File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/05/p05-1021_metho.xml
Size: 13,685 bytes
Last Modified: 2025-10-06 14:09:42
<?xml version="1.0" standalone="yes"?> <Paper uid="P05-1021"> <Title>Improving Pronoun Resolution Using Statistics-Based Semantic Compatibility Information</Title> <Section position="4" start_page="165" end_page="166" type="metho"> <SectionTitle> 2 Computing the Statistics-based Semantic Compatibility </SectionTitle> <Paragraph position="0"> In this section, we introduce in detail how to compute the semantic compatibility, using the predicate-argument statistics obtained from the corpus or the web.</Paragraph> <Section position="1" start_page="165" end_page="166" type="sub_section"> <SectionTitle> 2.1 Corpus-Based Semantic Compatibility </SectionTitle> <Paragraph position="0"> Three relationships, possessive-noun, subject-verb and verb-object, are considered in our work. Before resolution a large corpus is prepared. Documents in the corpus are processed by a shallow parser that could generate predicate-argument tuples of the above three relationships1.</Paragraph> <Paragraph position="1"> To reduce data sparseness, the following steps are applied in each resulting tuple, automatically: * Only the nominal or verbal heads are retained.</Paragraph> <Paragraph position="2"> * Each Named-Entity (NE) is replaced by a common noun which corresponds to the semantic category of the NE (e.g. &quot;IBM&quot; - &quot;company&quot;) 2.</Paragraph> <Paragraph position="3"> * All words are changed to their base morphologic forms (e.g. &quot;companies - company&quot;). During resolution, for an encountered anaphor, each of its antecedent candidates is substituted with the anaphor . According to the role and type of the anaphor in its context, a predicate-argument tuple is extracted and the above three steps for data-sparse reduction are applied. Consider the sentence (1), for example. The anaphors &quot;it1&quot; and &quot;it2&quot; indicate a subject verb and verb object relationship, respectively. Thus, the predicate-argument tuples for the two candidates &quot;government&quot; and &quot;money&quot; would be (collect (subject government)) and (collect (subject money)) for &quot;it1&quot;, and (collect (object government)) and (collect (object money)) for &quot;it2&quot;. Each extracted tuple is searched in the prepared tuples set of the corpus, and the times the tuple occurs are calculated. For each candidate, its semantic compatibility with the anaphor could be represented simply in terms of frequency</Paragraph> <Paragraph position="5"> where count(candi,ana) is the count of the tuple formed by candi and ana, or alternatively, in terms of conditional probability (P(candi,ana|candi)), where the count of the tuple is divided by the count of the single candidate in the corpus. That is</Paragraph> <Paragraph position="7"> In this way, the statistics would not bias candidates that occur frequently in isolation.</Paragraph> </Section> <Section position="2" start_page="166" end_page="166" type="sub_section"> <SectionTitle> 2.2 Web-Based Semantic Compatibility </SectionTitle> <Paragraph position="0"> Unlike documents in normal corpora, web pages could not be preprocessed to generate the predicate-argument reserve. Instead, the predicate-argument statistics has to be obtained via a web search engine like Google and Altavista. For the three types of predicate-argument relationships, queries are constructed in the forms of &quot;NPcandi VP&quot; (for subjectverb), &quot;VP NPcandi&quot; (for verb-object), and &quot;NPcandi 's NP&quot; or &quot;NP of NPcandi&quot; (for possessive-noun). Consider the following sentence: (2) Several experts suggested that IBM's accounting grew much more liberal since the mid 1980s as its business turned sour.</Paragraph> <Paragraph position="1"> For the pronoun &quot;its&quot; and the candidate &quot;IBM&quot;, the two generated queries are &quot;business of IBM&quot; and &quot;IBM's business&quot;.</Paragraph> <Paragraph position="2"> To reduce data sparseness, in an initial query only the nominal or verbal heads are retained. Also, each NE is replaced by the corresponding common noun.