| 李娜.基于深度学习的《方志物产》用途实体自动识别模型构建与应用[J].数字图书馆论坛,2022,(12):19~28 |
| 基于深度学习的《方志物产》用途实体自动识别模型构建与应用 |
| Construction of Automatic Recognition Model of Function Entities in Local Chronicles: Produce Based on Deep Learning |
| 投稿时间：2022-11-08 |
| DOI：10.3772/j.issn.1673-2286.2022.12.003 |
| 中文关键词: 深度学习；方志物产；命名实体识别；数字人文；用途实体 |
| 英文关键词: Deep Learning; Local Chronicles: Produce; Named Entity Recognition; Digital Humanities; Function Entities |
| 基金项目: |
|
| 摘要点击次数: 1167 |
| 全文下载次数: 995 |
| 中文摘要: |
| 以特色馆藏文献《方志物产》为研究语料，基于人工标注语料，运用Bi-LSTM、Bi-LSTM-CRF、BERT、Siku-BERT等4种深度学习模型开展实验，以精确率P、召回率R、调和平均数F作为测试指标，对模型的识别性能进行对比分析，促进物产知识的挖掘和利用。实验结果显示：相较于基于CRF的模型，4种深度学习模型的整体性能取得明显提升；Bi-LSTM、Bi-LSTM-CRF、BERT、Siku-BERT的最好R值分别为74.80%、78.05%、88.62%、89.74%；BERT、Siku-BERT注意力机制类深度学习模型的识别效果优于Bi-LSTM、Bi-LSTM-CRF循环类深度学习模型。由于方志类古籍文本结构复杂多样、人工标注精度存在误差、语料规模较小等因素，自动识别模型的实体抽取性能仍有较大的优化空间，但深度学习模型在方志类古籍的内容挖掘中表现出一定的优越性，且不同语料间预训练模型的迁移应用具有可行性。 |
| 英文摘要: |
| Taking the local chronicles as the research corpus, based on the manually labeled corpus, we use four deep learning models such as Bi-LSTM,Bi-LSTM-CRF, BERT and Siku-BERT to carry out experiments, and then use the accuracy rate P, recall rate R and F-value as test indicators to compare and analyze the recognition performance of the models, so as to promote the mining and utilization of product knowledge. The experimental results show that: Compared with the previous model based on CRF, the overall performances of the four deep learning models have been significantly improved; The best R-values of Bi-LSTM, Bi-LSTM-CRF, BERT and Siku-BERT are 74.80%, 78.05%, 88.62% and 89.74% respectively; The recognition effects of attention mechanism deep learning models such as BERT and Siku-BERT are better than those of cyclic deep learning models such as Bi-LSTM and Bi-LSTM CRF. Although there is still much room for further optimization of the model performance due to the structural characteristics of the local chronicle ancient books, the imperfect manual annotation of the corpus, the small scale of the corpus and other factors, the deep learning models show certain superiority in the content mining of local chronicles, and the migration and application of the pre-training model between different corpora are feasible. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
