Automatic identification of storytelling responses to past-behavior interview questions via machine learning

Abstract

Structured interviews often feature past-behavior questions, where applicants are asked to tell a story about past work experience. Applicants often experience difficulties producing such stories. Automatic analyses of applicant behavior in responding to past-behavior questions may constitute a basis for delivering feedback and thus helping them improve their performance. We used machine learning algorithms to predict storytelling in transcribed speech of participants responding to past-behavior questions in a simulated selection interview. Responses were coded as to whether they featured a story or not. For each story, utterances were also manually coded as to whether they described the situation, the task/action performed, or results obtained. The algorithms predicted whether a response features a story or not (best accuracy: 78%), as well as the count of situation, task/action, and response utterances. These findings contribute to better automatic identification of verbal responses to past-behavior questions and may support automatic provision of feedback to applicants about their interview performance.