Emora: An Inquisitive Social Chatbot Who Cares For You

Sarah E. Finch, James D. Finch, Ali Ahmadvand, Jason Ingyu Choi, Xiangjue Dong, Ruixiang Qi, Harshita Sahijwani, Sergey Volokhin, Zihan Wang, Wang Zihao, Jinho D. Choi, Choi, Jinho D.

Abstract

Inspired by studies on the overwhelming presence of experience-sharing in human-human conversations, Emora, the social chatbot developed by Emory University, aims to bring such experience-focused interaction to the current field of conversational AI. The traditional approach of information-sharing topic handlers is balanced with a focus on opinion-oriented exchanges that Emora delivers, and new conversational abilities are developed that support dialogues that consist of a collaborative understanding and learning process of the partner's life experiences. We present a curated dialogue system that leverages highly expressive natural language templates, powerful intent classification, and ontology resources to provide an engaging and interesting conversational experience to every user.

Introduction

Most conversational AIs today are transactional or information-oriented; while they excel at providing weather updates or movie trivia, they often fail to be truly "social." Human conversation is primarily driven by the sharing of experiences and opinions. Emora addresses this gap by aiming to be a "caring friend"—a chatbot that remembers personal details, expresses its own personality, and navigates complex, nested topics without losing conversational coherence.

System Architecture and Methods

Emora utilizes a hybrid, pipelined architecture that blends data-driven neural models with designer-controlled rule-based systems. The system includes components for Automatic Speech Recognition (ASR), a Natural Language Processing (NLP) pipeline, a Dialogue Manager, and various informational resources.

The overall system architecture of Emora, showing the flow from user input to response generation.

Dialogue Management via State Machines

Unlike end-to-end generative models, Emora uses a state machine where transitions represent dialogue actions. This allows for precise control over the bot's personality. To handle the non-linear nature of human chat, the system employs two key mechanisms:

  1. The State Stack: If a user initiates a sub-dialogue (e.g., moving from a general COVID-19 discussion to a specific sub-topic about remote learning), the previous state is pushed onto a stack. Once the sub-topic concludes, the state is popped, ensuring the bot can return to the original thread.
  2. Variable Table: A dedicated table stores learned user attributes (e.g., student=True). This allows Emora to personalize future turns, creating a sense of continuity and memory.
A sample dialogue transcript illustrating the use of the variable table and state stack to maintain context.

Natural Language Understanding (NLU)

Emora uses a Mixture of Experts (MoE) approach for intent and topic classification, combining Amazon’s built-in classifiers with custom-trained deep learning models. This improves coverage on niche topics and specific intents like "Reject." The NLU is further supported by an ontology of over 3.7 million entities derived from WordNet and Wikipedia.

Key Insights and Results

The development of Emora yielded several key insights into social AI:

  • Reciprocity: By sharing its own "profile" (e.g., claiming to be a student), Emora encourages users to open up, fostering a "collaborative understanding."
  • Opinions over Facts: Users are more engaged when the bot asks for their opinion rather than just reciting facts.
  • The "Life" Component: Topics focusing on personal experiences (School, Work, Relationships) were the most successful.

In the 2019 Alexa Prize semifinals, Emora achieved an average rating of 3.61. Controlled A/B testing showed that opinion-oriented openings (3.73) significantly outperformed fact-oriented ones (3.59). Furthermore, components focusing on personal experiences (the "Life" handler) achieved the highest ratings (3.85).

Conclusion and Future Work

While Emora represents a significant step toward social AI, challenges remain in scalability and common-sense reasoning. The heavy reliance on handcrafted templates makes it difficult to scale to an infinite number of topics. Future work will focus on automating character development and creating more dynamic methods for connecting distinct experience-oriented topics.

Summary updated

Abstract

Inspired by studies on the overwhelming presence of experience-sharing in human-human conversations, Emora, the social chatbot developed by Emory University, aims to bring such experience-focused interaction to the current field of conversational AI. The traditional approach of information-sharing topic handlers is balanced with a focus on opinion-oriented exchanges that Emora delivers, and new conversational abilities are developed that support dialogues that consist of a collaborative understanding and learning process of the partner's life experiences. We present a curated dialogue system that leverages highly expressive natural language templates, powerful intent classification, and ontology resources to provide an engaging and interesting conversational experience to every user.
BibTeX
@misc{finch2020emora,
  author = {Sarah E. Finch and James D. Finch and Ali Ahmadvand and Jason Ingyu Choi and Xiangjue Dong and Ruixiang Qi and Harshita Sahijwani and Sergey Volokhin and Zihan Wang and Wang Zihao and Jinho D. Choi and Choi, Jinho D.},
  title = {{Emora: An Inquisitive Social Chatbot Who Cares For You}},
  year = {2020},
  eprint = {2009.04617},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2009.04617},
}
Figure