Abstract: To resolve the problem of voice command recognition errors existing in the speech interaction system for YSU-Ⅱ lower limb rehabilitation robot,a Bidirectional Gated Recurrent Unit (Bi-GRU) based Seq2Seq model was proposed to detect and correct the errors in instruction text.In addition,a Contextual and Keywords-based Attention (CK Attention) mechanism was proposed to enhance the performance of instruction text proofreading model.To objectively evaluate the performance of the model,a corpus for rehabilitation robot training tasks was established,and five 5-fold cross-validation method was employed to conduct a series of experiments on the corpus.The experimental results demonstrated that the Bi-GRU based Seq2Seq model was applicable for the instruction text proofreading task,and the CK Attention mechanism contributed to improve the performance of the text proofreading model.The detection F1 and the correction F1 of the proposed model had reached 97.72% and 93.89% respectively.The processing time of the instruction text proofreading model for common instructions was 0.156 s ~0.391 s.

Key words: text proofreading, speech interaction, Seq2Seq, bidirectional gated recurrent unit, attention mechanism

摘要： 针对YSU-Ⅱ下肢康复机器人语音交互系统存在指令误识的问题,构建了基于双向门控循环单元的Seq2Seq模型来检测并纠正指令文本中的错误字符,提出一种基于指令上下文和关键字的注意力机制(CK Attention),用于捕获指令文本的上下文语义和关键字信息,以提升模型的文本校对能力。面向康复机器人的训练任务自行建立了语料库,并采用5次5折交叉验证法在该语料库上开展了相关实验,以客观评估模型性能。实验结果表明,所建模型适用于指令文本校对任务,CK Attention机制能够有效提升模型的文本校对性能,其检错F1值和纠错F1值分别达到97.72%和93.89%,对常用指令文本的校对时长在0.156 s~0.391 s之间。

关键词: 文本校对, 语音交互, Seq2Seq, 双向门控循环单元, 注意力机制