<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1112">
  <Title>Exploring Correlation of Dependency Relation Paths for Answer Extraction</Title>
  <Section position="4" start_page="889" end_page="890" type="relat">
    <SectionTitle>
2 Related Work
</SectionTitle>
    <Paragraph position="0"> In recent years' TREC Evaluation, most top ranked QA systems use syntactic information in answer extraction. Next, we will briefly discuss the main usages.</Paragraph>
    <Paragraph position="1"> (Kaisser and Becker, 2004) match a question into one of predefined patterns, such as &amp;quot;When did Jack Welch retire from GE?&amp;quot; to the pattern &amp;quot;When+did+NP+Verb+NPorPP&amp;quot;. For each question pattern, there is a set of syntactic structures for potential answer. Candidate answers are ranked by matching the syntactic structures. This method worked well on TREC questions. However, it is costing to manually construct question patterns and syntactic structures of the patterns.</Paragraph>
    <Paragraph position="2"> (Shen et al., 2005) classify question words into four classes target word, head word, subject word and verb. For each class, syntactic relation patterns which contain one question word and one proper answer are automatically extracted and scored from training sentences. Then, candidate answers are ranked by partial matching to the syntactic relation patterns using tree kernel. However, the criterion to classify the question words is not clear in their paper. Proper answers may have absolutely different relations with different subject words in sentences. They don't consider the corresponding relations in questions.</Paragraph>
    <Paragraph position="3"> (Tanev et al., 2004; Wu et al., 2005) compare syntactic relations in questions and those in answer sentences. (Tanev et al., 2004) reconstruct a basic syntactic template tree for a question, in which one of the nodes denotes expected answer position. Then, answer candidates for this question are ranked by matching sentence syntactic tree to the question template tree. Furthermore, the matching is weighted by lexical variations. (Wu et al., 2005) combine n-gram proximity search and syntactic relation matching. For syntactic relation matching, question tree and sentence subtree around a candidate answer are matched from node to node.</Paragraph>
    <Paragraph position="4"> Although the above systems apply the different methods to compare relations in question and answer sentences, they follow the same hypothesis that proper answers are more likely to have same relations in question and answer sentences. For example, in question &amp;quot;Who founded the Black Panthers organization?&amp;quot;, where, the question word &amp;quot;who&amp;quot; has the dependency relations &amp;quot;subj&amp;quot; with &amp;quot;found&amp;quot; and &amp;quot;subj obj nn&amp;quot; with &amp;quot;Black Panthers organization&amp;quot;, in sentence &amp;quot;Hilliard introduced Bobby Seale, who co-founded the Black Panther Party here ...&amp;quot;, the proper answer &amp;quot;Bobby Seale&amp;quot; has the same relations with most question phrases.</Paragraph>
    <Paragraph position="5"> These methods achieve high precision, but poor recall due to relation variations. One meaning is often represented as different relation combinations. In the above example, appositive rela- null tion frequently appears in answer sentences, such as &amp;quot;Black Panther Party co-founder Bobby Seale is ordered bound and gagged ...&amp;quot; and indicates proper answer Bobby Seale although it is asked in different way in the question.</Paragraph>
    <Paragraph position="6"> (Cui et al., 2004) propose an approximate dependency relation matching method for both passage retrieval and answer extraction. The similarity between two relations is measured by their co-occurrence rather than exact matching. They state that their method effectively overcomes the limitation of the previous exact matching methods. Lastly, they use the sum of similarities of all path pairs to rank candidate answers, which is based on the assumption that all paths have equal weights. However, it might not be true. For example, in question &amp;quot;What book did Rachel Carson write in 1962?&amp;quot;, the phrase &amp;quot;Rachel Carson&amp;quot; looks like more important than &amp;quot;1962&amp;quot; since the former is question topic and the latter is a constraint for expected answer. In addition, lexical variations are not well considered and a weak relation path alignment algorithm is used in their work.</Paragraph>
    <Paragraph position="7"> Based on the previous works, this paper explores correlation of dependency relation paths between questions and candidate sentences. Dynamic time warping algorithm is adapted to calculate path correlations and approximate phrase mapping is proposed to cope with phrase variations. Finally, maximum entropy-based ranking model is developed to incorporate the correlations and rank candidate answers.</Paragraph>
  </Section>
class="xml-element"></Paper>