MA-COIR: Leveraging Semantic Search Index and Generative Models for Ontology-Driven Biomedical Concept Recognition

TL;DR

MA-COIR addresses the challenge of recognizing biomedical concepts, including those not explicitly mentioned, by reframing recognition as an indexing-recognition task and assigning semantic search indexes (ssIDs) to ontology concepts. The framework leverages a BART-based encoder-decoder with a constrained decoder to generate ssID sequences, and augments learning with LLM-generated queries and semi-supervised data via data augmentation. It introduces multi-level queries derived from passages (including claims and concept names) to improve recognition in low-resource settings. Experimental results on CDR, HPO, and HOIP demonstrate strong performance, particularly with passage-level inputs, and show the benefits of ssID indexing, data augmentation, and combining information from multiple query levels for robust ontology-driven concept recognition. The work offers a practical pathway toward scalable, mention-agnostic biomedical concept extraction with reduced annotation requirements and improved adaptability to diverse ontologies.

Abstract

Recognizing biomedical concepts in the text is vital for ontology refinement, knowledge graph construction, and concept relationship discovery. However, traditional concept recognition methods, relying on explicit mention identification, often fail to capture complex concepts not explicitly stated in the text. To overcome this limitation, we introduce MA-COIR, a framework that reformulates concept recognition as an indexing-recognition task. By assigning semantic search indexes (ssIDs) to concepts, MA-COIR resolves ambiguities in ontology entries and enhances recognition efficiency. Using a pretrained BART-based model fine-tuned on small datasets, our approach reduces computational requirements to facilitate adoption by domain experts. Furthermore, we incorporate large language models (LLMs)-generated queries and synthetic data to improve recognition in low-resource settings. Experimental results on three scenarios (CDR, HPO, and HOIP) highlight the effectiveness of MA-COIR in recognizing both explicit and implicit concepts without the need for mention-level annotations during inference, advancing ontology-driven concept recognition in biomedical domain applications. Our code and constructed data are available at https://github.com/sl-633/macoir-master.

Figures (5)

• Figure 1: Concept recognition by MA-COIR follows the default workflow indicated by purple arrows. When an LLM generates simplified queries from a given passage, additional processes, denoted by blue arrows, are incorporated. When "6-2-8-0-5" is generated, "HOIP_0004832: TNF signalling" is predicted as a concept within the query.
• Figure 2: Indexing Phase in MA-COIR: A semantic search index (ssID) is assigned based on a label tree derived from the domain ontology. Through hierarchical clustering, the ssID for the concept "HOIP_0004832: TNF signaling" is "6-2-8-0-5".
• Figure 3: An example of constructing a claim-concept instance is as follows: Given a passage, we prompt the LLM to breakdown the passage into several claims. For one claim, we then perform excerpt mining. Next, we match these mined excerpts to the passage’s annotated concepts by assessing semantic similarity. If an excerpt closely aligns with an annotated concept, we pair the concept with the claim. In this example, seven concepts are paired with a single claim, forming a claim-concept instance.
• Figure 4: F1 scores by MA-COIR between complex query (passage) and the average of the simpler set of queries (claim/concept) from top-1 generated sequence using different indexes on the HOIP.
• Figure 5: Prompt template for generating concept names / claims for passage. A prompt consists of task instruction, output format instruction, several demonstrations and the query.