Generating Punctuation in Written Arguments Chris Reed
Derek Long
Department of Computer Science, University College London, Gower Street, London WC1E 6BT Tel. (44)(0)171 380 7214 Fax. (44)(0)171 387 1397
[email protected] http://www.cs.ucl.ac.uk/staff/C.Reed/
Department of Computer Science University of Durham South Road, Durham DH1 3LE Tel. (44)(0)191 374 2000 Fax. (44)(0)191 374 2560
[email protected]
Abstract The generation of punctuation (including formatting) in written text has not been explored extensively, despite general acknowledgment that punctuation is crucial not just for rare cases of disambiguation, but more generally for ease of comprehension and thus for the ultimate effectiveness of a text. A framework for the generation of natural language argument is summarized, and is then shown to be suited to the generation of a number of forms of punctuation which have not be adequately accounted for in previous work. In particular, it is proposed that chapter, section and paragraph breaks, footnotes, and quotations require an abstract, intention-laden representation for their adequate generation. Finally, it is demonstrated that explicit handling of focus shift operators can facilitate the generation of punctuation in general, and that the distinction between structural and rhetorical facets of text provides a means of accounting for context-specific punctuation.
Introduction The various types of punctuation in written text seem to be arranged in a hierarchy which is related to the hierarchical structure of the text itself (though, following Nunberg (1990), it is important to recognize that the term punctuation is used to refer to both the non-alphanumeric characters, commas, colons, etc. and also font, typeface, horizontal and vertical positioning, etc.). This paper is based around work which has focused upon the problem of generating persuasive text: generating coherent, effective, natural language arguments (Reed at al., 1996a). It has been found that the abstract representations designed to facilitate such generation are also highly suited to the production of a number of types of punctuation which have frequently posed problems for natural language generation (NLG) systems. It is posited that these types of punctuation (and in particular, chapter, section, and paragraph breaks, footnotes, appendices, and quotations) are dependent upon a high level representation of the text, in which the communicative intent is paramount. It is further suggested that these high level forms of punctuation are the direct realization of such intention-laden goals. Before examining further the hierarchy of punctuation forms, and its relationship to similar hierarchical structure in both free text and argument, a brief overview of the generation
framework is presented, followed by a summary of the planning approach adopted. At the highest level of abstraction, the Argument Structure (AS) level is responsible for the logical form of the argument, ie. the deductive, inductive and fallacial relationships holding between premises and conclusions. The operators which are available at the AS level are thus conceptually similar to a subset of those presented in (Maybury, 1993). Below the AS level, the Eloquence Generation (EG) level effects stylistic and rhetorical refinements, based upon contextual and hearer-dependent parameters. The EG level does not alter the plan produced by the AS level; rather, it introduces rhetorical statements, inserts lexical information for subsequent realization, marks stylistic preference (voice, mood, etc.), and so on. Importantly, much of the functionality at the EG level can be seen as a process of markup of the AS level plan. Crucial to the interaction between the AS and EG levels is the notion of salience. The structure planned by the AS level does not explicitly employ linguistic goals: rather, it uses goals expressing the salience of facts to the hearer. It is the task of the EG level to determine how information should be made salient to the hearer - often, items need not be uttered for them to be salient, resulting in enthymematic argument. Below the AS and EG levels, the inter-clause structure is resolved, employing relations similar to those of Rhetorical Structure Theory, RST (Mann and Thompson, 1987). Finally the grammatical, syntactic and morphological realization is performed through interface with the LOLITA system (Smith et al., 1994). The current work employs a concept of argument based upon the study of a corpus of natural arguments drawn from scientific papers, advertisements, editorial commentary and ‘letters to the editor’. It is similar to that referred to as the “standard approach” in argumentation theory (Freeman, 1991), in which an argument is seen as comprising one or more premises contributing to a single conclusion. Each premise can in turn be supported by a subargument, with the conclusion of the subargument acting as premise to the superargument. Finally, groups of premises contribute to a conclusion either independently (disjunct support) or in combination (conjunct support). This distinction is of crucial importance when attempting
to determine the validity of an argument, (Freeman, 1991), and it is therefore surprising to find the distinction eschewed in many generation systems (eg. (McConachy and Zukerman, 1996)).
