Source-linked AI summary
Configurable Semantic Chunking for Biomedical Information Extraction in Retrieval-Augmented Generation
Riya Ahuja, Tim Kacprowski, Roya Shiasi Sardoabi
TL;DR
Fixed-size chunking in BioMedRAG can fragment biomedical entities and relation-bearing evidence, motivating a more configurable approach. The paper replaces only chunk construction with semantic candidate generation and evaluates it across four biomedical benchmarks, finding gains on explicit-cue extraction tasks but competitive fixed chunking in denser or classification-oriented settings.
Problem
BioMedRAG’s fixed-size chunking can fragment semantic evidence and produce incomplete or misaligned retrieval context for biomedical information extraction.
Method
The framework combines entity-preserving windows, trigger-aware and proposition-first candidates, tiered prioritization, and relation hierarchies while preserving BioMedRAG’s other components.
Results
Semantic chunking improves extraction-focused benchmarks, achieving +8.4 F1 points on GM-CIHT (82.6%) and +1.0 F1 on DDI (79.2%), while fixed-size chunking remains competitive for ChemProt and ADE.
Takeaways & Limitations
Configuration files make chunking decisions explicit and adaptable, while the reported benefits are strongest for coarse-grained, trigger-explicit relations with moderate entity density.
Takeaways & Limitations
The conclusions are best interpreted for sentence-level biomedical relation extraction over published literature, and transfer requires task-specific symbolic resources.
Abstract
from arXiv · showhide
BioMedRAG introduced retrieval-augmented generation with a learned chunk scorer for biomedical information extraction. However, it relies on fixed-size chunking which can fragment semantic evidence. We propose a configurable semantic chunking framework that addresses this limitation by combining entity-preserving windows, trigger-centered chunking, proposition-first extraction, tiered trigger prioritization, and hierarchical relation resolution. The framework integrates with BioMedRAG by replacing only the chunk construction stage while preserving the embedding model, learned chunk scorer, generator, and evaluation protocol. We evaluate the framework on biomedical relation extraction benchmarks (GM-CIHT, DDI, ChemProt) and adverse event classification (ADE). On GM-CIHT, the full hybrid configuration achieves 82.6% F1, improving over the fixed-size baseline (74.2% F1) by 8.4 points under our experimental setup. Cross-dataset analysis shows that semantic chunking improves extraction datasets with explicit relation cues, such as GM-CIHT and DDI, while fixed chunking remains competitive or stronger for dense biochemical extraction and binary classification settings such as ChemProt and ADE. By externalizing chunking logic into configuration files, the framework provides an interpretable and adaptable alternative to rigid fixed-size chunking for biomedical RAG pipelines.
1 Introduction
The paper addresses semantic fragmentation in BioMedRAG’s fixed-size chunking by introducing configurable semantic chunking for biomedical information extraction. It evaluates the framework across four benchmarks and reports that its benefits depend on task characteristics.
- Significance: The framework targets reliable biomedical knowledge extraction relevant to clinical decision support, drug discovery, precision medicine, and translational medicine.The paper connects improved retrieval quality with automated extraction of drug-disease, gene-protein, and adverse-event information.
- Motivation: Fixed 5-word windows can separate relation-bearing expressions, reducing retrieval precision and forcing the language model to infer missing context.For example, fixed chunking can split a drug mechanism from its target.
- Approach: The proposed framework combines entity-aware segmentation, trigger-centered extraction, tiered trigger prioritization, hierarchical relation resolution, and proposition-first extraction.A unified scoring mechanism balances trigger confidence, relation specificity, and contextual alignment.
- Evaluation: The evaluation covers GM-CIHT, DDI, ChemProt, and ADE, with semantic chunking most effective for extraction tasks containing explicit relation cues and moderate entity density.Component-wise ablations examine segmentation, proposition extraction, trigger prioritization, relation resolution, ranking, and entity detection.
- Approach: The framework replaces fixed-size windows with entity-preserving, trigger-aware, and proposition-oriented evidence candidates.It is integrated into BioMedRAG while preserving the embedding model, learned chunk scorer, generator, preprocessing pipeline, and evaluation protocol.
2 Related Work
Related work establishes RAG as a retrieval-and-generation paradigm and identifies chunking granularity as important for retrieval quality. Existing chunking approaches do not specifically enforce biomedical entity preservation and relation-trigger modeling.
