TY - GEN

T1 - Knowledge-Aware Neural Networks for Medical Forum Question Classification

AU - Roy, Soumyadeep

AU - Chakraborty, Sudip

AU - Mandal, Aishik

AU - Balde, Gunjan

AU - Sharma, Prakhar

AU - Natarajan, Anandhavelu

AU - Khosla, Megha

AU - Sural, Shamik

AU - Ganguly, Niloy

N1 - Funding Information: We propose MedBERT, a novel application of transformers-based dual encoder model, for MFQC task, which is also medical domain knowledge-aware. We contribute a multi-label MFQC dataset; Med-BERT achieves state-of-the-art performance on ICHI (accuracy of 0.7) and CADEC dataset (accuracy of 0.9 and macro F1 score of 0.71), and generalizes very well in low-resource settings. Through extensive experimentation, we learn that incorporating medical concept-bearing terms as side information, contribute significantly to MedBERT. We learn that certain target classes heavily depend on keywords, while others require one to learn optimal representation of medical context. An interesting future direction will be to extend MedBERT to structured prediction tasks like entity and relation prediction, or broadly link prediction. Instead of BERT, we will work with BioBERT [16], which is a domain-specific pretrained model trained on biomedical articles. Acknowledgements. This work is supported in part by the Institute PhD Fellowship of IIT Kharagpur, the Federal Ministry of Education and Research (BMBF), Germany under the project Leib-nizKILabor (grant no. 01DD20003), the Adobe-funded project titled “Computational Aspects and Role of Content for Persuasive Brand Positioning”, and IMPRINT-1 Project RCO (project no. 6537).

PY - 2021/10/30

Y1 - 2021/10/30

N2 - Online medical forums have become a predominant platform for answering health-related information needs of consumers. However, with a significant rise in the number of queries and the limited availability of experts, it is necessary to automatically classify medical queries based on a consumer's intention, so that these questions may be directed to the right set of medical experts. Here, we develop a novel medical knowledge-aware BERT-based model (MedBERT) that explicitly gives more weightage to medical concept-bearing words, and utilize domain-specific side information obtained from a popular medical knowledge base. We also contribute a multi-label dataset for the Medical Forum Question Classification (MFQC) task. MedBERT achieves state-of-the-art performance on two benchmark datasets and performs very well in low resource settings.

AB - Online medical forums have become a predominant platform for answering health-related information needs of consumers. However, with a significant rise in the number of queries and the limited availability of experts, it is necessary to automatically classify medical queries based on a consumer's intention, so that these questions may be directed to the right set of medical experts. Here, we develop a novel medical knowledge-aware BERT-based model (MedBERT) that explicitly gives more weightage to medical concept-bearing words, and utilize domain-specific side information obtained from a popular medical knowledge base. We also contribute a multi-label dataset for the Medical Forum Question Classification (MFQC) task. MedBERT achieves state-of-the-art performance on two benchmark datasets and performs very well in low resource settings.

KW - clinical text classification

KW - online health communities

UR - http://www.scopus.com/inward/record.url?scp=85119178431&partnerID=8YFLogxK

U2 - 10.48550/arXiv.2109.13141

DO - 10.48550/arXiv.2109.13141

M3 - Conference contribution

AN - SCOPUS:85119178431

T3 - International Conference on Information and Knowledge Management, Proceedings

SP - 3398

EP - 3402

BT - CIKM '21

PB - Association for Computing Machinery (ACM)

T2 - 30th ACM International Conference on Information and Knowledge Management, CIKM 2021

Y2 - 1 November 2021 through 5 November 2021

ER -