of Informatics, University of Tsukuba • Research Engineer @Cookpad Inc. • HQ @nlpaper.challenge (https://nlpaper-challenge.connpass.com/) • ex-Researcher @Retrieva Inc. Interests: Natural Language Processing, Named Entity Recognition @himkt or @himakotsu @himkt For more detail: https://himkt.github.io
the North American Chapter of the Association for Computational Linguistics (NAACL2016) Guillaume Lample et al. (Carnegie Mellon University) (現在: Facebook AI Research) 書誌情報 3 著者実装: https://github.com/glample/tagger
or IOB) (B: Begin, I: Inside, O: Outside, E: End, S: Single) George B-PER Washington E-PER Visited O England S-LOC PERSON: George Washington, LOCATION: England 6
NEが連続する場合,2番目以降のNEの開始単語に 接頭辞 B を付与(NERの場合にはほとんどBのprefixは出てこない) •BIO (or IOB2) • NEが連続するかどうかにかかわらずNEの開始単語に接頭辞 B を付与 •BIOES: BIOの拡張(今回の実験で採用) • BIO の拡張: NE の終端単語に接頭辞 E を付与 • ただし, NE が単一の単語である場合,その単語には接頭辞 S を付与 7 I-PER E-PER S-PER O S-LOC O B-PER I-PER I-PER B-PER O I-LOC O B-PER I-PER I-PER B-PER O I-LOC O I-PER IOB BIO BIOES 系列タグスキーマ
NAACL2016. [2] End-to-end Sequence Labeling via Bi-directional LSTM-CNNs-CRF. Ma et al. ACL2016. [3] Semi-supervised sequence tagging with bidirectional language models. Peters et al. ACL2017. [4] Deep contextualized word representations. Peters et al. NAACL2017. [5] Empower Sequence Labeling with Task-Aware Neural Language Model. Liu et al. AAAI2018 [6] Not all contexts are created equal: Better word representations with variable attention. Ling et al. EMNLP2016 [7] Understanding the difficulty of training deep feedforward neural networks. Glorot et al. AISTATS2010. [8] Efficient Backprop. LeCun et al. Neural Network: Tricks of the Trade. [9] Adaptive Gradient Methods with Dynamic Bound of Learning Rate. Luo et al. ICLR2019. 40