CN112100405B - Veterinary drug residue knowledge graph construction method based on weighted LDA - Google Patents

Abstract

The invention discloses a veterinary drug residue knowledge graph construction method based on weighted LDA (Latent Dirichlet Allocation ). Firstly, constructing a veterinary drug knowledge framework, and carrying out deep search by combining a web crawler with the knowledge framework and downloading documents. Aiming at the topic noise and the bias problem of the feature words existing in the LDA topic model, a weighted LDA method is used for topic mining, and veterinary drug related documents are downloaded again. Named entity recognition and relationship extraction are accomplished using a dictionary-based model. And finally, constructing a veterinary drug knowledge graph by utilizing a Neo4j graph database. The invention can be used for constructing a veterinary drug residue knowledge graph, finding out the characteristic rule of veterinary drug residues and the reason of injury of veterinary drug residues to human bodies, and ensuring the quality safety of meat, eggs and milk, thereby protecting the physical health and life safety of people.

Description

A method for constructing veterinary drug residue knowledge graph based on weighted LDA

Technical field

The present invention relates to the field of natural language processing, and in particular to a method for constructing a knowledge graph of veterinary drug residues based on weighted LDA.

Background technique

Food safety issues have attracted more and more attention, among which the safety of meat, eggs, and dairy products is a top priority. Veterinary drugs play an important role in preventing and treating animal diseases and promoting animal growth. The breeding process of livestock products is inseparable from veterinary drugs. However, irregular, illegal use and abuse of veterinary drugs have led to excessive veterinary drug residues, leading to poisoning incidents. By constructing a knowledge map of veterinary drug residues, we can find out the characteristics and patterns of veterinary drug residues and the reasons why veterinary drug residues cause harm to the human body, ensuring the quality and safety of meat, egg and milk products, thereby protecting people's health and life safety.

Veterinary drug residue data involves the residue standards of veterinary drugs, sampling inspections of veterinary drugs that exceed the standards, animal toxicology experimental data of veterinary drug residues, and harmful symptoms to humans. These data include structured data and unstructured text data. Use these data to extract and classify knowledge, build a basic framework of veterinary drug residue knowledge, and then use the constructed veterinary drug knowledge framework to download relevant literature on veterinary drug knowledge.

Download the literature based on the basic framework of veterinary drug residue knowledge and obtain relevant literature on veterinary drug knowledge. LDA was used for topic mining to obtain potential information in veterinary drug literature. LDA (Latent Dirichlet Allocation, Dirichlet distribution) is an unsupervised machine learning technology that can be used to identify latent topic information in large-scale document sets or corpora. The LDA topic mining method believes that all words have the same weight. In fact, a large number of high-frequency irrelevant words do not contribute to topic mining. A weighted LDA topic mining method that combines veterinary drug knowledge level semantic similarity and TF-IDF is used to download documents.

Through data fusion, data integration, entity recognition and relationship extraction, we can then build a knowledge graph to find out the characteristics and patterns of veterinary drug residues and the reasons why veterinary drug residues cause harm to the human body, ensuring the quality and safety of meat, eggs and milk, thereby protecting people's health. Health and life safety.

Contents of the invention

The purpose of the present invention is to provide a method for constructing a knowledge map of veterinary drug residues based on weighted LDA. In order to solve the above technical problems, the main technical contents of the present invention are as follows:

A method for constructing a knowledge graph of veterinary drug residues based on weighted LDA, including the following steps:

(1) Construct a veterinary drug knowledge framework: use hierarchical analysis and rule-based methods to extract knowledge from veterinary pharmacology and veterinary toxicology books. Obtain veterinary drug toxicology-related knowledge from the Pubchem website using a wrapper-based approach. The jieba word segmentation tool is used to remove stop words, word segmentation, and part-of-speech tagging of these corpus, and finally form a dictionary to form a hierarchical veterinary drug knowledge framework;

(2) Download literature data: Use the dictionary obtained in the previous step, combined with the name of the veterinary drug, to conduct a multi-level search on the Web of science, that is, traverse every path from the root node to the leaf node, and perform a multi-layer search on all the words on each path. Search within layer results. The support vector machine (SVM) method was used to classify the obtained documents, including those related to veterinary drug knowledge and those not related to veterinary drug knowledge. For literature related to veterinary drug knowledge, the weighted LDA method is used for topic extraction;

(3) Information extraction: dictionary-based named entity recognition and relationship extraction;

(4) Build a knowledge graph: Import the entities of the above-mentioned knowledge in the field of veterinary medicine and the relationships between entities into the Neo4j database in csv format.

