File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/00/w00-1411_metho.xml
Size: 22,813 bytes
Last Modified: 2025-10-06 14:07:25
<?xml version="1.0" standalone="yes"?> <Paper uid="W00-1411"> <Title>An integrated framework for text planning and pronominalisation</Title> <Section position="3" start_page="77" end_page="78" type="metho"> <SectionTitle> 2 Reconstructing Centering for NLG </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="77" end_page="77" type="sub_section"> <SectionTitle> 2.1 Centering in a nutshell </SectionTitle> <Paragraph position="0"> The main assumptions of Centering theory are: 1. For each utterance in a discourse there is precisely one entity which is the centre of attention or center. The center in an utterance Un is the most grammatically salient entity realised in U~_i which is also realised in Un. This is also referred to as the backward-looking center or Cb. The notion of &quot;salience&quot; for the purposes of centering theory is most commonly defined according to a hierarchy of grammatical roles: SUBJECT > DIRECT OBJECT > INDIRECT OBJECT ~> OTHERS (see e.g., Brennan et al 1987)* For alternative approaches see e.g., (Strube and Hahn 1999), (Walker et al 1994).</Paragraph> <Paragraph position="1"> 2. There is a preference for consecutive utterances within a discourse segment to keep the same entity as the center, and for the center to be realised as Subject or preferred center (Cp).</Paragraph> <Paragraph position="2"> Kibble (1999) dubbed these principles cohesion and salience respectively. Pairs of successive utterances (U,~, U~+i} are classified into the transition types shown in Fig. 1, in the or.der of preference specified by Grosz et al's &quot;Rule ,, * 3. The center is the entity which is most likely to be pronominalised: Grosz et al's &quot;Rule 1&quot; in its weakest form states that if any entity is referred to by a pronoun, the Cb must be.</Paragraph> <Paragraph position="3"> CONTINUE.&quot; cohesion and salience .... both hold; same center (o r Cb(Un) undefined), realised as Subject in Un+l; RETAIN.&quot; cohesion only; i.e. center remains the same but is not realised as Subject in Un+l; SMOOTH SHIFT.&quot; salience only; center of Un+l realised as Subject but : , not equal .to,Cb(U~); ROUGH SHIFT.&quot; neither cohesion nor salience holds.</Paragraph> <Paragraph position="4"> NO CB: this transition is used by some researchers but is not discussed by GJW.</Paragraph> </Section> <Section position="2" start_page="77" end_page="78" type="sub_section"> <SectionTitle> 2.2 Pronominalisation </SectionTitle> <Paragraph position="0"> Text genres are known to vary in the extent to which pronouns are used. The CT-based framework allows us to experiment with different strategies for choosing when to use a pro- null noun, including: 1. Never use anaphoric pronouns -- for instance, in certain legal documents or technical manuals where there must be no possibility of ambiguity.</Paragraph> <Paragraph position="1"> 2. Always pronominalise the Cb.</Paragraph> <Paragraph position="2"> 3. Use a pronoun for non-Cbs only if the Cb is pronominalised.</Paragraph> <Paragraph position="3"> 4. Pronominalise the Cb only after a Continue transition.</Paragraph> <Paragraph position="4"> Strategy 3 is favoured by (GJW) who cite psychological evidence that &quot;the Cb is preferentially realised by a pronoun in English and by equivalent forms (i.e., ,.zero pronouns) in other languages&quot; (op cit., p. 214). However, in the implementation reported in section 3 we opt for the more restrictive strategy 4. The generation approach outlined below enables us to experiment with different strategies and compare the resulting texts.</Paragraph> </Section> <Section position="3" start_page="78" end_page="78" type="sub_section"> <SectionTitle> 2.3 Centering and discourse structure center after a sequence of CONTINUE. How- </SectionTitle> <Paragraph position="0"> The canonical formulation of CT is concerned ever, in a sequence CONTINUE-RETAIN-SHIFT .' with local* cohesion;--'specifying .aqment-,. t~rarisi~':&quot;::':the:sHng'~:is:p redicted:Am:its:Adegcal':cdegntext~ but tions between consecutive sentences in a dis- the RETAIN is not; whenever RETAIN is a cheap course segment and favouring sequences which maintain the same center. Our implementation incorporates two extensions which have implications for more structured discourse: Strube and Hahn's (1999) &quot;cheapness&quot; principle, which favours transitions that introduce a new topic * in. a salient position, ,and .Cristea's Veins Theory (Cristea et al 1998) which redefines transitions in terms of rhetorical hierarchy rather than linear sequence of clauses (see section 3.3 for discussion).