Named Entity Recognition Method for Chinese Public Opinion Based on Multi-Task Learning
Yunkai Zhang, Bo Hao Cheng · 2023
Using deep learning to extract public opinion data is a common and effective approach. Nevertheless, many existing named entity recognition models in specific domains [1][2] often depend on hand-engineering, such as constructing external knowledge bases, to improve model performance. Applying these models to different domain data requires the reconstruction of domain-specific knowledge bases, hindering the lateral transferability of the model across various domain tasks. In this paper, we design two subtasks under the framework of multi-task learning to enhance named entity recognition (NER), ultimately proposing a named entity recognition model, MtlTransNER. MtlTransNER employs multi-task learning to enhance not only the performance of named entity recognition but also to avoid complex hand-engineering, thus make the model’s application to data in different domains easier. We conducted experiments on the open-source public opinion dataset [3]. The experimental results demonstrate that our approach achieved an F1 score of 0.6941 on the validation set, outperforming BiLSTM-CRF [4], BERTBiLSTM-CRF [5], BERT-CRF [6] and LLL-WCM [7]. The ablation study also show the effectiveness of each main module in MtlTransNER.