Skip to content

Dialogue Generation

Dialogue generation systems produce contextually appropriate responses in conversations. Dialogue systems span two broad categories:

Task-oriented dialogue: Guided interactions toward achieving specific goals (e.g., restaurant booking, customer support, technical troubleshooting). Success is measured by task completion; hallucinations are problematic when they mislead users about task constraints or factual prerequisites.

Open-domain dialogue: Extended conversations on diverse topics without a specific task goal. Open-domain systems must balance faithfulness (grounding in provided context or knowledge) with engagement and diversity.

Hallucination in dialogue

Dialogue systems tolerate hallucination differently than summarization systems. While task-oriented systems require high faithfulness (especially in safety-critical domains), open-domain systems prioritize engagement—researchers often accept minor hallucinations if they make responses more interesting or diverse.

Intrinsic hallucinations: Response contradicts the dialogue history (e.g., claiming the user said something they didn't).

Extrinsic hallucinations: Response asserts facts unsupported by dialogue history or external knowledge (e.g., fabricating details about a restaurant's menu when not retrieved).

Dialogue systems show high tolerance toward hallucinations compared to summarization, because user expectations are different—users expect conversational systems to sometimes speculate or offer general knowledge beyond the immediate context.

Architectures

Modern dialogue generation uses transformer-based sequence-to-sequence models (BLENDERBOT, BART, GPT) and retrieval-augmented models that condition responses on retrieved context from knowledge bases or conversation history.

Key papers