Parallel Sentence Extraction from Comparable Corpora with Neural Network Features

Parallel Sentence Extraction from Comparable Corpora with Neural Network Features Chenhui Chu1, Raj Dabre2, Sadao Kurohashi2 1Japan Science and Technology Agency 2Graduate School of Informatics, Kyoto University E-mail: chu@pa.jst.jp, raj@nlp.ist.i.kyoto-u.ac.jp, kuro@i.kyoto-u.ac.jp Abstract Parallel corpora are crucial for machine translation (MT), however they are quite scarce for most language pairs and domains. As com- parable corpora are far more available, many studies have been conducted to extract parallel sentences from them for MT. In this paper, we exploit the neural network features acquired from neural MT for parallel sentence extraction. We observe significant improvements for both accuracy in sentence extraction and MT performance. Keywords: Parallel Sentence Extraction, Comparable Corpora, Neural Network 1. Introduction Neural machine translation (NMT) has achieved impres- sive results recently (Bahdanau et al., 2014). As the neural translation models can also be viewed as bilingual language models, they can be used to generate scores for candidate translations as neural network (NN) features for reranking the n-best lists produced by a statistical machine transla- tion (SMT) system, whose quality rivals the state of the art (Sutskever et al., 2014). Comparable corpora are a set of monolingual corpora that describe roughly the same topic in different languages. Al- though they are not exact translation equivalents of each other, there are a large amount of parallel sentences con- tained in the comparable texts. The task of parallel sen- tence extraction (Munteanu and Marcu, 2005) is to identify truly parallel sentences from the erroneous ones from com- parable corpora. Intuitively, because the NN features give a measure of the bilingual similarity of a sentence pair, they could be helpful for this task. However, this assumption has not been verified previously. In this paper, we incorporate the NN features into a ro- bust parallel sentence extraction system (Chu et al., 2014), which consists of a parallel sentence candidate filter and a binary classifier for parallel sentence identification. The NN features are naturally used as additional features for the classifier. Experiments conducted on Wikipedia data show that the NN features improve the strong baseline system significantly for both accuracy in sentence extraction and SMT performance. 2. Parallel Sentence Extraction System The overview of our parallel sentence extraction system is presented in Figure 1. We first align articles on the same topic in Wikipedia via the interlanguage links ((1) in Fig- ure 1). Then we generate all possible sentence pairs by the Cartesian product from the aligned articles, and discard the pairs that do not fulfill the conditions of a filter to reduce the candidates keeping more reliable sentences ((2) in Fig- ure 1). Next, we use a classifier trained on a small number of parallel sentences from a seed parallel corpus to identify the parallel sentences from the candidates ((3) in Figure 1). Finally, we train a NMT model on the extracted parallel sentences, and use it to obtain the NN features for the clas- sifier to further improve the performance ((4) in Figure 1). The strategy of the filter and the features used …

Content truncated to 500 words