</Paragraph> <Paragraph position="3"> (e.g, &quot;IBM's business&quot; - &quot;company's business&quot; and &quot;business of IBM&quot; - &quot;business of company&quot;). A set of inflected queries is generated by expanding a term into all its possible morphological forms. For example, in Sentence (1), &quot;collect money&quot; becomes &quot;collected|collecting|... money&quot;, and in (2) &quot;business of company&quot; becomes &quot;business of company|companies&quot;). Besides, determiners are inserted for every noun. If the noun is the candidate under consideration, only the definite article the is inserted. For other nouns, instead, a/an, the and the empty determiners (for bare plurals) would be added (e.g., &quot;the|a business of the company|companies&quot;). Queries are submitted to a particular web search engine (Google in our study). All queries are performed as exact matching. Similar to the corpus-based statistics, the compatibility for each candidate and anaphor pair could be represented using either frequency (Eq. 1) or probability (Eq. 2) metric. In such a situation, count(candi,ana) is the hit number of the inflected queries returned by the search engine, while count(candi) is the hit number of the query formed with only the head of the candidate (i.e.,&quot;the + candi&quot;).</Paragraph> </Section> </Section> <Section position="5" start_page="166" end_page="168" type="metho"> <SectionTitle> 3 Applying the Semantic Compatibility </SectionTitle> <Paragraph position="0"> In this section, we discuss how to incorporate the statistics-based semantic compatibility for pronoun resolution, in a machine learning framework.</Paragraph> <Section position="1" start_page="166" end_page="166" type="sub_section"> <SectionTitle> 3.1 The Single-Candidate Model </SectionTitle> <Paragraph position="0"> One way to utilize the semantic compatibility is to take it as a feature under the single-candidate learning model as employed by Ng and Cardie (2002).</Paragraph> <Paragraph position="1"> In such a learning model, each training or testing instance takes the form of i{C, ana}, where ana is the possible anaphor and C is its antecedent candidate. An instance is associated with a feature vector to describe their relationships.</Paragraph> <Paragraph position="2"> During training, for each anaphor in a given text, a positive instance is created by pairing the anaphor and its closest antecedent. Also a set of negative instances is formed by pairing the anaphor and each of the intervening candidates. Based on the training instances, a binary classifier is generated using a certain learning algorithm, like C5 (Quinlan, 1993) in our work.</Paragraph> <Paragraph position="3"> During resolution, given a new anaphor, a test instance is created for each candidate. This instance is presented to the classifier, which then returns a positive or negative result with a confidence value indicating the likelihood that they are co-referent. The candidate with the highest confidence value would be selected as the antecedent.</Paragraph> </Section> <Section position="2" start_page="166" end_page="167" type="sub_section"> <SectionTitle> 3.2 Features </SectionTitle> <Paragraph position="0"> In our study we only consider those domain-independent features that could be obtained with low</Paragraph> </Section> <Section position="3" start_page="167" end_page="167" type="sub_section"> <SectionTitle> Feature Description </SectionTitle> <Paragraph position="0"> DefNp 1 if the candidate is a definite NP; else 0 Pron 1 if the candidate is a pronoun; else 0 NE 1 if the candidate is a named entity; else 0 SameSent 1 if the candidate and the anaphor is in the same sentence; else 0 NearestNP 1 if the candidate is nearest to the anaphor; else 0 ParalStuct 1 if the candidate has an parallel structure with ana; else 0 FirstNP 1 if the candidate is the first NP in a sentence; else 0 Reflexive 1 if the anaphor is a reflexive pronoun; else 0 Type Type of the anaphor (0: Single neuter pronoun; 1: Plural neuter pronoun; 2: Male personal pronoun; 3: Female personal pronoun) StatSem[?] the statistics-base semantic compatibility of the candidate SemMag[?][?] the semantic compatibility difference between two competing candidates while **ed feature is only for the twin-candidate mode) computational cost but with high reliability. Table 1 summarizes the features with their respective possible values. The first three features represent the lexical properties of a candidate. The POS properties could indicate whether a candidate refers to a hearer-old entity that would have a higher preference to be selected as the antecedent (Strube, 1998). SameSent and NearestNP mark the distance relationships between an anaphor and the candidate, which would significantly affect the candidate selection (Hobbs, 1978). FirstNP aims to capture the salience of the candidate in the local discourse segment. ParalStuct marks whether a candidate and an anaphor have similar surrounding words, which is also a salience factor for the candidate evaluation (Mitkov, 1998).