TL;DR
This paper investigates whether fictional characters' speech patterns are distinct enough to be computationally classified, using modern drama as a testbed. They found that certain authors, particularly canonical ones, excel at creating unique character voices, and that a Sparse Additive Generative (SAGE) model is particularly effective for this classification task.
What Problem Does This Paper Address?
The paper delves into the concept of "dialogism" in literature, a term coined by Mikhail Bakhtin. Bakhtin argued that truly dialogic works feature characters with independent, distinct voices, rather than merely serving as mouthpieces for the author. While previous computational literary studies have explored dialogism, this paper takes a step back to ask fundamental questions: Can we actually distinguish these character voices computationally? If so, which methods work best? And are some authors demonstrably better at creating these distinct voices, perhaps correlating with their canonical status? The authors focus on modern drama because plays offer explicit speaker labels, simplifying the task of attributing dialogue.
Methods
The core of the methodology is text classification. The authors experiment with several approaches:
- Lexical and Syntactic Features: This standard approach involves extracting features like sentence/word length, type-token ratio, function word proportions, and n-gram frequencies (for words and part-of-speech tags). They also use dependency triples and mask proper nouns to avoid author-specific clues. Feature selection techniques like variance threshold and chi-squared statistic are employed to manage high dimensionality. A Support Vector Machine (SVM) classifier is used here.
- Sparse Additive Generative (SAGE) Models: This is a key novel contribution. SAGE models are generative models that represent a class's word distribution as deviations from a background distribution, with a bias towards rare and infrequent terms. The authors use SAGE to derive weights for each word in a sentence, and then average these weighted word vectors to create sentence vectors. These vectors are then fed into a logistic regression classifier.
- Distributed Word Vectors (word2vec): They also use word2vec embeddings, representing each dialogue as a weighted average of word vectors, with weights derived from TF-IDF or SAGE.
- Lexicon-Based Features: They incorporate lexicons that score words across stylistic dimensions (e.g., colloquial vs. literary, concrete vs. abstract, subjective vs. objective) and emotion intensity (anger, fear, sadness, joy) and valence, arousal, dominance. Principal Component Analysis (PCA) is used on these vectors to visualize authorial differences.
Datasets
The primary dataset consists of 63 plays published between 1880 and 1920, by authors including George Bernard Shaw, Oscar Wilde, Cale Young Rice, Sydney Grundy, Somerset Maugham, Arthur Wing Pinero, and Hermann Sudermann. This period was chosen to align with Bakhtin's theoretical development. The authors also create "artificial plays" by sampling characters from existing plays to test the robustness of their methods.
Key Insights and Ideas
- Character Voices are Distinguishable: Contrary to some skepticism, the paper strongly suggests that fictional character voices can be computationally distinguished, often with high precision. This supports the intuition behind dialogism.
- SAGE is a Powerful Tool: The SAGE model proves particularly effective for this task, outperforming other methods by effectively weighting distinctive and infrequent words that contribute to unique character styles.
- Canonical Authors Excel: Playwrights who are more canonical (like Shaw) tend to be more successful at creating characters with distinct voices, suggesting this ability might be linked to literary prominence or talent.
- Lexical and Dialectal Features are Key: The analysis highlights that simple lexical and dialectal features (e.g., specific vocabulary, common phrases) are very effective in distinguishing characters, even those with fewer lines.
- Word Embeddings Capture Subtle Style: While less interpretable, word2vec embeddings, when combined with SAGE weights, significantly boost classification performance, indicating they capture nuanced stylistic aspects.
Key Results
- Classification Accuracy: The paper reports F1 scores for character classification. For individual characters using lexical and syntactic features, Shaw achieved the highest average F1 score (.400), significantly above the baseline (.133). When using SAGE + word2vec, the average F1 score reached .605 for original plays.
- Authorial Distinction: PCA plots (Figure 1) show that authors like Shaw and Maugham have characters spread out across the principal components, indicating a wider range of distinct styles, while authors like Rice have more clustered characters.
- Artificial Plays: Artificial plays generally showed higher classification scores than original plays, suggesting that the controlled sampling might isolate stylistic differences more effectively, though the interpretation of this is complex.
Limitations and Future Work
- Small Sample Size: The authors acknowledge that their sample of authors and plays is small, and definitive conclusions about canonical status and distinctiveness require a larger dataset.
- Genre Specificity: The study is limited to modern drama. The authors suggest investigating the effect of genre (e.g., comedy) on distinctiveness.
- Interpretability of Word Embeddings: While effective, the dense vectors from word2vec are difficult to interpret directly, making it hard to pinpoint why they contribute to classification.
- Artificial Play Interpretation: The higher scores for artificial plays are noted as an interesting observation but require further controlled experiments for clear interpretation.
- Future Work: The authors suggest expanding the corpus to include more authors and genres, including novels, to draw more robust conclusions about dialogism and authorial distinctiveness.
Summary updated