Planning The task of generating persuasive discourse is seen in this work, as in much of NLG, as one of planning. NOAH (Sacerdoti, 1977) is frequently employed in text generation (Hovy, 1993), but is hampered by the fact that abstract operator bodies are primarily composed of operators. These operators are applied immediately the abstract operator is selected: the latter are thus acting as schema-like ‘recipes’, and systems are liable to suffer from reduced flexibility as a consequence. In contrast, the current work makes use of AbNLP (Fox and Long, 1995), a hierarchical planner based upon the concept of encapsulation, whereby the body of an abstract operator contains goals rather than operators, and further, that the body of an operator is not opened up until an entire abstract plan has been completed (ie. there are no goals left unfulfilled at that level of abstraction). On completion of an abstract plan (which can be seen, in discourse planning, as a skeletal outline of what is to be communicated), the refinement operation opens up all the abstract operator bodies, such that the structure and constraints determined at one level of abstraction are propagated to the next level down. As a consequence, many choices which might have been considered during planning of an argument at the detailed level can be pruned as they become inconsistent with the abstract plan. Such an approach has been shown to considerably improve upon the performance of a classical planner, (Bacchus and Yang, 1992). The use of AbNLP in a framework for argumentative discourse planning is discussed in more detail in (Reed et al, 1996a). The initial state in planning a persuasive text is described by two goals. Firstly, BEL (H, P), which represents the intention for the hearer, H, to believe some proposition, P. Thus it carries intentional rather than informational content (Young and Moore, 1994) and is classified as a communicative goal (Moore and Paris, 1994). Secondly, the initial state also includes the goal IS_SALIENT (H, P, _), also a communicative goal. This represents the fact that the hearer must not just believe P, but that P must be salient to him (in the general discourse context, ‘_‘- the use of this context parameter is discussed in more detail in (Reed et al., 1997)). In normal communication (ie. communication which obeys Grice’s maxims of co-operation, (Grice, 1975), and is therefore truthful and honest), every BEL goal will have a corresponding IS_SALIENT goal. It is the BEL goals which are planned for at the AS level: ultimately, the argumentation will have to reach beliefs which the speaker assumes the hearer either believes or will accept without further support. The IS_SALIENT goals, however, are usually satisfied by the MAKE_SALIENT operator which represents a primitive action at the AS level, and expresses the goal of making a proposition salient to the hearer (how this is achieved is determined by the EG level: some pieces
of information in an argument are obvious and can be left implicit; others must be realized explicitly, in varying degrees of detail). The BEL goals are satisfied by the deductive, inductive and fallacial operators available to the AS level. Argument components are frequently linked by support relations which can be characterized by a Modus Ponens operator, in which the inference between antecedent and consequent is weaker than that of strict deduction. The MP operator is given below, in Figure 1 (note that similar operators are available for situations in which the hearer believes ~P, as opposed to those in which the hearer has no opinion on P). MP (H, P) Shell: Precond: Add: Body: Goals:
∃X: (X → P) ~BEL (H, X) BEL (H, X) PUSH_TOPIC (P) BEL (H, X) IS_SALIENT (H, X, P) BEL (H, X → P) IS_SALIENT (H, X→P, P) POP_TOPIC (P)
Figure 1. Modus Ponens operator In addition to the BEL and IS_SALIENT goals corresponding to two premises of Modus Ponens, the body of the operator also contains the topic manipulation goals, PUSH_TOPIC and POP_TOPIC. These are typically satisfied by primitive operators (of the same name, for convenience), which represent actions to be performed on a topic stack similar that proposed by (Grosz and Sidner, 1986). The explicit operationalization of topic manipulation (rather than seeing focus as a constraint active over the planning process) offers a number of advantages in generating persuasive discourse. There is a close relationship between the position of topic manipulators and the occurrence of clue-words, which have been shown to be vital in ensuring the coherency of an argument (Cohen, 1987). Crucially, topic manipulators also co-occur with many forms of punctuation and formatting: this point is discussed in more detail below. Seeing PUSH_TOPIC and POP_TOPIC as goals suggests that it may also be appropriate to subdivide the class of communicative goals proposed by (Moore and Paris, 1994) into those which are intentional and those which are attentional (such an approach would also seem to be consistent with Moser and Moore’s (1996) attempt to reconcile (Grosz and Sidner, 1986) with RST, (Mann and Thompson, 1987)). The planning conducted at the AS level thus results in a highly parsimonious plan, in that it employs just three operators representing primitive actions at this level of abstraction: IS_SALIENT, PUSH_TOPIC and POP_TOPIC. For example, the goals in Figure 2 represents the abstract aim of persuading the hearer of a.
