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
No references match your search.