Conversational Agents: Goals, Technologies, Vision and Challenges

Merav Allouch, Amos Azaria, Rina Azoulay

TL;DR

This comprehensive review surveys the landscape of conversational agents (CAs), detailing their goals, underlying technologies, and future challenges. It provides a structured overview of CA design, applications across various domains, and the critical human-related aspects like emotion and personality, offering a valuable resource for researchers and developers in the field.

What Problem Does This Paper Address?

The paper addresses the growing ubiquity and complexity of conversational agents (CAs) by providing a structured and comprehensive overview of the field. As CAs become increasingly integrated into daily life, understanding their diverse goals, the technologies enabling them, their applications, and the associated challenges is crucial. The authors aim to clarify terminology, categorize CAs based on their abilities and domains, and highlight current and future research directions, including technical, human-related, and ethical considerations.

Methods

The paper doesn't introduce a novel method but rather synthesizes existing knowledge. It categorizes CAs based on several criteria:

  • Mode of Communication: Text-based (chatbots), voice-based virtual agents, and embodied agents (graphically or physically).
  • Action Capabilities: Communication-only agents vs. virtual/personal assistants with cyber and/or physical abilities.
  • Application Domain: Open domain/general purpose CAs vs. goal-oriented CAs.

It details the core technological components of CAs, including:

  • Text-Related Components: Natural Language Understanding (NLU), Dialogue Manager, and Natural Language Generation (NLG).
  • Voice-Related Components: Automatic Speech Recognition (ASR), Non-verbal Information Extraction, and Text-to-Speech (TTS).
  • Physical-Related Components: Perception, Shared Motor Knowledge, and Gesture Generator for embodied agents.
  • Task-Related Components: State Tracker, Policy Manager, and Action Manager for goal-oriented CAs.

The review also covers various machine learning techniques, deep learning advancements, and end-to-end models used in CA development.

Datasets

The paper discusses several publicly available conversation datasets used for CA development and evaluation, categorized by application:

  • General-Purpose CAs:
    • DailyDialog
    • SubTle corpus (derived from OpenSubtitles)
    • Movie dialogue datasets (Kaggle, Cornell Movie Dialogues Corpus)
    • Ubuntu dialogue corpus
  • Question-Answering Datasets:
    • Stanford Question Answering Dataset (SQuAD) v1.1 & v2
    • CNN/Daily Mail dataset
    • Natural Questions dataset
    • TriviaQA
  • Goal-Oriented CAs:
    • Schema-Guided Dialogue (SGD) dataset
    • MultiWOZ
    • Taskmaster-1
    • MultiDoGo
  • Supporting CAs (Healthcare, Education, Social Assistance):
    • COVID-19 dialogue dataset
    • MedDialog
    • SEMAINE
    • EmpatheticDialogues
    • Offensive response dataset
    • BURCHAK dataset
    • The CIMA collection

Key Insights and Ideas

  • The blurring lines between CA types: While distinct categories exist, modern CAs often integrate capabilities from multiple types (e.g., voice-based agents using text processing).
  • The critical role of NLU and NLG: These are foundational for any CA, enabling understanding and response generation, with deep learning significantly advancing their capabilities.
  • Human-centric aspects are paramount for engagement: Beyond technical prowess, CAs need to exhibit emotional intelligence, personality, and personalization to foster user connection and satisfaction, especially in sensitive domains like healthcare and education.
  • Ethical considerations are increasingly important: As CAs become more influential, issues of privacy, transparency (identifying as AI), potential for manipulation (influence agents), and fairness require careful attention and development of detection mechanisms.
  • Evaluation remains a challenge: While human evaluation is the gold standard, its cost drives the development of machine-based metrics, though achieving strong correlation with human judgment is an ongoing research area.

Key Results

The paper does not present specific quantitative results from a single experiment but rather synthesizes findings from numerous studies. Key qualitative findings include:

  • CAs have demonstrated success in various domains including commercial, healthcare, educational, and personal assistance.
  • Deep learning and advanced NLP techniques have significantly improved CA capabilities in understanding and generating natural language.
  • Adding personality and emotional awareness to CAs can lead to increased user engagement and satisfaction.
  • Personalization of CAs can enhance user experience but raises privacy concerns.
  • Influence agents in social networks are a growing concern, with studies showing a significant percentage of accounts exhibiting such behavior.

Limitations and Future Work

Limitations: * The paper acknowledges inconsistencies in terminology and definitions within the CA field. * It highlights the ongoing challenge of achieving truly human-like conversation, particularly in terms of thinking beyond just talking. * The evaluation of CAs remains complex, with machine metrics often not strongly correlating with human judgment. * Privacy concerns are a significant limitation for personalized CAs.

Future Work: * Addressing moral and security issues related to CA development. * Developing more robust methods for emotion detection and empathetic response generation. * Enhancing CA personality and adaptation to user needs and preferences. * Creating more effective and reliable evaluation metrics. * Developing protocols and algorithms for CAs to learn directly from humans and fine-tune their capabilities. * Investigating the ethical implications of CA identity disclosure and the potential for manipulation. * Developing technologies to detect and counter malicious influence agents.

Summary updated

BibTeX
@inproceedings{allouch2021conversational,
  author = {Merav Allouch and Amos Azaria and Rina Azoulay},
  title = {{Conversational Agents: Goals, Technologies, Vision and Challenges}},
  booktitle = {Sensors},
  year = {2021},
  url = {https://openreview.net/forum?id=ZETc_vLPWU-},
}
Figure