Active Learning for Statistical Phrase-based Machine Translation

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 415–423, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Active Learning for Statistical Phrase-based Machine Translation∗ Gholamreza Haffari and Maxim Roy and Anoop Sarkar School of Computing Science Simon Fraser University Burnaby, BC, Canada {ghaffar1,maximr,anoop}@cs.sfu.ca Abstract Statistical machine translation (SMT) mod- els need large bilingual corpora for train- ing, which are unavailable for some language pairs. This paper provides the first serious ex- perimental study of active learning for SMT. We use active learning to improve the qual- ity of a phrase-based SMT system, and show significant improvements in translation com- pared to a random sentence selection baseline, when test and training data are taken from the same or different domains. Experimental re- sults are shown in a simulated setting using three language pairs, and in a realistic situa- tion for Bangla-English, a language pair with limited translation resources. 1 Introduction Statistical machine translation (SMT) systems have made great strides in translation quality. However, high quality translation output is dependent on the availability of massive amounts of parallel text in the source and target language. However, there are a large number of languages that are considered “low- density”, either because the population speaking the language is not very large, or even if millions of peo- ple speak the language, insufficient amounts of par- allel text are available in that language. A statistical translation system can be improved or adapted by incorporating new training data in the form of parallel text. In this paper, we propose sev- eral novel active learning (AL) strategies for statis- tical machine translation in order to attack this prob- lem. Conventional techniques for AL of classifiers are problematic in the SMT setting. Selective sam- pling of sentences for AL may lead to a parallel cor- pus where each sentence does not share any phrase ∗We would like to thank Chris Callison-Burch for fruitful discussions. This research was partially supported by NSERC, Canada (RGPIN: 264905) and by an IBM Faculty Award to the third author. pairs with the others. Thus, new sentences cannot be translated since we lack evidence for how phrase pairs combine to form novel translations. In this pa- per, we take the approach of exploration vs. exploita- tion: where in some cases we pick sentences that are not entirely novel to improve translation statis- tics, while also injecting novel translation pairs to improve coverage. There may be evidence to show that AL is use- ful even when we have massive amounts of parallel training data. (Turchi et al., 2008) presents a com- prehensive learning curve analysis of a phrase-based SMT system, and one of the conclusions they draw is, “The first obvious approach is an effort to iden- tify or produce data sets on demand (active learning, where the learning system can request translations of specific sentences, to satisfy its information needs).” Despite the promise of active learning for SMT there has been very little experimental work …

Content truncated to 500 words