BEL (h, a) IS_SALIENT (h, a, _) Figure 2. Initial goals
If a could be supported by a single Modus Ponens argument, the BEL goal would be fulfilled by the MP operator, and the IS_SALIENT goal by MAKE_SALIENT, producing the abstract plan in Figure 3. MP (h, a) X = b MAKE_SALIENT (h, a, _) Figure 3. First complete abstract plan
During refinement, the MP body is opened up, from where the BEL goals are satisfied by matching those in the hearer model, the IS_SALIENT goals by MAKE_SALIENT operators and the topic manipulators by corresponding primitives. Thus the final plan of the AS level is as shown in Figure 4.
PUSH_TOPIC (a) MAKE_SALIENT (h, b, a) MAKE_SALIENT (h, b → a, a) POP_TOPIC (a) MAKE_SALIENT (h, a, _) Figure 4. Plan of primitives for single step MP argument
Such plans, after appropriate reordering of components (Reed and Long, 1997), form the basis of the realization process, with surface features being introduced at each stage in the framework. The following section examines the relationship between various classes of punctuation and the levels of abstraction in the proposed framework, concentrating on the punctuation which is a direct result of the functionality at the AS and EG levels.
Generating Punctuation Punctuation is a crucial component of written text. It functions to clarify text by delimiting semantic units, a process which does not just clarify meaning, but plays a vital role in comprehension (compare, for example, the growing use of punctuation analysis as an aid in parsing technology: (Briscoe, 1996; Jones, 1995; Peh and Ting, 1996; Simard, 1996) ). Nunberg (1990) describes the role of punctuation as one of indicating text-categories: in the same way that the lexical content of a text is based upon the lexical categories, so the graphological appearance is based upon text categories. Thus Nunberg proposes the category of text-sentence, which is typically delimited by a capital letter at the head, and a full stop at the tail. A textsentence does not necessarily coincide with a lexical sentence (of standard linguistic analysis): a text sentence may comprise several lexical sentences (conjoined, for
example, by a semicolon), or may alternatively not even contain a single lexical sentence1. Text generation has not formally distinguished between these two forms of grammar (ie. the textual and the lexical) - generation seems to have relied upon the linguistics of speech (despite the fact that realization is most commonly textual). As Nunberg persuasively argues, it is important to see written text as a linguistic system in itself - not just a transcript of speech (the role of punctuation, for example, is not one of indicating intonation patterns). Despite failing to make this distinction, NLG has nevertheless been able to approach the problem of producing effective punctuation due to the fact that many systems maintain an abstract representation of communicative intent. This representation is free of structural and textual constraints - these constraints are introduced as the planning process is conducted, and the communicative intentions are realized into surface textual form. As the planning proceeds, both lexical and textual structure can be imposed - the latter can include punctuation. Recent research in NLG has generally made the assumption that the generation process can be characterized by a pipeline model, wherein decisions made a lower levels do not affect those made at higher levels (though aspects of this model have been criticized - (Hovy and Wanner, 1996), for example). Accordingly, discourse structure is usually represented in the form of a hierarchical tree. This pipeline model, and the hierarchical structure generated as a result, lends itself to the introduction of punctuation, and in particular, offers a means of dealing with the different levels of punctuation. At the lowest levels of realization, lexical and morphological structure is determined (by a system such as LOLITA (Smith et al., 1994) ). It is here that punctuation such as hyphens and virgules (eg. and/or) are introduced. Above this sits further grammatical functionality, that which Hovy and Wanner (1996) have proposed as the sentence generation level: it is here that delimiting and separating commas are introduced, along with dash interpolation and parentheses. More abstract again, inter-clause relations, such as those of Rhetorical Structure Theory, RST (Mann and Thompson, 1987), link semantic units of text, and can give rise to punctuation appropriate to this level in the hierarchy. Full stops, colons and semicolons are all generated at this level, due to their dependence upon the notion of ‘clause’. However, it is also clear that they are rather close to the border with the previous level, at which it has been proposed that (text-)sentence content delimitation is performed (Hovy and Wanner, 1996). Given the recursive manner in which RST is applied, it is unsurprising that it can also be used to account for a number of other, ‘higher-level’ forms of punctuation. For example, Hovy and Arens (1991) discuss the generation of 1
Nunberg’s example (p22) offers three text sentences of which only one is classed as a lexical sentence: The L9000 delivers everything you wanted in a luxury sedan. With more power. At a price you can afford.
