基于Attention_DenseCNN的水稻问答系统问句分类

Classification Technology of Rice Questions in Question Answer System Based on Attention_DenseCNN

摘要:

为了解决“中国农技推广APP”问答社区中水稻提问数据快速自动分类的问题，提出一种基于Attention_DenseCNN的水稻文本分类方法。根据水稻文本具备的特征，采用Word2vec方法对文本数据进行处理与分析，并结合农业分词词典对文本数据进行向量化处理，采用Word2vec方法能够有效地解决文本的高维性和稀疏性问题。对卷积神经网络（CNN）上下游卷积块之间建立一条稠密的链接，并结合注意力机制（Attention），使文本中的关键词特征得以充分体现，使文本分类模型具有更好的文本特征提取精度，从而提高了分类精确率。试验表明：基于Attention_DenseCNN的水稻问句分类模型可以提高文本特征的利用率、减少特征丢失，能够快速、准确地对水稻问句文本进行自动分类，其分类精确率及F1值分别为95.6%和94.9%，与其他7种神经网络问句分类方法相比，分类效果明显提升。

Abstract:

In the QA community of Chinese Agricultural Technology Promotion APP, thousands of rice text data questions are added every day, and the rapid and automatic classification of questions is a key step to realize the intelligent QA system of rice. However, due to the high dimensional sparsity of text data and the particularity of agricultural problems, the classification of rice questions faces difficult challenges. In order to improve the classification performance of rice question text, a convolution text classification method with dense connection was proposed. A dense connection between upstream and downstream convolution blocks was established, which enabled the model to synthesize large-scale features from small-scale features. Combined with the agricultural word segmentation dictionary, the text data was segmented into 100-dimensional word vectors by Word2vec. Neural network model’s parameters for question classification in rice question answering system were obtained by training text data with dense concatenated convolution model and attention mechanism. The experimental results showed that the text classification model based on Attention_DenseCNN can optimize the text’s representation and feature extraction, and also it can automatically classify the rice question text with accuracy of 95.6% and F1 value of 94.9%. Compared with the other seven text classification methods, the classification performance had obvious advantages.