arXiv:2511.08029v2 Announce Type: replace-cross 
Abstract: Hard negatives are essential for training effective retrieval models. Hard-negative mining typically relies on ranking documents using cross-encoders or static embedding models based on similarity metrics such as cosine distance. Hard negative mining becomes challenging for biomedical and scientific domains due to the difficulty in distinguishing between source and hard negative documents. However, referenced documents naturally share contextual relevance with the source document but are not duplicates, making them well-suited as hard negatives. In this work, we propose BiCA: Biomedical Dense Retrieval with Citation-Aware Hard Negatives, an approach for hard-negative mining by utilizing citation links in 20,000 PubMed articles for improving a domain-specific small dense retriever. We fine-tune the GTE_small and GTE_Base models using these citation-informed negatives and observe consistent improvements in zero-shot dense retrieval using nDCG@10 for both in-domain and out-of-domain tasks on BEIR and outperform baselines on long-tailed topics in LoTTE using Success@5. Our findings highlight the potential of leveraging document link structure to generate highly informative negatives, enabling state-of-the-art performance with minimal fine-tuning and demonstrating a path towards highly data-efficient domain adaptation.

تم تقديم نهج جديد يسمى BiCA لاسترجاع البيانات الكثيفة في المجال الطبي الحيوي، مستفيدًا من السلبيات الصعبة الواعية بالاستشهادات من 20,000 مقال في PubMed لتحسين تدريب نماذج الاسترجاع. تعالج هذه الطريقة التحديات المتعلقة بتعدين السلبيات الصعبة في المجال الطبي الحيوي من خلال استخدام مستندات تشارك في الصلة السياقية دون أن تكون مكررة.

Se ha introducido un nuevo enfoque llamado BiCA para la recuperación densa biomédica, aprovechando negativos difíciles conscientes de citas de 20,000 artículos de PubMed para mejorar el entrenamiento de modelos de recuperación. Este método aborda los desafíos del minado de negativos difíciles en el campo biomédico al utilizar documentos que comparten relevancia contextual sin ser duplicados.

Une nouvelle approche appelée BiCA a été introduite pour la récupération dense biomédicale, en s'appuyant sur des négatifs difficiles conscients des citations provenant de 20 000 articles de PubMed pour améliorer l'entraînement des modèles de récupération. Cette méthode répond aux défis du minage de négatifs difficiles dans le domaine biomédical en utilisant des documents partageant une pertinence contextuelle sans être des doublons.

A new approach called BiCA has been introduced for biomedical dense retrieval, leveraging citation-aware hard negatives from 20,000 PubMed articles to enhance the training of retrieval models. This method addresses the challenges of hard-negative mining in the biomedical field by utilizing documents that share contextual relevance without being duplicates.

BiCA: Effective Biomedical Dense Retrieval with Citation-Aware Hard Negatives

arXiv:2505.07917v2 Announce Type: replace-cross 
Abstract: Biomedical question-answering (QA) systems require effective retrieval and generation components to ensure accuracy, efficiency, and scalability. This study systematically examines a Retrieval-Augmented Generation (RAG) system for biomedical QA, evaluating retrieval strategies and response time trade-offs. We first assess state-of-the-art retrieval methods, including BM25, BioBERT, MedCPT, and a hybrid approach, alongside common data stores such as Elasticsearch, MongoDB, and FAISS, on a ~10% subset of PubMed (2.4M documents) to measure indexing efficiency, retrieval latency, and retriever performance in the end-to-end RAG system. Based on these insights, we deploy the final RAG system on the full 24M PubMed corpus, comparing different retrievers' impact on overall performance. Evaluations of the retrieval depth show that retrieving 50 documents with BM25 before reranking with MedCPT optimally balances accuracy (0.90), recall (0.90), and response time (1.91s). BM25 retrieval time remains stable (82ms), while MedCPT incurs the main computational cost. These results highlight previously not well-known trade-offs in retrieval depth, efficiency, and scalability for biomedical QA. With open-source code, the system is fully reproducible and extensible.

قدمت دراسة نظامًا لتوليد الأسئلة البيولوجية المعززة بالاسترجاع (RAG) ، حيث تم تقييم استراتيجيات الاسترجاع المختلفة وكفاءتها على مجموعة فرعية من بيانات PubMed. تقيم الدراسة طرقًا متطورة مثل BM25 و BioBERT ، وتقيس كفاءة الفهرسة وأداء الاسترجاع قبل نشر النظام على مجموعة بيانات PubMed الكاملة.

Un estudio ha presentado un sistema de Generación Aumentada por Recuperación (RAG) para la respuesta a preguntas biomédicas, evaluando diversas estrategias de recuperación y su eficiencia en un subconjunto de datos de PubMed. La investigación evalúa métodos de vanguardia como BM25 y BioBERT, midiendo la eficiencia de indexación y el rendimiento de recuperación antes de implementar el sistema en el corpus completo de PubMed.

Une étude a introduit un système de génération augmentée par récupération (RAG) pour le questionnement biomédical, évaluant diverses stratégies de récupération et leur efficacité sur un sous-ensemble de données de PubMed. La recherche évalue des méthodes à la pointe comme BM25 et BioBERT, mesurant l'efficacité de l'indexation et la performance de récupération avant de déployer le système sur l'ensemble du corpus de PubMed.

A study has introduced a Retrieval-Augmented Generation (RAG) system for biomedical question answering, evaluating various retrieval strategies and their efficiency on a subset of PubMed data. The research assesses state-of-the-art methods like BM25 and BioBERT, measuring indexing efficiency and retrieval performance before deploying the system on the full PubMed corpus.

BiCA: Effective Biomedical Dense Retrieval with Citation-Aware Hard Negatives

Was this article worth reading? Share it

Airparser

Humanize AI

Sourcely

Meteoria

Grasp.info

ZeroGPT.org

Ready to build your own newsroom?