Automated Translation of Semantic Relationships

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 241–249, Beijing, August 2010 Automated Translation of Semantic Relationships Dmitry Davidov ICNC Hebrew University of Jerusalem Ari Rappoport Institute of Computer Science Hebrew University of Jerusalem arir@cs.huji.ac.il Abstract We present a method for translating se- mantic relationships between languages where relationships are defined as pattern clusters. Given a pattern set which rep- resents a semantic relationship, we use the web to extract sample term pairs of this relationship. We automatically trans- late the obtained term pairs using multi- lingual dictionaries and disambiguate the translated pairs using web counts. Finally we discover the set of most relevant tar- get language patterns for the given rela- tionship. The obtained pattern set can be utilized for extraction of new relationship examples for the target language. We evaluate our method on 11 diverse tar- get languages. To assess the quality of the discovered relationships, we use an au- tomatically generated cross-lingual SAT analogy test, WordNet relationships, and concept-specific relationships, achieving high precision. The proposed framework allows fully automated cross-lingual rela- tionship mining and construction of mul- tilingual pattern dictionaries without rely- ing on parallel corpora. 1 Introduction Acquiring and understanding semantic relation- ships is crucial for many NLP applications. In many cases, we would like to know if a given term pair participates in a specified semantic re- lationship or if two different term pairs encode the same (possibly unspecified) type of relation- ship. Beyond the well-known major relationship types such as hyponymy (is-a) and meronymy (part-of), there is a huge number of other rela- tionships between objects and concepts. Exam- ples include general relations such as larger-than, contained-in, liked-by and domain specific ones such as country-language, product-manufacturer, product-seller, drug-disease etc. The vast majority of NLP research is done in a few languages for which extensive corpora (in- cluding the web) are available. As a result, most relationship retrieval studies and lexical database compilation efforts target only a few languages. However, due to the substantial growth of the mul- tilingual web1 and a growing demand for NLP application coverage for less common languages, there is a need for relationship data in many less studied languages. In this paper we address the task of translating relationships between languages, which has two obvious benefits. First, it can directly help appli- cations such as machine translation, cross-lingual information retrieval, cross-lingual web mining and the construction and enrichment of seman- tic databases. Second, it can assist applications in a single language, especially when compensat- ing for a relative scarcity of resources in that lan- guage. We focus on relations between two enti- ties, which are the most common type. When discussing the translation of relation- ships, it is important to define how these are rep- resented and in what way the task differs from MT. While there is wide agreement on the def- inition and representation of major relationship types such as hypernymy and (to a lesser extent) meronymy, there is no single accepted method (or …

Content truncated to 500 words