This item is available under a Creative Commons License for non-commercial use only
Over the last ﬁfty years, conversational agent systems have evolved in their ability to understand natural language input. In recent years Natural Language Processing (NLP) and Machine Learning (ML) have allowed computer systems to make great strides in the area of natural language understanding. However, little research has been carried out in these areas within the context of conversational systems. This paper identiﬁes Convolutional Neural Network (CNN) and Support Vector Machine (SVM) as the two ML algorithms with the best record of performance in ex isting NLP literature, with CNN indicated as generating the better results of the two. A comprehensive experiment is deﬁned where the results of SVM models utilising sev eral kernels are compared to the results of a selection of CNN models. To contextualise the experiment to conversational agents a dataset based on conversational interactions is used. A state of the art NLP pipeline is also created to work with both algorithms in the context of the agent dataset. By conducting a detailed statistical analysis of the results, this paper proposes to provide an extensive indicator as to which algo rithm oﬀers better performance for agent-based systems. Ultimately the experimental results indicate that CNN models do not necessarily generate better results than SVM models. In fact, the SVM model utilising a Radial Basis Function kernel generates statistically better results than all other models considered under these experimental conditions.
O Sullivan, Kieran (2018). Comparig the effectiveness of support vector machines and convolutional neural networks for determing user intent in conversational agents. Masters dissertations, DIT, 2018.