False news recognition method and device based on knowledge graph, equipment and medium

Abstract

The application discloses a false news identification method and device based on a knowledge graph, a device, and a medium. Common sense knowledge graphs are constructed based on the Internet, common sense logical reasoning is performed by using a knowledge graph matching technology, and then the authenticity of news content is determined. The inference process is intuitive and easy to understand, and false news in different fields can be identified.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of information processing, and particularly relates to a false news identification method and device based on a knowledge graph, equipment and a medium. BACKGROUND

[0002] With the rapid development of mobile Internet, a piece of false news can cause great influence in a short time, and it is increasingly challenging to efficiently and accurately identify false news.

[0003] In recent years, storing external knowledge into a knowledge graph and combining external knowledge for false news detection is a common method. External knowledge often contains rich semantic information and objective facts, and using knowledge helps to understand the content of the news. At the same time, by comparing the news content with external knowledge, the authenticity of the news can be effectively judged.

[0004] The existing methods for combining external knowledge for false news detection mainly use attention mechanisms to integrate external knowledge. For example, some methods integrate visual information and text information into text representation through attention mechanisms, effectively helping the model to understand the content of the news text. Some methods use named entity recognition to align entities in the text with entities in the knowledge graph, and use a multi-head attention method to integrate news text information, entity information and entity context information to obtain a semantic-rich news text model. Some methods use graph neural networks to integrate external knowledge. Some methods construct a heterogeneous information network containing text information, image information and entity information in the knowledge graph, and use GCN to integrate information of each modality to obtain a news representation that integrates text information, external knowledge and visual information. However, the above methods directly integrate external knowledge into news content without considering comparing external knowledge with news content. Although the representation learning ability of deep learning can obtain the bottom-level spatial features of things, these features are obtained through a black box and are a continuous vector, which is impossible for humans to understand. Humans can only understand semantic scenarios. At the same time, compared with traditional methods, deep learning methods take longer to train and infer, and have higher hardware requirements. SUMMARY

[0005] The purpose of the present application is to overcome the defects of the prior art and provide a false news identification method, device, equipment and medium based on a knowledge graph. A common sense knowledge graph is constructed based on the Internet, common sense logical reasoning is performed using knowledge graph matching technology, and the authenticity of news content is determined.

[0006] The purpose of the present application is achieved by the following technical solutions:

[0007] A false news recognition method based on a knowledge graph, the method comprising:

[0008] extracting knowledge in the news to be recognized using a template matching method based on relation words to obtain common sense knowledge in the news;

[0009] extracting knowledge in the news to be recognized using a template-based information extraction model to obtain common sense knowledge in the news;

[0010] de-duplicating and reconstructing a news knowledge graph from the common sense knowledge obtained by the template matching method and the general information extraction method;

[0011] extracting a part related to the news knowledge graph from the common sense knowledge graph to obtain a common sense knowledge subgraph;

[0012] constructing a similarity matrix according to the common sense knowledge subgraph and the news knowledge graph point similarity information and edge similarity information;

[0013] obtaining a final graph matching result using a random walk-based graph matching algorithm according to the similarity matrix to determine whether the knowledge in the news knowledge graph is consistent with the knowledge in the common sense knowledge graph, and to determine the authenticity of the news to be recognized.

[0014] Further, the common sense knowledge graph construction method is as follows:

[0015] using a web crawler to crawl Internet entries to obtain common sense data;

[0016] performing named entity recognition on the common sense data, extracting entities and attributes, and storing the data set in the form of triples;

[0017] generating a knowledge graph from the common sense data.

[0018] Further, the extracting knowledge in the news to be recognized using a template matching method based on relation words to obtain common sense knowledge in the news specifically comprises:

[0019] abstracting sentence structure patterns appearing in the news into templates and establishing a relation word dictionary;

[0020] extracting common sense knowledge by performing word segmentation on the text and matching using an AC automaton.

