CN115563297A - A food safety knowledge map construction and completion method based on graph neural network - Google Patents

Abstract

The invention provides a food safety knowledge graph building and complementing method based on a graph neural network, which comprises the steps of acquiring national food safety standard files related to food types, food additives and pesticide residues in food, processing the national food safety standard files into triples capable of being applied to the knowledge graph through operations such as data cleaning, formatting and the like, and building a body layer mode framework of the food safety knowledge graph; designing a food safety knowledge map query system capable of realizing entity query and visual display; generating a word vector of an entity name in the food safety knowledge graph according to a pre-training language model BERT; and performing feature fusion on text information and graph structure information in the food safety knowledge graph by using a graph neural network architecture, and respectively applying the feature fusion to two downstream tasks of entity classification and link prediction so as to fulfill the aim of knowledge graph completion. The invention improves the integrity and the practicability of the food safety knowledge map and realizes the intelligent application of food safety standard information.

Description

A food safety knowledge map construction and completion method based on graph neural network

technical field

The invention belongs to the technical field of knowledge graph and graph neural network, and in particular relates to a method for constructing and completing a food safety knowledge graph based on graph neural network.

Background technique

With the continuous improvement of people's living standards, people pay more and more attention to food safety. Both regulators and ordinary people need a set of practical and effective standard guidance. But at present, the national food safety standards are still stored in the form of documents, and there are many types and numbers of documents, the format is not uniform enough, and there are references to each other. Although my country has been promoting the standardization and information management of food quality, there are still deficiencies in the structural integration of these standards and systematic analysis of the relationship.

With the development of science and technology, the form of knowledge graph can describe the concepts and relationships of the objective world as nodes and edges in the form of a graph structure, which is more conducive to querying the content of food standards and the relationship between them. At present, most of the food safety knowledge maps at home and abroad are focused on the construction of sub-fields, and there is no authoritative food safety standard knowledge base as a guide.

If only relying on manual extraction of standard content as data to construct a knowledge map, it is undoubtedly inefficient for such a huge field of food, and the information is often incomplete. The graph neural network has advantages in processing graph-structured data, and can try to be applied to entity classification and link prediction tasks of knowledge graph completion. However, due to the structure of the graph neural network itself, when the deep graph information is obtained, there will be over-smoothing, which will affect the effect of the model, and if only the graph structure information is considered, the information of the knowledge graph itself has not been fully utilized. It is also one of the issues that need to be considered now.

Contents of the invention

The technical problem to be solved by the present invention is to provide a method for constructing and completing a food safety knowledge map based on a graph neural network, which is used to improve the integrity of the food safety knowledge map.

The technical solution adopted by the present invention to solve the above technical problems is: a method for constructing and completing a food safety knowledge map based on a graph neural network, including the following steps:

(Here is consistent with the claim, to be filled after the claim is confirmed)

The beneficial effects of the present invention are:

1. A method for constructing and completing a food safety knowledge map based on a graph neural network of the present invention, constructing a knowledge map by extracting food safety standard information, and generating word vectors of entity names in the food safety knowledge map according to the pre-trained language model BERT; Using the graph neural network architecture to fuse the text information and graph structure information in the food safety knowledge map, and apply them to the two downstream tasks of entity classification and link prediction to complete the food safety knowledge map, and realize the improvement of food safety. The function of the integrity of the knowledge map; the invention introduces text representations into the graph neural network, enriches the node information, and alleviates the over-smoothing phenomenon caused by the structure of the multi-layer digital graph neural network itself, which is conducive to the model to learn more knowledge map information; the present invention improves the scoring function in the R-GCN link prediction model, and introduces the scoring function constructed by the idea of segmented block dot product, which is improved in the form of calculating the graph structure feature vector and the text feature vector respectively. effect on link prediction.

2. The present invention obtains national food safety standard documents related to food categories, food additives, and pesticide residues in food, and through operations such as data cleaning and formatting, the food safety standards for different food categories, food additives, and maximum limits of pesticide residues in food, etc. The national standard information is processed into a triplet form that can be applied to the knowledge map, and the ontology layer model structure of the food safety knowledge map is constructed, which makes up for the vacancy of the knowledge map data in the field of food safety.

