TUD: Semantic Relatedness for Relation Classification

Proceedings of the 5th International Workshop on Semantic Evaluation, ACL 2010, pages 210–213, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics TUD: semantic relatedness for relation classification Gy¨orgy Szarvas∗and Iryna Gurevych Ubiquitous Knowledge Processing (UKP) Lab Computer Science Department Technische Universit¨at Darmstadt Hochschulstraße 10., D-64289 Darmstadt, Germany http://www.ukp.tu-darmstadt.de/ Abstract In this paper, we describe the system sub- mitted by the team TUD to Task 8 at SemEval 2010. The challenge focused on the identification of semantic relations be- tween pairs of nominals in sentences col- lected from the web. We applied max- imum entropy classification using both lexical and syntactic features to describe the nominals and their context. In addi- tion, we experimented with features de- scribing the semantic relatedness (SR) be- tween the target nominals and a set of clue words characteristic to the relations. Our best submission with SR features achieved 69.23% macro-averaged F-measure, pro- viding 8.73% improvement over our base- line system. Thus, we think SR can serve as a natural way to incorporate external knowledge to relation classification. 1 Introduction Automatic extraction of typed semantic relations between sentence constituents is an important step towards deep semantic analysis and understand- ing the semantic content of natural language texts. Identification of relations between a nominal and the main verb, and between pairs of nominals are important steps for the extraction of structured se- mantic information from text, and can benefit vari- ous applications ranging from Information Extrac- tion and Information Retrieval to Machine Trans- lation or Question Answering. The Multi-Way Classification of Semantic Re- lations Between Pairs of Nominals challenge (Hendrickx et al., 2010) focused on the identi- fication of specific relation types between nomi- nals (nouns or base noun phrases) in natural lan- guage sentences collected from the web. The main ∗ On leave from the Research Group on Artificial Intelli- gence of the Hungarian Academy of Sciences task of the challenge was to identify and clas- sify instances of 9 abstract semantic relations be- tween noun phrases, i.e. Cause-Effect, Instrument- Agency, Product-Producer, Content-Container, Entity-Origin, Entity-Destination, Component- Whole, Member-Collection, Message-Topic. That is, given two nominals (e1 and e2) in a sentence, systems had to decide whether relation(e1,e2), re- lation(e2,e1) holds for one of the relation types or the nominals’ relation is other (falls to a category not listed above or they are unrelated). In this sense, the challenge was an important pilot task towards large scale semantic processing of text. In this paper, we describe the system we sub- mitted to Semeval 2010, Task 8. We applied max- imum entropy classification to the problem using both lexical and contextual features to describe the nominals themselves and their context (i.e. the sentence). In addition, we experimented with features exploiting the strength of association be- tween the target nominals and a predefined set of clue words characteristic to the nine relation types. In order to measure the semantic relatedness (SR) of targets and clues, we used the Explicit Seman- tic Analysis (Gabrilovich and Markovitch, 2007) SR measure …

Content truncated to 500 words