Bayesian Deep Convolution Belief Networks for Subjectivity Detection

Iti Chaturvedi, Erik Cambria, Soujanya Poria, Rajiv Bajpai

Abstract: Subjectivity detection aims to distinguish natural language as either opinionated (positive or negative) or neutral. In word vector based convolutional neural network models, a word meaning is simply a signal that helps to classify larger entities such as a document. Previous works do not usually consider prior distribution when using sliding windows to learn word embedding's and, hence, they are unable to capture higher-order and long-range features in text. In this paper, we employ dynamic Gaussian Bayesian networks to learn significant network motifs of words and concepts. These motifs are used to pre-train the convolutional neural network and capture the dynamics of discourse across several sentences.