Soybean planting field entity relationship extraction method and system based on knowledge distillation

Abstract

The invention discloses a soybean planting field entity relationship extraction method and system based on knowledge distillation, and the method comprises the steps: obtaining original text data of a soybean planting field, carrying out the cleaning and phrasing processing, and constructing an initial corpus; making an entity and relationship type labeling specification, performing manual labeling on the initial corpus, and constructing an entity relationship extraction data set in the soybean planting field; the method comprises the following steps: extracting a data set based on a data set entity relationship, and constructing and training an entity relationship joint extraction model GlobalRel based on table filling; a GlobalRel model is used as a teacher model, an LSTM-GlobalRel model is used as a student model, training is carried out through a knowledge distillation mechanism, and a lightweight Tiny-GlobalRel model is obtained; a Tiny-GlobalRel model is used for carrying out triple extraction of an entity relationship on an input text.

Description

Technical Field

[0001] This invention relates to the field of soybean cultivation, and more specifically to a method and system for extracting entity relationships in the field of soybean cultivation based on knowledge distillation. Background Technology

[0002] In the field of natural language processing, entity relation extraction is one of the key technologies for constructing knowledge graphs and supporting intelligent decision-making, especially in vertical fields such as agriculture. However, texts in the soybean cultivation field are characterized by dense technical terms, widespread entity nesting, and complex overlapping relationships. Traditional rule-based or statistical methods struggle to effectively capture semantic connections, while existing deep learning models, although high-performance, often rely on large amounts of computing resources and labeled data, making deployment in real-world agricultural environments difficult. Furthermore, agricultural texts often have loose sentence structures and diverse expressions, lacking unified annotation standards and high-quality datasets, further limiting the promotion and application of entity relation extraction models in this field. Therefore, there is an urgent need for an entity relation extraction method that maintains both high accuracy and lightweight characteristics to meet the practical needs of edge computing environments. Summary of the Invention

[0003] This invention provides a method for entity relation extraction in the soybean planting field based on knowledge distillation, comprising the following steps:

[0004] We acquired raw text data related to soybean cultivation, cleaned and segmented it to build an initial corpus.

[0005] Establish annotation standards for entity and relation types, manually annotate the initial corpus, and construct an entity relation extraction dataset in the soybean planting field;

[0006] Based on the entity relationship extraction dataset, a joint entity relationship extraction model GlobalRel based on table filling is constructed and trained;

[0007] Using the GlobalRel model as the teacher model and the LSTM-GlobalRel model as the student model, a lightweight Tiny-GlobalRel model is obtained through knowledge distillation.

[0008] The Tiny-GlobalRel model is used to extract entity relation triples from the input text.

[0009] Preferably, the method for constructing the entity relationship extraction dataset includes:

[0010] Retrieves soybean planting-related text from structured, semi-structured, and unstructured data sources;

[0011] The texts related to soybean planting were processed by denoising, deduplication, and sentence segmentation to form a sentence-level corpus;

[0012] Based on the preset entity and relation types, the annotation tool is used to annotate the sentences with entities and relations, and the annotation results are obtained;

[0013] Convert the annotation results into a structured data format that includes text and a list of triples.

[0014] Preferably, the GlobalRel model includes: a word fusion encoder, a head-and-tail entity interaction module, a global feature mining module, and a decoding module;

[0015] The word fusion encoder is used to fuse character-level and vocabulary-level features;

[0016] The head and tail entity interaction module is used to capture the association information between the head entity and the tail entity;

[0017] The global feature mining module is used to extract global dependency information from sentences;

[0018] The decoding module is used to decode triples based on the global entity table and the head and tail entity tables.

[0019] Preferably, the construction process of the word-word fusion encoder includes: directly introducing lexical information into the BERT pre-trained model, fusing lexical-level semantics on the basis of character representation, and obtaining the word-word fusion-based BERT encoder CWFE.

[0020] Preferably, the knowledge distillation process includes:

[0021] Use the trained GlobalRel model to generate soft labels for the input sentence;

[0022] The soft labels were used to guide the training of the student model LSTM-GlobalRel;

[0023] Distillation is achieved by optimizing the parameters of the student model LSTM-GlobalRel through a weighted sum of soft and hard losses.

