Skip to content
PaperFren

Biomedical text mining

Can expert models for each entity type help each other tag biomedical text?

Yoon W, So CH, Lee J, et al. · BMC bioinformatics · 2019

Open access · cc by · source: Europe PMC

Letting separately trained disease, chemical and gene taggers pass their predictions to one another raised precision and cut cases where one entity type was mistaken for another.

Save to libraryView originalAsk PaperFrenStudy view

Study at a glance

Design
Computational / modelling — New NER architecture (CollaboNet) compared with single-task and multi-task BiLSTM-CRF baselines on six public biomedical corpora, plus manual error analysis.
N
No single N: six benchmark datasets (BC2GM, BC4CHEMD, BC5CDR-chem, BC5CDR-disease, JNLPBA, NCBI); the error analysis covered 4334 single-model errors and 3966 CollaboNet errors.
Population
Sentences from MEDLINE abstracts annotated for disease, chemical or gene/protein entities
Outcome
Entity-level precision, recall and F1 (exact span match); counts of error types

Structured fields used in claim comparison tables when every cited study has a complete layer.

Key findings

CollaboNet achieved higher precision and F1 than most single-task models and beat the previous multi-task model on 4 of 6 datasets. In the error analysis, entity-type confusions made up 30.8% of the single model's errors but 18.6% of CollaboNet's errors. However, CollaboNet made more span (boundary) errors, and the authors note the overall gain may look modest given the added complexity.

Methodology

The authors built a single-task tagger (a bidirectional LSTM with a CRF output layer and CNN-based character embeddings) for each entity type. In CollaboNet, each tagger is first trained on its own dataset, then retrained while receiving the outputs of the other 'collaborator' taggers as extra input. They compared this against single-task models and a published multi-task model on six benchmark corpora and hand-classified the errors.

Limitations

The extracted text lacks the results tables, so exact F1 scores per dataset cannot be checked here. Improvements are reported against a small set of BiLSTM-era baselines, before transformer models like BioBERT became standard. Training is costly (about 3000 minutes for the collaborative phases), and inconsistent annotation conventions across datasets make exact-match scores partly a measure of labelling style rather than understanding. No significance tests across the ten random initialisations are described in the text.

How this study connects

Role on claims

Each row is a claim on a concept or method page where this paper supports, challenges, or qualifies the statement. Roles are hand-checked — not a model guess.

  • SupportsBiomedical named entity recognitionconcept

    Training across datasets is a reliable lever with small data.

    Sharing information across datasets or entity types helps, especially when data are scarce: multi-task training gave 16 significant gains in 27 same-entity experiments and helped most as training data shrank; CollaboNet cut entity-type confusions from 30.8% to 18.6% of errors; one LLaMA trained on three datasets beat single-dataset training (0.876 vs 0.762 F1 on NCBI disease).

    Evidence for the claim as stated.

Related papers in this topic

Same topic cluster — not a recommendation engine.