<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1043">
  <Title>DISCOURSE AND COItESION IN EXPOSITORY TEXTI ~</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
DISCOURSE AND COItESION IN EXPOSITORY TEXTI ~
</SectionTitle>
    <Paragraph position="0"> l, Background and Introduction This paper discusses tile role of disconrse in expository text; text which typically comprises published scholarly papers, textbooks, proceedings of conferences, and other highly stylized documents. Our purpose is to examine the extent to which those discourse-related phenomena that generally assist the analysis of dialogue text -- where speaker, hearer, and speech-act information are more actively inwllved in the identification of plans and goals - can be used to help with the analysis of expository text.</Paragraph>
    <Paragraph position="1"> In particular, we make the optimistic assmnption that expository text is strongly connected; i.e., that all adjacent pairs of clauses in such a text are connected by 'cohesion markers,' both explicit and implicit. We investigate the impact that this assmnption may have on the depth of understanding that can be achieved, rite nnderlying semantic structures, aud the supporting lcnowledge base for the analysis. An application of this work in designin~g the M-based machine translation nmdel, TRANSLA-TOR, is discussed in NIRENBURG ET AL (1986) which appears elsewhere in this volume.</Paragraph>
    <Paragraph position="2"> When we read all expository text, our intuition relies on some basic assumptions about its coherence. That is, we normally expect the series of concepts to flow naturally from one sentence to the next. Moreover, when a conceptual discontinuity ocmn's at some point within the text, we are sometimes given all explicit syntactic clue (like. 'on the other hand') that such will occnr. More often, however, we are not given snch a nine, we are expected to automatically detect this shift of focus without requiring , any explicit prompting.</Paragraph>
    <Paragraph position="3"> Most of the research in tile field of discourse analysis uses texts which are dialogues; two or more people are involved, speaker and hearer roles are constantly changing, and speech-act (speaker's intention) infermarion is a changing and essential factor in tile semantics of the dialogue. For instance, extensive work has been published by LONGRACE (1977), PHILLIPS (1977), REICHMAN (1984, 1985), JOSHI ET AL. (1981), and GRIMES (1978). Although expository text does not typically contain dialogues, techniques of discourse analysis appears nevertheless to contribute strongly to the Another area of research that directly bears upon the present prob lem is the notion of textual coherence. According to HOBBS (1976), an utterance is coherent if it is an action within the implementation of some plan. In particular, conversation may be characterized as all expression of planned behavior with goals, and is thus coherent in this sense. Hobbs describes four classes of coherent conversational moves that can occur in a dialogue: Occasion (cause or enablement), Evaluation, Explanation, and Expansion. In each of these moves, the speaker's goat is to manipulate the inference process of the hearer, so that tile latter links what he/she already knows with what is new in the message. We shall illustrate that tile same premise can serve as a starting point for identifying and characterizing coherence in an expository text.</Paragraph>
  </Section>
  <Section position="2" start_page="0" end_page="181" type="metho">
    <SectionTitle>
2. Overview of TRANSLATOR
</SectionTitle>
    <Paragraph position="0"> TRANSLATOR is file name given to an ongoing research project at Colgate University which attempts to define a basis for muttilingual machine translation by using a universal intermediate metalanguage, or 'interliugua,' at iis heart. The idea is to design an interlingua which is robust enough to represent sufficient syntactic, semantic, and pragmatic knowledge about a text in any source language, so that its translation into a different target language can proceed independently of the original text.</Paragraph>
    <Paragraph position="1"> A more thorough introduction to TRANSI.ATOR can be found th TUCKER AND NIRENBURG (1984) and NIRENBURG ET AL (1986).</Paragraph>
    <Paragraph position="2"> f This material is b&amp;sed ilDon work suplx~rted by tile National Science Poundation under Grant DCR-8407114.</Paragraph>
