ASABE Logo

Article Request Page ASABE Journal Article

Relation Extraction for Knowledge Graph Generation in the Agriculture Domain: A Case Study on Soybean Pests and Disease

Pengxiang Wang1, Cong Zhang1, Dingqian Wang1, Shaohua Zhang1, Jun Wang2, Xianzhi Wang3, Lan Huang1,*

Published in Applied Engineering in Agriculture 39(2): 215-224 (doi: 10.13031/aea.15124). Copyright 2023 American Society of Agricultural and Biological Engineers.

1Computer Science, Yangtze University, Jingzhou, Hubei, China.

2Agriculture, Yangtze University, Jingzhou, Hubei, China.

3Agriculture, Yunnan University, Kunming, Yunnan, China.

*Correspondence: lanhuang@yangtzeu.edu.cn

Submitted for review on 24 March 2022 as manuscript number ITSC 15124; approved for publication as a Research Article and as part of the Artificial Intelligence Applied to Agricultural and Food Systems Collection by Associate Editor Dr. Garey Fox and Community Editor Dr. Yiannis Ampatzidis of the Information Technology, Sensors, & Control Systems Community of ASABE on 3 March 2023.

Highlights

ABSTRACT. Precision agriculture is an emerging type of agriculture that intensively uses information technology to automate agricultural production. Soybean (Glycine max (L.) Merri.), is an important crop in China, with an annual demand of approximately 110 million tons. However, in China, soybean production is threatened by more than 30 kinds of disease and 100 kinds of pests. With the rapidly increasing specialized information in the literature, it is difficult for farmers to keep up. Relation extraction automatically identifies and extracts structured knowledge from natural language text and thus can help to alleviate the problem. In this study, we propose to employ relation extraction to systematically extract information from expert-written text, and generate a knowledge graph from the extracted information. This case study was planned in China, therefore we mainly used Chinese texts. Firstly, we carefully chose expert-written text on soybean pests and disease, labeled the entities, and classified their thematic relations into five categories. Then, we built and trained three relation extraction models using state-of-the-art deep learning architectures and evaluated their performance on our task. Finally, we constructed an example knowledge graph from the extracted information and demonstrated their potential usage for automatic reasoning and solution recommendation for pests and disease prevention. In total, this study sampled 1038 entities and 1569 relation instances. Experimental results showed that our best model achieved an F1 score of 98.49% on identifying relations from text. Experimental results also showed the effectiveness of the example knowledge graph.

Keywords. Bidirectional encoder representation from transformers, Knowledge graph, Relation extraction, Soybean pests and disease.

China has been constantly facing great threats from agricultural and biological disasters, especially with the increasing population of pests and disease (Li and Cai, 2017). Major breakouts can endanger agricultural production, affect food safety, and cause severe economic losses. Traditional cultivation heavily relies on human labor and farmers’ expert knowledge, resulting in less timely reaction to such breakouts. Recent advances in artificial intelligence and the Internet of Things have brought smart and precise agriculture into possibility. Computers can now automatically detect disease and pests in the fields using computer vision techniques. Yet, finding the proper treatment for the detected disease and pests still requires up-to-date expert knowledge, whose acquisition, exploration, and utilization remain to be automated. Relation extraction is a major approach for automatic acquisition of such knowledge from the rapidly growing literature and other domain specific text, while knowledge graph has become an emerging information retrieval form that supports friendly knowledge exploration and utilization. This article attempts to narrow this gap by investigating effective relation extraction and knowledge graph construction methods for the agricultural domain.

Soybean (Glycine max (L.) Merri.) is an important crop source in China (Kui and Qi, 2021), with an annual consumption of about 110 million tons. Soybean production can be affected by more than 30 kinds of disease and about 100 kinds of pests. Timely and effective detection and control of such threats can avoid losses of millions of dollars. While means of pest and disease control develop fast, farmers usually lack the time and skills to keep up with the rapidly growing scientific literature. In order to meet these challenges and strengthen the connection between academic research results and the actual fields, we first adopted knowledge graph, an effective visualization mechanism for organizing complex knowledge, as the connecting platform. Then we employed entity and relation extraction to automatically extract important information from relevant texts written by domain experts. Finally, the knowledge graph was automatically constructed from the extracted information, which then served as a friendly and effective information retrieval portal for farmers. Figure 1 shows an exemplary scenario of applying the proposed technique.

Figure 1. An example process of farmers using the proposed system to deal with pests and disease.

