MTLAT - A Multi-Task Learning Framework Based on Adversarial Training for Chinese Cybersecurity NER.

With the continuous development of cybersecurity texts, the importance of Chinese cybersecurity named entity recognition (NER) is increasing. However, Chinese cybersecurity texts contain not only a large number of professional security domain entities but also many English person and organization entities, as well as a large number of Chinese-English mixed entities. Chinese Cybersecurity NER is a domain-specific task, current models rarely focus on the cybersecurity domain and cannot extract these entities well. To tackle these issues, we propose a M ulti- T ask L earning framework based on A dversarial T raining (MTLAT) to improve the performance of Chinese cybersecurity NER. Extensive experimental results show that our model, which does not use any external resources except static word embedding, outperforms state-of-the-art systems on the Chinese cybersecurity dataset. Moreover, our model outperforms the BiLSTM-CRF method on Weibo, Resume, and MSRA Chinese general NER datasets by 4.1%, 1.04%, 1.79% F1 scores, which proves the universality of our model in different domains.