Consideration on Syntactic Analyses for a Speech Understanding System.
Seiichi Nakagawa, Yoshihisa Ohguro · Institutional Repositories DataBase (IRDB) · 1988
In this paper, we compared the left-to-right & top-down parsing strategy with the island-driven & botom-up strategy by using a simulated phoneme recognizer. The both strategies adopted the beam search. The syntactic constraints were represented by a context-free grammar. The word lattice for an utterrance was generated by a word spotting algorithm from an ambiguous phoneme sequence. These two parsing strategies were used for finding the best word sequence from the word lattice. We evaluated them through three different tasks (small, combined and large). From the parsing results by the two strategies, we can conclude in the following : (1) the recognition accuracy was almost the same for both strategies, (2) the left-to-right & top-down parsing strategy was superior to the island-driven & bottom-up strategy in terms of the processing time. And also we found that the sentence recognition rate was not so sensitive on the vocabulary size, or on the combined task in comparison with a single task. However, when the initial part of an utterance was noisy, the island-driven strategy became superior to the left-to-right strategy. Finally, we found that our proposed parsing machanism for missing function words was effective.