Learning to Exploit Structured Resources for Lexical Inference

Proceedings of the 19th Conference on Computational Language Learning, pages 175–184, Beijing, China, July 30-31, 2015. c⃝2015 Association for Computational Linguistics Learning to Exploit Structured Resources for Lexical Inference Vered Shwartz† Omer Levy† Ido Dagan† Jacob Goldberger§ † Computer Science Department, Bar-Ilan University § Faculty of Engineering, Bar-Ilan University vered1986@gmail.com {omerlevy,dagan}@cs.biu.ac.il jacob.goldberger@biu.ac.il Abstract Massive knowledge resources, such as Wikidata, can provide valuable informa- tion for lexical inference, especially for proper-names. Prior resource-based ap- proaches typically select the subset of each resource’s relations which are relevant for a particular given task. The selection process is done manually, limiting these approaches to smaller resources such as WordNet, which lacks coverage of proper- names and recent terminology. This paper presents a supervised framework for auto- matically selecting an optimized subset of resource relations for a given target infer- ence task. Our approach enables the use of large-scale knowledge resources, thus providing a rich source of high-precision inferences over proper-names.1 1 Introduction Recognizing lexical inference is an important component in semantic tasks. Various lexical- semantic relations, such as synonomy, class- membership, part-of, and causality may be used to infer the meaning of one word from another, in order to address lexical variability. For in- stance, a question answering system asked “which artist’s net worth is $450 million?” might re- trieve the candidates Beyonc´e Knowles and Lloyd Blankfein, who are both worth $450 mil- lion. To correctly answer the question, the appli- cation needs to know that Beyonc´e is an artist, and that Lloyd Blankfein is not. 1Our code and data are available at: https://github.com/vered1986/LinKeR Corpus-based methods are often employed to recognize lexical inferences, based on either co- occurrence patterns (Hearst, 1992; Turney, 2006) or distributional representations (Weeds and Weir, 2003; Kotlerman et al., 2010). While earlier meth- ods were mostly unsupervised, recent trends intro- duced supervised methods for the task (Baroni et al., 2012; Turney and Mohammad, 2015; Roller et al., 2014). In these settings, a targeted lexical inference relation is implicitly defined by a train- ing set of term-pairs, which are annotated as posi- tive or negative examples of this relation. Several such datasets have been created, each representing a somewhat different flavor of lexical inference. While corpus-based methods usually enjoy high recall, their precision is often limited, hinder- ing their applicability. An alternative common practice is to mine high-precision lexical in- ferences from structured resources, particularly WordNet (Fellbaum, 1998). Nevertheless, Word- Net is an ontology of the English language, which, by definition, does not cover many proper- names (Beyonc´e → artist) and recent termi- nology (Facebook → social network). A po- tential solution may lie in rich and up-to-date structured knowledge resources such as Wikidata (Vrandeˇci´c, 2012), DBPedia (Auer et al., 2007), and Yago (Suchanek et al., 2007). In this paper, we investigate how these resources can be exploited for lexical inference over proper-names. We begin by examining whether the common usage of WordNet for lexical inference can be ex- tended to larger resources. Typically, a subset of WordNet relations is manually …

Content truncated to 500 words