Precise Information Retrieval Exploiting Predicate-Argument Structures

International Joint Conference on Natural Language Processing, pages 37–45, Nagoya, Japan, 14-18 October 2013. Precise Information Retrieval Exploiting Predicate-Argument Structures Daisuke Kawahara† Keiji Shinzato‡ Tomohide Shibata† Sadao Kurohashi† †Graduate School of Informatics, Kyoto University ‡Rakuten Institute of Technology {dk, shibata, kuro}@i.kyoto-u.ac.jp, keiji.shinzato@mail.rakuten.com Abstract A concept can be linguistically expressed in various syntactic constructions. Such syntactic variations spoil the effectiveness of incorporating dependencies between words into information retrieval systems. This paper presents an information re- trieval method for normalizing syntactic variations via predicate-argument struc- tures. We conduct experiments on stan- dard test collections and show the effec- tiveness of our approach. Our proposed method significantly outperforms a base- line method based on word dependencies. 1 Introduction Most conventional approaches to information re- trieval (IR) deal with words as independent terms. In query sentences1 and documents, however, de- pendencies exist between words.2 To capture these dependencies, some extended IR models have been proposed in the last decade (Jones, 1999; Lee et al., 2006; Song et al., 2008; Shin- zato et al., 2008). These models, however, did not achieve consistent significant improvements over models based on independent words. One of the reasons for this is the linguistic vari- ations of syntax, that is, languages are syntacti- cally expressed in various ways. For instance, the same or similar meaning can be expressed using the passive voice or the active voice in a sentence. Previous approaches based on dependencies can- not identify such variations. This is because they use the output of a dependency parser, which gen- erates syntactic (grammatical) dependencies built 1In this paper, we handle queries written in natural lan- guage. 2While dependencies between words are sometimes con- sidered to be the co-occurrence of words in a sentence, in this paper we consider dependencies to be syntactic or semantic dependencies between words. upon surface word sequences. Consider, for ex- ample, the following sentence in a document: ? ? ? ? (1) YouTube was acquired by Google. Dependency parsers based on the Penn Tree- bank and the head percolation table (Collins, 1999) judge the head of “YouTube” as “was” (“YouTube←was”; hereafter, we denote a depen- dency by “modifier←head”). This dependency, however, cannot be matched with the dependency “YouTube←acquire” in a query like: (2) I want to know the details of the news that Google acquired YouTube. Furthermore, even if a dependency link in a query matches that in a document, a mismatch of dependency type can cause another problem. This is because previous models did not distinguish de- pendency types. For example, the dependency “YouTube←acquire” in query sentence (2) can be found in the following irrelevant document. (3) Google acquired PushLife for $25M ... YouTube acquired Green Parrot Pictures ... While this document does indeed contain the de- pendency “YouTube←acquire,” its type is dif- ferent; specifically, the query dependency is ac- cusative while the document dependency is nom- inative. That is to say, ignoring differences in de- pendency types can lead to inaccurate information retrieval. In this paper, we propose an IR method that …

Content truncated to 500 words