Hallucinating Phrase Translations for Low Resource MT

Proceedings of the Eighteenth Conference on Computational Language Learning, pages 160–170, Baltimore, Maryland USA, June 26-27 2014. c⃝2014 Association for Computational Linguistics Hallucinating Phrase Translations for Low Resource MT Ann Irvine Center for Language and Speech Processing Johns Hopkins University Chris Callison-Burch Computer and Information Science Dept. University of Pennsylvania Abstract We demonstrate that “hallucinating” phrasal translations can significantly im- prove the quality of machine translation in low resource conditions. Our hallucinated phrase tables consist of entries composed from multiple unigram translations drawn from the baseline phrase table and from translations that are induced from mono- lingual corpora. The hallucinated phrase table is very noisy. Its translations are low precision but high recall. We counter this by introducing 30 new feature functions (including a variety of monolingually- estimated features) and by aggressively pruning the phrase table. Our analysis evaluates the intrinsic quality of our hallucinated phrase pairs as well as their impact in end-to-end Spanish-English and Hindi-English MT. 1 Introduction In this work, we augment the translation model for a low-resource phrase-based SMT system by auto- matically expanding its phrase table. We “halluci- nate” new phrase table entries by composing the unigram translations from the baseline system’s phrase table and translations learned from compa- rable monolingual corpora. The composition pro- cess yields a very large number of new phrase pair translations, which are high recall but low preci- sion. We filter the phrase table using a new set of feature functions estimated from monolingual cor- pora. We evaluate the hallucinated phrase pairs in- trinsically as well as in end-to-end machine trans- lation. The augmented phrase table provides more coverage than the original phrase table, while be- ing high quality enough to improve translation per- formance. We propose a four-part approach to hallucinat- ing and using new phrase pair translations: 1. Learn potential translations for out-of- vocabulary (OOV) words from comparable monolingual corpora 2. “Hallucinate” a large, noisy set of phrase translations by composing unigram transla- tions from the baseline model and from the monolingually-induced bilingual dictionary 3. Use comparable monolingual corpora to score, rank, and prune the huge number of hallucinated translations 4. Augment the baseline phrase table with hal- lucinated translations and new feature func- tions estimated from monolingual corpora We define an algorithm for generating loosely compositional phrase pairs, which we use to hal- lucinate new translations. In oracle experiments, we show that such loosely compositional phrase pairs contribute substantially to the performance of end-to-end SMT, beyond that of component un- igram translations. In our non-oracle experiments, we show that adding a judiciously pruned set of automatically hallucinated phrase pairs to an end- to-end baseline SMT model results in a signifi- cant improvement in translation quality for both Spanish-English and Hindi-English. 2 Motivation Translation models learned over small amounts of parallel data suffer from the problem of low coverage. That is, they do not include trans- lations for many words and phrases. Unknown 160 words, or out-of-vocabulary (OOV) words, have been the focus of previous work on integrating …

Content truncated to 500 words