Intent Classification Using Machine Learning Algorithms and Augmented Data

An essential step in a conversational agent is an intent classification of user-generated text input. The purpose of building the intent classifier for a chatbot is to understand the intention of user queries to respond fast and accurately. Robust chatbots necessitate more utterances for an improved training model. Nevertheless, acquiring and annotating data is time-consuming and expensive. This paper investigates machine learning techniques and data augmentation for addressing intent classification. Experiments were conducted on office product's question answering of Amazon using Random Forest, Multinomial Naïve Bayes, Logistic Regression, and Support Vector Machine (SVM). Contextual word embedding with BERT was used for generating new synonym utterances. The main experiments are the comparison of the performance of these methods after augmenting new data. In general, SVM and random forest yield comparable results. Followed by logistic regression. However, the f1 score of the multinomial naïve base is the lowest. Additionally, we discovered that augmenting new utterances had a simple effect on the performance of models.