Thai variable-length question classification for e-commerce platform using machine learning with topic modeling feature

Abstract

Nowadays, e-commerce platform continuously grows every year and becomes a part of our daily life. However, the application changes from time to time. Either new users or experienced users could face a problem. Several channels, which are FAQ, email, live chat, and call, are provided by e-commerce platform to cope with the problem. FAQ is usually ignored because it is hard to search for the desired answer. The rest channels are applicable. However, the huge number of users causes a bottleneck especially in the special events which delays the users to receive help because customer service agent can reply to the user once at a time. Therefore, this thesis proposed Thai variable-length question classification for e-commerce platform. The proposed model is based on a fusion of two architectures, Latent Dirichlet Allocation (LDA) and Bidirectional Long Short-Term Memory (Bi-LSTM), as a feature extraction process. Then, the results are concatenated and fed into a multilayer perceptron (MLP) network with a softmax as an activation function to classify an incoming question. The experimental results indicated that the proposed model outperforms the existing classification models with an accuracy of 84.43%.