# EmpBot: A T5-based Empathetic Chatbot focusing on Sentiments

**Authors:** Emmanouil Zaranis, Georgios Paraskevopoulos, Athanasios Katsamanis, Alexandros Potamianos  
**arXiv:** [2111.00310](https://arxiv.org/abs/2111.00310)  
**Date:** 2021-10-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2111.00310v1/

## Key points

- Introduction of EmpBot, an end-to-end empathetic chatbot
- Goal of an empathetic conversational agent: understand content and respond to implied feelings
- Method proposed: transformer pretrained language model called T5
- Three objectives during finetuning: response language modeling, sentiment understanding, and empathy forcing
- Importance of relevant and coherent responses (first objective)
- Focus on understanding emotional state of conversational partner (second objective)
- Training the model to respond with empathy (third objective)
- Evaluation using EmpatheticDialogues dataset and automated metrics + human evaluation
- Results show more empathetic responses compared to existing models
- Previous efforts in dialogue systems focused on engaging conversations, diverse responses, personalized responses, detecting emotions, and incorporating empathy
- Various approaches proposed such as emotional embeddings, memory mechanisms, GAN-based frameworks, dual-decoders, and emotion-labeled datasets
- EmpBot combines sentiment understanding and empathy forcing with a transformer pretrained language model
- Effective in producing more empathetic interactions compared to existing methods

## Summary

In this paper, the authors introduce EmpBot, an end-to-end empathetic chatbot. The goal of an empathetic conversational agent is not only to understand the content of the conversation but also to acknowledge and respond appropriately to the implied feelings of the conversation partner. To achieve this, the authors propose a method based on a transformer pretrained language model called T5. During finetuning, the authors suggest using three objectives: response language modeling, sentiment understanding, and empathy forcing. The first objective is crucial for generating relevant and coherent responses. The second objective focuses on understanding the emotional state of the conversational partner, which is an important aspect of human communication. The third objective aims to favor empathetic responses by training the model to respond with empathy. To evaluate their model, the authors use the EmpatheticDialogues dataset and employ both automated metrics and human evaluation. The results show that including sentiment understanding and empathy forcing auxiliary losses leads to more empathetic responses compared to existing state-of-the-art models. Previous efforts in dialogue systems have focused on making conversations more engaging by considering conversational context or producing diverse responses. Personalized responses have also been explored by conditioning generation on a persona profile. Additionally, researchers have worked on detecting emotions and incorporating empathy into dialogue systems. Various approaches such as emotional embeddings, internal and external memory mechanisms, GAN-based frameworks, dual-decoders, and emotion-labeled datasets have been proposed. Overall, this paper presents EmpBot as an approach that combines sentiment understanding and empathy forcing with a transformer pretrained language model to generate empathetic responses in chatbot conversations. The evaluation results demonstrate its effectiveness in producing more empathetic interactions compared to existing methods.