</Paragraph> <Paragraph position="1"> &quot;Cheapness&quot; is satisfied by a transition pair ((Un-1, Un), (Un, Un+l)) if the preferred center of Un is the Cb of Un+l. For example, this test is satisfied by a RETAIN-SHIFT sequence but not by CONTINUE-SHIFT, so it is predicted that the former pattern will be used to introduce a new center. (This claim is consistent with the findings of Brennan 1998, Brennan et al 1987.) If we consider examples la-e below, the sequence cd'-e ~, including a RETAIN-SHIFT sequence, reads more fluently than c-d-e even though the latter scores better according to the canonical ranking. null . a. John has had trouble arranging his vacation. null b. He cannot find anyone to take over his responsibilities.</Paragraph> <Paragraph position="2"> c. He called up Mike yesterday to work out a plan. CONTINUE d. He has been pretty annoyed with Mike recently. CONTINUE e. Mike cancelled last week's project meeting at short notice.</Paragraph> </Section> </Section> <Section position="4" start_page="78" end_page="78" type="metho"> <SectionTitle> EXP-SMOOTH SHIFT </SectionTitle> <Paragraph position="0"> d'. Mike has mmoyed him a lot recently.</Paragraph> </Section> <Section position="5" start_page="78" end_page="80" type="metho"> <SectionTitle> RETAIN </SectionTitle> <Paragraph position="0"> e I. He cancelled last week's project meeting at short notice. SMOOTH SHIFT The &quot;cheapness&quot; principle illustrates the need for global opfimisation. We noted above that there is evidence that a RETAIN-SHIFT sequence is the preferred way of introducing a new transition following CONTINUE, another CONTINUE would be cheap as well. The RETAIN is motivated as it enables a &quot;cheap&quot; SMOOTH SHIFT, and so we need a way of evaluating the whole sequence CONTINUE-RETAIN-SHIFT ver-</Paragraph> <Paragraph position="2"> CT has developed primarily in the context of natural language interpretation, focussing on anaphora resolution (see e.g., Brennan et al 1987). Curiously, NLG researchers have tended to overlook GJW's proposal that Rule 2 provides a constraint on speakers, and on natural-language generation systems ...To empirically test the claim made by Rule 2 requires examination of differences in inference load of alternative multi-utterance sequences that differentially realize the same content.</Paragraph> <Paragraph position="3"> GJW, p. 215.</Paragraph> <Paragraph position="4"> With a few exceptions (e.g., Mittal et al 1998, Kibble 1999, Kibble and Power 1999, Cheng 2000) NLG researchers have interpreted CT as a theory of pronominalisation only (e.g., Dale 1992). In this paper we concentrate on planning, aiming to determine whether the prim ciples underlying the constraints and rules of the theory can be &quot;turned round&quot; and used as planning operators for generating coherent text. Text planning in conformity with CT will need to follow the following set of heuristics: 1. Plan tile order of clauses so that adjacent clauses have at least one referent in cornIlion. null 2. Cohesion: Prefer orderings which maintain the same Cb in successive clauses. ,3..- Salience: .Realise as=SubjeCt- of U;~ tile most grammatically salient entity in U,~-i which is mentioned in Un (the Cb).</Paragraph> <Paragraph position="5"> 4. Cheapness: Realise as Subject of Un an entity which is mentioned in U,~+l (and will therefore be Cb of U,,+i).</Paragraph> <Paragraph position="6"> Breaking down the problem like this reveals ferent transitions. We assume that certain that there are various ways the distinct tasks options for syntactic realisation can be pre...... can. be slotted, into.-an.NLG,system~Cohesion. . ........ ._~dicted.,ma::t~he,~basis~,of:,the~axgu~ment ~ ~str:ucnaturally comes under Text Planning: ordering a sequence of utterances to maintain the same entity as the center, within constraints on ordering determined by discourse relations. However, identifying the center depends on grammatical salience, which is normally determined by the Sentence Planner- for example, choice of active or passive voice. Three possibilities are: (r) &quot;Incremental&quot; sentence-by-sentence generation, where the syntactic structure of Un is determined before the semantic content of Un+l is planned. That is, the Text Planner would plan the content of Un+l by aiming to realise a proposition in the knowledge base which mentions an entity which is salient in Un. We axe not aware of any system which performs all stages of generation in a sentence-by-sentence way, and in any case this type of architecture would not allow the cheapness principle to be implemented as it would not support the necessary forward planning.