June 16th
I'm reading another ACL paper series. Disclaimer, all the review is based on my own opinion and not suppose to be referred outside this page or outside me !!!
1. Multi-Engine Machine Translation Guided By Explicit Word Matching, Shyamsundar Jayaraman, Alvon Lie, CMU ACL 2005.
This Multi-Engine MT combines the output of several machine translation engine and reconstruct the translation through several steps:
1. Word Alignment Matcher. It first sounds like another complicated statistical alignment model, but fortunately, it just only a word matcher. No parameter learning, no EM algorithm.. It's just try to construct alignment based on word matching.
2. Basic Hypothesis Generation. Here, we must recall ourselves that this paper is not about reranking multiple translation output, but about constructing the translation hypothesis from several translation output. This step can be summarized as the process to produce synthetic combinations of words and phrases from the original translations that satisfy a set of adequacy constraint. The notion of adequacy constraint refers the constraint that the words in the hypothesis adequately represent a translation.
3. Scoring. Yes, the final step is to attach a score to each hypothesis and pick the best one. In scoring the hypothesis, they are concern with the issue of translation of different length. To normalize the translation by calculating a geometric average per word score. They also exploit the confidence score of each translation out which is reflected on the confidence score. They also consider the language model score.
1. Multi-Engine Machine Translation Guided By Explicit Word Matching, Shyamsundar Jayaraman, Alvon Lie, CMU ACL 2005.
This Multi-Engine MT combines the output of several machine translation engine and reconstruct the translation through several steps:
1. Word Alignment Matcher. It first sounds like another complicated statistical alignment model, but fortunately, it just only a word matcher. No parameter learning, no EM algorithm.. It's just try to construct alignment based on word matching.
2. Basic Hypothesis Generation. Here, we must recall ourselves that this paper is not about reranking multiple translation output, but about constructing the translation hypothesis from several translation output. This step can be summarized as the process to produce synthetic combinations of words and phrases from the original translations that satisfy a set of adequacy constraint. The notion of adequacy constraint refers the constraint that the words in the hypothesis adequately represent a translation.
3. Scoring. Yes, the final step is to attach a score to each hypothesis and pick the best one. In scoring the hypothesis, they are concern with the issue of translation of different length. To normalize the translation by calculating a geometric average per word score. They also exploit the confidence score of each translation out which is reflected on the confidence score. They also consider the language model score.
posted by pahendra at 2:08 AM
0 Comments:
Post a Comment
<< Home