IBM Watson is a computer system capable of answering questions posed in natural language. It was developed as part of IBM's DeepQA project by a research team led by principal investigator David Ferrucci. Watson was named after IBM's founder and first CEO, industrialist Thomas J. Watson. The system was initially designed to compete on the quiz show Jeopardy!, where it defeated champions Brad Rutter and Ken Jennings in 2011, winning the first-place prize of US$1 million. Watson's success demonstrated the potential of Artificial intelligence in open-domain question answering and sparked commercial applications in fields such as healthcare.
Watson's architecture integrates advanced Natural language processing techniques, information retrieval, knowledge representation, and automated reasoning. IBM stated that Watson uses "more than 100 different techniques to analyze natural language, identify sources, find and generate hypotheses, find and score evidence, and merge and rank hypotheses." Over time, Watson's capabilities were extended, and its deployment shifted to IBM Cloud, with evolved machine learning and optimized hardware.
Software
Watson's core software is IBM's DeepQA, built on the Apache UIMA (Unstructured Information Management Architecture) framework. The system was written in multiple languages, including Java, C++, and Prolog, and ran on SUSE Linux Enterprise Server 11 with Apache Hadoop for distributed computing. Beyond DeepQA, Watson included strategy modules. For example, one module calculated Final Jeopardy wagers based on confidence scores and contestant standings. Another used Bayes' rule to estimate the probability of a Daily Double, drawing on historical data from the J! Archive. If a Daily Double was found, the wager was computed by a two-layered neural network similar to those used by TD-Gammon, a backgammon-playing program developed by Gerald Tesauro in the 1990s. These strategy parameters were tuned by benchmarking against a statistical model of human contestants fitted on J! Archive data.
Hardware
Watson was workload-optimized, integrating massively parallel POWER7 processors. It employed a cluster of ninety IBM Power 750 servers, each with a 3.5 GHz POWER7 eight-core processor and four threads per core, totaling 2,880 processor threads and 16 terabytes of RAM. According to John Rennie, Watson could process 500 gigabytes (equivalent to a million books) per second. IBM master inventor Tony Pearson estimated the hardware cost at about three million dollars. Its Linpack performance was 80 TeraFLOPs, about half the cutoff for the Top 500 Supercomputers list. For Jeopardy!, all content was stored in RAM because hard drives would be too slow for real-time responses.
Data
Watson's knowledge sources included encyclopedias, dictionaries, thesauri, newswire articles, and literary works. It also used databases, taxonomies, and ontologies such as DBpedia, WordNet, and YAGO. The IBM team provided millions of documents to build Watson's knowledge base, enabling it to answer a wide range of questions.
Operation
Watson parsed questions into keywords and sentence fragments to find statistically related phrases. Its main innovation was not a new algorithm but the ability to execute hundreds of proven language analysis algorithms simultaneously. The more algorithms that independently found the same answer, the higher the confidence. Once Watson had a few potential solutions, it checked them against its database to determine plausibility.
Comparison with human players
Watson had advantages and disadvantages versus human Jeopardy! players. It lacked contextual understanding, and humans often generated responses faster, especially for short clues. Watson's programming prevented it from buzzing before being confident, but it had faster reaction time once ready and was immune to psychological tactics like category jumping. In mock games, humans used the six to seven seconds Watson needed to process clues to decide whether to buzz. Watson's electronic circuitry received the ready signal and evaluated confidence before buzzing, making its reaction faster than humans except when they anticipated the signal. Watson spoke with an electronic voice synthesized from recordings by actor Jeff Woodman for IBM's text-to-speech program.
Commercial applications and divestiture
In February 2013, IBM announced Watson's first commercial application: utilization management decisions for lung cancer treatment at Memorial Sloan Kettering Cancer Center, in conjunction with WellPoint (now Elevance Health). In 2022, IBM divested its Watson Health division into Merative, which was sold to Francisco Partners, an American private equity firm. The division cost $4 billion to develop but sold for $1 billion. By 2023, Watson had resulted in IBM losing 10% of its stock value, costing four times more than it brought in, and leading to mass layoffs. Despite these setbacks, Watson remains a landmark in the history of Artificial intelligence and Machine learning.