A Table Fact Verification Method Based on a Tight Graph Inference Network

Abstract

The present invention relates to a tabular fact verification method based on a dense graph inference network, which includes the following steps: constructing a news dependency subgraph, a table subgraph, and a program subgraph for news sequences, tables, and program information respectively, splitting the table subgraph into multiple meta-graph networks, and immersively connecting the table subgraph with the news subgraph and the program subgraph respectively to form a dense news-table association graph and a program-table association graph network; designing two inference modeling mechanisms, namely local association modeling and global consistency modeling, to respectively capture deep evidence with association and consistency, and obtaining deep language evidence and deep logical evidence on the dense news-table association graph and the program-table association graph respectively; designing an interactive fusion layer to strengthen the interaction and sharing between the two pieces of evidence in order to obtain mutually corroborative evidence. The present invention not only improves the performance of fake news detection, but also provides effective evidence to achieve the interpretability of the detection results of structured data.

Description

Technical Field

[0001] The present invention belongs to the field of electronic information technology, and particularly relates to a method for verifying table facts based on a tight graph inference network. Background Art

[0002] In the current rapid development of the Internet and social media, traditional mainstream media has encountered challenges in all aspects. Social media, including Weibo, WeChat, short video platforms, forums, blogs, etc., has quickly integrated into people's daily lives, providing a platform for people to communicate with each other and share opinions and insights. While spreading and sharing true and reliable news, a large number of unverified or untrue information, including rumors, false news, misreported news, etc., is also scattered in social media. Fact-checking has become the key to curbing the spread of false news. How to detect false information spread in social media in a timely and effective manner has become one of the key issues in the field of social media.

[0003] Existing research mainly focuses on the detection of false news in unstructured data. Usually, evidence is collected from unstructured web pages or comments to verify the truth and reliability of false news. However, these methods usually ignore fact-checking for structured data, that is, it is difficult to collect evidence from structured table data to verify the falsity of news. As we know, structured data (such as tables, XML information, etc.) is everywhere on the network. The fact-checking task based on structured data is an indispensable part of the fact-checking task. How to perform fact verification based on structured data not only has a strong demand but also has great challenges.

[0004] Different from the traditional pattern of fact-checking by obtaining evidence from unstructured data through language reasoning, symbolic information (such as symbolic information like count and max) plays an important role in understanding structured data. Current research mainly falls into two major categories: methods relying on pre-trained models and graph structure-based methods. Specifically, 1) The pre-training-based methods are more suitable for the situation where structured data and unstructured data coexist. They usually use pre-trained models such as TAPAS and TAPEX to train tabular data to obtain structured evidence. The advantage of these methods is that they can learn deep context semantics, but it is difficult to perform more accurate reasoning on the table according to the given statement; 2) The graph structure-based methods usually create various heterogeneous graph networks for two aspects of reasoning, namely language reasoning and symbolic reasoning. In particular, researchers Chen et al. released the TABFACT dataset and designed Table-BERT for language reasoning and the Latent Program Algorithm (LPA) for logical reasoning respectively to verify the sequence of statements based on the table. Subsequently, a series of related studies focused on leveraging the advantages of language reasoning and symbolic reasoning in this way, designed different heterogeneous graph networks, and fused language and symbolic information to carry out fact-checking and verification work. Although these methods have all achieved initial development and achieved relatively satisfactory performance, they all have an obvious defect, that is, when constructing the heterogeneous graph network, the connection between nodes in subgraphs is usually established based on the same content. This connection method has serious sparsity. Especially when encountering subgraphs with limited semantics, this sparsity is particularly obvious. Such a connection method affects the interaction and communication between subgraphs and is difficult to cooperate to capture valuable evidence in structured new data. Therefore, designing a reasonable and tight heterogeneous graph structure and performing efficient reasoning modeling between subgraphs is the key to fact-checking and verification based on structured data. Summary of the Invention

[0005] Technical Problems to be Solved

[0006] In view of the defects existing in the current methods for detecting false information based on structured data, the present invention proposes a method for verifying table facts based on a tight graph reasoning network.

