Language Model Adaptation for Difficult to Translate Phrases

Proceedings of the 13th Annual Conference of the EAMT, pages 160–167, Barcelona, May 2009 Language Model Adaptation for Difficult to Translate Phrases Behrang Mohit ∗ Frank Liberato ∗∗ Intelligent Systems Program ∗, Department of Computer Science ∗∗ University of Pittsburgh, Pittsburgh, PA 15260, USA {behrang,liberato,hwa}@cs.pitt.edu Rebecca Hwa ∗,∗∗ Abstract This paper investigates the idea of adapt- ing language models for phrases that have poor translation quality. We apply a se- lective adaptation criterion which uses a classifier to locate the most difficult phrase of each source language sentence. A spe- cial adapted language model is constructed for the highlighted phrase. Our adapta- tion heuristic uses lexical features of the phrase to locate the relevant parts of the parallel corpus for language model train- ing. As we vary the experimental setup by changing the size of the SMT training data, our adaptation method consistently shows strong improvements over the baseline sys- tems. 1 Introduction Statistical Machine Translation (SMT) systems generally use the same setup to translate all sen- tences. During decoding, the SMT engine searches through a large set of model parameters. Many pa- rameters are sparse, irrelevant and noisy with re- spect to the individual sentence that is being trans- lated. The dominant solution to this problem is to train with a larger corpus. This addresses the data sparsity problem, but it creates more irrele- vant model parameters. Moreover, large volumes of training data may not always be available. In this paper, we consider ways of filtering the irrelevant and noisy parameters in order to improve translation quality. We propose a lan- guage model adaptation method for the transla- tion of phrases. We construct one adapted lan- guage model per source language phrase by using c⃝ 2009 European Association for Machine Translation. its lexical features. Our method uses these features along with the parallel corpora sentence associa- tions to locate the relevant target language training sentences. Furthermore, we examine the idea of adapting the language model based on the level of difficulty that a phrase presents to the SMT system. We estimate the translation difficulty of phrases in a pre-translation step with either gold standard la- beling or a trained classifier. We find that only phrases that are deemed difficult for the SMT sys- tem, benefit from the adaptation. Finally, we asses the feasibility of our selective adaptation within a complete SMT pipeline. 2 Motivation Previously we explored the estimation of the trans- lation difficulty of phrases, using an automatic MT evaluation (BLEU) score (Mohit and Hwa, 2007). We used our estimation method to label a set of gold standard phrases with the difficulty informa- tion. Moreover, we showed that it is possible to au- tomatically learn the translation difficulty by con- structing a phrase difficulty classifier. In this paper, our aim is to improve MT quality by focusing on what we call Difficult To Trans- late Phrases (DTPs) within source language sen- tences. We compiled a group of DTPs for a base- line SMT system. We then manually examined …

Content truncated to 500 words