3. The present invention builds a Python-based Flask framework and a web-side visualization system of Echarts.js, stores the data of the national food safety standard content in the form of a graph and applies it to downstream tasks, and realizes entity query and visual display of food Security knowledge graph query system.

4. The invention makes up for the shortcomings of the current technology, improves the integrity and practicability of the food safety knowledge map, and realizes the intelligent application of food safety standard information.

Description of drawings

Fig. 1 is a flowchart of the preliminary construction of a food safety knowledge map according to an embodiment of the present invention.

Fig. 2 is an ontology structure diagram of the food safety knowledge map of the embodiment of the present invention.

Fig. 3 is a flowchart of data extraction and processing in an embodiment of the present invention.

Fig. 4 is a schematic structural diagram of an entity classification model according to an embodiment of the present invention.

Fig. 5 is a schematic structural diagram of a link prediction model according to an embodiment of the present invention.

detailed description

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

Referring to Fig. 1, the embodiment of the present invention includes a preliminary building block of a food safety knowledge graph, a knowledge graph entity classification model based on a graph neural network and a text pre-training model, and an encoder-decoder-based knowledge graph link prediction model.

A method for constructing and completing a food safety knowledge map based on a graph neural network of the present invention comprises the following steps:

S1: Collect data; extract national food safety standard information by means including web crawler and table analysis, including semi-structured information and unstructured information in standard documents; the work flow chart of data extraction is shown in Figure 3;

S11: Obtain information on national food safety standards from websites and download national food safety standard documents through web crawlers;

S12: OCR identifies the downloaded national food safety standard file, and parses the table information in the file;

S13: Obtain semi-structured data and unstructured data based on rule matching and web page and table analysis.

S2: Construct ontology; based on the obtained semi-structured information and manual assisted annotation, construct the ontology framework of the food safety knowledge map, and use the obtained ontology framework to further improve the data extraction method. The resulting ontology framework diagram is shown in Figure 2. These include entity types and relationship types related to food categories, food standards, food additives, and pesticide residues;

S3: Data processing; according to the storage requirements of the constructed ontology framework and knowledge graph, format and clean semi-structured and unstructured data to obtain triplet form data in csv format; work flow chart of data processing As shown in Figure 3;

S4: Store the knowledge map and build a visual query system;

S41: storing the knowledge graph triplet data in csv format into the Neo4J graph database;

S42: Based on the Python Flask framework, call the Neo4j graph database, and display the knowledge graph with the force-directed graph of Echarts.js;

S43: Call the cypher statement of Neo4J through Python to realize the query of the knowledge map entity, and display it on the web page.

S5: Use the graph neural network and the text pre-training model BERT to complete the knowledge map entity classification task; Figure 4 shows a schematic diagram of the relational graph convolution network combined with text representation in the present invention, including the BERT text representation part and the R-GCN graph structure information part ;

S51: Construct a data set based on the triplet of the food safety knowledge map, divide the training set, verification set and test set, assign the food safety-related entity type to the corresponding type number, and abstract the entity itself into the entity number, based on the entity type table Construct the DGLDataset dataset with the triple table;

S52: Use the BERT pre-training model as the entity classification model (see Figure 4), extract the text word vector representation of the entity name of the knowledge graph, and perform dimensionality reduction processing on the word vector obtained from the pre-training;

The entity classification model is used to combine the structural information of the graph and the textual description of the entity, expand the knowledge graph representation from a single network structure representation to the common representation of structure and text, and build a model that integrates multi-source knowledge graph information; through the text pre-training model BERT obtains the word vector representation of the entity description, adds the word vector representation to the entity representation and fuses the original entity relationship structure information, and jointly learns the knowledge graph representation through the relational graph convolutional network model;

The BERT service bert-as-service is used to encapsulate the pre-trained BERT model based on the C/S architecture, and provide it to the client as a service, so as to provide support for engineering projects more conveniently. Because most of the entity names in the food safety knowledge map are Chinese words, the Chinese BERT pre-training model chinese_L-12_H-768_A-12 is selected to obtain the 768-dimensional word vectors of all entity names. Due to the high dimension of the word vector obtained through the pre-training model, if it is used directly, the calculation amount of the model will be greatly increased, and the dimensionality of the 768-dimensional word vector with a high dimension is reduced by the method of Principal Component Analysis (PCA). . Subsequent experiments proved that the reduction of dimensionality has little effect on the experimental results. Considering the computational complexity and the integrity of text information representation, the text representation dimension used uniformly by the entity classification model is 5-dimensional and the floating-point number is reserved for 4 bits. Decimals are convenient for subsequent use.

