Translation Quality-Based Supplementary Data Selection by Incremental Update of Translation Models

Proceedings of COLING 2012: Technical Papers, pages 149–166, COLING 2012, Mumbai, December 2012. Translation Quality-Based Supplementary Data Selection by Incremental Update of Translation Models Pratyush Banerjee1, Sudip Kumar Naskar1, Johann Roturier2, Andy Way3, Josef van Genabith1 (1) CNGL, School of Computing, DCU {pbanerjee,snaskar,josef}@computing.dcu.ie (2) Symantec Ireland, Dublin johann_roturier@symantec.com (3) Capita Translation and Interpreting, Delph, UK andy.way@capita-ti.com ABSTRACT Supplementary data selection from out-of-domain or related-domain data is a well established technique in domain adaptation of statistical machine translation. The selection criteria for such data are mostly based on measures of similarity with available in-domain data, but not directly in terms of translation quality. In this paper, we present a technique for selecting supplementary data to improve translation performance, directly in terms of translation quality, measured by automatic evaluation metric scores. Batches of data selected from out-of-domain corpora are incrementally added to an existing baseline system and evaluated in terms of translation quality on a development set. A batch is selected only if its inclusion improves translation quality. To assist the process, we present a novel translation model merging technique that allows rapid retraining of the translation models with incremental data. When incorporated into the ‘in-domain’ translation models, the final cumulatively selected datasets are found to provide statistically significant improvements for a number of different supplementary datasets. Furthermore, the translation model merging technique is found to perform on a par with state-of-the-art methods of phrase-table combination. KEYWORDS: Statistical Machine Translation, Domain Adaptation, Supplementary Data Selec- tion, Model Merging, Incremental Update. 149 1 Introduction Statistical Machine Translation (SMT) has grown to be the most dominant machine translation paradigm. A prerequisite for SMT is the availability of sizeable parallel training data. The popularity of SMT has led to the free availability of a variety of parallel corpora on the web. While some such corpora comprise data from wide-coverage domains such as politics or news, others are based on much more focused and narrower domains such as medical texts or software manuals. In order to create an SMT system for a specific domain which does not have sufficient parallel training data, domain adaptation methods are necessary to best utilise supplementary parallel training data from available out-of-domain or related-domain corpora. However, the conventional wisdom of more data being better does not always hold true for domain-specific systems (Axelrod et al., 2011). Adding a lot of out-of-domain data to an in-domain SMT system tends to overwhelm the in-domain characteristics, thereby negatively affecting translation quality. Thus relevant data selection from large supplementary out-of-domain data plays an important part in domain adaptation of SMT systems. In this paper we focus our efforts on creating an SMT system for translating user-generated forum content in Symantec web forums. Being a multinational company, Symantec supports web forums for its products and services in multiple languages with the English forum being both the oldest and (by far) the largest with considerable amounts of relevant information. Translating the forum content enables this information to be available across all languages. Moreover, these forums have also …

Content truncated to 500 words