Narrative Hermeneutic Circle: Improving Character Role Identification from Natural Language Text via Feedback Loops.

Josep Valls-Vargas, Jichen Zhu, Santiago OntañónIJCAI 2015 · OpenReview

Abstract

While most natural language understanding systems rely on a pipeline-based architecture, certain human text interpretation methods are based on a cyclic process between the whole text and its parts: the hermeneutic circle. In the task of automatically identifying characters and their narrative roles, we propose a feedback-loop-based approach where the output of later modules of the pipeline is fed back to earlier ones. We analyze this approach using a corpus of 21 Russian folktales. Initial results show that feeding back high-level narrative information improves the performance of some NLP tasks.

Introduction

In computational linguistics, most systems follow a linear "pipeline" architecture, processing text from basic syntax to high-level semantics. However, narratives present unique challenges; a single character may be referenced through various descriptors and pronouns (e.g., "the maiden," "she," "the hostess"). Resolving these mentions to a single entity—coreference resolution—remains a significant bottleneck in deep story understanding.

Humans interpret text through a "hermeneutic circle," where understanding the whole narrative informs the interpretation of its parts. This paper proposes a feedback-loop architecture for narrative extraction that mimics this process, feeding high-level narrative roles (like "Hero" or "Villain") back into low-level linguistic tasks to improve coreference resolution.

Methodology

The authors developed Voz, a system that moves beyond one-way pipelines toward an iterative, cyclic interpretation process inspired by the philosophical concept of the hermeneutic circle.

The Voz system architecture featuring a feedback loop from character roles back to coreference resolution.

Symbolic Extraction and Coreference Resolution

Voz begins with standard NLP preprocessing using the Stanford CoreNLP suite for POS tagging and dependency parsing. For coreference resolution, it employs six distinct "Coreference Preference" (CP) matrices ($m_1$ through $m_6$). These matrices represent the likelihood that any two mentions in the text refer to the same entity based on different linguistic features.

The coreference resolution process combines multiple preference matrices ($m_1$ to $m_6$) into a final entity graph $G$.

The Feedback Loop

In the initial pass, the system uses matrices based on syntax and gender/number. After predicting character roles (e.g., identifying a character as the "Helper"), this high-level information is fed back into matrices $m_5$ and $m_6$. If two mentions are predicted to share the same narrative role, the system increases the likelihood of grouping them as the same character. This process iterates until the results stabilize, typically by the third pass.

Experimental Setup

The system was evaluated on a corpus of 21 Russian folktales translated into English. Roles were defined based on Vladimir Propp’s narrative theory (1973), which identifies functional roles such as Hero, Villain, Provider/Donor, and Helper. The authors manually annotated 4,791 mentions to create a ground-truth dataset. To manage the small dataset size, Propp’s 31 sub-roles were merged into broader categories (e.g., Hero, Villain, Other).

Results and Discussion

Coreference Improvement

The feedback loop significantly refined the grouping of mentions. The initial pipeline pass generated 1,359 coreference groups, which the feedback loop reduced to 888 groups, moving closer to the ground-truth of 173 actual characters. This demonstrates that global narrative context can effectively correct local linguistic errors.

Role Accuracy and Stability

While coreference resolution improved dramatically, the improvement in role identification accuracy was marginal (from 0.64 to 0.65). However, character prediction reached an accuracy of 0.882 after the voting process, a significant improvement over previous baseline methods. Notably, the system's interpretations stabilized quickly, with the second and third iterations often yielding identical results.

Conclusion

This work demonstrates that breaking the traditional linear pipeline in favor of a hermeneutic feedback loop can alleviate the "authorial bottleneck" in story understanding. Future work will involve validating this approach on larger, more diverse literary datasets and exploring other narrative frameworks, such as Joseph Campbell’s "Hero’s Journey."

Summary updated

Abstract

While most natural language understanding systems rely on a pipeline-based architecture, certain human text interpretation methods are based on a cyclic process between the whole text and its parts: the hermeneutic circle. In the task of automatically identifying characters and their narrative roles, we propose a feedback-loop-based approach where the output of later modules of the pipeline is fed back to earlier ones. We analyze this approach using a corpus of 21 Russian folktales. Initial results show that feeding back high-level narrative information improves the performance of some NLP tasks.
BibTeX
@inproceedings{vallsvargas2015narrative,
  author = {Josep Valls-Vargas and Jichen Zhu and Santiago Ontañón},
  title = {{Narrative Hermeneutic Circle: Improving Character Role Identification from Natural Language Text via Feedback Loops.}},
  booktitle = {IJCAI},
  year = {2015},
  url = {https://openreview.net/forum?id=Csx-dVpJUOV8},
}
Figure