S53: Use R-GCN to aggregate information, update node representation, and obtain graph structure information;

S54: Embedding the text word vector of the entity name before the output layer of R-GCN;

S55: After learning the output layer, use softmax to classify entities, predict possible categories for each entity in the sample, and optimize by minimizing the cross-entropy loss function on all labeled nodes.

Experiments have proved that the R-GCN model combined with text information in the present invention achieves a higher entity classification accuracy rate when compared with whether text representation is used.

S6: The encoder-decoder-based graph neural network model completes the knowledge map link prediction task; Figure 5 shows the structure diagram of the encoder-decoder-based link prediction model, including the BERT text representation part and R-GCN graph structure information Part; the R-GCN model is used as the encoder to obtain the representation of the triplet, and the scoring function is used as the decoder;

The knowledge map constructed by triples and the entity classification based on graph neural network and text representation have improved the integrity of the food safety knowledge map in terms of entity types, but there are still some missing relationships in the food safety knowledge map. By predicting the missing triples in the knowledge graph and judging whether the triples meet the requirements based on the scoring function, link prediction is realized. For a triple of the type (head entity, relationship, tail entity), predict another entity based on one of the entities and the corresponding relationship, so as to solve the problem of head and tail entity prediction.

The link prediction model obtained in this way can also be applied to the reasoning of knowledge graphs, predicting other entities that may have a specific relationship with it according to an entity and its relationship structure. As far as food safety issues are concerned, the above methods can be used to infer other foods that have a certain relationship with it or other substances that may exceed the standard according to a certain food category, so as to accurately grasp the possible existence of food in the whole food chain (production, transportation, storage, etc.). potential safety hazards, discover problems in time, and prevent problems before they happen.

In the link prediction task, it mainly includes two aspects of work. On the one hand, in order for the model to learn the information of the knowledge graph, it is necessary to obtain the representation of the knowledge graph. Compared with entity classification models, more consideration needs to be given to the type and representation of edges in structural graphs. On the other hand, it is necessary to choose a triplet scoring function suitable for the current problem, which directly affects the training and results of the model.

S61: Divide the food safety knowledge map data set. In order to avoid the impact of type imbalance on the prediction effect, when extracting data, try to ensure that the relationship type ratios of the training set, verification set, and test set are basically the same. Food safety-related entity types and relationship types Abstract to the corresponding type number;

S62: Use the BERT pre-training model to perform text feature extraction on the description information of each node in the knowledge graph and obtain word vector representation;

S63: The encoder part of the link prediction model: update the representation of the node based on the hidden layer of R-GCN, including the description of the node as text information, obtain the word vector representation through the BERT pre-training model, and embed the text representation word vector into the last R- Before the GCN hidden layer, the graph structure information of the knowledge map itself is input into the R-GCN hidden layer after initialization and embedding to obtain the representation of the knowledge map; the two features are fused and represented before entering the last hidden layer, and then passed The last layer of R-GCN hidden layer;

S64: the decoder part of the link prediction model: realize the evaluation of triplets through the scoring function, and obtain the ranking of positive samples among all sample scores. Define the scoring function as f(s, r, o), where s, r, o denote head entity, relation and tail entity, respectively.

S641: The scoring function is based on the segmented block dot product idea, combined with the graph structure information and text information contained in the node representation, and is divided into f _graph (s, r, o) and f _word (s, r, o) part to calculate, that is, f(s, r, o) = f _graph (s, r, o) + f _word (s, r, o), divide the graph structure features and text features into two parts, f _graph (s, r, o) = ∑ _{0≤x, y<2} s _{x, y} <r _x , s _y , o _z >, f _word (s, r, o) = ∑ _{2≤x, y<4} <r _x , s _y , o _z >; where x, y, and z represent the segment numbers of the relationship representation, head entity representation, and tail entity representation segments, respectively.

