Abstract
Computers scientist have been grappling with the challenge of developing cognitive systems. Some of the challenges that have made the creation of such systems include natural language processing and machine learning. However, IBM has been able to produce a system known as Watson that has demonstrated human-like intelligence. The system was able to deal with the challenge of natural language processing by using a parsing process to derive the context of questions. This made IBM Watson have human-like understanding of natural language. The system is also built on active technologies that facilitate adaptation training through data and user changes. To find the best responses to questions genrerated by user, IBM Watson uses a combination of two algorithms. One algorithm returns whole documents while the other returns the specific phrase that answers the question.
Introduction
Since the invention of artificial intelligence, computer scientists have been trying to create machines that are able to think like humans. The increase in processing power and data storage of computers has made this goal feasible. Developing such systems could have a major advantage in many business sectors by amplifying human intelligence and for better decision-making (Deloitte, 2015). Currently, there are cognitive systems that can learn through interaction with humans and data. The systems have the ability to adapt and become smarter over time. An example of such a system is IBM’s Watson that uses the Internet and internal data as its extended mind. IBM Watson aims at effectively applying intelligent problem-solving techniques to increase the capabilities of current information technology. The system is mainly focused on unstructured data that makes up a majority of big data.
Challenges in Cognitive Learning
IBM Watson can be described as a deep natural language system that has question answering capabilities. For the system to be effective, it must have natural language processing and machine learning capabilities. These are the two main subfields of cognitive computing. The system should be able to accept questions from users in natural language and provides the users with answers. The provided answers should also be as accurate as possible. IBM Watson provides responses to questions by putting the questions into context as opposed to key-word search engines that simply retrieve relevant documents. The context is assessed from both from immediate information found within the question and IBM Watson’s knowledge base that is mostly in the form of unstructured information. This unstructured information is in natural language and makes up to 80% of data found on the internet today. This means that to find responses, the system should be able to process this information that is available to it. The other challenge is how to improve the accuracy of answers through machine learning.
Question Processing
The asking of questions and retrieving of responses by users is allowed by the Question and Answer Application Programming Interface (QAAPI) that is present in IBM Watson (High, 2012). The QAAPI allows users to post their questions in natural language and also submit feedback on the received responses. Once the system, receives the questions, it parses the question to derive its context. Although IBM Watson does not understand the individual words of human language, it understands its features. Understanding of these features is what allows the system to extract the question’s context as it is able to take apart the language and determine inferences between the question’s texts and the answer’s text. This inference identification process is done with human-like accuracy due to the temporal and spatial constraint of the context matter. This is enabled by the nearly infinite number of rules that is used to simulate every case that can be found in human language. The inference process is also done at speeds and scales exceeding any human capabilities.
Generation of Hypothesis
IBM Watson then generates a set of hypotheses by looking into its knowledge base known as the corpus that may contain valuable responses. The corpus contains all kind of unstructured information in the form of journals, news, text books, guidelines, and others. IBM Watson while ingesting the corpus goes through whole content and converts it into a form that will be easier for it to use. The system also performs curator that quickly and automatically manages contested information resulting in the collection of high-quality information. This involves removal of any information that comes from unreliable source, irrelevant, or out of date using a guided intuitive process. The process of comparing the question to the possible answers in the corpus is done by hundreds of reasoning algorithms. Each of these algorithms performs a different function, some match terms and find synonyms, others identify temporal and special features in the question, and others find relevant sources of information.
Response Selection
The output from the IBM Watson does not only produce the response but also the confidence ratings and supporting evidence. Every reasoning algorithm produces one or more confidence scores. The scores are based on the specific area that the reasoning algorithm focuses on and shows the level to which the potential response is inferred to the question. The results are then weighted against a statistical model that indicates how well the algorithm performed at establishing inferences between the response and the question. The statistical method is then used to calculate the confidence level that the system has concerning the evidence used in referring the response to the question. This process is repeated for all possible responses until IBM Watson can find an answer that is stronger that the rest. The number of responses to be returned by the system can also be varied by the user. This feature is useful especially when the user wants to see the top responses that IBM Watson has chosen.
Machine Learning
Deep Question Answering (QA) Algorithm
The DeepQA algorithm used by IBM Watson is a combination of two algorithms; the Information Retrieval (IR) algorithm and the Question Answering (QA) algorithm. The IR algorithm is similar to the ones the used by search engines and returns whole documents while QA algorithm returns just a single word or phrase as a response to an inquiry (Ferucci et. al, 2010). The retrieval of answers by the DeepQA algorithm basically follows three steps; linguistic preprocessing, finding possible answers and finding the best answer.
Linguistic Preprocessing
The first step in linguistic preprocessing is the tokenization that involves the splitting of human language into smaller units such as words and punctuations. Although most people understand this concept, there is no clear way for developing tokenizers. This is because different applications and natural languages have different needs. Also, the structural ambiguity present in natural language makes it difficult to determine which tokenization will be used for a given text. Due to this reason, the common representation of tokenized text involves the inserting of spaces between characters. The algorithm then does part of speech (POS)-tagging, which simply involves tagging words to identify their context. The process mostly involves the use of supervised machine learning with labeled data. Most human languages also have different words that have the same concept. The algorithm puts all these words into their base form so as to improve recall in information retrieval. This process of finding the base form of words is known as lemmatization.
Tokenization also allows the DeepQA algorithm to decompose the questions into clues and sub-clues. This decomposition process is done through a set of Pprolog rules such as using a conjunction to indicate a boundary between sub-clues. This separation of clues and sub-clues allows IBM Watson to generate separate parse trees for each sub-clue. The algorithm also performs named entity recognition for future use in the pipeline. This typically involves the automatic annotation of occurrences that carry important semantic information about a clue. However, to allocate the semantic information to a clue the algorithm differentiates the subject and the object in a verb phrase. The focus detection of the question is the done by the algorithm to allow it to replace the clue in the question with an answer without altering its meaning. This is sometimes followed by detection of lexical answer type (LAT (Swiezinski, 2011) This may prove to be helpful in semantically interpreting of questions. The IBM Watson builds a LAT-detector using a trained supervisor with a set of question-answer-pairs that is based on the processed linguistic information.
Finding Possible Answers
One second after the system understands the answer that is expected based on the clues, the algorithm performs this next step. Depending on the identified clues gathered from the linguistic preprocessing, the algorithm applies rules to develop queries for its corpus. The generated queries comprise of a combination of search queries of the clues. The possible answers are then listed depending on the corpus resource. In most cases, the answer is among the 250 top candidates. If the answer is not in this category, then there is a high probability that IBM Watson will not get the answer to the question.
Candidate Answer Generation
References
Deloitte. (2015). Deloitte’s Point of View on IBM Watson. Retrieved from https://www2.deloitte.com/content/dam/Deloitte/us/Documents/about-deloitte/us-ibm-watson-client.pdf
Ferucci, D., Brown, E., Chu-Carroll, J., Fan, J., & Gondek, D. (2010). Building Watson: An Overview of the DeepQA Project. AI Magazine
High, R. (2012). Front cover the era of cognitive systems: An inside look at IBM Watson and how it works. Retrieved from http://www.redbooks.ibm.com/redpapers/pdfs/redp4955.pdf
IBM. (2013). Artificial Intelligence: Learning Through Interactions and Big Data. Retrieved from http://www.redbooks.ibm.com/redpapers/pdfs/redp4974.pdf
Swiezinski, L. (2011). Lifecycle of a Jeopardy Question Answered by Watson DeepQA
