-LANG [3]. VIE-LANG uses an SI-Net [2] as internal r e p r e s e n t a t i o n . The SSL c o n t a i n s the rules according to which the mapping between net-structures and surface s t r u c t u r e s of a s e n t e n c e is carried out. This i n f o r m a t i o n is structured in a way that it can be evaluated from two sides. The parser interprets it as production-rules that control the analysis. S y n t a c t i c and s e m a n t i c features of the input s e n t e n c e are evaluated and i n d i v i d u a l s are created in the s e m a n t i c net. The g e n e r a t o r uses the same rules to express selected net-structures in adequate natural language e x p r e s s i o n s . It is shown how both processes can make e f f e c t i v e use of the SSL. The different p o s s i b i l i t i e s for evaluating the SSL are explained and illustrated by e x a m p l e s .

The second layer of the net contains the d y n a m i c knowledge which c o n s i s t s of individualized concepts. The parser creates individuals of those net s t r u c t u r e s which are addressed by the input words. As more input is a n a l y z e d more i n d i v i d u a l s and links are created. These i n d i v i d u a l s c o n s t i t u t e the e p i s o d i c layer of the net. The c o n c e p t u a l content of the net is organized according to the idea of semantic primitives [8] which are characterized by typical attributes. Action primitives have attributes which c o r r e s p o n d to cases of a case grammar (AGENT, OBJECT, RECIPIENT, LOCATION, etc.) [4], [ii].

B. I

A.

Parsin@

Our parser b e l o n g s to the class of s e m a n t i c p a r s e r s as suggested by [i], [7]. Since syntax carries a lot of i n f o r m a t i o n in G e r m a n it has to be considered in analysis: The syntactic role of a constituent cannot be determined by word-order, instead its morphological endings which indicate the surface case of the c o n s t i t u e n t have to be evaluated.

O V E R V I E W OF THE SYSTEM V I E - L A N G

Representation

In the system VIE-LANG real world k n o w l e d g e is represented within a s e m a n t i c net (SN) which is realized in the formalism of an SI-Net [2]. The net is o r g a n i z e d in two layers.

The parser is a d a t a - d r i v e n system, which combines syntactic and semantic processes. Syntax is used as the tool to gain information concerning the c o n s t i t u e n t s of the sentence, but the s y n t a c t i c p r o c e s s e s interact with s e m a n t i c ones in order to c o n f i r m their h y p o t h e s e s about a c o n s t i t u e n t . To recognize NPs and PPs the parser uses an ATN, w h i c h accepts s e m a n t i c a l l y valid interpretations only. The r e s u l t a n t s t r u c t u r e s include s y n t a c t i c and semantic information about the constituent. These structures are then c o l l e c t e d in a c o n s t i t u e n t list.

The g e n e r i c layer c o n t a i n s the static k n o w l e d g e of the system. At the generic level real world knowledge is r e p r e s e n t e d in the form of concepts and roles. A concept is defined by its a t t r i b u t e s which consist of two parts: role and value restriction. The value r e s t r i c t i o n is a concept which defines the range of p o s s i b l e fillers for the attribute, the role d e f i n e s the relation of the filler to the concept being defined.

96

the IMS. The current individual (the a r g u m e n t of this call) is then removed from the stack and control is r e t u r n e d to the calling p r o c e d u r e , thus allowing the next rule to be processed. The IMS w h i c h is created during this part of the p r o c e s s forms the input for the step two p r o c e s s o r which will finally produce the output sentence [6].

The s e m a n t i c representation of a sentence is built by linking the c o n s t i t u e n t s to the predicate. This p r o c e s s is c o n t r o l l e d by the S S L - e n t r y for the verb. F i r s t the d o m i n a n t verb has to be d i s a m b i g u a t e d [9]. SSL entries for verbs contain the information how verb-dependent constituents are mapped onto the cases r e p r e s e n t e d within the net. In a last step referents for m o d i f y i n g c o n s t i t u e n t s are d e t e r m i n e d and attached. A s e n t e n c e is c o n s i d e r e d parsed s u c c e s s f u l l y after all of the s e n t e n c e have been As a result the parser c o n f i g u r a t i o n of i n d i v i d u a l s the s e m a n t i c representation C.

II

to have been constituents incorporated. produces a in the net of the input.