    <Paragraph position="3"> In this paper', we limit ourselves to exploring those discourse-related phenomena which appear ill expository text, and suggesting how these phenomena may be captured during the analysis of a text and represented in tile intertingua itself. To support this exploration, we use those parts of tile interlingua for TRANSLATOR which are relevant to discourse mmlysis, and identify their rote in the analysis process. The use of italics in the paragraphs below denotes a concept which has a precise definition and connotation within iuterlingoa itself.</Paragraph>
    <Paragraph position="4"> An interlingua text may be either a single interlingua sentence or a series of sentences connected by discourse operators d. More formally: text :: = sentence \[ d (text text) The discourse operators d are enumerated and briefly described below; their' meanings are more fully described in a later section.</Paragraph>
    <Paragraph position="5"> Discourse Operator (d) Use in 'd (textl text2)' -simil change in topic from textl to text2 -I simiI continuation of same topic expan expansion -expan generalization temp temporal sequence condi conditional (cause or enablement) compare compa rison equ!y __ ~tiva!enPSe ................</Paragraph>
    <Paragraph position="6"> An interlingua sentence is con\]posed of a series of clauses, together with its own characteristic subworld, modality, focus, and speech-act thfbrmation. null Witltottt going into fro'thor detail \[see NIRENBURG ET At. (1986) for filrther description\], we note that this representation abandons tile traditional phrase-structure, dependency or other pnrely syntactic basis for representation, in favor of a far deeper level of representation for rnecharlical understanding.</Paragraph>
    <Paragraph position="7"> 3. Focus Shift ill F, xpository Text In expository text, the speaker and hearer roles are more or less permanendy assigned to the author and the reader, respectively. Tile exposition is permanently under the control of the author, add the reader plays a more or less passive role throughout. Still, speech act information plays a role in this setting, in the following ways: Definitions, as in 'Data that i,; stored more or less permanently in a computer we term a database. &amp;quot; Opinions, as in 'We agree with the point of view that software piracy is illegal.' Facts, as in 'The Symbolics LISP machine can have up to 8 megabytes of memory.' Promises, as ill 'We shall explain this subject more fully in Chapter 8.' Advice, as in 'If you are not interested in the theoretical foundations of database management systems, you may wish to skip the next section.' null Questions, as in 'What is the tradeoff between flexibility and efficiency in comparing the relational and hierarchical database models?' Commands, as in 'You should answer the following questions before proceeding to Chapter 2.' Some of these speech acts are directly related to tile topic under discussion, while others serve only to guide the reader through his/her planning and goal-setting activities while reading the text.</Paragraph>
    <Paragraph position="8"> Tile identification of focus shift is enabled by both the underlying knowledge base and the discourse-related phenomena that appear in the text itself. At the outset of analysis, the text is viewed as a sequence of sentences, made up of clauses, each one containing a single focus, which may be either an object or an event. Both objects and events have flamelike l'epresentations and are derived from information stored in an underlying knowledge base. Tile knowledge base is assumed to be structured, so that relationships among specific kinds of objects and events are  revealed. These include, for instance, 'isa,' 'part-of,' 'be-agent-of,' and other links that tend to explain how primitive and compound events and objects are interrelated in the world.</Paragraph>
    <Paragraph position="9"> A focus shift between adjacent sentences or clauses serves to signal the author's attempt to transfer the reader's attention from the given information to the new information that will be added to the presentation. The syntactic context within which such a shift might take place is arbitrary. For instance, consider the following two examples:  1. The data is shown below. Notice that some values are missing. 2. When data has missing values, it is called 'sparse'.</Paragraph>
    <Paragraph position="10">  The first shows a shift from the focus 'data' to the focus 'missing values.' The second shows a shift from the focus 'data' to the focus 'sparse.' These illustrations show that the kind of shift that takes place between two adjacent loci in a text may val~j. In the first sentence, the shift was one of expansion, while the shift in the second sentence was one of generalization. null From a strictly syntactic point of view, we see then that focus shift can take place regularly between adjacent clauses (sentence 2 above), adjacent sentences (sentences 1 above), and larger units of text which are adjacent. Thus, the network of focus shifts within a text may be complex.</Paragraph>
