A Transition-Based Neural Framework for Chinese Nested Entity and Relation Recognition

Entity and relation extraction for Chinese texts are typical performed in a pipelined fashion in the sense that by first segmenting sequence into words then recognizing entities and relations subsequently. However, this process often leads to the problem of error propagation and prevents cross-task information integration. To address this issue, we propose a novel transition-based model that performs nested entity recognition and relation extraction jointly without the need of word segmentation beforehand, which is achieved by leveraging the recent advance in using a lattice structure for Chinese sentence encodings. On standard ACE benchmarks, our model gives the best results in the literature. Further analyses show the effectiveness of the proposed architecture in capturing structural outputs.