An NLP model to classify abstract sentences into the role they play (e.g. objective, methods, results, etc..) to enable researchers to skim through the literature and dive deeper when necessary.
Try Demo; WEB APP
- More specificially, I'am going to replicate the deep learning model behind the 2017 paper PubMed 200k RCT: a Dataset for Sequenctial Sentence Classification in Medical Abstracts.
- The PubMed 200k RCT dataset is described in Franck Dernoncourt, Ji Young Lee. PubMed 200k RCT: a Dataset for Sequential Sentence Classification in Medical Abstracts. International Joint Conference on Natural Language Processing (IJCNLP). 2017.
Some miscellaneous information:
-
PubMed 20k is a subset of PubMed 200k. I.e., any abstract present in PubMed 20k is also present in PubMed 200k.
-
PubMed_200k_RCTis the same asPubMed_200k_RCT_numbers_replaced_with_at_sign, except that in the latter all numbers had been replaced by@. (same forPubMed_20k_RCTvs.PubMed_20k_RCT_numbers_replaced_with_at_sign). -
Count Plot
All the note books are availabel here
- NaiveBiase Model -> 72% Accuracy
- Conv1D Model -> 78% Accuracy
- Model using pretrained token embedding ( Universal sentence embedding ) -> 75% Accuracy
- Conv1D Model using character level embedding -> 73% Accuracy
- Model with both token and charcter level embedding -> 76% Accuracy
- Model with token, character and position level embedding ( https://arxiv.org/pdf/1612.05251.pdf ) -> 81% Accuracy
- Model described in this paper with bert embedding -> 88% Accuracy
- Tensorflow
- tensorflow_text
- tensorflow_hub
- sklearn
- Matplotlib
- numpy
- pandas
- spaCy
