Digital White Papers

38 WWW.ILTANET.ORG | ILTA WHITE PAPER KNOWLEDGE MANAGEMENT Automating Knowledge Delivery Using Knowledge Bots the minds of the many people who conversed with it." (Joseph Weizenbaum, Computer Power and Human Reason: From Judgment to Calculation, at 188-89 (1976)). Weizenbaum later wrote he was "startled" to find "clear evidence that people were conversing with the computer as if it were a person who could be appropriately and usefully addressed in intimate terms." Id. at 6-7. NLP techniques have continued to advance in the 50 years since ELIZA, and chatbots have become increasingly impressive and increasingly inexpensive to implement. As the result of advances in NLP techniques and people's willingness to anthropomorphize and converse with computers, chatbots are now ubiquitous in our society: on our phones, in our homes, on social media, on retail websites, in online tech support forums, etc. Given the widespread use of chatbots in other segments of daily life, the time seemed right to welcome a knowledge bot into our law firm. I accordingly began work on a Knowledge And Resource Locator for Ogletree Systems – KARLOS for short. "Teaching" the Bot to Answer FAQs I started work on KARLOS in January 2018 by creating a knowledge base in Microso QnA Maker. A knowledge base is essentially a table of questions and answers that the knowledge bot references when responding to user inputs. I populated our knowledge base with our KM Counsel team's most frequently asked questions, paired with the answers to those questions. Where appropriate I assigned multiple variations of a question to the same answer in the knowledge base, so, for example, KARLOS would provide the same response to the questions "Who are the firm's tax experts?", "Who are the firm's Internal Revenue Code experts?", etc. Answers stored in the knowledge base can include hyperlinks to other resources, email links, bullet-point lists, numbered lists, and other types of formaing, which allows for a richer and more helpful experience for users asking complex questions. Finding questions and answers to teach our bot was easy: we had already compiled many standard answers in our Quick Parts and in our email folders. We also had tables of helpful information available on the firm intranet that could easily be imported into the knowledge base: a Who Does What? list, a table of practice groups and their distribution lists, a table of administrative departments and their contact information, etc. Having begun work on our knowledge base, I then began implementing a chatbot to look up answers in that knowledge base in late February 2018. With about 30 hours of work and research, spread out over about four weeks into late March, I had a fully functional prototype bot that could answer a few hundred different questions and help with a few routine search tasks. When asked a question, KARLOS uses NLP to compare the user's question to the question-and-answer pairs in its knowledge base, trying to find a potential match. If KARLOS finds a potential match with a question in its knowledge base, and if the confidence score for that match exceeds the confidence threshold, KARLOS provides the corresponding answer to the user. If KARLOS is not sufficiently confident that a user's question is in its knowledge base, KARLOS will respond "Sorry, I can't help you with that," and direct the user to the KM department for further assistance. KARLOS also provides a link to suggest new questions and answers to include in his knowledge base. Although we started by teaching KARLOS the answers to the questions KM received, we knew KARLOS could also help other Finding questions and answers to teach our bot was easy: we had already compiled many standard answers in our Quick Parts and in our email folders.

• Cover