Topic Models + Word Alignment = A Flexible Framework for Extracting Bilingual Dictionary from Comparable Corpus

Proceedings of the Seventeenth Conference on Computational Natural Language Learning, pages 212–221, Sofia, Bulgaria, August 8-9 2013. c⃝2013 Association for Computational Linguistics Topic Models + Word Alignment = A Flexible Framework for Extracting Bilingual Dictionary from Comparable Corpus Xiaodong Liu, Kevin Duh and Yuji Matsumoto Graduate School of Information Science Nara Institute of Science and Technology 8916-5 Takayama, Ikoma, Nara 630-0192, Japan {xiaodong-l,kevinduh,matsu}@is.naist.jp Abstract We propose a flexible and effective frame- work for extracting a bilingual dictionary from comparable corpora. Our approach is based on a novel combination of topic modeling and word alignment techniques. Intuitively, our approach works by con- verting a comparable document-aligned corpus into a parallel topic-aligned cor- pus, then learning word alignments us- ing co-occurrence statistics. This topic- aligned corpus is similar in structure to the sentence-aligned corpus frequently used in statistical machine translation, enabling us to exploit advances in word alignment re- search. Unlike many previous work, our framework does not require any language- specific knowledge for initialization. Fur- thermore, our framework attempts to han- dle polysemy by allowing multiple trans- lation probability models for each word. On a large-scale Wikipedia corpus, we demonstrate that our framework reliably extracts high-precision translation pairs on a wide variety of comparable data condi- tions. 1 Introduction A machine-readable bilingual dictionary plays a very important role in many natural language pro- cessing tasks. In machine translation (MT), dic- tionaries can help in the domain adaptation set- ting (Daume III and Jagarlamudi, 2011). In cross-lingual information retrieval (CLIR), dictio- naries serve as efficient means for query trans- lation (Resnik et al., 2011). Many other multi- lingual applications also rely on bilingual dictio- naries as integral components. One approach for building a bilingual dictio- nary resource uses parallel sentence-aligned cor- pora. This is often done in the context of Statis- tical MT, using word alignment algorithms such as the IBM models (Brown et al., 1993; Och and Ney, 2003). Unfortunately, parallel corpora may be scarce for certain language-pairs or domains of interest (e.g., medical and microblog). Thus, the use of comparable corpora for bilin- gual dictionary extraction has become an active research topic (Haghighi et al., 2008; Vuli´c et al., 2011). Here, a comparable corpus is defined as collections of document pairs written in dif- ferent languages but talking about the same topic (Koehn, 2010), such as interconnected Wikipedia articles. The challenge with bilingual dictionary extraction from comparable corpus is that exist- ing word alignment methods developed for paral- lel corpus cannot be directly applied. We believe there are several desiderata for bilin- gual dictionary extraction algorithms: 1. Low Resource Requirement: The approach should not rely on language-specific knowl- edge or a large scale seed lexicon. 2. Polysemy Handling: One should handle the fact that a word form may have multiple meanings, and such meanings may be trans- lated differently. 3. Scalability: The approach should run effi- ciently an massively large-scale datasets. Our framework addresses the above desired points by exploiting a novel combination of topic models and word alignment, as …

Content truncated to 500 words