The construction of a veterinary drug knowledge framework in step (1) includes the following contents:

(a) Develop a knowledge structure for veterinary drug residues, which includes five parts: veterinary drug residues, introduction to toxicology, effects on organs and systems, properties and toxicity;

(b) Veterinary drug residues: including causes, effects and hazards. Hazards can be divided into three parts: harm to the human body, food and the environment;

(c) Veterinary drug attributes: category, physical and chemical properties, pharmacokinetics, effects, applications, maximum residue limits and adverse reactions;

(d) Toxicity of veterinary drugs: classification of toxic effects, commonly used parameters, special risk groups, exposure routes, preventive measures, inhalation methods, and animal experiments. The classification of toxic effects includes nature, time of occurrence, site and recovery. Commonly used parameters include acute toxicity, mutagenicity, carcinogenicity, teratogenicity, acute toxicity, etc. Animal experiments include mice, rats, rabbits, dogs, etc.;

(e) Introduction to veterinary drug toxicology: including purpose, content and methods. Methods can be divided into two parts: biological experiments and group surveys; effects on organs and systems: including eyes, skin, liver, kidneys, nervous system, blood system, immune system, gastrointestinal tract, endocrine system and respiratory system;

(f) If each part contains table content, place it under the corresponding category.

The multi-level search in step (2) includes the following steps:

(a) Selected veterinary drugs: The "National Food Safety Standard Maximum Residue Limits of Veterinary Drugs in Foods" standard stipulates 2191 residue limits and usage requirements for 267 kinds (categories) of veterinary drugs in livestock and poultry products, aquatic products, and bee products;

(b) Use Selenium and chrome driver to complete dynamic web page (Ajax) data capture. The scope of the search is all documents since the establishment of the web of science database. Considering that the amount of data in veterinary drug toxicology research is small, the search is not restricted to journals;

(c) According to the veterinary drug knowledge framework, from the root node to the leaf node. For all nodes on each path, these keywords are combined to search in multi-level results.

SVM text classification in step (2) includes the following steps:

(a) The purpose is to divide the obtained literature set into two categories, those related to veterinary drug knowledge and those not related to veterinary drug knowledge;

(b) The TF-IDF method calculates and expresses the importance of a certain keyword in the text through statistical methods. TF refers to word frequency, which indicates how often a specified term appears in a text, and IDF refers to inverse text frequency, which is a measure of the general importance of a word. The TF calculation method of entry t _i in text d _j :

Among them, n _{i, j} is the number of times the term t _i appears in the text d _j , and the denominator represents the sum of the number of times all words appear in the text d _j .

How to calculate IDF:

Among them, |D| is the total number of texts, |j: t _i ∈d _j | is the total number of texts containing the entry t _i . In order to prevent the denominator from being 0 due to the entry being not in the text, add 1 to the denominator. Finally, the TF-IDF value of the entry t _i is obtained:

tfidf _{i, j} = tf _{i, j} × idf _i

(c) First use the TF-IDF algorithm to extract the feature words of part of the abstract of the paper and generate a document vector. If you select the full text in the abstract, the vector dimension is too high, which will increase the complexity of the calculation and is not conducive to subsequent classification. According to the characteristics of literature related to veterinary drug knowledge, just select the short text in the conclusion part of the paper abstract. For this part of the text, use the TF-IDF algorithm to extract feature words and generate a document vector;

(d) Randomly select part of the data and manually label it. Set the ratio of training set to test set to 8:2;

(e) Adjust the SVM penalty parameter C and evaluate the model based on accuracy (a), precision (P), recall (R) and F1 value;

(f) Test set to verify the model;

(g) For literature data, use the trained model to obtain literature data related to the topic of veterinary drug residues.

