真龍主・星音,豊島・孝之
自動翻訳に向けた対話理解の問題
Abstract:This paper provides an examination of some of the known current problems in
the Interpreting Telephony task Such an examination is necessary in order to
determine where natural language understanding efforts (i.e., plan recognition,
dialog understanding, etc.) should be focussed. Examining the problems also
provides explicit recognition of some of the assumptions that have been made in
the design of the automatic interpretation architecture; these assumptions can
then either be confirmed, or reevaluated and modified.
The results were obtained by analyzing ATR's "conversations 1-10" corpus,
which is a current research target for automatic translation efforts. Thus, the
results reflect the immediately relevant problems for matchine translation.
Although undoubtedly more problems will be discovered as the state-of-the-art
advances, this study has attempted to be a complete listing of the problems in the
current corpus.
The paper first reviews an assumed architecture for automatic interpretation.
Next, the known outstanding problems, as derived from the corpus, are presented
in order of descending frequency, and are also tabulated in the appendix. The
frequency is presented as a rough measure of the importance of the problem. The
paper concludes with a brief discussion.