DocBAN: An Efficient Biaffine Attention Network for Document-Level Named Entity Recognition

Named entity recognition (NER) aims to identify prefined types of entities from the given text sequence. When the sequence is composed of multiple sentences, this problem is referred as document-level NER. Recently, most span-based methods exploit the biaffine attention to get a n × n score matrix, where n is the length of sequence and each entry denotes a span representation. However, when dealing with long sequences, such as doc-level NER, those models exhibit low efficiency due to enumerating all spans via Biaffine Attention Network (BAN), which scales quadratically with the sequence length. To address this limitation, we propose an efficient BAN for doc-level NER, called DocBAN, which scales linearly with sequence length and can be regarded as a parallel alternative to the standard biaffine attention. Specifically, we reduce its time and space complexity by introducing a sliding window mechanism, and use the U-Net to capture global features of multiple windows. Extensive experiments demonstrate that DocBAN we proposed, serving as an alternative for BAN, can significantly improve the efficiency of existing span-based methods while maintaining competitive performance.