[0007] Technical Solution

[0008] A method for verifying table facts based on a tight graph reasoning network, characterized in that:

[0009] S1: Context Encoding Module

[0010] Learn the context representations of three types of information with the help of a pre-trained BERT model, and the three types of information are news sequences, tables, and program information respectively;

[0011] S2: Tight Graph Construction Module

[0012] Construct a news dependency subgraph, a table subgraph, and a program subgraph for the news sequence, table, and program information respectively. Split the table subgraph into multiple meta-graph networks, and immerse-connect the table subgraph with the news subgraph and the program subgraph respectively to form a tight news-table association graph and a program-table association graph network;

[0013] S3: Local and Global Inference Modeling Module

[0014] Design two inference modeling mechanisms, local association modeling and global consistency modeling, to capture in-depth evidence with association and consistency respectively, and obtain deep language evidence and deep logical evidence on the tight news-table association graph and the program-table association graph respectively;

[0015] S4: Interactive Fusion Module

[0016] Design an interactive fusion layer to enhance the mutual fusion between the tight news-table association graph and the program-table association graph network, strengthen the interaction and sharing between the two evidences, so as to obtain their mutually corroborative evidences.

[0017] A further technical solution of the present invention: S2 includes the following steps:

[0018] S21: For the news sequence, convert the news sequence into a news dependency subgraph in the form of a dependency parsing tree; for the table information, each cell node in the table content is connected not only to the header of its column but also to other cells in its row, so as to construct a table subgraph; for the program, use the LPA algorithm for program synthesis and selection, and finally select the top 3 synthesized optimal programs as the program subgraph;

[0019] S22: Adopt a table decomposition mechanism to split the table subgraph. The table decomposition mechanism combines the header row and each row of data in the table body to form a two-row tight table, and then connects each row and each column to construct a meta-graph network; for an n-row table, it can be decomposed into n - 1 meta-graph networks;

[0020] S23: The nodes of the news subgraph or the program subgraph are connected not only to the nodes with the same content in these meta-graph networks but also to the corresponding header nodes, forming a tight news-table association graph and a program-table association graph network.

[0021] A further technical solution of the present invention: S3 includes the following steps:

[0022] S31: Use an association-aware GAN to model the association relationship between nodes in the subgraph. Given a subgraph, its adjacent nodes are {v1, v2, …, vI}, where I represents the number of adjacent nodes; the graph attention mechanism is used to measure the association degree α between nodes ij , which can be calculated as i and v j The weight of the edge between:

[0023]

[0024] Where f(·) is a LeakyReLU activation function, and w and W are trainable parameters;

[0025] S32: Node v i is encoded into a new representation with associated information

[0026]

[0027] Among them, W α With b α are all trainable parameters, l is the number of layers of the graph network;

[0028] S33: Given a boundary node v a and the adjacent node set {v1,v2,…,v I}, the global inter-graph consistency can be formulated as sharing the same boundary node v a Node set tend to have similar consistency characteristics, I′ represents the a The number of neighboring nodes;

[0029] S34: Use graph attention network to measure the consistency β between two nodes ia , which can be calculated as the boundary node v a and neighbor node v i The edge weights between:

[0030]

[0031] Among them, f(·) is the activation function and W is a trainable parameter;

[0032] S35: Neighbor node v i can be encoded into a new node representation with shared consistency information:

[0033]

[0034] in, Yes a The representation feature can be obtained through the neighbor node v iObtain weights by encoding the node consistency between different subgraphs; both W and b are trainable parameters;

[0035] S36: Since nodes may exist inside and between subgraphs, to cover all possibilities, the outputs of two inference modeling mechanisms are integrated, that is

[0036] S37: For the i-th node, the deep evidence semantics inferred by the inference modeling module is language evidence on the tight news-table association graph and logical evidence on the tight program-table association graph

[0037] S38: Use the gated attention mechanism on the two association graphs to obtain the deep language evidence v′ s and the deep logical evidence v′ p .

