Multi-View Co-Training of Transliteration Model

Multi-View Co-training of Transliteration Model Jin-Shea Kuo Haizhou Li Chung-Hwa Telecomm. Laboratories, Taiwan d8807302@gmail.com Institute for Infocomm Research, Singapore 119613 hli@i2r.a-star.edu.sg Abstract This paper discusses a new approach to training of transliteration model from unlabeled data for transliteration extraction. We start with an inquiry into the formulation of transliteration model by considering different transliteration strategies as a multi-view problem, where each view exploits a natural division of transliteration features, such as phoneme- based, grapheme-based or hybrid features. Then we introduce a multi-view Co- training algorithm, which leverages compatible and partially uncorrelated information across different views to effectively boost the model from unlabeled data. Applying this algorithm to transliteration extraction, the results show that it not only circumvents the need of data labeling, but also achieves performance close to that of supervised learning, where manual labeling is required for all training samples. 1 Introduction Named entities are important content words in text documents. In many applications, such as cross- language information retrieval (Meng et al., 2001; Virga and Khudanpur, 2003) and machine translation (Knight and Graehl, 1998; Chen et al., 2006), one of the fundamental tasks is to identify these words. Imported foreign proper names constitute a good portion of such words, which are newly translated into Chinese by transliteration. Transliteration is a process of translating a foreign word into the native language by preserving its pronunciation in the original language, otherwise known as translation-by-sound. As new words emerge everyday, no lexicon is able to cover all transliterations. It is desirable to find ways to harvest transliterations from real world corpora. In this paper, we are interested in the learning of English to Chinese (E-C) transliteration model for transliteration extraction from the Web. A statistical transliteration model is typically trained on a large amount of transliteration pairs, also referred to a bilingual corpus. The correspondence between a transliteration pair may be described by the mapping of different basic pronunciation units (BPUs) such as phoneme- based1, or grapheme-based one, or both. We can see each type of BPU mapping as a natural division of transliteration features, which represents a view to the phonetic mapping problem. By using different BPUs, we approach the transliteration modeling and extraction problems from different views. This paper is organized as follows. In Section 2, we briefly introduce previous work. In Section 3, we conduct an inquiry into the formulation of transliteration model or phonetic similarity model (PSM) and consider it as a multi-view problem. In Section 4, we propose a multi-view Co-training strategy for PSM training and transliteration extraction. In Section 5, we study the effectiveness of proposed algorithms. Finally, we conclude in Section 6. 2 Related Work Studies on transliteration have been focused on transliteration modeling and transliteration extraction. The transliteration modeling approach deduces either phoneme-based or grapheme-based mapping rules using a generative model that is 1 Both phoneme and syllable based approaches are referred to as phoneme-based in this paper. 373 trained from a large bilingual corpus. Most of the works are devoted to phoneme-based …

Content truncated to 500 words