Establishing the weighted LDA topic model in step (2) includes the following steps:

(a) LDA (latentdirichletallocation) is a 3-layer Bayesian model that describes the relationship between documents, topics, and vocabulary. Its graph model is shown in Figure 2. The meaning of each symbol in the figure: α is the hyperparameter of the Dirichlet distribution θ, β is the Dirichlet distribution The hyperparameters, θ is the polynomial distribution of "document-topic", /> is the polynomial distribution of "topic-vocabulary", z is the topic distribution of words, w is the word, K is the number of topics, M is the number of documents, and N is the number of words in a document;

(b)LDA process:

1. Randomly assign a topic number Z to each word in each document in the corpus;

2. Re-scan the corpus, use the Gibbs Sampling formula to sample each word, find its theme, and update it in the corpus;

3. Repeat step 2 until Gibbs Sampling converges;

4. The topic word co-occurrence frequency matrix of the statistical corpus, this matrix is the model of LDA.

(c) Gibbs Sampling fitting θ, the process of:

1. Scan the article and randomly assign a topic Z _j to each word w _n ;

2. Initialize Z _j to an integer between 1 and K;

3. Re-scan each article, use the LDA model to perform topic modeling on the corpus, use GibbsSampling for parameter inference to continuously iterate, and record the value of Z _j at the same time. Parameter θ, The calculation formula is as follows:

in, is the number of words of topic j in article d,/> is the number of words of all topics in article d,/> is the number of times word w appears under topic j,/> is the total number of words of topic j in article d.

(d) The LDA algorithm does not combine relevant semantic information well in the topic modeling process, which seriously affects the semantic coherence, interpretability and accuracy of text semantic representation of the topic. In view of the vocabulary distribution characteristics of veterinary drug knowledge, this paper uses a hierarchical semantic similarity calculation formula to calculate the similarity based on the semantic similarity between each word and the seed word of veterinary drug knowledge, gives different weights to the words, and integrates the weight information into Gibbs sampling process;

p1 and p2 represent two words, and d represents the path distance between p1 and p2 in the veterinary medicine knowledge hierarchy system. The larger d, the smaller the similarity. The value range of similarity is [0,1]. k is an adjustable parameter, usually set to 20 by default.

(e) LDA favors the extraction of high-frequency words during the parameter estimation process, while some low-frequency feature words of hidden topics are drowned. Taking into account both high-frequency words and low-frequency feature words of hidden topics, consider using the TF-IDF method for optimization. TF-IDF uses statistical methods to calculate and express the importance of a certain keyword in the text. The topic-word matrix iteratively generated by the topic model is weighted by calculating the TF-IDF value of the word, which effectively weakens the influence of high-frequency noise words.

(f) Steps of weighted LDA:

1. Process the paper abstract data set into word segmentation and remove stop words;

2. Perform Gibbs sampling on the corpus to generate document-topic distribution and topic-word distribution;

3. Calculate the similarity, sort according to the similarity, retain the top K/2 topics as candidate topics, and construct a new document-topic distribution and topic-word distribution based on the candidate topics;

4. Use TF-IDF to weight the topic-word distribution to obtain the weighted probability, and then select the 20 feature words with the highest weight based on the topic-word distribution.

(g) Determination of the number of topics in LDA: When training the model, it is necessary to set the number of topics in advance and manually adjust the parameters based on the training results. The number of topics is 40, the hyperparameter α is 0.25, and β is 0.1;

(h) The documents in the corpus are veterinary drug knowledge-related documents obtained by SVM classification in the previous step. Topic mining is performed on these documents to obtain related topic vocabulary. Then search again using these topic words.

The establishment of weighted information extraction in step (3) includes the following steps:

(a) Named entity recognition: Use open source word segmentation tools for word segmentation and remove stop words, and use the veterinary drug knowledge dictionary for named entity recognition;

(b) Relationship extraction: Define the relationship extraction model between veterinary drug knowledge entities in advance and perform relationship extraction.