[0024] The present invention also provides an entity relation extraction system for soybean planting based on knowledge distillation. The system is used to implement the above method and includes: a collection module, a labeling module, a construction module, a training module, and an extraction module.

[0025] The acquisition module is used to acquire raw text data in the soybean planting field, and to perform cleaning and sentence segmentation to build an initial corpus.

[0026] The annotation module is used to formulate annotation specifications for entity and relation types, manually annotate the initial corpus, and construct an entity relation extraction dataset in the soybean planting field.

[0027] The building module is used to construct and train the entity relationship joint extraction model GlobalRel based on table filling based on the entity relationship extraction dataset;

[0028] The training module is used to train the lightweight Tiny-GlobalRel model by using the GlobalRel model as the teacher model and the LSTM-GlobalRel model as the student model through a knowledge distillation mechanism.

[0029] The extraction module is used to extract entity relation triples from the input text using the Tiny-GlobalRel model.

[0030] Preferably, the workflow of the acquisition module includes:

[0031] Retrieves soybean planting-related text from structured, semi-structured, and unstructured data sources;

[0032] The texts related to soybean planting were processed by denoising, deduplication, and sentence segmentation to form a sentence-level corpus;

[0033] Based on the preset entity and relation types, the annotation tool is used to annotate the sentences with entities and relations, and the annotation results are obtained;

[0034] Convert the annotation results into a structured data format that includes text and a list of triples.

[0035] Preferably, the GlobalRel model includes: a word fusion encoder, a head-and-tail entity interaction module, a global feature mining module, and a decoding module;

[0036] The word fusion encoder is used to fuse character-level and vocabulary-level features;

[0037] The head and tail entity interaction module is used to capture the association information between the head entity and the tail entity;

[0038] The global feature mining module is used to extract global dependency information from sentences;

[0039] The decoding module is used to decode triples based on the global entity table and the head and tail entity tables.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] This invention can significantly improve the accuracy and robustness of entity relation extraction in the soybean planting field, effectively overcome the challenges of entity nesting, relation overlap and complex professional terminology in agricultural texts, and at the same time greatly reduce the model's dependence on computing resources and memory, enabling the high-performance entity relation extraction model to run stably in edge environments such as embedded devices and mobile terminals with limited computing power, and promote the widespread application of agricultural knowledge graph construction and intelligent decision-making systems in actual production. Attached Figure Description

[0042] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart of the distillation model according to an embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram of entities and entity relationships in an embodiment of the present invention;

[0045] Figure 3 This is an example of a dataset for extracting entity relationships in the soybean planting field according to an embodiment of the present invention.

[0046] Figure 4 This is an embodiment of the present invention of a joint entity relationship extraction model based on table filling in the soybean planting field;

[0047] Figure 5 This is an example of a table filling and decoding strategy according to an embodiment of the present invention; wherein, (a) is a global entity table; and (b) is a head and tail entity table;

[0048] Figure 6 This is a schematic diagram of a word fusion encoder according to an embodiment of the present invention;

[0049] Figure 7 This is a schematic diagram of the head-and-tail entity interaction and global feature mining in an embodiment of the present invention;

[0050] Figure 8 This is a schematic diagram of the knowledge distillation process according to an embodiment of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] Example 1

[0054] This embodiment provides a method for entity relation extraction in the soybean planting field based on knowledge distillation. The overall distillation process is as follows: Figure 1 As shown, it includes the following steps:

[0055] S1. Obtain raw text data in the soybean planting field, clean and segment it to build an initial corpus.

[0056] To meet the research needs of entity relation extraction in the soybean planting field, this embodiment constructs an entity relation extraction dataset for the soybean planting scenario and systematically analyzes and summarizes its corpus features.

[0057] The corpus used in this embodiment mainly comes from online resources and agricultural professional literature. To ensure the authority and accuracy of the data, the online data portion selected highly credible content platforms such as Baidu Encyclopedia, Wikipedia, the China Agricultural Encyclopedia Media Knowledge Service Platform, and the National Agricultural Science Data Center as the main information sources. These platforms cover the core knowledge content of each stage of soybean cultivation, providing a solid foundation for data collection.

