entity and relation extraction

This work addresses the high computational overhead and latency of existing span-based named entity recognition methods, which enumerate numerous candidate spans and rely on token-level augmentation, rendering them inefficient for industrial applications demanding real-time performance. To overcome these limitations, the authors propose SpanDec, a framework that concentrates span representation interaction exclusively in the final Transformer layer and introduces a lightweight decoder coupled with a dynamic candidate filtering mechanism. This design enables early pruning of low-quality spans, thereby eliminating redundant computations in earlier layers. SpanDec achieves accuracy comparable to state-of-the-art span-based models while significantly improving throughput and reducing computational cost, making it well-suited for high-concurrency services and edge deployment.

To address the challenge of jointly preserving semantic integrity and structural consistency in nested and overlapping named entity recognition (NER), this paper proposes a structure-aware decoding framework. The method leverages pretrained language model representations and integrates multi-granularity span composition with hierarchical decoding. Its core contributions are: (1) a novel collaborative mechanism between candidate span generation and structured attention, explicitly modeling entity boundaries, hierarchical nesting, and cross-entity dependencies; and (2) a joint optimization objective incorporating hierarchical structural constraints and semantic–structural consistency, combining classification loss with structure-consistency loss. Experimental results on ACE 2005 demonstrate significant F1-score improvements over prior work. The approach achieves superior precision, recall, and boundary localization for both nested and overlapping entities, and exhibits strong robustness on long sentences and scenarios with multiple co-occurring entities.

This work addresses few-shot named entity recognition (NER) by challenging the end-to-end joint modeling paradigm. Drawing on generative grammar theory, we propose a “extract-then-classify” decoupled framework: entity extraction is treated as a syntactic task—requiring no semantic information—while classification is delegated to pre-trained language models (PLMs) or large language models (LLMs) as a semantic task. Empirical analysis reveals that rare words—particularly proper nouns—serve as critical syntactic cues; high-precision extraction is achieved using only shallow syntactic features (e.g., POS tags, dependency relations, and n-grams), with word embeddings or contextualized semantic representations yielding no performance gain. On benchmarks including CoNLL-2003, our extraction module achieves state-of-the-art F1 scores; ablation studies confirm that incorporating semantic features does not improve extraction accuracy. To our knowledge, this is the first study grounding syntactic–semantic separation in formal linguistics, providing both theoretical justification and empirical validation for decoupled modeling, while elucidating the root cause of failures in multi-task joint parsing.

To address the inefficiency, slow inference, and high resource consumption in jointly modeling entity linking (EL) and relation extraction (RE) for large-scale NLP tasks, this paper proposes cIE, a retrieval-reading dual-stage joint framework. Methodologically, cIE introduces a novel unified input representation that jointly encodes candidate entities and relations, enabling end-to-end alignment via a single forward pass through a shared reader; it further incorporates a contextualized candidate injection mechanism and leverages lightweight fine-tuning of pretrained language models for efficient joint decoding. Contributions include: (1) achieving cross-domain EL+RE joint SOTA under limited academic computational resources for the first time; (2) attaining state-of-the-art performance on multiple in-domain and cross-domain benchmarks; (3) accelerating inference by up to 40× compared to prior joint models; and (4) setting new records across multiple key metrics on the cIE task.

To address the susceptibility of large language models (LLMs) to training data bias—leading to hallucinations—and their heavy reliance on extensive annotated data and computational resources in relation extraction (RE), this paper proposes RAG4RE, the first systematic framework integrating retrieval-augmented generation (RAG) into RE. RAG4RE requires no fine-tuning and synergistically combines dense passage retrieval (DPR), semantic matching, and LLMs (e.g., Flan-T5, Llama2, Mistral) for robust relation identification under few-shot or zero-shot settings. Evaluated on standard benchmarks including TACRED and TACREV, RAG4RE achieves up to a 7.2% absolute F1 improvement over strong LLM-only baselines and conventional supervised methods. These results empirically validate that retrieval-guided generation effectively mitigates hallucination, reduces dependency on labeled data, and enhances generalization across diverse RE tasks.

This work addresses the scarcity of fine-grained annotated resources for named entity recognition (NER) and relation extraction (RE) in art history. To bridge this gap, the authors introduce FRAME, a novel dataset comprising descriptions of individual artworks sourced from museum catalogs and auction records. The dataset features a three-tier manual annotation scheme—metadata, content, and coreference layers—covering 37 entity types aligned with Wikidata. Annotations are provided in stand-off format using UIMA XMI CAS, facilitating tasks such as entity linking, knowledge graph construction, and fine-tuning or evaluation of large language models. As the first open-source, fine-grained, multi-layer annotated resource tailored to art historical research, FRAME establishes a benchmark for NER, RE, and few-shot or zero-shot modeling in this domain.

This work addresses the challenges of multi-domain named entity recognition (NER) under conditions of scarce labeled data, where conventional approaches suffer from domain discrepancy, data sparsity, and overfitting. To overcome these limitations, the paper proposes a unified framework that integrates unsupervised pre-training, transfer learning, data augmentation, few-shot learning, and domain-adversarial training. This approach enables effective adaptation to target domains without requiring any annotated data therein, significantly enhancing model generalization and robustness in low-resource, multi-domain settings. Experimental results demonstrate that the proposed method substantially outperforms existing baselines, offering a novel and practical pathway toward efficient and transferable NER in resource-constrained environments.

This work addresses the limitations of traditional approaches to named entity recognition (NER) and relation extraction (RE), which typically rely on separate models and struggle to support zero-shot joint inference over arbitrary entity and relation types. The authors propose GLiNER-Relex, the first extension of GLiNER to unified NER and RE, introducing an end-to-end framework built upon a shared bidirectional Transformer encoder. By leveraging span-level entity representations, configurable relation type embeddings, and a dedicated relation scoring module, GLiNER-Relex enables single-model, zero-shot extraction of knowledge triples with arbitrary labels. Evaluated on CoNLL04, DocRED, FewRel, and CrossRE benchmarks, it matches or surpasses both task-specific models and large language models in performance while maintaining high computational efficiency, and includes a clean, open-source inference API.

This work addresses the absence of high-quality, multi-domain named entity recognition and linking (NERL) datasets for historical Italian by introducing ENEIDE, the first publicly available silver-standard dataset for this language variety. ENEIDE comprises 2,111 documents from two scholarly digital collections spanning the 18th to 20th centuries, annotated with over 8,000 entities and partitioned into training, development, and test sets. The dataset incorporates an innovative NIL (not-in-lexicon) handling mechanism and leverages semi-automatic annotation, Wikidata entity linking, and rigorous quality control. Baseline experiments demonstrate that ENEIDE poses a substantial challenge to current NERL models, revealing a significant performance gap between zero-shot and fine-tuned approaches, while also enabling temporal disambiguation and cross-domain evaluation.