- Retrieval-Augmented Generation: RAG retrieves relevant passages and conditions generation on that evidence to improve factual accuracy and interpretability.The paradigm has expanded from open-domain question answering to structured tasks, including biomedical relation extraction.
- Biomedical RAG: BioMedRAG applies RAG to biomedical relation extraction, using retrieved evidence to support structured predictions from current literature.The paper extends this system rather than replacing its overall retrieval-augmented design.
- Chunking Strategies: Chunking strategies influence RAG retrieval quality, with prior work comparing document, passage, sentence, and proposition-level retrieval units.Fixed-size and sliding-window methods provide uniform or overlapping context but are not tailored to biomedical relations.
- Chunking Strategies: Proposition-based methods decompose text into atomic units, while contextual chunk embeddings use long-context models.These approaches are largely domain-agnostic and do not enforce biomedical entity preservation or explicitly model relation triggers.
3 Method
The method replaces BioMedRAG’s fixed chunk-construction stage with configurable candidates that preserve entities and relation structure. Candidate generation combines semantic sources, followed by similarity ranking with proposition bias while leaving downstream components unchanged.
- BioMedRAG Pipeline: BioMedRAG originally splits each sentence into non-overlapping fixed 5-word windows, which can fragment entities and relation-bearing propositions.The proposed method changes this chunk-construction stage while retaining the embedding model, learned chunk scorer, generator, and evaluation protocol.
- Problem Formulation: The framework addresses triple extraction and relation classification by identifying relation-bearing evidence within biomedical sentences.Triple extraction produces head-relation-tail structures, while classification predicts the relation between marked entity spans.
- Candidate Generation: A hybrid candidate pool combines entity-aware sliding windows, proposition-first spans, and fixed-width fallback windows.Fallback windows preserve non-empty coverage when entity or trigger signals are unavailable.
- Candidate Selection: Candidates are ranked by cosine similarity to the target relation definition, then proposition candidates in the top-N are promoted into the final top-k set.This introduces an explicit preference for structurally complete relation spans.
- Configurability: Dataset-specific triggers, tiers, relation hierarchies, negation rules, and context markers are stored in JSON configuration files.The configuration-driven design makes chunking decisions inspectable and supports adaptation without code changes.
3.4 Entity-Aware Segmentation
Entity-aware segmentation preserves biomedical entity boundaries before trigger-based methods focus on relation-bearing content. It uses lightweight span detection and adjusted sliding windows, but remains dependent on the quality of trigger and entity signals.
- Entity Preservation: Biomedical entities often span multiple tokens or contain hyphenation, so fixed-width chunking can produce partial strings and degrade retrieval evidence.Examples include alpha-ketoglutarate and acetyl-CoA carboxylase.
- Entity Detection: Lightweight pattern rules detect bracketed markup, capitalized multi-token terms, and hyphenated compounds as character spans.These spans are used to adjust boundaries and construct proposition candidates rather than to perform complete named entity recognition.
- Boundary Adjustment: Window ends intersecting entities are extended, while window starts intersecting entities are shifted past them.This boundary adjustment guarantees that detected entities are not split across chunks.
- Boundary Adjustment: The method keeps a hyphenated entity such as tumor necrosis factor alpha receptor-complex intact within one chunk.This contrasts with a fixed 5-token window that can divide receptor- and complex across adjacent chunks.
- Trigger-Based Extension: Entity-aware segmentation preserves boundaries but does not determine whether a chunk expresses a semantic relation.Trigger-based detection is therefore added to emphasize relation-bearing content.
- Trigger-Based Extension: Trigger candidates are learned from annotated entity pairs and prioritized by relation specificity, with fallback to entity-aware and fixed-width chunks when no trigger is detected.Trigger-centered windows use ±4 words around detected triggers to capture nearby entities and modifiers.
3.6 Tiered Trigger Prioritization
Tiered trigger prioritization ranks relation triggers by semantic specificity and directional clarity, favoring precise relational evidence during proposition extraction.
- 3.6 Tiered Trigger Prioritization: Triggers are organized into three tiers according to semantic specificity and directional clarity.Primary triggers are explicit actions, secondary triggers are moderately specific, and tertiary triggers are generic or weaker signals.
