| Summary: | 碩士 === 國立雲林科技大學 === 資訊工程研究所 === 95 === In this study, we use a combination of lexical and statistical method to do sentence alignment for an English-Chinese corpus. Past research reveals that alignment using a dictionary involves a lot of word matching and dictionary look ups. To address these two issues, we first restrict the range of candidate target sentences, based on the location of the source sentence relative to the beginning of the text. Moreover, careful empirical selection of stop words, based on word frequencies in the source text, helps to reduce the number of dictionary look ups. Experimental results show that the amount of word matching can be cut down by 75% and that of dictionary look ups by as much as 43% without sacrificing precision and recall. Another experiment was also done with twenty New York Times articles with 598 sentences and 18395 words. The resulted precision is 95.6% and the recall is 93.8%. Among all predicted alignment, 86% of the alignment is 1:1 (one source sentence to one target sentence), 8% is 1:2, and 6% is 2:1. Further analysis shows that most errors occur in alignments of types 1:2 and 2:1. Future work should focus on problems with these two alignment types. A language learning system is also developed to train learners in reading and writing English. The system provides two types of exercises. In the first exercise, learners are asked to do sentence alignment when given an English article and its Chinese translation. In the second exercise, learners are asked to produce an English sentence when given both its Chinesetranslation and the English translation of each word in the Chinese sentence. The system provides some scaffolding to help learners when they encounter difficulties. Experimental results show that the learners enjoyed doing the exercises and felt that they learned some useful skills from the exercises.
|
|---|
