Deep Multi-cultural Graph Representation Learning

摘要

This research aims at the development of a knowledge representation that will elucidate and visualize the differences and similarities between concepts expressed in different languages and cultures. Wikipedia graph structure is considered around one concept namely "Nazism" in two languages, English and German for the purpose of understanding how online knowledge crowdsourcing platforms will be affected by different language groups and their cultures. The solution is divided into capturing structure of weighted graph representation learning via random surfing, cross-lingual document similarity via Jaccard similarity, multi-view representation learning by deploying Deep Canonical Correlation Autoencoder (DCCAE) and sentiment classification task via SVM. Our method shows superior performance on word similarity task. Based on our best knowledge, it is the first application of DCCAE in this context.