Knowledge graph (Liu et al., 2020) is a semantic network that records entities and their semantic relations, usually consisting of triplets <e1, r, e2>, where e1 and e2 denote two distinct entities and r denotes the relation between them. Hundreds and thousands of such triplets form a complex graph, which becomes the basis for information retrieval, automatic reasoning and decision-making. Automatic construction of such a knowledge graph requires entity relation extraction (Wan et al., 2005), a sub-task of information extraction (Guo and He, 2015) that automatically identifies entities and their relations from natural language text. Methods for entity relation extraction can be categorized by the underlying methodology: pattern matching, machine learning and deep learning. Pattern matching requires manually constructing features and rules that can describe mentions of entities and their relations in text, thus is labor intensive and error-prone. Manually created patterns also lacked variety. Machine-learning-based methods focus on capturing the statistical traits of relations using feature vectors (Kambhatla, 2004), kernel functions (Zelenko et al., 2003), bootstrapping (Brin, 1998), clustering (Hasegawa et al., 2004), and logistic regression (Wang, 2008). These methods still heavily depend on manually designed features, therefore cannot adapt to large-scale prediction in the open domain. With the emergence of big data, deep neural networks have demonstrated their capabilities in feature learning. Models like Convolutional Neural Networks (CNN) (Nguyen and Grishman, 2015) and variants (Zeng et al., 2015), Recurrent Neural Networks (RNN) (Socher et al., 2012), Long Short-Term Memory (LSTM) networks (Xu et al., 2015), Graph Convolutional Networks (GCN) (Schlichtkrull et al., 2017), and the Bidirectional Encoder Representation from Transformers (BERT) (Devlin et al., 2019) have demonstrated their capabilities in learning task-specific patterns together with the general patterns.

Advances in artificial intelligence have also impacted modern agriculture. Some researched on using computer vision to improve soybean planting (Karlekar and Seal, 2020; Tetila et al., 2020; Zhang et al., 2021; Zhao et al., 2021) and other agricultural activities (Gao et al., 2013; Abdalla et al., 2019; Zhou et al., 2020; Ren et al., 2020; Ahmad et al., 2021). Natural language processing for improving agriculture has also gradually gained research interest. For example, Borja Espejo-Garcia et al. (2018, 2019) developed end-to-end sequence labelers for phytosanitary regulations using different deep-learning architectures and investigated the automatic translation of Spanish agricultural regulations into machine processable representations. Li et al. (2021) analyzed agricultural Text-to-Speech (TTS) and discussed text analysis, rhythm generation, and speech synthesis of the TTS system. Wang et al. (2021, 2022) proposed a BERT based similarity matching model for agricultural questions, which rapidly and automatically detects questions that have the same semantic content, and proposed classification model based on the CNN structure to recognize questions about rice in particular. Jin et al. (2020) proposed a classification model, also based on the CNN structure, to distinguish agricultural questions and their short text answers. Yang and Yang (2020) also investigated the question answering task, with the focus on common crop disease.

Different from previous studies, we focused on the task of automatically extracting important information from text, organizing and presenting it to farmers to facilitate their capabilities in preventing and dealing with pests and disease in fields. We focused on four important types of relations: for a given pest or a disease, the reside_in relation denotes its affecting location, the symptom_for relation denotes its classic symptons, the characterized_by relation denotes its morphological characteristics, and the solution_for relation denotes its solution. We collected and constructed a dataset from expert-written texts, then investigated the effectiveness of several state-of-the-art deep learning models for automatically identifying such relations, built a knowledge graph from the extracted relations and made it publicly available (http://yuanyvki.xyz/). Both experimental results and the prototype knowledge graph demonstrated the effectiveness of the proposed method.

Datasets

Datasets Source

The experimental data came from The Colored Atlas of Primary Soybean Pests and disease edited by Li and Cai (2017), which records 19 disease and 41 pests that are commonly seen in China and have a great impact on soybean planting (sup. table 1). Each item is explained in great detail, for example, its distribution in China, morphological characteristics, occurrence patterns and control methods.

Datasets Processing

Before labeling the text, we consulted recent work on relation extraction that involves generating new datasets for annotation. Ren et al. (2021) screened 2000 scientific papers and web pages on tomato disease and pests and labeled 13 relation types. Yuan et al. (2021) labeled six relations in rice pest control from 1581 text segments. Wu et al. (2020) collected 14 relations from 1619 sentences on crop pests and disease from the crop germplasm information network of China.

Following related work, a total of 1569 sentences were screened from the source book, with an average sentence length of 36 words. From the text, four relation categories most relevant to disease and pest control were identified: reside_in, symptom_for, characterized_by, and solution_for plus another category for relations that do not belong to the preceding classes. Table 1 shows the number of instances identified from the text and an example sentence for each relation type.

