File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/02/w02-1704_metho.xml
Size: 19,862 bytes
Last Modified: 2025-10-06 14:08:08
<?xml version="1.0" standalone="yes"?> <Paper uid="W02-1704"> <Title>Web References</Title> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> 3 Discourse markers </SectionTitle> <Paragraph position="0"> Several contemporary discourse theories posit that important aspects of a text's coherence can be formally described (and represented) by means of discourse relations holding between adjacent spans of text (e.g. Asher, 1993; Mann, Thompson, 1988). We use the term discourse marker for those lexical items that (in addition to non-lexical means such as punctuation, aspectual and focus shifts, etc.) can signal the presence of such a relation at the linguistic surface. Typically, a discourse relation is associated with a wide range of such markers; consider, for instance, the following variety of Concessions, which all express the same underlying propositional content. The words that we treat as discourse markers are underlined.</Paragraph> <Paragraph position="1"> We were in SoHo;fneverthelessjnonetheless j however j still j yetg, we found a cheap bar.</Paragraph> <Paragraph position="2"> We were in SoHo, but we found a cheap bar anyway.</Paragraph> <Paragraph position="3"> Despite the fact that we were in SoHo, we found a cheap bar.</Paragraph> <Paragraph position="4"> Notwithstanding the fact that we were in SoHo, we found a cheap bar.</Paragraph> <Paragraph position="5"> Although we were in SoHo, we found a cheap bar.</Paragraph> <Paragraph position="6"> If one accepts these sentences as paraphrases, then the various discourse markers all need to be associated with the information that they signal a concessive relationship between the two propositions involved. Notice that the markers belong to di erent syntactic categories and thus impose quite di erent syntactic constraints on their environment in the sentence. Discourse markers do not form a homogeneous class from the syntactican's viewpoint, but from a functional perspective they should nonetheless be treated as alternatives in a paradigmatic choice. A detailled characterization of discourse markers, together with a test procedure for identifying them in text, has been provided for English by (Knott, 1996). Recently, (Grote, to appear) adapted Knott's procedure for the German language. Very brie y, to identify a discourse marker (e.g., because) in a text, isolate the clause containing a candidate from the text, resolve any anaphors and make elided items explicit; if the resulting text is incomplete (e.g., because the woman bought a Macintosh), then the candidate is indeed a 'relational phrase', or for our purposes, a two-place discourse marker.</Paragraph> <Paragraph position="7"> In addition to the syntactic features, the differences in meaning and style between similar markers need to be discerned; one such di erence is the degree of speci city: for example, but can mark a general Contrast or a more speci c Concession. Another one is the notable di erence in formality between, say but ... anyway and notwithstanding.</Paragraph> <Paragraph position="8"> From the perspective of text generation, not all paraphrases listed above are equally felicitous in speci c contexts. In order to choose the most appropriate variant, a generator needs knowledge about the ne-grained di erences between similar markers for the same relation.</Paragraph> <Paragraph position="9"> Furthermore, it needs to account for the interactions between marker choice and other generation decisions and hence needs knowledge about the syntagmatic constraints associated with different markers. We will discuss this perspective in Section 4.1 From the perspective of text understanding, discourse markers can be used as one source of information for guessing the rhetorical structure of a text, or automatic rhetorical parsing. We will characterize this application in Section 4.2.</Paragraph> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> 4 Requirements on a discourse </SectionTitle> <Paragraph position="0"> marker lexicon As the following two subsections will show, text generation and understanding have quite different preferences on the information coded in a discourse marker lexicon, or \DiMLex&quot; for short. In addition, di erent systems employ different programming languages, and the format of the lexicon has to be adapted accordingly.</Paragraph> <Paragraph position="1"> Yet we want to avoid coding di erent lexicons manually and thus seek a common \core representation&quot; for DiMLex from which the various application-speci c instantiations can be derived. Before proposing such a representation, though, we have to examine in more detail the di erent requirements.