融合动态词典特征和CBAM的苹果病虫害命名实体识别方法

Named Entity Recognition of Apple Diseases and Pests Based on Dynamic Dictionary Features and CBAM

摘要:

在苹果病虫害命名实体识别中，针对罕见字语义特征提取不充分，实体类别相似难以区分的问题，本文提出一种融合动态词典和卷积块注意力模块(Convolutional block attention module, CBAM)的实体识别方法。首先，基于字的双向长短时记忆-条件随机场模型(Bidirectional long short-term memory-conditional random field, BiLSTM-CRF)，在嵌入层利用通道注意力网络(Channel attention module, CAM)动态融合词典信息，同步集成字的四角号码信息，以提高对罕见字表征能力。随后，对序列编码层输出序列特征，基于空间注意力网络(Spatial attention module, SAM)，新增并行连接的空间注意力(Parallel connection spatial attention, PCSA)模块，提高模型对上下文信息提取能力。最后，使用含有6大类标签、127574个标注字符的苹果病虫害数据集进行验证测试。结果显示模型精确率、召回率和F1值分别达到95.76%、92.46%、94.08%，较现有的常用同类模型性能显著提升，实现了对农业病虫害命名实体的精准识别。

Abstract:

In the named entity recognition of apple diseases and pests, a entity recognition model was proposed to address the problems of insufficient semantic feature extraction for rare words and difficulties in distinguishing entities due to similar entity categories. This model integrated dynamic lexicon and convolutional block attention module (CBAM). Firstly, based on the bidirectional long short-term memory-conditional random field model (BiLSTM-CRF), a channel attention module (CAM) was used to dynamically obtain lexicon information for the words, and the fourcorner code information of Chinese characters was simultaneously fused to enhance the representation ability for rare words. Then after the sequence features output by the sequence encoding layer, a parallel connection spatial attention (PCSA) module based on the spatial attention module (SAM) was added to improve the model’s ability to extract contextual information. Finally, the model was validated and tested by using an apple disease and pest dataset which contained six major classes and 127574 annotated characters. The results showed that the precision, recall, and F1 value could reach 95.76%, 92.46% and 94.08%, respectively，indicating a significant improvement in performance compared with existing commonly used similar models, which achieved accurate recognition of agricultural disease and pest named entities.