enumerated lists from consecutive SEQUENCE relations, and of itemized lists from clauses connected by a JOINT relation. They also suggest that section headings are the result of the IDENTIFICATION relation and appendices and footnotes of the BACKGROUND relation. Lastly, they note that some structures are the result of concepts not explicitly handled by RST - quotation, for example, relying 2 upon PROJECTION . A number of punctuation forms, however, seem to be the result of structure which is more abstract yet, lying superordinate to the representation offered by RST. The pragmatic organization developed at the Argument Structure level of the framework proposed in (Reed et al., 1996a) corresponds more closely to these forms of punctuation than the less abstract organization discussed above. The plan produced by the AS level is highly suited to high level document formatting since the focus shifts which are represented explicitly, are naturally aligned with breaks of chapter, section and paragraph. It also has explicit notions of support and authority required for handling quotation. Furthermore, the effect of coherency constraints (Reed and Long, 1997) corresponds intuitively to heuristics used naturally in the generation of footnotes, and appendices.
text identified as X; The author organizes X into two parts identified by A and B; The author assigns the level of chapter to A and B; etc.). This intentionality, identified by Pascual and Virbel as lying behind high level discourse breaks, cannot be accounted for by RST alone. For it has been recognized that RST cannot easily be extended to deal with the distinction between the informational and intentional content (Moore and Pollack, 1992). Paragraph breaks also pose a related problem for an account based on RST. For although such breaks occur at a single level in the hierarchical text plan, they do not occur with a single type of relation; rather with almost any relation which happens to be included at the appropriate level in the hierarchy. The abstraction offered in the AS level plan allows paragraph breaks to be tied to a single operator (namely, a PUSH_TOPIC), whenever that operator 3 is found at a particular level in the hierarchy . As discussed in (Moore and Pollack, 1992), the mapping between informational RST relations and corresponding intentions is not a one-to-one mapping. Similarly here, the intention behind the generation of a paragraph break can co-occur with a number of different informational RST relations.
Footnotes Breaks The use of chapter and section breaks is termed by Hovy and Arens (1991) separation, referring to the vertical displacement often associated with these formatting conventions. However, the organization of chapters, sections and subsections relies upon more than the vertical distance that such headings are afforded (since headings may or may not be numbered, and may or may not be inline, and may or may not be in bold or italic face, and so on). In the text of extended argument (such as a monograph), the chapter and section breaks correspond to focus shifts at the highest levels of abstraction in the plan of primitives generated by the AS level. However, as mentioned above, the PUSH_TOPIC operator in a plan represents a communicative goal, in the terminology of Moore and Paris (1994). The intention can be realized through the application of, say, a section break. That is to say that the speaker intends the hearer to appreciate that a new topic is being introduced (at this level of abstraction in the discourse), and that one means of achieving this is to state the fact using an explicit section break. This is precisely the notion implicit in the analysis of chapter and section breaks offered by Pascual and Virbel (1996), in which the use of such formatting expresses the underlying architecture of the text, which they term a metadiscourse. A metadiscourse consists of statements made by the speaker, though the use of formatting, regarding the structure of the discourse itself (eg. The author creates a 2
The discussion here concentrates upon RST simply because most of the work on generating punctuation is itself RST-based. However, comparable theories also offer scope for investigation of punctuation - (Say and Akman, 1995) using Discourse Representation Theory, DRT (Kamp and Reyle, 1993), for example.
