SGER: Experiments on Integrating Speech Recognition and Natural Language Processing

***本项目解决了语音识别和自然语言处理组件在语音理解系统中的集成。目前大多数语音识别器都将词级隐马尔可夫模型（hmm）与n-gram语言模型结合使用。通过使用手机作为基本识别单元，可以实现额外的灵活性。这项研究提出了一种新的技术，称为SOHMM随机观察隐马尔可夫模型-用于识别来自电话候选人的单词。目前大多数口语系统使用概率语言模型来模拟句法和语义模式。该项目将集成基于SOHMM的语音识别器和约束依赖语法（CDG）解析器，该解析器支持在统一的模块化框架中使用词汇、韵律、句法和语义知识来源，以及使用特定领域的约束。本研究将完善SOHMM-CDG口语处理系统的模型，并解决构建和修剪词图的问题，作为两个组件之间的接口。本文将通过实验验证该集成系统对TIMIT和Resource Management语料库中话语的识别和解析的有效性。与仅使用统计语言模型相比，句子的准确性应该得到显著提高，尤其是在使用特定领域的约束时。

*** This project addresses the integration of speech recognition and natural language processing components in a speech understanding system. Most current speech recognizers use word-level hidden Markov models (HMMs)in conjunction with n-gram language models. Additional flexibility can be achieved by using phones as the basic recognition unit. This research proposes a new technique called a SOHMM Stochastic Observation Hidden Markov Model - for recognizing words from phone candidates. Most current spoken language systems use probabilistic language models to model syntactic and semantic patterns. This project will integrate a SOHMM- based speech recognizer with a Constraint-Dependency Grammar (CDG) parser, which supports the use of lexical, prosodic, syntactic, and semantic knowledge sources in a uniform modular framework, as well as the use of domain- specific constraints. This work will refine the model for the SOHMM-CDG spoken language processing system and address the problem of building and pruning word graphs, which act as the interface between the two components. An experiment will be conducted to demonstrate the effectiveness of the integrated system for recognizing and parsing utterances in the TIMIT and Resource Management corpora. Sentence accuracy should be significantly improved over what is possible using only statistical language models, especially when domain-specific constraints are used ***.