[0021] Further, the extracting knowledge in the news to be recognized using a template-based information extraction model to obtain common sense knowledge in the news specifically comprises:

[0022] filtering entities according to the type of the news to be recognized and combining the common sense knowledge graph to construct a corresponding entity dictionary;

[0023] extracting the entity to be recognized in the sentence based on an entity dictionary using an AC automatic machine;

[0024] In combination with the common sense knowledge graph, knowledge in the news is extracted through a UIE model.

[0025] Further, the constructing a similarity matrix according to the common sense knowledge subgraph and the news knowledge graph point similarity information and edge similarity information specifically includes:

[0026]

[0027] wherein s is the similarity, X represents the character set corresponding to the string in the common sense knowledge subgraph, Y represents the character set corresponding to the string in the news knowledge graph, |X| represents the number of characters in set X, and |Y| represents the number of characters in set Y.

[0028] Further, the constructing a similarity matrix according to the common sense knowledge subgraph and the news knowledge graph point similarity information and edge similarity information specifically includes:

[0029] Using a pre-trained language model, the common sense knowledge subgraph and the news knowledge graph point similarity information and edge similarity information are inputted;

[0030] The inputted string of the common sense knowledge subgraph and the news knowledge graph are respectively outputted as a 768-dimensional word vector, and then the similarity between the two word vectors is judged according to the Euclidean distance between them to obtain the similarity matrix.

[0031] Further, the judging whether the knowledge in the news knowledge graph and the knowledge in the common sense knowledge graph are consistent according to the similarity matrix and using a graph matching algorithm based on random walk to obtain the final graph matching result specifically includes:

[0032] Using a graph matching algorithm based on random walk to match the common sense knowledge graph and the news knowledge graph, the entity nodes matched by the news knowledge graph in the common sense knowledge graph are obtained;

[0033] According to the matched entity node pair, the edges in the news knowledge graph are traversed to judge whether there is an edge in the common sense knowledge graph corresponding to the edge in the news knowledge graph, if not, it is determined as not consistent, if there is an edge corresponding, the semantic matching of the entity nodes and the relationship in the corresponding edge is performed using a pre-trained language model to judge whether they belong to the same knowledge, only when the head and tail entity and the relationship in the two edges are matched, it is determined as consistent, otherwise it is determined as inconsistent.

[0034] In another aspect, the present application also provides a false news recognition device based on a knowledge graph, which is used to implement the false news recognition method based on a knowledge graph as described above, and comprises:

[0035] A first extraction module is configured to extract knowledge in the news to be recognized using a template matching method based on a relation word, and obtain common sense knowledge in the news.

[0036] A second extraction module is configured to extract knowledge in the news to be recognized using a template-based information extraction model, and obtain common sense knowledge in the news.

[0037] A news knowledge graph establishing module is configured to reconstruct a news knowledge graph by removing duplicate common sense knowledge obtained by the template matching method and the general information extraction method.

[0038] A common sense knowledge subgraph establishing module is configured to extract a part related to the news knowledge graph from the common sense knowledge graph, and obtain a common sense knowledge subgraph.

[0039] A similarity matrix constructing module is configured to construct a similarity matrix according to the common sense knowledge subgraph and point similarity information and edge similarity information of the news knowledge graph.

[0040] A true-false recognition module is configured to obtain a final graph matching result by using a graph matching algorithm based on random walk according to the similarity matrix, and determine whether the knowledge in the news knowledge graph is consistent with the knowledge in the common sense knowledge graph, so as to determine the truth or falsity of the news to be recognized.

[0041] In another aspect, the present application also provides a computer device comprising a processor and a memory, wherein the memory stores a computer program, the computer program is loaded and executed by the processor to implement any of the false news recognition methods based on a knowledge graph as described above.

[0042] In another aspect, the present application also provides a computer readable storage medium, wherein the storage medium stores a computer program, the computer program is loaded and executed by a processor to implement any of the false news recognition methods based on a knowledge graph as described above.

