LEXICAL CHAIN BASED DOCUMENT-LEVEL MACHINE TRANSLATION

Abstract

Most statistical machine translation (SMT) systems translate each source sentence independently without considering document-level context. As a result, it often leads to inconsistent or inappropriate word selections throughout the translation of the entire document. In the meanwhile, considering larger context often suffers from the high computational complexity and noisy information. We argue that an efficient and wise extraction of related information can help determine the choice of words in the target language. Inspired by previous approaches in document summarization, we propose to leverage lexical chains as skeletons of a document to identify related words and perform translation under the constraint of generated target lexical chains. Experimental results have shown that the proposed models based on lexical chains can substantially improve the translation quality in terms of BLEU on large-scale datasets.