Alignment by Bilingual Generation and Monolingual Derivation

Proceedings of COLING 2012: Technical Papers, pages 1963–1978, COLING 2012, Mumbai, December 2012. Alignment by Bilingual Generation and Monolingual Derivation Toshiaki Nakazawa Sadao Kurohashi Graduate School of Informatics, Kyoto University Yoshida-honmachi, Sakyo-ku Kyoto, 606-8501, Japan nakazawa@nlp.ist.i.kyoto-u.ac.jp, kuro@i.kyoto-u.ac.jp ABSTRACT One of the main issues in a word alignment task is the difficulty of handling function words that do not have direct translations which we call unique function words. They are often aligned to some words in the other language incorrectly. This is prominent in language pairs with very different sentence structures. In this paper, we propose a novel approach for handling unique function words. The proposed model monolingually derives unique function words from bilin- gually generated treelet pairs. The monolingual derivation prevents incorrect alignments for unique function words. The derivation probabilities are estimated from a large monolingual corpus, which is much easier to acquire than a parallel corpus. Also, the proposed alignment model uses semantic-head dependency trees where dependency relations between words be- come similar in each language. Experimental results on an English-Japanese corpus show that the proposed model achieves better alignment and translation quality compared with the baseline models. TITLE AND ABSTRACT IN JAPANESE 二言語の生成と単言語の派生によるアライメント 単語アライメントタスクにおける主な問題の一つは、機能語の中でも相手言語に対応する語 が存在しない機能語の扱いの困難さである。我々はこのような語を孤立機能語と呼ぶ。孤立 機能語は、相手言語の何らかの単語に不適切に対応付けられることが多く、これは特に文構 造が大きく異なる言語対において顕著である。本論文では、孤立機能語を扱うための新しい 手法を提案する。提案モデルは、二言語で生成された部分木ペアから、孤立機能語をそれぞ れ単言語で派生することにより、孤立機能語が誤って対応付けられることを防ぐ。派生確率 は、対訳コーパスに比べて入手が容易である大規模単言語コーパスから推定する。また提案 モデルは、単語同士の依存関係が各言語で近くなるように、意味主辞依存構造木を用いる。 英日コーパスでの実験結果から、提案モデルはベースラインモデルと比べてより良いアライ メントおよび翻訳精度を実現した。 KEYWORDS: monolingual derivation, semantic-head dependency tree, treelet alignment. KEYWORDS IN JAPANESE: 単言語の派生, 意味主辞依存構造木, 木構造アライメント. 1963 1 Introduction Alignment accuracy is crucial for providing high quality corpus-based machine translation sys- tems because translation knowledge is acquired from an aligned training corpus. For similar language pairs, alignment accuracy is high. Less than 10% alignment error rate (AER) for French-English has been achieved by the conventional word alignment tool GIZA++, an im- plementation of the alignment models called the IBM models (Brown et al., 1993), with some heuristic symmetrization rules. However, for distant language pairs such as English-Japanese, the conventional alignment method is quite inadequate (achieving an AER of about 20%). There are two main issues in a word alignment task for distant language pairs: one is the word order difference, while the other relates to function words. The word order issue has to some extent been solved by using word dependency trees in the alignment model (Nakazawa and Kurohashi, 2011). Most of the remaining alignment errors are related to func- tion words such as English articles and Japanese case markers (Wu et al., 2011) because they do not have counterparts in the other language. As an example, most of the errors in Figure 1 are related to function words: “has” and “は (topic-marker)” in example (A), and “although”, “は (topic-marker)”, “を (ACC)” and “が (but)” in example (B). Several previous works focused on alignment errors of function words. Isozaki et al. (2010) inserted pseudo nodes in English sentences for Japanese function words. Wu et al. (2011) removed the alignment of some function words to effectively acquire translation rules using the underlying word alignment by GIZA++. Nevertheless, these methods for dealing with function words are ad-hoc and based on hand-crafted rules. In this paper, we propose a novel …

Content truncated to 500 words