[0043] The present application has the following advantages:

[0044] (1) The present application provides a connection for deep learning and semantic space by constructing a knowledge graph, and determines the truth or falsity of news by judging whether the knowledge in the news is consistent with the common sense knowledge graph, so that the inference process is intuitive and easy to understand.

[0045] (2) The application can identify false news in different fields by virtue of the wide knowledge field involved in the common sense knowledge graph, and the application scene is significantly better than the existing false news detection method. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a false news identification method based on a knowledge graph provided by an embodiment of the application;

[0047] Figure 2 is a flowchart of knowledge extraction by a template matching method based on a relation word according to an embodiment of the application;

[0048] Figure 3 is a structure block diagram of a false news identification device based on a knowledge graph provided by an embodiment of the application. DETAILED DESCRIPTION

[0049] The embodiments of the application will be described in detail below with specific, concrete examples. Those skilled in the art can easily understand other advantages and effects of the application from the disclosure of the specification. The application can also be implemented or applied by means of other different specific embodiments, and each detail in the specification can be modified or changed based on different views and applications without departing from the spirit of the application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0050] Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0051] The existing false news identification method directly integrates external knowledge into news content without considering comparing external knowledge and news content. Although the representation learning ability of deep learning can obtain the bottom space features of things, these features are obtained through a black box and are a continuous vector, which cannot be understood by humans. Humans can only understand the semantic scene. At the same time, compared with traditional methods, the training and reasoning of the deep learning method is time-consuming and requires high hardware.

[0052] In order to solve the above technical problems, the present application provides the following embodiments of a false news identification method, device, equipment and medium based on a knowledge graph.

[0053] Embodiment 1

[0054] Reference Figure 1 As Figure 1 shown is a false news identification method based on a knowledge graph provided by the embodiment, which specifically includes the following steps:

[0055] Step 1: Use a template matching method based on relational terms to extract knowledge from the news to be identified, and obtain common sense knowledge from the news.

[0056] Specifically, refer to Figure 2 ,like Figure 2 The diagram shows a flowchart of the knowledge extraction method based on relational words in this embodiment. By studying frequently occurring sentence structure patterns in news articles, these patterns are abstracted into a template. Let p be the envisioned general sentence structure pattern, typically composed of three parts: n, r, and v, corresponding to the head entity, relation, and tail entity in a triple, respectively. n and v need to have certain identifiers. Next, a relational word dictionary is built for the relations that p can identify. The text is segmented using the LAC model, relational words are selected using the AC automaton algorithm, and knowledge is extracted from the text based on the identifier rules in p.

[0057] Step 2: Use the UIE (Universal Information Extraction) model to extract knowledge from the news to be identified, and obtain common sense knowledge from the news.

[0058] Specifically, based on the type of news to be identified, entities are filtered using a commonsense knowledge graph to construct a corresponding entity dictionary. Based on this entity dictionary, an Aho-Corasick automaton (AC) is used to extract the entities to be identified from the sentences. Then, combined with the commonsense knowledge graph, the UIE model is used to extract knowledge from the news articles.

[0059] It should be noted that, in addition to the UIE model, other template-based information extraction models such as casrel and gplinker can also be used in this embodiment to extract knowledge from the news to be identified.

[0060] As one implementation method, the common sense knowledge graph is constructed in this embodiment as follows:

[0061] We used a web crawler to scrape Baidu Encyclopedia entries and obtain common knowledge data. We then used Jieba to perform named entity recognition on the scraped data, extracting entities and attributes, and storing the dataset as triples (entity, attribute, attribute value).

[0062] After obtaining the original data of entities and their attribute relations ownthink.csv, the abnormal data in the original data is processed, each node is numbered to generate the entity node dataset ownthink_entity.csv file, the attribute value node dataset ownthink_value.csv file, and the dataset ownthink_deal.csv file of the relationship between nodes. The names of the table headers are named according to the requirements of the neo4j database import file. Then, the entity nodes and attribute value nodes are linked through their relationship by using the neo4j-admin import tool, and are batch imported into the neo4j graph database to generate the Baidu Baike knowledge graph.