[0058] After acquiring the raw text, a comprehensive cleaning process is first performed to improve the quality of the corpus. On the one hand, noisy characters and irrelevant symbols are removed, such as HTML tags, special control characters, and abnormally encoded symbols. On the other hand, information unrelated to soybean cultivation, such as advertising slogans, navigation content, and repetitive paragraphs, is eliminated to ensure that the corpus focuses on core domain knowledge.

[0059] To improve the efficiency and standardization of subsequent data annotation, long text segments need to be segmented into semantically complete sentences. Sentence segmentation is performed using regular expressions based on common punctuation marks such as periods, question marks, and semicolons, ensuring the semantic coherence and completeness of the segmentation results.

[0060] After deduplication, noise reduction, formatting and sentence segmentation, a total of 15,126 sentences were finally obtained, with a total word count of approximately 300,000 words.

[0061] S2. Develop annotation standards for entity and relation types, manually annotate the initial corpus, and construct an entity relation extraction dataset for the soybean planting field.

[0062] Before conducting entity relation extraction research in the soybean planting field, it is necessary to first determine the entity types and relation categories involved in the soybean data text. However, due to the generally loose sentence structure and diverse expressions in agricultural texts, there is currently a lack of unified and widely accepted annotation standards in the agricultural field. Therefore, this invention takes the soybean planting field as the research object, and under the guidance of agricultural experts, combines agricultural professional books and authoritative agricultural knowledge websites to summarize and analyze relevant corpora. Based on domain knowledge, eight core entities involved in the soybean planting process are summarized: crops, parts of plants, pests and diseases, pesticides, fertilizers, natural environment, agronomic measures, and symptoms. Their detailed classification is shown in Table 1.

[0063] Table 1

[0064]

[0065] The relationship between entities refers to the interaction and influence between agricultural entities. Figure 2 This invention demonstrates typical relationships between various entities. Combining agricultural knowledge systems and expert advice, it designs and standardizes 16 types of relationships between entities, specifically including: harmful relationships, inclusion relationships, adaptive relationships, prevention and control relationships, promoting relationships, inducing relationships, mitigating relationships, causing relationships, resistance relationships, inhibiting relationships, characterizing relationships, manifesting relationships, belonging relationships, acting upon relationships, appearing in relationships, and improving relationships.

[0066] This embodiment uses LabelStudio as the annotation tool. This open-source platform supports annotation of various data formats, including text, images, audio, and video, and is particularly suitable for natural language processing tasks such as named entity recognition and relation extraction, offering high flexibility and scalability. To ensure the accuracy and consistency of the annotation results, this invention divides the data annotation process into three stages: annotation specification design, initial annotation, and expert review.

[0067] The original JSON annotation results were structured and converted to a format similar to that of existing public datasets, referencing standard formats. Figure 3 The normalized structure is shown. The transformed data format consists of the original text and a list of triples. The list of triples includes the relation type, head entity, tail entity, and the index range of the head and tail entities in the original sentence, with the index starting from 0.

[0068] The unlabeled dataset after sentence segmentation was divided into training, validation, and test sets in an 8:1:1 ratio. Then, through the above three stages of annotation, 20,652 entities and 20,026 relations were finally obtained.

[0069] S3. Extract entity relationships from the dataset, construct and train the GlobalRel joint entity relationship extraction model based on table filling.

[0070] like Figure 4 As shown, the GlobalRel table-filling entity relation joint extraction model proposed in this embodiment, based on global features and word fusion, consists of an encoding layer, a head-and-tail entity interaction layer, a global feature mining layer, and a decoding layer. The encoding layer embeds the word fusion module into the BERT pre-trained language model to obtain a fused vector representation of character and word vectors. The head-and-tail entity interaction layer interacts with the feature vectors obtained from the encoding layer to capture the relationship between the subject and object. The global feature mining module considers all relations simultaneously, fully capturing the global dependency information within the table. The decoding layer jointly decodes the global entity table and the head-and-tail entity table according to a set decoding strategy, extracting all triples in the sentence.

[0071] This embodiment proposes a novel table annotation strategy, which designs two sub-tasks for entity recognition and relation extraction: a global entity recognition task and a head-and-tail entity recognition task. The former is used to annotate the boundary information of all entities in a sentence, while the latter annotates the specific positions of head-and-tail entity pairs for each relation type.