Description of the drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only These are embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on the provided drawings without exerting creative efforts.

Figure 1 is the veterinary drug knowledge dictionary framework of the present invention;

Figure 2 is a schematic diagram of the LDA algorithm;

Figure 3 is a flow chart of a weighted LDA document search algorithm.

Detailed ways

In order to further elaborate on the technical means and effects adopted by the present invention to achieve the intended inventive purpose, the specific implementation manner, structure, characteristics and effects proposed according to the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments. back.

A method for constructing a knowledge graph of veterinary drug residues based on weighted LDA, including the following steps:

(1) Construct a veterinary drug knowledge framework: use hierarchical analysis and rule-based methods to extract knowledge from veterinary pharmacology and veterinary toxicology books. Obtain veterinary drug toxicology-related knowledge from the Pubchem website using a wrapper-based approach. The jieba word segmentation tool is used to remove stop words, word segmentation, and part-of-speech tagging of these corpus, and finally form a dictionary to form a hierarchical veterinary drug knowledge framework;

(2) Download literature data: Use the dictionary obtained in the previous step, combined with the name of the veterinary drug, to conduct a multi-level search on the Web of science, that is, traverse every path from the root node to the leaf node, and perform a multi-layer search on all the words on each path. Search within layer results. The support vector machine (SVM) method was used to classify the obtained documents, including those related to veterinary drug knowledge and those not related to veterinary drug knowledge. For literature related to veterinary drug knowledge, the improved LDA method was used for topic extraction;

(3) Information extraction: dictionary-based named entity recognition and relationship extraction;

(4) Build a knowledge graph: Import the entities of the above-mentioned knowledge in the field of veterinary medicine and the relationships between entities into the Neo4j database in csv format.

The veterinary drug knowledge framework used in step (1) includes the following contents:

(a) Develop a knowledge structure for veterinary drug residues, which includes five parts: veterinary drug residues, introduction to toxicology, effects on organs and systems, properties and toxicity;

(b) Veterinary drug residues: including causes, effects and hazards. Hazards can be divided into three parts: harm to the human body, food and the environment;

(c) Veterinary drug attributes: category, physical and chemical properties, pharmacokinetics, effects, applications, maximum residue limits and adverse reactions;

(d) Toxicity of veterinary drugs: classification of toxic effects, commonly used parameters, special risk groups, exposure routes, preventive measures, inhalation methods, and animal experiments. The classification of toxic effects includes nature, time of occurrence, site and recovery. Commonly used parameters include acute toxicity, mutagenicity, carcinogenicity, teratogenicity, acute toxicity, etc. Animal experiments include mice, rats, rabbits, dogs, etc.;

(e) Introduction to veterinary drug toxicology: including purpose, content and methods. Methods can be divided into two parts: biological experiments and group surveys; effects on organs and systems: including eyes, skin, liver, kidneys, nervous system, blood system, immune system, gastrointestinal tract, endocrine system and respiratory system;

(f) If each part contains table content, place it under the corresponding category.

The multi-level search in step (2) includes the following steps:

(a) Selected veterinary drugs: The "National Food Safety Standard Maximum Residue Limits of Veterinary Drugs in Foods" standard stipulates 2191 residue limits and usage requirements for 267 kinds (categories) of veterinary drugs in livestock and poultry products, aquatic products, and bee products;

(b) Use Selenium and chrome driver to complete dynamic web page (Ajax) data capture. The scope of the search is all documents since the establishment of the web of science database. Considering that the amount of data in veterinary drug toxicology research is small, the search is not restricted to journals;

(c) According to the veterinary drug knowledge framework, from the root node to the leaf node. For all nodes on each path, these keywords are combined to search in multi-level results.