Table 1. The number of instances and an example sentence translated into English for each relation category.
Relation CategoryNo. InstancesExample Sentence
reside_in 237Soybean anthrax mainly damages stems and pods.
symptom_for 350Soybean virus disease can cause grain mottle.
characterized_by 293Riptortus pedestris bug: 15~17mm long, narrow, yellowish brown to dark brown.
solution_for 591Insect resistant varieties or insect resistant varieties should be selected for the control of soybean heartworm.
others 98Soybean sheath blight disease is easy to occur when soybean seedlings encounter low temperature and heavy rain.

Framework

Figure 1 shows the general steps of our framework. Constructing such a knowledge graph usually has four steps. Using a domain-specific corpus, e.g., a reference book, entities are identified and their relations are extracted with pre-trained relation extraction models. The extracted relations, in form of triplets, are integrated and stored in files or databases. When the end-user queries the knowledge graph, automatic reasoning will be performed and critical paths related to the submitted query will be retrieved and highlighted, e.g. suggesting possible treatments for a disease or a pest. The subgraphs containing the relevant paths are then retrieved and visualized on the fly, as a friendly answer to the end user’s search queries.

Relation Extraction Models

Relation extraction has long been an important problem in natural language processing. Traditionally, relation extraction tasks used models like eigenvector (Kambhatla, 2004), kernel functions (Zelenko et al., 2003), bootstrapping (Brin, 1998), clustering (Hasegawa et al., 2004), and logistic regression (Wang, 2008). In recent years, deep neural networks have achieved wide success in many natural language tasks. Deep neural networks can learn the importance of word semantics and liguistic features such as grammar structures from training text, avoiding the tedious and often stereotyped manual encoding. In this paper, we built relation extraction models on top of three representative frameworks: convolutional neural network (CNN) (Nguyen and Grishman, 2015), piecewise CNN (PCNN) (Zeng et al., 2015) and BERT (Devlin et al., 2019), and evaluated their performances on identifying important relations on soybean disease and pests.

The CNN-Based Models

The convolutional neural network (CNN) relation extraction model mainly consists of five layers: the embedding layer, the convolution layer, the pooling layer, the fully-connected layer, and the output layer (fig. 2). The embedding layer generates input representation and is usually the same for different models, therefore it is described separately from the other layers. The PCNN model only differs from the CNN model in the pooling layer and therefore is also described here.

The embedding layer generates the necessary representations of the input sentences. In this study, we used both word embedding and position embedding. Word embedding represents each word with a vector with continuous values, mapping words into a new space where semantically similar words become closer and the similarity between two words can be quantified. Let n be the length of the relation mentioned, for each word xi, it is mapped to vector ei through word embedding matrix W with dimension me. Matrix W can be initialized either by a random process or by pre-trained word embeddings, usually generated from open domain corpora. We used the Global Vectors for Word Representation (GloVe) model (Pennington et al., 2014) as the pre-trained word embeddings.

Figure 2. The general structure of a CNN-based relation extraction model.

Position embeddings encode the relative distances between words in a sentence. For each word xi, its relative distances to the two entities xi1 and xi2 are i-i1 and i-i2. The two distances are mapped into two vectors di1 and di2 through the position embedding matrix D. D can be initialized randomly. Taking the following sentence as an example, the relative distances of damage to the two entities soybean sheath blight and seedlings are 2 and -1, respectively.

S:[Soybean sheath blight]0 [mainly]1 [damages]2 [seedlings]3)

Finally, the word embedding eiand the position embeddings d1 and d2 are concatenated into a single vector Xi = [ei, di1, di2]T to represent word xi, and the sentence x with n words becomes a matrix X = [X1, X2, ..., Xn].

Matrix X is then input into the convolution layer, which surpassingly extracts higher-level features. Given the convolution window size w, the filter is seen as a weight matrix f = [f1, f2,..., fw] and is applied to X, producing a score sequence S = [S1, S2,..., Sn-w+1]. Given b as the bias, and g as the activation function, Si is computed as:

(1)

The score sequences are then output to the max pooling layer, which retains the largest score in S:

(2)

Let the number of filters be m, the output of the convolution layer is Z = [p1, p2,..., pm]. Dropout regularization is then performed on Z to improve robustness. The adjusted Z then passes through a fully connected layer and the softmax layer. The softmax layer builds a softmax classifier to predict the probability of the input entity pair belonging to each relation category. Given the category labels y ? {1, 2,..., N} and the input x, the probability of x belonging to category c is:

(3)

In theory, the convolution and the pooling layers can be seen as a combined structure and can be replicated multiple times to form a deep layered neural network. In this article, as in previous studies, we used one such structure.

In the standard CNN model, the max pooling filters retain the largest value in a given window, thereby can lose the structural information between two entities. Piecewise max pooling, on the other hand, divides the output of the convolution layer into three parts based on the location of entities, and segments the output vector into three sub-vectors to offer a sentence level representation. Finally, all convolution kernels are spliced into segmented pooling layers to output P1:n:

(4)

The Bert-Based Model

BERT builds upon the transformer structures. Transformer is a deep neural network framework that uses self-attention instead of RNN to capture both local and global sequential information, thereby can benefit natural language tasks. Such a model usually consists of three parts: input, transformers and output.

