Multi-level Bootstrapping For Extracting Parallel Sentences From a Quasi-Comparable Corpus

Multi-level Bootstrapping for Extracting Parallel Sentences from a Quasi-Comparable Corpus Pascale Fung and Percy Cheung Human Language Technology Center, Department of Electrical & Electronic Engineering, HKUST, Clear Water Bay, Hong Kong {pascale,eepercy}@ee.ust.hk Abstract We propose a completely unsupervised method for mining parallel sentences from quasi-comparable bilingual texts which have very different sizes, and which include both in-topic and off-topic documents. We discuss and analyze different bilingual corpora with various levels of comparability. We propose that while better document matching leads to better parallel sentence extraction, better sentence matching also leads to better document matching. Based on this, we use multi-level bootstrapping to improve the alignments between documents, sentences, and bilingual word pairs, iteratively. Our method is the first method that does not rely on any supervised training data, such as a sentence-aligned corpus, or temporal information, such as the publishing date of a news article. It is validated by experimental results that show a 23% improvement over a method without multilevel bootstrapping. 1 Introduction Sentence-aligned parallel corpus is an important resource for empirical natural language tasks such as statistical machine translation and cross-lingual information retrieval. Recent work has shown that even parallel sentences extracted from comparable corpora helps improve machine translation qualities (Munteanu and Marcu, 2004). Many different methods have been previously proposed to mine parallel sentences from multilingual corpora. Many of these algorithms are described in detail in (Manning and Schűtze, 1999, Dale et al., 2000, Veronis 2001). The challenge of these tasks varies according to the degree of comparability of the input multilingual documents. Existing work extract parallel sentences from parallel, noisy parallel or comparable corpora based on the assumption that parallel sentences should be similar in sentence length, sentence order and bi- lexical context. In our work, we try to find parallel sentences from a quasi-comparable corpus, and we find that many of assumptions in previous work are no longer applicable in this case. Alternatively, we propose an effective, multi-level bootstrapping approach to accomplish this task (Figure 1). Figure1. Multi-level bootstrapping for parallel sentence extraction Extraction of matching bilingual segments from non-parallel data has remained a challenging task after almost a decade. Previously, the author and other researchers had suggested that bi-lexical information based on context can still be used to find correspondences between passages, sentences, or words, in non-parallel, comparable texts of the same topic (Fung and McKeown 1995, Rapp 1995, Grefenstette 1998, Fung and Lo 1998, Kikui 1999). More recent works on parallel sentence extraction from comparable data align documents first, before extracting sentences from the aligned documents (Munteanu and Marcu, 2002, Zhao and Vogel, 2002). Both work used a translation model trained from parallel corpus and adaptively extract more parallel sentences and bilingual lexicon in the comparable corpus. In Zhao and Vogel (2002), the comparable corpus consists of Chinese and English versions of new stories from the Xinhua News agency. Munteanu and Marcu (2002) used unaligned segments from the French-English Hansard corpus and finds parallel sentences among them. Zhao and Vogel (2002) used a …

Content truncated to 500 words