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.1 The text generation perspective </SectionTitle> <Paragraph position="0"> Present text generation systems are typically not very good at choosing discourse markers. Even though a few systems have incorporated some more sophisticated mappings for speci c relations (e.g., in DRAFTER (Paris et al., 1995)), there is still a general tendency to treat discourse marker selection as a task to be performed as a \side e ect&quot; by the grammar, much like for other function words such as prepositions.</Paragraph> <Paragraph position="1"> To improve this situation, we propose to view discourse marker selection as one subtask of the general lexical choice process, so that |to continue the example given above |one or another form of Concession can be produced in the light of the speci c utterance parameters and the context. Obviously, marker selection also includes the decision whether to use any marker at all or leave the relation implicit. When these decisions can be systematically controlled, the text can be tailored much better to the speci c goals of the generation process.</Paragraph> <Paragraph position="2"> The generation task imposes a particular view of the information coded in DiMLex: the entry point to the lexicon is the discourse relation to be realized, and the lookup yields the range of alternatives. But many markers have more semantic and pragmatic constraints associated with them, which have to be veri ed in the generator's input representation for the marker to be a candidate. Then, discourse markers place (predominantly syntactic) constraints on their immediate context, which a ects the interactions between marker choice and other realization decisions. And nally, markers that are still equivalent after evaluating these constraints are subject to a choice process that can utilize preferential (e.g. stylistic or length-based) criteria. Therefore, under the generation view, the information in DiMLex is grouped into the following three classes: |Applicability conditions: The necessary conditions for using a discourse marker, i.e., the features or structural con gurations that need to be present in the input speci cation.</Paragraph> <Paragraph position="3"> |Syntagmatic constraints: The constraints regarding the combination of a marker and the neighbouring constituents; most of them are syntactic and appear at the beginning of the list given above (part of speech, linear order, etc.).</Paragraph> <Paragraph position="4"> |Paradigmatic features: Features that label the di erences between similar markers sharing the same applicability conditions, such as stylistic features and degrees of emphasis.</Paragraph> <Paragraph position="5"> Very brie y, we see discourse marker choice as one aspect of the sentence planning task (e.g. (Wanner, Hovy, 1996)). In order to account for the intricate interactions between marker choice and other generation decisions, the idea is to employ DiMLex as a declarative resource supporting the sentence planning process, which comprises determining sentence boundaries and sentence structure, linear ordering of constituents (e.g. thematizations), and lexical choice. All these decisions are heavily interdependent, and in order to produce truly adequate text, the various realization options need to be weighted against each other (in contrast to a simple, xed sequence of making the types of decisions), which presupposes a exible computational mechanism based on resources as declarative as possible. This generation approach is described in more detail in (Grote, Stede, 1998).</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 4.2 The text understanding perspective </SectionTitle> <Paragraph position="0"> In text understanding, discourse markers serve as cues for inferring the rhetorical or semantic structure of the text. In the approach proposed in (Marcu, 1997), for example, the presence of discourse markers is used to hypothesize individual textual units and relations holding between them. Then, the overall discourse structure tree is built using constraint satisfaction techniques. Our analysis method uses the lexicon for an initial identi cation and disambiguation of discourse markers. They serve as one of several other shallow features that determine through a statistical, learned language model the optimal rhetorical analysis.</Paragraph> <Paragraph position="1"> In contrast to the use of markers in generation, the list of cues is signi cantly longer and includes phrasal items like aus diesem Grund (for this reason) or genauer genommen (more precisely).</Paragraph> </Section> </Section> <Section position="6" start_page="0" end_page="0" type="metho"> <SectionTitle> 5 Our XML/XSL solution </SectionTitle> <Paragraph position="0"> In the following we show some sample representations and style sheets that have been abridged for presentation purposes.