Entity Set Expansion using Interactive Topic Information

Copyright 2012 by Kugatsu Sadamitsu Sadamitsu, Kuniko Saito, Kenji Imamura, and Yoshihiro Matsuo 26th Pacific Asia Conference on Language,Information and Computation pages 108–116 Entity Set Expansion using Interactive Topic Information Kugatsu Sadamitsu, Kuniko Saito, Kenji Imamura and Yoshihiro Matsuo NTT Media Intelligence Laboratories, NTT Corporation 1-1 Hikarinooka, Yokosuka-shi, Kanagawa, 239-0847, Japan {sadamitsu.kugatsu, saito.kuniko, imamura.kenji, matsuo.yoshihiro} @lab.ntt.co.jp Abstract We propose a new method for entity set expan- sion that achieves highly accurate extraction by suppressing the effect of semantic drift; it requires a small amount of interactive infor- mation. We supplement interactive informa- tion to re-train the topic models (based on in- teractive Unigram Mixtures) not only the con- textual information. Although the topic infor- mation extracted from an unsupervised corpus is effective for reducing the effect of seman- tic drift, the topic models and target entities sometimes suffer grain mismatch. Interactive Unigram Mixtures can, with very few interac- tive words, ease the mismatch between topic and target entities. We incorporate the inter- active topic information into a two-stage dis- criminative system for stable set expansion. Experiments confirm that the proposal raises the accuracy of the set expansion system from the baselines examined. 1 Introduction The task of this paper is entity set expansion in which the lexicons are expanded from just a few seed entities (Pantel et al., 2009). For example, the user inputs the words “Apple”, “Google” and “IBM”, and the system outputs “Microsoft”, “Face- book” and “Intel”. Many set expansion and relation extraction algorithms are based on bootstrapping al- gorithms (Thelen and Riloff, 2002; Pantel and Pen- nacchiotti, 2006), which iteratively acquire new en- tities from corpora. These algorithms suffer from the general problem of “semantic drift”. Semantic drift moves the extraction criteria away from the ini- tial criteria demanded by the user and so reduces the accuracy of extraction. Recently, topic information is being used to alle- viate semantic drift. Topic information means the genre of each document as estimated by statistical topic models. Sadamitsu et al. (2011) proposed a bootstrapping method that uses unsupervised topic information estimated by Latent Dirichlet Alloca- tion (LDA) (Blei et al., 2003) to alleviate semantic drift. They use a discriminative method (Bellare et al., 2006) in order to incorporates topic information. They showed that the use of topic information im- proves the accuracy of the extracted entities. Although unsupervised topic information has been confirmed to be effective, the topic models and target entity sometimes demonstrate grain mis- match. To avoid this mismatch, we refine the topic models to match the target entity grain. Deciding the entity grain from only positive seeds is diffi- cult (Vyas et al., 2009). For example, the positive seed words are “Prius” and “Civic”. In this situa- tion, whether “Cadillac” is positive or negative de- pends on the user’s definition. If the user thinks that “Japanese car” is positive grain, “Cadillac” should be placed into the negative class but if “car” is the positive grain it should be placed into the positive class. Note that we use …

Content truncated to 500 words