Studies on Speech Recognition, Closed Caption and Summarization of Broadcast News

众所周知，仅基于基本结构的HASH不能充分地捕捉连续帧之间的相关性。在我们以前的工作中，为了解决这一问题，分段单元的HALTORY被引入，并显示其有效性。同时，将Δ倒谱和ΔΔ倒谱结合到HSPs的分段单元中，也提高了识别性能。首先，我们比较了基于帧的模型和基于段的模型。结果表明，使用分段特征作为输入向量的有效性。其次，我们比较了基于音节和基于三音子的Herring。识别实验表明，基于音节的障碍语适用于日语。接下来，我们开发了一种使用任务适应策略和新闻文章的习惯表达来构建语言模型的方法。首先，我们使用有限数量的目标文章，研究了N-gram语言模型的任务适应方法的效果。其次，我们使用最新的文章研究了语言模型自适应方法的效果。第三，我们考察了习语作为语素单位使用的效果，因为一些特定的表达和习语经常出现在新闻文章中。最后，我们提出了一种对日语电视新闻文本中的每一个句子进行摘要的方法，并对该方法进行了评价。摘要新闻语篇的句子数量少，句子又长，不宜只选取重要的句子进行摘要。然后，我们试图减少多余的部分，其中包括修改器等，每一个句子。我们使用了一种专门用于新闻文本的简单解析方法，这样就不会破坏句法结构。通过对32名被试的问卷调查，对这种总结方法进行了评价。

It is well-known that HMMs only of the basic structure can not capture the correlation among successive frames adequately. In our previous work, to solve this problem, segmental unit HMMs were introduced and their effectiveness was shown. And the integration of Δ cepstrum and ΔΔ cepstrum into the segmental unit HMMs was also found to improve the recognition performance in the work. Firstly, we compared frame-based models and segment-based models. Results showed the effectiveness of the use of segmental features as input vectors. Secondly, we compared syllable-based HMMs and triphone-based HMMs. Recognition experiments showed that syllable-based HMMs are suitable for Japanese.Next, we developed a method that constructs language models using a task adaptation strategy and idiomatic expressions of news articles. First, we investigated the effect of a task adaptation method of N-gram language model using a limited amount of target articles. Second, we investigated the effect of the language model adaptation method using the latest articles. Third, we investigated the effect of the use of idiomatic expressions as morpheme units, since some specific expressions and idiomatic expressions are frequently observed in news articles. We showed that our proposed three methods were effective for constructing N-gram language models.Finally, we proposed and evaluated a method for summarizing each sentence in TV news texts written in Japanese. It is not appropriate to select important sentences for abstracting news text, because a news text consists of only a few and long sentences. Then, we tried to reduce redundant parts, which consisted of modifier etc., of each sentence. We used a simple parsing method specialized for news texts so that the syntactical structure was not destroyed. We evaluated this summarizing method by obtaining information by means of questionnaires to 32 examinees.