N-ARY TREE-CNN FOR ARABIC SENTIMENT ANALYSIS

Abstract

Distributed document and sentence representation is an essential step in text classification. Several models have been studied to compose sentences into a fixed length representation. Such models range from simple order-insensitive models, like Bag-of-Words, to sequence based models, like RNN. In this thesis we propose an architecture that takes into account the hierarchal nature of the language, by building on binary Recursive Neural Nets, using CNN as an internal representation building block for N-ary trees. The algorithm is applied on Arabic sentiment analysis as an example text classification task and reduces the error rate by up to 15-20% for several standard datasets.