Automated COVID-19 Dialogue System Using a New Deep Learning Network

Abstract

During the coronavirus disease 2019 (COVID-19) pandemic outbreak, it is necessary to apply social distancing measurements and search for an alternative to physical contact due to the spread of viral infection. The interest in task-oriented dialogue systems has grown remarkably in healthcare, using natural language in the dialogue between patients and doctors. However, the doctor’s advice is implicit and unclear in most conversations, and the patient may also be nervous when describing symptoms or may have difficulty describing them. Therefore, the patient’s description of symptoms is insufficient for a diagnosis by doctors. This study aims to provide suitable medical advice based on the patients’ symptoms during the conversation between doctors and patients by proposing a new deep learning method for automated medical dialogue systems. The model is based on an encoder and two stages of learning to make reliable decisions. The encoder extracts important words using text normalization, resulting in two vectors: symptom vectors and doctor utterance vectors. The symptom vectors are represented as a weighted bag-of-words feature. The first stage is used to cluster the patients’ utterances by applying Hopfield network while considering the semantic similarity, whereas the second stage extracts an implicit label as a template of advice using clustering. Additionally, the external evaluation model used the applied feedforward neural network classification algorithm using labels obtained in the second stage. The CovidDialog-English dataset is used to evaluate the model. The experimental results indicate the high performance of the feedforward neural network with an F1-score of 0.972 and presents a comparison of three clusters using the k-nearest neighbours and naïve Bayes-based models.