Unsupervised Discovery of Relations and Discriminative Extraction Patterns

Proceedings of COLING 2012: Technical Papers, pages 17–32, COLING 2012, Mumbai, December 2012. Unsupervised Discovery of Relations and Discriminative Extraction Patterns Alan Akbik1 Larysa Visengeriyeva1 Priska Herger2 Holmer Hemsen1 Alexander Löser1 Technische Univeristät Berlin Database Systems and Information Management Group Einsteinufer 17, 10587 Berlin, Germany {firstname.lastname}@tu-berlin.de1, priska@campus.tu-berlin.de2 ABSTRACT Unsupervised Relation Extraction (URE) is the task of extracting relations of a priori unknown semantic types using clustering methods on a vector space model of entity pairs and patterns. In this paper, we show that an informed feature generation technique based on dependency trees significantly improves clustering quality, as measured by the F-score, and therefore the ability of the URE method to discover relations in text. Furthermore, we extend URE to produce a set of weighted patterns for each identified relation that can be used by an information extraction system to find further instances of this relation. Each pattern is assigned to one or multiple relations with different confidence strengths, indicating how reliably a pattern evokes a relation, using the theory of Discriminative Category Matching. We evaluate our findings in two tasks against strong baselines and show significant improvements both in relation discovery and information extraction. KEYWORDS: Unsupervised Relation Extraction, Clustering, Vector Space Models. 17 1 Introduction Recently, there has been great interest in broadening information extraction methods to allow for unsupervised discovery of relational information in large document collections of unknown content. Contrary to classic information extraction in which relationship types (such as BORNIN or MARRIEDTO) are specified in advance, such methods automatically identify a priori unknown relationship types in a given corpus. For these identified semantic relations1, they subsequently or simultaneously perform an information extraction step, thereby transforming the corpus into structured, relational data without any supervision or previous knowledge about its content. One such approach, Unsupervised Relation Extraction (URE), addresses this challenge by building on the latent relation hypothesis which states that pairs of words that co-occur in similar patterns tend to have similar relations (Turney, 2008; Rosenfeld and Feldman, 2007). Current techniques capture this in a vector space model by computing a pair-pattern matrix in which each row represents an entity pair and each column a distinct pattern, with co- occurrence counts as cell values. This representation allows us to compute the similarity of two entity pairs by comparing the distribution over observed patterns. Using such a similarity metric, clustering methods can find clusters of entity pairs that share similar patterns and can therefore be assumed to represent a relation. Ideally, a clustering method returns three kinds of structured information, each of which is highly relevant to information discovery and extraction in unknown corpora: Firstly, a set of clusters, each of which represents one distinct relation. Secondly, for each cluster a set of entity pairs between which this relation holds. Thirdly, for each cluster a set of discriminative patterns that extensionally describe the relation and may be used by an information extraction system to find further relation instances of this relation. The choice of patterns as well as …

Content truncated to 500 words