The return of the chatbots: Conversing about time with Aristotle

Abstract

This Master's Thesis examines the multidisciplinary field of dialog systems in the context of the recent reappearance of chatbots, their exponential proliferation and massive adaptation by third-party companies inside the current digital ecosystem. Following a timeline from the conceptualization of the Turing Test in 1950 and leading up to the establishment of the Alexa Prize in 2016 and on, the thesis explores the current trends in the rapidly developing chatbots ecosystem. A classification of dialog systems based on their main features is presented, along with an introduction to their basic components and processing models.

The thesis addresses a core research question in the field, which is related to the current limitations of commercial chatbots and the challenges they face in processing verbal interaction and responding appropriately. To this end, it employs Aristotle, a commercial, business-analytics chatbot developed by the Silicon Valley based start-up, Bouquet.ai. This is a real case-study, through which hands-on testing of the natural language understanding capabilities of a commercial chatbot was undertaken. The initial phase of the testing pointed -among others- to the need for deeper exploration of Temporal Expression Recognition (TER) by the chatbot; this is a well-known sub-task in Natural Language Processing (NLP), which is not only challenging at a research level in the NLP field, but also highly important for a chatbot in the Business Analytics domain where time information is critical. The extensive testing comprised the creation and processing of a thorough set of potential user utterances involving highly diverse temporal expressions, as documented in the Linguistics and Computational Linguistics literature.

The findings indicate that Aristotle is capable of recognizing simple, as well as more complex temporal expressions in user utterances and returning the appropriate answer. However, the variety of temporal expressions that can be handled can be greatly extended, providing for more natural and effective verbal interaction with the user. The thesis includes concrete suggestions on extending the current back-end system of Aristotle for achieving wide-coverage of time-references in verbal interaction with its users. These suggestions have been formulated in the template structure used by Bouquet.ai, for immediate integration in the chatbot and A/B testing.