Hovy and Arens (1991) suggest that footnotes and other offpage formatting (appendices, side bars, marginal notes, etc.) are the result of the RST relation BACKGROUND, - as are parentheses - with a greater or lesser amount of material included in the Satellite. However, it would be more convenient to permit a wider variety of relations to be expressed outside the main flow of the text (ELABORATION, ANALOGY, etc.), but in such a way that the resultant text (both in the body of the text and in the footnote) is coherent. The framework in (Reed and Long, 1997) employs a notion of coherency which is not determined simply by the existence (or otherwise) of an RST description for text. For text which is incoherent to most readers, may well have a valid RST representation (Marcu, 1996). Furthermore, it is possible to take an incoherent text and by merely introducing clue words (ie. not affecting the RST representation at all) render the text coherent again examples of this process are provided by Cohen (1987). In addition, the counterintuitive use of a dichotomous notion of coherency is eschewed in favour of a process, in text generation, of maximizing - rather than ensuring coherency. The use of offpage text can be seen as the result of the AS level process of maximizing coherency. One method of improving coherency is to remove superfluous text (ie. text which, in an argument, contributes only a small degree of support). However, as discussed in (Reed and Long, 1997), 3
The paragraph break is further complicated by the fact that it is not solely dependent upon focus shifts at appropriate levels of abstraction: it also affected by heuristics effecting restrictions on, for example, minimum paragraph length (- it is usually inappropriate to have a paragraph of just one sentence). Further discussion of these heuristics is provided below.
this pruning may be prohibited by concurrent, conflicting heuristics aiming to maximize the persuasive effect of an argument (by including as much support as possible). One resolution of this conflict is to marginalize the text by marking it as a footnote (or parenthetical if there is very little material, or appendix if there is a great deal). This process is conducted entirely at the AS level, again abstracted from the later realization into particular RST relations. As a result, footnotes are not restricted to expressing any single RST relation.
Quotations The explicit notions of support and authority employed at the AS level facilitate the generation and indication of quoted passages. As discussed in (Reed et al., 1996b), an argument which employs the affidavit of some third party needs to convince the hearer that the authority is reliable (ie. to show clearly the grounding of the belief). Although it is the AS level which determines the inclusion of secondary evidence, it is the subordinate Eloquence Generation (EG) level which marks areas of the plan for quotation formatting, on detection of MAKE_SALIENT goals which express direct secondary evidence. It is also the EG level which heuristically selects between the two quotation styles (viz. inline and italicized, or indented and vertically separated), on the basis of quotation length. Due to the hierarchical nature of the plan, the quotation markup performed on the topmost node (ie. the MAKE_SALIENT) gets propagated through the planning 4 process . This results in each component (at any level of abstraction) of a quotation being marked as part of that quotation. Commonly, the lower levels will realize a quotation by placing quotes before the first linguistic primitive contained within the quote, and after the last primitive (and perhaps following the close quote with a reference). The markup indicating quotation also licenses the introduction of square brackets and ellipsis (though generation of such phenomena is not currently considered within this work). Alternatively, if the EG specified that the quote should be inline, the lower levels would have the option of realizing the goal into indirect speech (though the sophistication required to generate structures such as those analyzed in (Doran, 1996) would rest entirely below the functionality of the AS and EG levels).
