KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Retriever-Reader
Two-step QA architecture where a 'retriever' first selects relevant documents, then a 'reader' extracts the precise answer from these documents to maximize efficiency and accuracy.
Answer Span Prediction
Task of predicting the start and end positions of the exact answer in a given context, typically performed by classification layers on the text tokens.
Knowledge-Intensive QA
QA systems requiring access to extensive external knowledge bases to answer complex questions that cannot be solved by reasoning alone on the provided context.
Open-Domain Fact Verification
Process of validating the accuracy of a claim by searching and analyzing evidence across unrestricted sources on the web or in large document collections.
Question Decomposition
Technique that divides a complex question into multiple simpler and individually tractable sub-questions, facilitating information retrieval and composition of complete answers.
Answer Triggering
Binary classification mechanism that determines whether a retrieved passage actually contains an answer to the asked question, avoiding generating incorrect answers from irrelevant contexts.
Reader Confidence Score
Metric calculated by the 'reader' model indicating the certainty level of the extracted answer, often used for reranking or filtering candidate answers in QA pipelines.
Reranking Model
Specialized model that reevaluates and reorders an initial list of retrieved documents or passages to improve the precision of the most relevant results before the answer extraction step.
Passage Overlap
Phenomenon where several retrieved passages contain redundant or overlapping information, requiring deduplication or merging techniques to optimize answer presentation.
Knowledge Graph Integration
Incorporation of structured knowledge graphs into QA systems to provide verifiable factual information and complement answers based on unstructured text.
Answer Aggregation
Process of combining multiple candidate answers from different passages or documents to form a single, coherent, and complete answer to the asked question.