Multi-Task Learning for Depression Detection in Dialogs

This is the source code repository for the paper Multi-Task Learning for Depression Detection in Dialogs (SIGDial 2022).

Requirements

• allennlp >= 2.0
• pytorch
• numpy
• disarray: used in utility.py, calculates metrics derived from a confusion matrix, info here

Datasets

DAIC-WOZ

Our main task depression detection uses DAIC-WOZ (part of the Distress Analysis Interview Corpus) (Gratch et al.,2014). We show one example in data/daic/. Download the whole dataset is available here.

DailyDialog

Our auxiliary tasks use Dailydialog (Li et al., 2017). We show a few examples in data/dailydialog/. We use the original separation of train, validation, and test. Download from here.

Source code

• main.py: choose MODE for train and test, modify arguments for different multi-task settings. We also provide a pretrained model in repo model/
• model.py: hierarchical structure modeling
• dataset_reader.py: read daic-woz and dailydialog
• utility.py: store auxiliary functions
• constant.py: store hard-coded paths, labels, etc.

Dialog act annotation

We provide annotation of dialog act for Ellie's utterances in DAIC (repo analysis/). Note that it's a corase-grained annotation with 5 classes: question, opening, comment, backchannel, other.

Citation

@inproceedings{li2022multi,
  title={Multi-Task Learning for Depression Detection in Dialogs},
  author={Li, Chuyuan and Braud, Chlo{\'e} and Amblard, Maxime},
  booktitle={SIGDIAL 2022-The 23rd Annual Meeting of the Special Interest Group on Discourse and Dialogue},
  year={2022}
}