Code for paper

Yubo Xie and Pearl Pu. Empathetic Dialog Generation with Fine-Grained Intents. CoNLL 2021. PDF Link.

The project was developed using the following packages:

tqdm==4.49.0
numpy==1.19.3
scipy==1.5.2
pandas==1.1.0
tensorflow==2.3.1
pytorch_transformers==1.2.0

• datasets.py: read the data and tokenize the text;
• model_basics.py: implementation of Transformer basic components;
• model_emo_pred.py: implementation of the response emotion/intent predictor;
• model.py: implementation of the empathetic dialog model;
• model_utils.py: utility functions for the model implementation;
• train_os.py: pre-train the model on the OS dataset;
• train_edos.py: fine-tune the model on the EDOS dataset;
• train_ed.py: fine-tune the model on the ED dataset;
• train_emo_os.py: pre-train the response emotion/intent predictor on the OS dataset;
• train_emo_edos.py: fine-tune the response emotion/intent predictor on the EDOS dataset;
• train_emo_ed.py: fine-tune the response emotion/intent predictor on the ED dataset;
• predict_emo.py: predict the response emotion/intent;
• beam_search.py: implementation of the beam search algorithm;
• predict.py: generate the responses.

The preprocessed data needed for training (tokenization and emotion/intent distribution for each utterance) can be found here (inside the folder data).

TensorFlow checkpoints can be found here (inside the folder checkpoints).

The raw OS and EDOS datasets can be found here.

See the LICENSE file in the root repo folder for more details.