Graph-based Semi-Supervised Learning of Translation Models from Monolingual Data

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 676–686, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Graph-based Semi-Supervised Learning of Translation Models from Monolingual Data Avneesh Saluja⇤ Carnegie Mellon University Pittsburgh, PA 15213, USA avneesh@cs.cmu.edu Hany Hassan, Kristina Toutanova, Chris Quirk Microsoft Research Redmond, WA 98502, USA hanyh,kristout,chrisq@microsoft.com Abstract Statistical phrase-based translation learns translation rules from bilingual corpora, and has traditionally only used monolin- gual evidence to construct features that rescore existing translation candidates. In this work, we present a semi-supervised graph-based approach for generating new translation rules that leverages bilingual and monolingual data. The proposed tech- nique first constructs phrase graphs using both source and target language mono- lingual corpora. Next, graph propaga- tion identifies translations of phrases that were not observed in the bilingual cor- pus, assuming that similar phrases have similar translations. We report results on a large Arabic-English system and a medium-sized Urdu-English system. Our proposed approach significantly improves the performance of competitive phrase- based systems, leading to consistent im- provements between 1 and 4 BLEU points on standard evaluation sets. 1 Introduction Statistical approaches to machine translation (SMT) use sentence-aligned, parallel corpora to learn translation rules along with their probabil- ities. With large amounts of data, phrase-based translation systems (Koehn et al., 2003; Chiang, 2007) achieve state-of-the-art results in many ty- pologically diverse language pairs (Bojar et al., 2013). However, the limiting factor in the suc- cess of these techniques is parallel data availabil- ity. Even in resource-rich languages, learning re- liable translations of multiword phrases is a chal- lenge, and an adequate phrasal inventory is crucial ⇤This work was done while the first author was interning at Microsoft Research for effective translation. This problem is exacer- bated in the many language pairs for which par- allel resources are either limited or nonexistent. While parallel data is generally scarce, monolin- gual resources exist in abundance and are being created at accelerating rates. Can we use monolin- gual data to augment the phrasal translations ac- quired from parallel data? The challenge of learning translations from monolingual data is of long standing interest, and has been approached in several ways (Rapp, 1995; Callison-Burch et al., 2006; Haghighi et al., 2008; Ravi and Knight, 2011). Our work in- troduces a new take on the problem using graph- based semi-supervised learning to acquire trans- lation rules and probabilities by leveraging both monolingual and parallel data resources. On the source side, labeled phrases (those with known translations) are extracted from bilingual corpora, and unlabeled phrases are extracted from mono- lingual corpora; together they are embedded as nodes in a graph, with the monolingual data de- termining edge strengths between nodes (§2.2). Unlike previous work (Irvine and Callison-Burch, 2013a; Razmara et al., 2013), we use higher order n-grams instead of restricting to unigrams, since our approach goes beyond OOV mitigation and can enrich the entire translation model by using evidence from monolingual text. This enhance- ment alone results in an improvement of …

Content truncated to 500 words