File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/04/w04-2322_metho.xml
Size: 23,896 bytes
Last Modified: 2025-10-06 14:09:23
<?xml version="1.0" standalone="yes"?> <Paper uid="W04-2322"> <Title>A Rule Based Approach to Discourse Parsing</Title> <Section position="4" start_page="1" end_page="1" type="metho"> <SectionTitle> 2 The Classical Linguistic Dis- </SectionTitle> <Paragraph position="0"> course Model (C-LDM) Unlike the Discourse Structures Model (DSM) of Grosz and Sidner (1986), a pragmatic and psychological theory that aims to clarify the relationship between speakers' intentions and their focus of attention in discourse, or the rhetorical model of Rhetorical Structures Theory (Mann and Thompson, 1988) that is designed to identify the coherence relations between segments of text, the Linguistic Discourse Model (LDM) (Polanyi and Scha, 1984; Polanyi, 1988; Polanyi and van den Berg, 1996) is a syntactically informed, semantically driven model developed to provide proper semantic interpretation for every utterance in a discourse despite the apparent discontinuities that are present even in well structured written texts. In its focus on understanding discourse meaning, the LDM is close in spirit to Structured Discourse Representation Theory (S-DRT) (Asher, 1993). While S-DRT attempts to account for discourse structure purely semantically, the LDM framework is concerned to maintain a separation between discourse &quot;syntactic&quot; structure, on the one hand, and discourse interpretation on the other. Therefore, like DSM and RST, the LDM incorporates an explicit tree structured model of relationships between discourse segments as its model of discourse &quot;syntax&quot;. In discourse parsing under the LDM, any attachment to the developing discourse tree of a textual unit is treated as an instruction to update an appropriate semantic representation. We construct dynamic semantic representations (DSRs), similar to the Discourse Representation Structures (Kamp, 1981; Kamp and Reyle, 1993) used in S-DRT as its model of discourse semantics.</Paragraph> <Paragraph position="1"> The DSRs correspond to the contexts relative to which subsequent segments can be interpreted.</Paragraph> <Paragraph position="2"> The analysis of intra-sentential structure is done by sentential syntax which identifies the syntactic and semantic structures within the sentence and makes the resulting analysis available for discourse processing.</Paragraph> <Section position="1" start_page="1" end_page="1" type="sub_section"> <SectionTitle> 2.1 Overview of the Classic LDM </SectionTitle> <Paragraph position="0"> In the Linguistic Discourse Model (LDM) discourse is formed through the recursive combination of discourse constituent units (DCUs). The structure of a discourse is represented by an open right tree of DCUs. Basic discourse units (BDUs), resulting from a segmentation of the discourse according to rules of discourse segmentation, form the content of the leaves of the tree. Once a text has been segmented into BDUs, an open right tree representing the structure of the discourse is built up. The completed tree shows, for any given point in the discourse, which discourse units (DCUs) remain available for continuation and which DCUs are no longer available. Because discourse anaphora resolution is critically constrained by discourse structure, the tree representation makes clear the domain in which the antecedent for a given anaphoric referential expression is to be found.</Paragraph> <Paragraph position="1"> Antecedents must be available at a node along the right edge of the discourse tree. (Polanyi, 1985; Grosz and Sidner 1986; Webber, 1991) The LDM posits three structural relations be- null tween discourse units: 1. discourse coordination a. Units related by bearing a similar relationship to an existing or newly formed common parent in the tree (lists, narratives). null b. Available at the C-node is information common to all child nodes.</Paragraph> <Paragraph position="2"> 2. discourse subordination a. Units related by an elaboration relationship in which the subordinated unit provides more information about an entity or situation described in the subordinating unit.</Paragraph> <Paragraph position="3"> b. Units unrelated to existing units available on the right edge of the tree, viewed as intrusions or interruptions.</Paragraph> <Paragraph position="4"> c. Available at the S-node is information specific only to the subordinating or dominant constituent (usually the left child).</Paragraph> <Paragraph position="5"> 3. n-ary constructions a. Units related by logical or rhetorical, genre or interactional conventions specific to a given language.