Comparisons of word representations for convolutional neural network: An exploratory study on tourism Weibo classification
Rui-Hong Sun, Jin‐Xing Hao · 2017
Word representations, which are critical to the performance of convolutional neural network, has attracted considerable attention from many researchers. Two popular categories of word representations for convolutional neural network are the pre-trained representation, which requires training on external web documents, and the internal presentation, which relies on the internal text features. Although prior studies always claim that their models are better than previous ones based on various evaluation criteria, very little research compares the two categories of word representations. In this study, we reported our initial attempt to compare the two categories of word representation models for conventional neural network in the context of tourism Weibo classification. We have designed two experiments to examine their differences on classification performance. Results show that the performance of the pre-trained representation depends on the domain and the size of the training corpus, while the performance of the internal representation without the training process, can achieve equivalent performance for the domain of tourism Weibo classification. We also discussed the theoretical and practical implications for this study.