基于CNN-Transformer的农作物病虫害知识问答意图识别与槽位填充联合模型

意图识别与槽位填充是农作物病虫害知识问答中问题理解的两个重要任务。在已有面向农业领域的研究中，上述任务仍被视为两个完全独立的子任务，并且未充分利用意图识别与槽位填充的语义信息。为此，该研究提出一种基于CNN-Transformer的意图识别与槽位填充联合模型（CDPCT-IDSF）。该模型根据农作物病虫害文本语义复杂设计CNN网络与多层Transformer结合强调局部的有用信息以缓解语义缺失问题；然后在Transformer解码器中引入对齐保证输入与输出一对一关系以提高识别正确槽位标签的能力。此外，进一步构建了包含20个意图类别、12个槽位类别和11242条标注样本的农业病虫害知识问答数据集进行对比试验，CDPCT-IDSF模型在该语料库上的槽位填充F1值为94.36%，意图识别精度为92.99%，整体识别精度为87.23%，优于其他对比模型，结果证明了所提模型在农作物病虫害意图识别与槽位填充任务上的有效性，可为面向农作物病虫害的知识问答研究提供理论支撑。

Joint model of intent detection and slot filling of knowledge question for crop diseases and pests using CNN-Transformer

Intent detection and slot filling have been two of the most vital tasks to question the crop diseases and pests using knowledge graph. Among them, intent detection is to determine the user''s intention or purpose from the text entered by the user in the crop pest field. Slot filling is to identify and extract specific information from the user''s text, especially for the slot, for example, "what is the control of tomato bacterial wilt?". The task identifies the semantic slot "tomato bacterial wilt" (head entity) and the intention "control " (relation) to form a triple (tomato bacterial wilt, control, -). Where"-" represents the semantic slot and the intention to find the answer from the knowledge graph. A graph convolution encoder has been proposed to capture the spatiotemporal semantic features of conversation discourse in recent years, in order to learn the co-occurrence relationship between intention detection and slot filling. The slot decoding context has improved the accuracy of intention detection. In turn, the target of intention detection can be optimized to further improve the performance of slot filling. The best performance has been achieved in the experiments on public data sets. Sub word attention adapter (SAA) has also been used to effectively extract the context features from complex tags, and then retain the overall discourse information. In addition, an intention attention adapter (IAA) has been proposed to obtain the comprehensive sentence features, in order to predict the users’ intention. In experiments, the knowledge question and answer have been focused mainly on the pipeline in the existing agricultural research. The named entity recognition has been firstly used to identify the head entity in the question. The text classification can be then used to classify the question for the intention. However, such a pipeline can be vulnerable to error propagation that caused by word segmentation errors. It is very necessary to make full use of the correlation between intent detection and slot filling in the agricultural field. In this study, a joint model of intent detection and slot filling was proposed to fully utilize the semantic information using CNN-Transformer (CDPCT-IDSF). According to the semantic complexity of crop pest text, the CNN network and multi-layer Transformer were designed to emphasize the local useful information. The semantic loss was alleviated to introduce the alignment guarantee one-to-one relationship between input and output in the Transformer decoder. As such, the model was improved to identify the correct slot labels. In addition, a question-and-answer dataset was constructed for the agricultural pest knowledge, including 20 intention categories, 12 slot categories, and 11242 labeled samples. Comparative experiments show that the slot-filling F1 value of the CDPCT-IDSF model on the corpus was 94.36%, the intent detection accuracy was 92.99%, and the overall recognition accuracy was 87.23%, which was better than other comparison models. The improved model was also performed better for the intent detection and slot filling in crop diseases and pests. The finding can provide the theoretical support to knowledge question and answer research on crop diseases and pests. Moreover, the experiments were also carried out on two public datasets. CDPCT-IDSF was also suitable for English corpus scenarios. On ATIS datasets, the slot filling F1 value was 95.43%, intent detection accuracy was 97.95%, and overall recognition accuracy was 87.68%. On SNIPS datasets, the slot filling F1 value, intent detection accuracy and overall recognition accuracy of CDPCT-IDSF were 95.70%, 98.57% and 89.96%, respectively. At the same time, the better generalization and robustness were achieved in the rest corpus.