Automatic sentence partitioning of TV news sentences for closed caption service to hearing impaired people

摘要

The authors are investigating automatic summarization and automatic captioning technologies to help TV program production with closed-captions for hearing impaired people. Japanese TV news articles include rather long sentences than news-paper articles. Because long sentences are difficult to read on closed-captions, automatic partitioning of long sentences is effective in both automatic summarization and automatic captioning. The paper shows a automatic partitioning system of long sentences in TV news articles and evaluation results of the partitioning. The system is composed of three phases. First of all, the system partitions direct quotation parts which are quoted by "「" and "」". The second phase partitions several sentences separated by "。 " in direct quotations. The third phase partitions composed sentences at each clauses using partitioning rules. After the third phase, to keep the coherency of each partitioned sentences, the system inserts appropriate conjunctions at the partitioning points and adjusts politeness, tense and aspect for each sentences. Using 9205 sentences from TV news articles in 1991 which are longer than 100 characters, experiments shows that mean sentence length are from 133 characters to 50 characters