MIDAS Dialog Act

Implementation for the paper MIDAS, A Dialog Act Annotation Scheme for Open-Domain Human-Machine Spoken Conversations. The implementation is based on Huggingface transformers

Dataset preparation

Example Input format in the including file folder ./da_data:

chatbot utterance : previous user utterance > current user utterance ## dialog act 1 of current user utterance;dialog act 2 of current user utterance

EMPTY means that there is no previous user utterance and the current user utterance is the first utterance responding to the chatbot

For example the dialog below with user current utterance's dialog act annotated as pos_answer can be formatted as: "did you know that : yes > i did ## pos_answer;"

Chatbot: did you know that?

User: yes

User: i did(dialog act: pos_answer)

Another example dialog shown below can be formatted as : "do you want to hear some fun facts about cats instead : EMPTY > yes ## pos_answer;command". In this case, there is only current user utterance so the previous utterance is input as EMPTY

Chatbot: do you want to hear some fun facts about cats instead

User: yes(dialog act: pos_answer and command)

Train and evaluate MIDAS dialog act prediction model

python run_classifier.py --data_dir da_data/ --bert_model bert-base-uncased --task_name da --output_dir output --do_train --do_eval --binary_pred

• --data_dir: the data directory where training and evaluating data file is stored. By default, the training data file is named as 'train.txt', and the evaluating data file is named as 'dev.txt'
• --bert_model: bert pre-trained model selected in the list: bert-base-uncased, bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, bert-base-multilingual-cased, bert-base-chinese, or bert pretrained lm finetuned on dialog data
• --task_name: the name of the task to train, for example da represents dialog act
• --output_dir: the output directory where the model predictions and checkpoints will be written
• --do_eval: whether to evaluate on the evaluation file
• --do_train: whether to run training
• --binary_pred: whether to use Binary-Cross-Entropy for binary prediction instead of only one tag(in case of multi-dialog-act

Evaluate MIDAS dialog act prediction model

python run_classifier.py --data_dir da_data/ --bert_model output --task_name da --output_dir output --do_eval --binary_pred

To use pretrained model, please download the model "pytorch_model.bin" to the dir output. Note: The pretrained model "pytorch_model.bin" in the output dir is trained with BERT finetuned on human-machine conversational data using text as context.

Test MIDAS dialog act prediction model on customized data

python run_classifier.py --data_dir da_data/ --bert_model output --task_name da --output_dir output --do_inference --binary_pred

• --data_dir: the data directory where the customized inference data file is stored. By default, the inference data file is named as 'inference.txt'. A sample format of the inference data can be found in da_data/inference.txt. The dialog act prediction for the customized data is stored in $output_dir$/inference_results.txt