Gene Ontology as a Teacher for Biomedical Discourse Relations

Discourse parsers are models that label how neighbouring sentences relate: one causes the other, one contrasts with it, or one adds detail to it. In biomedicine they are starved of training data. The main biomedical treebank, BioDRB, has roughly 2,000 implicit relations, meaning relations without a signalling word like "because" or "however". The best reported model on it reaches 69.57% accuracy (Shi and Demberg). Meanwhile, the resources behind functional enrichment analysis contain millions of curated facts about what genes do and how they connect. Functional enrichment analysis is the statistical practice of asking whether a list of genes is unusually concentrated in known pathways. Its knowledge bases include the Gene Ontology (GO), a hierarchy of biological functions, and KEGG, a collection of directed pathway diagrams. When two adjacent sentences in a paper mention genes that KEGG connects with a directed "activates" edge, the sentences are probably in a causal relation. When the genes merely share many GO annotations, the second sentence is probably listing or elaborating.

We propose to turn the enrichment test itself into a labelling function. For each unannotated sentence pair mentioning genes, a hypergeometric test over their ontology annotations, together with pathway edges and GO parent–child links, proposes a discourse relation or abstains. The resulting silver corpus, drawn from PubMed and filtered for agreement with a connective-prediction model, is used to train an implicit-relation classifier. During that training we hide gene names so the classifier must learn the wording that marks each relation, not the knowledge base.

Core Claim

Structured biological knowledge supplies a relation-type signal for distant supervision, which sentiment could not. In MEGA-DT, Huber and Carenini trained attention models on review star ratings and built 250,000 silver RST trees, where RST (Rhetorical Structure Theory) marks each text span as central ("nucleus") or supporting ("satellite"). Their signal could recover tree shape and which spans are central, but not fine-grained relation types. Ontology structure distinguishes kinds of relatedness: a directed edge, shared membership, a hierarchical specialisation, or opposite regulation. We claim these map, noisily but above chance, onto causal, conjunction or list, instantiation, and contrast senses. We also claim that silver data built this way helps most on rare biomedical relations, where Jiang, Carenini and Ng showed performance is limited mainly by frequency (Pearson 0.87 between class frequency and F-score on news text).

The claim is falsifiable at the first step. If the ontology labels do not agree with gold BioDRB labels more often than the class prior predicts, the idea fails before any model is trained.

Relation to Prior Work

Two papers motivate the design by showing the pieces work separately. Poon, Toutanova and Quirk used the Pathway Interaction Database to distantly supervise protein–protein interaction extraction within single sentences across 22 million abstracts, reaching 25.6 F1 on GENIA against 33.2 for supervised models. We reuse the idea of a knowledge base as a noisy teacher, but apply it to relations between sentences rather than between entities. MEGA-DT shows that a document-level signal can seed large silver discourse corpora. We replace sentiment, which is absent from scientific prose, with ontology structure.

The closest existing work is Shi and Demberg's, which both motivates and constrains this proposal. They showed that biomedical discourse relations track entity semantics. Their approach fed flat subject–predicate–object triples into a knowledge-enhanced BERT (K-BERT) at input time. They also transferred labels from explicit relations sharing the same entities, and that unsupervised voting reached only 41.95%. Our approach differs in three ways. Knowledge enters as training labels rather than input features, so the trained classifier needs no knowledge base at test time. The labels use hierarchy and direction, not flat links. And the labels are produced for implicit pairs directly rather than borrowed from explicit ones.

GOLF supplies a method we reuse. It is a hierarchy-aware contrastive classifier that lifted PDTB Expansion.List from about 9% to over 43% F1, and we use it as the student architecture. For silver-label construction we reuse the agreement filtering of Jiang, Carenini and Ng, and the connective-prediction approach of Zhou et al. and Liu et al..

Research Program

The pipeline starts from PubMed abstracts and open-access full texts. We normalise gene mentions to identifiers, split text into argument pairs following BioDRB conventions, and keep adjacent pairs with no explicit connective in which both arguments mention at least one gene or a GO term name.

For a pair with gene sets $A_1$ and $A_2$, we first test functional overlap. Let $N$ be the number of GO terms in the background, $M$ the number annotated to $A_1$, $n$ the number annotated to $A_2$, and $k$ the number shared. The overlap score is the hypergeometric tail

$$ s(A_1, A_2) = -\log \Pr[X \ge k], \quad X \sim \mathrm{Hypergeometric}(N, M, n), $$

corrected for GO's parent–child redundancy by counting only the most specific shared terms. This is exactly the overrepresentation test used in enrichment analysis, pointed at a sentence pair instead of an experiment.

