Rethinking Boundaries: End-To-End Recognition of Discontinuous Mentions with Pointer Networks

A majority of research interests in irregular (e.g., nested or discontinuous) named entity recognition (NER) have been paid on entities, while discontinuous entities received limited attention. Existing work for NER, however, either suffers from decoding ambiguity predicting using token-level local features. In this work, we present an innovative model NER based pointer networks, where the simultaneously decides whether a token at each frame constitutes mention and next constituent is. Our has three major merits compared with previous work: (1) The mechanism is memory-augmented, which enhances boundary detection interactions between current decision prior recognized mentions. (2) encoder-decoder architecture can linearize complexity structure prediction, thus reduce search costs. (3) makes every global information, i.e., by consulting all input, encoder decoder output view. Experimental results CADEC ShARe13 datasets show that our outperforms flat hypergraph models as well state-of-the-art transition-based NER. Further in-depth analysis demonstrates performs recognizing various including flat, overlapping ones. More crucially, effective detection, kernel source to