Neural and Statistical Methods for Leveraging Meta-information in Machine Translation

Neural and Statistical Methods for Leveraging Meta-information in Machine Translation Shahram Khadivi skhadivi@ebay.com eBay Inc., Aachen, Germany Patrick Wilken∗ patrick.wilken@rwth-aachen.de RWTH Aachen University, Aachen, Germany Leonard Dahlmann fdahlmann@ebay.com Evgeny Matusov ematusov@ebay.com eBay Inc., Aachen, Germany Abstract In this paper, we discuss different methods which use meta information and richer context that may accompany source language input to improve machine translation quality. We focus on category information of input text as meta information, but the proposed methods can be extended to all textual and non-textual meta information that might be available for the input text or automatically predicted using the text content. The main novelty of this work is to use state-of-the-art neural network methods to tackle this problem within a statistical machine translation (SMT) framework. We observe translation quality improvements up to 3% in terms of BLEU score in some text categories. 1 Introduction Using larger context in machine translation to improve its quality, including selection of correct word meaning, has been a challenging task. Correct translation of polysemous words is vital to transfer important information from source sentence to the translation. To find the correct sense of polysemous words and phrases, usually only the context of the source sentence is available. Depending on the use case, the context can be extended to the surrounding sentences, or external signals about the text, like its topic or genre. Therefore, we can consider all methods that try to use a larger context for translation as methods that can help MT system select the right translation for polysemous words. In e-commerce, the problem of polysemous words is more severe. For example, incorrect literal translation of a brand name like Apple, Coach, Diesel, Affliction, Avenue, Cables To Go, Free People, etc. misleads a potential buyer. It also may create legal issues when e.g. a wrong translation directs buyers to a competitive brand name. In e-commerce MT scenarios, one of the main tasks is often the translation of the titles of the items offered for sale. Such titles are short, non-grammatical, and the local context of a given word is very variational. Since item title translation has a crucial importance in cross-border trade for e-commerce, we are trying to leverage meta data available for each item to deal with these irregularities. Items in e-commerce inventory are usually classified according to a hierarchical taxonomy. The hierarchy itself contains top-level categories (like Clothing, Electronics) with varying degrees of depth in each top category. Although such hierarchy is created based on ∗Patrick Wilken has contributed to this work during his internship at eBay Inc. Proceedings of MT Summit XVI, vol.1: Research Track Nagoya, Sep. 18-22, 2017 | p. 41 business insights, it implicitly groups objects which can be described in semantically similar terms. Therefore, we expect less variation in word senses within a category, with ambiguity decreasing deeper in the tree. For instance, Apple is very likely to be a brand name, not a fruit in Smartphones. Therefore, item category information can potentially provide a strong signal to …

Content truncated to 500 words