PATTY: A Taxonomy of Relational Patterns with Semantic Types

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1135–1145, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics PATTY: A Taxonomy of Relational Patterns with Semantic Types Ndapandula Nakashole, Gerhard Weikum, Fabian Suchanek Max Planck Institute for Informatics Saarbr¨ucken, Germany {nnakasho,weikum,suchanek}@mpi-inf.mpg.de Abstract This paper presents PATTY: a large resource for textual patterns that denote binary relations between entities. The patterns are semanti- cally typed and organized into a subsumption taxonomy. The PATTY system is based on ef- ficient algorithms for frequent itemset mining and can process Web-scale corpora. It har- nesses the rich type system and entity popu- lation of large knowledge bases. The PATTY taxonomy comprises 350,569 pattern synsets. Random-sampling-based evaluation shows a pattern accuracy of 84.7%. PATTY has 8,162 subsumptions, with a random-sampling-based precision of 75%. The PATTY resource is freely available for interactive access and download. 1 Introduction Motivation. WordNet (Fellbaum 1998) is one of the most widely used lexical resources in computer sci- ence. It groups nouns, verbs, and adjectives into sets of synonyms, and arranges these synonyms in a tax- onomy of hypernyms. WordNet is limited to single words. It does not contain entire phrases or pat- terns. For example, WordNet does not contain the pattern X is romantically involved with Y. Just like words, patterns can be synonymous, and they can subsume each other. The pattern X is romantically involved with Y is synonymous with the pattern X is dating Y. Both are subsumed by X knows Y. Patterns for relations are a vital ingredient for many appli- cations, including information extraction and ques- tion answering. If a large-scale resource of relational patterns were available, this could boost progress in NLP and AI tasks. Yet, existing large-scale knowledge bases are mostly limited to abstract binary relationships be- tween entities, such as “bornIn” (Auer 2007; Bol- lacker 2008; Nastase 2010; Suchanek 2007). These do not correspond to real text phrases. Only the Re- Verb system (Fader 2011) yields a larger number of relational textual patterns. However, no attempt is made to organize these patterns into synonymous patterns, let alone into a taxonomy. Thus, the pat- terns themselves do not exhibit semantics. Goal. Our goal in this paper is to systematically compile relational patterns from a corpus, and to im- pose a semantically typed structure on them. The result we aim at is a WordNet-style taxonomy of binary relations. In particular, we aim at patterns that contain semantic types, such as ⟨singer⟩ sings ⟨song⟩. We also want to automatically generalize syntactic variations such as sings her ⟨song⟩ and sings his ⟨song⟩, into a more general pattern sings [prp] ⟨song⟩ with POS tag [prp]. Analogously but more demandingly, we want to automatically infer that the above patterns are semantically subsumed by the pattern ⟨musician⟩ performs on ⟨musical composition⟩ with more general types for the entity arguments in the pattern. Compiling and organizing such patterns is chal- lenging for the following reasons. 1) The number …

Content truncated to 500 words