[0038] A further technical solution of the present invention: S4 includes the following steps:

[0039] S41: Input the deep language evidence v′ s and the deep logical evidence v′ p into two BiLSTM modules respectively. The outputs of the BiLSTM modules are respectively denoted as m′ s and m′ p ;

[0040] S42: Use the attention mechanism to match the significant features between the two evidences. For the j-th word v′ pj of the deep logical evidence, the following calculation is obtained:

[0041]

[0042]

[0043] where and are respectively regarded as the original association degree and the regularized association degree measurement between the j-th word in the logical evidence and the news sequence;

[0044] S43: Obtain the language-guided logical evidence o′ p by the following formula:

[0045]

[0046] S44: First combine o′ p with the encoded logical evidence m′ p , and input them into the fully connected layer to ensure the consistency representation between the two evidences:

[0047]

[0048] F′ sp = MLP(m′ s ; n′ p ) (9)

[0049] where n′ p is a new synthetic representation for logical evidence, and is the sum between elements;

[0050] S45: Predict the probability distribution and perform minimum cross - entropy training based on the true label y:

[0051] p = softmax(W sp F′ sp + b sp ) (10)

[0052] loss = -∑ylogp (11).

[0053] A computer system, characterized by comprising: one or more processors, and a computer - readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above - mentioned method.

[0054] A computer - readable storage medium, characterized by storing computer - executable instructions, which are used to implement the above - mentioned method when executed.

[0055] A computer program, characterized by comprising computer - executable instructions, which are used to implement the above - mentioned method when executed.

[0056] Beneficial effects

[0057] The present invention is a method for fact - checking and verification for tables based on graph - structure reasoning. Aiming at the defects in the research of false news detection under tabular data conditions, a tabular fact - verification method based on a tight graph reasoning network (TGRN) is proposed. It constructs a tightly - associated graph network to establish tight connections between different sub - graphs, and constructs an inference and modeling mechanism from local and global perspectives to mine the mutually - supporting evidence semantics between language and logical evidence for fact - checking and verification work based on structured tables. The present invention not only improves the performance of false news detection, but also provides effective evidence to realize the interpretability of the detection results of structured data. Compared with the prior art, the present invention has the following innovations:

[0058] 1: The tight graph constructed by the present invention exposes more available nodes connected to other sub - graphs, and these nodes can be more tightly connected to different sub - graphs, alleviating the sparsity of sub - graphs;

[0059] 2: The two inference modeling layers designed by the present invention from local and global perspectives can learn deep evidence through the correlation modeling within subgraphs and the consistency modeling between subgraphs.

[0060] 3: The graph construction method constructed by the present invention - the tight graph network and the proposed novel inference modeling mechanism are organically integrated, and the evidence that corroborates each other between linguistic evidence and logical evidence can effectively improve the task performance.

[0061] The extensive experiments of the present invention on two competitive fake news detection datasets have confirmed the effectiveness and interpretability of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The drawings are only for the purpose of illustrating specific embodiments and are not considered as limiting the present invention. Throughout the drawings, the same reference signs denote the same components.

[0063] Figure 1 is the architecture diagram of the present invention;

[0064] Figure 2 is the experimental performance diagram of the present invention under two datasets of TABFACT and FEVEROUS;

[0065] Figure 3 is the separation performance comparison diagram of different modules of the present invention under two datasets of TABFACT and FEVEROUS. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0066] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0067] The present invention proposes a table fact verification method (TGRN) based on a tight graph reasoning network, which mainly captures collaborative evidence from two perspectives: language and logic for fact-checking work. Specifically, first, to solve the problem of severe sparsity, the present invention innovatively constructs a tight-fitting graph network (TF graph). It first divides a complete table into multiple meta-graph networks composed of two rows of content, thus exposing more nodes connected to other sub-graphs. Then, when other sub-graphs are connected to the entire table, the elements in other sub-graphs are connected not only to the cell nodes with the same content in the meta-graph network of the table but also to the corresponding header nodes of the cells. Second, the present invention proposes an inference modeling layer from two perspectives: local and global. It uses intra-sub-graph association modeling and inter-sub-graph consistency modeling to learn deep evidence, and captures deep language evidence and deep logic evidence on the tight news-table association graph and the program-table association graph respectively. Finally, the present invention designs an interaction fusion layer to strengthen the interaction and sharing between the two pieces of evidence to obtain their mutually corroborative evidence. The architecture diagram of the present invention is as shown in Figure 1 shown, and includes the following four modules:

