Watson (Artificial Intelligence)

from Wikipedia, the free encyclopedia

Watson is a computer program from the field of artificial intelligence . It was developed by IBM to provide answers to questions entered in digital form in natural language . The program named after Thomas J. Watson , one of the first presidents of IBM, was developed as part of the DeepQA research project.

To demonstrate its capabilities, the program competed in three episodes of the quiz program Jeopardy! Which were broadcast from February 14 to 16, 2011 ! with two human opponents who had previously won record sums on the show. The game, for which a prize money of one million dollars was awarded, was therefore compared in the media to the duel between the world chess champion Garry Kasparov and the Deep Blue computer . The system won the game with a final score of $ 77,147 versus $ 24,000 and $ 21,600 of its human competitors, respectively. In January 2017, a Japanese insurance company replaced more than 30 employees with the Watson platform. The AI ​​should check the names and details of the insured as well as their medical history and assess injuries.

Background and purpose

The ultimate goal of the project is to create a high-quality semantic search engine . This should capture the meaning of a question asked in natural language and find the relevant passages and facts within a short time in a large database, which also includes texts in natural language. Such software could support complex decisions in many areas, such as medical diagnostics, especially if they have to be made under time pressure.

Watson implements algorithms for natural language processing and information retrieval , based on methods of machine learning , knowledge representation and automatic inference . The system contains software modules for the creation of hypotheses, their analysis and evaluation. It draws on a collection of statements and extensive text collections, but is not connected to the Internet. Inquiries to Watson have so far been made in writing. Unlike current systems such as B. Wolfram Alpha does not require a formal query language . Since February 2011, IBM has been working with Nuance, a leading manufacturer of speech recognition software . The planned ability to also deal with spoken questions is intended to facilitate the use of a specialized version of Watsons in healthcare.

IBM plans to commercialize Watson-based systems over the next several years. The head of the responsible research laboratory assumes that the costs of the entire system could initially amount to several million US dollars, since the necessary hardware already costs at least one million dollars. In the context of pilot studies, the system has so far been used, among other things, to make predictions about which drugs could be effective against certain diseases; by integrating numerous sensor data and information on environmental influences, predicting which components of complex industrial machines are at risk of premature failure and should therefore be serviced; but also to suggest innovative combinations of ingredients for tasty recipes . In addition, neuromorphic chips are planned in the future , such as B. TrueNorth , to enable inputs in the form of natural language, images and videos, as well as any sensors. In addition, Watson is to relieve lawyers in the future with legal research in legal databases.

Appearance at Jeopardy!

Ken Jennings , a 74-time winner at Jeopardy! , was subject to the Watson computer program in February 2011

The quiz show Jeopardy! presents systems for the automatic answering of natural language questions with an interesting challenge, since the tasks given as answers are usually formulated in a deliberately ambiguous manner, often necessitate the linking of several facts and the right question has to be found within five seconds. The developers of the System Watson therefore set themselves the goal of beating human candidates in this game.

In the first test runs in 2006, Watson only found about 15% of 500 descriptions of previous Jeopardy! -Sends the correct question. Jeopardy's best candidates ! achieve about 95% accuracy in comparison. Over the next few years, Watson was equipped with a database of approximately 100 gigabytes of texts, including dictionaries, encyclopedias, such as: B. the entire Wikipedia , and other reference material. However, Watson has no connection to the Internet, so, like his human opponents, is on his own . Among other things, the information is statistically evaluated in order to derive meaningful references. Instead of relying on a single algorithm, Watson uses hundreds of them simultaneously to find a potentially correct answer along a path. The more algorithms independently achieve the same answer, the more likely it is considered that Watson has found the correct solution. When the system has developed a small number of potential solutions for a task, these are checked using a database in order to assess which of them can be considered potentially useful. For this purpose z. B. Checked times.

In a sequence of 20 practice games, the human candidates used the 6 to 8 seconds duration while reading the advice term to operate the buzzer and give the correct answer. The Watson system, optimized for this period of time, evaluates a response and decides whether there is sufficient certainty about the result to trigger the buzzer.

Since February 2010, Watson has been able to play human Jeopardy! -Beat candidates. IBM first recreated a practice situation in a conference room at the Thomas J. Watson Research Center in Yorktown Heights, New York , that mimicked the situation at Jeopardy , and had individuals, including previous Jeopardy candidates, participate in auditions against Watson with Todd Alan Crain from The Onion as a quiz master . The computer system on which Watson was running was electronically transmitted the guessing terms (in response to a question) and was able to operate the buzzer and use an electronic voice to give the answers in Jeopardy's own question format.

