Research on Medical Dialogue Generation based on Pre-trained Models

Na Liu ； Xiaohui Su ； Feng Huang

《International Core Journal of Engineering》 9卷9期 (2023/09) Pp. 159-170

https://doi.org/10.6919/ICJE.202309_9(9).0020

摘要

China is gradually entering an aging society, the online medical industry is becoming more and more important, and the requirements for intelligent online medical dialogue are getting higher and higher. The traditional dialogue system built using template-based or sentence-based planning methods is simple, but it requires high-skilled talents to provide different templates in different fields, which requires a lot of material, human and financial resources, poor portability and scalability, and is difficult to transfer to other fields. Based on the previous research on large-scale pre-training models to build a medical dialogue generation system, this thesis builds an intelligent dialogue system based on the Seq2Seq method, adds a fusion knowledge module, and adds a noise filter mechanism in the decoding process to screen out medical knowledge that does not match well with the history of doctor-patient dialogue. Finally, experiments are carried out on the large-scale medical dialogue dataset KaMed and COVID-19, and the experimental data show that compared with the traditional human-computer dialogue generation Seq2Seq model, the Perplexity value of the proposed method is reduced by 12.7%, and the B@2 value is increased by 5.72%, which greatly improves the accuracy of the model, makes its response more accurate and has a higher medical reference value.

關鍵字

Medical Conversation Generation ； Pre-trained Models ； Task-oriented Conversation Generation ； Medical Knowledge

參考文獻

QIN Libo, LI Zhouyang, LOU Jieming, YU Qiying, CHI Wanxiang. Review of research progress on natural language generation in task-based dialogue systems[J].Journal of Chinese Information Technology, 2022,36(01):1-11+20.

Wen T H, Gasic M, Kim D, et al. Stochastic language generation in dialogue using recurrent neural networks with convolutional sentence reranking[J]. arXiv preprint arXiv:1508.01755, 2015.

Wen T H, Gasic M, Mrksic N, et al. Semantically conditioned lstm-based natural language generation for spoken dialogue systems[J]. arXiv preprint arXiv:1508.01745, 2015.

Dušek O, Jurčíček F. Sequence-to-sequence generation for spoken dialogue via deep syntax trees and strings[J]. arXiv preprint arXiv:1606.05491, 2016.

Dušek O, Jurčíček F. A context-aware natural language generator for dialogue systems[J]. arXiv preprint arXiv:1608.07076, 2016.