[0068] Module 1: Context Encoding Module. The invention uses a pre-trained BERT model to learn the context representations of three types of information, namely news sequences, tables, and program information;

[0069] Module 2: Tight Graph Construction Module. The present invention first constructs three types of sub-graphs based on different methods, and then innovatively proposes a tight graph construction mode, which divides the table sub-graph into multiple tight meta-graph networks to strengthen the connection tightness between different sub-graphs.

[0070] Module 3: Local and Global Based Inference Modeling Module. This module designs two modeling inference mechanisms: local association modeling and global consistency modeling, to respectively mine specific language and logic evidence, and the consistency evidence between the two types of evidence.

[0071] Module 4: Interaction Fusion Module. This module enhances the mutual interaction and fusion between language evidence and logic evidence, and learns the mutually corroborative evidence between the two through encodings and attention with different structures.

[0072] The method flow of the present invention is specifically as follows:

[0073] Module 1: Context Encoding Module.

[0074] Step 1: The present invention uses a pre-trained BERT model as the basic model for context encoding. To make the input of the present invention more suitable for the BERT model, the present invention first linearly serializes a two-dimensional table into a table sequence T′ in the form of a table template, and then the news sequence S, the table sequence T′, and the function names F of the program (such as functions like eq, count, max, etc.) are input into the BERT model in such a way [[CLS]; S; [SEP]; T′; [SEP]; F; [SEP]]. Among them, [CLS] is a special command to represent the classifier, and [SEP] represents the terminator. Finally, the present invention obtains the encoding information of each basic word in the sequence as

[0075] Module 2: Dense graph construction module.

[0076] Step 2: In the dense graph construction and modeling module, the present invention first constructs three types of subgraphs, and then proposes a dense graph form fusion to fuse these three types of subgraphs.

[0077] Step 3: Subgraph construction. For news, tables, and program information, the present invention first constructs different subgraphs to depict their respective internal structures. For the news sequence, the present invention converts the news sequence into a news dependency subgraph in the form of a dependency parsing tree. For table information, each cell node in the table content is connected not only to the header of its column but also to other cells in its row, thereby constructing a table subgraph. For the program, the present invention uses the LPA algorithm for program synthesis and selection, and finally the present invention selects the top 3 optimal synthesized programs as the program subgraph.

[0078] Step 4: Dense graph construction. The traditional connection method between subgraphs is usually to directly connect the nodes with the same content in two subgraphs. Such a method has obvious defects, that is, there is a serious sparsity problem between subgraphs. For this reason, the present invention proposes a tight-fitting graph network (TF graph), as Figure 1 (ii) shown. The TF graph designs a table decomposition mechanism to split the table subgraph, and the table subgraph can be immersively connected to the news subgraph and the program subgraph respectively.

[0079] Step 5: Generally, a table consists of a table header and a table body. The table decomposition mechanism of the present invention combines each row of data in the table header row and the table body to form a small two-row compact table, and then connects each row and each column to construct a meta-graph network. For an n-row table, it can be decomposed into n-1 meta-graph networks. Next, the nodes of other subgraphs (news subgraph or program subgraph) are connected not only to the nodes with the same content in these meta-graph networks, but also to the corresponding table header nodes.

[0080] Step 6: In the manner of Step 5, the method constructs a compact news-table association graph and a program-table association graph network respectively, making the connection between subgraphs tighter and more comprehensive, and reducing the sparsity of the connection between subgraphs.

[0081] Module 3: Local and global inference modeling module.

[0082] Step 7: The present invention designs two inference modeling mechanisms of local association modeling and global consistency modeling, captures deep evidence with association and consistency respectively, and obtains deep language evidence and deep logical evidence on the compact news-table association graph and the program-table association graph respectively.

