Does focusing on nearby words help find gene and disease names?
A convolutional network that looks only at the words around each term matched or beat most existing tools at spotting biomedical names in text, without hand-made features.
Source
GRAM-CNN: a deep learning approach with local context for named entity recognition in biomedical text
Study at a glance
- Design
- Computational / modelling — New neural architecture (word/character/POS embeddings, multi-size CNN kernels, CRF output) compared with published NER systems on three benchmark corpora.
- N
- No single N: BC2 has 15 000 training and 5000 test sentences; NCBI has 6892 disease mentions from 793 abstracts; JNLPBA has 20 546 training and 4260 test sentences.
- Population
- Biomedical text corpora (gene mentions, disease mentions, and five entity classes such as protein, DNA and cell type)
- Outcome
- Precision, recall and F1-score for entity recognition
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
The authors built GRAM-CNN, which represents each word by a pretrained biomedical word embedding, a character-level embedding and a part-of-speech tag embedding, then runs convolution kernels of several widths over the neighbouring words. A conditional random field layer on top picks the most likely sequence of labels. They trained and tested it on three public biomedical named-entity datasets and compared it with previously published systems.
What they found
GRAM-CNN reached an F1-score of 87.26% on the gene-mention set, ranking second overall and first among single (non-ensemble) models. On the disease corpus it ranked first with the same F1 and the highest recall. On the harder five-class JNLPBA corpus it scored 72.57%, just behind a rule-based system built specifically for that task, and it was the only method near the top on all three datasets.
The limits
What it doesn't show
There is no ablation isolating how much the local-context idea itself (versus the embeddings or the CRF) drives the gains, and no direct head-to-head retraining of an LSTM under identical settings; comparisons rely on numbers reported by other papers. The model cannot handle overlapping, disjoint or table-based mentions, needs thousands of labelled training examples and took days to train. Error analysis suggests some test-set labels are inconsistent, so the scores partly reflect annotation noise.
Key terms
- Named entity recognition (NER)
- The task of finding and labelling mentions of things such as genes, diseases or proteins in text.
- Convolutional kernel size
- How many neighbouring words a convolution filter looks at at once; several sizes capture n-gram patterns at different scales.
- Conditional random field (CRF)
- An output layer that scores whole label sequences so that impossible transitions (like an inside tag with no beginning) are discouraged.
- F1-score
- The harmonic mean of precision and recall, a single measure of tagging accuracy.
- IOB2 tagging
- A labelling scheme marking each token as Beginning, Inside or Outside an entity.
Flashcards
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.
Quiz yourself
What is the main architectural idea of GRAM-CNN?
Common questions
Why would ignoring the rest of the sentence help?
Biomedical sentences are long and often mix several topics, so the authors argue that distant words add noise; nearby words usually carry the clues that identify an entity.
Does this prove CNNs beat LSTMs for biomedical NER?
No. The comparison uses results reported by other papers under different setups, and there is no controlled experiment swapping only the encoder.
Why is JNLPBA harder?
It has five entity classes and, as the error analysis shows, inconsistent annotations, so all systems score lower on it.
More on Biomedical text mining