Feature enrichment through multi-gram models

Abstract

We introduce a feature enrichment approach, by developing multi-gram cosine similarity classification models. Our approach combines cosine similarity features of different N-gram word models, and unsupervised sentiment features, into models with a richer feature set than any of the approaches alone can provide. We test the classification models using different machine learning algorithms on categories of hateful and violent web content, and show that our multi-gram models give across-the-board performance improvements, for all categories tested, compared to combinations of baseline unigram, N-gram, and sentiment classification models. Our multi-gram models perform significantly better on highly imbalanced sets than the comparison methods, while this enrichment approach leaves room for further improvements, by adding instead of exhausting optimization options.