</Paragraph> <Paragraph position="7"> * A pipelined system where the &quot;topic&quot; or &quot;theme&quot; of a sentence is designated independently as part of the semantic input, and centering rules reflect the information structure of a discourse. This approach was suggested by Kibble (1999), proposing that text and sentence planning should be driven by the goal of realising the designated topic in positions where it will be interpreted as the Cb. However, this is not really a solution so much as a refinement of the problem, since it simply shifts the problem of identifying the topic. Prince (1999) notes that definitions of &quot;topic&quot; in the literature do not provide objective tests for topichood and proposes that the topic should be identified with the centre of attention as defined by CT; however, what would be needed here would be a more fimdamental definition which would, account for a particular entity being chosen to be tile centre of attention.</Paragraph> <Paragraph position="8"> o The solution we adopt is to treat tile task of identifying Cbs as an optimisation problem, giving different weightings to t, he difture of predicates, which means that centering transitions can be calculated as part The text planner has been developed within ICONOCLAST, a project which investigates applications of constraint-based reasoning in Natural Language Generation using as subject-matter the domain of medical information leaflets. Following (Scott and de Souza 1990), we represent rhetorical structure by graphs like figure 2, in which non-terminal nodes represent RST relations, terminal nodes represent propositions, and linear order is unspecified. The task of the text planner is to realize the rhetorical structure as a text structure in which propositions are ordered, assigned to textual units (e.g., sentences, paragraphs, vertical lists), and linked where appropriate by discourse connectives (e.g., 'since', 'however').</Paragraph> <Paragraph position="9"> Even for a simple rhetorical input like figure 2 many reasonable text structures call be generated. Since there are two nucleus-satellite relations, tile elementary propositions can be ordered in four ways; several discourse connectives can be employed to realize each rhetorical relation (e.g. concession can be realized by 'although', 'but' and '.however'); at one extreme, the text can be spread out over several paragraphs, while at the other extreme it can be squeezed into a single sentence. With fairly restrictive constraint settings, the system generates 24 text-structure patterns for figure 2, including the following (shown schematically): A. Since contain(elixir, gestodene), ban(fda, 3.1 Choosing centers elixir). However, approve(fda, elixirplus). Given a text structure, we enumerate all per-B. approve(fda, elixirplus), although since contain(elixir, gestodene ) , ban ( f da, elixir). The final output texts will depend on how the propositions are realized syntactically; among other things this will depend on centering choices within each proposition.</Paragraph> <Paragraph position="10"> In outline, the procedure that we propose is as follows: ~ .</Paragraph> <Paragraph position="11"> . Enumerate all text structures that are acceptable realizations of the rhetorical structure. null</Paragraph> <Paragraph position="13"> For each text structure, enumerate all permissible choices for the Cb and Cp of each proposition.</Paragraph> <Paragraph position="14"> Evaluate the solutions, taking account of referential coherence among other considerations, and choose the best.</Paragraph> <Paragraph position="15"> For the example in figure 2, centers can be assigned in four ways for each text-structure pattern, making a total of 96 solutions.</Paragraph> <Paragraph position="16"> As will probably be obvious, such a procedure could not be applied for rhetorical structures with many propositions. For examples of this kind, based on the relations 'cause' and 'concession' (each of which can be marked by several different connectives), we find that the total number of text-structures is approximately</Paragraph> </Section> <Section position="6" start_page="80" end_page="83" type="metho"> <SectionTitle> 5 N-~ for N propositions. Hence with N = 4 we </SectionTitle> <Paragraph position="0"> would expect around 600 text structures; with perhaps 5-10 ways of assigning centers to each text structure, the total number of solutions would approximate to 5000. Global optimization of the solution therefore becomes impracticable for texts longer than about five propositions; we address this problem by a technique of partial optimization in which a macro-planner fixes the large-scale structure of the text, thus defining a set of micro-planning problems each small enough to be tackled by the methods described here.</Paragraph> <Paragraph position="1"> Stage 1 of the planning procedure is described elsewhere (Power, 2000); we focus here on stages 2 and 3, in which the text planner enumerates the possible assignments of centers and evaluates which is the best.</Paragraph> <Paragraph position="2"> missible centering assignments as foil0ws: &quot; ..... 1. Determine the predecessor Yn-1 (if any) of each proposition Un.</Paragraph> <Paragraph position="3"> 2. List the potential Cbs and Cps of each proposition, henceforth denoted by ECb and ECp.</Paragraph> <Paragraph position="4"> 3. Compute ~li combinations from ECb and ECp that respect the fundamental centering constraint that Cb(Un) should be the most salient candidate in Un-1.</Paragraph> <Paragraph position="5"> Two criteria for determining the predecessor have been implemented; the user can select one or other criterion, thus using the NLG system to test different approaches. Following a linear criterion, the predecessor is simply the proposition that precedes the current proposition in the text, regardless of structural considerations. Following a hierarchical criterion, the predecessor is the most accessible previous proposition, in the sense defined by Veins Theory (Cristea et al 1998). We will return to this issue later; for now we assume the criterion is linear.</Paragraph> <Paragraph position="6"> ECb(Un) (potential Cbs of proposition Un) is given by the intersection between Cf(U,~) and Cf(Un-1) -- i.e., all the referents they have in common. The potential Cps are those referents in the current proposition that can be realized as most salient. Obviously this should depend on the linguistic resources available to the generator; the system actually uses a simpler rule based oil case roles within the proposition. Figure 3 shows the potential Cbs and Cps for the proposition sequence in solution A.</Paragraph> <Paragraph position="7"> Our treatment of salience here simplifies ill tWO ways. First, we assume that syntactic realization serves only to distinguish the Cp from all other referents, which are ranked on the same level: thus effectively SUBJECT > OTHERS.</Paragraph> <Paragraph position="8"> Secondly, we assume that the system already .knows, from the event.class,of the proposition, which of its case roles can occur in subject position: thus for ban, both arguments are potentim Cps because active and passive realizations are both allowed; for contain, only elixir is a potential Cp because we disallow 'Gestodene is contained by Elixir'.</Paragraph> <Paragraph position="9"> With these simplifications, the enumeration of centering assignments is straightforward; in the above example, four combinations are possible, since there are two choices each for Cp(U2) and Cp(U3), none of which leads to any violation of the basic centering constraint. This constraint only comes into play if there are several choices for Cb(Un), one of which coincides with Cp(Un-1).</Paragraph> <Section position="1" start_page="81" end_page="82" type="sub_section"> <SectionTitle> 3.2 Evaluating solutions </SectionTitle> <Paragraph position="0"> Various metrics could be used in order to evaluate centering choices. One possibility, for example, would be to associate a cost with each transition, so that perhaps 'Continue' (the best transition) has zero cost, while 'No Cb' (the worst transtion) has the highest cost. However, we have preferred the approach mentioned earlier in which cohesion and salience are evaluated separately; this allows us to include the further criterion of cheapness.</Paragraph> <Paragraph position="1"> Although this paper focusses on centering issues, it is important to remember that other aspects of text quality are evaluated at the same time: the aim is to compute a global measure so that disadvantages in one factor can be weighed against advantages in another. For instance, text pattern B is bound to yield poor continuity of reference because it orders the propositions so that U1 and U2 have no referents in coinmon. Text pattern A avoids this defect, but this does not necessarily mean that A is better than B overall; there may be other reasons, unconnected with centering, for preferring B to A.