SVM text classification in step (2) includes the following steps:

(a) The purpose is to divide the obtained literature set into two categories, those related to veterinary drug knowledge and those not related to veterinary drug knowledge;

(b) The TF-IDF method calculates and expresses the importance of a certain keyword in the text through statistical methods. TF refers to word frequency, which indicates the frequency of a specified term appearing in the text, and IDF refers to inverse text frequency, which is a measure of the general importance of a word. The TF calculation method of entry t_i in text d_j:

Among them, n _{i, j} is the number of times the term t _i appears in the text d _j , and the denominator represents the sum of the number of times all words appear in the text d _j .

How to calculate IDF:

Among them, |D| is the total number of texts, |j:t _i ∈d _j | is the total number of texts containing the entry t _i . In order to prevent the denominator from being 0 due to the entry being not in the text, add 1 to the denominator. Finally, the TF-IDF value of the entry t _i is obtained:

tfidf _{i, j} = tf _{i, j} × idf _i

(a) First, use the TF-IDF algorithm to extract the feature words of part of the abstract of the paper and generate a document vector. If the full text of the abstract is selected, the vector dimension is too high, which will increase the complexity of the calculation and is not conducive to subsequent classification. According to the characteristics of literature related to veterinary drug knowledge, just select the short text in the conclusion part of the paper abstract. For this part of the text, use the TF-IDF algorithm to extract feature words and generate a document vector;

(b) Randomly select part of the data and manually label it. Set the ratio of training set to test set to 8:2;

(c) Adjust the SVM penalty parameter C and evaluate the model based on accuracy (a), precision (P), recall (R) and F1 value;

(d) Test set to verify the model;

(e) For literature data, use the trained model to obtain literature data related to the topic of veterinary drug residues.

Establishing the weighted LDA topic model in step (2) includes the following steps:

(a) LDA (latentdirichletallocation) is a 3-layer Bayesian model that describes the relationship between documents, topics, and vocabulary. Its graph model is shown in Figure 2. The meaning of each symbol in the figure: α is the hyperparameter of the Dirichlet distribution θ, β is the Dirichlet distribution The hyperparameters, θ is the polynomial distribution of "document-topic", /> is the polynomial distribution of "topic-vocabulary", z is the topic distribution of words, w is the word, K is the number of topics, M is the number of documents, and N is the number of words in a document;

(b)LDA process:

1. Randomly assign a topic number Z to each word in each document in the corpus;

2. Re-scan the corpus, use the Gibbs Sampling formula to sample each word, find its theme, and update it in the corpus;

3. Repeat step 2 until Gibbs Sampling converges;

4. The topic word co-occurrence frequency matrix of the statistical corpus, this matrix is the model of LDA.

(c) Gibbs Sampling fitting θ, the process of:

1. Scan the article and randomly assign a topic Z _j to each word w _n ;

2. Initialize Z _j to an integer between 1 and K;

3. Re-scan each article, use the LDA model to perform topic modeling on the corpus, use GibbsSampling for parameter inference to continuously iterate, and record the value of Z _j at the same time. Parameter θ, The calculation formula is as follows:

in, is the number of words of topic j in article d,/> is the number of words of all topics in article d,/> is the number of times word w appears under topic j,/> is the total number of words of topic j in article d.

(d) The LDA algorithm does not combine relevant semantic information well in the topic modeling process, which seriously affects the semantic coherence, interpretability and accuracy of text semantic representation of the topic. In view of the vocabulary distribution characteristics of veterinary drug knowledge, this paper uses a hierarchical semantic similarity calculation formula to calculate the similarity based on the semantic similarity between each word and the seed word of veterinary drug knowledge, gives different weights to the words, and integrates the weight information into Gibbs sampling process;

p1 and p2 represent two words, and d represents the path distance between p1 and p2 in the veterinary medicine knowledge hierarchy system. The larger d, the smaller the similarity. The value range of similarity is [0,1]. k is an adjustable parameter, usually set to 20 by default.

(e) LDA favors the extraction of high-frequency words during the parameter estimation process, while some low-frequency feature words of hidden topics are drowned. Taking into account both high-frequency words and low-frequency feature words of hidden topics, consider using the TF-IDF method for optimization. TF-IDF uses statistical methods to calculate and express the importance of a certain keyword in the text. The topic-word matrix iteratively generated by the topic model is weighted by calculating the TF-IDF value of the word, which effectively weakens the influence of high-frequency noise words.

