基于语境感知和迁移学习的口语理解方法

Oral comprehension method based on context perception and transfer learning

口语理解是语言处理中的重要组成部分。在语言翻译中,口语理解可以将自然语言用户的话语解析为语义框架。为了提高口语理解的性能,在多回合对话中可以加入上下文信息。但是收集大规模的人工标记的多轮对话语料库是复杂,昂贵的,且不现实的。所以在此提出了一个上下文编码语言转换器(CLT)模型,以便于利用各种上下文信息。实验结果表明,在两个大规模的单轮对话基准测试和一个大规模多回合对话测试中,该模型与现有的转移学习方法相比,性能提高显著。

Oral comprehension is an important part of language processing.In language translation,oral comprehension can parse the utterances of natural language users into semantic frames.In order to improve the performance of oral comprehension,contextual information can be added to the multi-round dialogue.However,it is complicated,expensive,and unrealistic to collect large-scale,manually marked multi-round dialogue corpus.Therefore,a contextual coding language converter(CLT)model is proposed here to facilitate the use of various contextual information.The experimental results show that in two large-scale single-round dialogue benchmark tests and a large-scale multi-round dialogue test,the performance of this model is significantly improved compared with the existing transfer learning methods.