An Effective Method of Using Web Based Information for Relation Extraction

An effective method of using Web based information for Relation Extraction Yong Wai Keong, Stanley Institute for Inforcomm Research 21 Heng Mui Keng Terrace, Singapore 119613 wkyong@i2r.a-star.edu.sg Su Jian Institute for Inforcomm Research 21 Heng Mui Keng Terrace, Singapore 119613 sujian@i2r.a-star.edu.sg Abstract We propose a method that incorporates paraphrase information from the Web to boost the performance of a supervised re- lation extraction system. Contextual infor- mation is extracted from the Web using a semi-supervised process, and summarized by skip-bigram overlap measures over the entire extract. This allows the capture of lo- cal contextual information as well as more distant associations. We observe a statisti- cally significant boost in relation extraction performance. We investigate two extensions, thematic clustering and hypernym expansion. In tan- dem with thematic clustering to reduce noise in our paraphrase extraction, we attempt to increase the coverage of our search for para- phrases using hypernym expansion. Evaluation of our method on the ACE 2004 corpus shows that it out-performs the base- line SVM-based supervised learning algo- rithm across almost all major ACE relation types, by a margin of up to 31%. 1 Introduction and motivation In this paper, we shall be primarily dealing with the sort of relations defined in the NIST’s Auto- matic Content Extraction program, specifically for the Relation Detection and Characterization (RDC) task (Doddington et al., 2004). These are links be- tween two entities mentioned in the same sentence, and further restrict our consideration to those rela- tionships clearly supported by evidence in the scope of the same document. The ACE’s annotators mark all mentions of re- lations where there is a direct syntactic connection between the entities, i.e. when one entity mention modifies another one, or when two entity mentions are arguments of the same event. Relations between entities that are implied in the text but which do not satisfy either requirement are considered to be im- plicit, and are marked only once. Our work sits squarely in the realm of work on regular IE done by (Zelenko et al., 2003; Zhou et al., 2005; Chen et al., 2006). Here, the corpus of interest is a well defined set of texts, such as news articles, and we have to detect and classify all ap- pearances of relations from a set of given relation types in the documents. In line with assumptions in the related work, we assert that the differences in the markup for implicit and explicit relations does not significantly affect our performance. Supervised learning methods have proved to be some of the most effective for regular IE. They do however, need large volumes of expensively anno- tated examples to perform robustly. As a result, even the large ACE compilation has deficiencies in the number of instances available for some of the rela- tion types. (Zhang et al., 2005) reports an F-score of 50% for categories of intermediate size and only 30% for the least common relation types. It appears that there are hard limits on the performance of rela- tion extraction …

Content truncated to 500 words