  </Section>
  <Section position="3" start_page="181" end_page="182" type="metho">
    <SectionTitle>
4. Defining Discourse Cohesion Relations
</SectionTitle>
    <Paragraph position="0"> The relations defined below are designed to provide a vehicle exposing the discourse structure of expository text. These relations are a variation of those developed by REICHMAN (1984) and HOBBS (1976); they differ because they are especially adapted for use in expository, rather than dialogue, types of text. The 'discourse cohesion relations' that can exist between two adjacent units of text cl and c2 (which in turn may be clauses, sentences, or larger texts) are defined and illustrated as follows: TEMPORAL: temp(cl,c2) is true if there is a temporal relationship between cl and c2. For instance, the sentences 'It became overcast. It began to rain.' exhibit a link between the concepts of cloud cover and raining, in the sense that one happened before the other.</Paragraph>
    <Paragraph position="1"> CONDITIONAL: condi(cl,c2) is true if cl either causes or enables c2 to occur. For instance, the adjacent sentences 'It began to rain.</Paragraph>
    <Paragraph position="2"> John went indoors.' exhibit a cause-and-effect relationship between two conceptual actions, raining and going indoors.</Paragraph>
    <Paragraph position="3"> EXPANSION: +expan(cl,c2) is true if c2 serves as an example or a further explanation of cl. For instance, the sentences 'The data is shown below. Notice that some values are missing.' exhibit this conceptual relationship.</Paragraph>
    <Paragraph position="4"> GENERALIZATION: -expan(cl,c2) is true if c2 serves as a generalization of cl, such as in a definition. In the sentence, 'The software that allows a person to use and/or modify this data is called a DBMS,' the new concept DBMS is defined for the first time in the text, using refinements of another concept 'software' that occur through the discourse cohesion relation +expau. That is, if we identify 'software' as concept cl, 'allowing a person to use and/or modify data' as concept c2, and 'DBMS' as concept c3, then we see that rite refined concept, say cl', results from +expan(cl,c2), and the new concept c3 results as from cl' through generalization; that is, -expan(cl',c3), or -expan( + expan(cl ,c2),c3).</Paragraph>
    <Paragraph position="5"> CONTRASTIVE: -simil(cl,c2) is true if c2 is either dissimilar or opposite from cl. For instance, consider the sentence, 'In accessing a database, the user gives English-like commands rather than Pascal-like algorithms.' Let cl denote the concept of 'accessing a database,' c2 denote the (refined) concept of 'the user giving English-like commands,' and c3 denote the concept of'the user giving Pascal-like algorithms.' Then we have the contrastive relation appearing in the following conceptual refinements: cl'= +expan(cl,c2) and cl&amp;quot;=-expan(cl',c3). That is, c3 serves to refine the concept cl' by providing a counterexample from that which was provided in the original refinement of cl by c2.</Paragraph>
    <Paragraph position="6"> SIMILAR: +simil(cl,c2) is true if c2 is similar, but not explicitly identical, to cl. For example, consider the two sentences, ' One role of a DBMS is to provide quick access. That is, we want the user to be able to access any item in the database within a few seconds of response time.' If we tel these two represent the  concepts cl and e2, respectively, we see that c2 is an approximately identical restatement of cl, and so + simil(cl,c2) is true.</Paragraph>
    <Paragraph position="7"> EQUIVALENT: equiv(cl,c2) is true if we can further ascertain that c2 is equivalent, or conceptually identical, to cl. Often this equivalence is marked by an explicit sign of synonymy, such as the parentheses in the following example. 'The software that allows the user to access this data is called a database management system (DBMS).' Here, equivalence is marked between the newly-defined concept 'database management system' and the acronym DBMS.</Paragraph>
    <Paragraph position="8"> DIGRESSION: none(cl,c2) is true if none of the other relations listed above exist between cl and c2.</Paragraph>
  </Section>
  <Section position="4" start_page="182" end_page="182" type="metho">
    <SectionTitle>
5. Inferring Focus Shift and Discourse Relations
</SectionTitle>
    <Paragraph position="0"> Following the definition of these discourse cohesion classes, it is necessary to identify some principles upon which the discourse structure may be revealed in the text as analysis progresses from the first sentence forward. That is, at any point in the reading of a text, the system must understand 'what's going on' in the sense of its discourse structure.</Paragraph>
