基于关键词注意力的细粒度面试评价方法

海量的在线面试视频数据为智能面试评价提供了重要的数据基础.随着目前全球疫情的蔓延，网络在线面试的需求程度上升，对智能面试评价工具的需求也随之上升.结构化面试中，面试官需要依据评价标准，观察面试者所做的回答，并形成面试者人格特性、沟通技能以及领导力等方面的画像评估，以此判断面试者的特质是否与应聘职位相匹配.其中人格特性评估是公司间广泛使用的一种评估方法，因为人格特性影响着人们的语言表达、人际交往等多个方面，是辅助面试官决策该面试者是否符合其应聘岗位需求的重要参考.基于此，提出了基于循环神经网络长短期记忆(long short term memory, LSTM)以及关键词-问题注意力机制的多层次(hierarchical keyword-question attention LSTM, HKQA-LSTM)细粒度面试评价方法，旨在针对面试者的不同人格特性维度进行打分，并据此得到综合面试得分.首先，通过引入关键词注意力机制有效筛选出面试对话中与人格特性密切相关的重要词句；然后，在此基础上采用了关键词-问题层次注意力机制和2阶段的模型学习机制，充分结合面试者表述文本的多尺度上下文特征，对人格特性进行准确预测；最后通过融合人格特性得到具有较高解释性的面试综合评价结果.基于真实面试场景数据的实验结果表明，该方法能有效地评价面试者的不同人格特性得分，并准确地预测面试者总体得分.

Fine-Grained Interview Evaluation Method Based on Keyword Attention

Massive online interview video data provides an important data basis for intelligent interview evaluation. With the spread of the current global epidemic, the demand for online interviews has increased, as well as the intelligent interview evaluation tools. In a structured interview, the interviewer needs to observe the interviewee’s answers based on the evaluation criteria, and form a profile evaluation of the interviewee’s personality traits, communication skills, and leadership, so as to judge whether the interviewee’s characteristics match the position. Among them, personality evaluation is a widely accepted evaluation method among companies. Because personality traits affect people’s language expression, interpersonal communication and other aspects, it is an important reference to assist the interviewer to decide whether an interviewee meets their job requirements. Based on this, a fine-grained interview evaluation method based on the long short term memory (LSTM) and the hierarchical keyword-question attention mechanism (HKQA-LSTM) is proposed, which aims to score the different personality dimensions of the interviewees and obtain a comprehensive interview score based on this. First, we effectively filter out important words and sentences that are closely related to personality traits in the interview dialogue by introducing a keyword attention mechanism. Then, we use keyword-question level attention mechanism and two-stage model learning mechanism on this basis, and fully combine the multi-scale contextual features of the texts expressed by interviewees to accurately predict personality traits. Finally, through the fusion of personality traits, a comprehensive interpretive evaluation result of the interview is obtained. The experimental results based on real interview scene data show that this method can effectively evaluate the interviewees’ different personality traits scores and accurately predict the interviewees’ overall scores.