named entity recognition

To address the inefficiency and fine-tuning dependency of entity mention detection in large language models (LLMs), this paper proposes ToMMeR—a lightweight probe model (<300K parameters) leveraging early-layer hidden representations of LLMs. Methodologically, ToMMeR exploits the emergent, structured entity representations naturally present in early Transformer layers, enabling zero-shot, high-recall mention boundary identification. It integrates a span classification head with the LLM’s intrinsic discriminative capability to filter and calibrate candidate mentions. Evaluated across 13 NER benchmarks, ToMMeR achieves 93% zero-shot recall; after LLM-based discrimination, precision exceeds 90%. With an extended classification head, F1 scores reach 80–87%, approaching state-of-the-art performance. This work is the first systematic demonstration that early LLM layers encode transferable, trainable-free mention detection capacity—establishing a new paradigm for efficient, training-free entity recognition.

Existing natural language processing resources often lack task-specific information for niche or emerging entities, hindering accurate classification in domains such as business or healthcare provider categorization. To address this limitation, this work proposes a dynamic classification framework that requires no additional labeled text: given only entity names and their corresponding labels, the method retrieves web-based information and leverages large language models (LLMs) to generate task-relevant descriptions, which are then used to train a text classifier. This end-to-end approach achieves strong performance on low-resource entity classification, attaining macro-averaged F1 scores of 82.3% on Standard Industrial Classification (SIC) coding and 72.9% on healthcare provider categorization, thereby demonstrating its effectiveness and practical utility.

Existing masked language models (MLMs) struggle with narrative role classification in long documents due to their limited context windows, hindering effective capture of discriminative contextual cues. Method: We propose a lightweight, entity-centric heuristic context selection framework that dynamically identifies salient text segments via entity localization rules, and further enhances information density through context truncation and semantic reweighting. This enables efficient supervised fine-tuning on compact MLMs such as XLM-RoBERTa. Results: Our approach achieves 89.7% accuracy on the entity-frame classification subtask (Subtask 1) of SemEval-2025 Task 10—matching or exceeding the performance of large generative models (e.g., GPT-4, Claude-3) while reducing inference cost by two orders of magnitude. The core contribution is the first integration of entity-driven context refinement into MLM-based long-document narrative understanding, offering a high-efficiency, resource-frugal solution for constrained deployment scenarios.

To address the low accuracy of structured data extraction from materials science literature, this paper proposes a hybrid text-mining framework. First, symbolic entity markers are introduced to enhance named entity recognition (NER) performance; subsequently, a joint modeling approach integrates sequence labeling with structured generation to enable collaborative extraction of entities and relations. This method innovatively combines the strengths of multi-stage and end-to-end paradigms, overcoming traditional limitations in fine-grained entity identification and complex relational modeling. Evaluated on three authoritative benchmark datasets—MatScholar, SOFC, and one additional domain-specific corpus—the framework achieves a 58% improvement in entity-level F1 score and an 83% improvement in relation-level F1 score over state-of-the-art methods. The proposed approach establishes a new, efficient, and robust paradigm for constructing scientific literature knowledge graphs.

To address the sensitivity of in-context learning (ICL) to example selection, the scarcity of human annotations, and the poorly understood generalization mechanisms in low-resource named entity detection (NED), this paper proposes PICLe: a framework that requires no manual labeling. PICLe leverages large language models (LLMs) for zero-shot generation of noisy pseudo-labeled in-context examples; integrates semantic clustering-based sampling with multi-path parallel inference; and introduces a self-verification mechanism to dynamically filter high-confidence predictions and fuse decisions. A key finding is that partially correct pseudo-labeled examples achieve ICL performance comparable to fully correct ones. Evaluated on five biomedical NER datasets, PICLe significantly outperforms standard ICL under zero-shot and few-shot settings—especially when only a small number of gold examples are available—demonstrating robust, scalable contextual generalization for low-resource NED without human annotation.

Large language models are prone to generating hallucinations or dubious citations in academic writing, undermining research credibility. This study presents the first systematic evaluation and comparison of mainstream citation verification tools—CheckIfExist, HalluCiteChecker, Hallucinator, HalRef, and RefChecker—on real-world academic documents. The analysis reveals significant limitations in current approaches, particularly concerning citation extraction accuracy, breadth of database coverage, and consistency in verification. While these tools can offer preliminary alerts for potentially fabricated references, their overall effectiveness remains constrained. This work provides an empirical foundation and clear directions for improving the verification of citation authenticity in scholarly communication.

This study addresses the challenge of identifying substance use in low-resource, unstructured Spanish electronic health records by proposing a multi-task ensemble deep learning framework that jointly models toxicant named entity recognition (ToxNER) and usage context detection (ToxUse). The approach integrates the BETO pretrained language model with conditional random field (CRF) decoding, diverse training strategies, and a sentence-level filtering mechanism to enhance model accuracy and robustness under scarce labeled data. Experimental results demonstrate that the system achieves an F1 score of 0.94 and precision of 0.97 for trigger detection, and an F1 score of 0.91 for argument detection, significantly outperforming baseline methods.

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 limitations of existing ultra-fine entity typing methods, which rely solely on sentence-level context and struggle to accurately classify long-tail entity types. To overcome this, the authors propose a novel approach that leverages controllable generation to produce coherent multi-sentence narratives that expand entity context. They introduce the first Narrative-UFET dataset and design two variants of synthetic narratives—type-preserving and type-altering—to systematically investigate how discourse structure influences entity type disambiguation. Experimental results demonstrate that multi-sentence contexts derived from synthetic narratives substantially outperform sentence-level baselines, particularly for long-tail types. Notably, type-altering narratives provide stronger discriminative signals, and synthetic narratives consistently surpass natural contexts in classification performance, confirming the effectiveness and potential of controllable narrative generation for enhancing ultra-fine-grained entity typing.

This work addresses document-level conspiracy theory detection by proposing a joint framework that integrates multi-label span classification with sequence classification. For extracting conspiracy-related markers—such as roles and actions—the approach formulates the task as boundary-aware multi-label span classification, incorporating IoU-based positive labeling, hard negative sampling, and an inclusion-aware non-maximum suppression strategy, while distinguishing between entity-like and abstract roles. Document-level classification is performed using a RoBERTa model enhanced with label smoothing. Evaluated on SemEval-2026 Task 10, the method achieves 7th place in Subtask 1 (macro F1 = 0.2251) and 11th place in Subtask 2 (weighted F1 = 0.7694), demonstrating the effectiveness of the proposed techniques.