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 5.1 Source representation </SectionTitle> <Paragraph position="0"> In our hierarchical XML structure, the <dictionary> root tag encloses the entire le, and every single entry rests in an <entry> tag, which unambigously identi es every entry with its id attribute. Within the <entry> tag there are four subordinate tags: <form>, <syn>, <sem>, and <examples>.</Paragraph> <Paragraph position="1"> The <form> tag contains the orthographic form of the headword; at present this amounts to two slots for alternative orthographies. The <syn> area contains the syntactic information about the headword. In this shortened example, there is only the <init field> tag present, in the initial eld of a sentence. Correspondingly, <sem> contains semantic features such as the <function> tag, which contains the semantic/discourse relation expressed by the headword. Finally, <examples>, contains one or more <example> tags that may each give an example sentence.</Paragraph> <Paragraph position="2"> We have shortened this presentation considerably; the full lexicon contains more ne-grained features for all three areas: within <form>, information on pronounciation, syllable structure, and hyphenation; within <syn>, information about syntactic subcategorization and possible positions in the clause; within <sem>, for example the feature whether the information subordinated by the marker is presupposed or not.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 5.2 HTML views </SectionTitle> <Paragraph position="0"> The listing in Figure 4 shows a style sheet that provides an HTML by listing the XML data in a format that roughly resembles a print dictionary. Figure 2 shows the output that results from applying this XSL le to the XML source in gure 1.</Paragraph> <Paragraph position="1"> 05: denn Occurrences: middle eld / Nullstelle Semantics: kausal Related markers: weil da Examples: Das Konzert muss ausfallen, *denn* die S angerin ist erkrankt.</Paragraph> <Paragraph position="2"> Die Blumen auf dem Balkon sind erfroren, *denn* es hat We assume that the general structure of the formatting part of XSL is familiar to the reader. We would like to highlight some details. XLINK is used to ensure that the entry contains an HTML-anchor named after the head-word (ll. 14-20). This way it is possible to link to a certain entry from the <rel> tag of a different entry (39-45).</Paragraph> <Paragraph position="3"> We also employ the XSL equivalent to an if/then/else construct (24-31). The <xsl:choose> tag encloses the choices to be made. The <xsl:when> tag contains the condition match=&quot;.[alt orth='none']&quot; that does nothing if the <alt orth> tag contains the data none. Every other case is covered by the <xsl:otherwise> tag that prints out the <alt orth> information if it is not no entry. Entry alt orth init field mid field . . .</Paragraph> <Paragraph position="4"> Figure 3 shows another possible view for the data. In this case the data is presented in table form, ordered by the value of the mid field tag. It would be easy to show that it is possible to use a <xsl:choose> construct as shown in the example before to print out only those entries that satisfy a certain condition.</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 5.3 The text generation view </SectionTitle> <Paragraph position="0"> For the lexicon to be applied in our text generation system 'Polibox' (Stede, 2002), we need a Lisp-based version of DiM-Lex. Using the (defstruct <name> <slot1> shown in Figure 2 .. <slotn>) construct, we de ne a class of objects for discourse markers, where the features needed for generation are stored in the slots. Again, we abbreviate slightly: Now, a Lisp-object for each individual discourse marker entry is created with the function make-Discmarker, which provides the values for the slots. Figure 5 shows the shape of the entry for denn, whose XML-source was given in gure 1.</Paragraph> <Paragraph position="1"> Again, we aim at deriving these entries automatically via an XSL sheet (which we do not show here). Notice that the mapping task is now somewhat di erent from the HTML cases, since the transformation part of XSL (XSLT) comes into play here. Instead of merely displaying the data in a web browser as in the examples before, an XSLT processor may transform data for use in some XML based client application. As explained in Section 4.1, in the generation scenario we are given a tree fragment consisting of a discourse relation node and two daughters representing the related material, the nucleus and the satellite of the relation. In order to decide whether a particular marker can be used, one important constraint is the \size&quot; of the daughter material, which can be a single proposition or an entire sub-tree. The generator needs to estimate whether it will t into a single phrase, clause, sentence, or into a sequence of sentences; a subordinating conjunction, for instance, can only be used if the material can be expressed within a clause. Thus, the Lisp-entry contains slots N-Complexity and S-Complexity, which are highly application-speci c and thus do not have a simple corresponding feature in the XML source representation of the dictionary. The XSL sheet thus inspects certain combinations of daughter attributes of <syn> and maps them to new names for the llers of the two Complexity slots in the Lisp structure. (Similar mappings occur in other places as well, which we do not show here.)