Deep learning approaches to classify the relevance and sentiment of news articles to the economy

We consider a text classification task over an open source dataset involving news snippets and their relevance to the US economy. Text classification and sentiment analysis have been performed using nine different classifiers among which three are the traditional machine learning models, namely, support vector machine, extreme gradient boosting and logistic regression, and six neural network-based methods. The neural net frameworks include long short-term memory (LSTM), bidirectional long short-term memory (BiLSTM) and an ensemble of one dimensional convolution network (1D CNN) with LSTM/BiLSTM. Both word-to-vector and term-frequency inverse-document-frequency vectors are used in our analysis with text and sentiment classification tasks. A detailed comparative study is provided to assess the relative performance of different classification approaches. It is observed that the ensemble with 1D CNN performs better in both binary and multiclass classifications. Specifically, in the multinomial sentiment classification, 1D CNN with BiLSTM has the best performance as opposed to 1D CNN with LSTM in the binary text classification. BiLSTM architecture which incorporates the backward dependencies turns out as superior to LSTM by a margin of 30% in multiclass classification even though the considered dataset is small and inherently challenging. Further analysis to evaluate the impact of successive increases in percentage of augmented data reveals that such augmentation has a limit up to 180% in this dataset beyond which the performance starts decreasing.