We argue that the machine translation community is overly reliant on the Bleu machine translation evaluation metric. We show that an improved Bleu score is neither necessary nor sufficient for achieving an actual improvement in translation quality, and give two significant counterexamples to Bleu’s correlation with human judgments of quality. This offers new potential for research which was previously deemed unpromising by an inability to improve upon Bleu scores. 1

Evaluation results recently reported by Callison-Burch et al. (2006) and Koehn and Monz (2006), revealed that, in certain cases, the BLEU metric may not be a reliable MT quality indicator. This happens, for instance, when the systems under evaluation are based on different paradigms, and therefore, do not share the same lexicon. The reason is that, while MT quality aspects are diverse, BLEU limits its scope to the lexical dimension. In this work, we suggest using metrics which take into account linguistic features at more abstract levels. We provide experimental results showing that metrics based on deeper linguistic information (syntactic/shallow-semantic) are able to produce more reliable system rankings than metrics based on lexical matching alone, specially when the systems under evaluation are of a different nature. 1

A number of approaches to Automatic MT Evaluation based on deep linguistic knowledge have been suggested. However, n-gram based metrics are still today the dominant approach. The main reason is that the advantages of employing deeper linguistic information have not been clarified yet. In this work, we propose a novel approach for meta-evaluation of MT evaluation metrics, since correlation cofficient against human judges do not reveal details about the advantages and disadvantages of particular metrics. We then use this approach to investigate the benefits of introducing linguistic features into evaluation metrics. Overall, our experiments show that (i) both lexical and linguistic metrics present complementary advantages and (ii) combining both kinds of metrics yields the most robust metaevaluation performance. 1