Conditional Adversarial Learning For Empathetic Dialogue System

Abstract

As the population ages the mental health and physical health of the elderly become an increasingly crucial issue Their sense of loneliness and loss require us to listen more patiently and give them empathetic responses To this end an empathetic dialogue system for the elderly is proposed Through the use of empathy the system can obtain a greater understanding of the user’s emotion and situation which in turn assist the system to help the user to understand themselves Past empathetic dialogue systems did not consider personal experience or verify the emotional appropriateness of the generated response Neither did past systems employ empathetic techniques drawn from psychology to deal with user’s emotions and their situation Therefore the proposed dialogue system aims to provide a more empathetic response in common health topics for the elderly The creation of such a response is done by generating a template sentence and filling it in using a Slot-Value table The proposed system utilizes BERT to decide on the system dialogue act and emotion of the user response Using the decision made by BERT a controlled generation of a template is performed with a transformer trained with adversarial training Finally Slot-Value table is used to fill the generated template Due to the lack of empathetic dialogue dataset 1740 turns dialogues from the elderly containing empathetic response as defined in psychology were collected The collected data were labeled for the user’s emotion experience and system dialogue act Each user input sentence was processed with maximum matching to extract the Slot-Value table then the input sentence was converted into a template sentence by emptying out the slots These template sentences were used for training the transformer The experimental results from 5-fold cross-validation showed the conditional adversarial training approach yielded an 86 4% accuracy in response emotion selection and 90 5% accuracy in event selection which was slightly better than the transformer and a proposed transformer with conditional input On top of that the BLEU score for the proposed conditional adversarial training approach was also higher

Date of Award	2019
Original language	English
Supervisor	Chung-Hsien Wu (Supervisor)