Italics Other than their use for inline quotations, italics are also employed for emphasizing phrases, defining terms, and indicating foreign words and titles. As Hovy and Arens (1991) point out, non-quote italics act at a level too finegrained for accounts based upon theories of text coherence such as RST to impinge upon. However, although rather less principled than the structuring of breaks and footnotes, the appropriate introduction of some italic markup can be 4
In a manner comparable to feature propagation of linguistic analysis.
performed by heuristics at the EG level. For example, italics for definitions rely in part upon parameters active at the EG level (such as hearer’s technical competence), and in part upon the presence of MAKE_SALIENT goals with definitional content. In addition, although italicizing for emphasis is particularly difficult to account for, detection of antithetical construction offers one means for the EG level to introduce emphasis through italics (eg. “This is mine; that is yours”). The EG level is generally responsible for making hearerdependent decisions for the realisation of most forms of high level punctuation. As Hovy and Arens (1991) note, “After all, the same sentence can properly be a footnote in one text, and a parenthesized part of the text proper in another”. Where such options are the result of differences between hearers (ability to follow complex argument, attention span, etc.), it is heuristics at the EG level which determine the most appropriate surface realisation. Similarly, the use of breaks may also be hearer-dependent (for similar reasons: a hearer who is content to follow complex argumentation may require fewer paragraph breaks, for example). Thus the realisation of higher level punctuation is ultimately the responsibility of the EG level, for it is vital that punctuation, as with all other aspects of the text, is tailored to the audience.
Conclusions This work has shown that classes of punctuation can be seen to fall into a hierarchy, comparable with the stages of the pipelined approach of NLG. At the highest levels of this hierarchy, the generation of punctuation relies upon abstract goals expressing intentional content. RST is recognized to be inadequate for representing intentional goals, and this is posited as a reason why RST-based systems have encountered problems in the generation of some forms of punctuation. The approach outlined in this paper relies upon explicit handling of topic manipulation. This facilitates straightforward introduction of coarse-grained formatting such as chapter, section and paragraph breaks. Furthermore, the abstract representation employed also enables a less restricted use of footnotes than previously available. Lastly, consideration of hearer-specific parameters has been shown to be crucial in the final surface realisation of punctuation. The framework described has been designed for the generation of natural language argument, and it was to this end that an abstract, intention-laden representation was employed to which RST was subordinate. However, it has been demonstrated that such a representation is appropriate for the flexible generation of the full range of formatting and punctuation devices, regardless of domain.
Acknowledgments This work has been partly funded by EPSRC grant no. 94313824.
References Baccus F. and Yang, Q. (1992). The expected value of hierarchical problem-solving. In Proceedings, National Conference on AI (AAAI’92). Briscoe, E. (1996). The Syntax and Semantics of Punctuation and its Use in Interpretation. In Working Notes of the ACL-SIGPARSE Workshop on Punctuation in Computational Linguistics, Santa Cruz, CA. Cohen, R. (1987). Analyzing the Structure of Argumentative Discourse. Computational Linguistics, 13(1): 11-24. Doran, C. (1996). Punctuation in Quoted Speech. In Working Notes of the ACL-SIGPARSE Workshop on Punctuation in Computational Linguistics, Santa Cruz, CA. Freeman, J.B. (1991). Dialectics and the Macrostructure of Arguments. Foris, Berlin. Fox, M. and Long, D.P. (1995). Hierarchical Planning using Abstraction, IEE Proceedings on Control Theory and Applications 142(3). Grice, H.P. (1975). Logic and Conversation. In P.Cole and J.L. Morgan, editors, Syntax and