Before inputting to the learning module, we used BERT-base-Chinese to pre-train text. For annotation, each sentence is labeled with [CLS] and [SEP] at the beginning and end of the sentence respectively, and each entity is labeled with [E1start], [E1end], [E2start] and [E2end] at the start and the end of the first and the second entity. An exemplary annotated sentences is:

S:[CLS]Soybean [E1start] phyllosticta leaf spot [E1end] mainly damages [E2start] leaves [E2end][SEP]

A BERT model usually consists of multiple transformer encoders (fig. 3). Each encoder encodes local information, and the model as a whole can potentially capture global information within the input text.

Transformer can be used as an encoder or a decoder. BERT uses the encoder to encode both the semantic information and the location information. The encoder first embeds words into vectors and composes them into a vector sequence. Then it uses self-attention to capture relations between all words in the sequence. The processed sequence is then summed and normalized through the residual network and delivered to the fully connected feed-forward network, and standardized before output. Figure 4 shows the structure of a transformer encoder.

Figure 3. The general structure of a BERT model.
Figure 4. Structure of a single transformer encoder.

The self-attention mechanism is the cornerstone of the transformer encoder. The attention mechanism originated from cognitive science. It can identify the most relevant parts from the input. BERT uses bidirectional self-attention, a hierarchical multi-level architecture that models the representation of context paragraphs at different granularities. Basically, bidirectional self-attention is a mapping from query Q to corresponding key value pairs <K, V>, where dk is vector Q’s dimension.

(5)

Firstly, the similarity between the query Q and each key K is calculated. The similarity function can be simply the dot product function. Secondly, the softmax function is used to normalize the weight. Finally, the weight and the key value are weighted by the median value V.

As for the output layer, two methods were used, one used the [CLS] output, and the other used the concatenation of the corresponding output from the hidden layers of the two entity’s start tags. Figure 5 shows the two different methods.

Experiments and Results

Experimental Setup

Following common practices and related work (Wu et al., 2020; Ren et al., 2021; Yuan et al., 2021), we split the dataset into three subsets for training, validation, and testing respectively, with the proportion of 7:1:2. All samples are stratified to ensure different categories share similar distributions in all subsets. Training set was used for learning the model parameters, validation set was used for validating the hyper-parameters and the test set was used for evaluating the learned model's predictive performance.

Training of the CNN-based models involve several important hyper-parameters, for example the learning rate and the dropout rate. Learning rate determines the convergence speed of the learning process. Dropout rate decides the proportion of nodes in each hidden layer being discarded, and can avoid overfitting and increase the generalizability of the learned model. We adopted a learning rate of 2×10-5 and the ADAWM optimizer (Loshchlov and Hutter, 2017). Table 2 lists the major hyper-parameters used in CNN, PCNN, and BERT. All experiments were conducted using an Intel Corei5-10600KF processor @4.10GHz, 16GB memory and 1 TGB hard disk on Linux.

Performance Evaluation Indexes

Precision (P), recall (R), and F1 measures were used to evaluate and compare the performance of different models. Precision indicates how many of the predicted positive samples are true positive. Recall indicates how many positive examples in the sample are predicted correctly. Precision and recall are computed as:

(6)

(7)

where TP, FP, FN, and TN represent the number of true positives, false positives, false negatives, and true negatives respectively. There usually exists a trade-off between precision and recall: the higher the precision the lower the recall and vice versa. F value is used to balance the two and ideally reaches the maximum when both precision and recall obtain maximum. The general equation is:

(8)

ß measures the relative importance of precision to recall. In practical applications, precision, and recall are usually equally important, and ß is generally set to 1, resulting the F1 measure:

Figure 5. Two output layer formats of BERT (Above: uses [CLS] as output; Below: uses entity start as output).

(9)

Relation Extraction Results and Analysis

Table 2. Hyper-parameters in training CNN-based and BERT-based models.
ModelHyperparameterValue
BERTBatch size4
Gradient descent optimizerADAWM
Hidden layer size768
Learning rate decay2×10-5
Maximum sequence length128
CNN/PCNNActivation FunctionReLU
Convolution kernel size3
Dropout ratio0.5
Hidden layers230
Learning rate decay0.1
Padding size1
Weight decay1×10-5
Word vector dimension100

Table 3 shows the overall performance in terms of precision, recall, and F1 of the three models, with the best results noted in bold. The BERT model with different output schemes is noted separately as BERT[CLS] and BERTentity. Experimental results showed that PCNN outperformed CNN, indicating semantics above word levels are helpful for the entity relation identification task since the major difference between the two models is that PCNN uses piecewise max pooling. BERTentity attained the best performance on this task, which further demonstrated the usefulness of including sentence-level semantics into the model. The self-attention mechanism further enables BERT to execute concurrently and extract the relational information between words and within the sentence, capturing relational features and sentence semantics at multiple levels. Since BERTentity was the best performing model, all following reported experimental results were produced using this model.

