基于BERT在税务公文系统中实现纠错功能

Implementation of Error Correction Function in Tax Official Document System Based on BERT

税务公文作为社会政治的产物,具有鲜明的政治性。而撰制公文是一项严肃的工作,必须保持准确、严肃的文体特点。为减轻撰制者和审核者的负担,该实验针对税务系统,利用基于BERT-BiLSTM-CRF的序列标注模型和BERT掩码语言模型的特点,对公文句子中常见的单个字错误进行了检错、纠错实验。准确率、召回率和F1值相比传统的纠错方法有着明显的提升。结果表明,基于BERT-BiLSTM-CRF的序列标注模型和BERT掩码语言模型在税务公文检错纠错应用中具有较大价值。

As a product of social politics,tax official documents have a distinct political nature.Writing official documents is a serious work,and it must maintain accurate and serious style characteristics.In order to reduce the burden of writers and reviewers,this experiment is aimed at the tax system and uses the advantages of the BERT-BiLSTM-CRF-based sequence labeling and BERT mask language model to detect and correct single word errors in official document sentences.Compared with traditional error correction methods,the accuracy rate,recall rate and F1 value are significantly improved.The results show that the BERT-BiLSTM-CRF-based sequence labeling and BERT mask language model have great value in the error detection and correction of tax administrative documents.