Tourist Areas In Question Speech Recognition Language Model Rules To Automatically Build Research

As a significant part of Speech Recognition System, the language model can provide all kinds of linguistic knowledge for the speech recognition. And the structure of language model will directly influence the integrated quality of the system as well as accuracy rates of the recognition. Compared with statistical language model, rule-based language models are much more widely used in some special Speech Recognition Systems. Currently, rule-based language models need to be manually structured under the specialist’s participation. But, owning to the inconstant grammatical rule and complicated words and phrases in Chinese, it is difficult to cover all the linguistic conditions only by the specialist’s knowledge. Most seriously, structuring manually not only waste time and energy, but also has higher error rates. So, this paper focuses on the automatic acquisition of basic elements which constitutes the rule-based language model, including rules of sentence pattern as well as words and phrases. And it makes some achievements from following aspects.1. The Automatic Acquisition of Sentence Pattern Rules in rule-based language model. For the problems of flexibility and coverage in sentence pattern rules, this paper uses an automatic acquisition and statistical method based on the sentence pattern tree to resolve. First, sentence pattern trees in the Treebank are selected to be compared with each other by the calculation, and the result of their similarities will constitute the candidates of structure gatherings. Second, the frequency of each structure appeared in the Treebank will be counted and higher frequency ones continue to constitute the candidates of sentence pattern gatherings. At last, all sentence patterns in those candidate gatherings will be calculated and distributed different weights. Those owning a higher value than the threshold are sentences satisfying the requirement. According to the experiment, this method will improve the flexibility and coverage of sentence pattern rules in the rule-based language model.2. The Automatic Acquisition of Words and Phrases in rule-based language model. For the low coverage and accuracy rates of words and phrases, this paper uses a method combining statistics and rules to get the recognition and acquisition of words and phrases. First, the MI value of adjacent words property will be used to predict the boundary of phrases. And then, the summarized phrases boundary information and the inner constituting rule will be used to made adjustments to the boundary and get words as well as phrases. Finally, it has been tested that this method will improve the accuracy rate and coverage of phrases in the rule-based language model.3. Based on the previous work, a rule-based language model about tourism consultant will be structured and the new language model will be added into ASR. Finally, the new Speech Recognition System will be given analysis and evaluation.