Source-linked AI summary
Aligning Biomedical Texts and Knowledge Graphs: A Systematic Comparison of Lightweight Alignment Strategies
Artem Bisliouk, Elizaveta Nosova, Heiko Paulheim, Andreea Iana, Rita T. Sousa
TL;DR
Biomedical literature and knowledge graphs express overlapping facts differently, motivating explicit text–triple alignment for biomedical retrieval and grounding. The paper compares lightweight contrastive alignment choices with frozen encoders and finds that triple composition and retrieval direction dominate, while a simple text-to-KG projection performs best.
Problem
Biomedical literature and knowledge graphs encode overlapping facts in unstructured text and symbolic triples without explicit correspondence, limiting direct text–triple alignment.
Method
The paper freezes a text encoder and KG embedding model, trains only a lightweight contrastive projection, and compares six alignment design dimensions.
Results
Triple composition and training direction explain the largest retrieval-variance shares, while projecting text into KG space with a linear concatenation head yields the best alignment.
Takeaways & Limitations
Lightweight alignment offers a practical foundation for bridging biomedical text and KGs for knowledge grounding, evidence retrieval, and KG completion.
Takeaways & Limitations
Evaluation covers one CTD relation family and strict one-to-one pairs, so transfer to broader relation sets and many-to-many evidence remains untested.
Abstract
from arXiv · showhide
Biomedical knowledge exists in two complementary but distinct forms: unstructured scientific literature and structured knowledge graphs (KGs). Aligning them is essential for knowledge grounding, evidence retrieval, and KG completion, yet existing methods do not explicitly align free-text evidence with KG triples. We present a unified framework for systematically studying design choices for aligning biomedical text and KGs. With a text encoder and a KG embedding model both frozen, we learn only a lightweight projection between their spaces via a contrastive objective. This enables a fair comparison across six design dimensions: text encoder, KG embedding model, projection head, triple composition, training direction, and hard-negatives sampling. We construct CTD-Align, a corpus of over 22K one-to-one tripledocument pairs linking chemical-gene interactions from the Comparative Toxicogenomics Database to supporting PubMed passages. We evaluate alignment on it in two retrieval settings: document-to-triple and triple-to-document. We find that the triple composition and the training direction (i.e., shared retrieval space) have the greatest impact, whereas the text encoder and hard-negatives sampling matter little. Overall, simple choices win: projecting text into the KG space with a linear head over concatenated subject, predicate, and object embeddings performs best. These findings establish lightweight contrastive alignment as an effective, practical foundation for bridging biomedical text and KGs.
1. Introduction
Biomedical literature and knowledge graphs encode overlapping biomedical facts in fundamentally different forms, without explicit document–triple correspondence. The paper introduces a frozen-encoder, lightweight contrastive alignment framework and finds that simple design choices perform best.
- Biomedical literature is unstructured, whereas knowledge graphs encode curated biomedical facts as symbolic triples.
- Existing LM–KG integration methods primarily optimize downstream tasks rather than explicitly mapping free-text evidence to KG triples.
- The framework freezes text and KG encoders, trains only a lightweight contrastive projection, and compares six alignment design dimensions.
- CTD-Align contains 22,293 one-to-one pairs linking CTD chemical-gene interactions with supporting PubMed evidence.
- Triple composition and training direction have the greatest impact, while text-encoder and hard-negative choices matter little.
- Projecting text into KG space with a linear head over concatenated subject, predicate, and object embeddings performs best in both retrieval settings.
- Lightweight alignment provides a practical foundation for knowledge grounding, evidence retrieval, and KG completion.
2. Related Work
Prior LM–KG work commonly injects structured knowledge into language models or uses language to support KG tasks. This paper instead aligns frozen text and KG representations for bidirectional triple–document retrieval.
- Many LM–KG methods train underlying encoders for downstream tasks such as relation classification, question answering, or KG completion.
- The paper freezes both encoders, learns only a lightweight projection, and targets retrieval between evidence documents and triples.
- Related alignment methods often train encoders or align individual nodes and graph inputs rather than KG triples with evidence documents.
