基於BERT任務模型之低誤報率中文別字偵測模型

Low False Alarm Rate Chinese Misspelling Detection Model Based on BERT Task Model

中文別字偵測技術可以應用在教育及出版等許多實務領域。雖然近期許多研究提出了一些能提高效能的模型，但這些模型卻有著誤報率偏高的缺點。在真實的應用中，減少誤報情況的發生是很重要的，因為使用者在操作時，一直出現誤報的情況，會讓使用者體驗不佳，所以如何生成低誤報率且高效率的模型，成為一個要處理的問題。本文採用BERT在Single Sentence Tagging任務模型來解決中文別字偵測的問題，並配合此模型設計了訓練資料的大量生成方法。實驗顯示本文所提方法對SIGHAN 2015測試資料集的誤報率（False Alarm Rate）達到0.0297。與先前其他低誤報率方法相比，此方法有最低的誤報率以及最高的召回率。

Chinese misspelling detection technology can be applied in fields such as education and publishing. This research topic has garnered considerable attention. Recently, although many studies have proposed models that are based on deep learning and that are capable of improving detection accuracy, these models have the disadvantage of high false alarm rates. In real application scenarios, it is important to reduce the occurrence of false alarms because false alarms, while using the system, lead to poor user experience. Therefore, it is important to create a model with low false alarm rate and high efficiency. In this paper, BERT Single Sentence Tagging task model is used to solve the Chinese misspelling detection problem. To work with this model, mass training data generation methods were designed. Experiments showed that the method employed in this study has a false alarm rate of 0.0297 for the SIGHAN 2015 test set. Compared to other previous methods with low false alarm rates, this method has the lowest false alarm rate and the highest recall rate.