S642: The graph structure information part needs to consider the symmetry, use an odd number to indicate an asymmetric relationship, and an even number to indicate a symmetric relationship, introduce parameters s _{x, y} , and do not add relationship constraints in the calculation of the word vector part. When x is an odd number and x+y≥2, s _{x, y} is -1, and the rest are 1.

S643: In order to simplify the calculation, the index w _{x, y} of the tail entity is introduced to reduce the complexity.

S65: The data set constructed in the previous stage does not contain negative data, so the method of randomly destroying positive data is used for negative sampling, that is, a certain proportion of negative samples are generated by replacing the head entity or tail entity with other random entities in the data set, And optimized by cross-entropy loss.

The above embodiments are only used to illustrate the design concept and characteristics of the present invention, and its purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims (9)

1. A method for constructing and completing a food safety knowledge map based on a graph neural network, characterized in that: comprising the following steps: S1: Use methods including web crawlers and table analysis to extract national food safety standard information. National food safety standard information includes semi-structured data and unstructured data in standard documents; S2: Based on semi-structured data and artificially assisted labeling, construct the ontology framework of the food safety knowledge map, and use the ontology framework to further improve the data extraction method; S3: According to the storage requirements of the ontology framework and knowledge graph, format and clean semi-structured and unstructured data to obtain triple data of food safety knowledge graph in csv format; S4: Store food safety knowledge map triplet data and build a visual query system; S5: Combining text representation relational graph convolutional network, including BERT text representation part and R-GCN graph structure information part; using graph neural network and text pre-training model BERT to build an entity classification model to complete the knowledge map entity classification task; entity classification model It is used to combine the structural information of the graph and the text description of the entity, expand the knowledge graph representation from a single network structure representation to the common representation of structure and text, and build a model that integrates multi-source knowledge graph information; the entity is obtained through the text pre-training model BERT The word vector representation of the description, after adding the word vector representation to the entity representation and merging with the original entity relationship structure information, jointly learn the knowledge graph representation through the relational graph convolutional network model; S6: Use the R-GCN model as the encoder to obtain the representation of the triplet, and use the scoring function as the decoder; construct the structure of the link prediction model based on the encoder-decoder, including the BERT text representation part and the R-GCN graph structure information Part; the encoder-decoder-based link prediction model is used to predict other entities that may have a certain relationship with it according to an entity and its relationship structure, by predicting missing triples in the knowledge graph, based on the scoring function judgment Whether the triplet meets the requirements, so as to complete the knowledge graph link prediction task. 2. A method for constructing and completing a food safety knowledge map based on a graph neural network according to claim 1, characterized in that: in the step S1, the specific steps are: S11: Obtain information on national food safety standards from the Internet through a crawler program, and download national food safety standard documents; S12: OCR identifies the downloaded national food safety standard file, and parses the table information in the file; S13: Obtain semi-structured data and unstructured data based on rule matching and web page and table analysis. 3. A graph neural network-based food safety knowledge map construction and completion method according to claim 1, characterized in that: in the step S2, the ontology framework includes food categories, food standards, food additives and Entity types and relationship types related to pesticide residues. 4. A food safety knowledge map construction and completion method based on graph neural network according to claim 3, characterized in that: in the step S4, the specific steps are: S41: storing the knowledge graph triplet data in csv format into the Neo4J graph database; S42: Based on the Python Flask framework, call the Neo4j graph database, and display the knowledge graph with the force-directed graph of Echarts.js; S43: Call the cypher statement of Neo4J through Python to realize the query of the knowledge map entity, and display it on the web page. 5. A method for constructing and completing a food safety knowledge map based on a graph neural network according to claim 4, characterized in that: in the step S5, the specific steps are: S51: Construct a data set based on the triplet data of the food safety knowledge map, divide the training set, verification set and test set, use the entity type related to food safety as the corresponding type number, and abstract the entity itself into an entity number, based on the entity type table Construct the DGLDataset dataset with the triple table; S52: Using the BERT pre-training model as the entity classification model, extracting the text word vector representation of the entity name of the knowledge graph, and performing dimensionality reduction processing on the word vector obtained through pre-training; S53: Utilizing the R-GCN aggregation information to update the node representation, and obtain graph structure information; S54: Embedding the text word vector of the entity name before the output layer of R-GCN; S55: After learning the output layer, use softmax to classify entities, predict possible categories for each entity in the sample, and optimize by minimizing the cross-entropy loss function on all labeled nodes. 6. A food safety knowledge map construction and completion method based on graph neural network according to claim 5, characterized in that: in the step S6, the entity classification model includes an input layer, several hidden layers and an output layer; Each layer of the entity classification model includes a relational graph convolutional network layer R-GCN, which is used to calculate each node on the training set according to the node representation and edge type through the message function Output information, and aggregate information to obtain new node representations; The input layer of the model uses the type and number of entities as features; the input layer and hidden layer use ReLU as the activation function; the output of the last layer uses softmax for classification; In the hidden layer, the node representation obtained by the input layer is calculated by multiple relational graph convolutional network layers R-GCN and activation function ReLU respectively; let Y be the index set of nodes, K be the total number of categories,

