Cross Language Information Integration Bridging the Gap

Integrating information in multiple natural languages is a challenging task that often requires manually created linguistic resources such as a bilingual dictionary or examples of direct translations of text. Comparable coprora have important properties that can be exploited to infer word translations without a bilingual dictionary. The main premise underlying comparable corpora is that translations of two co-occurring words in a source language also co-occur in the target language. Past work has made use of this property for directly extracting a lexicon from comparable corpora. A major drawback of this work is the number of parameters (translation probabilities) to be estimated which increases quadratically with the vocabulary sizes. In this paper, we propose a novel cluster based approach that tries to map groups of related words, rather than the individual words themselves. This method has the advantage that the number of parameters to be estimated remains indepdendent of the vocabulary size. Experiments show that the computational demands of this problem are really very high and the method seems to work well for small data sets. Approximating the multinomials produced by the clustering algorithm (LDA) to speed up computation degrades the performance of the model.