基于注意力池化和堆叠式结构的病虫害文献识别模型

Diseases and Pests Articles Identification Model Based on Attention Pooling and Stacked Structure

为解决病虫害文献识别过程中存在语义特征学习不够、上下文信息不能充分利用等问题,以病虫害相关文献摘要为研究对象,提出一种基于注意力池化策略和堆叠式双向长短期记忆(Bi-directional long-short term memory,BiLSTM)的神经网络模型(AP-LSTM)。该模型采用堆叠式长短期记忆结构,提高了对语义特征的学习能力,在进行堆叠操作时,通过将输入向量与输出向量拼接,进一步加强了对语义信息的表征;然后采用基于注意力机制的池化策略为不同的词分配不同权重,使模型在抓住重点的同时能够充分利用上下文信息。本文在包含1439条正例、1061条负例的自标注数据集上进行了实验,所提出的AP-LSTM模型在该数据集上的精确率、召回率、F1值和准确率分别为92.67%、97.20%、94.88%和94.00%,实验结果表明,AP-LSTM模型能够有效识别病虫害文献。

Diseases and pests articles identification is an important pre-task of natural language processing in the field of diseases and pests.It is of great significance to develop a fast and accurate method for diseases and pests articles identification.In order to solve the problems of insufficient learning of semantic features and insufficient use of context information in the process of diseases and pests articles identification,a neural network model of attention pooling based bi-directional long-short term memory(AP-LSTM)was proposed,which was based on attention pooling strategy and bi-directional long-short term memory(BiLSTM).The model adopted the stacked LSTM structure,which improved the learning ability of semantic features.In the stacking operation,the input vector and output vector were concatenated to further enhance the representation of semantic information.Then,a pooling strategy based on the attention mechanism was used to assign different weights to different words,so that the model can make full use of context information while grasping the keywords.The experiments were carried out on a self annotated dataset with 2500 labeled samples,including 1439 positive cases and 1061 negative cases.The precision,recall,F1 score,and accuracy of the proposed AP-LSTM model on the dataset were 92.67%,97.20%,94.88%,and 94.00%,respectively.The experimental results showed that the proposed AP-LSTM model can effectively identify pest literature.