    <Paragraph position="1"> Letting cl and c2 again denote a pair of items which appear adjacent to each other in a text, the following principles can be used to identify focus shift, based on the discourse cohesion relations that can occur  between cl and c2.</Paragraph>
    <Paragraph position="2"> 1. If cl is followed by c2 and + expan(cl,e2) is true, then a focus shift from cl to cl' takes place. That is, c1' is an embellishment of cl due to the relationship + expan and the supporting concept c2.</Paragraph>
    <Paragraph position="3"> 2. Similarly, the relation -simil(cl,c2) yields the focus shift from el to the embellishment cl'.</Paragraph>
    <Paragraph position="4"> 3. If cl is followed by c2 and -expan(cl,c2) is true, then the focus shift from cl to c2 takes place. That is, cl relinquishes its role as the focus of discourse to c2 by the process of generalization.</Paragraph>
    <Paragraph position="5"> 4. Similarly, each one of the relations condi(cl,c2), temp(cl,c2), and none(el,c2) yields a..focus shift from cl to c2.</Paragraph>
    <Paragraph position="6"> 5. On the other hand, the relations +simil(cl,c2) and equiv(cl,c2)  cause no shift to take place; that is, cl remains the focus of discourse after e2 has been processed in each case.</Paragraph>
    <Paragraph position="7"> Connectivity between adjacent concepts in a text is sometimes explicitly revealed by the presence of 'clue words' and other markers. The use of clue words for discourse analysis is common (eg REICHMAN (1984)). The example text discussed in the following section contains several such clue words. Sometimes the marker appears as a punctuation mark (such as a parenthetical which signals the relation +equiv), oilier instances appear as single words (such as 'However&amp;quot; signaling -simil), while still others are complete clauses (such as 'there may be far less' signaling + simil).</Paragraph>
    <Paragraph position="8"> Yet, many instances of conceptual connectivity are not cued by the presence of such markers; the are revealed instead by general syntactic structure (such as the appearance of a relative clause, signaling +expan) or by semantic properties that are possessed by the underlying concepts and stored in the knowledge base. The following discussion suggests how such knowledge can be used to mark instances of conceptual connectivity in expository text.</Paragraph>
    <Paragraph position="9"> Intuitively, some of the conceptual properties that reveal discourse cohesion relations are the following:  Merging these conceptual clues with the explicit syntactic clues for discourse connectivity, leads ~ the following table. This table reveals some of the clues (both explicit and implici0 that lead to exposure of the cohesion relation d(cl c2), where cl attd c2 are adjacent concepts (processes or objects) within the text.</Paragraph>
    <Paragraph position="11"> A simple algorithm to infer such relations between pairs of concepts in the text, ci atnl cj, can be given. However, space does not permit its further elaboration in this paper.</Paragraph>
  </Section>
  <Section position="5" start_page="182" end_page="182" type="metho">
    <SectionTitle>
6, An Example
</SectionTitle>
    <Paragraph position="0"> To illustrate the application of these ideas, we have analyzed the five sentences of a paragraph taken from the first page of Jeffrey Ullman's book, Principles of Database Systems, given below in a specially annotated form. The annotations C, S, and D on the left denote clauses, sentences, and discourse cohesion markers that are uncovered in a parse of this paragraph. null  rather than ... \[to allow the user to deal with the data\] as the computer stores the data.</Paragraph>
    <Paragraph position="1"> In this sense, C13 the DBMS acts as C14 an interpreter for a lfigh-level programming language, C15 ideally allowing the user to specify what must be done, with little or no attention on the user's part C16 to the detailed algorithms or data  C18 in the case of a DBMS, D there may be far less C19 relationship between the data as seen by the user and ...\[the data\] as stored by the computer D than C20 ...\[the relationship\] between, say, arrays as defined in a typical programming language and the representation of those arrays in memory. While space does not permit a detailed description of the analysis of this text, below is a summarization of the final result of such an analysis.  Here, we note that each sentence has inherited a focus, and file remaining connectives and semantic properties can later be used to expose the overall discourse structure of the paragraph.</Paragraph>
  </Section>
class="xml-element"></Paper>