A Controlled Attention for Nested Named Entity Recognition

Traditional methods to recognize named entities are conducted as sequence labelling or span classification. They are usually implemented on a raw input without any cue about possible named entities. This method cannot be aware of entity boundaries and learn semantic dependencies between them. Cognitive neuroscience has revealed that foveating stimuli improves the efficiency of processing in terms of acuity. Inspired by this phenomenon, we propose a controlled attention mechanism for recognizing named entities. In our method, instead of feeding a raw input into a neural network, task-related cues are implanted into each sentence to indicate boundaries of possible named entities. Then, the modified sentence is sent into a deep network to learn a discriminative entity-relevant sentence representation. In our experiments, the controlled attention is evaluated on English and Chinese corpora. Comparing with existing models, it shows significant improvement for nested named entity recognition. We achieve the state-of-the-art performance in all evaluation datasets. The controlled attention has three advantages for named entity recognition. First, it enables a neural network to become aware of entity boundaries and construct semantic dependencies relevant to possible entities. Second, implanting entity cues enables a neural network to concentrate on the task-related semantic features while disregarding nonessential information in a sentence. Third, the controlled attention also has the potentiality to be extended for other NLP tasks, e.g., entity relation extraction and event extraction.