Immediate and accessible grief treatment via cold reading chatbots

Tracey, Peter

Abstract

This thesis presents a potential solution for prolonged grief disorder (PGD) sufferers waiting for psychological aid, by simulating the cold reading process through a chatbot model. PGD occurs in approximately 10% of all bereavements, and there is currently overwhelming demand for psychiatric aid meaning that 50% of patients wait over 3 months for treatment and 10% of patients wait over a year. This is likely to worsen during and immediately following the coronavirus pandemic. Therefore, an alternative is needed to treat PGD sufferers sooner. An existing solution is the use of a griefbot, a chatbot designed to resemble the deceased. However, current griefbots rely on pre-existing data from the deceased. Some people may not have this pre-existing data if the death was unanticipated, and the deceased did not leave behind sufficient messaging data. Therefore, another alternative is required to support PGD patients waiting for grief treatment who do not have the required pre-existing data for a griefbot. This research presents the solution as a chatbot that imitates a psychic medium who purports to communicate with the deceased. The proposed chatbot would not require pre-existing data from the deceased and could therefore be used by any PGD sufferer. Multiple approaches to building a chatbot were tried, including rules-based, retrieval-based, and generative models. Rules-based models use pre-written pattern-template pairs to produce predetermined responses to anticipated inputs. The rules-based cold reading chatbot works well in delivering a simple cold reading from start to end but is limited in its conversational range to the script it has been written to follow. Retrieval-based models calculate the distance between a user’s input and each line in a dialogue corpus and upon finding the closest line, returns the response to that line from the corpus. A retrieval-based cold reading chatbot returns messages that are suitable for a psychic medium, but its inflexibility leads to a lot of repetition in its
2
responses despite small, yet significant changes in the user’s inputs. Generative models use artificial neural networks to learn the connections between conversational inputs and outputs in order to create new responses to unanticipated inputs. The generative cold reading chatbot is able to learn some general conversational skills but struggles with learning the full cold reading technique from the available corpora. In order to improve the generative model, more training data would need to be obtained. Three experts in psychology provided their feedback to the overall premise, the rules-based chatbot and the generative chatbot. All experts supported the use of chatbots in PGD treatment and two of the experts supported the premise of a chatbot psychic medium to help PGD patients, while the other suggested a chatbot that simulates a grief specific therapist.

• Thesis Type: Thesis
• Deposit Date: Jul 3, 2023
• Publicly Available Date: Jul 31, 2023
• Award Date: Jun 30, 2023