HEAL: A Knowledge Graph for Distress Management Conversations

Authors: Anuradha Welivita, Pearl Pu11459-11467

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 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.
Researcher Affiliation Academia Anuradha Welivita, Pearl Pu School of Computer and Communication Sciences Ecole Polytechnique F ed erale de Lausanne Switzerland {kalpani.welivita, pearl.pu}@epfl.ch
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Code and data available at github.com/anuradha1992/HEAL.
Open Datasets Yes Thus, we curated a new dataset from Reddit, containing dialogues that discuss real-world distressful situations. We used the Pushshift API (Baumgartner et al. 2020) to collect and process dialogue threads from a carefully selected set of 8 subreddits: mentalhealthsupport; offmychest; sad; suicidewatch; anxietyhelp; depression; depressed; and depression help... We used 80% of the dialogues to derive the knowledge graph and retained 10% of the dialogues each for validation and testing downstream tasks.
Dataset Splits Yes We used 80% of the dialogues to derive the knowledge graph and retained 10% of the dialogues each for validation and testing downstream tasks.
Hardware Specification No The paper does not provide specific hardware details (like GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper mentions various software components and models (e.g., Sentence BERT, BART, GPT-2, NLTK, RoBERTa, vis.js) but does not provide specific version numbers for them.
Experiment Setup Yes We experimented with 8 similarity thresholds from 0.6 to 0.95 with 0.05 increments to cluster distress narratives. ... This resulted in selecting an optimal threshold of 0.85. ... we selected 0.7, 0.75, and 0.7 as the optimal thresholds for clustering expectations, responses and feedback, respectively.