Domain Adaptation in SMT of User-Generated Forum Content Guided by OOV Word Reduction: Normalization and/or Supplementary Data

Domain Adaptation in SMT of User-Generated Forum Content Guided by OOV Word Reduction: Normalization and/or Supplementary Data? Pratyush Banerjee, Sudip Kumar Naskar, Johann Roturier1, Andy Way2, Josef van Genabith CNGL, School of Computing, Dublin City University, Dublin, Ireland {pbanerjee,snaskar,josef}@computing.dcu.ie 1 Symantec Limited, Dublin, Ireland johann roturier@symantec.com 2 Applied Language Solutions, Delph, UK andy.way@appliedlanguage.com Abstract This paper reports a set of domain adap- tation techniques for improving Statisti- cal Machine Translation (SMT) for user- generated web forum content. We inves- tigate both normalization and supplemen- tary training data acquisition techniques, all guided by the aim of reducing the num- ber of Out-Of-Vocabulary (OOV) items in the target language with respect to the training data. We classify OOVs into a set of types, and address each through ded- icated normalization and/or supplemen- tary training material selection-based ap- proaches. We investigate the effect of these methods both in an additive as well as a contrastive scenario. Our findings show that (i) normalization and supple- mentary training material techniques can be complementary, (ii) for general forum data, fully automatic supplementary train- ing data acquisition can perform as well or sometimes better than semi-automatic normalization (although tackling different types of OOVs) and (iii) for very noisy data, normalization really pays off. 1 Introduction Web-forums are rich sources of user-generated content on the web. The increasing popularity of technical forums have motivated major IT compa- nies like Symantec to create and support forums around their products and services. For individual users or larger customers, such forums provide an easy source of information and a viable alternative to traditional customer service options. Being a c⃝ 2012 European Association for Machine Translation. multinational company, Symantec hosts its forums in different languages (English, German, French etc), but currently the content is siloed in each language. Clearly, translating the forums to make information available across languages would be beneficial for Symantec as well as its multilingual customer base. This forms the primary motivation of techniques presented here. Despite growing interest in translation of forum data (Flournoy and Rueppel, 2010), to date, sur- prisingly little research has actually focussed on forum data translation (Roturier and Bensadoun, 2011). Compared to professionally edited text, user-generated forum data is often more noisy, tak- ing some liberty with commonly established gram- mar, punctuation and spelling norms. For our re- search, we use translation memory (TM) data from Symantec, which is part of their corporate doc- umentation, professionally edited and generally conforming to the Symantec controlled language guidelines. On the other hand, our target data (fo- rum) is only lightly moderated and does not con- form to any publication quality guidelines. Hence despite being from the same IT domain, there is a significant difference in style between the training and the test data. In this paper, we focus our efforts on systematically reducing this difference through the use of both normalization and supplementary training material acquisition techniques. Our research was conducted on English to Ger- man (En–De) and English to French (En–Fr) lan- guage directions. To …

Content truncated to 500 words