[0072] The global entity task constructs an n×n two-dimensional table for each sentence to indicate whether an entity exists. For each character position (x... i x j If from x i To x j If a cell constitutes a complete entity, it is marked as "1"; otherwise, it is marked as "0".

[0073] In the head-and-tail entity task, assuming the number of relation categories in the dataset is R, a separate n×n two-dimensional table needs to be constructed for each relation to identify the positional boundaries of the head-and-tail entity pairs under the current relation. This invention defines a label set L = {"N / A", "B", "E", "S"}, where "B" represents (x... i x j ) represents the starting position of the head and tail entities; "E" indicates (x i x j The top-left and bottom-right corners of a rectangular area in the table mark the end positions of the header and tail entities. This means the header and tail entities in the current relationship can be determined using the coordinates of their top-left and bottom-right corners. If both the header and tail entities consist of a single character, the rectangular area degenerates into a single point, and is marked with "S". Cells not falling into this category are marked with "N / A".

[0074] Taking the sentence "Aphids feed on soybean leaves, causing the leaves to turn yellow" as an example, the global entity table and the head and tail entity tables in this example are as follows: Figure 5 As shown, this contains four entities: "aphid," "soybean leaf," "soybean," and "yellowing leaves." A global entity table is constructed based on the start and end positions of each entity in the sentence. For example, "aphid" starts at position 1 and ends at position 2 in the sentence, so the position in the first row and second column of the global entity table is marked as 1; "soybean leaf" starts at position 5 and ends at position 8 in the sentence, so the position at coordinates (5,8) in the global entity table is marked as 1, and so on for the other entities. The marking results are as follows. Figure 5 As shown in (a), the portion marked "0" is omitted from the diagram. In the example above, these four entities form five triples: (aphid, damage, soybean leaves), (aphid, damage, soybean), (aphid, cause, leaves turn yellow), (soybean, contain, soybean leaves), and (soybean leaves, belong to, soybean). According to the head-and-tail entity table filling strategy, the start position of the head-and-tail entities is marked as "B", and the end position is marked as "E".

[0075] In Chinese entity relation extraction tasks, entity boundaries typically correspond to lexical boundaries. However, the BERT model uses characters as the smallest processing unit during the encoding stage, generating character-level embeddings with limited semantic expressive power. Traditional approaches often rely on Chinese word segmentation tools to segment the input text, introducing lexical-level semantic information. However, word segmentation has inherent errors, which can severely impact the model's accuracy in recognizing entity boundaries. To more effectively integrate lexical features, this invention directly incorporates lexical information into the BERT pre-trained model, fusing lexical-level semantics on top of character representations, resulting in the character-word fusion-based BERT encoder CWFE (Character-Word FusionEncoder). The overall structure of this module is as follows: Figure 6 As shown.

[0076] A multi-head attention mechanism is introduced to model the bidirectional interaction between the feature vectors of the head and tail entities. This approach allows the two feature vectors to interact, fully extracting and utilizing the potential correlation information between the head and tail entities, helping the model to more accurately and quickly determine the specific relationship type between entity pairs.

[0077] To address issues such as ambiguous entity boundaries, nested entities, overlapping relationships, and insufficient entity interaction and a focus on only local information while neglecting global information in current table-filling relationship extraction, this invention constructs a word vector library for the soybean planting field and introduces a word fusion module to encode the input vectors to solve the problem of ambiguous entity boundaries; it uses a head-tail entity interaction module to extract as much relevant information as possible between head and tail entities to help the model determine entity types more quickly; and it employs a global feature mining module (e.g., based on the Transformer multi-head attention concept) to address these problems. Figure 7 As shown, after multiple iterations of optimization of the interaction information between the head and tail entities and the global feature inference information, potential relation triples in the sentence are mined; and a set of table annotation methods and decoding strategies are designed for the table filling model, which solve the problems of entity nesting and relation overlap through the global entity table and the head and tail entity tables.

[0078] S4. Using the GlobalRel model as the teacher model and the LSTM-GlobalRel model as the student model, a lightweight Tiny-GlobalRel model is obtained through knowledge distillation.

