AI 용어집
인공지능 완전 사전
Dialogue State Management
Core component of a conversational agent that tracks and updates the conversation context, including user intents, collected entities, and exchange history to ensure consistency of interactions across multiple turns.
Natural Language Generation (NLG)
Module that converts a semantic representation or dialogue action into a coherent, grammatically correct phrase or text adapted to the conversation tone for smooth human interaction.
Conversational History Tracking
Process of storing and analyzing the complete sequence of exchanges between the user and the agent, enabling reference to past elements and maintaining long-term memory of the conversation.
Anaphora and Coreference Resolution
System's ability to identify and interpret pronouns or expressions that refer to entities mentioned earlier in the dialogue, essential for understanding implicit requests and maintaining context.
Turn Management
Set of rules and strategies governing the alternation of interventions between the user and the agent, including end-of-turn detection, interruption management, and taking initiative in the conversation.
Frame-based Dialogue System
Architecture where dialogue is structured around predefined 'frames' or templates representing information needed to resolve a specific task, such as booking a flight or ordering a product.
Dialogue Language Model
Type of language model (e.g., GPT, BERT) specifically trained or fine-tuned on large conversational data corpora to understand and generate text adapted to the multi-turn context of dialogues.
User Intent Detection
Classification task aimed at identifying the underlying goal or need of the user from their utterance, constituting the crucial first step to guide the agent's response.
Nested Entity Resolution
Advanced information extraction process that identifies entities (names, dates, locations) that may be contained within each other within the same statement, complicating the understanding of context.
Clarification Strategy
Set of tactics used by the agent to request clarifications when the user's input is ambiguous, incomplete, or contradictory, in order to gather necessary information before proceeding.
Long-term Coherence Management
Challenge of maintaining logical, factual, and stylistic coherence of a conversational agent over very long interactions, avoiding contradictions and forgetting of the initial context.
Hybrid Dialogue System
Architecture combining rule-based approaches (for predictability and security) and machine learning model-based approaches (for flexibility and natural language understanding).
Dialogue Quality Evaluation
Process of measuring the performance of a conversational agent using automatic metrics (e.g., success rate, number of turns) and human evaluation (relevance, coherence, engagement) to quantify its effectiveness.
Fallback Mechanism
Safety strategy activated when the system fails to understand the user's intent or execute an action, typically involving asking a clarification question or transferring to a human operator.
Task-oriented vs. Social Dialogue
Distinction between agents designed to accomplish a specific objective (booking, technical support) and those aiming to simulate human conversation for entertainment or companionship, with very different design requirements.