Parsing with graph convolutional networks and clustering

Abstract

When designing machine translation systems, an important task is to represent data using graphs, where words act as vertices, and relations between words in a sentence act as edges. One of these tasks at the first stage of the analysis is the classification of words as parts of speech, and at the next stage of the analysis is to determine the belonging of words to the sentence members’ classes. A robust approach to carry out such a classification is to determine words embeddings by using Graph Convolutional Networks at the beginning of the analysis and then to apply k-means clustering which is an algorithm that splits objects (words) into groups. To determine weights an ordinary network is applied to obtained hidden layers in order to use these weights in subsequent analysis.