Effective Selectional Restrictions for Unsupervised Relation Extraction

International Joint Conference on Natural Language Processing, pages 1312–1320, Nagoya, Japan, 14-18 October 2013. Effective Selectional Restrictions for Unsupervised Relation Extraction Alan Akbik Larysa Visengeriyeva Johannes Kirschnick Alexander L¨oser Technische Univerist¨at Berlin Databases and Information Systems Group Einsteinufer 17, 10587 Berlin, Germany firstname.lastname@tu-berlin.de Abstract Unsupervised Relation Extraction (URE) methods automatically discover semantic relations in text corpora of unknown con- tent and extract for each discovered rela- tion a set of relation instances. Due to the sparsity of the feature space, URE is vulnerable to ambiguities and underspeci- fication in patterns. In this paper, we pro- pose to increase the discriminative power of patterns in URE using selectional re- strictions (SR). We propose a method that utilizes a Web-derived soft clustering of n-grams to model selectional restrictions in the open domain. We comparatively evaluate our method against a baseline without SR, a setup in which standard 7- class Named Entity types are used as SR and a setup that models SR using a fine- grained entity type system. Our results in- dicate that modeling SR into patterns sig- nificantly improves the ability of URE to discover relations and enables the discov- ery of more fine-granular relations. 1 Introduction In traditional approaches for Relation Extrac- tion (RE), all target relations (such as BORNIN or HASWONPRIZE) need to be specified in ad- vance. For each relation, an extractor is trained or manually created that finds relation instances in text (Jiang and Zhai, 2007). This process is expensive and usually involves manually labeling large amounts of training data, making it difficult to scale RE to large sets of relations. Worse, be- cause target relations must be manually defined in advance, the usefulness of RE in corpora of un- known content is limited (Akbik et al., 2012). This limits their applicability to the open domain where a potentially unbounded number of relations may be expressed in text. In contrast, Unsupervised Relation Extraction (URE) approaches do not require target relations to be pre-specified, and require no labeled train- ing data (Rosenfeld and Feldman, 2007). Instead, they automatically discover prominent relations in a given text corpus and extract for each identi- fied relation a list of relation instances. Current methods (Akbik et al., 2012; Yao et al., 2012) uti- lize a vector space model of semantics in which they group co-occurring pairs of entities (referred to as entity pairs) into clusters based on distribu- tional evidence over observed patterns. Each clus- ter is interpreted as one discovered semantic re- lation and the entity pairs in each cluster as the instances of this relation. Pattern ambiguities. However, a problem for such approaches is that patterns may be ambigu- ous in the sense that they point to more than one relation. The pattern “[X] GET [Y]”1 for exam- ple may be observed for a person and a product (“Jim got a new VW Beetle.”), a person and a dis- ease (“Jim got H1N1.”) or (colloquially) between a person and a difficult-to-understand topic (“Jim finally got Game Theory.”). Such ambiguous …

Content truncated to 500 words