Knowledge intelligence service can improve the production efficiency and decision-making in smart agriculture. Agricultural question-answering (QA) systems can integrate natural language processing and knowledge retrieval for precise cultivation. However, the conventional QA systems cannot resolve the ambiguous queries due to the complexity and domain specificity of the agricultural knowledge. In this study, a maize knowledge graph QA framework was proposed to clarify the QA systems using large language models (LLMs). The framework was composed of the knowledge retriever, fact analyzer, clarification question generator, and answer generator. These modules were interacted with via action tokens to control the clarification. Knowledge retrieval was implemented by a semantic relation matching model. The ComplEx was used to calculate the relevance score between entities and the question for the candidate answer retrieval. An encoder-decoder structure was employed to generate both clarification question templates and disambiguating options. A contextual Transformer was employed to jointly encode the user query and candidate answer entities during encoding. An example was taken as the technical QA in the context of maize cultivation. A maize knowledge graph with the LLMs was constructed to extract the fine-grained attributes and relations from the variety websites. An ambiguous QA multi-turn dialogue dataset and a clarification QA trajectory dataset were constructed to evaluate the proactive interaction of the QA system using the maize knowledge graph. A comparison was made on the ambiguous QA task in the existing LLMs (ChatGLM-4, Qwen2.5, Llama3.3, and ShizishanGPT). The structured knowledge graph was used to retrieve the domain-specific knowledge. The users were asked to generate the contextually clarification questions in a multi-turn manner. Thereby, an accurate identification was realized to fully meet their actual information needs. Compared with the optimal baseline models, there were improvements of 21.07, 15.53, 24.38, and 6.63 percentage points on the ROUGE-1, ROUGE-2, ROUGE-L, and BLEU-4 metrics, respectively, on the ambiguous QA multi-turn dialogue dataset. Furthermore, the candidate answer selection model demonstrated that high performance was achieved to effectively and accurately retrieve the possible answers corresponding to the ambiguous questions from the knowledge graph in most cases, which was improved by 23.7, 11.8, and 9.4 percentage points on the P@1, R@1, and mAP metrics, respectively. The clarification question generation model also outperformed the best baseline model, thus achieving improvements of 3.45, 6.57, 3.67, and 6.52 percentage points on the ROUGE-1, ROUGE-2, ROUGE-L, and BLEU-4, respectively. The ambiguities within user inputs were accurately detected to generate the clarification questions on the uncertainties. Furthermore, ablation experiments were performed to assess the contribution rate of the knowledge retrieval module, the question clarification module, and the trajectory learning mechanism. Both modules enhanced the overall performance of the QA system. Finally, the performance of the various models was verified using real-world agricultural QA cases. These practical evaluations were validated to effectively identify the ambiguities within the user queries. The contextually appropriate clarification questions were generated to provide accurate answers, indicating the applicability in the complex agricultural knowledge scenarios. In summary, the agricultural knowledge graph QA framework can be expected for proactive interaction in the QA systems. The interactivity and usability are enhanced in practical applications, such as agricultural consultation and intelligent agricultural customer service. Particularly in knowledge-intensive domains, such as agriculture, healthcare, and law, this approach can effectively bridge the knowledge gaps for the QA system to understand and accurately respond to ambiguous queries.

A knowledge graph is often required to collect and integrate the vast amount of data with the rapid development of agricultural informatization in recent years. But the data can be characterized by dispersal and fragmentation. Relevant data is usually stored separately in different institutions and databases, leading to data sparsity and missing information in the agricultural knowledge system. Among them, knowledge graph completion can be used to extract the potential semantic information using a neural network. The domain-specific data features can also be obtained after special domain knowledge graph applications using convolution and pooling operations. However, existing knowledge completion models still struggle to effectively capture the long-range dependencies in the complex relational scenarios. In this study, a two-stage knowledge graph completion model (RASE-ARKGC) was proposed using large language models with rule augmentation and channel attention. The wheat was taken as an example. In the first stage, the large language models (LLMs) were used for the rule augmentation (RA). In particular, both the semantic and structural information of the knowledge graph were utilized to generate the logical rules. Firstly, the paths were sampled from the knowledge graph to represent the structural information. Secondly, the rule generator was adopted to mine the potential rules using semantic and structural information. Thirdly, the logical rule sequence was used to evaluate the quality of the rules and then filter out the outliers. As such, the rule augmentation module effectively expanded the dataset in order to alleviate the data sparsity in the knowledge graph. The channel attention mechanism and dilated convolution (SE-ARKGC) were introduced in the second stage. Specifically, the channel attention mechanism was enhanced to capture the long-range dependencies. While the dilated convolution was used to adjust the dilation rate for the multi-scale receptive fields, in order to capture the multi-level interactions between entities and relationships. In addition, the residual module was also used to effectively mitigate the gradient vanishing in the deep network, in order to further enhance the expression of the model. The complete transfer of feature information was then realized in the multilayer network. A series of experiments was conducted on the self-constructed wheat knowledge graph (WheatSeedBiz) in order to validate the effectiveness of the two-stage knowledge graph completion. The results showed that the RASE-ARKGC model achieved 0.482 in MRR and 0.555 in Hits@10, which were improved by 9.8% and 10.2%, respectively, compared with the ConvE model. In addition, the experiments were conducted on the open-source datasets WN18 and FB15k to evaluate the generalization of the SE-ARKGC model. The experimental results showed that the optimal or sub-optimal performance was achieved over multiple baseline models. Ablation experiments were also carried out to evaluate the rule augmentation and the SE-Block-based knowledge graph completion. Each module was enhanced the performance of the model. Overall, the RASE-ARKGC model effectively expanded the dataset for optimal performance in the knowledge completion task. The finding can provide an improved approach for the knowledge graph completion in the wheat domain. The generalization of the model can be applicable to the rest domains.

Deep convolution neural networks constitute a breakthrough in computer vision. Based on this, the Convolutional Neural Network (CNN) models offer enormous potential for crop disease classification. However, significant training data are required to realize their potential. In the case of crop disease image recognition, especially with complex backgrounds, it is sometimes difficult to acquire adequately labeled large datasets. This research proposed a solution to this problem that integrates multi-attention modules, i.e., channel and position block (CPB) module. Given an intermediate feature map, the CPB module can infer attention maps in parallel with the channel and position. The attention maps can then be multiplied to form input feature maps for adaptive feature refinement. This provides a simple yet effective intermediate attention structure for CNNs. The module is also lightweight and produces little overhead. Some experiments on cucumber and rice image datasets with complex backgrounds were conducted to validate the effectiveness of the CPB module. The experiments included different module locations and class activation map display characteristics. The classification accuracy reached 96.67% on the cucumber disease image dataset and 95.29% on the rice disease image dataset. The results show that the CPB module can effectively classify crop disease images with complex backgrounds, even on small-scale datasets, which providing a reference for crop disease image recognition method under complex background conditions in the field.