Finally, Watson competed on Jeopardy in three shows that aired between February 14 and 16, 2011, against former champions Ken Jennings and Brad Rutter, who had previously won record sums on the show. After the system Watson and the candidate Rutter were still tied after the first round, Watson emerged from the other two as the clear winner. IBM donated the US $ 1 million prize money to charitable causes. Jennings and Rutter announced they would donate half of their prizes of $ 300,000 and $ 200,000, respectively.

construction

The software engine of Watson's DeepQA. At Watson, this runs on the SUSE Linux Enterprise Server operating system .

The computer network is composed of 90 Power 750 servers 16  T B RAM . Each server has a Power7 8-core processor clocked at 3.5 GHz , with each core executing up to 4 threads simultaneously.

DeepQA was written in different programming languages; including Java , C ++ and Prolog . DeepQA is here in the form of annotators a UIMA - Pipeline implemented.

The use of UIMA Asynchronous Scaleout and Hadoop enables massive parallelization. Special UIMA annotators allow doing a figure on Hadoop MapReduce - scheme to a large number of text documents to be able to search in parallel.

functionality

How IBM Watson works

Watson takes on a jeopardy! -Answer (the question) of the moderator in electronic form using a keyboard. Such a Jeopardy! Answer can be very complex and consist of several sentences, riddles and word jokes.

Linguistic preprocessor

! The Jeopardy response is from the DeepQA engine using a linguistic - preprocessor analyzed. The logical structure is mapped as a tree in Prolog with the help of a parser of the sentence .

A tokenizer , consisting of UIMA annotators for pattern matching , takes care of mapping to lexical response types (LAT). The relationship between the parts of the sentence (the grammar ) is analyzed. This applies in particular to the pronoun (to which Watson must refer with the question to be generated by him), as well as words that indicate which class of answer (e.g. poet, country, epoch, etc.) is sought.

The pronoun - if it is not recognizable as such - is found by removing it from the question and making a statement. This is the part of the sentence that DeepQA focuses on when evaluating candidates.

Candidate generation

The candidate generation takes the prologue code of the linguistic preprocessor and forwards it to various search engines . Here, as is INDRI and Lucene used for the search of unstructured documents, which in a HDFS are stored. In addition, there are special engines that receive the LAT Prolog code and carry out SPARQL queries on semantic databases ( triplestores ) or SQL queries on relational databases that are based on DB2 . The documents cover a broader area of ​​knowledge and can be searched more quickly, while the structured and, in particular, semantic data sources offer greater accuracy.

The data comes from various sources, such as DBpedia , W quantities , Yago , Cyc , Freebase , Wikipedia , IMDB , World Book Encyclopedia , the Bible as well as various taxonomies and ontologies , literary works and articles by PR Newswire and New York Times . In addition, websites are analyzed and saved in the form of text snippets in Watson's databases.

DeepQA generates between 100 and 250 search results. These results ( candidates ) represent hypotheses for the possible answer.

In Jeopardy! Watson has no access to the Internet, only to the data in the internal databases. In principle, however, DeepQA also has the option of obtaining further information from the Internet for future applications and also taking real-time data into account with the help of web services .

Candidate evaluation

The most likely results of the search are analyzed in more detail. For this purpose, DeepQA has several thousand software agents, each of which carries out a very special analysis in parallel. Above all, this includes agents for the analysis of temporal (temporal) and spatial (spatial) relationships, taxonomies, simple calculations for arithmetic puzzles, evaluation of the acoustics for words that sound similar, Scrabble evaluation for words whose letters have been swapped, agents that search results perform a more detailed semantic analysis, as well as many others.

This analysis often encompasses a very broad spectrum of knowledge, with different candidates and knowledge domains being analyzed independently and massively in parallel by the respective agents. Since each search result is analyzed by up to a thousand agents, the number of evidence fragments analyzed at the same time multiplies. From the 100 to 250 hypotheses, up to 100,000 fragments of evidence are analyzed in independent threads. A software filter eliminates all results from agents who have not provided evidence of the correctness of a search result.

Application examples

"Olli" at CeBIT 2017

In late August 2016, 20th Century Fox released a trailer for the film The Morgan Project , which was made by Watson. It is the first trailer in film history to be created using an algorithm. The IBM manager John R. Smith stated in a blog entry that Watson had analyzed a total of 100 trailers of horror films in order to produce the approximately 60-second trailer. Watson divided these into segments, and after a visual analysis, an audio analysis and an analysis of the composition of the scenes, the artificial intelligence analyzed the film Morgan and filtered out the appropriate parts. Ultimately, the system decided on ten sequences, from which a film team then put together the trailer.

At CeBIT 2017, IBM presented an autonomous bus called Olli, which is controlled by Watson. Watson and Olli are networked, the computing power comes from IBM's data center in Frankfurt.

Watson is now also available to end users in the form of various applications. An example of this is Cognos Analytics , software for intelligent data analysis and visualization, or Watson Assistant, which can be used to create intelligent chatbots and digital assistants. Numerous other Watson services can sometimes even be used free of charge via the IBM Cloud and range from image and speech recognition to machine learning models.

