TY - CHAP
T1 - CREAM: Named Entity Recognition with Concise query and REgion-Aware Minimization
AU - Yao, Xun
AU - Yang, Qihang
AU - Hu, Xinrong
AU - Yang, Jie
AU - Guo, Yi
PY - 2023
Y1 - 2023
N2 - Recent advancements in Machine Reading Comprehension (MRC) models have sparked interest in the field of Named Entity Recognition (NER), where entities are extracted as answers of given queries. Yet, existing MRC-based models face several challenges, including high computational costs, limited consideration of entity content information, and the tendency to generate sharp boundaries, that hinder their generalizability. To alleviate these issues, this paper introduces CREAM, an enhanced model leveraging Concise query and REgion-Aware Minimization. First, we propose a simple yet effective strategy of generating concise queries based primarily on entity categories. Second, we propose to go beyond existing methods by identifying entire entities, instead of just their boundaries (start and end positions), with an efficient continuous cross-entropy loss. An in-depth analysis is further provided to reveal their benefit. The proposed method is evaluated on six well-known NER benchmarks. Experimental results demonstrate its remarkable effectiveness by surpassing the current state-of-the-art models, with the substantial averaged improvement of 2.74, 1.12, and 2.38 absolute percentage points in Precision, Recall, and F1 metrics, respectively.
AB - Recent advancements in Machine Reading Comprehension (MRC) models have sparked interest in the field of Named Entity Recognition (NER), where entities are extracted as answers of given queries. Yet, existing MRC-based models face several challenges, including high computational costs, limited consideration of entity content information, and the tendency to generate sharp boundaries, that hinder their generalizability. To alleviate these issues, this paper introduces CREAM, an enhanced model leveraging Concise query and REgion-Aware Minimization. First, we propose a simple yet effective strategy of generating concise queries based primarily on entity categories. Second, we propose to go beyond existing methods by identifying entire entities, instead of just their boundaries (start and end positions), with an efficient continuous cross-entropy loss. An in-depth analysis is further provided to reveal their benefit. The proposed method is evaluated on six well-known NER benchmarks. Experimental results demonstrate its remarkable effectiveness by surpassing the current state-of-the-art models, with the substantial averaged improvement of 2.74, 1.12, and 2.38 absolute percentage points in Precision, Recall, and F1 metrics, respectively.
KW - Machine Reading Comprehension
KW - Named Entity Recognition
KW - Query Optimization
KW - Region-Aware Loss
UR - http://www.scopus.com/inward/record.url?scp=85175965507&partnerID=8YFLogxK
UR - https://ezproxy.uws.edu.au/login?url=https://doi.org/10.1007/978-981-99-7254-8_59
U2 - 10.1007/978-981-99-7254-8_59
DO - 10.1007/978-981-99-7254-8_59
M3 - Chapter
AN - SCOPUS:85175965507
SN - 9789819972531
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 763
EP - 777
BT - Web Information Systems Engineering – WISE 2023: 24th International Conference, Proceedings
A2 - Zhang, Feng
A2 - Wang, Hua
A2 - Barhamgi, Mahmoud
A2 - Chen, Lu
A2 - Zhou, Rui
PB - Springer Nature Singapore
CY - Singapore
ER -
