Cognitive computing represents a different paradigm in which a software solution is not statically programmed, but rather is “learned” through a combination of available data sources and training activities. It is particularly well suited to addressing complex situations involving unstructured content and uncertainty. Both the capabilities and requirements are quite different from those of more traditional software based solutions. Typically it is not seen as a replacement for existing solutions, but rather as an augmentation to address parts of the overall problem that are difficult or impossible using other approaches.
Historically, computing systems have been extremely adept at well-understood numerical or procedural operations, but it has been difficult to apply them to problems involving natural language, either for user interaction or to extract knowledge from that type of content.
For most existing software systems the human user has to learn to interact with the system on its terms. Although in the hands of an experienced user, this can be an efficient interface, allowing a user to interact with the system simply by talking with it in the form of an interactive conversation is more intuitive. This form of interaction using natural language can it open a tool to a wider, potentially untrained, set of users and is also an inherently mobile-friendly approach.
Cognitive computing is particularly well suited to deal with the huge amount of natural language that is currently underutilized. Some of the more obvious examples of this kind of content include books, journal articles, white papers, blogs, tweets, etc. As valuable as the knowledge contained in this content might be, until recently the potential to leverage it with computing systems has been limited to simple things like indexing and keywords searches. Although these forms of access are not without value, they fall well short of what a human expert could do. Cognitive computing is a significant step forward to provide the capability to fully leverage this knowledge. This is important because there is a common shortage of experts in any given domain, and the quantity of the content is growing much faster than a human or even teams of humans could ever read, understand, remember and apply.
Perhaps the best-known example of a cognitive computing platform is the IBM technology that played and won the Jeopardy game in 2011. That particular technology has since evolved from a specific game-playing engine to a general-purpose tool currently focused on unstructured natural language content.
For better or worse, IBM has elected to brand a number of different products “Watson”, some of them are much more directly related to the technology developed for Jeopardy than others. Although over time it is expected that various capabilities will migrate between the various Watson branded products, currently they exist in separate silos with distinct functionality. Out of the myriad of relatively new Watson branded technologies we have been heavily involved in leveraging three of them:
Watson Question and Answer [Watson QA or WQA] is the product that most people think of when hearing about “Watson” in the context of cognitive computing. Currently everything about WQA is focused on providing a “single best answer” to a query presented in the form of a natural language question. This includes not only the processing pipeline, but also the development tools, training methodology, and user interaction model. Although the technology is evolving rapidly, currently WQA requires that the desired answer exists somewhere within a single document in it’s previously ingested corpus. Furthermore, the nature of the question must be answerable using primarily linguistic logic. For example it can answer questions relating to facts or definitions, but can’t perform even simple numerical calculations. WQA is currently available to Watson Ecosystem Partners and in very limited form on the Bluemix platform.
Watson Discovery Advisor [WDA], like WQA, utilizes a corpus of ingested natural language material. It currently differs from WQA in purpose, processing pipeline and business model. Unlike WQA, WDA has been designed to interactively explore complex topics for which there might not be a “single best answer”. It is particularly adept at allowing the user to discover and understand complex relationships and patterns. WDA is provided in a Software as a Service (SAAS) model.
Although Watson Analytics [WA] also has a natural language interface, unlike WQA and WDA, it is primarily targeted at numerical analysis and presentation and answering questions about tabular data rather than natural language content. (Think spreadsheets vs. journal articles.) It is significantly different from other tools in that the tool internally performs many operations that have previously been left up to the user. Activities such as evaluating the quality of the data are performed automatically. Novice users without a strong data science background should be able to use it to produce valuable insight while still allowing more advanced capabilities for those who do understand the underlying math.
WA is currently available in a freemium model as a hosted web application. You work with the tool on the web after first uploading your data. You can use the basic features for free, but if you outgrow the limits for data size or complexity, there is a paid subscription tier.