Semantics III: Speech Acts, pp 41-58, Academic Press, NY. Grosz, B.J. and Sidner, C.L. (1986). Attention, Intentions and the Structure of Discourse. Computational Linguistics 12(3): 175-204. Hovy, E.H. (1993). Automated Discourse Generation Using Discourse Structure Relations. Artificial Intelligence 63: 341-385. Hovy, E.H. and Arens, Y. (1991). Automatic Generation of Formatted Text. Proceedings of the 8th National Conference on AI (AAAI’91). Anaheim, CA, pp92-96. Hovy, E.H. and Wanner, L. (1996). Managing Sentence Planning Requirements. In Working Notes of the ECAI’96 Workshop on Planning and Natural Language Generation, Budapest, pp53-58. Jones, B. (1995). Exploring the Variety and Use of Punctuation. In Proceedings of the 17th Conference of the Cognitive Science Society, pp619-624. Kamp, H. and Reyle, U. (1993). From Discourse to Logic. Parts 1 and 2. Kluwer Academic, Dordrecht. Mann, W.C. and Thompson, S.A. (1987). Rhetorical structure theory: description and construction of text structures. In G. Kempen, editor, Natural Language Generation: New Results in Artificial Intelligence, Psychology and Linguistics. Martinus Nijhoff, pp 85-95 Maybury, M.T. (1993). Communicative Acts for Generating Natural Language Arguments. In Proceedings of the 11th National Conference on Artificial Intelligence (AAAI’93), pp 357-364. Marcu, D. (1996). Distinguishing between Coherent and Incoherent Texts. In Proceedings of the Student Conference on Computational Linguistics in Montreal. Montreal, pp 136-143. McConachy, R. and Zukerman, I. (1996). Using Argument Graphs to Generate Arguments. In Proceedings of the 12th European Conference on AI (ECAI’96), Budapest, John Wiley, pp592-596
Moore, J.D. and Pollack, M.E. (1992). A Problem for RST: The Need for Multi-Level Discourse Analysis. Computational Linguistics 18 (4): 537-544 Moore, J.D. and Paris, C.L. (1994). Planning Text for Advisory Dialogues: Capturing Intentional and Rhetorical Information. Computational Linguistics 20(4): 651-694. Moser, M. and Moore, J.D. (1996). Towards a Synthesis of Two Accounts of Discourse Structure. Computational Linguistics 22(3):409-419 Nunberg, G. (1990). The Linguistics of Punctuation. CSLI Lecture Notes No. 18, Menlo Park, CA. Pascual, E. and Virbel, J. (1996). Semantic and Layout Properties of Text Punctuation. In Working Notes of the ACL-SIGPARSE Workshop on Punctuation in Computational Linguistics, Santa Cruz, CA. Peh, L.S. and Ting, C. (1996). A Divide-and-Conquer Strategy for Parsing. In Working Notes of the ACLSIGPARSE Workshop on Punctuation in Computational Linguistics, Santa Cruz, CA. Reed, C.A., Long, D.P. and Fox, M. (1996a). An Architecture for Argumentative Discourse Planning. In Practical Reasoning, edited by D. Gabbay, and H.J. Ohlbach. Springer Verlag, Berlin, 555-566. Reed, C.A., Long, D.P., Fox, M. and Garagnani, M. (1996b). Persuasion as a Form of Inter-Agent Negotiation. In Working Notes, 2nd Australian Workshop on Distributed AI, Cairns, 13-27. Reed, C.A., Long, D.P. and Fox, M. (1997, to appear). Context and Focusing in Argumentative Discourse Planning. Proceedings of the International Conference on Context, Rio de Janeiro. Reed, C.A. and Long, D.P. (1997, to appear). Ordering and Focusing in an Architecture for Persuasive Discourse Planning. In Proceedings of the 6th European Workshop on Natural Language Generation (EWNLG’97), Duisberg, Germany Sacerdoti, E.D. (1977). A Structure for Plans and Behaviour, Elsevier, Amsterdam. Say, B. and Akman, V. (1996) Information-Based Aspects of Punctuation. In Working Notes of the ACL-SIGPARSE Workshop on Punctuation in Computational Linguistics, Santa Cruz, CA. Simad, M. (1996). Considerations on Parsing a Poorly Punctuated Text in French. In Working Notes of the ACL-SIGPARSE Workshop on Punctuation in Computational Linguistics, Santa Cruz, CA. Smith, M.H., Garigliano, R. and Morgan, R.C. (1994). Generation in the LOLITA System: An Engineering Approach. In Proceedings, 7th International Workshop on Natural Language Generation (INLG'94), Kennebunkport, Maine. Young, R.M. and Moore, J.D. (1994). DPOCL: A principled approach to discourse planning. In Proceedings, 7th International Workshop on Natural Language Generation (INLG'94), Kennebunkport, Maine, 13-20.
