人工智慧面試官透過面試者之面部表情模擬多元評分者之人格特質評價之研究

Abstract

過往在人員甄選環節中，人格特質分數常被納入甄選標準內，過往常見評價方式分為兩種，應徵者之自我評價及面試官評價，而此兩種評價方法皆存在著評分偏誤的疑慮；有研究指出多元評分者（如：主管、同事、下屬或是客戶等）的分數在實務上更具意義，然而卻不適用於人員甄選環節。基於上述發現，本研究旨在建立一套能推估多元評分者（主管、同事、下屬與客戶）評價面試者之五大人格特質分數之人工智慧面試官。本研究邀請50位在職之業務人員和30位在職之業務基層主管擔任模擬應徵者，並蒐集其自我評價五大人格特質資料與面試影像資料，同時邀請前者之主管、同事及客戶擔任多元評分者，邀請後者之下屬及客戶擔任多元評分者。接著，透過自我評價與多元評分者評估之人格特質評分，與模擬應徵者之面部表情建立模型，模型將可透過面試者之面部表情推估自我評價與其多元評分者之人格特質評分。研究結果顯示，本研究所建之人工智慧面試官推估主管評價之準確度落在93.4% - 95.2%、同事評價之準確度落在94.4% - 95.2%、客戶評價之準確度落在93.6% - 95.9%、自我評價之準確度落在86.2% - 89.5%。During personnel selection, personality traits are usually included in the selection criteria. There are two common methods: self-report and the interviewers’ judgement. However, the abovementioned methods may suffer from biases. Previous studies have reported that judgements of multi-raters in the workplace, such as supervisors, peers, or customers, are more useful in practice. Unfortunately, this method is not applicable in personnel selection. Based on the aforementioned findings, the purpose of this study is to use an intelligence interview agent to infer personality trait ratings by multiple observers , including supervisors, peers, subordinates or customers.I collected the data of personality judgements from self-report and video records from 50 salespeople and 30 sales managers (interviewees). I also collected the data of personality judgements from the salespeople’s supervisors, peers, and customers, and from the managers’ subordinates and customers (multi-raters). The model that was established by video records and personality judgements from interviewees and multi-raters was used to detect the interviewees’ facial expressions to infer personality judgements of the interviewees and multi-raters. The results show accuracies between 93.4%– 95.2% in the supervisor model, 94.4%–95.2% in the peer model, 93.6%–95.9% in the customer model, 86.2% - 89.5% in the interviewee model.