@inproceedings { , title = {Towards idea mining: Problem-solution phrase extraction from text}, abstract = {This paper investigates the feasibility of problem-solution phrases extraction from scientific publications using neural network approaches. Bidirectional Long Short-Term Memory with Conditional Random Fields (Bi-LSTM-CRFs) and Bidirectional Encoder Representations from Transformers (BERT) were evaluated on two datasets, one of which was created by University of Cambridge Computer Laboratory containing 1000 positive examples of problems and solutions (UCCL1000) with the corresponding phrases annotated. The F1-scores computed on the UCCL1000 dataset indicate that BERT is an effective approach to extract solution phrases (with an F1-score of 97\%) and problem phrases (with an F1-score of 83\%). To test the model’s robustness on a different corpus with a different annotation scheme, a dataset consisting of 488 problem-solution samples from the Conference on Neural Information Processing Systems (NIPS488) was collected and annotated by human readers. Both Bi-LSTM-CRFs and BERT performances were dramatically lower for NIPS488 in comparison with UCCL1000.}, conference = {ADMA 2022: International Conference on Advanced Data Mining and Applications}, doi = {10.1007/978-3-031-22137-8\_1}, isbn = {9783031221361}, pages = {3-14}, publicationstatus = {Published}, publisher = {Springer Verlag}, url = {https://uwe-repository.worktribe.com/output/10339175}, volume = {13726 LNAI}, keyword = {Text mining, Problem-solution extraction, NLP}, year = {2023}, author = {Liu, Haixia and Brailsford, Tim and Goulding, James and Maul, Tomas and Tan, Tao and Chaudhuri, Debanjan} editor = {Chen, Weitong and Yao, Lina and Cai, Taotao and Pan, Shirui and Shen, Tao and Li, Xue} }