In the field of oncology, Watson for Oncology advises cancer doctors in 230 hospitals around the world on the search for the best therapy in each case (as of mid-2018). However, in 2017 the head of the cancer department at Copenhagen's Imperial Hospital criticized the system sharply and stopped the experiment at his clinic.

Film documentaries

Web links

Videos

Individual evidence

  1. ^ The DeepQA Project. IBM , accessed October 14, 2012 .
  2. ^ Dylan Loeb McClain: First Came the Machine That Defeated a Chess Champion. In: New York Times. February 16, 2011, accessed September 10, 2011 .
  3. ^ A b John Markoff: On 'Jeopardy,' Watson's a Natural. In: New York Times. February 16, 2011, accessed February 16, 2011 .
  4. ↑ Watson makes Quiz Kings look old. In: Spiegel Online netzwelt. Retrieved February 16, 2011 .
  5. ↑ Till Haase: IBM software in a Japanese company: Insurance replaces 34 employees with artificial intelligence. In: DRadio knowledge. January 4, 2017, accessed January 4, 2017 .
  6. ↑ DeepQA FAQs. IBM , accessed October 14, 2012 .
  7. ^ Press release from Nuance ( Memento of February 20, 2011 in the Internet Archive ) (English) of February 17, 2011, accessed on February 24, 2011.
  8. ↑ a b c Clive Thompson: What Is IBM's Watson? In: New York Times . June 14, 2010, accessed July 2, 2010 .
  9. ↑ Steve Lohr: And Now, From IBM, Chef Watson. In: New York Times. February 27, 2013, accessed March 4, 2013 .
  10. ↑ Dharmendra S. Modha : Introducing a Brain-Inspired Computer: TrueNorth's neurons to revolutionize system architecture. IBM Research, accessed August 7, 2014 .
  11. ↑ Artificial intelligence is revolutionizing legal advice. (No longer available online.) In: Strafjournal. Archived from the original on April 13, 2016 ; accessed on April 13, 2016 .
  12. ↑ a b Computer: Watson and the Pet Shop Boys. In: Spiegel Online . July 12, 2010. Retrieved October 14, 2012 .
  13. ^ Jon Brodkin: IBM's Jeopardy-playing machine can now beat human contestants. (No longer available online.) Network World , February 10, 2010, archived from the original on June 3, 2013 ; accessed on October 14, 2012 (English).
  14. ↑ IBM Watson runs SUSE Linux Enterprise Server. Novell, accessed September 10, 2011 .
  15. ^ Amara D. Angelicas: How Watson works: a conversation with Eric Brown, IBM Research Manager. January 31, 2011, accessed February 16, 2011 .
  16. ↑ Is Watson the smartest machine on earth? February 10, 2011, accessed October 14, 2012 .
  17. ↑ Dean Takahashi: IBM researcher explains what Watson gets right and wrong. VentureBeat, February 17, 2011, accessed February 18, 2011 .
  18. ^ Adam Lally, Paul Fodor: Natural Language Processing With Prolog in the IBM Watson System. The Association for Logic Programming, March 31, 2011, accessed October 18, 2001 .
  19. ^ Adam Lally, Paul Fodor: Natural Language Processing With Prolog in the IBM Watson System. (PDF; 67 kB) May 24, 2011, accessed on October 14, 2012 (English).
  20. ↑ a b c d David Ferrucci, Eric Brown, Jennifer Chu-Carroll, James Fan, David Gondek, Aditya A. Kalyanpur, Adam Lally, J. William Murdock, Eric Nyberg, John Prager, Nico Schlaefer, Chris Welty: The AI ​​behind Watson - The Technical Article. In: AI Magazine . AAAI , 2010, accessed October 14, 2012 .
  21. ^ Howard Baldwin: Smarter is: Making Watson smarter… faster. (No longer available online.) IBM Data Management magazine, April 15, 2011, archived from the original on January 28, 2015 ; Retrieved October 28, 2012 .
  22. ↑ Julia Alexander: Watch the first ever movie trailer made by artificial intelligence In: polygon.com, September 1, 2016.
  23. ↑ Giuseppe Rondinella: KI 'Watson' develops the perfect film trailer for 'Morgan' In: horizont.net, September 2, 2016.
  24. ↑ A bus that doesn't need a driver In: Süddeutsche Zeitung, December 13, 2016.
  25. ↑ The Olli bus doesn't need a driver In: golem.de, March 22, 2017
  26. ↑ Olli: Local Motors builds autonomous buses for the railways In: cebit.de, January 2, 2017
  27. ↑ What is a chatbot. IBM, accessed May 7, 2019 .
  28. ↑ Watson Services in the IBM Cloud. IBM, accessed May 7, 2019 .
  29. ↑ Christian J. Meier: IBM's virtual doctor makes mistakes. In: www.nzz.ch. August 9, 2018, accessed August 9, 2018 .