(f) Steps of weighted LDA:

1. Process the paper abstract data set into word segmentation and remove stop words;

2. Perform Gibbs sampling on the corpus to generate document-topic distribution and topic-word distribution;

3. Calculate the similarity, sort according to the similarity, retain the top K/2 topics as candidate topics, and construct a new document-topic distribution and topic-word distribution based on the candidate topics;

4. Use TF-IDF to weight the topic-word distribution to obtain the weighted probability, and then select the 20 feature words with the highest weight based on the topic-word distribution.

(g) Determination of the number of topics in LDA: When training the model, it is necessary to set the number of topics in advance and manually adjust the parameters based on the training results. The number of topics is 40, the hyperparameter α is 0.25, and β is 0.1;

(h) The documents in the corpus are veterinary drug knowledge-related documents obtained by SVM classification in the previous step. Topic mining is performed on these documents to obtain related topic vocabulary. Then search again using these topic words.

The establishment of weighted information extraction in step (3) includes the following steps:

(a) Named entity recognition: Use open source word segmentation tools for word segmentation and remove stop words, and use the veterinary drug knowledge dictionary for named entity recognition;

(b) Relationship extraction: Define the relationship extraction model between veterinary drug knowledge entities in advance and perform relationship extraction.

All parts not involved in the present invention are the same as the prior art or can be implemented using the prior art.

Each embodiment in this specification is described in a progressive manner. Each embodiment focuses on its differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other. As for the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple. For relevant details, please refer to the description in the method section.

The above description of the disclosed embodiments enables those skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be practiced in other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (3)