</Paragraph> <Paragraph position="6"> b. Preposed modifier, sentence initial adverbial, &quot;cue word&quot;, (reported speech) attribution phrase.</Paragraph> <Paragraph position="7"> c. Available at N-nodes is information about each constituent and the relationship connecting them.</Paragraph> <Paragraph position="8"> Although we believe that the general approach to discourse structure captured by the Classical LDM is essentially sound, there are three critical problems with the existing framework: 1. Segmenting the incoming text into BDUs 2. Determining the existing or new node at which to attach an incoming BDU 3. Determining the relationship between the incoming BDU and the attachment node Although very difficult challenges associated with each of these discourse parsing tasks remain, in developing the Unified Linguistic Discourse Model (U-LDM) we have made significant progress recently on solving them. These are discussed in Sections 3 and 4 below.</Paragraph> </Section> </Section> <Section position="5" start_page="1" end_page="5" type="metho"> <SectionTitle> 3 Discourse Segmentation </SectionTitle> <Paragraph position="0"> The problem of segmenting discourse into the elementary units appropriate for building up the structure of the discourse is an extremely difficult one. Each discourse theory must specify how &quot;segments&quot; should be identified in light of the questions the theory is set up to answer.</Paragraph> <Paragraph position="1"> Models based on Grosz and Sidner's 1986 work, especially those which form the basis of spoken language systems, define segments in terms of the intentions of the speaker: when the speaker's intention shifts, the segment associated with that intention ends and immediately following talk is included in new (or resumed) segments. While very useful in dealing with task oriented talk where speakers move between asking questions, informing others and giving commands, this model is less applicable to determining discourse segments within a sentence. The problem is an acute one for the analysis of written texts because often a subsequent, not necessarily adjacent, segment will continue the development of material introduced in a sub-sentential, often subordinate, constituent. Construction of the appropriate representation of the rhetorical or semantic structure of discourse must therefore keep sub-sentential units available for attachment at independent nodes on the tree along. The entire sentence or sentential main constituent must also be available to be continued after any continuation on sub-sentential units has been completed. As reported by Carlson et al. (2003), under RST , lexical and syntactic information used to segment discourse into Elementary Discourse Units (EDUs) is based on verbal constituents including clauses and infinitives.</Paragraph> <Paragraph position="2"> As we show below, the approach taken to segmentation under the U-LDM, while it includes as segments (and non-segments) many of the constructions currently used in RST, provides a rationalization for the choice of units. Rather than posit which syntactic objects function as discourse segments, we started by establishing the semantic basis for functioning as a segment and then identified which syntactic constructions carry the semantic information needed for discourse segment status. We then identified as Basic Discourse Units (BDUs) segments that have the potential to independently establish an anchor point for future continuation. We then drew a further distinction between BDUs as a class of syntactic structures with the potential to Under S-DRT, no explicit structural tree is constructed and no explicit segmentation criteria have been proposed in the literature.</Paragraph> <Paragraph position="3"> Although some clauses are not treated as elementary units and &quot;a small number of phrasal EDUs are allowed, provided that the phrase begins with a strong discourse marker.&quot; establish anchor points and the actual BDUs in a given sentence which can function as indexical anchor points in a specific discourse. We believe these distinctions, while cumbersome, are necessary for both theoretical and practical text analysis. null</Paragraph> <Section position="1" start_page="3" end_page="5" type="sub_section"> <SectionTitle> 3.1 Discourse Segments under the U-LDM </SectionTitle> <Paragraph position="0"> As a semantic theory, the U-LDM must account for the interpretation of utterances. Specifically, we must account for the availability for update of appropriate discourse contexts or sub-contexts introduced in earlier text. In order to do so, we must be able to match incoming discourse utterances with their target contexts, some of which may have been introduced in syntactically subordinated positions within a sentence. Therefore, in designing U-LDM discourse segmentation, we have identified the syntactic reflexes of the semantic content of the linguistic or paralinguistic phenomena making up discourse.