Table 3. Overall performance of different models.
ModelPRF1
CNN0.95650.95970.9581
PCNN0.96950.95970.9649
BERT[CLS]0.96010.96970.9649
BERTentity0.98320.98650.9849

As an important component of the model, the impact of different optimization methods were investigated. Figure 6 compares three gradient descent algorithms ADAMW, ADAM (Kingma and Ba, 2014), and SGD (Bottou, 2010), when their batch size were all set to 4. It can be seen that ADAMW and ADAM were comparable and both significantly better than SGD.

Figure 6. Performance of different optimization methods.

Apart from the optimization method, the batch size is another important factor that can affect learning performance. We experimented with 1, 2, and 4 respectively, using ADAMW and BERT. Figure 7 compared their performances. Batch size of 4 obtained the best results. This suggests more information from other sentences might also help for identifying entities within a sentence.

Table 4 shows the performance of the best performing model on different relation types. On all categories, BERTentity achieved extraordinary results, indicating the quality of annotated data. Specifically, the characterized_by relation type obtained perfect precision and recall scores. Table 5 shows the sentences with classification errors. There are only five such cases.

Figure 7. Performance of different batch sizes.
Table 4. Performance on different relation types.
Relation TypePRF1Size
reside_in0.95920.97920.9691237
symptom_for0.98590.95890.9722350
characterized_by111293
solution_for10.99170.9958591
others0.904810.9598
Table 5. Sentences where the entity extraction model failed (entity1: “”, entity2: “”).
SentencesGround truthClassified as
Soybean virus disease will lead to pod-setting stage. The symptoms are round
or irregular brown patches on the pods, and the pods are often deformed.		othersreside_in
The adults and nymphs of whitefly suck the plant leaf juice with a stinging mouthpiece,
causing the leaves to fade green, turn yellow, wilt, and even wither the whole plant.		reside_in symptom_for
Disease resistant varieties should be selected for the control of soybean bacterial macular disease,
disease-free seeds should be selected, and seeds should be disinfected before sowing.		reside_insymptom_for
The symptoms caused by rhizoictona can occur from the germination of seeds, causing
rotten seeds, and the symptoms of bacterial wilt appear a few days after emergence.		symptom_forreside_in
Soybean epidemic disease extend to the petiole, making the base of the petiole brown
and concave, and the leaves droop and wither in an “eight” shape, but do not fall off.		symptom_forreside_in

Knowledge Graph Construction Results and Analysis

Figure 8. The overall knowledge graph.

The extracted relations need to be organized to serve the end users. Knowledge graph is a powerful tool for such a purpose. In recent years, many knowledge graph tools have been developed, supporting a variety of data formats and graph scales. In agronomy, common tools include Pajek, CiteSpace, Ucinet, and Neo4j. For example, Qin (2009) used Pajek to generate a knowledge graph on Chinese agricultural history, Dai and Cai (2015) used CiteSpace to study and analyze agricultural water saving, Yang and Zhou (2013) analyzed agricultural e-commerce both in China and worldwide with Ucinet, and Li and Zhu (2020) chose Neo4j to visualize their pesticides knowledge graph. We used Neo4j to store entity and relational data. Neo4j stores records as triplets and can adapt to different problem scales. Each entity pair associated with its relation type form a triplet. These triplets are then rendered into a graph, where entities become nodes and relations become edges. Different entities and relations interconnect with each other, thereby forming a complex graph structure. Figure 8 shows the overall knowledge graph.

Figure 9. Example knowledge graph for query the “root rot disease”. (a) the resulting subgraph with all relations; (b) subgraph with the reside_in and sympton_for relations; (c) subgraph with the solution_for relations.

