Library
Conversational AI: Dialogue Systems, Conversational Agents, and Chatbots
ai

Conversational AI: Dialogue Systems, Conversational Agents, and Chatbots

Michael McTear 2021 13 references

McTear's comprehensive framework for dialogue systems covering rule-based, statistical, and neural approaches — use when building, evaluating, or reasoning about conversational AI systems.

dialogue-systems chatbots natural-language-processing conversational-agents neural-dialogue speech-processing evaluation-metrics

Overview

The Core Framework

  • Conversational AI evolved through three paradigms: rule-based (controllable but rigid), statistical/RL (robust under uncertainty but opaque), and neural end-to-end (fluent but unpredictable)
  • Each paradigm solves its predecessor's problems while creating new ones — this progression is cumulative, not replacement
  • Hybrid architectures combining all three are the practical state of the art for deployed systems
  • The pipeline (ASR→NLU→DM→NLG→TTS) remains the dominant architecture; end-to-end eliminates error propagation but loses interpretability
  • Evaluation is the persistent bottleneck — no consensus metric exists for open-domain dialogue

Quick Lookup

Situation Do This Avoid This
Starting a new dialogue system Begin rule-based, add ML incrementally Jumping straight to end-to-end neural
Handling ASR/NLU uncertainty Use POMDP belief states, not single-best Treating noisy input as deterministic
Neural responses are bland Apply MMI objective or nucleus sampling Relying on beam search with small k
Evaluating open-domain dialogue Use SSA or ACUTE-EVAL with human judges Using BLEU (fails for dialogue)
System contradicts itself Add persona grounding or memory networks Ignoring semantic consistency
Choosing a toolkit Match paradigm to data availability and safety needs Selecting based on popularity alone
Deploying to real users Prototype with Wizard of Oz first Skipping user simulation and WoZ stages

The Key Insight

"Current research in Conversational AI focuses mainly on the application of data-driven approaches... However, it is important to be aware of previous achievements in dialogue technology and to consider to what extent they contribute to current research and development." — Michael McTear, p. 1

References