</Paragraph> <Paragraph position="1"> Since elementary discourse units are needed to build up discourse structure recursively, we have identified as discourse segments the syntactic constructions that encode a minimum unit of meaning and/or discourse function interpretable relative to a set of contexts. We understand a minimum unit of meaning to communicate information about not more than one &quot;event&quot;, &quot;event-type&quot; or state of affairs in a &quot;possible world&quot; of some type . Clauses, and many other verb based structures, carry indexical information that ties the content to the context in which it is to be interpreted. Minimal functional units, on the other hand encode information about how previously occurring (or possibly subsequent) linguistic gestures relate structurally, semantically, interactionally or rhetorically to other units in the discourse or to information in the context in which the discourse takes place der the U-LDM are the syntactic reflex of a linguistically realized semantic &quot;gesture&quot; interpreted relative to context, they need not be contiguous, but may completely surround another segment (e.g. an appositive, or non-restrictive relative clause.) Discontinuous seg- null Roughly speaking an &quot;elementary proposition&quot;, &quot;event-type predicate&quot; etc. In a Davidsonian style semantics, quantification over an event variable signals a separate unit of meaning.</Paragraph> <Paragraph position="2"> Greetings, discourse PUSH/POP markers and other &quot;cue phrases&quot;, connectives etc. are all functional segments.</Paragraph> <Paragraph position="3"> ments occur when there is overt material on both sides of the intervening segment. With fragmentary segments, the full interpretation remains unrecoverable from surrounding context. For example: a single word answer to a question is a complete segment, whereas the same word uttered but &quot;left hanging&quot; would be an uninterpretable fragment. (See Appendix for extensive example of a segmented text.)</Paragraph> </Section> <Section position="2" start_page="5" end_page="5" type="sub_section"> <SectionTitle> 3.2 Basic Discourse Units </SectionTitle> <Paragraph position="0"> An important contrast between the U-LDM and other approaches to segmentation concerns the distinction made in the U-LDM between discourse segments such as those we have identified above and Basic Discourse Units (BDUs). While all BDUs under the U-LDM are segments, not all segments are BDUs. BDUs, under this model, are discourse segments of a type that can be independently continued: operator segments are one example of non-BDU segments. Other verb bases constituents that might be expected to be segments are not because they do not establish an interpretation context independent of other segments that can be updated by subsequent units. In general, these &quot;notable non-segments&quot;, summarized in Table 2, are heavily integrated into other nominal or verbal constructions and cannot be accessed for independent continuation.</Paragraph> <Paragraph position="1"> In answer to a reviewer who asked if in &quot;Singing is fun&quot;, singing should not be an independent In order to account for continuation in specific sentences, we further identify one class of instances of BDU: Active BDUs (A-BDUs) are BDUs on the right edge of a discourse tree. The main clause of any sentence will be an A-BDU and, depending on the deployment of BDU segments within a given sentence, other BDUs may also be accessible for continuation. (See Section</Paragraph> </Section> </Section> <Section position="6" start_page="5" end_page="11" type="metho"> <SectionTitle> 4 below.) 4 Discourse Parsing with the U-LDM </SectionTitle> <Paragraph position="0"> Ascertaining the relationship of a BDU to the discourse is a complex parsing process involving lexical, semantic, structural and syntactic information null . For the case of written prose we are concentrating on here, the unit of analysis is the sentence (or sentence fragment). Sentences are attached to the DPT of the text as a unit . Discourse attachment of the sentence involves two decisions: where along the right edge to attach, and what is the relationship to the attachment point. The process, which includes constructing a BDU tree of the sentence, can be summarized as follows: segment, we would answer that this sentences concerns one eventuality (something being fun), not two. Since any noun can be referred to by a pronoun in the next sentence simply referring to the noun is not equivalent to referring to the eventuality in which the referent of the noun is a participant.