Abstract:
Automatic personality detection from text has gained interest since researchers discovered that linguistic style can be an indicator of personality. However, accurate personality classification remains a challenging task, often lacking data and robust evaluation metrics. This thesis investigates the ability of various machine learning models to predict the Big Five personality traits from text. We evaluate our models using two datasets. The first is the existing Stream of Consciousness Essays (SoCE) dataset, containing essays written by college students about their thoughts. The second is our newly collected Behavioral Interview Data (BID), featuring an annotated corpus tailored for this research. This new dataset includes university students' responses to behavioral questions similar to those asked in job interviews. In our experiments with both datasets, we explore different Natural Language Processing (NLP) techniques, focusing particularly on the Generative Pre-trained Transformer (GPT), using various parameters and testing methods. We compare GPT’s performance with a wide range of traditional and deep learning classifiers, including the BERT base model. Our key findings indicate that our data provides stronger indicators for detecting the Big Five traits than the SoCE dataset. Among the models tested, GPT-based approaches, notably GPT-4 (the latest version of GPT), consistently outperformed other approaches in identifying all five traits, even without prior training on the datasets. Additionally, we observe that fine-tuning GPT enhances its performance, particularly with the SoCE dataset. While achieving accuracy and F1 scores that are comparable to those in related studies, our research offers a more reliable evaluation of model performance by employing the Area Under the Curve (AUC) score, a metric that is more robust against data imbalance and sensitive model parameters. Moreover, our work underscores the practical applications of these models in real-world contexts, like behavioral job interviews, providing valuable insights for future research and applications in this field.