</Paragraph> </Section> <Section position="4" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 5.4 The text understanding view </SectionTitle> <Paragraph position="0"> Our analysis method recasts rhetorical parsing as a set of classi cation decisions, where a pars- null dictionary entry for denn (abridged) ing framework builds a tree structured analysis. Each of the decisions is based on a set of features. Feature types range from syntactical con guration to the presence of a certain discourse marker. The mapping from a pattern of observed features to a rhetorical relation may be learned automatically by a classi cation learning algorithm.</Paragraph> <Paragraph position="1"> Learning and analysis applications use a parsing framework that gives us a set of text span pairs. Every two text spans are subject to a classi cation learning algorithm (during training) or the actual classi er. So, a rhetorical relation is assigned to these two spans of text along with a score so that the parsing framework may decide which of several competing classi cations to accept.</Paragraph> <Paragraph position="2"> Learning and actual rhetorical analysis are accomplished by a set of distinct tools that add speci c annotations to a given input text, before resulting relations are learned or guessed. These tools include a data mining component, a part-of-speech tagger and a segmenter. They all access data organized in an XML syntax. The central learning and parsing application makes use of a Document Object Model (DOM) representation of the corpus. This data structure is e ectively used for information interchange between several components, because it allows us to easily visualize and modify the current data at each processing step during development.</Paragraph> <Paragraph position="3"> With the present corpus data, the learning algorithm is theoretically able to identify rhetorical markers automatically and could thus compile a marker lexicon. However, markers are highly ambiguous. Even though many of them can be tagged as adverbials or conjunctions, markers often lack distinctive syntactic and/or positional properties; some of them are phrasal, some are discontinuous. To identify signi cant cue - relation correlations, a lot of annotated data is necessary: more than is usually available. In a sparse data situation, we want to easen the learning task for the rhetorical language model: It makes sense to use a discourse marker lexicon.</Paragraph> <Paragraph position="4"> On the other hand, we do not expect a hand-crafted lexicon to contain all contextual constraints that would enable us to assign a single rhetorical relation. These constraints can be very subtle; some of them should be represented as probabilistic scalar information.</Paragraph> <Paragraph position="5"> Thus, DiMLex contributes to initial discourse marker disambiguation. From each entry, we interpret syntactic positioning information, morphosyntactic contextual information and a scope class (sentential, phrasal, discourselevel) as a conjunction of constraints. The presence of a certain discourse marker in a speci ed con guration is one of the features to be observed in the text.</Paragraph> <Paragraph position="6"> Depending on the depth of the syntactic and semantic analysis carried out by the text understanding system, di erent features provided by DiMLex can be taken into account. Certain structural con gurations can be tested without any deep understanding; for instance, the German marker w ahrend is generally ambiguous between a Contrast and a Temporal-Cooccurrence reading, but when followed by a noun phrase, only the latter reading is available (w ahrend corresponds not only to the English while but also to during).</Paragraph> <Paragraph position="7"> In the parsing client application, DiMLex serves as resource for the identi cation of cue phrases in speci c structural con gurations.</Paragraph> <Paragraph position="8"> Rhetorical information from the DiMLex entries may serve as one of several cues for the classication engine. The nal linking from cue patterns to rhetorical relations is learned from a corpus annotated with rhetorical structures.</Paragraph> </Section> </Section> class="xml-element"></Paper>