ECNU: Effective Semantic Relations Classification without Complicated Features or Multiple External Corpora

Proceedings of the 5th International Workshop on Semantic Evaluation, ACL 2010, pages 226–229, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics ECNU: Effective Semantic Relations Classification without Complicated Features or Multiple External Corpora Yuan Chen†, Man Lan†,§, Jian Su§, Zhi Min Zhou†, Yu Xu† †East China Normal University, Shanghai, PRC. §Institute for Infocomm Research, Singapore. lanman.sg@gmail.com Abstract This paper describes our approach to the automatic identification of semantic rela- tions between nominals in English sen- tences. The basic idea of our strategy is to develop machine-learning classifiers which: (1) make use of class-independent features and classifier; (2) make use of a simple and effective feature set without high computational cost; (3) make no use of external annotated or unannotated cor- pus at all. At SemEval 2010 Task 8 our system achieved an F-measure of 75.43% and a accuracy of 70.22%. 1 Introduction Knowledge extraction of semantic relations be- tween pairs of nominals from English text is one important application both as an end in itself and as an intermediate step in various downstream NLP applications, such as information extraction, summarization, machine translation, QA etc. It is also useful for many auxiliary tasks such as word sense disambiguation, language modeling, para- phrasing and discourse relation processing. In the past decade, semantic relation classifica- tion has attracted a lot of interest from researchers and a wide variety of relation classification schemes exist in the literature. However, most research work is quite different in definition of relations and granularities of various applications. That is, there is little agreement on relation inventories. SemEval 2010 Task 8 (Hendrickx et al., 2008) provides a new standard benchmark for semantic relation classification to a wider community, where it defines 9 relations includ- ing CAUSE-EFFECT, COMPONENT-WHOLE, CONTENT-CONTAINER, ENTITY-DESTINATION, ENTITY-ORIGIN, MEMBER-COLLECTION,INSTRUMENT-AGENCY, MESSAGE-TOPIC, PRODUCT-PRODUCER, and a tenth pseudo- relation OTHER (where relation is not one of the 9 annotated relations). Unlike the previous semantic relation task in SemEval 2007 Task 4, the current evaluation pro- vides neither query pattern for each sentence nor manually annotated word sense (in WordNet se- mantic) for each nominals. Since its initiative is to provide a more realistic real-world application design that is practical, any classification system must be usable without too much effort. It needs to be easily computable. So we need to take into account the following special considerations. 1. The extracted features for relation are ex- pected to be easily computable. That is, the steps in the feature extraction process are to be simple and direct for the purpose of reduc- ing errors possibly introduced by many NLP tools. Furthermore, a unified (global) feature set is set up for all relations rather than for each relation. 2. Most previous work at SemEval 2007 Task 4 leveraged on external theauri or corpora (whether unannotated or annotated) (Davi- dov and Rappoport, 2008), (Costello, 2007), (Beamer et al., 2007) and (Nakov and Hearst, 2008) that make the task adaption to different domains and languages more difficult, since they would not have such manually …

Content truncated to 500 words