Skip to content

๐Ÿ“„ information_extraction

๐Ÿ’ฌ EMNLP2026 ยท 9 paper notes

๐Ÿ”ฅ Top topics: LLM ร—2 ยท Adversarial Robustness ร—2

EMER-Guard: Protocol-Grounded Verification for LLM Emergency Instructions

Addressing the severe trade-off between life-threatening instruction errors and excessive refusal in emergency NLP, this paper constructs EMER-GuardBenchโ€”a 69,000-example counterfactual auditing benchmark with instruction-group-disjoint splitsโ€”and presents EGIS, an evidence-conditioned multi-task verifier that delivers 89.9% Harm F1 while eliminating benign-stress false positives entirely under the Severity-Weighted Miss Rate (SWMR) metric.

All Relations at Once: Efficient Open-World Tabular Relation Discovery with Hierarchy-Aware Reward Shaping

Addressing the core bottlenecks of closed-set label reliance, rigid binary penalties on near-miss predictions, and quadratic pair-wise inference latency in column relation extraction, this paper introduces an open-world reinforcement learning framework with hierarchy-aware reward shaping and span-level advantage assignment that generates all column-pair relations in a single structured response, markedly outperforming both domain-specific models and frontier LLMs.

SSR: Symbolic-Neural Soft Reasoning for Robust and Verifiable Thinking Chains

Addressing the catastrophic brittleness of deterministic symbolic solvers under imperfect LLM auto-formalization, SSR introduces confidence-weighted MaxSAT relaxation, ambiguity-resolving targeted retrieval, and bidirectional deduction tree generation, boosting multi-step reasoning accuracy and producing replayable, verifiable thinking chains.

IdioSteer: Controlling Figurative-vs-Literal Interpretation via Residual-Stream Activation Steering

Addressing large language models' deep-seated bias toward figurative interpretations of idioms during open-ended generation, this paper proposes residual-stream activation steering to achieve precise inference-time control: by injecting a cross-layer mean-difference vector at the prompt's final idiom token, models are systematically shifted toward literal continuations while maintaining fluency, with the steering direction transferring seamlessly to unseen idioms.

MelonBench: A Benchmark for Coded Entity Resolution over Noisy Web Evidence

Addressing the reliance of existing browsing benchmarks on explicit queries and clean authoritative sources, this paper presents MelonBench, a 665-item benchmark instantiated in Chinese entertainment blind-item discourse to evaluate search agents under intentional semantic obfuscation and noisy user-generated content, where the best search-augmented system achieves only 49.9% accuracy and reveals critical bottlenecks in semantic de-obfuscation and open-ended entity recall.

ObSchema: From Local Observations to Global Entity Attribute Schemas in Scientific Literature

Addressing scattered, heterogeneous, and condition-entangled entity properties in scientific literature, ObSchema introduces a batch-wise observation-to-schema framework that decouples semantic classification from deterministic evidence-based promotion, integrating value-aware duplicate merging and adaptive grouping to induce high-fidelity hierarchical attribute schemas.

Personalization over Generalization: History-Aware User Modeling for Dialogue Agents

Existing sentence-level AI detectors treat all authors identically and fail to capture individual writing boundaries; this paper proposes PHAS-ID, a personalized detector that grounds sentence predictions in an author's labeled writing history via dual cross-attention fusion, alongside GhostFill, a style-aware synthetic generation pipeline mirroring real collaborative interleaving patterns.

Measuring Source-Resolution Distortion in LLM Output Arbitration: A Source-Bounded Evaluation Framework

Addressing the pervasive loss and misattribution of critical source structures when LLMs arbitrate among multiple candidate answers, this paper formalizes Source-Resolution Distortion (SRD), introduces a source-bounded benchmark, and proposes A2Rโ€”a framework that builds a sparse decision ledger before deterministic rendering to substantially cut conflict erasure and boost provenance recoverability.

The Harder Text Embedding Benchmark (HTEB): Beyond One-dimensional Static Robustness

Addressing the fundamental blind spots of static, single-scalar embedding leaderboards like MTEB, this paper introduces HTEB, a dynamic evaluation framework that stochastically injects 8 LLM-generated transformations along Lexical/Stylistic, Length, and Language axes, demonstrating that embedding robustness is multidimensional and that parameter scaling fails to close robustness gaps.