Unsupervised Induction of Cross-Lingual Semantic Relations

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 681–692, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Unsupervised Induction of Cross-lingual Semantic Relations Mike Lewis School of Informatics University of Edinburgh Edinburgh, EH8 9AB, UK mike.lewis@ed.ac.uk Mark Steedman School of Informatics University of Edinburgh Edinburgh, EH8 9AB, UK steedman@inf.ed.ac.uk Abstract Creating a language-independent meaning representation would benefit many cross- lingual NLP tasks. We introduce the first un- supervised approach to this problem, learn- ing clusters of semantically equivalent English and French relations between referring expres- sions, based on their named-entity arguments in large monolingual corpora. The clusters can be used as language-independent semantic relations, by mapping clustered expressions in different languages onto the same relation. Our approach needs no parallel text for train- ing, but outperforms a baseline that uses ma- chine translation on a cross-lingual question answering task. We also show how to use the semantics to improve the accuracy of machine translation, by using it in a simple reranker. 1 Introduction Identifying a language-independent semantics is a major long term goal of computational linguistics, and is interesting both theoretically and for practical applications. It assumes that semantically equiva- lent sentences in any language can be mapped onto a common meaning representation. Such a repre- sentation would be of great utility for tasks such as translation, relation extraction, summarization, question answering, and information retrieval. Re- gardless of whether it is even possible to create such a semantics, we show that an incomplete version can be useful for downstream tasks. Semantic machine translation aims to map a source language to a language-independent meaning representation, and then generate the target language translation from this. It is hoped this would allevi- ate the difficulties of simpler models when translat- ing between languages with very different word or- dering and syntax (Vauquois, 1968). Despite many attempts to define interlingual representations (Mi- tamura et al., 1991; Beale et al., 1995; Banarescu et al., 2013), state-of-the-art machine translation still uses phrase-based models (Koehn et al., 2007). The major obstacle to defining interlinguas has been de- vising a meaning representation that is language- independent, but capable of expressing the limitless number of meanings that natural languages can ex- press (Dorr et al., 2004). Our approach avoids this problem by utilizing the methods of distributional semantics. Recent work has shown that paraphrases of expressions can be learned by clustering those with similar arguments (Poon and Domingos, 2009; Yao et al., 2011; Lewis and Steedman, 2013)—for example learning that X wrote Y and X is the author of Y are equivalent if they appear in a corpus with similar (X, Y) argument- pairs such as {(Shakespeare, Macbeth), (Dickens, Oliver Twist)}. We extend this to the multilingual case, aiming to also map the French equivalents X a ´ecrit Y and Y est un roman de X on to the same cluster as the English paraphrases. Conceptually, we treat a foreign expression as a paraphrase of an English expression. The cluster identifier …

Content truncated to 500 words