IBM Watson is an AI tool that one can use to improve applications and
services offered by a company or organization. According to the IBM
Watson Dashboard, one can use the tool to predict outcomes more
accurately, automate business processes and decision making, and
reduce the time needed to complete mundane tasks (Watson Is AI For
Business, n.d.).
IBM uses many different methods and algorithms to offer the Watson
AI tool to the masses. We have studied some of these algorithms in
the MSDA program, such as machine learning and deep learning. For
example, IBM initially used a deep learning algorithm to train a
computer to defeat Gary Kasparov, a famous world chess champion
(Wang, 2017).
When an IBM team trained a computer to answer questions similar to
those posed in the "Jeopardy" game show, IBM also experimented
with several machine learning algorithms that we have also studied.
(Wang, 2017). For example, IBM tried several algorithms to optimally
train the computer, including decision trees, deep learning, support
vector machines, and logistic regression. (Wang, 2017).
I found it interesting that the team tried to use deep learning but
found it was not as optimal as the logistic regression model that the
team settled on for a final solution (Wang, 2017). Mr. Wang says that
they probably did not have enough training data to use a deep learning
algorithm (Wang, 2017). As we learned in DATA 640, deep learning
algorithms typically require a lot of data, as evidenced by real-world
applications that accurately classify images. Deep learning also
requires a lot of computing power to iterate through massive datasets
for the computer to "learn" optimal actions to take based on varying
states in the environment. This need for computing capacity and
massive datasets makes IBM Watson well suited for applications that
leverage deep learning algorithms.
Switching gears, IBM Watson also has powerful Natural Language
Processing (NLP) capabilities. The role of NLP in Watson applications
depends on the application. For example, a text-to-speech application
can leverage NLP to convert written text into audio that a human can
hear. This type of application could help individuals who are visually
impaired and unable to read text on the screen. In contrast, a speech-
to-text application can leverage NLP to convert spoken words into
text that one can read, which could help the hearing impaired.
When I was first assigned to lead our agency's first AI initiatives, one
of the first services I experimented with was IBM Watson Discovery
and Knowledge Studio. Our agency publishes the results of many
audits and also research reports as part of our mission to help ensure
the accountability, integrity, and efficiency of the United States Postal
Service. You can access our audit reports here: Document Library |
USPS Office of Inspector General (uspsoig.gov).
One of my tasks was to experiment with importing several of these
audit reports in PDF format to see if I could gain additional insights
into the reports, such as mentions of specific people or locations,
currency amounts, and keywords. I was impressed with the
capabilities and how well Knowledge Studio could extract such data
and enable one to query using keywords, a person's name, or location.
However, I also found some nuances which are difficult for NLP to
overcome. For example, one of our audit directors has the last name of
"Poland." As best as I can recall, the NLP algorithm picked this up as a
location instead of a person. Another challenge was around security.
Our Office of General Counsel requires redaction of certain
information that does not get published to the public. It would have
been difficult, if not impossible, for us to leverage IBM Watson
Knowledge Studio to analyze non-redacted reports since the IBM
Cloud is a public-facing service. We have stringent information
security standards that a company must meet in the Federal
government. These requirements can sometimes preclude us from
using cloud-based services available to the private sector.
Another NLP application is machine reading comprehension, where
the computer parses a long passage of text and then summarizes the
key points in the text. For example, our Office of General Counsel
often has to read through voluminous legislation to understand its
applicability and impact on our operations. One could use NLP to
summarize the text's key points and even use text analytics to extract
critical data, such as keywords, currency amounts, locations, and
entities such as other federal agencies or persons of interest. This
capability could save our agency a lot of time. One could even feed the
extracted data into a machine learning pipeline to perform tasks such
as classification/labeling of the data. Microsoft has an exciting demo
on machine reading comprehension for anyone interested: Machine
Reading Comprehension - Microsoft AI Lab. Using the IBM Watson
Knowledge Studio demo, one can also see similar text analytics
capabilities. f
I'm interested to hear what other IBM Watson services my classmates
can use in a specific industry domain. It is always enjoyable to read
about other industry domains outside the Federal government.
Reference:
Wang, J. (2017, July 12). ARK Invest. Retrieved from https://ark-
invest.com/articles/analyst-research/ibm-watson/#
Watson is AI for Business. (n.d.). IBM.com. Retrieved from
https://cloud.ibm.com/developer/watson/dashboard
