A Novel Technique for Extracting Entity Relations

Entity relation extraction is a task that aims to extract the semantic relationship of entity pairs from a given text. Due to the importance of entity semantics to relation extraction, many methods pay too much attention to the semantic characteristics of entities and ignore other useful information in the text. In recent years, pre-trained language models, such as Bert and XLNET, have achieved great success in multiple tasks in the field of natural language processing by fine-tuning with downstream task models. However, the original standard task of the pre-trained language model does not include relation extraction. In this paper, we extract the shortest dependency path (SDP) of the input text, splice the original text and the SDP as the input of the pre-trained language model, splice the corresponding features of the entities and the SDP in the output semantic features of pre-trained language model as the input of the downstream relation extraction network, and use a new loss function based on cross entropy(CE) in the training process, which optimizes the model. On this basis, a new entity relation extraction method is proposed. The performance comparison experiment in several 2010 Task 8 data sets shows that the proposed method is better than most methods.