[0063] It should be noted that the Baidu Baike entries captured in the embodiment are only an example, and other entries on the Internet can also be captured as basic data for constructing the common sense knowledge graph.

[0064] Step 3: The common sense knowledge obtained by the template matching method and the general information extraction method is de-duplicated to construct a news knowledge graph.

[0065] Step 4: Extracting the part related to the news knowledge graph from the common sense knowledge graph to obtain a common sense knowledge subgraph.

[0066] Step 5: Constructing a similarity matrix according to the common sense knowledge subgraph and the news knowledge graph point similarity information and edge similarity information.

[0067] Specifically, in the embodiment, character-based similarity calculation criteria and Ernie model-based word vector generation similarity calculation criteria are designed when calculating the similarity. The character-based similarity calculation criteria function uses the Sorensen Dice similarity coefficient. The Dice coefficient is a simple calculation method for calculating the similarity between two sets, and the calculation formula is as follows:

[0068]

[0069] Where s is the similarity, X and Y represent the character sets corresponding to the two strings respectively, |X| represents the number of characters in set X, and |Y| represents the number of characters in set Y.

[0070] The method of calculating the similarity based on the Ernie model word vector generation uses a pre-trained language model to output a 768-dimensional word vector for the input string. Then, the Euclidean distance between the two word vectors is used to determine the similarity between them. In the present application, any one of the above two methods can be selected for similarity matching to obtain the similarity matrix of the graph.

[0071] It should be noted that in addition to using the Ernie model, other pre-trained language models can also be used in the embodiment.

[0072] Step 6: According to the similarity matrix, the final graph matching result is obtained using a random walk-based graph matching algorithm, to determine whether the knowledge in the news knowledge graph matches the knowledge in the common sense knowledge graph, and to determine whether the to-be-identified news is true or false.

[0073] Specifically, the embodiment uses a random walk-based graph matching algorithm to match the two knowledge graphs, obtains the entity nodes in the Baidu Baike knowledge graph that match the news knowledge graph, traverses the edges in the news knowledge graph according to these matched entity node pairs, determines whether there is an edge in the Baidu Baike knowledge graph that corresponds to the edge, and if not, considers that the knowledge does not conform to the Baidu Baike knowledge, and the news is fake news; if there is an edge that corresponds to the edge, the Ernie model is used to perform corresponding semantic matching on the entity nodes and the relationship in the two edges, to determine whether they belong to the same knowledge, and only when the head and tail entities and the relationship in the two edges all match is it determined that they are the same, otherwise it is determined that the knowledge does not conform to the Baidu Baike knowledge, and thus the news belongs to fake news.

[0074] Existing fake news identification technologies are mostly based on deep learning methods. Although the representation learning capability of deep learning can obtain the bottom layer spatial features of things, these features are obtained through a black box and are a continuous vector that humans cannot understand. Humans can only understand semantic scenarios. Knowledge graphs provide a connection between deep learning and semantic space and bridge the gap. The embodiment uses a knowledge graph to represent and model common sense knowledge, constructs a Baidu Baike knowledge graph, identifies the true or false of news by determining whether the knowledge in the news matches the Baidu Baike knowledge graph, and the inference process is intuitive and easy to understand. At present, most research on fake news detection only focuses on a certain field, and the fake news encountered in daily life often comes from multiple fields. The detection model obtained by the existing technical method can only identify fake news in a specific field, but with the advantage of the wide range of knowledge involved in the Baidu Baike knowledge base, fake news in different fields can be identified.

[0075] Embodiment 2

[0076] Referring to Figure 3 As Figure 3 shown is a structure block diagram of a fake news identification device based on a knowledge graph provided by the embodiment. The device specifically includes the following structures:

[0077] A first extraction module is configured to extract knowledge in the to-be-identified news using a template matching method based on a relationship word, and obtain common sense knowledge in the news.

