CeER: A Nested Name Entity Recognition Model Incorporating Gaze Feature

Nested name entity recognition (NER) is a fundamental information extraction task. Although many model architectures have been proposed to solve this task, human annotators still have strength in recognition of complex structures and professional fields. In this work, we propose a Cognition-enhancing Entity Recognition model (CeER), which introduces cognition-based data to improve the performance of nested name entity recognition. Specifically, we extract the gaze feature from eye-tracking data and build a cognition-enhancing encoder to represent human cognitive information. First, we adopt a set of binning rules to convert useful information extracted from eye-tracking data into gaze features of words. Second, we construct the Gaze Feature Learning module to encode words as high-dimensional embeddings and then decode them into gaze feature vectors thereby capturing gaze features of words which reflect their importance in the reading cognitive process. Finally, we utilize the encoder improved by gaze feature learning and follow the question-answering architecture to identify all possible nested entities. We select three public eye-tracking datasets and two nested NER datasets, GENIA and SciERC, to perform evaluation experiments. Experimental results demonstrate that incorporating gaze feature improves the nested name entity recognition task.