Coping with the document frequency bias in sentiment classification

Abstract

In this article, we study the polarity detection problem using linear supervised classifiers. We show the interest of penalizing the document frequencies in the regularization process to increase the accuracy. We propose a systematic comparison of different loss and regularization functions on this particular task using the Amazon dataset. Then, we evaluate our models according to three criteria: accuracy, sparsity and subjectivity. The subjectivity is measured by projecting our dictionary and optimized weight vector on the SentiWordNet lexicon. This original approach highlights a bias in the selection of the relevant terms during the regularization procedure: frequent terms are overweighted compared to their intrinsic subjectivities. We show that this bias appears whatever the chosen loss or regularization and on all datasets: it is closely link to the gradient descent technique. Penalizing the document frequency during the learning step enables us to improve significantly our performances. A lot of sentimental markers appear rarely and thus, are unappreciated by statistical learning algorithms. Explicitly boosting their influences leads to increasing the accuracy in the sentiment classification task. Copyright © 2012, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.