Study on utilization and effectiveness of linguistic information in the word recognition based on phoneme, syllable or character sequence with errors

人们普遍认为，在基于音素、音节或字符识别的单词识别阶段，必须使用语言信息。然而，以下问题尚未得到很好的研究：1)什么样的音素、音节或字符识别错误会导致单词识别错误？2)当使用词典作为语言信息时，音素、音节或字符识别分数与单词识别分数和词汇量之间的关系是什么？3)当使用转移概率作为语言信息时，它们之间的关系是什么？为了回答这些问题，特别重要的是研究音素、音节或字符对在短距离内区分单词的特性，其中一对单词之间的距离是基于两个单词之间的音素、音节或字符对的差异。本研究以5317个日语流行词为研究对象，考察了音位、音节或字对在区分短距离词对时的特点。然后利用词典作为语言学家的信息，对存在错误的音素、音节或字符序列进行单词识别的性能评估。接下来，我们提出了以转移概率作为语言信息的短距离词对个数的快速计算算法。同样，我们在使用转移概率时也进行了性能评估。

It is widely recognized that linguistic information must be used in the stage of word recognition based on phoneme, syllable or character recognition. However,the following questions have not been well investigated:1)What kinds of errors in phoneme,syllable or character recognition cause errors in word recognition? 2)When using a word dictionary as linguistic information, what is the relationship between the relationship between the phoneme,syllable or character recognition score, the word recognition score, and the vocabulary size? 3)When using transition probability as linguistic information, what is the relationship between them? In order to answer these problems, ,.it is particularly important to investigate the properties of phoneme,syllable or character pairs in distinguishing words within a short distance, where the distance between a pair of words is based on the phoneme,syllable or character pair differences between the two words. In this research, we investigate the properties of phoneme, syllable or character pairs in distinguishing word pairs with short distance in the 5317 Japanese popular words. Then we carry out performance evaluation in word recognition based on phoneme, syllable or character sequence with errors when using a word dictionary as linguist information. Next, we propose high speed algorithm for computing number of word pairs with short distance when using transition probability as linguistic information. In a similar fashion, we carry out the performance evaluation when using the transition probability.