A NER Method Based on Location-Aware Multi-Feature Fusion

Abstract

Named Entity Recognition (NER) is a fundamental and critical task in natural language processing, which mainly includes entity boundary recognition and entity classification. Previous studies of NER have made a lot of achievements, but most of them ignore the important feature of location information, and the extraction of text span feature is not enough. For the above shortcomings, we propose a location-aware multi-feature fusion entity recognition method, which enhances the sensitivity of the model to the span length and span by explicitly adding rotation position coding. Then the feature extraction and fusion of span are carried out by conditional layer normalization (CLN) fusion dilated convolution and multiplicative attention respectively, which strengthens the richness of span features. Thanks to the inclusion of the above methods, our model outperforms several SOTA baseline models on several widely used public datasets. Our model is also applicable to both flat NER and nested NER.