Watson (computer)
Watson is a computer program in the field of artificial intelligence. It was developed by IBM to provide answers to questions that are entered in digital form in natural language. Named after Thomas J. Watson, one of the first president of IBM, named program was developed as part of the DeepQA research project.
To demonstrate its efficiency, the program competed in three radiated from 14 to February 16, 2011 Follow the game show Jeopardy with two human opponents who had won on the show previously record sums. The game, for a prize of one million dollars was awarded, was therefore compared in the media with the duel of world chess champion Garry Kasparov against the computer Deep Blue. The system won the game with a final score of $ 77,147 compared to $ 24,000 or $ 21,600 of his human competitors.
- 5.1 videos
Background and Objectives
The project's goal is ultimately to create a high-quality semantic search engine. This is to grasp the meaning of questions in natural language question in a large database, which also includes texts in natural language in a short time find relevant passages and facts. Such software could in many fields, such as medical diagnosis, support complex decisions, especially if these have to be made under time pressure.
Watson implemented algorithms of the Natural language processing and information retrieval, based on methods of machine learning, knowledge representation and the automatic inference. The system includes software modules for the creation of hypotheses, their analysis and evaluation. It draws on a collection of statements and large text collections, but is not connected to the Internet. Inquiries to Watson are currently provided in text form. Unlike current systems such as Wolfram Alpha, however, it does not require any formal query language. Since February 2011, IBM is working with the company Nuance, a leading supplier of software for voice recognition. The planned ability to work well spoken questions, is to facilitate the use of a specialized version of Watson in healthcare.
IBM plans to market commercially on Watson -based systems over the next few years. The head of the relevant research laboratories assumes that the cost of the overall system could initially amount to several million dollars because already have the necessary hardware costs at least a million dollars. As part of the pilot studies, the system was previously used, among other things, to predict which drugs might be effective against certain diseases; predict by integration of multiple sensor data and information on environmental influences which components of complex industrial machines run the risk of prematurely fail and therefore should be maintained; but also to propose innovative combinations of ingredients for tasty recipes.
Appearance on Jeopardy!
The quiz show Jeopardy! provides systems for automatically answering natural language questions before an interesting challenge because the tasks set as responses are usually formulated deliberately ambiguous, often make the linkage of several facts required and the appropriate question within a time limit of five seconds must be found. The developers of Watson system therefore set themselves the goal to beat in this game human candidates.
In a sequence of 20 practice games the human candidates took the 6 to 8 seconds in duration during the reading of the rate concept to press the buzzer and give the correct answer. The optimized system at this time Watson evaluates an answer and decides whether there is enough certainty about the outcome to trigger the buzzer.
Since February 2010, Watson is able to beat! Candidates under rule just matches human Jeopardy. IBM introduced initially an exercise situation in a conference room at the Thomas J. Watson Research Center in Yorktown Heights, New York, after which mimics the situation in Jeopardy, and let individuals, including former Jeopardy candidates to participate in auditions against Watson, Todd Alan Crain from The Onion as quizmaster. The computer system on which Watson has been executed, the rate terms have been transmitted (in response to a question ) electronically and it was able to press the buzzer and give the answers in Jeopardy own question format with an electronic voice.
Finally, Watson joined Jeopardy in the three programs that were broadcast between 14 and 16 February 2011 against the former champions Ken Jennings and Brad Rutter, who had won on the show earlier record sums. After the system Watson and the candidate Rutter were still after the first round on a par, Watson went from the other two emerged as the clear winner. The prize money of one million U.S. dollars, IBM charitable purposes. Jennings and Rutter announced plans to donate half of their prices of $ 300,000 and $ 200,000.
Construction
The software engine of Watson is DeepQA. This runs at Watson on the operating system SUSE Linux Enterprise Server.
The computer network consists of 90 Power 750 servers with 16 TB of RAM. Each server has a clocked with 3.5 GHz POWER7 eight core processor, with each core running up to four threads simultaneously.
DeepQA was written in different programming languages ; including Java, C and Prolog. DeepQA here is implemented in the form of a UIMA annotators pipeline.
Through the use of UIMA Asynchronous Scaleout and Hadoop a massive parallelization is possible. Special UIMA annotators make this a picture on Hadoop MapReduce scheme to a large number of text documents to be able to search in parallel.
Operation
Watson takes a Jeopardy! Response ( the question ) in electronic form. Such Jeopardy! Response can be very complex and consist of several sentences, word puzzles and jokes.
The Jeopardy! Response is analyzed by the DeepQA engine with the help of a linguistic preprocessor. The logical structure is imaged by means of a parser of the sentence as a tree in Prolog.
A tokenizer, consisting of UIMA annotators for pattern matching, takes care of the mapping to lexical answer type (LAT ). The relationship of the parts of sentences to one another is analyzed ( the grammar). This particularly applies to the pronoun, as well as words that indicate which class of response (eg poet, country, era, etc.) is sought.
The pronoun is - if this is not recognized as such - thus found that by its removal from the question is a statement. In this part of the sentence DeepQA focuses on the candidate evaluation.
Candidate generation
The candidate generation accepts the Prolog code of the linguistic preprocessor and passes them to various search engines on. Here about INDRI and Lucene for searching of unstructured documents is used, which are stored in a HDFS. In addition, there are special engines which accept the LAT prolog code and perform SPARQL queries on semantic databases (triple gate ) or SQL queries on relational databases, which are based on DB2. The documents cover this from a wider field of knowledge, and are quickly searchable, while the structured and in particular semantic data sources provide higher accuracy.
DeepQA generated while between 100 and 250 search results. These results ( candidates) making a hypothesis for the possible s response
In Jeopardy! Watson has no access to the Internet, but only on the data in the internal databases. In principle DeepQA has for future applications but also the possibility of further information about it from the Internet and to consider using web services but also real time data.
Candidate rating
The finding likely results will be analyzed in more detail. For this DeepQA has several thousand software agents each performing a special analysis in parallel. This primarily includes agents for the analysis of temporal ( temporal ) and spatial ( spartiellen ) contexts, taxonomies, simple calculations for computational puzzle, evaluation of acoustics for words that sound similar, Scrabble Rating for words whose letters have been mixed, agents, the search results a more detailed semantic analysis perform as well as many others.
This analysis often includes a very broad range of knowledge, with different candidates and knowledge domains are analyzed by each agent independently and massively parallel. Since each search result is analyzed by up to a thousand agents, the number of simultaneously analyzed evidence fragments multiplied. Thus, up to 100,000 fragments of evidence are to be analyzed in independent threads from the 100 to 250 hypotheses. A software filter eliminates all the results of agents who have not provided any proof of the correctness of a search result.