When end user queries the knowledge graph, relevant nodes and edges, i.e., entities and relations, will be identified and presented, generating a sub-graph on the fly. Neo4j uses Cypher (Holzschuher and Peinl, 2013) as the query language, which is an efficient query language for graph databases. The resulting graphical interface succinctly presents only the most relevant parts from the complete knowledge graph, greatly facilitating information seeking. Taking root rot disease as an example, when end user queries for this disease, the resulting succinct knowledge graph is shown in figure 9a. Different types of related information were noted with different colors. Blue, yellow, green and purple colors correspond to the symptom_for, the reside_in, the solution_for and the characterized_by categories respectively. Using the graph in figure 9a, farmers can easily retrieve various types of information relating to the disease. For example, it shows all the possible locations on a soybean plant that could be affected: it can not only affect the lower part of the main root, but also affect the plant surface, shoot and leaf. It also showed the different symptoms once different areas become affected with multi-hop relations: leaf would turn yellow, skin would turn rotten and black, root top would be water soaked, the lower part of the main root would become rotten and have brown strips, the entire plant would be dwarfed or even dead. Figure 9b shows the location and the resulting symptoms. Meanwhile, a diversity of agrochemicals and techniques that could help to prevent and treat the root rot disease are also shown in the graph, as detailed in figure 9c. The applicable agrochemicals include phosphate fertilizers, morpholycin aqueous, four particular seed coating suspension products, two types of carbendazim products and two anti-fungal aqueous medicines. Treating and preventing techniques include rational crop rotation, control of sowing depth, soil ventilation, deep soil loosening, moderate plowing and removing standing water in the fields. This is one example scenario of how the knowledge graph could be used. When encountering a particular symptom, for example yellowed leaves, the end user can also query the knowledge graph for possible causes. When clicking on one cause node, more symptom nodes associated with this cause will appear. End user can then reason and make decisions based on multiple pieces of information.

Conclusion

In this study, we proposed a relation extraction model for extracting knowledge about soybean pests and disease from natural language text. To capture the rich semantics hidden in text, word embeddings and position embeddings were used to encode word semantics and sentence-level information. We experimented the BERT model and two CNN-based models. Experimental results showed that after fine-tuning the hyper-parameters like learning rates, the BERT model obtained significant accuracy in identifying these relations, with an overall F1 score of 98.49%. The extracted knowledge was organized into a knowledge graph to provide succinct and efficient access for end users in the fields. At this stage, reasoning on the knowledge graph is mainly performed manually. In future, we plan to develop the automatic reasoning mechanism on the knowledge graph. Meanwhile, an image-based query interface will further tighten the connection between the research literature and the fields. Field images will be taken and analyzed, upon which decisions for further treatments can be made using the knowledge graph.

Acknowledgments

This work was funded by the National Natural Science Foundation of China (31560396, 31860385) and the Natural Science Foundation of Yunnan Province (2018FB061). We would like to thank the anonymous reviewers for their comments that greatly improved the manuscript.

References

Abdalla, A., Cen, H., El-manawy, A., & He, Y. (2019). Infield oilseed rape images segmentation via improved unsupervised learning models combined with supreme color features. Comput. Electron. Agric., 162, 1057-1068. https://doi.org/10.1016/j.compag.2019.05.051

Ahmad, A., Saraswat, D., Aggarwal, V., Etienne, A., & Hancock, B. (2021). Performance of deep learning models for classifying and detecting common weeds in corn and soybean production systems. Comput. Electron. Agric., 184, 106081. https://doi.org/10.1016/j.compag.2021.106081

Bottou, L. (2010). Large-scale machine learning with stochastic gradient descent. Proc. 19th Int. Conf. on Computational Statistics (pp. 177-186). Physica-Verlag. https://doi.org/10.1007/978-3-7908-2604-3_16

Brin, S. (1998). Extracting patterns and relations from the World Wide Web. Proc. International Workshop on the World Wide Web and Databases (pp. 172-183). Berlin: Springer. https://doi.org/10.1007/10704656_11

Dai, F. G., & Cai, H. J. (2015). Research progress and development trend of agricultural water saving based on CiteSpace [in Chinese]. J. Water Resour. Water Eng., 26(1), 212-218. https://doi.org/10.11705/j.issn.1672-643X.2015.01.042

Devlin, J., Chang, M., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. Proc. 2019 Conf. of the North Am. Chapter of the Assoc. for Computational Linguistics: Human Language Technol. (pp.4171-4186). Stroudsburg, PA: Assoc. for Computational Linguistics.

https://doi.org/10.18653/v1/N19-1423

Espejo-Garcia, B., Lopez-Pellicer, F. J., Lacasta, J., Moreno, R. P., & Zarazaga-Soria, F. J. (2019). End-to-end sequence labeling via deep learning for automatic extraction of agricultural regulations. Comput. Electron. Agric., 162, 106-111. https://doi.org/10.1016/j.compag.2019.03.027

Espejo-Garcia, B., Martinez-Guanter, J., Pérez-Ruiz, M., Lopez-Pellicer, F. J., & Zarazaga-Soria, F. J. (2018). Machine learning for automatic rule classification of agricultural regulations: A case study in Spain. Comput. Electron. Agric., 150, 343-352. https://doi.org/10.1016/j.compag.2018.05.007

Gao, J., Li, X., Zhu, F., & He, Y. (2013). Application of hyperspectral imaging technology to discriminate different geographical origins of Jatropha curcas L. seeds. Comput. Electron. Agric., 99, 186-193. https://doi.org/10.1016/j.compag.2013.09.011

