Skip to content

GEM is a Graph Attention Encoder for Multi-task Depression Severity Detection in Multi-party Conversations.

License

Notifications You must be signed in to change notification settings

KUAS-ubicomp-lab/GEM

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

29 Commits
 
 
 
 
 
 

Repository files navigation

GEM

Graph Attention Encoder for Multi-task Depression Severity Detection in Multi-party Conversations.

System Design

GEM High-level Design drawio

Approach

Multi-party conversations (MPCs) consists of complex discourse-level structures among multiple interlocutors and respective utterances. Compared to two-party conversations, MPCs significantly enhances the discourse-base natural language understanding allowing to model specific downstream tasks including emotion recognition of the interlocutors. Recent advancements of graph attention networks (GATs) have been utilized to discover various downstream tasks on MPC modelling using the root-level and sub-level utterance structures. GEM, incorporates these root-level and sub-level MPC utterances to model depression detection and severity classification downstream tasks. To our knowledge, we make the first attempt to utilize root-sub relationships between root-level and sub-level utterances in an MPC to develop a GAT-based encoder for depression detection. As shown in the system design, three variants of GATs are proposed.

  • Vanilla GATs
  • GATs with long short-term memory (LSTM)
  • GATs with skip connections

Downstream Tasks Formation

Three downstream tasks, Depressed Utterance Detection (DUD), Depressed Interlocutor Recognition (DIR), and Depression Severity Classification (DSC), are formed based on the outputs of the task-specific layers of GEM.

Experimental Setup

Requirements:

  • Python 3.8
  • PyTorch 2.0
  • OpenAI

Setting OpenAI API key: export OPENAI_API_KEY=yourkey

About

GEM is a Graph Attention Encoder for Multi-task Depression Severity Detection in Multi-party Conversations.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages