محمدمهدي كامكار

This study focuses on the design, implementation, an‎d eva‎luation of an intelligent interview system for analyzing anger in job interviews of construction supervisors using natural language processing (NLP) an‎d automatic speech recognition (ASR) technologies. The objective was to develop an automated system capable of identifying an‎d eva‎luating supervisors’ emotional responses in hypothetical job-related scenarios in order to assist construction managers in selec‎ting can‎didates with effective anger management skills. In the first phase, in collaboration with psychology experts an‎d using the Delphi method, eight psychological questions were designed to elicit emotional reactions—particularly anger—from supervisors. Possible responses were collected through a questionnaire administered to 75 construction supervisors (54 men an‎d 21 women, aged 22–55) an‎d were labeled into three categories: positive (+1, appropriate anger management), negative (–1, angry reaction), an‎d neutral (0, indifference).For speech-to-text conversion, a hybrid system was implemented consisting of Google’s model (primary, online) an‎d the Whisper model (backup, offline). The Whisper model was fine-tuned using 1,407 sentences (1,125 training samples an‎d 282 validation samples; 80% construction-industry-related an‎d 20% general-domain sentences), reducing the word error rate (WER) from 45% to 17.63%. Subsequently, 18 interviews were conducted with construction supervisors using the system, along with 2 interviews generated via ChatGPT for comparative purposes. eva‎luation results based on precision, recall, an‎d accuracy metrics showed that speech-to-text accuracy reached 96% for Google an‎d 82% for Whisper. For response label classification, accuracy was 78% an‎d 73%, respectively. Semantic analysis of responses was performed using a multilingual transformer model, which assigned appropriate labels based on high semantic similarity. The final score for each interviewee was calculated as the algebraic sum of the labels across the eight questions to assess their anger management ability. The results indicate that the proposed system provides reliable performance compared to manual assessment methods (approximately 85–90% accuracy based on clinical detection studies) an‎d advanced models such as GPT-4 (80–87% accuracy in similar analyses). The primary innovation of this research lies in optimizing the system for the Persian language an‎d the construction industry, as well as employing the Whisper model in offline mode to reduce dependence on computational resources. The proposed system can serve as an effective tool for construction managers in the screening an‎d selec‎tion process of supervisors with appropriate emotional competencies, while also offering the potential for future expansion into multimodal (audio-visual) analysis.

پردازش زبان طبيعي , تشخيص خودكار گفتار , تحليل احساس

Natural Language Processing (NLP) , Automatic Speech Recognition (ASR) , Sentiment Analysis

Muhammad Mehdi Comcar

Mostafa Khanzadi