In 2007, IBM Research took on the grand challenge of building a computer system that could compete with champions at the game of Jeopardy!, an American nationally televised quiz show. In 2011, the open-domain question-answering system, named Watson, beat the two highest ranked players in a two-game Jeopardy! match. DeepQA is the name given to the question-answering (QA) system architecture, which was developed for Watson to play the Jeopardy! game. In the tradition of IBM’s grand challenges such as Deep Blue with chess and Blue Gene with the human genome, Watson was presented as the next scientific breakthrough with natural language processing on Jeopardy!
The Watson services available today evolved from the original DeepQA architecture that defeated the Jeopardy! quiz show champions to a set of RESTful APIs delivered on the IBM Cloud which is the platform as a service (PaaS) offering from IBM.
This section discusses key capabilities that had to be developed to make question-answering systems practical in real-world applications (and why the DeepQA architecture had to evolve).
At a high level, DeepQA generates and scores many hypotheses by using an extensible collection of natural language processing, machine learning, and reasoning algorithms, which gather and weigh evidence over both unstructured and structured content to determine the answer with the best confidence.
The primary computational principle supported by DeepQA was to assume and pursue multiple interpretations of the question, to generate many plausible answers or hypotheses, and to collect and evaluate many competing evidence paths that might support or refute those hypotheses.
![[Pasted image 20230613195021.png]]
In the question analysis step, parsing algorithms decompose the question into its grammatical components. Other algorithms in this step identify and tag specific semantic entities, such as names, places, or dates. In particular, the type of thing being asked for, if it is indicated at all, will be identified. This is called lexical answer type (LAT) and it is the word in the question that indicates the type of answer to look for.
In the hypothesis generation step, DeepQA does a variety of broad searches for each of several interpretations of the question. These searches are performed over a combination of unstructured data (natural language documents) and structured data (available databases and knowledge bases) fed to Watson during training. The focus, at this point, is on generating a broad set of hypotheses, which for this application, are called candidate answers.
In the hypothesis and evidence scoring step, the candidate answers are first scored independently of any additional evidence by deeper analysis algorithms.
In the merging and ranking step, the many possible answers are scored by many algorithms to produce hundreds of feature scores. Trained models are applied to weigh the relative importance of these feature scores. These models are trained with machine learning methods to predict, based on past performance, how best to combine all these scores to produce final, single-confidence numbers for each candidate answer and to produce the final ranking of all candidates. The answer with the strongest confidence is Watson’s final answer.
Approximately the same time that Watson was competing on the Jeopardy! quiz show, the technology world was being disrupted by three fundamental forces:
- The advancement in machine learning capabilities, opening up new applications of predictive analytics, natural language processing (NLP), speech recognition, and computer vision.
- The ability for these capabilities to be offered through APIs on the cloud, massively decreasing the time-to-value of cognitive computing.
- The amount of data that is available and the potential to harness it beyond traditional analytics.
Data has become the competitive advantage and most of it is invisible to traditional computing platforms. In healthcare, each year seems to produce data in an exponential amount compared to prior years. Understanding unstructured data in the form of text documents, images, videos, and raw sensor output provides the vast proportion of the opportunity.
IBM soon realized that the research efforts that led to the design of the DeepQA architecture and its success playing the Jeopardy! game, had to be capitalized and evolved to solve real world problems.
In 2014, IBM formed the Watson Group to commercialize Watson technology. Work done with Oncologists at Memorial Sloan Kettering Cancer Center became the Watson Oncology Advisor. Efforts to organize and unlock the value of text documents became Watson Explorer and Watson Discovery Advisor. These were the first solutions built with what we know today as IBM Watson services.
Overtime, the steps in the DeepQA pipeline described in DeepQA architecture became microservices that greatly enhanced the ability to customize the system for a particular domain and to integrate individual services to build a complete solution.
The Watson Conversation service was developed to enable a developer to automate branching conversations between a user and the application. The Conversation service, renamed Watson Assistant in 2018, enables applications to use natural language to automatically respond to user questions, cross-sell and up-sell, guide users through processes or applications, or even “hand-hold” users through difficult tasks. The Watson Assistant service can track and store user profile information to learn more about that user, guide that user through processes based on the user’s unique situation, or pass the user information to a back-end system to help the user take action and get the help needed.
The Watson Assistant service allows you to quickly build, test, and deploy a bot or virtual agent across mobile devices, messaging platforms such as Slack, or even on a physical robot. Watson Assistant has a visual dialog builder to help you create natural conversations between your apps and users, without requiring any coding experience.
As these services are implemented in production applications, handling multiple modes of communication is necessary. Speech is common in mobile, home automation, and wearable applications. Being able to understand and translate between languages is critical to expanding globally. And finally, the ability to understand images and videos can greatly expand the potential of engagement applications in domains, like robotics and connected cars.
![[Pasted image 20230613200658.png]]
![[Pasted image 20230613200716.png]]
![[Pasted image 20230613200759.png]]
Speech to Text, Text to Speech, Visual Recognition, Language Translator, Tone analyzer, and Personality Insights services were developed to round out machine perception capabilities of Watson in addition to the NLP tools.
Watson Assistant orchestrates the capabilities represented by Watson Natural Language Classifier, Watson Natural Language Understanding (which are NLP services), and dialog capabilities and exposes them through a single tool. Additional capabilities, like Watson Tone Analyzer, Speech to Text, and Text to Speech integrations, are planned in the future (although developers can orchestrate these services themselves now in the applications layer).
Watson Discovery service allows users to extract value from unstructured data by converting, normalizing, and enriching it. By using a simplified query language, a user can explore the data or quickly tap into pre-enriched data sets, such as the Discovery News collection. Discovery News primarily includes English language news sources that are updated continuously, with over 300,000 new articles and blogs added daily, from more than 100,000 sources.
Besides the Watson services (AI APIs), IBM introduced the following tools to help you build intelligent applications:
- Watson Knowledge Studio to teach Watson the language of a domain and help domain experts discover meaningful insights in unstructured text.
- Watson Studio to build and train AI models, and prepare and analyze data, all in one integrated environment.
- Watson Machine Learning to use domain data to create, train, and deploy self-learning models.
- Watson Knowledge Catalog is a machine learning data catalog (MLDC) that enables you to access, curate, categorize, and share data, knowledge assets, and their relationships, wherever they are. Dynamic masking of sensitive data elements unlocks information for AI application development.
![[Pasted image 20230613200839.png]]