How do you Converse with an Analytical Chatbot? Revisiting Gricean Maxims for Designing Analytical Conversational Behavior

Title of the Paper

How to Converse with Analytical Chatbots? Revisiting Analytical Conversational Behavior Design through Gricean Maxims

Paper Information

• Subject Area: Human-Computer Interaction, Chatbot Design, Data Visualization and Analysis
• Keywords: Chatbot, Intent Understanding, Visual Analysis, Ambiguity, Repair, Optimization

Research Background and Issues

• Identified Problems or Challenges:

  1. Analytical chatbots aim to facilitate data-driven conversations through natural language, but there is limited research on their design principles and alignment with user expectations.
  2. Current natural language interfaces (NLI) primarily focus on language as an input mechanism, lacking in-depth exploration of integrating language output with data visualization.
  3. The characteristics of different platforms (e.g., text-based vs. voice-based interaction) and their impact on user behavior remain unclear.

• Significance: Analytical chatbots have the potential to enhance the data exploration experience. However, the complexity of conversational data interaction behaviors necessitates understanding and adapting to these behaviors for effective design.

• Research Motivation and Related Work:

  1. This study revisits the applicability of Gricean Maxims in data-related conversations.
  2. By analyzing existing natural language interfaces and chatbot systems, the study extracts design principles tailored for analytical chatbots.

Solution

• Methods or Solutions:

  1. Propose design principles for analytical chatbots based on Gricean Maxims (quantity, quality, relation, and manner).
  2. Design and implement three prototype chatbots (Slack + charts, Echo Show + voice + charts, and voice-only Echo devices), each focusing on different interaction modalities and platforms.

• Innovations: Combining Gricean Maxims with the requirements of analytical data conversations to explore design features that support cooperative conversational behaviors; providing solutions for ambiguity resolution and improving data insights.

• Implementation Steps and Key Technologies:

  1. Conduct "Wizard of Oz Studies" to observe real user behaviors.
  2. Evaluate and compare user responses across three different platforms and interaction modalities, including analysis of natural language query characteristics, user expectations for system responses, and behaviors during failures.
  3. Develop a chatbot design framework incorporating natural language parsing, visualization modules, and template-based natural language generation techniques.

Research Outcomes

• Specific Results:

  1. Identified behavioral differences between voice and text interaction modalities. Voice interactions tend to focus on lookup queries, while text combined with charts supports deeper analytical exploration.
  2. Extracted six fundamental design patterns (e.g., data exploration-oriented greetings and descriptions, transition mechanisms, repair mechanisms).
  3. Proposed effective methods for disambiguation and refinement of ambiguous or unspecified user queries.

• Advantages Compared to Existing Solutions:

  1. Provides multimodal support (text, voice, charts), offering broader applicability compared to single-modal systems.
  2. Implements intuitive and user-friendly query correction through interactive widgets (e.g., dropdown menus in Slack).
  3. The design framework based on Gricean Maxims aligns more closely with human conversational habits, enhancing user experience.

• Experimental or Evaluation Results:

  1. In the Slack prototype system, interactive threads and widgets significantly increased multi-turn conversation frequency, with an average session length of 3.7 turns.
  2. Voice interactions on Echo devices were convenient but lacked depth, with users primarily engaging in simple fact-finding queries.
  3. Echo Show devices with charts supported more analytical needs but require improved consistency between voice and chart interactions.

• Limitations and Future Directions:

  1. Current corpora and system conversational capabilities are limited, especially in handling complex ambiguity and reasoning tasks.
  2. Multi-user collaboration and emotional/context-aware analytical chatbots remain unexplored.
  3. Developing personalized chatbots and ensuring data privacy and security are critical issues for future research.

Conclusion

This study provides a framework supported by theoretical and empirical research for designing analytical chatbots, bridging gaps between academic research and practical data interaction. It also proposes solutions and future directions to further enhance interaction efficiency and user experience.