- Biomedical alignment approaches commonly target entities or concepts for generation or classification, whereas this work evaluates bidirectional triple-level retrieval.
- The paper positions its contribution as a systematic study of triple-level biomedical text–KG alignment.
3. Methodology
The method embeds paired documents and KG triples with frozen encoders, projects one modality into the other using contrastive learning, and evaluates retrieval in both directions across six design dimensions.
- Documents and triples expressing the same fact are encoded separately, while only a lightweight projection between their spaces is trained.
- Triple composition combines subject, predicate, and object embeddings using concatenation, Hadamard, L1, or L2 operators.
- Projection Heads: Projection heads include linear, MLP, and cross-attention architectures of increasing complexity.
- Training Direction: Training direction determines whether documents map to KG space, triples map to text space, or batches alternate between directions.
- Training Objective: InfoNCE contrasts each matched pair with other minibatch items and optionally with corrupted-triple hard negatives.
- Training Objective: Hard negatives replace the predicate, object, subject, or both entities, with no hard negatives generated for KG-to-text targets.
- Evaluation: Retrieval ranks the full target-modality candidate pool using cosine similarity for document-to-triple and triple-to-document tasks.
4. Biomedical Corpora & Dataset Construction
The dataset combines CTD’s curated biomedical relationships and PubMed evidence into a strict one-to-one corpus for alignment experiments. Construction filters ambiguous interactions, selects entity-containing evidence, and resolves collisions.
- CTD Knowledge Graph: CTD manually curates relationships among chemicals, genes, phenotypes, and diseases, with PubMed identifiers providing textual provenance.
- CTD Knowledge Graph: The core KG centers on single-action chemical-gene interactions, while additional disease and pathway associations enrich graph connectivity.
- CTD Knowledge Graph: CTD interactions are dereified by collapsing reified interaction patterns into direct edges for KG embedding.
- Figure 2 separates relation-type distributions for core and enriched CTD KGs, with four rare enriched-KG relations grouped as “Others.”
- Document Evidence: Entity co-occurrence spans are extracted from PubMed abstracts and ranked for semantic relevance to each chemical-gene triple.
- Candidate Document Selection: The corpus retains 1,546,310 unique single-action triples after discarding approximately 1.2 million multi-action records, or 40.5%.
- One-to-one Mapping: CTD-Align contains 22,293 one-to-one triple-document pairs from 17,530 publications after resolving shared-span collisions.
5. Experimental Setup
The experiments compare biomedical text–KG alignment components using domain-specific encoders, KG embedding models, projection settings, and retrieval evaluation. Training uses frozen representations with lightweight contrastive alignment and reports performance over complete candidate pools.
- Knowledge Graph Embeddings: Three KG embedding models—RotatE, TuckER, and RDF2Vec—capture graph structure through rotations, tensor factorization, or random-walk Word2Vec training.RotatE models relations as complex-space rotations; TuckER uses Tucker decomposition; RDF2Vec learns entity embeddings from graph random walks.
- Document Embeddings: BioBERT and PubMedBERT produce 768-dimensional document embeddings through mean pooling over token-level representations.BioBERT is pretrained on PubMed and PubMed Central text, whereas PubMedBERT is trained from scratch on biomedical corpora.
- Alignment Model: The alignment model varies projection heads, text encoders, KG models, triple compositions, hard-negative strategies, and training directions through separate hyperparameter optimization.The projector search selected hidden size 256, dropout 0.273, and weight decay 1.33 × 10^-3; the linear projector has no hidden-size hyperparameter.
- Alignment Model: InfoNCE training uses in-batch negatives, temperature 0.07, optional four corrupted KG negatives, AdamW, early stopping, batch size 256, and gradient clipping at 1.0.Training runs for at most 100 epochs with patience 15 for early stopping.
- Evaluation: Evaluation reports five-fold mean and standard deviation for MRR and Hits@k, plus median rank against the full target-modality candidate pool.Document-to-triple queries rank against all KG triples, while triple-to-document queries rank against all documents.
6. Results and Discussion
Alignment performance is driven primarily by the shared retrieval space and triple representation: projecting text into KG space with concatenated triple embeddings performs best. Triple composition explains most variance, while projection head, encoder, and hard-negative choices have smaller or conditional effects.
