Latent Semantic Analysis Boosted Convolutional Neural Networks for Document Classification

摘要

Convolutional neural networks (CNNs) have been shown to be effective in document classification tasks. CNNs can be setup using various architectures with many different parameter settings, which may make them difficult to implement. For many document classification tasks, data transformed with ngrams (typically using uni, bi, and trigrams) and term-frequency inverse-document-frequency (TFIDF) weighting are still considered effective baseline models when used with linear classifiers such as logistic regression, especially in smaller datasets with less than 500K observations. A parsimonious CNN baseline model for sentiment classification should replicate the easy use of linear methods. In this study, we introduce a Latent Semantic Analysis (LSA) based CNN model, in which natively trained LSA word vectors are used as input into parallel 1-dimensional convolutional layers (1D-CNNs). The LSA word vector model is obtained by applying singular value decomposition (SVD) on the data transformed by a unigram and TFIDF weighting. Thus, the convolutional layers are designed with window sizes that are best suited for LSA word vectors. This parsimonious LSA -based CNN model exceeds the accuracy of all linear classifiers utilizing ngrams with TFIDF on all analyzed datasets, with average improvement of 0.73% by the top performing LSA-based CNN models. This may be due to the fact that CNNs are better adept at capturing word relationships in phrases and sentences that are not necessarily in the training corpus. Furthermore, the LSA-based CNN model exceeds the performance of word2vec-based CNN models as well. Thus, the success of LSA-based CNNs may potentiate their use as a baseline in classification tasks alongside linear models. Also, we provide guiding principles to simplify the application of LSA-based CNNs in document classification tasks.