1. A method for constructing a knowledge graph of veterinary drug residues based on weighted LDA, which is characterized by including the following steps: (1) Build a veterinary drug knowledge framework: use hierarchical analysis and rule-based methods to extract knowledge from veterinary pharmacology and veterinary toxicology books, use a packager-based method to obtain veterinary drug toxicology-related knowledge from the Pubchem website, and use the jieba word segmentation tool to The described knowledge is used to remove stop words, word segmentation, and part-of-speech tagging to finally form a dictionary, forming a hierarchical veterinary drug knowledge framework; (2) Download literature data: Use the dictionary obtained in the previous step, combined with the names of veterinary drugs, to conduct a multi-layer search on the Web of science, that is, traverse every path from the root node to the leaf node, and conduct a search for all words on each path. Search in multi-layer results, use the support vector machine SVM method to classify the obtained documents, including two categories: related to veterinary drug knowledge and irrelevant to veterinary drug knowledge. For documents related to veterinary drug knowledge, use the weighted LDA method for topic extraction; Establishing the weighted LDA topic model in step (2) includes the following steps: (4a) LDA is a 3-layer Bayesian model, which describes the relationship between documents, topics, and words. α is the hyperparameter of the Dirichlet distribution θ, and β is the Dirichlet distribution. The hyperparameters, θ is the polynomial distribution of "document-topic", /> is the polynomial distribution of "topic-vocabulary", z is the topic distribution of words, w is the word, K is the number of topics, M is the number of documents, and N is the number of words in a document; (4b) LDA process: (1) Randomly assign a topic number Z to each word in each document in the corpus; (2) Re-scan the corpus, use the Gibbs Sampling formula to sample each word, find its topic, and update it in the corpus; (3) Repeat step (2) until Gibbs Sampling converges; (4) The topic word co-occurrence frequency matrix of the statistical corpus, which is the model of LDA; (4c) Gibbs Sampling fitting θ, the process of: (1) Scan the article and randomly assign a topic Z _j to each word w _n ; (2) Initialize Z _j to an integer between 1 and K; (3) Re-scan each article, use the LDA model to perform topic modeling on the corpus, use GibbsSampling for parameter inference to continuously iterate, and record the value of Z _j , parameters θ, The calculation formula is as follows: in, is the number of words of topic j in article d,/> is the number of words of all topics in article d,/> is the number of times word w appears under topic j,/> is the total number of words of topic j in article d; (4d) Based on the vocabulary distribution characteristics of veterinary drug knowledge, based on the semantic similarity between each word and the seed word of veterinary drug knowledge, use a hierarchical semantic similarity calculation formula to calculate the similarity, assign different weights to the vocabulary, and integrate the weight information into the vocabulary. Booth sampling process; p1 and p2 represent two words, d represents the path distance between p1 and p2 in the veterinary drug knowledge hierarchy system. The larger d, the smaller the similarity. The value range of similarity is [0,1], and k is set to 20; (4e) Use the TF-IDF method for optimization. TF-IDF calculates and expresses the importance of a keyword in the text through statistical methods, and calculates the TF-IDF value of the word to iteratively generate the topic-word matrix of the topic model. Weighting is performed to weaken the influence of high-frequency noise words; (4f) Steps of weighted LDA: (1) Process the paper abstract data set into word segmentation and remove stop words; (2) Perform Gibbs sampling on the corpus to generate document-topic distribution and topic-word distribution; (3) Calculate the similarity, sort according to the similarity, retain the top K/2 topics as candidate topics, and construct a new document-topic distribution and topic-word distribution based on the candidate topics; (4) Use TF-IDF to weight the topic-word distribution to obtain the weighted probability, and then select the 20 feature words with the highest weight based on the topic-word distribution; (4g) Determination of the number of topics in LDA: When training the model, the number of topics needs to be set in advance. According to the training results, the parameters are manually adjusted. The hyperparameter α is set to 0.25 and β is set to 0.1; (4h) The documents in the corpus are veterinary drug knowledge-related documents obtained by SVM classification in the previous step. Perform topic mining on these documents to obtain related topic words, and then use these topic words to search again; (3) Information extraction: dictionary-based named entity recognition and relationship extraction; (4) Build a knowledge graph: Import the entities of the above-mentioned knowledge in the field of veterinary medicine and the relationships between entities into the Neo4j database in csv format. 2. The method for constructing a knowledge map of veterinary drug residues based on weighted LDA according to claim 1, characterized in that constructing a veterinary drug knowledge framework in step (1) includes the following content: (2a) Formulate a veterinary drug residue knowledge system structure, which includes five parts: veterinary drug residues, toxicology introduction, effects on organs and systems, properties and toxicity; (2b) Veterinary drug residues: including causes, effects and hazards. Harms can be divided into three parts: harm to the human body, food and the environment; (2c) Veterinary drug attributes: category, physical and chemical properties, pharmacokinetics, effects, applications, maximum residue limits and adverse reactions; (2d) Toxicity of veterinary drugs: toxic effect classification, common parameters, special risk groups, exposure routes, preventive measures, inhalation methods, animal experiments. Toxic effect classification includes nature, occurrence time, site and recovery status. Common parameters include acute toxicity, Mutagenicity, carcinogenicity, teratogenicity, and acute toxicity. The subjects of animal experiments include mice, rats, rabbits, and dogs; (2e) Introduction to veterinary drug toxicology: including purpose, content and methods, which are divided into two parts: biological experiments and group surveys; effects on organs and systems include eyes, skin, liver, kidneys, nervous system, blood system, immune system, gastrointestinal tract, endocrine system, and respiratory system; (2f) If each part contains table content, place it under the corresponding category. 3. The method for constructing a knowledge map of veterinary drug residues based on weighted LDA according to claim 1, characterized in that the multi-layer search in step (2) includes the following steps: (3a) Selected veterinary drugs: The "National Food Safety Standard Maximum Residue Limits of Veterinary Drugs in Foods" standard stipulates 2191 residue limits and usage requirements for 267 types/categories of veterinary drugs in livestock and poultry products, aquatic products, and bee products; (3b) Use Selenium and chrome driver to capture dynamic web page data. The search scope is all the literature since webofscience established the database. Considering that the amount of data in veterinary drug toxicology research is small, the search is not restricted to journals; (3c) According to the veterinary drug knowledge framework, from the root node to the leaf node, for all nodes on each path, these keywords are combined to search in multi-layer results.