Note: This work the project was part of the course work for Advanced Natural Language Processing (CS685) at UMass, Amherst.
Our approach is based on the assumption that an empathetic conversational agent should mirror the emotion of the speaker . We propose to use a mode that favors sentiment understanding and empathetic response generation using the sentiment of each dialogue context. It is based on the Text-to-Text Transformer (T5) and we extend it with a sentiment analysis model and weighted loss during training, in order to apply sentiment understanding and enforce empathetic response generation.
| Input | T5 (Baseline) | Novel - T5 | Gold Response |
|---|---|---|---|
| I was bowling last week and I accidentally used the person's custom ball that was next |
Oh no! That's so embarrassing! | Oh no! What happened? | oh, how he reacted next |
| I found an injured bird in my yard today, it was only a baby too! |
Oh no! What happened? | I am sorry to hear that. Did you get it to come home? | How terrifying. I hope you were able to find and kill it |
| I woke up needing to use the bathroom at midnight. As soon as I sit down a giant bug flew into my face! I screamed so loud that I woke up my husband and he though we were being burgled! |
I am sure he was scared. | That is horrible. I hope you were able to sleep well. | Did you help it? |
| I am terrified that my wife is going to leave me. Why would she leave you? Because she is talking to other men behind my back |
I am sure she will. | That is a very scary thought. | oh no, i'm so sorry to hear that. I know your heart is broken. Did you try to talk to her about it? |
Dataset - Empathetic Dialogues
Train Novel T5 - Training Novel T5 model
Test Novel T5 - Testing Novel T5 model
Output - Generated Text
gpt2.zip - It is model where picked up most of the base code and modified it to work for T5 ( Source )
paper - Refer paper for more details.
Step 1. - pip install -r requirements.txt
Step 2. - python ./experiments/train_novel_t5.py --batch-size=4 --epochs=50 --multitask1=0.8 --ckpt=yourcheckpoint
Step 3. - python ./experiments/test_novel_t5.py --modelckpt=yourcheckpoint
Subramanya N - snagabhushan@umass.edu
Shashank Srigiri
Venkata Bramara Parthiv Dupakuntla