[0083] Step 8: Local association modeling. The motivation of this module is that the degree of association between nodes within a subgraph is usually stronger than that between nodes in different subgraphs. Since the TF graph adds duplicate boundary nodes between subgraphs, if we use traditional GAN to model these subgraphs, it will lead to the phenomenon that the duplicate boundary nodes between subgraphs are enhanced while the nodes inside the subgraphs are weakened. To eliminate this defect, we design a local intra-graph association model to enhance the degree of association between nodes within the subgraph, so as to obtain association evidence.

[0084] Step 9: The present invention uses an association-aware GAN to model the association relationship between nodes in a subgraph. Given a subgraph (such as a news subgraph or a table subgraph), its adjacent nodes are {v1, v2, …, v I}(where I represents the number of adjacent nodes), the present invention adopts a graph attention mechanism to measure the degree of association α ij , which can be calculated as the weight of the edge between nodes v i and v j :

[0085]

[0086] where f(·) is a LeakyReLU activation function, and w and W are trainable parameters.

[0087] Step 10: Then, node v i is encoded into a new representation with association information

[0088]

[0089] Among them, W α and b α are both trainable parameters. l is the number of layers of the graph network.

[0090] Step 11: Global consistency modeling. The present invention proposes a consistency-aware GAN to model the consistency of boundary nodes between subgraphs, aiming to consolidate information interaction and sharing between subgraphs, so as to obtain consistent evidence. Given the boundary node v a and the adjacent node set {v1, v2,..., v I}, the global inter-graph consistency can be formulated as the node sets a sharing the same boundary node v tending to have similar consistency features. Here, I′ represents the number of adjacent nodes of the node v a .

[0091] Step 12: Similarly, the present invention uses a graph attention network to measure the degree of consistency β ia between two nodes, so as to strengthen the interaction and sharing between two boundary nodes, which can be calculated as the edge weight between the boundary node v a and the neighbor node v i :

[0092]

[0093] where f(·) is an activation function and W is a trainable parameter.

[0094] Step 13: The neighbor node v i can be encoded into a new node representation with shared consistency information:

[0095]

[0096] where is the representation feature of v a , which can be weighted by encoding the node consistency degree between different subgraphs through the neighbor node v i . Both W and b are trainable parameters.

[0097] Step 14: Since nodes may exist inside and between subgraphs, in order to cover all possibilities, the present invention integrates the outputs of two inference modeling mechanisms, namely

[0098] Step 15: For the i-th node, the deep evidence semantics inferred by the inference modeling module is language evidence on the tight news-table association graph and logical evidence on the tight program-table association graph

[0099] Step 16: Finally, for all nodes, in order to aggregate the evidence information on the two tight association graphs, this module uses the gated attention mechanism on the two association graphs to obtain the language deep evidence v′ s and the logical deep evidence v′ p .

[0100] Module 4: Interactive Fusion Module.

[0101] Step 17: In order to learn the mutually corroborative evidence between the language and logical depth evidence, the present invention designs an interactive fusion layer to enhance the mutual fusion between the two TF graphs.

[0102] Step 18: This module first uses two BiLSTM modules to encode the language deep evidence v′ s and the logical deep evidence v′ p , and then represents the outputs of the BiLSTM modules as m′ s and m′ p .

[0103] Step 19: Then, this module uses the attention mechanism to match the significant features between the two evidences. For the j-th word v′ pj of the logical evidence, the following calculation can be obtained:

[0104]

[0105]

[0106] where and are respectively regarded as the original association degree and the regularized association degree measurement between the j-th word in the logical evidence and the news sequence.

[0107] Step 20: In this way, the present invention can obtain the language-guided logical evidence o′ p as:

[0108]

[0109] Step 21: In order to strengthen the effective interaction between the language deep evidence and the logical deep evidence, we first combine o′ p with the encoded logical evidence m′ p , and then input them into the fully connected layer to ensure the consistent representation between the two evidences:

[0110]

[0111] F′ sp = MLP(m′ s ; n′ p ) (9)

[0112] where n′ p is a new synthetic representation for logical evidence, and is the sum between elements.

