Abstract: The Chinese structure is flexible and diverse,and the same semantics can be expressed in multiple ways.When the Chinese instruction is too long or there are multiple sets of actions,it is difficult for the home service robot to understand and respond correctly.In addition,research on Chinese instruction parsing in home settings is still scarce in terms of data.Existing methods cannot accurately extract the entire sequence of actions contained in an instruction.An Action Sequence Extraction model based on ERNIE (ASE-ERNIE) method to extract executable action sequences in Chinese instructions was proposed.The model extracted instruction information through the relationship between action types and parameters,transforming the extraction of action sequences into the sequence annotation tasks.When there were multiple groups of actions and multiple parameters in the instruction,the whole action sequence could still be extracted.The effectiveness of the proposed method was verified by experiments,and the F1 value of the action sequence extraction task reached 80.37%.

Key words: instruction parsing, action sequence, ERNIE method, sequence annotation

摘要： 中文结构灵活多样,同样的语义有多种表达方式,在中文指令过长或存在多组动作的情况下,家庭服务机器人很难正确理解并做出响应。此外,家庭环境的中文指令解析研究在数据方面仍然稀缺。现有的方法不能准确提取出指令中包含的全部动作序列。由此提出一种基于ERNIE方法的动作序列提取(ASE-ERNIE)模型来提取中文指令中的可执行动作序列,该模型通过动作类型和参数之间的关系提取指令信息,将动作序列的提取工作转化为序列标注任务。在指令中存在多组动作并且存在多个参数的情况下,依然能够提取出全部动作序列。通过实验验证了所提方法的有效性,动作序列提取任务的F1值达到了80.37%。

关键词: 指令解析, 动作序列, ERNIE方法, 序列标注