TL;DR
This paper tackles the challenge of automatically identifying and linking different mentions of fictional characters in Chinese novels, a crucial task for digital humanities and literary analysis. They adapt a state-of-the-art anaphora resolution model (CHEN-Model) for this specific domain, incorporating speaker information from dialogue and addressing unique Chinese language features like zero pronouns.
What Problem Does This Paper Address?
Digital humanities research increasingly relies on automatic text analysis to process vast literary archives. However, existing Natural Language Processing (NLP) tools, often trained on general-domain text, struggle with the nuances of literary works. Fictional characters are central to novels, but their representations are diverse and complex, including proper names, aliases, pronouns, and especially in Chinese, "zero pronouns" (where a pronoun is omitted but understood from context). Resolving these varied mentions is a prerequisite for understanding character arcs, relationships, and plot development. This paper focuses on adapting NLP techniques to accurately resolve these character representations in Chinese novels, a task that goes beyond standard coreference resolution.
Methods
The authors propose ARMCC (Anaphora Resolution Model for fictional Characters in Chinese novels), an adaptation of the CHEN-Model, which jointly handles zero pronoun resolution and non-zero coreference resolution. Key modifications include:
- Speaker Extraction (SECN): They train a BERT-based model to extract speaker information from dialogue. This is crucial because dialogue constitutes a significant portion of novels, and speaker attribution helps disambiguate character references.
- Handling Chinese Quotation Marks: They replace Chinese quotation marks with standard pairwise marks to ensure proper tokenization and processing.
- Word Boundary Filtering: They refine the model to exclude units that appear within or across word boundaries, which are not valid character representations in Chinese.
- MacBERT for Embeddings: They leverage MacBERT, a pre-trained model known for its strong performance on text-level NLP tasks, for learning initial token and gap embeddings.
- Two-Stage Unit Scoring: ARMCC uses a two-stage process to score units (spans or gaps) for their likelihood of referring to a character, incorporating information from relevant units.
- Unit-Pair Scoring with Speaker Information: The model then scores pairs of units for their likelihood of referring to the same character, explicitly incorporating speaker information as a feature.
Datasets
- OntoNotes-Ch: Used as a base for training and developing ARMCC. The authors created a new dataset by filtering OntoNotes-Ch to focus specifically on person names and their coreferences.
- DJY Corpus: A corpus of 28,597 quotes from Jin Yong's novels, manually annotated for speaker identification.
- DWoP Corpus: A corpus of 1,835 quotes from the Chinese novel "World of Plainness," also annotated for speaker identification.
- DSECN: A combined and SQuAD-formatted dataset derived from DJY and DWoP, used for training the speaker extraction model (SECN).
- White Deer Plain: A specific Chinese novel used for end-to-end evaluation of the proposed ARMCC and SECN models.
Key Insights and Ideas
- Speaker Information is Crucial: Dialogue is a rich source of information for character resolution. Incorporating speaker attribution significantly improves performance, especially in genres with high dialogue proportions.
- Domain Adaptation is Necessary: General-domain NLP models are insufficient for literary analysis. ARMCC's modifications, tailored to Chinese novels, demonstrate the need for domain-specific adaptations.
- Zero Pronouns are a Major Hurdle: While ARMCC shows promise, resolving zero pronouns remains a significant challenge, highlighting the complexity of Chinese grammar in literary contexts.
- Character Representations are More Than Names: The model's ability to identify nearly twice as many character representations as just person names underscores the importance of resolving aliases, pronouns, and zero pronouns for a comprehensive understanding of characters.
Key Results
- ARMCC Performance: Achieved an F1-score of 70.62% for non-zero coreference resolution and 32.25% for zero pronoun resolution on the development set.
- SECN Performance: The speaker extraction model (SECN) achieved an Exact Match (EM) accuracy of up to 96.80% on Chinese novels, outperforming baseline methods significantly.
- Ablation Studies: Demonstrated that all proposed modifications (speaker information, word boundary filtering, quotation mark handling, and using MacBERT) contributed to ARMCC's performance, with word boundary filtering and MacBERT showing particular benefits for zero pronoun resolution.
- White Deer Plain Analysis: Case studies showed ARMCC successfully resolved many character representations, including pronouns and zero pronouns, and could identify character attributes. However, it still struggled with complex coreference relationships and certain zero pronoun cases.
Limitations and Future Work
- Zero Pronoun Resolution: The paper acknowledges that zero pronoun resolution remains a significant challenge, with lower F1-scores compared to non-zero coreference. They suggest that dependency parsers could offer temporary relief.
- Complex Coreference: ARMCC struggles with situations where multiple characters are referred to by different representations simultaneously, leading to incorrect clustering. They propose using character name lists and simple rules for splitting clusters, or dynamically adjusting model parameters.
- Novel-Specific Resources: The limited availability of annotated Chinese novel datasets hinders model performance. The authors emphasize the need for building more novel-specific annotated resources.
- Hidden Identity Plots: Resolving characters in plots involving hidden identities, pretenders, or substitutes is difficult and may require manual assistance.
- Syntactic Parsers and Heuristic Rules: The authors suggest that combining ARMCC with syntactic parsers and simple heuristic rules could further optimize character resolution.
- Dynamic Parameter Adjustment: For SECN, they suggest that dynamically adjusting parameters like 'm' and 'k' based on the number of different characters could alleviate errors in complex scenarios.
Summary updated