Paraphrasing Depending on Bilingual Context Toward Generalization of Translation Knowledge

Paraphrasing depending on Bilingual Context Toward Generalization of Translation Knowledge Young-Sook Hwang ETRI 161, Yuseong-gu, Daejeon 305-700, KOREA yshwang7@etri.re.kr YoungKil Kim ETRI 161, Yuseong-gu, Daejeon 305-700, KOREA kimyk@etri.re.kr SangKyu Park ETRI 161, Yuseong-gu, Daejeon 305-700, KOREA parksk@etri.re.kr Abstract This study presents a method to automat- ically acquire paraphrases using bilingual corpora, which utilizes the bilingual de- pendency relations obtained by projecting a monolingual dependency parse onto the other language sentence based on statisti- cal alignment techniques. Since the para- phrasing method is capable of clearly disam- biguating the sense of an original phrase us- ing the bilingual context of dependency re- lation, it would be possible to obtain inter- changeable paraphrases under a given con- text. Also, we provide an advanced method to acquire generalized translation knowl- edge using the extracted paraphrases. We applied the method to acquire the gener- alized translation knowledge for Korean- English translation. Through experiments with parallel corpora of a Korean and En- glish language pairs, we show that our para- phrasing method effectively extracts para- phrases with high precision, 94.3% and 84.6% respectively for Korean and English, and the translation knowledge extracted from the bilingual corpora could be general- ized successfully using the paraphrases with the 12.5% compression ratio. 1 Introduction Approaches based on bilingual corpora are promis- ing for the automatic acquisition of translation knowledge. Phrase-based SMT(Statistical Machine Translation) models have advanced the state of the art in machine translation by expanding the basic unit of translation from words to phrases, which al- lows the local reordering of words and translation of multi-word expressions(Chiang, 2007) (Koehn et al., 2003) (Och and Ney, 2004). However phrase-based SMT techniques suffer from data sparseness problems, that is; unreliable translation probabilities of low frequency phrases and low coverage in that many phrases encountered at run-time are not observed in the training data. An alternative for these problems is to utilize para- phrases. An unknown phrase can be replaced with its paraphrase that is already known. Moreover, we can smooth the phrase translation probability using the class of paraphrases. On the other hand, EBMT or PBMT systems might translate a given sentence fast and robustly geared by sentence translation patterns or general- ized transfer rules. Since it costs too much to con- struct the translation knowledge, they suffer from the problem of knowledge acquisition bottleneck. In this study, we present a method of automat- ically extracting paraphrases from bilingual cor- pora. Furthermore, we introduce a new method for acquiring the generalized translation knowledge. The translation knowledge is a kind of verb sub- categorization pattern composed of bilingual depen- dency relations. We obtain the generalized trans- lation knowledge by grouping the equivalent con- stituent phrases. The task of identifying the phrases equivalent to each other is defined as paraphrasing. Our paraphrasing method utilizes bilingual cor- pora and alignment techniques in SMT. Unlike pre- 327 vious approaches which identify paraphrases using a phrase in another language as a pivot without con- text information (Bannard et al., 2005), or apply the distributional …

Content truncated to 500 words