</Paragraph> <Paragraph position="1"> Feature StatSem records the statistics-based semantic compatibility computed, from the corpus or the web, by either frequency or probability metric, as described in the previous section. If a candidate is a pronoun, this feature value would be set to that of its closest nominal antecedent.</Paragraph> <Paragraph position="2"> As described, the semantic compatibility of a candidate is computed under the context of the current anaphor. Consider two occurrences of anaphors &quot;. . .it1 collected . . . &quot; and &quot;. . .it2 said . . . &quot;. As &quot;NP collected&quot; should occur less frequently than &quot;NP said&quot;, the candidates of it1 would generally have predicate-argument statistics lower than those of it2.</Paragraph> <Paragraph position="3"> That is, a positive instance for it1 might bear a lower semantic feature value than a negative instance for it2. The consequence is that the learning algorithm would think such a feature is not that &quot;indicative&quot; and reduce its salience in the resulting classifier.</Paragraph> <Paragraph position="4"> One way to tackle this problem is to normalize the feature by the frequencies of the anaphor's context, e.g., &quot;count(collected)&quot; and &quot;count(said)&quot;. This, however, would require extra calculation. In fact, as candidates of a specific anaphor share the same anaphor context, we can just normalize the semantic feature of a candidate by that of its competitor:</Paragraph> <Paragraph position="6"> The value (0 [?] 1) represents the rank of the semantic compatibility of the candidate C among candi set(ana), the current candidates of ana.</Paragraph> </Section> <Section position="4" start_page="167" end_page="168" type="sub_section"> <SectionTitle> 3.3 The Twin-Candidate Model </SectionTitle> <Paragraph position="0"> Yang et al. (2003) proposed an alternative twin-candidate model for anaphora resolution task. The strength of such a model is that unlike the single-candidate model, it could capture the preference relationships between competing candidates. In the model, candidates for an anaphor are paired and features from two competing candidates are put together for consideration. This property could nicely deal with the above mentioned training problem of different anaphor contexts, because the semantic feature would be considered under the current candidate set only. In fact, as semantic compatibility is a preference-based factor for anaphor resolution, it would be incorporated in the twin-candidate model more naturally.</Paragraph> <Paragraph position="1"> In the twin-candidate model, an instance takes a form like i{C1, C2, ana}, where C1 and C2 are two candidates. We stipulate that C2 should be closer to ana than C1 in distance. The instance is labelled as &quot;10&quot; if C1 the antecedent, or &quot;01&quot; if C2 is. During training, for each anaphor, we find its closest antecedent, Cante. A set of &quot;10&quot; instances, i{Cante, C, ana}, is generated by pairing Cante and each of the interning candidates C. Also a set of &quot;01&quot; instances, i{C, Cante, ana}, is created by pairing Cante with each candidate before Cante until another antecedent, if any, is reached.</Paragraph> <Paragraph position="2"> The resulting pairwise classifier would return &quot;10&quot; or &quot;01&quot; indicating which candidate is preferred to the other. During resolution, candidates are paired one by one. The score of a candidate is the total number of the competitors that the candidate wins over. The candidate with the highest score would be selected as the antecedent.</Paragraph> <Paragraph position="3"> Features The features for the twin-candidate model are similar to those for the single-candidate model except that a duplicate set of features has to be prepared for the additional candidate. Besides, a new feature, SemMag, is used in place of StatSem to represent the difference magnitude between the semantic compatibility of two candidates. Let</Paragraph> <Paragraph position="5"> The positive or negative value marks the times that the statistics of C1 is larger or smaller than C2.</Paragraph> </Section> </Section> class="xml-element"></Paper>