Represents the predicted value of the k-th category of the i-th node with a label in the l-th layer, and t _ik represents the real label, then the output layer obtains the minimum cross-entropy loss function Loss on all labeled nodes:

In the entity classification task, the calculation amount of the model is reduced by means of basis function decomposition; the word vector obtained by the BERT pre-training model is embedded into the output layer of the R-GCN model to obtain the best effect. 7. A food safety knowledge map construction and completion method based on graph neural network according to claim 5, characterized in that: in the step S6, the specific steps are: S61: Divide the food safety knowledge map data set, and make the relationship type ratios of the training set, verification set, and test set consistent when extracting data to avoid type imbalance from affecting the prediction effect; abstract food safety-related entity types and relationship types into correspondence type number of S62: Use the BERT pre-training model to perform text feature extraction on the description information of each node in the knowledge graph, and obtain word vector representation; S63: The encoder part of the link prediction model updates the representation of the node based on the hidden layer of R-GCN, uses the description of the node as text information to obtain the word vector representation through the BERT pre-training model, and embeds the text representation word vector into the last R-GCN Before the GCN hidden layer, the graph structure information of the knowledge map itself is input into the R-GCN hidden layer after initialization and embedding to obtain the representation of the knowledge map; the two features are fused and represented before entering the last hidden layer, and then passed The last layer of R-GCN hidden layer; S64: Let s, r, and o denote the head entity, relation, and tail entity respectively, and define the scoring function as f(s, r, o); the decoder part of the link prediction model realizes the evaluation of triples through the scoring function, and obtains The rank of positive samples among all sample scores; S65: The data set constructed in the previous stage does not contain negative example data, and the method of randomly destroying positive example data is used for negative sampling, and a certain proportion of negative samples is generated by replacing the head entity or tail entity with other random entities in the data set, and through crossover Entropy loss is optimized. 8. A graph neural network-based food safety knowledge map construction and completion method according to claim 7, characterized in that: in the step S64, the specific steps are: S641: Let the graph structure feature of the scoring function be f _graph (s, r, o), and the text feature be f _word (s, r, o), then: f(s, r, o) = f _graph (s, r, o) + f _word (s, r, o); Let x, y, and z be the segment numbers of the relationship representation, head entity representation, and tail entity representation segments respectively, and divide the graph structure features and text features into two parts, as follows: f _graph (s, r, o) = ∑ _{0≤x, y<2} s _{x, y} <r _x , s _y , o _z >, f _word (s, r, o) = ∑ _{2≤x, y<4} <r _x , s _y , o _z >; S642: Consider symmetry in the graph structure information part, use odd numbers to indicate asymmetric relationships, and even numbers to indicate symmetric relationships, and introduce parameters s _{x, y} ; do not add relationship constraints in the calculation of the word vector part; when x is an odd number and x+y≥ 2, s _{x, y} is -1, and the rest are 1; S643: Introduce the index w _{x, y} of the tail entity to simplify calculation and reduce complexity. 9. A computer storage medium, characterized in that: a computer program that can be executed by a computer processor is stored therein, and the computer program executes a graph-based neural network as described in any one of claims 1 to 8. Construction and completion method of food safety knowledge graph in network.

Images