THE

SYNTACTICO-SEMANTIC

LEXICON

By m e a n s of the SSL the mapping between surface expressions in natural language and structures of the r e p r e s e n t a t i o n is achieved. For an NLU d i a l o g u e system the relation between surface and r e p r e s e n t a t i o n is of i n t e r e s t in the c o n t e x t of parsing and the context of g e n e r a t i n g . The s t r u c t u r e of the SSL allows i n t e r p r e t a t i o n by both processes.

Generation

The task of the generator is to convert a selected part of the e p i s o d i c layer of the semantic net into surface sentences. This part a root node and v e r t i c e s and nodes attached to it form a c o h e r e n t graph is assumed to have been determined previously by the dialogue component. Generation is a c c o m p l i s h e d in two steps: step one p e r f o r m s a m a p p i n g of the SN to an i n t e r m e d i a t e s t r u c t u r e (IMS) containing words together with s y n t a c t i c a l and morphological data, and step two t r a n s f o r m s the IMS to surface s e n t e n c e s by ap p l y i n g syntactical transformations, linearizations and a morphological synthesis.

A t t r i b u t e s of actions realize the ideas of a case grammar. This leads to a c o r r e s p o n d e n c e between roles in the net and surface cases within the sentence. Cases of a case g r a m m a r at the one hand show r e g u l a r i t i e s in their relation to s y n t a c t i c c o n s t i t u e n t s (subject -> A G E N T ) , at the other hand the r e l a t i o n between a role and a surface case is v e r b - d e p e n d e n t . E.g. the verb 'bekommen' (to get) relates the subject to the role RECIPIENT, the verb 'geben' (to give) relates the s u b j e c t to the role SOURCE. The verb 'geben' requires the R E C I P I E N T to be e x p r e s s e d by a dative. Such d e p e n d e n c i e s are captured in the entries of the SSL w h e r e a s the r e g u l a r i t i e s are treated by d e f a u l t s .

To p r o d u c e a single sentence, the d o m i n a t i n g verb is selected first, as it plays a central role in a sentence. The semantic p r i m i t i v e s of which the SN is composed imply that there is no o n e - t o - o n e c o r r e s p o n d e n c e between concepts in the net and words of the language. Therefore the d e c i s i o n which verb to select d e p e n d s on the p a t t e r n of i n d i v i d u a l s in the e p is o d i c layer of the net. The c r i t e r i a for this selection are attached to the generic concept of the root node in form of a d i s c r i m i n a t i o n net (DN) [5]. Its tests ev a l u a t e the filled a t t r i b u t e s of the root primitive. The evaluation of this DN results not only in a verb, but in a ve r b - s e n s e .

A.

Structure

of the

SSL

The basic unit in the SSL is the e n t r y for a word-sense. Associated to each w o r d - s e n s e is an optional number of pairs which we will d e s c r i b e by the terms 'Left Side' (LS) and 'Right Side' (RS). A pair d e s c r i b e s how a word (phrase) of the sentence is represented within the semantic net. LSs d e s c r i b e features of the surface sentence. Most features refer to syntactic properties, e.g. constituents of a given surface case, infinitive c o n s t r u c t i o n s , lexical c a t e g o r i e s , surface words, and some features indicate selectional restrictions. If a LS c o n t a i n s more than one feature they are combined with an operator. One of the most frequent p a t t e r n s that is used in LSs c o m b i n e s a s y n t a c t i c feature with a net concept which is interpreted as s e l e c t i o n a l restriction. This c o m b i n a t i o n reflects our general parsing approach to combine syntax with semantics.

The g e n e r a t o r accesses the SSL entry for this verb-sense and continues by p r o c e s s i n g the d i f f e r e n t rules of which it is composed. The rules are e v a l u a t e d from right to left. Right sides mainly deal with entities in the SN, especially individuals. If an individual is relevant to g e n e r a t i o n , it is put on a stack ("current individual"). When the left side is processed, syntactical data along with the result of a r e c u r s i v e call of this part of the g e n e r a t o r is passed to

97

interpreted as actions and evaluated sequentially. For the parser the s t r u c t u r e s in the RS are interpreted as r e p r e s e n t a t i o n of the word, t h e r e f o r e the indicated net-structures are individualized. The complete structure that has been created after all RSs have been executed is used as the r e p r e s e n t a t i o n of the input-word.

RSs refer m o s t l y to s t r u c t u r e s w i t h i n the semantic net. There is no one-to-one correspondence between word-senses and conceptual p r i m i t i v e s . To r e p r e s e n t word (or phrase) m e a n i n g s p r i m i t i v e s are linked forming more complex structures. By d e f i n i t i o n there is one distinguished concept in each RS 'the root concept' which is the central element of the representation. All other structures r e f e r e n c e d in an RS are linked to it.

