Unsupervised Translation Sense Clustering

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 773–782, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Unsupervised Translation Sense Clustering Mohit Bansal∗ UC Berkeley mbansal@cs.berkeley.edu John DeNero Google denero@google.com Dekang Lin Google lindek@google.com Abstract We propose an unsupervised method for clus- tering the translations of a word, such that the translations in each cluster share a com- mon semantic sense. Words are assigned to clusters based on their usage distribution in large monolingual and parallel corpora using the soft K-Means algorithm. In addition to de- scribing our approach, we formalize the task of translation sense clustering and describe a procedure that leverages WordNet for evalu- ation. By comparing our induced clusters to reference clusters generated from WordNet, we demonstrate that our method effectively identifies sense-based translation clusters and benefits from both monolingual and parallel corpora. Finally, we describe a method for an- notating clusters with usage examples. 1 Introduction The ability to learn a bilingual lexicon from a parallel corpus was an early and influential area of success for statistical modeling techniques in natural language processing. Probabilistic word alignment models can induce bilexical distributions over target-language translations of source-language words (Brown et al., 1993). However, word-to-word correspondences do not capture the full structure of a bilingual lexicon. Consider the example bilingual dictionary entry in Figure 1; in addition to enumerat- ing the translations of a word, the dictionary author has grouped those translations into three sense clus- ters. Inducing such a clustering would prove use- ful in generating bilingual dictionaries automatically or building tools to assist bilingual lexicographers. ∗Author was a summer intern with Google Research while conducting this research project. Colocar [co·lo·car´], va. 1. To arrange, to put in due place or order. 2. To place, to put in any place, rank condition or office, to provide a place or em- ployment. 3. To collocate, to locate, to lay. Figure 1: This excerpt from a bilingual dictionary groups English translations of the polysemous Spanish word colocar into three clusters that correspond to different word senses (Vel´azquez de la Cadena et al., 1965). This paper formalizes the task of clustering a set of translations by sense, as might appear in a pub- lished bilingual dictionary, and proposes an unsu- pervised method for inducing such clusters. We also show how to add usage examples for the translation sense clusters, hence providing complete structure to a bilingual dictionary. The input to this task is a set of source words and a set of target translations for each source word. Our proposed method clusters these translations in two steps. First, we induce a global clustering of the en- tire target vocabulary using the soft K-Means algo- rithm, which identifies groups of words that appear in similar contexts (in a monolingual corpus) and are translated in similar ways (in a parallel corpus). Sec- ond, we derive clusters over the translations of each source word by projecting the global clusters. We evaluate these …

Content truncated to 500 words