@inproceedings{klila-etal-2026-injecting,
title = "Injecting Structured Biomedical Knowledge into Language Models:Continual Pretraining vs. {G}raph{RAG}",
author = "Klila, Jaafer and
Bannour Souihi, Sondes and
Boujelbane, Rahma and
Semmar, Nasredine and
Hadrich-Belguith, Lamia",
editor = "Piperidis, Stelios and
Bel, N{\'u}ria and
van den Heuvel, Henk and
Ide, Nancy and
Krek, Simon and
Toral, Antonio",
booktitle = "Proceedings of the Fifteenth Language Resources and Evaluation Conference",
month = may,
year = "2026",
address = "Palma de Mallorca, Spain",
publisher = "ELRA Language Resource Association",
url = "https://aclanthology.org/2026.lrec-1.855/",
doi = "10.63317/35ddm5i6bjyd",
pages = "10927--10936",
abstract = "The injection of domain-specific knowledge is crucial for adapting language models (LMs) to specialized fields such as biomedicine. While most current approaches rely on unstructured text corpora, this study explores two complementary strategies for leveraging structured knowledge from the UMLS Metathesaurus: (i) Continual pretraining that embeds knowledge into model parameters, and (ii) Graph Retrieval-Augmented Generation (GraphRAG) that consults a knowledge graph at inference time. We first construct a large-scale biomedical knowledge graph from UMLS (3.4 million concepts and 34.2 million relations), stored in Neo4j for efficient querying. We then derive a {\textasciitilde}100-million-token textual corpus from this graph to continually pretrain two models: BERTUMLS (from BERT) and BioBERTUMLS (from BioBERT). We evaluate these models on six BLURB (Biomedical Language Understanding and Reasoning Benchmark) datasets spanning five task types and evaluate GraphRAG on the two QA (Question Answering) datasets (PubMedQA, BioASQ). On BLURB tasks, BERTUMLS improves over BERT, with the largest gains on knowledge-intensive QA. Effects on BioBERT are more nuanced, suggesting diminishing returns when the base model already encodes substantial biomedical text knowledge. Finally, augmenting LLaMA 3-8B with our GraphRAG pipeline yields over than 3 points accuracy on PubMedQA and 5 points on BioASQ without any retraining, delivering transparent, multi-hop, and easily updated knowledge access. We release the processed UMLS Neo4j graph to support reproducibility."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="klila-etal-2026-injecting">
<titleInfo>
<title>Injecting Structured Biomedical Knowledge into Language Models:Continual Pretraining vs. GraphRAG</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jaafer</namePart>
<namePart type="family">Klila</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sondes</namePart>
<namePart type="family">Bannour Souihi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rahma</namePart>
<namePart type="family">Boujelbane</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nasredine</namePart>
<namePart type="family">Semmar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lamia</namePart>
<namePart type="family">Hadrich-Belguith</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Fifteenth Language Resources and Evaluation Conference</title>
</titleInfo>
<name type="personal">
<namePart type="given">Stelios</namePart>
<namePart type="family">Piperidis</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Núria</namePart>
<namePart type="family">Bel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Henk</namePart>
<namePart type="family">van den Heuvel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nancy</namePart>
<namePart type="family">Ide</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Simon</namePart>
<namePart type="family">Krek</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Antonio</namePart>
<namePart type="family">Toral</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>ELRA Language Resource Association</publisher>
<place>
