This document discusses evaluating the effectiveness of incorporating case frame information into neural Japanese named entity recognition. It proposes adding modulation gate to integrate case frame vectors (MVC) into a baseline Bi-LSTM-CRF model. The experiment shows adding MVC to the output of LSTM (posterior insertion) achieved the best performance compared to other insertion methods or no MVC. While MVC helped most entity types, location names saw less improvement likely due to lower MVC attribution rates. Overall, incorporating syntactic case frame data enhanced the baseline model's ability to recognize named entities.
5. ベースラインモデル:
Bi-LSTM-CRF(Misawa et al. 2017)
l NERで多く採用されている
基本的なニューラルNERモデル
l 文脈から予測を行うことができる
l 文字と単語を同時に考慮
5
Shotaro Misawa, Motoki Taniguchi, Yasuhide Miura,
and Tomoko Ohkuma. Character-based bidirectional lstm-crf
with words and characters for japanese named entity
recognition. In Proceedings of the First Workshop on Subword
and Character Level Models in NLP, pp. 97‒102, 2017.
9. Modulation Gate(Lu et al.2018)
l テキスト情報と外部情報を組み合わせる
l 入力:j番目のテキスト情報 𝒉7
対応する外部情報 𝒗7
l 出力:9
𝒉7
9
Di Lu, Leonardo Neves, Vitor Carvalho, Ning Zhang,
and Heng Ji. Visual attention model for name tagging in
multimodal social media. In Proceedings of the 56th
Annual
Meeting of the Association for Computational Linguistics
(Volume1:LongPapers), pp. 1990‒1999, 2018.
10. データセット
l 拡張固有表現タグ付きコーパス(橋本ら 2010)
l 新聞、白書、Webテキストなど様々なジャンル
l 文書数:およそ9,000
l 固有表現クラス
l 上記コーパスで頻出の6種類を使用
l 固有表現数:36万
10
Train Dev Test All
文書数 7,383 901 900 9,184
PRODUCT 76,956 10,130 9,631 96,717
NUMBER 60,469 6,633 7,876 74,978
LOCATION 45,601 5,704 5,207 56,512
TIME 43,001 5,354 5,477 53,832
ORGNIZATION 34,774 4,115 4,028 42,917
PERSON 32,844 4,297 3,989 41,130
ALL 293,645 36,233 36,208 366,086