Details of Research Outputs

Typing entities in structured sources such as Wikipedia has been well studied to construct English knowledge bases automatically. However, there still remain two tough challenges in typing entities in Chinese-pedia. The first one is that structured information from Chinese-pedia cannot assign entities fine-grained types due to its inaccuracy and coarseness. The other challenge is the incompletion of Chinese-pedia, which means we can only use limited attribute fields to type entities. In this paper, we propose a novel Hierarchical Neural System (HNS) to infer fine-grained types for entities in Chinese-pedia. The HNS contains three main models which are hierarchical attention model, feature fusion model and hierarchical classification model. The hierarchical attention model extracts features from entity description based on a bi-LSTM network with hierarchical attention mechanism to break the limitation of inaccurate Chinese-pedia. To deal with the incompletion of Chinese-pedia, the feature fusion model is presented to obtain type features from multi-source such as descriptions, info-boxes, and categories. Through this model, we fuse all the features from different sources together and reduce the features to low-dimensional and dense vectors. Finally, the hierarchical classification model is designed to infer fine-grained types for entities in Chinese-pedia with features obtained from the other two models. The experiments illustrate that HNS outperforms the start-of-art work by 15.6% on f1-score. © Springer International Publishing AG, part of Springer Nature 2018.