Leveraging Linguistic Structures for Named Entity Recognition with Bidirectional Recursive Neural Networks

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2664–2669 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Leveraging Linguistic Structures for Named Entity Recognition with Bidirectional Recursive Neural Networks Peng-Hsuan Li National Taiwan University No. 1, Sec. 4, Roosevelt Rd. Taipei 10617, Taiwan jacobvsdanniel@gmail.com Ruo-Ping Dong National Tsing Hua University No. 101, Sec. 2, Kuang-Fu Rd. Hsinchu 30013, Taiwan dongruoping@gmail.com Yu-Siang Wang National Taiwan University No. 1, Sec. 4, Roosevelt Rd. Taipei 10617, Taiwan b03202047@ntu.edu.tw Ju-Chieh Chou National Taiwan University No. 1, Sec. 4, Roosevelt Rd. Taipei 10617, Taiwan jjery2243542@gmail.com Wei-Yun Ma Academia Sinica No. 128, Sec. 2, Academia Rd. Taipei 11529, Taiwan ma@iis.sinica.edu.tw Abstract In this paper, we utilize the linguistic structures of texts to improve named en- tity recognition by BRNN-CNN, a spe- cial bidirectional recursive network at- tached with a convolutional network. Mo- tivated by the observation that named en- tities are highly related to linguistic con- stituents, we propose a constituent-based BRNN-CNN for named entity recogni- tion. In contrast to classical sequential la- beling methods, the system first identifies which text chunks are possible named en- tities by whether they are linguistic con- stituents. Then it classifies these chunks with a constituency tree structure by recur- sively propagating syntactic and semantic information to each constituent node. This method surpasses current state-of-the-art on OntoNotes 5.0 with automatically gen- erated parses. 1 Introduction Named Entity Recognition (NER) can be seen as a combined task of locating named entity chunks of texts and classifying which named entity cat- egory a chunk falls into. Traditional approaches label each token in texts as a part of a named en- tity chunk, e.g. “person begin”, and achieve high performances in several benchmark datasets (Rati- nov and Roth, 2009; Passos et al., 2014; Chiu and Nichols, 2016). Being formulated as a sequential labeling prob- lem, NER systems could be naturally imple- mented by recurrent neural networks. These net- works process a token at a time, taking, for each token, the hidden features of its previous token as well as its raw features to compute its own hidden features. Then they classify each token by these hidden features. With both forward and backward directions, networks learn how to prop- agate the information of a token sequence to each token. Chiu and Nichols (2016) utilize a varia- tion of recurrent networks, bidirectional LSTM, attached with a CNN, which learns character-level features instead of handcrafting. They accom- plish state-of-the-art results on both CoNLL-2003 (Tjong Kim Sang and De Meulder, 2003) and OntoNotes 5.0 (Hovy et al., 2006; Pradhan et al., 2013) datasets. Classical sequential labeling approaches take little information about phrase structures of sen- tences. However, according to our analysis, most named entity chunks are actually linguistic con- stituents, e.g. noun phrases. This motivates us to focus on a constituent-based approach for NER where the NER problem is transformed into a named entity classification task on every node of a 2664 Algorithm 1 Binarization 1: function BINARIZE(node) 2: n …

Content truncated to 500 words