</Paragraph> <Paragraph position="1"> Although the linguistic (and lexical) information we discuss could be augmented with processes relying upon high level world knowledge and inference, we believe that it is extremely significant to see how far one can get with discourse parsing without invoking non-linguistic information. * Identify potential BDUs within sentence using sentential syntax * Construct a BDU-tree from the segments of the sentence, using sentential syntactic information and discourse rules to map segments and relationships among them. This BDU-tree is itself an Open Right Tree dominated by the node corresponding to the Main clause of sentence 9 . (This is the Main BDU or MBDU).</Paragraph> <Paragraph position="2"> * Attach the BDU-tree as a unit to the Discourse Parse Tree by computing the relationship of MBDU and preposed modifiers, if any, to accessible DCUs aligned along the right edge of the tree using rules of discourse relations (See Section 4.1 below). Lexical information used for attachment decisions can come from anywhere in the BDU tree.</Paragraph> <Paragraph position="3"> * Once the BDU-tree is attached, its terminal leaves are terminal nodes of the Discourse Parse Tree (DPT) and any terminal or intermediary nodes on the right edge of the BDU tree are DCUs on the DPT accessible for attachment in the next iteration of the process. In order to determine which accessible DCUs are candidates for M-BDU attachment and what relationship obtains between the incoming unit and the selected DCU, a number of distinct types of evidence are used, including: 1) lexical information reuse somewhere in the BDU tree of the same lexeme, synonym/antonym, hypernym, or participation in the same lexical frame or &quot;semantic field&quot; as item in target node. 2) syntactic information parallel syntactic structure; topic/focus and centering information, syntactic status of re-used lexemes, pre-posed adverbial constituents, etc.</Paragraph> <Paragraph position="4"> 3) semantic information realis status, genericity, tense, aspect, point of view etc. in the MBDU This process is too complex to describe in detail here but it involves looking at both the F-structure of the sentential parsing information returned by the XLE and applying discourse rules to the BDUs identified. Soricut and Marcu (2003) also build up RST sentential trees to use in discourse parsing. Both the information and methods used to construct RST trees as well as the trees themselves differ from ours.</Paragraph> <Paragraph position="5"> 4) constituents of incomplete n-ary constructions on the right edge Questions, initial greetings, genre-internal units like sections and sub-sections, etc. 5) structure of both the local attachment point and the BDU-tree While we are still experimenting with understanding the complexities involved in attachment, we believe that different types of evidence have different weights and that the combined weight of evidence determines the attachment point. We have noted, however even at this stage of our investigations, that the weight given to each type of information differs for attachment site selection and relationship determination. Lexical information, for example, is often very important in determining site, while semantic and syntactic information is most relevant in determining relationship. In the remainder of this section we will give a small set of robust rules for determining the attachment site and relationship of an incoming BDU-tree to the existing parse tree of the discourse.</Paragraph> <Section position="1" start_page="10" end_page="11" type="sub_section"> <SectionTitle> 4.1 Rules for Determining Discourse Attachment Site Candidates and Attachment Relations </SectionTitle> <Paragraph position="0"> Both the attachment site choice and the actual attachment process rely on partially ordered sets of hybrid rules, each of which are conditioned on a set of constraints. Constraints for rules used in attachment site selection are primarily lexical constraints, although other information is also relevant.</Paragraph> <Paragraph position="1"> All types of evidence play a role in choosing the attachment relation. A rule is a pair: Rule <C, O> where C is the set of constraints that enable the rule and O is the associated operation. The operation associated with a rule can therefore be either the markup of a DCU as a possible attachment site, or an actual discourse relation, such as Subordination, Coordination or N-ary. A rule is enabled when all sub-conditions in C are satisfied and no other rules having priority are enabled. Rules may combine different sources of evidential information (semantic, syntactic, structural and lexical). If more than one rule is enabled at the same time, ambiguous parses are produced . Some rules are listed in Table 3.