- 3.6 Tiered Trigger Prioritization: Primary triggers receive bonus 3, secondary triggers bonus 2, and tertiary triggers bonus 1 during proposition ranking.Examples include inhibits as primary, blocks as secondary, and medication as tertiary.
- 3.6 Tiered Trigger Prioritization: When propositions overlap the same text span, the proposition containing the higher-tier trigger is retained.This passes the most semantically precise relational evidence to the selection stage.
3.7 Relation Hierarchy Resolution
Relation hierarchy resolution disambiguates sentences expressing multiple biomedical relations by applying configurable, task-specific priorities alongside trigger strength.
- 3.7 Relation Hierarchy Resolution: Sentences can express multiple relations simultaneously, creating ambiguity about which proposition best serves the extraction objective.The example combines TREATS for aspirin–inflammation with INHIBITS for aspirin–COX-2.
- 3.7 Relation Hierarchy Resolution: Configurable relation hierarchies encode dataset- and task-specific priorities for resolving competing propositions.For GM-CIHT, the hierarchy is defined in Table 1.
- 3.7 Relation Hierarchy Resolution: GM-CIHT prioritizes TREATS above INHIBITS and STIMULATES, while COEXISTS WITH represents weaker associative evidence.Weights reflect semantic specificity and relevance to the target extraction objective.
- 3.7 Relation Hierarchy Resolution: Relation orderings were tuned on the development set, with the reported hierarchy outperforming flat and inverted orderings by 1.2–2.1% F1.Other datasets use configuration files while retaining the same core chunking and selection strategy.
- 3.7 Relation Hierarchy Resolution: Combined ranking prefers higher-priority relations, resolves ties by trigger tier, and retains only non-overlapping propositions with highest precedence.In the example, the therapeutic proposition is retained over mechanistic or associative alternatives.
3.8 Proposition-First Extraction
Proposition-first extraction creates structurally complete candidate chunks by isolating atomic subject–trigger–object spans, while sliding windows provide fallback coverage.
- 3.8 Proposition-First Extraction: Sliding windows may separate entity context from relation context, leaving neither chunk as a complete triple.This fragmentation motivates extracting proposition-level spans.
- 3.8 Proposition-First Extraction: Proposition-first extraction isolates self-contained units containing a subject, relation trigger, and object.These atomic propositions are used as candidate chunks.
- 3.8 Proposition-First Extraction: The method detects infix, prefix, and postfix syntactic patterns for explicit relation expressions.Examples include Entity1–TRIGGER–Entity2, TRIGGER–Entity1 ...Entity2, and Entity1 ...Entity2–TRIGGER.
- 3.8 Proposition-First Extraction: For each trigger, nearest entities within ±25 tokens define the minimal span, which expands by ±3 tokens for negation and uncertainty markers.The expansion preserves polarity information for downstream processing.
- 3.8 Proposition-First Extraction: Sliding-window candidates provide fallback coverage when proposition patterns miss nested or coordinated constructions.Proposition chunks therefore complement rather than replace broader windows.
- 3.8 Proposition-First Extraction: Overlapping propositions are deduplicated using combined relation-hierarchy and trigger-tier priority.The INHIBITS proposition scores 6 versus 3 for the AFFECTS proposition in the example, and passive constructions invert semantic roles when required.
3.9 Hybrid Selection Strategy
The hybrid strategy combines semantic similarity with structural proposition priority, preserving broad context while ensuring structurally complete evidence reaches final selection.
- 3.9 Hybrid Selection Strategy: Candidates come from entity-aware sliding windows, proposition spans, and fixed-width fallback windows.Proposition spans are internally prioritized by trigger specificity and relation hierarchy.
- 3.9 Hybrid Selection Strategy: The semantic-best slot selects the candidate with highest similarity, while the structural-best slot promotes a top-N proposition candidate when available.If no qualifying proposition exists, the second-highest similarity chunk is selected as fallback.
- 3.9 Hybrid Selection Strategy: Proposition bias preserves precise relational evidence where competing triggers co-occur, while entity-aware windows retain broader context.The design avoids over-filtering contextual chunks that lack explicit triggers.
- 3.9 Hybrid Selection Strategy: Selected chunks are passed to BioMedRAG’s trained chunk scorer for final re-ranking, maintaining pipeline compatibility.The integration leaves the downstream learned scorer unchanged.
