Towards a Combination of Metrics for Machine Translation

Abstract

In this scholar, we compare three metrics for machine translation, from English toFrench and vice versa, and we give some combination formulas based on someschemes, algorithms, and machine learning tools. As an experimental dataset, weconsider 10 English and French theses abstracts published in the web with four free incharge machine translation systems. Five combinations, with the same implicit weights,are considered namely: (BLEU+NIST), (BLEU+ (1-WER)), (NIST+(1-WER)),(BLEU+NIST+(1-WER)), and (FR(BLEU)+FR(NIST)+FR(WER)). Thesecombinations are also considered differently through generating weights parameters onthe basis of regression. The results of 12 formulas are computed and compared then intotal. According to the obtained results, average regression combinations based onmachine learning step are the best, especially with the three basic metrics, followed byaverage WER metric in the case of English to French. For French to English,(FR(BLEU)+FR(NIST)+FR(WER)) combination is the best followed respectively bythe average regression combination with both first parameters (Reg(α,β)) and averageBLEU basic metric. Another performance criterion is considered here, in the secondposition, namely: the number of times, over the 10 abstracts, where the formula is thebest. Based on the obtained results, combination with regression based on the first andthe last parameters (Reg(α,γ)) outperforms the others, in the case of English to French,with 3 times followed by Reg(β,γ), Reg(α,β,γ), NIST+(1-WER), and the basic metrics(BLEU, NIST, and WER) with 2 times for each of them. For French to English, thebasic WER metric outperforms the others with three times followed by BLEU, (BLEU+(1-WER)), (FR(BLEU)+FR(NIST)+FR(WER)), and Reg(α,γ) with 2 times for each ofthem. To note that there is a room of improvement for the combinations with1.0914 inthe case of English to French and 1.01 in the case of French to English.