Bilingual system for solving medical questions: contribution with different architectures of language models
The present work focuses on the implementation and training of a bilingual Artificial Intelligence language model capable of solving questions of the medical Médico Interno Residente (MIR) exams with a variable number of possible answers, using state-of-the-art techniques from Natural Language Processing and Deep Learning.
Based on previous research in this task, the experimentation of this work has a solid knowledge base for training the models. In contrast to previous research, the models developed in this work receive information solely through examples of MIR exams without consulting any external information sources.
The language models developed in this work are able to learn the singularities of the MIR exam questions. To this end, different model architectures capable of answering exam questions with a variable number of possible options are proposed by modifying architectures from previous research.
By making use of exam sets in Spanish and English, the developed models in this project demonstrate a bilingual capability to answer these exams. This contrasts with most of the models developed to date in this type of task, which use only English.
2 language model architectures had been experimented, baseline and MedMCQA architecture.
Baseline architecture:
MedMCQA architecture:
The two datasets used are:
- MedMCQA dataset:
- Taken from github.com/medmcqa/medmcqa. Also available: huggingface/medmcqa
- CasiMedicos dataset: taken from github.com/ixa-ehu/antidote-casimedicos
The pre-trained model to use is EriBERTa: A Bilingual Pre-Trained Language Model for Clinical Natural Language Processing . Available in HiTZ/EriBERTa-base.
The experiments are listed into the src directory. Each one has its directory with its readme file.
