Causal Relationships between Emotions and Dialogue Acts in Human Conversation

Setup

Python

The causal discovery experiments are implemented using Causal Discovery Toolbox https://fentechsolutions.github.io/CausalDiscoveryToolbox/html/index.html

Additional R libraries are required while using the cdt framework such as bnlearn, kpcalg, pcalg and RCIT. Please see CausalDiscoveryToolbox for further details.

R

The FCI model is implemented using standard R package pcalg. Install the latest pcalg package https://cran.r-project.org/web/packages/pcalg/index.html

Datasets

DailyDialogue

The DailyDialogue data used for this project is parsed using https://github.com/jojonki/DailyDialog-Parser/blob/master/parser.py.

Switchboard Dialogue Act Corpus (SwDA)

The SwDA data used for replication of the Dialog Act tagger model is a parsed version from https://github.com/NathanDuran/Switchboard-Corpus.

Automatic Dailog Act Classification

To obtain DA labels for the DailyDialogue data, we use the hierarchical encoder with CRF, proposed by Kumar et al. (2017). The original source code can be found at https://github.com/YanWenqiang/HBLSTM-CRF.

Slight changes has been made on the code to adapt to our own research. Please refer to the forked repo.

Experiments

Data

The input data folder contains 4 csv files:

• annotated.csv - 65 manually annotated DailyDialogue sessions
• pred.csv - The rest of all dialogue sessions from the parsed DailyDialogue dataset with predicted DA labels through automatic dailog act classification
• trans_curr.csv - transformed binary data including current utterances' emotion and DAs, where each column represents an individual type of emotion/DA and each row represents a single utterance (0 for non-existence and 1 or existence)
• trans_prev.csv - transformed binary data including previous utterance's emotion and current utterance's DA, same format with trans_curr.csv.

Qualitative analysis

The case study.py file contains our case study counting co-occurences of emotion-DA pairs for each utterance of annotated data.

The case study fisher.R file contains the Fisher's exact test performed on discovered relationships from the case study.

Quantitative analysis

The causal_variable_fci.R file contains our FCI model for discovering causal relationships between emotion and DA as two random variables in different contexts.

The causal_variable_gies_mmpc.py file contains our MMPC and GIES models for same experiment as in causal_variable_fci.R.

The causal_specific_types.py file performs causal discovery on relationships between specific types of emotion and DA, using GIES and MMPC algorithms in two contexts:

• current utterance's emotion vs. current utterance's DA
• previous utterance's emotion vs. current utterance's DA

References

Juraksky, D. (1997). Switchboard SWBD-DAMSL Shallow-Discourse-Function Annotation Coders Manual. Retrieved from https://web.stanford.edu/~jurafsky/ws97/manual.august1.html.

Kalainathan, D. & Goudet, O. (2019). Causal Discovery Toolbox: Uncover Causal Relationships in Python. arXiv preprint arXiv: 1903.02278.

Kumar, H., Agarwal, A., Dasgupta, R., Joshi, S. & Kumar, A. (2017). Dialogue Act Sequence Labeling using Hierarchical encoder with CRF. arXiv preprint arXiv:1709.04250.

Li, Y., Shen, X., Li, W., Cao, Z. & Niu, S. (2017). DailyDialog: A Manually Labelled Multi-turn Dialogue Dataset. arXiv preprint arXiv:1710.03957.