Jul 2026· Annual International Computer Software and Applications Conference· pp. 1813-1818· 0 citations· 31 references
Abstract
Adapting large language models (LLMs) to new domains using techniques like pre-training or fine-tuning can be computationally intensive. This challenge has motivated researchers and practitioners to explore alternative strategies, such as Retrieval-based augmentation mechanisms, to reduce training cost while preserving model adaptability. These mechanisms enable LLMs to generate responses grounded in external and previously unseen data. Among these approaches, RetrievalAugmented Generation (RAG) is the widely adopted due to its relatively low computational overhead. However, conventional RAG pipelines exhibit some limitations, including weak semantic coherence across retrieved contexts, limited structural reasoning capability, and the absence of explicit relational representations. To overcome these limitations, recent research has shifted toward the integration of Knowledge Graphs (KGs), which provide structured semantic representations and explicit relational modeling. KG-based approaches looks promising but the typically depend on manually designed ontologies, resulting in scalability constraints and domain adaptation challenges. To address these limitations, we propose OF-BIO-KG, a framework for automated, ontology-free knowledge graph construction. Our approach begins by projecting domain data into a vector database to identify semantically coherent clusters. An LLM then labels these clusters as conceptual nodes and infers both hierarchical and cross-cluster relationships. These nodes and links are persisted in a graph database. By eliminating the need for predefined schemas, our framework offers a scalable, computationally efficient solution for domain-specific LLM adaptation.
Knowledge Organization Systems like Ontologies and taxonomies are fundamental for structuring scientific knowledge, yet their manual curation presents a persistent bottleneck in knowledge management. While Large Language Models (LLMs) offer a scalable mechanism for automated ontology generation, their capacity to classify complex, domain-specific semantics requires systematic evaluation. In this paper, we assess the performance of five small, open-source LLMs (up to 9 billion parameters) in identifying semantic relationships between biomedical concepts. To support this evaluation, we introduce MeSH-Rel-4K, a dataset comprising 4K semantic relationships extracted from the Medical Subject Headings (MeSH). We analyse three adaptation strategies: standard prompting, Chain-of-Thought prompting, and fine-tuning. While parameter-constrained models traditionally struggle with the nuances of in-context logic, our results reveal that targeted fine-tuning increases the average F1-score by 34.1 percentage points. These results confirm that direct fine-tuning effectively exceeds the reasoning bottlenecks of smaller LLMs, providing an accurate, automated methodology for the construction and evolution of specialised biomedical ontologies.
Tanay Aggarwal, Angelo Salatino, Francesco Osborne et al.· 0 citations
Biomedical knowledge discovery and clinical natural language processing increasingly rely on foundation models, but these approaches face persistent challenges, including incomplete curated knowledge, instability in language model outputs, and ambiguity in clinical interpretation. This dissertation argues that reliable biomedical AI requires integrating structured knowledge and foundation models through complementary control mechanisms rather than relying on any single method. To support this argument, the dissertation presents a unified framework across molecular and clinical domains. It develops an iterative prompt refinement approach with retrieval-augmented generation to improve the reliability of biomedical relation extraction, introduces graph transformer models to infer gene–gene relations and augment incomplete pathway knowledge, and proposes ontology-grounded retrieval and normalization pipelines to improve semantic alignment in multidisciplinary clinical notes. Together, these contributions define a layered reliability framework in which graph-based inference addresses knowledge incompleteness, retrieval-augmented prompt control mitigates instability, and ontology grounding reduces semantic ambiguity, providing a foundation for more reliable biomedical AI systems.
Biomedical entity linking (BioEL) aims to map textual mentions to standardized concepts in biomedical ontologies. While existing methods have made significant progress, most rely on binary supervision, indiscriminately penalizing all non-gold candidates and overlooking the rich hierarchical structure inherent in biomedical taxonomies. This restricts their ability to capture nuanced semantic relationships and adapt to varying degrees of entity similarity. We propose Taxonomy-Enhanced Entity Linking (TaxEL), a novel framework that unifies taxonomy-guided candidate sampling and structure-aware distributional supervision for BioEL. Specifically, TaxEL introduces: (1) Taxonomy-Guided Contrastive Sampling (TGCS), which systematically integrates both local ontology structure and global semantic similarity to generate informative positive and hard negative samples for each mention; and (2) Structured Semantic Alignment Loss (SSAL), which enforces alignment between model predictions and fine-grained semantic distributions derived from the taxonomy, enabling explicit control over prediction granularity. Experimental results on five public BioEL benchmarks demonstrate that TaxEL achieves state-of-the-art performance in Acc@1, and ablation studies confirm the critical role of both TGCS and SSAL. The TaxEL web service is publicly accessible at http://www.el.tcmkg.com, and all associated data and code can be obtained from https://github.com/TCMAI-BJTU/TaxEL.
