This paper describes a small-scale but organized attempt to evaluate output quality of several Japanese MT systems. The project also served as the first experiment of the implementation of the in-house MT evaluation guidelines created in 2000. Since time was limited and the budget was not infinite, it was launched with the following compact components: Five people; 300 source sentences per language pair; and 160 hours per evaluator. The quantitative results showed noteworthy phenomena. Although the test materials had been presented in a way that evaluators could not identify the performance of any particular system, the results were quite consistent. The scoring ratio that the two E-to-J evaluators employed was almost identical, while that of the J-to-E evaluators was similar. This indicates that high-quality output has universal appeal. Additionally, the evaluators noted that stronger systems, regardless of language pair, tended to be superior in source sentence analysis, target sentence arrangement, word choice, and lexicon entries whereas weaker systems tended to be inferior in these areas. As for language-pair comparison, the results indicate that English-to-Japanese systems may require more improvement than their counterparts, judging from the scores given and the number of unfound words recorded.

The reliable detection of sentence boundaries in running text is one of the first important steps in preparing an input document for translation. Although this is often neglected, it is necessary to obtain a translation with a high degree of quality. In this paper, we present a comparison of different paradigms for the detection of sentence boundaries in written text. We compare three different approaches: Directly encoding the knowledge in a program, a rule-based system relying on regular expressions to describe boundaries, and a statistical maximum-entropy learning algorithm to obtain knowledge about boundaries. Using the statistical system, we obtain a recall of 98.14%, classifying boundaries of six types, and using a training corpus of under 10,000 sentences.