Rules over this score and the graph structure then propose a label. A directed KEGG or GO regulation edge from a gene in the first argument to a gene or process in the second proposes a causal relation. High overlap without an edge proposes conjunction or list. A GO term in the second argument that descends from a term in the first proposes instantiation or specification. Opposite regulation of a shared target proposes contrast. Anything else abstains. We expect most pairs to abstain, and that is intended.

Each proposal is then compared with the relation a masked language model predicts by inserting the most likely connective. Pairs are kept when the two sources agree, or when the ontology score is extreme. Before training, gene and term mentions are replaced by type placeholders. The student GOLF model is pretrained on the silver pairs, then fine-tuned on BioDRB.

Baselines

The naive reference points are the BioDRB majority class and fine-tuned BioBERT. Shi and Demberg report BioBERT at 2.0 points below their best model and base BERT at 6.5 points below.

The strongest existing method is Shi and Demberg's K-BERT at 69.57% accuracy on BioDRB. We will rerun it under the evaluation protocol they describe, reproducing their number before comparing. We found no reported macro-F1 on BioDRB on Lacuna, so we will compute macro-F1 for all systems ourselves. This matters because rare senses are where the claim lives.

A second strong competitor is K-BERT given the same GO and KEGG knowledge as injected triples. It isolates whether knowledge helps more as a teacher than as an input. A third is GOLF trained on an equal-sized silver corpus labelled only by connective prediction, which isolates the ontology signal.

The ablations remove, one at a time, the hypergeometric score (keeping raw edges), the hierarchy-based instantiation rule, agreement filtering, and gene masking. All systems use the same encoder size, the same unlabelled pool, and matched hyperparameter search budgets. Beating K-BERT means higher mean accuracy across seeds with non-overlapping confidence intervals, together with a macro-F1 gain driven by the rare senses.

Why not just…?

Why not just prompt a large language model? GOLF's authors report ChatGPT trailing specialised classifiers by about 30% macro-F1 on implicit relations. GenoMAS (Liu, Li and Wang) reports 32% accuracy extracting metadata from biomedical free text even with frontier models. Neither result is on BioDRB, so we will run a few-shot baseline ourselves. We will also test the model as an alternative labelling source for the silver corpus, which is the more serious version of this objection.

Why not just harvest explicit connectives and strip them? Shi and Demberg's entity-matched version of this idea reached only 41.95%, which suggests the relations transfer poorly from explicit to implicit uses. It remains our connective-only silver baseline. If it matches the ontology-labelled corpus, the ontology contributes nothing and we will say so.

Why not just self-train or co-train on unlabelled biomedical text? This is a genuine threat. Jiang, Carenini and Ng gained up to 22% on rare relations this way, and self-training worked for BioCause causal triggers (Mihăilă and Ananiadou). Self-training, however, amplifies what the seed model already predicts, and rare senses are exactly where the seed model is weakest. We will run co-training with the same unlabelled pool as a baseline, and also combine it with ontology labels. If co-training alone equals our method on rare senses, the ontology's value is refuted.

Proposed Experiments

The decisive first check requires no training. We apply the labelling function to gold BioDRB pairs that mention genes and measure per-sense precision and coverage against the gold labels and against the class prior. Precision near the prior for every sense ends the project.

The main experiment compares all baselines on BioDRB accuracy and macro-F1, reporting per-sense F1 against sense frequency. The goal is to test whether silver data flattens the frequency–performance slope reported by Jiang, Carenini and Ng.

The key leakage check evaluates separately on BioDRB pairs containing no gene mentions. Gains there show that the student learned transferable wording rather than a gene lookup, and the gene-masking ablation should lose those gains if masking is what forces it.

A secondary transfer test measures cause-versus-effect assignment on BioCause, where the reported self-trained result is 83.79% F1 for role assignment.

Finally, we vary silver-corpus size and the ontology-score threshold, trading label noise against volume. Coverage is also a concern: GO and KEGG are incomplete, and Poon et al. extracted far more interactions from text than exist in curated databases. We will therefore report how much of PubMed receives a label at all.