LACNNER: Lexicon-Aware Character Representation for Chinese Nested Named Entity Recognition

Named Entity Recognition (NER) is one of fundamental researches in natural language processing. Chinese nested-NER is even more challenging. Recently, studies on NER have generally focused on the extraction of flat structures by sequence annotation strategy while ignoring nested structures. In this paper, we propose a novel model, named LACNNER, that utilizing lexicon-aware character representation for Chinese nested NER. We select the typical character-level framework to overcome error propagation problem caused by incorrect word separation. Considering the situation that Chinese words always contain much richer semantic information than single characters do, it firstly obtains more significant matching words through external lexicon in our LACNNER model, and then generates lexicon-aware character representations that make full use of word-level knowledge for nested named entity. We also evaluate the effectiveness of LACNNER by taking ACE-2005-Zh dataset as a benchmark. The experimental results fully verified the positive effect of incorporating lexicon-aware character-representation in recognition of Chinese nested entity structure.