acceptodds
Under review as a conference paper at ICLR 2027

An Evidence-Aligned Framework for In-Context Chinese Legal Named Entity Recognition

Abstract

Named entity recognition (NER) is a foundational information extraction task, and it must recover exact entity spans and types from limited demonstrations in the few-shot in-context settings. Chinese legal NER is particularly challenging because implicit word boundaries hinder exact span localization, while similar text spans may belong to different entity types depending on context. Existing methods often rely on whole-sequence or word-level retrieval, generate detached mention lists, creating a mismatch between the evidence provided and the evidence required for boundary- and role-sensitive decisions. In this paper, we propose C-NER, an in-context learning framework with a frozen LLM and no task-specific gradient updates. It retrieves demonstrations that contain local cues for locating entity boundaries, preserves predicted spans at their exact positions in the original text throughout extraction, and uses surrounding context to distinguish spans that may belong to different entity types. Experiments on four NER benchmarks, including complete-system and controlled-retrieval comparisons, show that C-NER achieves 90.99 F1 on CAIL2021 and 91.56 F1 on MSRA, outperforming the strongest reproduced baseline DEER by 3.65 and 1.87 F1 points, respectively, while ranking second on CLUENER2020 and OntoNotes. On CAIL2021, it uses one LLM call per instance and reduces LLM calls by 55.1% and token usage by 45.3% relative to DEER. Code, configurations, and reproducibility artifacts are available at https://anonymous.4open.science/r/charcon-ner.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
named entity recognition, in-context learning, Chinese legal text, demonstration retrieval, evidence alignment

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_64uDOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes C³-NER, a training-free framework for Chinese legal NER that integrates character-pattern retrieval, copy-preserving annotation, and contrastive guidance. Experiments on four NER benchmarks demonstrate improvements over existing methods on CAIL2021 and MSRA, supported by ablation studies and inference-cost comparisons.

Critical strengths
  1. The proposed framework combines retrieval, source-preserving annotation, and type disambiguation without requiring model fine-tuning.

  2. The method achieves competitive performance across multiple benchmarks and LLM backbones while reducing LLM calls and token consumption compared with existing methods.

Critical weaknesses
  1. The experimental evaluation is limited in scope. Only one legal NER benchmark is evaluated, while the other datasets use subsets of just 200 samples. Moreover, the proposed method underperforms DEER on CLUENER2020 and OntoNotes, raising concerns about its generalizability.

  2. The proposed framework relies on labeled training data to construct its Pattern Bank and manually defined confusion groups. However, the paper does not investigate sensitivity to the amount of available labeled data or the selection of confusion groups. This limits the evidence for its effectiveness in genuinely low-resource settings.

Questions for authors
  1. Can the authors provide results on the full benchmark test sets or report performance variance across different evaluation subsets to demonstrate the robustness of the reported improvements?

  2. How does performance change when the amount of labeled data available for demonstration retrieval and Pattern Bank construction is reduced?

  3. How are the confusion groups selected?

Reviewer_gAf7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes C³-NER, a framework for Chinese legal named entity recognition using a frozen LLM. It combines character-based demonstration retrieval, entity tags inserted into the original text, and training-derived hints for distinguishing confusing entity types. It improves over DEER by 3.65 F1 on CAIL2021 and 1.87 on MSRA, but performs worse on CLUENER2020 and OntoNotes. On CAIL2021, the gains persist across three backbones while requiring fewer model calls and tokens.

Critical strengths
  • The method's accuracy gains come with lower inference cost. On CAIL2021, C³-NER improves over DEER by 3.65 F1 while using 55.1% fewer calls and 45.3% fewer tokens. Gains across three backbones strengthen this practical result.
  • The retrieval experiments isolate a useful design choice. Holding the remaining pipeline fixed, character-based retrieval outperforms word-based retrieval on CAIL2021 and MSRA. This provides concrete evidence for avoiding fixed word segmentation in these settings.
Critical weaknesses
  • The methodological novelty is limited. Source-preserving annotation was already explored in TANL [1] and GPT-NER [2], while DEER already uses label-derived local evidence to guide extraction. The remaining contribution is chiefly their combination with character retrieval and pre-generation contrastive hints. The paper needs to distinguish this contribution more clearly from existing approaches.
  • The ablations do not distinguish the proposed explanation from simpler alternatives. The paper argues that overlapping character patterns retrieve short constructions that help locate entity boundaries. Yet single-character retrieval, which cannot encode character order within those constructions, already achieves 90.799 F1, compared with 90.940 for the full retriever (Table 7). The results could therefore reflect lexical overlap and avoiding segmentation errors, rather than retrieval of boundary-sensitive patterns. An analysis of boundary-only errors would help distinguish these explanations.
  • Similarly, removing contrastive guidance lowers F1 from 90.99 to 90.42 (Table 3), but aggregate F1 cannot show whether the guidance corrects confusion between legal roles or simply changes which spans are extracted. Reporting type errors on correctly localized spans, particularly within the predefined confusion groups should better show whether the guidance actually prevents one legal role from being mislabeled as another, rather than merely changing which entity spans are extracted.

Overall I think the methodological contribution appears incremental, and the claimed mechanisms need stronger validation.

[1] Paolini et al., 2021. Structured Prediction as Translation between Augmented Natural Languages. [2] Wang et al., 2025. GPT-NER: Named Entity Recognition via Large Language Models.

Questions for authors
  • What distinguishes Copy-Preserving Annotation from TANL and GPT-NER beyond formatting and validation details?
  • Are the small gains from multi-character patterns and contrastive guidance statistically reliable, and which boundary or type errors do they correct?
Reviewer_cPj2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for in-context Chinese legal named entity recognition to address three important challenges: selecting demonstrations with relevant entity boundary information, preserving the correspondence between predicted entities and their original text, and distinguishing context-dependent legal entity types. The proposed approach integrates character-pattern retrieval, copy-preserving annotation, and contrastive guidance, achieving strong empirical performance on the primary benchmark. However, the technical novelty appears limited, as the framework mainly combines existing techniques. Additionally, including a concrete Chinese legal NER example would help illustrate the unique challenges of the task and improve the clarity of the proposed approach. Finally, a comparison with the same LLM backbone using simple prompting without the proposed components would provide a stronger baseline and help demonstrate the effectiveness of the proposed framework.

Critical strengths
  1. This paper identifies three important challenges in in-context Chinese legal NER: selecting demonstrations that provide relevant entity boundary information, preserving the correspondence between predicted entities and their original occurrences, and distinguishing context-dependent legal entity types.
  2. The proposed approach demonstrates strong empirical performance on the primary benchmark, supporting its effectiveness for Chinese legal NER.
Critical weaknesses
  1. The technical novelty of the proposed approach appears limited, as it mainly combines three existing techniques: character-pattern retrieval, copy-preserving annotation, and contrastive guidance. The authors should further clarify the methodological contributions beyond the integration of these components.
  2. Including a concrete example of Chinese legal NER with Chinese characters would help readers better understand the unique characteristics and challenges of the task, particularly entity boundary identification and context-dependent entity classification.
  3. A comparison using the same LLM backbone with simple prompting, without any of the proposed components, would further strengthen the experimental evaluation. Such a comparison would help demonstrate the effectiveness of the proposed approach and clarify the performance improvements achieved beyond the capabilities of the underlying LLM.
Questions for authors

refers to the weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.