# Towards Coherent and Engaging Spoken Dialog Response Generation Using Automatic Conversation Evaluators

**Authors:** Sanghyun Yi, Rahul Goel, Chandra Khatri, Tagyoung Chung, Behnam Hedayatnia, Anu Venkatesh, Raefer Gabriel, Dilek Hakkani-Tur  
**arXiv:** [1904.13015](https://arxiv.org/abs/1904.13015)  
**Date:** 2019-04-30  
**Summary page:** https://www.summarizepaper.com/en/paper/1904.13015v1/

## Key points

- Existing end-to-end open domain dialog systems trained with maximum likelihood objective have limitations
- Lack of generalizability and generic response problem are common issues in these systems
- Authors propose a system that evaluates chatbot responses for coherence and engagement at each turn
- Turn-level dialog quality feedback is highly correlated with human evaluation
- Feedback is used to mitigate problems and improve the overall quality of the dialog system
- Two mechanisms presented: reranking and direct modification of loss function during training
- Studies show that incorporating explicit feedback improves response generation models
- Models incorporating both reranking and direct modification produce more engaging and coherent responses compared to traditional models
- Improvement observed through automatic evaluation metrics and human evaluation
- Proposed approach enhances end-to-end open domain dialog systems by incorporating explicit feedback on coherence and engagement at each turn

## Summary

The paper titled "Towards Coherent and Engaging Spoken Dialog Response Generation Using Automatic Conversation Evaluators" discusses the limitations of existing end-to-end open domain dialog systems, which are primarily trained with a maximum likelihood objective. These systems often suffer from lack of generalizability and the generic response problem, where the system generates responses that can be applied to a wide range of user utterances without providing specific or engaging answers. To address these issues, the authors propose a system that evaluates chatbot responses at each turn for coherence and engagement. This system provides explicit turn-level dialog quality feedback, which has been found to be highly correlated with human evaluation. The goal is to use this feedback as a signal for mitigating the aforementioned problems and improving the overall quality of the dialog system by selecting responses from different approaches. To demonstrate the effectiveness of incorporating explicit feedback into neural response generation models, two mechanisms are presented: reranking and direct modification of the loss function during training. These mechanisms aim to improve dialog quality by incorporating the feedback into the model's decision-making process. The authors conducted studies to evaluate their proposed approach in an open-domain spoken dialog setting. The results show that response generation models that incorporate both reranking and direct modification produce more engaging and coherent responses compared to traditional models trained with maximum likelihood objective alone. This improvement is observed through both automatic evaluation metrics and human evaluation. Overall, this paper presents a novel approach to enhance end-to-end open domain dialog systems by incorporating explicit feedback on coherence and engagement at each turn. The proposed mechanisms effectively improve response quality, addressing issues related to lack of generalizability and generic responses commonly seen in such systems.