A Simplification-Translation-Restoration Framework for Cross-Domain SMT Applications

Proceedings of COLING 2012: Technical Papers, pages 545–560, COLING 2012, Mumbai, December 2012. A Simplification-Translation-Restoration Framework for Cross-Domain SMT Applications Han-Bin Chen1, Hen-Hsen Huang1, Hsin-Hsi Chen1, and Ching-Ting Tan2 (1) Department of Computer Science and Information Engineering, National Taiwan University, Taipei, Taiwan (2) National Taiwan University Hospital, Taipei, Taiwan {hbchen, hhhuang}@nlg.csie.ntu.edu.tw, {hhchen, tanct5222}@ntu.edu.tw ABSTRACT Integration of domain specific knowledge into a general purpose statistical machine translation (SMT) system poses challenges due to insufficient bilingual corpora. In this paper we propose a simplification-translation-restoration (STR) framework for domain adaptation in SMT by simplifying domain specific segments of a text. For an in-domain text, we identify the critical segments and modify them to alleviate the data sparseness problem in the out-domain SMT system. After we receive the translation result, these critical segments are then restored according to the provided in-domain knowledge. We conduct experiments on an English-to- Chinese translation task in the medical domain and evaluate each step of the STR framework. The translation results show significant improvement of our approach over the out-domain and the naïve in-domain SMT systems. 用於跨領域統計式機器翻譯系統之簡化-翻譯-還原架構 摘要 因為雙語語料的不足，將特定領域知識整合到一般用途的統計式機器翻譯(SMT)系統具有 相當挑戰性。在本篇論文中，我們提出一個簡化-翻譯-還原(STR)的架構，藉由簡化特定領 域的片段來達成SMT的領域調適。對於一篇領域內的文字，我們首先辨識其中重要的片段 再做修改以減輕領域外SMT系統的資料稀疏問題。我們取得翻譯結果後，根據提供的領域 內知識將這些重要片段還原。最後我們進行了醫療領域的英中翻譯的實驗，並且評估STR 架構內的每一步驟。翻譯結果顯示我們的方法顯著地優於領域外的SMT系統，以及簡易型 的領域內SMT系統。 KEYWORDS : Cross-Domain SMT, Domain Adaptation, Statistical Machine Translation 關鍵詞 : 跨領域統計式機器翻譯, 領域調適, 統計式機器翻譯 545 1 Introduction Over the past decades, the rapid growth of available parallel corpus makes SMT development feasible, and SMT system has gradually moved toward practical use because of its relatively acceptable translation speed and quality. A phrase-based SMT system (Koehn et al., 2003; Koehn, 2004), for example, trains a phrase table from a large bilingual corpus as its translation model, and decodes source language input in polynomial time with greedy algorithms such as beam search. It translates phrases as basic units, and thus captures short-range reordering phenomena between source and target languages. Generally phrase-based SMT models outperform word-based ones (Koehn et al., 2003). However, an SMT system fails to capture long-range contextual knowledge due to the limited horizon and the sparseness nature of lexical n-grams. These drawbacks reduce the translation quality in terms of translation and reordering errors, especially when limited bilingual corpus is available for estimating translation model. The data sparseness problem may worsen when we build an SMT system for a specific domain but have small or no bilingual in-domain corpus. The translation performance could be seriously degraded since the SMT system cannot gather the statistical evidence of a segment containing out-of-vocabulary (OOV) words. For some domains, a bilingual dictionary containing source- target term pairs is available. With such in-domain knowledge, one can force an SMT system to translate OOV terms according to the dictionary. In this way, we correctly translate in-domain terms. However, translation quality may still be unsatisfying under this naïve integration because in-domain terms are rare or unseen in the background SMT model. Hence the context of these in-domain terms can hardly be captured. FIGURE 1 shows an incorrect English-Chinese translation of a sentence in a medical record by the online Google Translate service. In this example, it correctly translates the diagnosis term "crystal …

Content truncated to 500 words