V e r b - e n t r i e s for example s p e c i f y the r e l a t i o n b e t w e e n surface constituents and the cases w h i c h are a t t r i b u t e s of the action concept. Each v e r b - s e n s e calls for a typical s e n t e n t i a l pattern in w h i c h e a c h c o n s t i t u e n t has to fulfil certain s e m a n t i c restrictions. The parser selects a v e r b - s e n s e if the features of c o n s t i t u e n t s in the constituent list satisfy the c o n d i t i o n s of the LSs. After having selected one word-sense its RSs are evaluated and the c o n s t i t u e n t s are linked to the a c t i o n as c a s e - f i l l e r s .

Although the number of action-primitives is relatively small (14), the net provides possibilities to express d i f f e r e n c e s b e t w e e n related verbs. This is done by filling a t t r i b u t e s with certain v a l u e s by default. Such an attribute does not correspond to a c o n s t i t u e n t of the s e n t e n c e but is 'part' of the verb-sense, e,g. 'gehen' (to go) is represented by the concept CHANGE OF LOCATION, 'laufen' (to run) addresSes-the same concept, but its a t t r i b u t e SPEED is filled by a d i f f e r e n t value.

The parser uses the SSL entries to d isambiguate verbs. The LSs incorporate the factors by which w o r d - s e n s e s can be d i s c r i m i n a t e d from each other. For many verbs the selectional r e s t r i c t i o n of the direct object is a decisive factor. E.g. the verb 'bekommen' (to get) is i n t e r p r e t e d as O B J T R A N S iff the semantic r e s t r i c t i o n of the d i r e c t o b j e c t belongs to the class OWNABLE-OBJECT (see Fig. i). The m e c h a n i s m s by which d i s a m b i g u a tion is carried out if the LS is not met is e x p l a i n e d e l s e w h e r e [10].

Not all SSL entries are relevant to parser and generator some entries are r e l e v a n t to one process only. This should not be regarded as a disadvantage, on the contrary, such entries support e f f i c i e n t use of the SSL. Since each s u b s y s t e m has its own typical way of interpreting entries (LS and RS), process-specific entries are simply disregarded by the other system.

B.

E v a l u a t i o n of the SSL (BEKOMMEN (i [(AND (CASE ACC) (RESTR O W N A B L E - O B J E C T ) )

Parser and generator treat the entries in the SSL as p r o d u c t i o n - r u l e s , each i n t e r p r e t i n g LS and RS in its own way. The parser works from LSs to RSs whereas the generator works in the opposite direction. i.

((IND OBJTRANS) (VAL + O B J E C T *))] [ (T (CASE NOM))

Parsin@

---->

((VAL + R E C I P I E N T

The parser needs to map s u r f a c e - c o n s t i t u e n t s onto e l e m e n t s of the semantic net. To produce the semantic r e p r e s e n t a t i o n of an input word the parser accesses the SSL entry of this word. For each word there may be several word-sense entries. The LSs of all word-sense entries for a word incorporate the information n e c e s s a r y to d i s t i n g u i s h one sense from the others. The parser interprets the LSs as c o n d i t i o n s that have to be fulfilled by the input sentence. The SSL c o n t a i n s at least one pair LS - RS for each w o r d - s e n s e . In order to choose the c o r r e c t i n t e r p r e t a t i o n the LSs of the different word-senses are evaluated. After the parser has chosen a w o r d - s e n s e by matching sentence-patterns and LS-conditions the associated RSs are

[(AND

(PP VON) (RESTR PERSON

((VAL + SOURCE

*))] INSTITUTION))

*))]))

Fig. 1 SSL e n t r y for 'bekommen', w o r d - s e n s e - i ('to get an object') When the parser analyses the sentence 'Hans bekommt von dieser Frau ein Buch.' (John gets a book from this woman.) there are three constituents on the c o n s t i t u e n t list.

98

In the default cases the selectional restrictions are checked implicitly. The net does not allow i n s t a n t i a t i o n of structures that do not c o r r e s p o n d to the p a t t e r n s given in the g e n e r i c concepts. If this o c c u r s e.g. in the s e n t e n c e 'He will eat his hat.' an e r r o r - m e s s a g e is g e n e r a t e d because the semantic concept for 'hat', GARMENT, is not compatible with the restriction S U B S T A N C E for the OBJECT of the concept INGEST. At this stage of development we do not loosen selectional r e s t r i c t i o n s as suggested by W i l k ' s preference semantics [12].

