Question Answering and Dialogue Systems
Systems that answer natural-language questions and hold conversations, spanning retrieval and reading-comprehension question answering and both task-oriented and open-domain dialogue agents.
Definition
Question answering returns a direct answer to a natural-language question, while a dialogue system sustains a multi-turn conversation to inform or assist a user.
Scope
Covers question answering — factoid, retrieval-based, and reading-comprehension approaches — and dialogue systems, both task-oriented agents with dialogue-state tracking and open-domain conversational models. It addresses the role of pretrained models and the evaluation of correctness and coherence. Underlying transformer architectures are covered in the statistical-and-neural area.
Core questions
- How do retrieval-based and reading-comprehension question answering differ?
- How do task-oriented dialogue systems track state and choose actions?
- What distinguishes open-domain conversational agents?
- How are answers and conversations evaluated for quality?
Key concepts
- factoid question answering
- reading comprehension
- retrieval
- task-oriented dialogue
- dialogue-state tracking
- open-domain dialogue
- conversational agent
- evaluation
Key theories
- Reading-comprehension question answering
- Answering questions by locating or generating the answer from a passage, a task transformed by pretrained transformers fine-tuned on comprehension datasets.
- Dialogue-state tracking
- Maintaining a structured representation of the user's goals across turns so a task-oriented system can decide what to ask, confirm, or execute.
History
Conversational systems date to Weizenbaum's ELIZA (1966), which used simple pattern matching. Question answering matured through evaluation campaigns, and the arrival of large pretrained models such as BERT and subsequent generative models dramatically improved reading comprehension and open-domain dialogue.
Debates
- Genuine understanding versus pattern matching
- Whether fluent conversational systems understand language or, like ELIZA, exploit surface patterns; the question gains urgency as large models produce convincing but sometimes ungrounded responses.
Key figures
- Joseph Weizenbaum
- Daniel Jurafsky
- Jacob Devlin
Related topics
Seminal works
- weizenbaum1966
- devlin2019
Frequently asked questions
- What is the difference between task-oriented and open-domain dialogue?
- Task-oriented systems help a user accomplish a specific goal, like booking a flight, and track structured state. Open-domain systems aim to converse about anything, prioritizing coherence and engagement over completing a defined task.