4 Experiments
Experiments compare the semantic framework with fixed 5-word BioMedRAG chunking across four biomedical benchmarks and component variants. Semantic chunking performs best on GM-CIHT and DDI, while fixed chunking remains stronger on ChemProt and ADE under this setup.
- Experimental setup: The evaluation covers GM-CIHT, DDI, ChemProt, and ADE, using fixed 5-word BioMedRAG chunking as the baseline while preserving the remaining pipeline.GM-CIHT, DDI, and ChemProt use triple extraction; ADE uses binary relation classification.
- Main results: 82.6% F1 on GM-CIHT improves over the fixed baseline’s 74.2% by 8.4 F1 points, while DDI improves by 1.0 F1 from 78.2 to 79.2.The gains occur on extraction datasets with diverse relation types and explicit cues.
- Main results: Fixed chunking outperforms semantic chunking by 0.8% F1 on ChemProt and 1.5% F1 on ADE.ChemProt benefits from broader context for fine-grained biochemical distinctions, whereas ADE relies on heterogeneous sentence-level context for binary classification.
- Ablation studies: Relation hierarchy resolution provides the largest incremental GM-CIHT gain at +3.9 F1, followed by tiered triggers at +2.4 F1 and entity-aware chunking at +1.9 F1.Proposition extraction adds +0.2 F1, while hierarchy resolution prioritizes TREATS over weaker associations such as COEXISTS WITH.
- Ablation studies: With in-context examples, fixed chunking gains +0.9 F1, whereas semantic chunking performs best with one example and degrades by −5.4 F1 with three.The authors suggest coherent chunks may require fewer demonstrations, while the three-example drop may reflect context limits or demonstration sensitivity.
- Ablation studies: The default cosine ranking with proposition bias reaches 82.6% F1, compared with 79.4% for a 70/30 cosine-structural ranking and 81.5% for NER-based entity spans.Under this setup, proposition bias avoids diluting cosine similarity, and regex-based spans outperform NER by 1.1 F1.
- Cross-dataset interpretation: Semantic chunking is most effective for coarse-grained, trigger-explicit relation extraction with moderate entity density, particularly within sentence-level biomedical literature benchmarks.The paper cautions that these conclusions may not extend to clinical notes or document-level retrieval.
5 Conclusion
The framework replaces fixed-size biomedical RAG chunks with configurable semantic candidates, improving extraction benchmarks while remaining less effective for dense or classification-oriented tasks. Its conclusions are bounded by below-published fixed-baseline scores and the need for future extensions.
- The framework combines entity-aware sliding windows, proposition-first extraction, tiered trigger prioritization, and relation hierarchy resolution.All components are externalized in configuration files for adaptation without code modification.
- +8.4 F1 points on GM-CIHT (82.6%) and +1.0 F1 on DDI (79.2%) demonstrate gains on extraction-focused benchmarks.The fixed GM-CIHT baseline was 74.2% F1 under the study setup.
- Ablations identify relation hierarchy resolution as the largest incremental contributor in the GM-CIHT stack, adding +3.9 F1.Additional analyses found proposition bias superior to mixing structural scores into cosine ranking, while pattern-based spans matched or beat NER on GM-CIHT.
- The fixed baseline’s absolute F1 is below published BioMedRAG figures, although fixed and semantic conditions share identical training and preprocessing.The authors retain the within-setup comparison as the primary evidence for their claims.
- Semantic chunking is most effective for coarse-grained relations with moderate entity density, whereas fixed-size chunking remains competitive for ChemProt and ADE.In these settings, entity preservation may group multiple targets and proposition extraction may over-constrain contextual evidence.
- Future work targets adaptive chunking, dataset-specific hierarchy design, clinical-note and drug-label extension, and entity-pair-specific isolation.These directions address high-density and multi-target scenarios and support broader reproducibility.
Code and Data Availability
The study uses publicly available datasets and plans to release implementation, configuration, preprocessing, conversion, and evaluation resources upon publication.
- All datasets are publicly available through the BioMedRAG benchmark setting and their original sources.
- The planned release includes semantic chunking code, dataset configuration files, preprocessing scripts, unified JSONL conversion, and evaluation instructions.Configurations include trigger vocabularies, tier assignments, relation hierarchies, negation rules, context markers, and chunking parameters.