I n t e r p r e t a t i o n of the first pair of the entry for bekommen-i leads to the i n s t a n t i a t i o n of the root concept O B J T R A N S (RS: (IND OBJTRANS)) and the c r e a t i o n of the value OBJECT filled by the r e p r e s e n t a t i o n of book. The parameter '+' refers to the root individual f o r all pairs of the w o r d - s e n s e entry. For the parser the parameter '*' in the SSL refers to the r e p r e s e n t a t i o n of the c o n s t i t u e n t selected by the LS w h i c h is local to one pair. The second pair leads to the instantiation of the value RECIPIENT filled by the representation for 'Hans' and the third one finally instantiates SOURCE filled by the representation of 'Frau'. The resulting r e p r e s e n t a t i o n of the sentence is shown in Fig. 2.

ect

2.

When g e n e r a t i n g a sentence, the g e n e r a t o r starts by regarding the root node which has been passed to it by the d i a l o g u e component. Normally, this root node will, together with the attributes a t t a c h e d to it, c o r r e s p o n d to a verb, so this verb is selected first. As m e n t i o n e d above, a d i s c r i m i n a t i o n net is used to a c c o m p l i s h this task. The DN s e l e c t s a v e r b - s e n s e according to the a t t r i b u t e s of the root node.

/ ~ u r c e We will show the further p r o c e s s i n g by means of the e x a m p l e shown in Fig. 2. Let us assume the v e r b - s e n s e 1 of the verb 'bekommen' (Fig. I) has a l r e a d y been selected. The e n t r i e s of the SSL are treated from right to left by the generator, so we start with (IND OBJTRANS). This will result in a null action for the g e n e r a t o r , as an instance of OBJTRANS (OBJTRANS-II) is a l r e a d y known as c u r r e n t root node and it has been put as first element onto the stack for the c u r r e n t individual. (VAL + O B J E C T *) is c o n s i d e r e d next. + denotes the root node, * the~ individual attached to that role of it which is specified by the second parameter, i.e. OBJECT. This element, n a m e l y BOOK-4, is put on the stack.

/~recipient

name

Fig. 2 Net structure 'Hans bekommt yon dieser

for Frau ein Buch.'

Now the g e n e r a t o r p r o c e e d s with the LS: (CASE ACC) is a r e c u r s i v e call to the g e n e r a t o r with the current individual, BOOK-4, as new root node together with the i n f o r m a t i o n that the result shall bear a c c u s a t i v e case endings. The generator p r o c e s s e s the DN for the concept 'BOOK' and returns 'Buch'. This lexeme together with the case information now forms part of the IMS. After having processed the current individual BOOK-4, it is removed from the stack. The action (RESTR OWNABLE OBJECT) results in a no-op for the g e n e r a t ~ r , as this information has a l r e a d y been p r o c e s s e d in the DN when d e c i d i n g to use the verb-sense 'bekommen-l' (see below).

A c t i o n p r i m i t i v e s t y p i c a l l y have an AGENT and an OBJECT attribute. In most cases their surface equivalents are subject and direct object respectively. T h e r e f o r e it would be redundant to include these r e l a t i o n s for every verb. In these cases only the root concept is g i v e n in the RS (see Fig. 3). The mapping is carried out by default m e c h a n i s m s which are applied whenever the LSs do not refer to subject or direct object. (ESSEN

(i [(T) - - > ( ( I N D

SSL e n t r y for

Generator

INGEST))]))

Fig. 3 'essen' , w o r d - s e n s e - i ('to eat')

The second RS-LS-pair is treated in a similar way: The individual a t t a c h e d

99

to R E C I P I E N T is put on the stack, (CASE NOM) calls the g e n e r a t o r with PERSON-9 as new root node and says that the resultant structure shall be rendered as a nominative. The DN of PERSON s u p p l i e s the information that persons are best specified by their names (if p r e s e n t in the net if not, other criteria are considered) , and so the word 'Hans' c o m p l e t e s the s t r u c t u r e being passed to the IMS.

between a n a l y s i s and s y n t h e s i s is realized in the SSL by pairing off LSs and RSs. F l e x i b i l i t y is insured by the fact that parser as well as g e n e r a t o r treat LS and RS each in an i d i o s y n c r a t i c way.

ACKNOWLEDGEMENTS This r e s e a r c h was sponsored by Austrian 'Fonds zur Foerderung w i s s e n s c h a f t l i c h e n Forschung', grant 4158 (supervision Robert Trappl).