[0079] Traditional deep learning models typically rely on a large number of parameters and expensive computing resources to achieve excellent performance. However, in real-world applications with limited computing resources, such as embedded devices, mobile terminals, and wearable devices, these models are difficult to deploy efficiently. Knowledge distillation is a model compression technique whose basic framework usually includes two types of models: a teacher model, which is a complex, high-performance model with high computational cost and a large number of parameters; and a student model, which has a relatively simple structure and fewer parameters, making it suitable for deployment in resource-constrained environments. The core idea of ​​this method is to use the knowledge learned by the teacher model to guide the training of the student model through the output prediction distribution (i.e., "soft labels"), thereby significantly reducing model complexity while maintaining the original performance level as much as possible. In this way, the student model can approach the expressive power of the teacher model with a smaller number of model parameters, achieving a balance between lightweight design and high performance.

[0080] Knowledge distillation process such as Figure 8As shown, the process begins by training a GlobalRel model using the dataset to obtain a teacher model. The model parameters that perform best on the validation set are then selected as the fixed teacher model in the distillation process. During the decoding phase, the decoding module outputs a corresponding score vector for each coordinate position (e.g., (i, j) in the global entity table and (i, j, r) in the head and tail entity tables). Softmax is used to heat this vector, smoothing the output class probabilities to obtain soft predictions with richer information. The student model uses a BiLSTM-based LSTM-GlobalRel model, replacing the BERT-based word fusion encoder in the GlobalRel model with a BiLSTM network. The same dataset is then used to train the LSTM-GlobalRel student model, which is then heated using Softmax to obtain soft predictions. A standard Softmax with T=1 is then applied to obtain hard predictions. Finally, the soft predictions of the teacher model and student model are cross-entropy to obtain a soft loss. In the student model branch, T=1 is set to obtain hard predictions. Then, the loss for hard labeling and soft prediction is calculated, which is called hard loss. After obtaining the soft loss and hard loss, they can be weighted and summed to obtain the total loss.

[0081] In the above process, the GlobalRel model transfers its learned knowledge to the LSTM-GlobalRel model through knowledge distillation, enabling the GlobalRel model to pass on its predictive and generalization capabilities to the LSTM-GlobalRel model. This process aims to provide the LSTM-GlobalRel model with a more efficient and accurate learning path, allowing it to fully absorb the knowledge from the GlobalRel model during learning. The final result is the lightweight, high-performance Tiny-GlobalRel model.

[0082] S5. Use the Tiny-GlobalRel model to extract entity relation triples from the input text.

[0083] Example 2

[0084] This embodiment also provides a knowledge distillation-based entity relation extraction system for soybean cultivation, including: a collection module, an annotation module, a construction module, a training module, and an extraction module. The collection module is used to acquire raw text data in the soybean cultivation field, and perform cleaning and sentence segmentation to construct an initial corpus. The annotation module is used to formulate annotation specifications for entity and relation types, manually annotate the initial corpus, and construct an entity relation extraction dataset for the soybean cultivation field. The construction module is used to construct and train a table-filling-based joint entity relation extraction model, GlobalRel, based on the entity relation extraction dataset. The training module is used to train a lightweight Tiny-GlobalRel model by using the GlobalRel model as the teacher model and the LSTM-GlobalRel model as the student model through a knowledge distillation mechanism. The extraction module is used to extract triples of entity relations from the input text using the Tiny-GlobalRel model.

[0085] The workflow of the data acquisition module includes: acquiring soybean planting-related text from structured, semi-structured, and unstructured data sources; performing noise reduction, deduplication, and sentence segmentation on the soybean planting-related text to form sentence-level corpus; using annotation tools to annotate the sentences with entities and relations based on preset entity types and relation types to obtain annotation results; and converting the annotation results into a structured data format containing text and a list of triples.

[0086] The GlobalRel model comprises: a word-level fusion encoder, a head-and-tail entity interaction module, a global feature mining module, and a decoding module. The word-level fusion encoder fuses character-level and word-level features; the head-and-tail entity interaction module captures the association information between head and tail entities; the global feature mining module extracts global dependency information from the sentence; and the decoding module performs triple decoding based on the global entity table and the head-and-tail entity table.

