Bridge Relation Extraction: New Chinese Dataset and Model

As a transportation infrastructure, bridge plays an important role in the region connectivity and urban prosperity, and brings tremendous convenience to people's daily life and travel. The operation and management systems of bridges are more convenient owing to the popularization of artificial intelligence. Specifically, knowledge bases in bridge field not only contribute to the intelligent management and maintenance of bridges but are conducive to enhancing the efficiency of operation decisions. One of the predominant methods to construct knowledge bases automatically is relation extraction which aims to extract structured information from unstructured texts. There are, however, two major challenges in bridge relation extraction study: 1) no bridge-specific dataset for this task; 2) no specific relation extraction models for bridges. In this paper, we introduce a manually-crafted Chinese dataset called Bridge-RE for bridge relation extraction and propose an excellent model called BERT-Bridge that can extract the relations effectively from the sentences in bridge field. Specifically, we carefully select 2500 sentences from the PDF format articles which are strongly related to bridges. The entities and relations in sentences are annotated manually and the subjects of Bridge-RE are components, defects and maintenance methods about bridges. Furthermore, we intensively investigate BERT-based models and make a considerable improvement on bridge relation extraction task by appending entity information to the model's input. Experimental results have demonstrated that our model obtains better performance than baseline model on Bridge-RE dataset.