July 2024: ProDepDet wins the Honorable Mentioned Award at the 2024 IEEE CIS Student Grant Competition in Computational Intelligence in Biomedicine and Healthcare for the proposal entitled Vishadha: Early Smart Depression Detector.
ProDepDet is a framework which is specifically designed to use the knowledge of a pre-trained language model (PLM) in structure and semantic modelling in multi-party conversations to perform depression detection which is an unseen out-of-domain task. To our knowledge, this study is the first attempt to adapt the acquired knowledge of a PLM for out-of-domain task modelling using prompt tuning (PT)-based cross-task transferability.
The main contributions are:
- A novel method is proposed for enhancing out-of-domain task transferability of PT.
- A soft verbalizer is introduced along with a soft PT template for PT transferring for the first time.
- Multiple downstream tasks including depressed utterance classification (DUC) and depressed speaker identification (DSI) are used to evaluate the generalization and interpretability of the novel methods.
Python 3.8 and PyTorch 2.0 were used as the main programming language and machine learning framework, respectively . We separated MPC data into three categories based on the session length such as Len-5, Len-10, and Len-15 and used three different prompt lengths (l) such as 25, 50, and 75. Hyper-parameters were used such as GELU activations, Adam optimizer, with learning rate 0.0005, warmup proportion 0.1, and frozen model hyper-parameters, θ1 and θ2 both True.
We adopted several pre-trained models and large language models as source frozen baselines. To evaluate DUC , we used WSW, BERT, RoBERTa, SA-BERT, MPC-BERT, and DisorBERT as pre-trained models. For the evaluations of DSI , WSW, BERT, RoBERTa, ELECTRA, SA-BERT, MDFN, and MPC-BERT were used as pre-trained models. GPT-3, ChatGPT, and GPT-4 were adopted as large language models to evaluate both DUC and DSI .
- Download and extract the Reddit SDD Corpus
- Download and use the Reddit eRisk 18 T2 2018
- Download and use the Reddit eRisk 22 T2 2022
- Download and extract the Twitter Depression 2022
Evaluation results of DUC in terms of R10@1 which denotes the first correctly classified depressed utterances from 10 candidates. Ablation results are shown in the last two rows.
Evaluation results of DSI in terms of F1 score. Ablation results are shown in the last two rows.