</Paragraph> <Paragraph position="2"> The system evaluates candidate solutions by applying a battery of tests to each.node of the text plan. Each test identifies whether the node suffers from a particular defect. For instance, one stylistic defect (at least for the rhetorical relations occurring in figure 2) is that of placing nucleus before satellite; in general, the text reads better if important material is placed at the end. For each type of defect, we specify a weight indicating its importance: in evaluating continuity of reference, for example, the defect 'No Cb' might be regarded as more significant than other defects. Summing the weighted costs for all defects, we obtain a total cost for the solution; our aim is to find the solution with the lowest total cost.</Paragraph> <Paragraph position="3"> Regarding centering, the tests currently applied are as follows.</Paragraph> <Paragraph position="4"> This defect is recorded for any proposition with no Cb, except the first proposition in the sequence (which by definition cannot have a Cb).</Paragraph> <Paragraph position="5"> Applied to the four solutions to text structure A, with all weights equal to 1, these definitions yield costs, shown in Figure 4.-According to our metric, solutions A1 and A2 should be preferred to A3 and A4 because they incur less cost. This resutt=cml be assessed, by comparing -the following output texts, in which the generator has followed the policy of pronominalizing the Cb only after a 'Continue' transition: A1. Since Elixir contains gestodene, the FDA bans Elixir. However, it approves ElixirPlus.</Paragraph> <Paragraph position="6"> A2. Since Elixir contains gestodene, it is banned by the FDA. However, the FDA approves ElixirPlus. null A3. Since Elixir contains gestodene, the FDA bans Elixir. However, ElixirPlus is approved by the FDA.</Paragraph> <Paragraph position="7"> A4. Since Elixir contains gestodene, it is banned by the FDA. However, ElixirPlus is approved by the FDA.</Paragraph> <Paragraph position="8"> Of course we are not satisfied that this metric is the best; an advantage of the generation approach is that different evaluation methods can easily be compared.</Paragraph> </Section> <Section position="2" start_page="82" end_page="83" type="sub_section"> <SectionTitle> 3.3 Hierarchical centering </SectionTitle> <Paragraph position="0"> The linear approach, illustrated above, assigns centers on the basis of a proposition sequence, flattening the original hierarchy and ignoring nucleus-satellite relations. This means, for example, that in a text of two paragraphs, proposition U2.1 (the first proposition in the second paragraph) has to be treated as the successor to Ui.N (the final proposition of the first paragraph): even if Ui.:\, is relatively insignificant (the satellite of a satellite, perhaps). One's intuition in such cases is that some more significant proposition in the first paragraph should become the focus against which continuity of reference in the second paragraph is judged.</Paragraph> <Paragraph position="1"> Veins Theory (Cristea et al 1998) provides a possible formalization of this intuition, in which some earlier propositions become inaccessible as a rhetorical boundary is crossed. The theory could be applied to centering in various ways; we have implemented perhaps the simplest approach, in which centering transitions are assessed in relation to the nearest accessible predecessor. In many cases the linear and hierarchical definitions give tile same result, but sometimes they diverge, as in the following alternative to solutions A and B: C. ban(fda, elixir) since contain(elixir, gestodene).</Paragraph> <Paragraph position="2"> However, approve(f tin, elixirplus).</Paragraph> <Paragraph position="3"> Following Veins Theory, the predecessor of approve(f da, elixirplus) is ban(f da, elixir); its linear predecessor contain( elixir, ge.stodene ) (an embedded satellite) is inaccessible. This makes a considerable difference: under a hierarchical approach, fda can be the Cb of the final proposition; under a linear approach, this Proceedings of ANLP-NAACL 2000. proposition has no Cb. D Cristea, N Ide and L Romary, 1998. Veins ~ '. Iheory: ~ A :model of,:gtobat: discourse :cohesion</Paragraph> </Section> </Section> class="xml-element"></Paper>