[0087] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A soybean planting field entity relation extraction method based on knowledge distillation, characterized by, The method comprises the following steps: obtaining original text data in the soybean planting field, and performing cleaning and sentence processing to construct an initial corpus; developing annotation specifications for entity and relationship types, and manually annotating the initial corpus to construct an entity relationship extraction dataset in the soybean planting field; based on the entity relationship extraction dataset, constructing and training an entity relationship joint extraction model GlobalRel based on table filling; using the GlobalRel model as a teacher model and an LSTM-GlobalRel model as a student model, training through a knowledge distillation mechanism to obtain a lightweight Tiny-GlobalRel model; using the Tiny-GlobalRel model to extract triplets of entities and relationships from input text.

2. The knowledge distillation-based soybean planting field entity relation extraction method according to claim 1, wherein, The method for constructing the entity relationship extraction dataset comprises: obtaining soybean planting related texts from structured, semi-structured and unstructured data sources; performing denoising, deduplication and sentence processing on the soybean planting related texts to form a sentence-level corpus; according to preset entity types and relationship types, using an annotation tool to annotate entities and relationships in the sentences to obtain annotation results; converting the annotation results into a structured data format containing text and triplet lists.

3. The knowledge distillation-based soybean planting field entity relation extraction method according to claim 1, characterized in that, The GlobalRel model comprises a word fusion encoder, a head-tail entity interaction module, a global feature mining module and a decoding module; the word fusion encoder is used to fuse character-level and word-level features; the head-tail entity interaction module is used to capture the association information between head entities and tail entities; the global feature mining module is used to extract global dependency information in the sentence; the decoding module is used to decode triplets according to the global entity table and the head-tail entity table.

4. The knowledge distillation-based entity relation extraction method in the soybean planting field according to claim 3, characterized in that, The construction process of the word fusion encoder comprises: introducing word information directly into the BERT pre-training model, fusing word-level semantics on the basis of character representation, and obtaining a BERT encoder CWFE based on word fusion.

5. The knowledge distillation-based entity relation extraction method in the soybean planting field according to claim 1, wherein, The process of knowledge distillation comprises: using the trained GlobalRel model to generate soft labels for input sentences; using the soft labels to guide the training of the student model LSTM-GlobalRel; optimizing the parameters of the student model LSTM-GlobalRel through the weighted sum of soft loss and hard loss to complete distillation.

6. A soybean planting field entity relation extraction system based on knowledge distillation, the system is used to implement the method of any one of claims 1-5, characterized in that, comprises: a collection module, an annotation module, a construction module, a training module and an extraction module; the collection module is used to obtain original text data in the soybean planting field, and perform cleaning and sentence processing to construct an initial corpus; the annotation module is used to develop annotation specifications for entity and relationship types, and manually annotate the initial corpus to construct an entity relationship extraction dataset in the soybean planting field; the construction module is used to construct and train an entity relationship joint extraction model GlobalRel based on table filling based on the entity relationship extraction dataset; The training module is configured to use the GlobalRel model as a teacher model, use an LSTM-GlobalRel model as a student model, and train the Tiny-GlobalRel model through a knowledge distillation mechanism. The extraction module is configured to use the Tiny-GlobalRel model to perform triple extraction of entity relationships on input text.

7. The knowledge distillation based soybean planting domain entity relation extraction system according to claim 6, wherein, The workflow of the collection module includes: Obtaining soybean planting related texts from structured, semi-structured and unstructured data sources; Performing denoising, deduplication and sentence segmentation on the soybean planting related texts to form a sentence-level corpus; According to the preset entity type and relationship type, using a labeling tool to label entities and relationships of the sentences to obtain labeling results; Converting the labeling results into a structured data format containing text and a triple list.

8. The knowledge distillation based soybean planting domain entity relation extraction system according to claim 6, wherein, The GlobalRel model includes a word fusion encoder, a head-tail entity interaction module, a global feature mining module and a decoding module. The word fusion encoder is configured to fuse character-level and word-level features. The head-tail entity interaction module is configured to capture the association information between the head entity and the tail entity. The global feature mining module is configured to extract global dependency information in the sentence. The decoding module is configured to perform triple decoding according to a global entity table and a head-tail entity table.

Images