基于上下文信息的口语意图检测方法

近年来,随着人工智能的发展与智能设备的普及,人机智能对话技术得到了广泛的关注。口语语义理解是口语对话系统中的一项重要任务,而口语意图检测是口语语义理解中的关键环节。由于多轮对话中存在语义缺失、框架表示以及意图转换等复杂的语言现象,因此面向多轮对话的意图检测任务十分具有挑战性。为了解决上述难题,文中提出了基于门控机制的信息共享网络,充分利用了多轮对话中的上下文信息来提升检测性能。具体而言,首先结合字音特征构建当前轮文本和上下文文本的初始表示,以减小语音识别错误对语义表示的影响;其次,使用基于层级化注意力机制的语义编码器得到当前轮和上下文文本的深层语义表示,包含由字到句再到多轮文本的多级语义信息;最后,通过在多任务学习框架中引入门控机制来构建基于门控机制的信息共享网络,使用上下文语义信息辅助当前轮文本的意图检测。实验结果表明,所提方法能够高效地利用上下文信息来提升口语意图检测效果,在全国知识图谱与语义计算大会(CCKS2018)技术评测任务2的数据集上达到了88.1%的准确率(Acc值)和88.0%的综合正确率(F1值),相比于已有的方法显著提升了性能。

Intention Detection in Spoken Language Based on Context Information

In recent years,with the development of artificial intelligence and the popularization of smart devices,human-computer intelligent dialogue technology has received extensive attention.Spoken language understanding is an important task dialogue system,and spoken language intention detection is a key technology in spoken language understanding.Due to complex language phenomena such as semantic missing,frame representation and intent conversion in multiple rounds of dialogue,the intent detection task for spoken language is very challenging.In order to solve the above problems,a gated mechanism based information sharing neural network method was proposed in this paper,which can take advantages of contextual information in multiple rounds of dialogue to improve detection performance.Specifically,first the current round text and context text initial representation are constructed in combination with the phonetic features to reduce the impact of speech recognition errors on semantic representation.Secondly,a semantic encoder based on hierarchical attention mechanism is used to obtain deep semantic representations of the current round and contextual text,including multi-level semantic information from word to sentence to multiple rounds of text.Finally,the gated mechaniam based information sharing neural network is constructed to use the context semantic information to help the intent detection of the current round of text.The experimental results show that the proposed method can effectively use context information to improve the detection of spoken language intentions,and achieves 88.1%accuracy and 88.0%F1 value in dataset of CCKS2018 shared task-2,which is significantly improved performance compared with the existing methods.