<placeTerm type="text">Palma de Mallorca, Spain</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>The injection of domain-specific knowledge is crucial for adapting language models (LMs) to specialized fields such as biomedicine. While most current approaches rely on unstructured text corpora, this study explores two complementary strategies for leveraging structured knowledge from the UMLS Metathesaurus: (i) Continual pretraining that embeds knowledge into model parameters, and (ii) Graph Retrieval-Augmented Generation (GraphRAG) that consults a knowledge graph at inference time. We first construct a large-scale biomedical knowledge graph from UMLS (3.4 million concepts and 34.2 million relations), stored in Neo4j for efficient querying. We then derive a ~100-million-token textual corpus from this graph to continually pretrain two models: BERTUMLS (from BERT) and BioBERTUMLS (from BioBERT). We evaluate these models on six BLURB (Biomedical Language Understanding and Reasoning Benchmark) datasets spanning five task types and evaluate GraphRAG on the two QA (Question Answering) datasets (PubMedQA, BioASQ). On BLURB tasks, BERTUMLS improves over BERT, with the largest gains on knowledge-intensive QA. Effects on BioBERT are more nuanced, suggesting diminishing returns when the base model already encodes substantial biomedical text knowledge. Finally, augmenting LLaMA 3-8B with our GraphRAG pipeline yields over than 3 points accuracy on PubMedQA and 5 points on BioASQ without any retraining, delivering transparent, multi-hop, and easily updated knowledge access. We release the processed UMLS Neo4j graph to support reproducibility.</abstract>
<identifier type="citekey">klila-etal-2026-injecting</identifier>
<identifier type="doi">10.63317/35ddm5i6bjyd</identifier>
<location>
<url>https://aclanthology.org/2026.lrec-1.855/</url>
</location>
<part>
<date>2026-05</date>
<extent unit="page">
<start>10927</start>
<end>10936</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Injecting Structured Biomedical Knowledge into Language Models:Continual Pretraining vs. GraphRAG
%A Klila, Jaafer
%A Bannour Souihi, Sondes
%A Boujelbane, Rahma
%A Semmar, Nasredine
%A Hadrich-Belguith, Lamia
%Y Piperidis, Stelios
%Y Bel, Núria
%Y van den Heuvel, Henk
%Y Ide, Nancy
%Y Krek, Simon
%Y Toral, Antonio
%S Proceedings of the Fifteenth Language Resources and Evaluation Conference
%D 2026
%8 May
%I ELRA Language Resource Association
%C Palma de Mallorca, Spain
%F klila-etal-2026-injecting
%X The injection of domain-specific knowledge is crucial for adapting language models (LMs) to specialized fields such as biomedicine. While most current approaches rely on unstructured text corpora, this study explores two complementary strategies for leveraging structured knowledge from the UMLS Metathesaurus: (i) Continual pretraining that embeds knowledge into model parameters, and (ii) Graph Retrieval-Augmented Generation (GraphRAG) that consults a knowledge graph at inference time. We first construct a large-scale biomedical knowledge graph from UMLS (3.4 million concepts and 34.2 million relations), stored in Neo4j for efficient querying. We then derive a ~100-million-token textual corpus from this graph to continually pretrain two models: BERTUMLS (from BERT) and BioBERTUMLS (from BioBERT). We evaluate these models on six BLURB (Biomedical Language Understanding and Reasoning Benchmark) datasets spanning five task types and evaluate GraphRAG on the two QA (Question Answering) datasets (PubMedQA, BioASQ). On BLURB tasks, BERTUMLS improves over BERT, with the largest gains on knowledge-intensive QA. Effects on BioBERT are more nuanced, suggesting diminishing returns when the base model already encodes substantial biomedical text knowledge. Finally, augmenting LLaMA 3-8B with our GraphRAG pipeline yields over than 3 points accuracy on PubMedQA and 5 points on BioASQ without any retraining, delivering transparent, multi-hop, and easily updated knowledge access. We release the processed UMLS Neo4j graph to support reproducibility.
%R 10.63317/35ddm5i6bjyd
%U https://aclanthology.org/2026.lrec-1.855/
%U https://doi.org/10.63317/35ddm5i6bjyd
%P 10927-10936
Markdown (Informal)
[Injecting Structured Biomedical Knowledge into Language Models:Continual Pretraining vs. GraphRAG](https://aclanthology.org/2026.lrec-1.855/) (Klila et al., LREC 2026)
ACL