DeepQA

DeepQA is an IBM research architecture developed for the Watson question-answering system. Introduced in connection with IBM’s Jeopardy! challenge, DeepQA represents a probabilistic, evidence-based approach to answering natural-language questions by evaluating a large number of candidate hypotheses against a broad corpus of information.

The architecture is modular and highly parallel, composed of thousands of independent components that perform stages such as question analysis, hypothesis generation, evidence retrieval, and hypothesis scoring. Each candidate answer is assessed by multiple evidence models that measure features such as linguistic relevance, answer-type compatibility, source credibility, and the amount and quality of supporting text. A fusion or confidence module aggregates these scores to produce a single confidence score for each hypothesis, after which the top-scoring answer is selected for presentation, with accompanying justification drawn from retrieved evidence.

DeepQA integrates natural-language processing, information retrieval, machine learning, and probabilistic inference, relying on diverse signals rather than a single algorithm. It is designed to scale to large data sets and to operate under uncertainty and ambiguity in questions. The DeepQA approach has influenced later research in open-domain question answering and the design of large-scale QA systems beyond the Jeopardy! application.