MythNER: Multi-Agent NER Extraction and Benchmarking in Chinese Myth Narratives

ACL ARR 2026 January Submission8178 Authors

Keywords: Named Entity Recognition, Information Extraction, Benchmark Dataset, Chinese, Mythology, Multi-Agent Systems

Abstract:

Named entity recognition (NER) performs strongly in well‑studied domains, yet mythological narratives pose a long‑tail setting where entityhood is defined by referable instances, mentions vary through titles and aliases, and correct extraction requires stable long‑span boundaries. We introduce MythNER, a culturally grounded coarse‑grained NER benchmark for Chinese mythology built from Ne Zha subtitle text. MythNER uses four flat labels (PER/LOC/ORG/OBJ) with conservative annotation criteria; in particular, OBJ is restricted to named artifacts and fixed technique titles rather than generic concepts. We evaluate a zero‑shot Chinese spaCy model, a supervised BERT token‑classification baseline, and a multi‑agent LLM extraction pipeline with chunk/context tuning on the held‑out Ne_Zha_P2 test set.
Results show substantial domain shift for off‑the‑shelf NER, strong gains from supervised adaptation, and further improvements from agentic extraction under well‑chosen constraints. Our analysis characterizes dominant failure modes—boundary drift under exact‑span scoring, mythology‑specific type ambiguity, and over‑extraction of generic nouns as OBJ—highlighting the need for iterative consistency checks in narrative‑domain NER.

Paper Type: Long

Research Area: Resources and Evaluation

Research Area Keywords: Information Extraction, Resources and Evaluation, AI/LLM Agents, Computational Social Science, Cultural Analytics, and NLP for Social Good, NLP Applications

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models, Data resources, Data analysis

Languages Studied: Chinese

Submission Number: 8178