Entity Extraction Research with Relative Position Encoding and Masked Conditional Random Fields : Named Entity Recognition

Traditional named entity recognition methods lack relative positional information interaction and pay little attention to the distance information between entities, which makes it difficult to determine the entity boundaries when recognizing entities and affects recognition performance. To address these issues, we propose an entity recognition method that combines relative positional encoding with masked conditional random fields. First, we introduce rotation positional encoding at the encoding layer to enhance the span information between characters, and use a bidirectional long short-term memory (BiLSTM) to capture contextual representations and a convolutional neural network (CNN) to capture local key information in the text to construct a multi-dimensional feature-rich text representation. Second, in the decoding stage, MaskedCRF is used to legally constrain the label results to make them more in line with the correct label distribution. Finally, a lightweight data augmentation method for fragment entity deletion and replacement is proposed to expand the generalizability of the text enhancement model. The method is based on RoBERTa-BiLSTM-MaskedCRF. The experimental results on the Weibo and Resume datasets show that the F1 reached 72.85% and 96.30% respectively, which is an improvement of 3.80% and 0.62% compared to the baseline model. It has better recognition performance, is simple and convenient to implement, and has good portability. It can provide a way of thinking for subsequent entity recognition research.