Sentiment Analysis on Banking Chatbot using Graph-based Machine Learning Model

Many banking companies use a chatbot that can answer questions about their products. Knowing the sentiments of users can help companies evaluate the chatbots. However, the chatbot is made using a decision dialogue method so it has no knowledge of the sentiment of a sentence. Traditional data mining methods such as converting words to vectors are no longer efficient because of the limited representation of sentences. Therefore, we need a method that can connect words between words so that the model can understand it better, namely using graph-based representation. We represent a graph with words as nodes along with their node properties in the form of vectors and edges using a dependency parser to connect these nodes. With graphs we can represent text better such as being able to include the part of speech of the word, the number of letters, and so on. Data that has been converted to graph will be processed with a graph-based model. In this study, we compare traditional methods with graph-based methods in terms of the performance and inference time in detecting the sentiment of a sentence. The results of the experiment demonstrate that the graph-based model achieves a comparable level of accuracy (0.7173) to that of the tree-based model. However, when it comes to inference time, the graph-based models outperform the tree-based models, being three times faster.