Guo, X. Y., & He, T. T. (2015). Survey about research on information extraction [in Chinese]. Comput. Sci., 42(2), 14-17. https://doi.org/10.11896/j.issn.1002-137X.2015.02.003

Hasegawa, T., Sekine, S., & Grishman, R. (2004). Discovering relations among named entities from large corpora. Proc. 42nd Annu. Meeting on Assoc. for Computational Linguistics. Stroudsburg, PA: Assoc. for Computational Linguistics. https://doi.org/10.3115/1218955.1219008

Holzschuher, F., & Peinl, R. (2013). Performance of graph query languages: Comparison of cypher, gremlin and native access in Neo4j. Proc. Joint EDBT/ICDT 2013 Workshops (pp. 195–204). New York, NY: Association for Computing Machinery. https://doi.org/10.1145/2457317.2457351

Jin, N., Zhao, C. J., Wu, H. R., Li, S., & Yang, B. Z. (2020). Classification technology of agricultural questions based on BiGRU_MulCNN [in Chinese]. J. Agric. Mach.51(5), 199-206.

http://10.6041/j.issn.1000-1298.2020.05.022

Kambhatla, N. (2004). Combining lexical, syntactic, and semantic features with maximum entropy models for extracting relations. Proc. ACL 2004 on Interactive Poster and Demonstration Sessions (pp. 22–es). Stroudsburg, PA: Assoc. for Computational Linguistics. https://doi.org/10.3115/1219044.1219066

Karlekar, A., & Seal, A. (2020). SoyNet: Soybean leaf diseases classification. Comput. Electron. Agric., 172, 105342. https://doi.org/10.1016/j.compag.2020.105342

Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. https://doi.org/10.48550/arXiv.1412.6980

Kui, G. X., & Qi, C. J. (2021). Study on the resource effect and environmental effect of China’s import trade of major grain [in Chinese]. Wld. Agric. 505(5), 16-25+126. https://doi.org/ 10.13856/j.cn11-1097/s.2021.05.002

Li, Q. Z., & Cai, J. X. (2017). The colored atlas of primary soybean pests and disease [in Chinese]. Zhengzhou, Henan Province, China: Henan Science and Technology Press.

Li, X., Ma, D., & Yin, B. (2021). Advance research in agricultural text-to-speech: The word segmentation of analytic language and the deep learning-based end-to-end system. Comput. Electron. Agric., 180, 105908. https://doi.org/10.1016/j.compag.2020.105908

Li, Y. Z., & Zhu, H. M. (2020). Study on Cytoscape.js based web visualization of Neo4j pesticide knowledge graph [in Chinese]. Softw., 41(4), 10-13.

https://doi.org/10.3969/j.issn.1003-6970.2020.04.003

Liu, Z. Y., Han, X., & Sun, M. S. (2020). Knowledge graph and deep learning [in Chinese]. Beijing, China: Tsinghua University Press.

Loshchilov, I., & Hutter, F. (2017). Decoupled weight decay regularization. Proc. Int. Conf. on Learning Representations.

Nguyen, T. H., & Grishman, R. (2015). Relation extraction: Perspective from convolutional neural networks. Proc. 1st Workshop on Vector Space Modeling for Natural Language Processing (pp. 39-48). Stroudsburg, PA: Assoc. for Computational Linguistics.

Pennington, J., Socher, R., & Manning, C. (2014). GloVe: Global vectors for word representation. Proc. 2014 Conf. on Empirical Methods in Natural Language Processing (EMNLP) (pp. 1532–1543). Stroudsburg, PA: Assoc. for Computational Linguistics. https://doi.org/10.3115/v1/D14-1162

Qin, C. L. (2009). The study of construction knowledge domains map on Chinese agricultural history subject by co-occurrence analysis on scientometrics [in Chinese]. PhD diss. Nanjing, China: Nanjing Agriculture University.

Ren, G., Lin, T., Ying, Y., Chowdhary, G., & Ting, K. C. (2020). Agricultural robotics research applicable to poultry production: A review. Comput. Electron. Agric., 169, 105216. https://doi.org/10.1016/j.compag.2020.105216

Ren, N., Sun, Y. W., & Bao, T. (2021). Research on ontology construction method in agriculture: Case study of tomato disease and insect pests [in Chinese]. Info. Res., 282(4), 51-57.

http://10.3969/j.issn.1005-8095.2021.04.008

Schlichtkrull, M., Kipf, T. N., Bloem, P., van den Berg, R., Titov, I., & Welling, M. (2017). Modeling relational data with graph convolutional networks. Proc. Euro. Semantic Web Conf. (pp. 593-607). Berlin: Springer.

http://10.1007/978-3-319-93417-4_38

Socher, R., Huval, B., Manning, C. D., & Ng, A. (2012). Semantic compositionality through recursive matrix-vector spaces. Proc. 2012 Joint Conf. on Empirical Methods in Natural Language Processing and Comput. Natural Language Learning. Stroudsburg, PA: Assoc. for Computational Linguistics.

