〈Title〉
   A MT system evaluation method 
<Authors〉
     Nobuyoshi Terashima
     Satoru Ikehara
     Masahiro Miyazaki
   Natural Language Processing Laboratory, NTT Communications and Information   Processing Laboratories, Nippon Telegraph and Telephone Corporation.
〈Key Words〉
   Machine Translation, Evaluation method, Japanese Newspaper, Japanese to 
English Translation.
〈Outline〉
   Many technological problems have to be solved in order to develop
a Japanese-English translation system capable of translating source text without
pre-editing. To clarify these problems, an evaluation method has to be
developed. However no evaluation method has yet been developed, which
estimates translation quality based on translation technologies and source
text characteristics.
   This paper propose an evaluation method for machine translation quality. 
In the first place, the translation process is divided into sentence analysis
phase and the conversion and generation phase. The capability of the
former phase is represented by the capability of the morphemic and dependency
analysis phases , and that of the latter phase is derived from the capability
of English expression generation for several characteristic expressions in
Japanese. An evaluation method is proposed from this point of view and is used
to evaluate our MT system and also to identify the problems to develop the 
non-pre-editing translation system.
   Specifically, 1,000 sentences of lead texts (which are typical introductoryremarks summarizing a newspaper article immediately following the headlines)were taken from the Nikkei Sangyo Shinbun. Sentence structure and
characteristics in expression that would pose problems in machine translation 
were studied. Regarding the characteristics of sentence structuring, formal
characteristics of typical newspaper articles, such as length of articles,
number of "bunsetsu" (which are Japanese phrases, in most cases, consisting of
a noun and "joshi" or post-positional word) and other such statistical 
features are studied together with the ratios of types of sentences such as 
simple and complex sentences. As characteristics of expression, we consider 9 types of expression may pose problems in formulation of algorithms and
investigate the frequency of their appearance in newspaper articles.
   Then, these characteristics of newspaper articles and quality data for eachindividual technology are applied to the evaluation method. Estimates of final
translation quality are made and compared with actual measured values for
confirming accuracy of this evaluation method and examining technical problems
which remain for solution in the future.
   The outline of the results is as follows.
   Assuming 70% acceptability to be a practical level of quality, the number
of words and bunsetsu indicate that the morphemic and dependency analysis
technologies would require accuracy rates per word and per bunsetsu as high as
99.8% and 99.4%, respectively.  Current technology has already reached this
level. But the precision of translation for individual expressions is lower
than expected values and acceptability rates for the entire sentence drops to
40% to 50%. To achieve an acceptability rate of 70%, it is necessary to achievean average 10% improvement in all 9 types of expressions mentioned in this
study and thus achieve a translation ratio of 96%.
   To upgrade the translation rate of these expression, it will be necessary
to establish techniques for meaning analysis of post-positional words,
translation of expressions that combine noun clauses and compounded words,
handling connections and embedded clauses, and verification of ellipsis
and paratactical construction.