This paper addresses the challenging problem of automating the human's intelligent ability to evaluate output from machine translation (MT) systems, which are subsystems of Speech-to-Speech MT (SSMT) systems. Conventional automatic MT evaluation methods include BLEU, which MT researchers have frequently used. BLEU is unsuitable for SSMT evaluation for two reasons. First, BLEU assesses errors lightly at the beginning or ending of translations and heavily in the middle, although the assessments should be independent from the positions. Second, BLEU lacks tolerance in accepting colloquial sentences with small errors, although such errors do not prevent us from continuing conversation. In this paper, the authors report a new evaluation method called RED that automatically grades each MT output by using a decision tree (DT). The DT is learned from training examples that are encoded by using multiple edit distances and their grades. The multiple edit distances are normal edit dista nce (ED) defined by insertion, deletion, and replacement, as well as extensions of ED. The use of multiple edit distances allows more tolerance than either ED or BLEU. Each evaluated MT output is assigned a grade by using the DT. RED and BLEU were compared for the task of evaluating SSMT systems, which have various performances, on a spoken language corpus, ATR's Basic Travel Expression Corpus (BTEC). Experimental results showed that RED significantly outperformed BLEU.
展开▼
