AI Glossary
The complete dictionary of Artificial Intelligence
Question Answering System
Computer system capable of understanding natural language questions and providing relevant answers based on a knowledge base or documents. These systems generally combine NLP techniques, information retrieval, and automated reasoning.
Retrieval-Based QA
Question-answering approach that identifies and extracts answers directly from a corpus of pre-existing documents without generating new content. This method focuses on searching and selecting the most relevant passages to answer the asked question.
Generative QA
Question-answering system that generates new answers in natural language rather than extracting existing passages, using advanced language models like GPT. This approach allows formulating original and coherent answers based on context understanding.
Open-Domain QA
Question-answering system capable of answering questions on any topic without domain restrictions, generally relying on large corpora like Wikipedia or the web. These systems must face much greater ambiguity and diversity than closed-domain systems.
Closed-Domain QA
Question-answering system specialized in a specific domain like medicine, law, or finance, with a limited and well-defined knowledge base. This specialization allows achieving greater accuracy and depth in the provided answers.
Reading Comprehension
Ability of a system to understand and interpret a given text to answer specific questions about this text, evaluated on datasets like SQuAD. This fundamental skill in NLP tests semantic understanding and reasoning on documents.
Knowledge Graph QA
Question-answering approach that queries structured knowledge graphs to find factual answers, often requiring the conversion of natural questions into formal queries. This method excels for precise factual questions about entities and their relationships.
Neural Question Answering
Use of deep neural networks to implement question-answering systems, allowing automatic learning of complex language representations. These approaches have largely surpassed rule-based or manual feature methods.
Transformer Models for QA
Neural network architecture based on attention mechanisms that currently dominates modern question-answering systems like BERT and RoBERTa. Transformers enable efficient capturing of long-distance dependencies in text.
BERT for QA
Application of the BERT (Bidirectional Encoder Representations from Transformers) model, specifically fine-tuned for question-answering tasks, becoming a benchmark in the field. BERT excels in bidirectional contextual understanding of language.
Passage Retrieval
Crucial step in QA systems that involves identifying and selecting the most relevant text passages from a large corpus to answer a question. This step often uses semantic search and vector similarity techniques.
Answer Extraction
Process that identifies and isolates the exact answer within a relevant text passage, often treated as a sequence classification or named entity recognition task. This technique is central to extractive QA systems.
Factoid QA
Type of question-answering that focuses on factual questions with short, precise answers like names, dates, or numbers. These systems are evaluated on their ability to retrieve exact factual information from documents.
Non-Factoid QA
Category of question-answering that requires descriptive answers, explanations, or summaries rather than isolated facts, such as 'why' or 'how' questions. These systems require deeper understanding and reasoning capabilities.
Conversational QA
Question-answering system capable of maintaining a coherent dialogue over multiple turns, taking into account the conversational context and previous questions. This approach requires managing history and anaphoric references.
Multi-Hop QA
Advanced question-answering system that requires combining information from multiple documents or passages to formulate a complete answer. This approach simulates multi-step reasoning to answer complex questions.
Zero-Shot QA
Ability of a system to answer questions in domains or on topics for which it has not been specifically trained, relying on its general knowledge. This skill demonstrates true understanding and generalization of language.
Fine-Tuning for QA
Process of adapting a pre-trained language model specifically for question-answering tasks using annotated datasets like SQuAD or Natural Questions. This technique allows achieving optimal performance on specific QA tasks.
Attention Mechanism in QA
Essential component of modern QA models that allows the system to focus on the most relevant parts of the context to answer a given question. The attention mechanism weighs the importance of each word in the context relative to the question.
Question Classification
Preliminary step in QA systems that categorizes questions according to their type (who, what, when, where, why, how) to adapt the answer search strategy. This classification helps guide the system towards the right sources and extraction methods.