Tetila, E. C., Machado, B. B., Astolfi, G., Belete, N. A., Amorim, W. P., Roel, A. R., & Pistori, H. (2020). Detection and classification of soybean pests using deep learning with UAV images. Comput. Electron. Agric., 179, 105836. https://doi.org/10.1016/j.compag.2020.105836

Wan, C., Liu, T., & Li, S. (2005). Automatic entity relation extraction. J. Chinese Info. Process..2, 1-6.

Wang, H. R., Wang, X. M., Miao, Y. S., Xv, T. Y., Liu, Z. C., & Wu, H. R. (2022). Densely connected BiGRU neural network based on BERT and attention mechanism for Chinese agriculture-related question similarity matching [in Chinese]. J. Agric. Mach., 53(1), 244-252.

http://10.6041/j.issn.1000-1298.2022.01.027

Wang, H. R., Wu, H. R., Feng, S., Liu, Z. C., & Xu, T. Y. (2021). Classification technology of rice questions in question answer system based on Attention_DenseCNN [in Chinese]. J. Agric. Mach., 52(7), 237-243.

http://10.6041/j.issn.1000-1298.2021.07.025

Wang, M. (2008). A re-examination of dependency path kernels for relation extraction. Proc. Third International Joint Conference on Natural Language Processing: Volume-II. Stroudsburg, PA: Assoc. for Computational Linguistics.

Wu, S. S., Zhou, A. L., Xie, N. F., Liang, X. H., Wang, H. J., Li, X. Y., & Cheng, G. P. (2020). Construction of visualization domain-specific knowledge graph of crop disease and pests based on deep learning [in Chinese]. Trans. Chinese Soc. Agric. Eng., 36(24), 177-185.

https://doi.org/10.11975/j.issn.1002-6819.2020.24.021

Xu, Y., Mou, L., Li, G., Chen, Y., Peng, H., & Jin, Z. (n.d.). Classifying relations via long short term memory networks along shortest dependency paths. Proc. 2015 Conf. on Empirical Methods in Natural Language Processing (pp. 1785-1794). Stroudsburg, PA: Assoc. for Computational Linguistics. https://doi.org/10.18653/v1/D15-1206

Yang, G. F., & Yang, Y. (2020). Question classification of common crop diseases question answering system based on BERT [in Chinese]. J. Comput. Appl., 40(6), 1580-1586. https://doi.org/10.11772/j.issn.1001-9081.2019111951

Yang, M. M., & Zhou, Q. H. (2013). Knowledge graph analysis of agricultural e-commerce research at home and abroad [in Chinese]. Log. Eng. Manag. 10, 148-152.

https://doi.org/10.3969/j.issn.1674-4993.2013.10.050

Yuan, S. P., Li, R. L., Wang, C., & Xu, H. L. (2021). Relationship extraction from rice phenotype knowledge graph based on BERT [in Chinese]. Trans. CSAM, 52(5), 151-158. https://doi.org/10.6041/j.issn.1000-1298.2021.05.016

Zelenko, D., Aone, C., & Richardella, A. (2003). Kernel methods for relation extraction. J. Mach. Learn. Res., 3, 71–78.

Zeng, D., Liu, K., Chen, Y., & Zhao, J. (2015). Distant supervision for relation extraction via piecewise convolutional neural networks. Proc. 2015 Conf. on Empirical Methods in Natural Language Processing (pp. 1753-1762). Stroudsburg, PA: Assoc. for Computational Linguistics.

https://doi.org/10.18653/v1/D15-1203

Zhang, K., Wu, Q., & Chen, Y. (2021). Detecting soybean leaf disease from synthetic image using multi-feature fusion faster R-CNN. Comput. Electron. Agric., 183, 106064. https://doi.org/10.1016/j.compag.2021.106064

Zhao, G., Quan, L., Li, H., Feng, H., Li, S., Zhang, S., & Liu, R. (2021). Real-time recognition system of soybean seed full-surface defects based on deep learning. Comput. Electron. Agric., 187, 106230. https://doi.org/10.1016/j.compag.2021.106230

Zhou, Z., Song, Z., Fu, L., Gao, F., Li, R., & Cui, Y. (2020). Real-time kiwifruit detection in orchard using deep learning on Android™ smartphones for yield estimation. Comput. Electron. Agric., 179, 105856. https://doi.org/10.1016/j.compag.2020.105856