Relevance Detection using Text Entailment for Health-related Question-Answer Texts with Imbalanced Data

The health consultation website has questions repeated with similar topics. Thus hence high volume the questions were referred to by the users and given answers by the doctor. Moreover, the answers obtained by doctors can give different answers to similar questions. Otherwise, users find it difficult to judge whether the given answers including relevant or not despite the questions referred to being similar. The research goal is to use the finetuning IndoBERT-QA model for entailment detection and then yield the prediction entailment, contradiction, and neutral to judge the answer based on referred to the question. Strength finetuning the IndoBERT-QA model can manage the word out-of-vocabulary in the Indonesian language and can learn the context from word sequence or the context of a whole sentence that is unknown complexly in the corpus that has a large size. Otherwise BERT model is unable to identify the semantic correlation between text questions and answers using the Indonesian language despite having different syntactic. The experiment result for comparison finetuning IndoBERT-QA model achieved more better accuracy is 0,91 than the BERT model only yielded an accuracy of 0,87. Effectively finetuning the IndoBERT-QA model can classify the class label entailment, contradiction, and neutral than the BERT model. Thus hence the BERT model achieves a small recognize the anatomic structure and medical terms specifically that are important for humans.