[0078] The second extraction module is configured to extract knowledge in the to-be-identified news using a template-based information extraction model to obtain common sense knowledge in the news.

[0079] The news knowledge graph establishing module is configured to de-duplicate and establish a news knowledge graph based on the common sense knowledge obtained by the template matching method and the general information extraction method.

[0080] The common sense knowledge subgraph establishing module is configured to extract a part related to the news knowledge graph from the common sense knowledge graph to obtain a common sense knowledge subgraph.

[0081] The similarity matrix constructing module is configured to construct a similarity matrix based on the common sense knowledge subgraph and the point similarity information and the edge similarity information of the news knowledge graph.

[0082] The true-or-false identification module is configured to use a random walk-based graph matching algorithm to obtain a final graph matching result based on the similarity matrix, determine whether the knowledge in the news knowledge graph is consistent with the knowledge in the common sense knowledge graph, and determine the true or false of the to-be-identified news.

[0083] As an implementation form, the false news identification device based on a knowledge graph provided in the embodiment further includes a common sense knowledge graph constructing module, which is configured to use a web crawler to crawl Internet entries to obtain common sense data, perform named entity recognition on the common sense data, extract entities and attributes, store the data set in the form of triples, and generate a knowledge graph based on the common sense data.

[0084] The false news identification device based on a knowledge graph provided in the embodiment provides a connection for deep learning and semantic space by constructing a knowledge graph, identifies the true or false of news by determining whether the knowledge in the news is consistent with the common sense knowledge graph, and the inference process is intuitive and easy to understand. With the advantage of the common sense knowledge graph involving a wide range of knowledge fields, the device can identify false news in different fields, and the application scenario is significantly better than existing false news detection methods.

[0085] Embodiment 3

[0086] The preferred embodiment provides a computer device that can implement the steps in any embodiment of the false news identification method based on a knowledge graph provided in the embodiments of the application, and thus can implement the beneficial effects of the false news identification method based on a knowledge graph provided in the embodiments of the application. Details are described in the previous embodiments, which will not be repeated here.

[0087] Embodiment 4

[0088] Those skilled in the art can understand that all or part of the steps of various methods of the above embodiments can be completed by instructions or by controlling relevant hardware by the instructions, and the instructions can be stored in a computer readable storage medium and loaded and executed by a processor. To this end, the embodiments of the present application provide a storage medium, wherein a plurality of instructions are stored, and the instructions can be loaded by a processor to execute the steps of any embodiment of the knowledge graph based fake news identification method provided by the embodiments of the present application.

[0089] The storage medium can include a read only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0090] Since the instructions stored in the storage medium can execute the steps in any embodiment of the knowledge graph based fake news identification method provided by the embodiments of the present application, the beneficial effects of any knowledge graph based fake news identification method provided by the embodiments of the present application can be achieved, which are described in detail in the foregoing embodiments and will not be described here.

[0091] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for identifying fake news based on knowledge graphs, characterized in that, The method includes: A template matching method based on relational words is used to extract knowledge from the news to be identified, thereby obtaining common sense knowledge from the news. A template-based information extraction model is used to extract knowledge from the news to be identified, thereby obtaining common sense knowledge from the news. The common sense knowledge obtained by the template matching method and the general information extraction method is de-duplicated to reconstruct the news knowledge graph; Extract the parts related to the news knowledge graph from the common sense knowledge graph to obtain the common sense knowledge subgraph; A similarity matrix is ​​constructed based on the point similarity information and edge similarity information of the common sense knowledge subgraph and the news knowledge graph, specifically including: ， Where s is the similarity, X represents the set of characters corresponding to the strings in the common sense knowledge subgraph, Y represents the set of characters corresponding to the strings in the news knowledge graph, |X| represents the number of characters in set X, and |Y| represents the number of characters in set Y; Based on the similarity matrix, a graph matching algorithm based on random walks is used to obtain the final graph matching result. This determines whether the knowledge in the news knowledge graph matches the knowledge in the common sense knowledge graph, thus determining the veracity of the news to be identified. Specifically, this includes: A graph matching algorithm based on random walks is used to match the common sense knowledge graph and the news knowledge graph to obtain the entity nodes of the news knowledge graph that are matched in the common sense knowledge graph. Based on the matched entity node pairs, the edges in the news knowledge graph are traversed to determine whether there are any edges in the common sense knowledge graph that correspond to the edges in the news knowledge graph. If not, it is determined that they do not match. If there are corresponding edges, the pre-trained language model is used to perform semantic matching on the entity nodes and relations in the corresponding edges to determine whether they belong to the same knowledge. Only when the head and tail entities and relations in the two edges match are they determined to match; otherwise, they are determined to not match.

