Skip to content
PaperFren

Does training on many biomedical datasets at once improve NER?

Open paper intelligence

Letting one neural network learn many biomedical name-tagging datasets together gave modest average gains, larger gains on some datasets, and the biggest benefit when training data was scarce.

Source

A neural network multi-task learning approach to biomedical named entity recognition

Crichton G, Pyysalo S, Chiu B, et al. · BMC bioinformatics · 2017

doi.org/10.1186/s12859-017-1776-8Read the full paper ↗67 citationscc by

Study at a glance

Design
Computational / modelling — Controlled comparison of a single-task CNN tagger against two multi-task CNN variants (shared layers with private outputs; POS-tagger-fed dependent model) across 15 NER corpora, plus pairwise, grouped and reduced-training-data experiments.
N
No single N: 15 biomedical NER corpora plus one POS-tagging corpus; pairwise experiment trained 15 x 14 models.
Population
Public biomedical text corpora annotated for anatomy, chemicals, diseases, genes/proteins and species.
Outcome
Exact mention-level macro F1 score on each corpus's test set.

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

What they did

The authors built a convolutional neural network that tags each word in a sentence as part of a named entity (such as a gene or disease name) or not. They compared it with two multi-task versions: one sharing all layers except a separate output per dataset, and one feeding in features from a part-of-speech tagger. They trained on 15 public biomedical corpora, tested every pair of datasets, grouped datasets by entity type, and cut the training data to 50%, 25% and 10% to see how multi-task learning behaves with less data.

What they found

Training on all datasets together raised F-scores significantly on five datasets and lowered one for the shared-output model, while the part-of-speech model improved six and hurt none. Grouping datasets that annotate the same entity type produced 16 significant gains out of 27 experiments. The advantage of multi-task learning grew as training data shrank; for two datasets, half the training data plus a helper dataset beat the full data alone. Which dataset helped which was sometimes surprising, e.g. a gene/protein corpus best helped a chemical corpus.

The limits

What it doesn't show

Average improvements were small (under one F-score point overall), and the authors admit it is hard to predict in advance which dataset combinations will help. The models deliberately skip task-specific tricks such as gazetteers, so they are not state-of-the-art systems, and several corpora have no established NER benchmark to compare against. The models cannot handle overlapping annotations of the same term, and the static-embedding CNN setup predates transformer language models.

Key terms

Named entity recognition (NER)
Finding and labelling mentions of things like genes, chemicals or diseases in text.
Multi-task learning (MTL)
Training one model on several related tasks at once so a shared representation can transfer knowledge between them.
F-score (F1)
The harmonic mean of precision and recall; here computed on exact entity matches.
Word embeddings
Fixed numerical vectors representing words, pre-trained on biomedical text and used as model input.
Auxiliary task
A secondary task (here part-of-speech tagging) trained to provide helpful information for the main task.

Flashcards

1 / 11

0 of 11 answers reviewed

Research intelligence for this paper

See its role on concept claims, tensions it is part of, placement history, and related discoveries.

Open paper intelligence

Quiz yourself

1 / 5

What is the core idea of multi-task learning?

Common questions

Why not just merge all the datasets into one big training set?

Different groups annotate differently; earlier work found naive merging dropped performance sharply, so the authors kept each dataset as its own task.

When is multi-task learning most useful here?

When a dataset is small: the gap over single-task training widened as training data was cut.

Which multi-task model is safer?

The dependent (POS-fed) model had smaller average gains but never significantly hurt a dataset, whereas the fully shared model had bigger wins and a loss.

More on Biomedical text mining