Information entropy-informed sentence representation for question classification

The traditional question classification methods generally employ a large number of features extracted from labelled questions to achieve good classification result However, the high dimensionality of the feature space may lead to a higher training cost In this paper, we propose a novel and effective question semantic representation-based method for question classification, avoiding the complicated feature extraction process. We first exploit the neural network-based language model to learn the distributed representation of words which can capture the semantic relations between words. We then introduce the information entropy to measure the importance of the word to question classification, i.e. use the information entropy to adjust the weight of the word. Subsequently, the question vectors are fed into an SVM classifier as semantic features to obtain the classification result Experimental results demonstrate the effectiveness of our method with the improvement of 3.62% on open domain dataset and 6.50% on agricultural dataset over baseline.