Neural Language Models Evaluate Human Performance: The Role of Language and Prosody in Predicting Job Interview Scores

Anonymous

Abstract: In this work we test the use of state-of-the-art neural language model representations to predict behavioral traits that cannot be easily extracted from the textual input alone. We take the task of automated job interview scoring and make predictions on behavioral traits such as hirability, engagement, or friendliness. We find that representing text using neural models trained only on text already leads to better overall prediction results compared to a feature engineering approach that uses a combination of linguistic and extra-linguistic materials. Moreover, we show that combining word embeddings and prosodic features improves the results even further, highlighting the value of adding information from modalities other than text when evaluating human performance.