Lightweight Person-Place Relation Extraction from Historical Newspapers with Dependency Graphs and Proximity Features

DistilledGemma: Balanced Efficiency-Accuracy for Person-Place Relation Extraction from Multilingual Historical Articles

We present DistilledGemma, an efficient and accurate system for the HIPE-2026 shared task on person-place relation extraction from multilingual historical newspaper articles in English, German, and French. Our approach adopts a three-stage knowledge distillation pipeline designed to balance classification accuracy with computational efficiency. In the first stage, we systematically explored prompt engineering strategies across eight large language models to identify the most effective reasoning architecture for this challenging task. In the second stage, we applied supervised fine-tuning (SFT) via QLoRA to a Gemma 4 26B A4B teacher model, leveraging its strong multilingual capabilities to generate silver-standard chain-of-thought traces across the training corpus. In the final stage, we performed response-level distillation to transfer these learned reasoning patterns into a compact Gemma 4 E2B student model. In the official evaluation, our team WHEREAMI ranked 3rd on the standard test set with an accuracy profile mean score of 0.688, and 2nd on the binary test set with a mean score of 0.8156. Notably, by distilling knowledge from the 26B teacher to the 2.3B student, we preserved strong reasoning capabilities while reducing the deployed model size to approximately 2.3B effective parameters; the LoRA adapters used during training were merged into the student for inference. This configuration ranked 2nd in the balanced efficiency-accuracy profile across both the standard and binary test sets. These results demonstrate that knowledge distillation provides a practical and scalable solution for historical document processing, achieving competitive performance without excessive computational cost.

ReTaT: A Unified Benchmark for Relation Extraction across Text and Table

LCDS: Logic-Constrained Distant Supervision for Document-Level Relation Extraction

Document-level relation extraction (DocRE) identifies semantic relations between entity pairs in a document, but its progress is limited by the scarcity of high-quality annotations. Distant supervision alleviates this issue by automatically constructing large-scale training data, yet it also introduces noisy and incomplete labels. To address these problems, we propose LCDS, a logic-constrained distant supervision framework for DocRE. LCDS first applies Monte Carlo dropout to estimate prediction uncertainty and filter unreliable pseudo labels. It then mines high-confidence logical rules from human-annotated data and uses them to complement missing relations and correct conflicts. Through iterative training, LCDS progressively improves pseudo-label quality. Experiments on DocRED and Re-DocRED show that LCDS achieves F1 scores of 67.39% and $\mathbf{8 1. 4 2 \%}$, respectively, outperforming previous state-of-theart models and demonstrating the effectiveness of combining uncertainty estimation with logical reasoning.

Optimizing sample selection for large language model-based entity matching using AssistEM

The meteoric rise of large language models (LLMs) has reshaped natural language processing, inspiring new approaches to data integration tasks such as entity matching (EM). While proprietary models like GPT-4 deliver strong performance, open-source alternatives (e.g., Mistral, DeciLM, Gemma-3) require supervised fine-tuning (SFT) to excel in specialized domains. However, naïvely training billion-parameter scale LLMs on uncurated corpora is computationally prohibitive and environmentally costly, often taking days while introducing inefficiencies that limit rapid application deployment. We introduce AssistEM, a framework for efficient LLM adaptation to EM via principled data selection. AssistEM integrates NER and TF-IDF signals into a composite score, filters entity pairs with domain-level thresholds, constructs denoised semantic embeddings, and applies graph-based clustering with class-aware sampling to obtain a compact yet representative set of ∼\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim $$\end{document}8000 high-quality training instances. This yields rapid specialization: AssistEM-trained LLMs converge in under 4.5 hours–versus 12+ hours for DeciLM-7B and 24+ hours for Yi-34B–while surpassing GPT-4-0613 by 4.3 F1 points and Ditto by 11 F1 points across diverse EM benchmarks. By aligning data quality with model specialization, AssistEM demonstrates that selective fine-tuning not only accelerates adaptation but also improves training efficiency (requiring fewer GPU hours), enabling open-source LLMs to rival–and in some cases outperform–closed-source models. These results highlight data selection as a critical lever for sustainable, domain-specific LLM deployment.

A unified transformer-BiLSTM and graph attention network framework for explainable multilingual opinion mining and relationship inference in complex social-citation networks

The paper takes into consideration the rising requirement for accurate mining of opinion and inferring relations among entities as the amount of multilingual online information increases rapidly. Thus, the paper seeks a uniform statistical learning methodology for processing multiple languages and exploiting valuable relational discoveries. The paper introduces an innovative model which combines transformer-based context embedding, BiLSTM for capturing of sentiment flows, and GAT for examining relational data. Transformers model cross-lingual context probability distribution, BiLSTM models temporal transitions in sentiment, whereas GAT offers an attention-based statistical weighting on data structured as a graph. Examples include social networks and citation graphs. The model yields a sentiment classification accuracy and macro F1-score of 92.4% over the four multilingual datasets (English, Spanish, and Hindi), surpasses all baseline methods. The model has achieved state-of-the-art result 86.4% in relation inference with graph attention mechanism, and shows excellent reliability and generalization. In this work, we propose a statistically sound and integrated framework which incorporates contextual, sequential and relational modelling of multilingual opinion mining. Graph structured learning integrated with probabilistic encoding and temporal modeling is a novel technique for such tasks, which also shows a promising direction in terms of scalability and generalizability across different domains.

Semantic Structure and Importance Extraction from Sequential Conversational Data via Dimensional Reduction

The analysis of spoken data from panel discussions, policy dialogues, and educational meetings has gained increasing importance in both academic research and professional practice. However, conventional approaches to Japanese conversation analysis have relied heavily on keyword matching or surface‑level text similarity, making it difficult to capture deeper semantic relationships, topic transitions, and latent discourse structures. In addition, Japanese natural language processing pipelines often rely on environment-sensitive morphological analyzers, which hinder reproducibility and large-scale processing. To address these limitations, this study proposes a robust and semantically enriched framework for conversation understanding based on a composite distributed representation. The proposed method integrates three layers of linguistic information: (1) contextual sentence embeddings generated by a multilingual transformer model, (2) word embeddings obtained from fastText, and (3) co‑occurrence vectors that capture lexical association patterns within the conversation. Sudachi is employed for Japanese text preprocessing to ensure stable and reproducible morphological analysis. By combining these components into a unified composite vector, the framework simultaneously represents global sentence‑level meaning and local lexical relationships. Using this representation, a directed graph is constructed that incorporates both temporal adjacency and semantic proximity between utterances, enabling the visualization of key conversational connections. To evaluate the effectiveness of the composite representation, dimensionality‑reduction algorithms are applied to examine whether semantically similar utterances naturally form coherent clusters in low‑dimensional space. The resulting clusters are assessed for consistency and interpretability, demonstrating that the proposed representation successfully captures meaningful conversational structure.