- Overall Retrieval Performance: The best configuration projects text into KG space with a linear head over concatenated subject, predicate, and object embeddings.It improves MRR over the BioBERT verbalization baseline by 90% on D2T and 64% on T2D.
- Overall Retrieval Performance: Median rank drops from 927 to 38 on D2T and from 845 to 55 on T2D, while Hits@10 roughly doubles in both settings.These gains compare the best aligned model with the stronger verbalization baseline.
- The Alignment Target Space: Projecting text into KG space outperforms other training directions in both retrieval settings.Bidirectional training matches text→KG on D2T but collapses on T2D, while KG→text underperforms in both settings.
- Impact of Design Dimensions: Triple composition explains 87% of D2T and 62% of T2D MRR variance, whereas the text encoder and negative-sampling scheme each explain less than 1%.The analysis uses one-way η2 on a balanced grid combining text→KG and bidirectional training directions.
- Impact of Design Dimensions: Concatenation reaches 14.0 D2T MRR within text→KG, compared with 3.9 or less for L1, L2, and Hadamard operators.Concatenation preserves separable subject-predicate-object structure in a 3d vector, unlike the other operators’ d-dimensional representations.
- Impact of Design Dimensions: Under concatenation, RDF2Vec and TuckER reach 14.4 D2T MRR, outperforming RotatE’s 13.1, while RotatE falls to 4.2 versus RDF2Vec’s 10.6 on T2D.Marginal averages over composition operators are misleading because KG embedding quality interacts strongly with triple composition.
- Secondary Design Choices: With concatenation fixed, the linear projection outperforms MLP and cross-attention, explaining 69% of variance among concatenation-based D2T configurations.On T2D, KG embedding choice dominates because of RotatE’s collapse.
- Secondary Design Choices: In-batch negatives suffice, and BioBERT versus PubMedBERT has only a small impact after projection learning.This contrasts with the verbalization baseline, where BioBERT substantially outperforms PubMedBERT.
7. Limitations and Future Work
The study prioritizes controlled comparison over breadth by evaluating one chemical-gene relation family, one-to-one pairs, and intrinsic retrieval. Future work should test broader graphs, many-to-many alignment, and downstream utility.
- Scope Boundaries: The evaluation covers one chemical-gene relation family from CTD, limiting evidence about transfer to broader biomedical knowledge graphs.The authors suggest testing richer relation sets such as Hetionet and PrimeKG.
- Scope Boundaries: Restricting evaluation to one-to-one pairs simplifies alignment but excludes documents with multiple interactions and multiple documents supporting one triple.The authors identify multi-positive supervised contrastive learning as a natural extension for many-to-many alignment.
- Future Work: Intrinsic retrieval evaluation does not directly measure downstream utility for tasks such as KG completion or evidence-grounded retrieval.The paper proposes evaluating learned representations on such downstream tasks.
8. Conclusion
The paper presents a unified framework for comparing lightweight biomedical text–KG alignment strategies and finds that simple design choices achieve the strongest retrieval alignment.
- The framework compares biomedical text–KG alignment across six design dimensions.These dimensions are the text encoder, KG embedding model, triple-composition operator, projection head, training direction, and hard-negatives selection strategy.
- Over 22K one-to-one chemical-gene triple-document pairs support extensive retrieval experiments on CTD-Align.The corpus links chemical-gene triples with supporting biomedical documents.
- Triple composition and training direction have the greatest impact on retrieval performance, while text encoder and hard-negatives selection matter comparatively little.
- Projecting text into KG space with a linear head over concatenated subject, predicate, and object embeddings yields the best alignment.
- The KG→text setting uses only in-batch negatives and is excluded from the hard-negatives comparison.
Declaration on Generative AI
The authors disclose using Opus 4.8 and Codex (GPT 5.5) for several publication-related tasks, followed by review and editing.
- Opus 4.8 and Codex (GPT 5.5) were used for image generation, abstract generation, rewriting, language checks, peer-review simulation, and coding support.
- The authors reviewed and edited the generated content and take full responsibility for the publication.