🔥🔥🔥 Read the paper
Unzip the data.zip file.
Train the T5 and T5-GloVe empathy classifier and sentiment regression models using the following commands:
## T5 Models ##
CUDA_VISIBLE_DEVICES=0 python train_empathy_classifier.py --epochs 12 --model "t5" --lr 1e-5 --dim ["emo"|"exp"|"int"]
CUDA_VISIBLE_DEVICES=0 python train_sentiment_regressor.py --epochs 12 --model "t5" --lr 3e-5
## T5-GloVe Models ##
CUDA_VISIBLE_DEVICES=0 python train_empathy_classifier.py --epochs 15 --model "glove-t5" --lr 2e-5 --dim ["emo"|"exp"|"int"]
CUDA_VISIBLE_DEVICES=0 python train_sentiment_regressor.py --epochs 15 --model "glove-t5" --lr 2e-5
You can downlaod our empathy and sentiment models from the link given here. These pre-trained weights are used for training the main empathetic response generator models. The model paths are hardcoded in ERGMainModel in models.py and ERGGloVeMainModel in glove_models.py.
The main T5 and the T5-GloVe emapthetic response generator models can be trained using:
## T5 Empathatic Response Generator Model ##
CUDA_VISIBLE_DEVICES=0 python train_t5.py --epochs 15 --lr 1e-4
## T5-GloVe Empathatic Response Generator Model ##
CUDA_VISIBLE_DEVICES=0 python train_glove_t5.py --epochs 50 --lr 1e-4 --add-exemplars "glove-t5"
The code for fine-tuning the DPR model is provided in the DPR/ directory. You can follow the instructions in DPR/ directory to fine-tune a DPR model on the Empathetic Dialogues and/or Empathy Mental Health Dataset. Then you can use the fine-tuned model path to retrieve the exemplars using:
CUDA_VISIBLE_DEVICES=0 python dpr_exempler_retriever.py --path DPR/outputs/yyyy-mm-dd/aa-bb-cc/saved/empd/dpr_biencoder.0
If you do not pass a --path then the non fine-tuned DPR model will be used for retrieval. We have provided the fine-tuned DPR retrieved examples in the *_dpr.csv files in the data/empathetic_dialogues/ directory.
Comparing efficacy of our model LEMPEx against the baseline models on various automated and human-evaluated metrics.
Comparing responses between models.
Comparison of responses with and without DPR exemplars.
Comparison of responses with and without empathetic losses.
Top exemplars from the DPR model fine-tuned on Empathetic Dialogs, and the original pre-trained DPR checkpoint without any further training. The exemplars from the fine-tuned DPR model are considerably more empathetic, diverse and contextually relevant. Notably, exemplars from the fine-tuned DPR are not always semantically similar to the references, although they are stylistically plausible and relevant with respect to the context.
Majumder, Navonil, Deepanway Ghosal, Devamanyu Hazarika, Alexander Gelbukh, Rada Mihalcea and Soujanya Poria. “Exemplars-guided Empathetic Response Generation Controlled by the Elements of Human Communication.” IEEE Access (2022).