TL;DR
This paper explores the feasibility of using computational linguistics techniques, specifically machine learning models trained on personality trait datasets, to automatically infer the personality of characters in literary texts, focusing on Shakespearean plays. The authors demonstrate that while current models can achieve reasonable performance, the unique nature of literary language and character portrayal presents significant challenges compared to analyzing social media text.
What Problem Does This Paper Address?
Understanding and profiling personality is a key aspect of comprehending human behavior and interaction. While computational methods have made strides in inferring personality from modern communication (like social media), applying these techniques to literary texts, particularly fictional characters, is less explored. Literary characters are often crafted with nuanced personalities through dialogue and action, and it's unclear how well existing personality models, designed for real-world communication, can capture these fictional portrayals. The authors aim to bridge this gap by investigating the adaptability of personality trait recognition frameworks to literary works, specifically theater scripts, to gain insights into character psychology.
Methods
The authors frame personality trait recognition as a multi-output supervised learning problem. They initially train models on a labeled dataset of "essays" (stream-of-consciousness texts) that were annotated with personality traits based on the Five Factor Model (Big5). Their modeling pipeline incorporates a bag-of-words approach combined with linguistic features extracted by the TextPro tool (including part-of-speech tags, chunking, named entities, and token types). This pipeline is then applied to their literary dataset. For evaluation, they use F1-scores, a common metric for classification tasks, and compare their results to existing work.
Datasets
- Training Dataset: "Essays" dataset, originally from Pennebaker and King (1999). This dataset consists of approximately 2400 stream-of-consciousness texts from students who also completed Big5 personality questionnaires. Labels are derived from z-scores and median splits.
- Literary Dataset: Shakespeare's plays, specifically using the Open Source Shakespeare (OSS) corpus, which contains 38 plays. Dialogues are extracted, with each character's turns treated as a single document.
Key Insights and Ideas
- Lexical Hypothesis in Fiction: The core assumption is that personality traits can be inferred from language, even in fictional contexts. The paper tests the robustness of this "lexical hypothesis" when applied to carefully crafted literary dialogues.
- Theater Scripts as a Data Source: Theatre scripts are identified as a convenient source for character dialogue due to explicit character attribution, reducing the need for manual annotation.
- Big5 Model Limitations for Literary Nuance: While the Big5 model is a standard, the authors suggest it might be too broad for capturing the subtle personality nuances of literary characters. They note that literary critics often prefer deeper analysis over simple categorization.
- Dataset Mismatch Challenges: The "essays" dataset, while providing gold labels, differs significantly from literary texts in terms of purpose, grammar, and language evolution (diachronicity), potentially impacting model transferability.
- Character Archetypes vs. Nuance: The results suggest that characters with similar dialogue patterns might be assigned similar traits, even if their roles or motivations differ greatly (e.g., protagonists and antagonists in "Hamlet"). This highlights a potential limitation in distinguishing complex character arcs.
Key Results
The paper reports that their models achieve state-of-the-art performance on the "essays" dataset. When applied to Shakespeare's plays, the models produce trait attributions for verbose characters. For instance, in "Hamlet," most verbose characters are assigned traits like Openness (OPN) and Agreeableness (AGR). Shylock from "Merchant of Venice" is predicted to have neuroticism, aligning with the authors' shallow knowledge of the play. However, they observe low variability for Agreeableness, Extraversion, and Openness traits across characters, which they find counterintuitive.
Limitations and Future Work
- Dataset Mismatch: The significant differences between the "essays" training data and the Shakespearean literary text (purpose, grammar, language diachronicity) are acknowledged as a major limitation.
- Model Granularity: The Big5 model's broadness is a concern; the authors mention the existence of "facets" within the Big5 but lack suitable datasets. They also suggest that more specific trait categories might be more appropriate for literary characters.
- Dialogue-Centric Approach: Focusing solely on dialogue and ignoring actions or stage directions might lead to information loss.
- Character Evolution: The static nature of the current models doesn't account for potential personality changes within a character throughout a play.
- Contextual Understanding: The model doesn't consider the recipient of a character's speech, which can influence personality expression.
Future work could involve exploring finer-grained personality models (facets), incorporating non-dialogue elements (actions), developing methods to track character personality evolution, and investigating how speech context (who is being spoken to) affects trait attribution. The authors also suggest dispersion plots to visualize character trait changes over time.
Summary updated