动词驱动事件的共指关系中文语料库构建及大模型评测

Summary

Most event coreference work is closed-domain: it only links mentions of events that fit a fixed type inventory (e.g., ACE2005, ECB+, MAVEN-ERE), so many real-world actions and static states go unaddressed. The authors avoid type inventories altogether and define coreference through the event's linguistic carrier, starting with verb–verb links in Chinese. Their annotation guideline uses HowNet sememes for verb meaning and Yuan Yulin's argument-role system for semantic roles. It separates strong coreference (identical meaning and matching relevant roles) from weak coreference, where one event contains or modifies the other. Using this scheme, they manually annotated a Chinese corpus built from People's Daily texts. They then tested a two-stage pipeline. Mention detection relies on constituency parsing, and coreference prediction uses large language models adapted through prompting, instruction tuning or knowledge distillation. Distillation performed best. Fine-tuned small models came close to the LLMs while running much faster. The work provides an open-domain Chinese benchmark and shows how far current LLMs still are from solving the task.

Key points

• Corpus: 1,000 documents (380k+ characters), 29,250 verbs, 1,858 strong coreference chains and 994 weak coreference verb pairs. It is released on GitHub under NP-NET-research (Reference-Resolution-for-Chinese-Verbs).
• Weak coreference has three subtypes: verb-sense inclusion, role inclusion and role change. Roles that affect coreference are agent/patient, time, location, range and proposition. Roles such as instrument, material, manner and cause are treated as irrelevant.
• Mention detection via constituency parsing reaches 96.9% F1.
• Average coreference-prediction F1 was 50.0% with prompt engineering, 64.4% with instruction tuning and 71.7% with knowledge distillation. The tested models were Zhipu Qingyan, ChatGLM3-6B and QWen-7B-Chat.
• Supervised BERT-style small models score slightly below instruction-tuned LLMs but are far more efficient at inference. The paper appears in vol. 39, no. 11 (Nov. 2025) of the journal.