Image2Type: Entity Type Prediction via Image Analysis

Entity type prediction is the task to assign an entity in a Knowledge Graph (KG) its semantic type. However, the type information in most open KGs such as DBpedia or Wikidata are incomplete and noisy. Thereby the type of an entity is a fundamental information.
Currently, most state-of-the-art entity type prediction models use only structured data to perform entity type prediction. This reduces prediction accuracy, creates incomplete and noisy type information. At the same time 51.1% of entities in the data set
DBpedia630k have at least one associated image. Hence it is obvious to extract the additional information incorporated in these images to improve accuracy of entity type prediction to complete type information. This thesis presents an approach which creates
a combined entity representation consisting of image features and structural information.
On top of it a fully connected neural network is deployed for classification. The model performance is measured on two newly created real-world benchmark datasets. The results suggest that the approach is suitable to perform entity type prediction, however optimization is needed to improve performance.