PaErHaTi-DUTiR/FRAME-main

Failure-aware Knowledge Bridging for Low-Resource Multilingual Named Entity Recognition with Large Language Models # FRAME **Failure-aware Knowledge Bridging for Low-Resource Multilingual Named Entity Recognition with Large Language Models** ## Overview Large language models enable in-context multilingual named entity recognition (NER) without per-language annotation or retraining, but flexibility alone does not guarantee reliability. FRAME identifies three distinct, separable failure modes in LLM-based multilingual NER — cross-lingual boundary uncertainty, non-English type ambiguity, and generative omission — and assigns each one a dedicated, read-only knowledge bridge at inference time: - **Cross-lingual semantic bridge** (`SpanAgent`) — retrieves demonstrations through dense multilingual similarity to stabilize span boundaries. - **English semantic bridge** (`TypeAgent`) — routes entity typing through translation with an explicit reject option to reduce type confusion. - **Lexical memory bridge** (`Gazetteer`) — applies a corpus-derived gazetteer as a non-overlapping, decision-level completion step. None of the three bridges updates model parameters or requires additional annotation beyond what the training splits already provide. FRAME runs two LLM calls per sentence (SpanAgent, then TypeAgent) plus a rule-based gazetteer merge. ## Key results DeepSeek-V3, few-shot, entity-level micro-F1 (full comparison across GPT-4o, Qwen-2.5-72B, and zero-shot settings is in the paper): | Method | WikiANN | CoNLL-03/02 | MasakhaNER2.0 | |---|---|---|---| | GPT-NER | 59.03 | 73.91 | 66.32 | | CodeIE | 60.65 | 85.67 | 71.90 | | KDR-Agent | 70.44 | 84.49 | 70.21 | | **FRAME (Ours)** | **79.81** | **87.27** | **74.81** | Ablation (DeepSeek-V3, k=3) — relative F1 decline from removing one bridge at a time: | Variant | WikiANN | CoNLL-03/02 | MasakhaNER2.0 | |---|---|---|---| | **FRAME (Ours)** | **79.81** | **87.27** | **74.81** | | w/o English bridge | 77.22 (−3.25%) | 86.28 (−1.13%) | 69.78 (−6.72%) | | w/o lexical bridge | 75.64 (−5.22%) | 86.77 (−0.57%) | 72.87 (−2.59%) | ## R …