2. The method for identifying fake news based on knowledge graphs as described in claim 1, characterized in that, The method for constructing the common sense knowledge graph is as follows: Use web crawlers to scrape internet entries and obtain common knowledge data; Named entity recognition is performed on the common sense data to extract entities and attributes, and the dataset is stored in the form of triples. Generate a knowledge graph from the common sense data.

3. The method for identifying fake news based on knowledge graphs as described in claim 1, characterized in that, The method of extracting knowledge from the news to be identified using a relational word-based template matching method, specifically including: Abstract the sentence structure patterns appearing in news articles into templates and build a dictionary of relational terms; Common sense knowledge is extracted by segmenting the text and matching it using the AC automaton.

4. The knowledge graph-based fake news identification method as described in claim 2, characterized in that, The step of using a template-based information extraction model to extract knowledge from the news to be identified, and obtaining common sense knowledge from the news, specifically includes: Based on the news type to be identified, entities are filtered using the common sense knowledge graph to construct a corresponding entity dictionary; Based on an entity dictionary, the AC automaton is used to extract the entities to be identified from the sentence; Based on the aforementioned common sense knowledge graph, knowledge is extracted from news articles using the UIE model.

5. The method for identifying fake news based on knowledge graphs as described in claim 1, characterized in that, The construction of a similarity matrix based on the common sense knowledge subgraph and the point similarity information and edge similarity information of the news knowledge graph specifically includes: Using a pre-trained language model, input the point similarity information and edge similarity information of the common sense knowledge subgraph and the news knowledge graph; The input strings of the common sense knowledge subgraph and the news knowledge graph are output as 768-dimensional word vectors. Then, the similarity between the two word vectors is determined based on the Euclidean distance between them, and the similarity matrix is ​​obtained.

6. A knowledge graph-based fake news identification device, the device being used to implement the knowledge graph-based fake news identification method of claim 1, the device comprising: The first extraction module is used to extract knowledge from the news to be identified using a template matching method based on relational words, thereby obtaining common sense knowledge from the news. The second extraction module is used to extract knowledge from the news to be identified using a template-based information extraction model, thereby obtaining common knowledge from the news. The news knowledge graph building module is used to reconstruct the news knowledge graph by deduplicating the common sense knowledge obtained by the template matching method and the general information extraction method. The common sense knowledge subgraph building module is used to extract the parts related to the news knowledge graph from the common sense knowledge graph to obtain the common sense knowledge subgraph. The similarity matrix construction module is used to construct a similarity matrix based on the point similarity information and edge similarity information of the common sense knowledge subgraph and the news knowledge graph; The authenticity identification module is used to obtain the final graph matching result based on the similarity matrix using a graph matching algorithm based on random walk, to determine whether the knowledge in the news knowledge graph matches the knowledge in the common sense knowledge graph, and thus determine the authenticity of the news to be identified.

7. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program, which is loaded and executed by the processor to implement the knowledge graph-based fake news identification method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which is loaded and executed by a processor to implement the knowledge graph-based fake news identification method as described in any one of claims 1 to 5.

Images