Michelle Cheatham, Ph.D. (Advisor); Michael Raymer, Ph.D. (Committee Member); Matthew Molineaux, Ph.D. (Committee Member)
Master of Science (MS)
Conversations are more than just a sequence of text, it is where two or more participants interact in order to achieve their goals. Conversation Understanding (CU) requires all participants to understand each others intent. In the past decade, CU has been extended from automated human-human text processing to build automated conversational agents for human-machine interactions. Despite their popularity, these automated conversational agents (like Siri, Alexa, etc) can't handle more than one or two utterances, and they don't recognize conversations as intents. The development of approaches that extract intents behind an utterance is essential for the advancements of Question Answering (QA) systems, personal robot assistants, etc. Intent prediction is seen as text classification problem and most classification approaches are based on string similarity measures, which fail to incorporate context of previous utterances. In this thesis, we explore the utility of incorporating context using the sequential nature of intents in the conversations by applying Sequential Pattern Mining (SPM) algorithm. For this purpose, we have proposed a novel Frequent Ordered Pattern (FOP) based text classification approach. We used MSdialog-IntentPred dataset to compare the performance our approach with off-the-shelf basic implementations of string similarity based text classifiers. Note that we are not trying to achieve the state-of-the-art performance but rather to test the efficacy of our approach for intent classification. Based on the evaluations, our FOP based text classifiers were able to approach the performance of string similarity based text classifiers. In fact adding the contextual information from our FOP classifiers have improved the performance of string similarity based text classifiers. Finally, we discussed how both string similarity and FOP based text classification fail to incorporate the following features: multiparty behavior, and characteristics of goals/intents (goal decomposition, and goal causation). We also presented Hierarchical Intent Network (HIN) which is our attempt at incorporating the above mentioned features of a conversation. We believe that our HIN can help improve the performance of intent classification in MSdialog-IntentPred dataset.
Department or Program
Department of Computer Science and Engineering
Year Degree Awarded
Copyright 2021, some rights reserved. My ETD may be copied and distributed only for non-commercial purposes and may not be modified. All use must give me credit as the original author.
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License.