Rui Hua, Zeyu Liu, Zixin Shu et al.· IEEE journal of biomedical a...· 0 citations
The exponential growth of biomedical regulatory documentation has created an urgent need for automated, scalable knowledge structuring mechanisms that do not rely on rigid, expert-curated ontologies. This paper presents the system implementation and empirical analysis of the Ontology-Free Biomedical Knowledge Graph Induction (OF-Bio-KG) framework, focusing on three dimensions: (i) the end-to-end computational pipeline, (ii) the structural and statistical properties of the induced knowledge graph, and (iii) a proposed crowd-sourced expert evaluation strategy intended for future deployment. The OF-Bio-KG framework operates on a corpus of 1,068 European Medicines Agency (EMA) pharmaceutical product leaflets totalling over 131 million tokens. Using the nomic-embed-text model for long-context semantic encoding, a Balanced K-Means Constrained clustering algorithm for thematic organization, and LLaMA-3 8B for ontology-free relational inference, the system produces a multi-level knowledge graph comprising 256,148 leaf nodes, 3,742 cluster-level concept nodes, over 1.6 million intra-cluster edges, and 147,000 inter-cluster edges, all persisted in Neo4j. A central challenge in this work is the absence of a groundtruth biomedical knowledge graph against which the induced graph can be evaluated using standard metrics. To address this, we propose a human-in-the-loop validation strategy planned for future deployment: a purpose-built web application will be made available to Master's students at ENSIT, who will act as domain-proficient annotators to explore and validate graph components. This paper describes the current system, its empirical results, and the design of this forthcoming evaluation platform.
Ahlem Ben Younes, Baha Eddine Kalai, Laila Ben Ayed et al.· Annual International Compute...· 0 citations
Knowledge graph question answering (KGQA) focuses on answering questions with data retrieved from knowledge graphs (KGs). Although Large Language Models (LLMs) have enabled the development of innovative KGQA systems, serious challenges remain. Approaches that utilize supervised finetuning are accurate and responsive, but also require large training datasets that provide good coverage of both the target knowledge graph and expected user questions. Likewise, integrating KG-specific optimizations improves accuracy and responsiveness but limits versatility and is susceptible to updates to the KG's ontology. LLMs have facilitated the development of training-free and KG-agnostic solutions that utilize reasoning and real-time decision making to explore the target KG, but the semi-structured nature of KGs makes this a complex and unpredictable task, leading to performance degradation and longer response times. In this paper we present PYTHIA, a training-free, plug-and-play solution for KGQA over any RDF KG. At its core, PYTHIA consists of an LLM agent guided by a relation-centric conceptual model of the KGQA task which is acted upon through a suite of tools for entity linking, graph exploration and query execution. The agent is responsible for interpreting user intent, orchestrating tool calls, and evaluating tool output, while the tools themselves present simple interfaces that abstract complex processes, reducing cognitive load on the agent. To assess its viability, we perform a comprehensive evaluation of PYTHIA and prior systems across four knowledge graphs and six datasets. The results show PYTHIA achieving a median performance improvement of 15.31% while improving cost and time efficiency at the performance frontier.
Sergios-Anestis Kefalidis, Konstantinos Plas, Manolis Koubarakis· Proceedings of the 32nd ACM...· 0 citations
Integrating newly emerging terms into existing ontologies is a recurring maintenance problem in knowledge engineering, particularly in biomedical domains where terminology evolves faster than manual curation can accommodate. This paper addresses the placement of concepts that are absent from the target ontology—the out-of-knowledge-base setting—in which a textual mention must be assigned one or more insertion positions in the subsumption hierarchy rather than linked to an existing node. We propose a three-stage framework that extends the conventional retrieve-then-select paradigm with an intermediate stage of edge generation and enrichment, which expands the candidate set by traversing the local structure of the ontology. Stage 1 retrieves candidate edges using a fine-tuned bi-encoder trained with a max-margin objective; Stage 2 constructs and structurally enriches candidate edges; Stage 3 selects among them using either a fine-tuned cross-encoder or a large language model under explainable instruction tuning. We evaluate on two datasets derived from SNOMED CT, MM-S14-Disease and MM-S14-CPP, under a strict out-of-knowledge-base protocol. Fine-tuned pre-trained language models outperform zero-shot and instruction-tuned large language models on ranking accuracy, while the instruction-tuned configuration produces expert-auditable justifications at a modest cost in accuracy. On MM-S14-Disease, the strongest configuration places a correct insertion edge among the ten highest-ranked candidates for 38.7% of test mentions and recovers the complete gold edge set for 16.4%, against 26.1% and 9.2% for retrieval alone. The framework is positioned as decision support for ontology curators rather than as an autonomous ontology generator.
Zhanna B. Sadirmekova, M. Sambetbayeva, B. Abdygalym et al.· Information· 0 citations