Chinese Fine-Grained Entity Typing Based on Box Embedding

In the task of fine-grained entity typing, there is a need for better modeling of the relationships between entity types due to the large number of entity types and the complexity of these relationships. This paper investigates box embedding methods on a Chinese dataset, where both types and entities are embedded as hyper-rectangle boxes. Reprensentations of an entity mention and its context are generated by the entity typing model and then embedded into the box space. Through training, representations of types are obtained in the box space. During prediction, the range of the boxes is utilized to estimate the posterior probability of each entity belonging to a specific type. The box-based method achieved further performance improvement on the CFET dataset, indicating its stronger expressive power compared to vector-based type modeling. In addition, this paper visualizes the intervals and calculates the average length of types and entities in each embedding dimension. The relationships between types and between entities and types are shown and analyzed, demonstrating the rationality of box embedding in entity type modeling.