The Domain Restriction Hypothesis: Relating Term Similarity and Semantic Consistency

Proceedings of NAACL HLT 2007, pages 131–138, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics The Domain Restriction Hypothesis: Relating Term Similarity and Semantic Consistency Alfio Massimiliano Gliozzo ITC-irst Trento, Italy gliozzo@itc.it Marco Pennacchiotti University of Rome Tor Vergata Rome, Italy pennacchiotti@info.uniroma2.it Patrick Pantel USC, Information Sciences Institute Marina del Rey, CA pantel@isi.edu Abstract In this paper, we empirically demonstrate what we call the domain restriction hy- pothesis, claiming that semantically re- lated terms extracted from a corpus tend to be semantically coherent. We apply this hypothesis to define a post-processing module for the output of Espresso, a state of the art relation extraction system, show- ing that irrelevant and erroneous relations can be filtered out by our module, in- creasing the precision of the final output. Results are confirmed by both quantita- tive and qualitative analyses, showing that very high precision can be reached. 1 Introduction Relation extraction is a fundamental step many natural language processing applications suchin as learning ontologies from texts (Buitelaar et al., 2005) and Question Answering (Pasca and Harabagiu, 2001). The most common approach for acquiring con- cepts, instances and relations is to harvest semantic knowledge from texts. These techniques have been largely explored and today they achieve reasonable accuracy. Harvested lexical resources, such as con- cept lists (Pantel and Lin, 2002), facts (Etzioni et al., 2002) and semantic relations (Pantel and Pen- nacchiotti, 2006) could be then successfully used in different frameworks and applications. The state of the art technology for relation extrac- tion primarily relies on pattern-based approaches (Snow et al., 2006). These techniques are based on the recognition of the typical patterns that express a particular relation in text (e.g. “X such as Y” usually expresses an is-a relation). Yet, text-based algorithms for relation extraction, in particular pattern-based algorithms, still suffer from a number of limitations due to complexities of natural lan- guage, some of which we describe below. Irrelevant relations. These are valid relations that are not of interest in the domain at hand. For example, in a political domain, “Condoleezza Rice is a football fan” is not as relevant as “Condoleezza Rice is the Secretary of State of the United States”. Irrelevant relations are ubiquitous, and affect ontol- ogy reliability, if used to populate it, as the relation drives the wrong type of ontological knowledge. Erroneous or false relations. These are particu- larly harmful, since they directly affect algorithm precision. A pattern-based relation extraction algorithm is particularly likely to extract erroneous relations if it uses generic patterns, which are defined in (Pantel and Pennacchiotti, 2006) as broad coverage, noisy patterns with high recall and low precision (e.g. “X of Y” for part-of relation). Harvesting algorithms either ignore generic patterns (Hearst, 1992) (affecting system recall) or use man- ually supervised filtering approaches (Girju et al., 2006) or use completely unsupervised Web-filtering methods (Pantel and Pennacchiotti, 2006). Yet, these methods still do not sufficiently mitigate the problem of erroneous relations. Background knowledge. Another aspect that makes relation harvesting difficult is related …

Content truncated to 500 words