HEAL: A Knowledge Graph for Distress Management Conversations

Abstract

The demands of the modern world are increasingly responsible for causing psychological burdens and bringing adverse impacts on our mental health. As a result, neural conversational agents with empathetic responding and distress management capabilities have recently gained popularity. However, existing end-to-end empathetic conversational agents often generate generic and repetitive empathetic statements such as "I am sorry to hear that", which fail to convey specificity to a given situation. Due to the lack of controllability in such models, they also impose the risk of generating toxic responses. Chatbots leveraging reasoning over knowledge graphs is seen as an efficient and fail-safe solution over end-to-end models. However, such resources are limited in the context of emotional distress. To address this, we introduce HEAL, a knowledge graph developed based on 1M distress narratives and their corresponding consoling responses curated from Reddit. It consists of 22K nodes identifying different types of stressors, speaker expectations, responses, and feedback types associated with distress dialogues and forms 104K connections between different types of nodes. Each node is associated with one of 41 affective states. Statistical and visual analysis conducted on HEAL reveals emotional dynamics between speakers and listeners in distress-oriented conversations and identifies useful response patterns leading to emotional relief. Automatic and human evaluation experiments show that HEAL's responses are more diverse, empathetic, and reliable compared to the baselines.

A review of the explainability and safety of conversational agents for mental health to identify avenues for improvement

“…Knowledge graph-guided conversations Question answering using KG is seeing tremendous interest from AI and NLP community through various technological improvements in query understanding, query rewriting, knowledge retrieval, question generation, response shaping, and others (Wang et al, 2017). For example, the HEAL KG developed by Welivita and Pu (2022b) allows LLMs to enhance their empathetic responses by incorporating empathy, expectations, affect, stressors, and feedback types from distressing conversations. By leveraging HEAL, the model identifies a suitable phrase from the user's query, effectively tailoring its response.…”

EMO-KNOW: A Large Scale Dataset on Emotion-Cause

“…For instance, Welivita et al (2021) produced EDOS by fine-tuning RoBERTa on a small label set created through crowdsourcing, and then used it to label 1 million conversations in the Opensubtitles dataset (Lison et al, 2018). Similarly, Welivita and Pu (2022) created HEAL -a knowledge graph consisting of 1 million distress narratives, annotated for emotion, post summarization, and node clustering using a variety of pre-existing models. A prevalent shortcoming of emotion labeling with current models is their limited accuracy; the emotion classifiers utilized in both HEAL and EDOS only achieved accuracies ranging from 65% to 65.88%.…”

Grafting Fine-Tuning and Reinforcement Learning for Empathetic Emotion Elicitation in Dialog Generation

“…In addition to context-specific responses, we find that some generic responses such as "I'm sorry to 1 github.com/declare-lab/exemplary-empathy hear that" can also appease users. So our model introduces a knowledge graph HEAL 2 [35] involving user feedback for distress management conversations. It consists of five types of nodes: (1) stressors: the cause of distress, such as suicidal ideation; (2) expectations: questions asked by the speakers usually; (3) response types: the most common types of responses given by the listeners with different stressors, such as "I understand how you feel.…”

“…Besides empathizing with users, some approaches aim to elicit the user's positive emotions (i.e., emotion elicitation dialog system [21]), especially under some specific situations that need emo- tional intervention, e.g., depression treatment and complaint handling [6,35]. This scenario requires the systems to take the user feedback into account dynamically at every dialog turn and then generate an encouraging response when a negative emotion is identified from the user side.…”

“…We use the Dense Passage Retrieval model (DPR) [10] to retrieve a set of semantically similar responses to the input context from the training set as the empathy-special exemplars. Also, we use semantic search to retrieve generic responses that elicit users' positive emotions from the Distress Management Conversations Knowledge Graph (HEAL) [35], which contains user feedback information to form the empathy-generic exemplars. This way, the empathetic and emotion elicitation exemplars can provide the following step with sufficient implicit knowledge.…”