Classification of Connectives Based on Combination Neural Network

The overlap rate of connectives between different categories of connectives can be viewed directly from the heat map. Therefore, it is necessary to explore a classification method to minimize ambiguity of connectives. A series of methods for automatic classification of connectives in sentence-level are compared in this paper. A combination method is chosen by which we can predict the categories of connectives in a sentence. In the process of automatic classification of connectives, a supervised machine learning algorithm---K-Nearest Neighbor (KNN) was used to classify connectives in order to reduce the ambiguous connectives. However, this method has not gained good effect of automatic classification of connectives. Therefore, authors took another three methods, including Forward Feed Neural Network (FFNN) algorithm, Backward Propagation Neural Network (BPNN) algorithm and a combination method of FFNN&BPNN algorithm. All of these three methods can get better accuracy than the result of KNN algorithm, while the highest accuracy gained by the combination method of FFNN&BPNN among these methods of the experiments. These works can find a maximally disambiguated approach for the classification of connectives that was abstract from Penn Discourse Treebank 3.0 (PDTB3.0).