Topic adaptation for machine translation of e-commerce content

Topic Adaptation for Machine Translation of E-commerce Content Prashant Mathur prashant@fbk.eu Marcello Federico∗ federico@fbk.eu FBK, Trento, Italy Selc¸uk K¨opr¨u skopru@ebay.com Sharam Khadivi skhadivi@ebay.com Hassan Sawaf hsawaf@ebay.com eBay Inc., San Jose (CA), USA Abstract We describe effort to improve machine translation of item titles found in a large e-commerce inventory through topic modeling and adaptation. Item titles are short texts which typically contain brand names that do not have to be translated, and item attributes whose translation often depends on the context. Both issues call for robust methods to integrate context infor- mation in the machine translation process in order to reduce translation ambiguity. We survey both existing topic adaptation approaches and propose novel methods that augment the standard phrase-table models with sparse features and dense features measuring the topic match between each phrase-pair and the input text. We report extensive experiments on the translation of item titles from English into Brazilian Portuguese, and show the impact of topic adaptation both with and without domain adaptation. 1 Introduction Domain adaptation and topic adaptation can be seen as complementary methods to cope with the variability between training and testing data in machine translation. Under this perspective, domain modeling typically assumes training data partitioned according to human defined labels, while topic modeling builds on fuzzy clustering of the data and automatically learned labels. In statistical machine translation, knowledge about the domain or topics of the input can be lever- aged to bias the system towards training data matching the same labels of the input. Potential advantages of topic adaptation over domain adaptation is that fuzzy clustering can better cope with data sparseness than hard clustering, and that automatic labels do not require any man- ual intervention. On the other side, however, domain adaptation becomes difficult to beat when training data can be effectively and naturally partitioned into in-domain and out-of-domain data. In this paper we discuss the application of domain and topic adaptation to an e-commerce online MT system (Guha and Heger, 2014), whose target is the translation of user queries and all item titles, descriptions, and specifics shown in the search result pages. In particular, our investigation focuses on the translation of item titles, which consists of concise user-generated texts describing each item put on sale. Item titles differ in several ways from text genres typ- ically considered in machine translation research. Titles are usually short texts, of maximum ∗Most of the work was carried out during a visiting period of the first two authors at eBay Inc. Proceedings of MT Summit XV, vol.1: MT Researchers' Track Miami, Oct 30 - Nov 3, 2015 | p. 270 100 characters, with a simple syntactic structure, and containing brand names, feature values, as well as specific jargon. Their translation poses several challenges (Sanchez and Badeka, 2014), such as the correct rendering of proper names, which can often be confused with com- mon names, and the correct translation of product features, which often depends on the context. From a statistical learning perspective, MT of item …

Content truncated to 500 words