</Paragraph> <Paragraph position="2"> We assign weights heuristically at this point. At this stage in our research, we rely only on a partial order among the rules. In future work, we will investigate (1) how evidence is weighed and combined in order to make better attachment deci-The parsing process at the Discourse Parse Tree (DPT) level works as follows. When a BDU-Tree has been constructed and is ready to be attached to the right edge of the DPT, each DCU along the right edge is examined and the lexical information in the right-edge DPT nodes are compared with the lexical evidence retrieved sions and (2) the extent to which discourse ambiguity generated in this fashion is legitimate and how to reduce grammar overgeneration by more efficient handling of interactions among rules and the weighing of the linguistic evidence.</Paragraph> <Paragraph position="3"> from the incoming BDU-Tree. This process, guided by the set of discourse rules, produces an ordered set of active DCUs, representing the possible attachment points in order of likelihood. The set can then be pruned of its n lowest scoring constituents, according to an appropriate policy such as a threshold.</Paragraph> <Paragraph position="4"> In a second stage, each attachment rule is checked against possible attachment sites. Rules that fire successfully attach the BDU-Tree to the DPT at the chosen site with the relationship specified by the rule. Local semantic, lexical and syntactic information is then percolated up to the DCU consisting of the parent of both attachment point and incoming MBDU according to constraints of the discourse relation selected. If multiple attachments at different sites are possible, ambiguous parses are generated; less preferred attachments are discarded and the remaining attachment choices generate valid parse trees.</Paragraph> </Section> </Section> <Section position="7" start_page="11" end_page="13" type="metho"> <SectionTitle> 5 PALSUMM Text Summarization </SectionTitle> <Paragraph position="0"> So far, we have described the U-LDM only as a theoretical approach to discourse parsing. We now turn briefly to describe a computational implementation of these methods. The PALSUMM Text Summarization System is a domain independent symbolic sentence extraction system that produces high level readable summaries that preserve the language and style of the original text and eliminate problems with unresolved or incorrect reference. Our system is currently used to summarize a corpus of 300 technical reports produced by our laboratory.</Paragraph> <Paragraph position="1"> The PALSUMM System relies on the Xerox Linguistic Environment (XLE) to parse the sentences of our source texts. The f-structure output of the XLE parser is segmented into units according to the criteria identified above. The segments are then combined into a BDU-tree. Using syntactic information about syntactic coordination and subordination relations, lexical ontological information taken from WordNet and a customized lexical domain ontology as well as discourse rules, the M-BDU of the sentence along with any other BDUs that must be accessible along the right edge of the discourse tree to accommodate possible continuations are identified, Both the site of attachment and the attachment relation are then computed using discourse attachment rules of the type presented above.</Paragraph> <Paragraph position="2"> Text summarization algorithms are then applied to the resulting tree.</Paragraph> <Paragraph position="3"> Running in purely symbolic mode, the tree is pruned at a given level of embeddedness to produce a summary of a desired length or degree of summarization.</Paragraph> <Paragraph position="4"> Because the resulting summa- null For illustration purposes, we present in Appendix A a summary of a document that was hand coded using the rules given and then summarized automatically using the PALSUMM tree pruning algorithm. The PALSUMM Summaries were judged to be significantly more readable than summaries produced by MEAD in a small comparative study.</Paragraph> <Paragraph position="5"> In Appendix B, we present a diagram of the PALSUMM system.</Paragraph> <Paragraph position="6"> Although closely related to methods reported by Marcu (1999, 2000) for summarization using ries may be longer than desired, alternatively we also use statistical methods to identify salient information (see discussion and references in Marcu 2003) and then construct a partial discourse tree that includes only information identified as most salient and the text at all nodes dominating that salient information.</Paragraph> </Section> class="xml-element"></Paper>