Improving Word Alignment Using Linguistic Code Switching Data

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 1–9, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Improving Word Alignment Using Linguistic Code Switching Data Fei Huang∗and Alexander Yates Temple University Computer and Information Sciences 324 Wachman Hall Philadelphia, PA 19122 {fei.huang,yates}@temple.edu Abstract Linguist Code Switching (LCS) is a situation where two or more languages show up in the context of a single conversation. For example, in English- Chinese code switching, there might be a sentence like “· ‚15© ¨ � k ‡meeting (We will have a meeting in 15 minutes)”. Traditional machine translation (MT) systems treat LCS data as noise, or just as regular sentences. However, if LCS data is processed intelligently, it can provide a useful signal for training word alignment and MT models. Moreover, LCS data is from non-news sources which can enhance the diversity of training data for MT. In this paper, we first extract constraints from this code switching data and then incorporate them into a word alignment model training procedure. We also show that by using the code switching data, we can jointly train a word alignment model and a language model using co- training. Our techniques for incorporating LCS data improve by 2.64 in BLEU score over a baseline MT system trained using only standard sentence-aligned corpora. 1 Introduction Many language users are competent in multiple languages, and they often use elements of multiple languages in conversations with other speakers with competence in the same set of languages. For example, native Mandarin speakers who also speak English might use English words in a Chinese sentence, like “\ • � ù ‡ ¯ K �solutioní º(Do you know the solution to this problem ?)”. This phenomenon of mixing ∗*The author is working at Raytheon BBN Technologies now languages within a single utterance is known as Linguistic Code Switching (LCS). Examples of these utterances are common in communities of speakers with a shared competency in multiple languages, such as Web forums for Chinese emigr´es to the United States. For example, more than 50% of the sentences we collected from a Web forum (MITBBS.com) contains both Chinese and English. Traditional word alignment models take a sentence-level aligned corpus as input and gener- ate word-level alignments for each pair of parallel sentences. Automatically-gathered LCS data typically contains no sentence-level alignments, but it still has some advantages for training word alignment models and machine translation (MT) systems which are worth exploring. First, because it contains multiple languages in the same sentence and still has a valid meaning, it will tell the relationship between the words from different languages to some extent. Second, most LCS data is formed during people’s daily conversation, and thus it contains a diversity of topics that people care about, such as home furnishings, cars, entertainment, etc, that may not show up in standard parallel corpora. Moreover, LCS data is easily accessible from Web communities, such as MITBBS.com, Sina Weibo, Twitter, etc. However, like most …

Content truncated to 500 words