Reducing Phrase-based SMT Translation Tables by Phrase Pair Coverage Rate

[[abstract]]傳統以詞組為本之統計式機器翻譯 (Phrase-based SMT) 在產生詞組翻譯對照表 (Phrase Translation Table) 前，會先做詞彙對應 (word alignment)，然後再以直覺 (heuristic) 的方式去找出可能的詞組對應 (phrase alignment)，以產生 Phrase Translation Table。但是這樣的 phrase alignment 缺乏客觀的詞組分斷 (phrase segmentation) 標準，所以有可能產生大量且雜亂的詞組。 為此，本文提出一個依詞組對涵蓋率來縮減詞組為本之統計式機器翻譯之雙語對照表的方法來改良 Phrase-based SMT。本方法先運用 EM 演算法 (EM Algorithm) 找出語料庫 (Corpus) 最佳的 phrase segmentation，然後在進行 Moses 的翻譯模型訓練 (TM Training) 這步驟時，再利用phrase segmentation 的資訊計算詞組對涵蓋率以縮減 Phrase Translation Table 的大小。過濾完之後的 Phrase Translation Table 即與語言模型 (LM Model) 合併用於解碼器 (Decoding) 進行翻譯。最後用 BLEU 來對翻譯過的文章作翻譯的評分，並且與 Moses 所建構的翻譯系統作效能的評估比較。 實驗的結果顯示，Phrase Translation Table 的大小可減少大約60~70%左右，而且 BLEU的分數可以非常接近 Moses 的分數。[[abstract]]Before generating the phrase translation tables in traditional phrase-based statistical machine translation (PB SMT), word alignment will be conducted first. Heuristics are then used to find the possible phrase pairs, thus producing the Phrase Translation Tables. Since the phrases and phrase pairs are induced from word alignment without phrase segmentation criteria, it is possible to produce a large number of messy phrases. In this paper, we propose a “Phrase Pair Coverage Rate” measure to help reduce the Phrase Translation Table. We firstly use an EM algorithm to find the best phrase segmentation of the source and target sentences, later in the Translation Model training (TM training) step of Moses, the phrase segmentation information is used, based on the “Phrase Pair Coverage Rate” to reduce the size of the Phrase Translation Table. The reduced Phrase Translation Table is then used with the Language Model (LM) to decode (i.e., to translate) the source sentences. Finally, the BLEU score is estimated to evaluate the translation quality, and comparison is made against the Moses SMT system. The experimental results show that phrase table size can be reduced by about 60~70%, and the BLEU score is close to Moses performance.[[note]]碩