As for the last test-action-pair, Pp causes a p r e p o s i t i o n a l phrase, 'yon der Frau', to be created. In German, the p r e p o s i t i o n 'yon' implies dative case, so no a d d i t i o n a l entry (CASE DAT) is r e q u i r e d in the SSL. (Note that this o m i s s i o n e n a b l e s the parser to ignore case e r r o r s in the input s e n t e n c e that do not influence the semantics.) 3.

Creating

REFERENCES [i]

B o g u r a e v B.K.: Automatic Resolution of L i n g u i s t i c A m b i g u i t i e s , Univ. of C a m b r i d g e , Comp. Laboratory, TR-II ; 1979. [2] B r a c h m a n , R.J.: A Structural P a r a d i g m for Representing Knowledge, Bolt, B e r a n e k and Newman; 1978. [3] B u c h b e r g e r E., S t e i n a c k e r I., Trappl R., Trost H., Leinfellner E.: VIE-LANG A German Language Understanding System, in: Trappl R.(ed.), Cybernetics and Systems R e s e a r c h , North Holland, Amsterdam; 1982. [4] F i l l m o r e C.: The Case for Case, in: Bach E., Harms R.T. (eds.) : Univ e r s a l s in Linguistic Theory, Holt, Rinehart & Winston, New York, 1968. [s] G o l d m a n N.M.: C o m p u t e r G e n e r a t i o n of Natural Language from a Deep C o n c e p t u a l Base, Stanford AI Lab Memo AIM-247; 1974. [6] H o r a c e k H.: Generierung im System VIE-LANG: Linguistischer Teil, TR 83-04, Dept. of Medical Cybernetics, Univ. of Vienna, Austria; 1983. [7] R i e s b e c k , C.K. and Schank, R.C.: Comprehension by Computer: Expectation-based Analysis of Sentences in Context, Yale Univ., RR-78; 1976. [8] S c h a n k R.C.: Conceptual Information P r o c e s s i n g , N o r t h - H o l l a n d , Amsterdam; 1975. [9] S t e i n a c k e r I., Trost H., Leinfellner E. : Di s a m b i g u a t i o n in German, in: Trappl R. (ed.), Cybernetics and Systems Research, North Holland, Amsterdam; 1982. [10] S t e i n a c k e r I., Trost H.: Structural Relations - A Case Against Case, Proceedings of the IJCAI 83, K a r l s r u h e , 1983. [ii] Trost H.: E r s t e l l e n der i n h a l t l i c h e n Komponenten eines Semantischen Netzes, TR 81-03, Dept. of Medical Cybernetics, Univ. of Vienna, Austria; 1983. [12] Wilks Y.: Making Preferences more Active, University of Edinburgh, D.A.I., RR-32,1977.

D i s c r i m i n a t i o n Nets

So far, the use of the SSL has been d e m o n s t r a t e d only p a r t i a l l y : in the example above some of the e l e m e n t s in RSs and LSs have been treated as no-ops, especially INDIV and RESTR. These elements, instead of being used in the process of g e n e r a t i o n , p r o v i d e i n f o r m a t i o n for b u i l d i n g data structures for the g e n e r a t o r , n a m e l y the above m e n t i o n e d DNs. AS an example, consider the entry for 'bekommen' (Fig. i), (INDIV OBJTRANS) informs us about a correspondence between the concept OBJTRANS and the v e r b - s e n s e 'bekommen-l' This correspondence leads to the i n c o r p o r a t i o n of 'bekommen-l' as a leaf node in the DN for the concept OBJTRANS. Other clues for c o n s t r u c t i n g the DNs are provided by the VALs, thus giving them a double usage: (VAL + R E C I P I E N T *) in the SSL entry for 'bekommen-l' (Fig. i) implies that an individual a t t a c h e d to the R E C I P I E N T role of an OBJTRANS individual is a prerequisite for selecting this verb-sense. (The absence of a recipient in the net would lead to the s e l e c t i o n of 'weggeben' (to give away).)

III

the der no

SUMMARY

We have shown how a lexicon that includes syntactic and semantic i n f o r m a t i o n has to be structured to allow e f f i c i e n t use by two processes, parser and generator. Whereas both must have access to k n o w l e d g e about syntax as well as representation, their starting position differs: The parser is c o n f r o n t e d with surface expressions, therefore LSs are evaluated first. The generator has to process net structures, so it b e g i n s by evaluating RSs. The reciprocal relation

i00