[0113] Step 22: Finally, the present invention predicts the probability distribution and trains the model by minimizing the cross entropy based on the true label y:

[0114] p = softmax(W sp F′ sp + b sp ) (10)

[0115] loss = -∑ylogp (11)

[0116] The method of the present invention is applicable to the social network environment and can be provided in the social media network environment based on structured tabular evidence.

[0117] As described above, only the specific embodiments of the present invention are shown, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for verifying table facts based on a tight graph inference network, characterized in that: S1: Context encoding module Learn the context representations of three types of information with the help of a pre-trained BERT model, and the three types of information are news sequences, tables, and program information respectively; S2: Tight graph construction module Construct a news dependency subgraph, a table subgraph, and a program subgraph for news sequences, tables, and program information respectively. The table subgraph is divided into multiple meta-graph networks, and the table subgraph is connected to the news subgraph and the program subgraph in an immersive manner to form a tight news-table association graph and a program-table association graph network; S2 includes the following steps: S21: For news sequences, convert the news sequences into a news dependency subgraph in the form of a dependency parsing tree; for table information, each cell node in the table content is connected not only to the header of its column but also to other cells in its row, thereby constructing a table subgraph; for programs, use the LPA algorithm for program synthesis and selection, and finally select the top 3 optimal synthesized programs as the program subgraph; S22: Adopt a table decomposition mechanism to split the table subgraph. The table decomposition mechanism combines the header row and each row of data in the table body to form a two-row tight table, and then connects each row and each column to construct a meta-graph network; for an n-row table, it can be decomposed into n−1 meta-graph networks; S23: The nodes of the news subgraph or the program subgraph are connected not only to the nodes with the same content in these meta-graph networks but also to the corresponding header nodes to form a tight news-table association graph and a program-table association graph network; S3: Local and global inference modeling module Design two inference modeling mechanisms, local association modeling and global consistency modeling, to capture deep evidence with association and consistency respectively, and obtain deep language evidence and deep logical evidence on the tight news-table association graph and the program-table association graph respectively; S3 includes the following steps: S31: Use the correlation-aware GAN to model the correlation relationship between nodes in the subgraph. Given a subgraph, its adjacent nodes are , where represents the number of adjacent nodes; adopt the graph attention mechanism to measure the correlation degree between nodes , which can be calculated as the weight of the edge between node and : Among them, is a LeakyReLU activation function, and W are trainable parameters; S32: Node is encoded as a completely new representation with associated information : Among them, and are both trainable parameters, is the number of layers of the graph network; S33: Given boundary nodes and the set of adjacent nodes , the global graph consistency can be formulated as the set of nodes sharing the same boundary nodes tend to have similar consistency characteristics, denoting the number of adjacent nodes for the node; S34: Use a graph attention network to measure the consistency between two nodes , which can be calculated as the edge weight between the boundary node and the neighbor node : (3) Among them, is an activation function, and W is a trainable parameter; S35: Neighbor node can be encoded as a completely new node representation with shared consistency information: (4) Among them, is a representation feature, which can be weighted by encoding the node consistency between different subgraphs through neighbor nodes ; both W and b are trainable parameters; S36: Since nodes may exist inside and between subgraphs, in order to cover all possibilities, the outputs of two inference modeling mechanisms are integrated, namely ; S37: For the i th node, the deep evidence semantics inferred by the inference modeling module is linguistic evidence on the tight news-table association graph , and is logical evidence on the tight program-table association graph ; S38: Use the gated attention mechanism on two associated graphs to obtain deep linguistic evidence and deep logical evidence ; S4: Interactive fusion module Design an interactive fusion layer to enhance the mutual fusion between the tight news-table association graph and the program-table association graph network, strengthen the interaction and sharing between the two evidences, and obtain their mutually corroborative evidences.

2. The method for verifying tabular facts based on a tight graph inference network according to claim 1, characterized in that S4 includes the following steps: S41: Input the deep linguistic evidence and the deep logical evidence into two BiLSTM modules respectively. The respective outputs of the BiLSTM modules are denoted as and ; S42: Use the attention mechanism to match the significant features between the two evidences, and for the j nth word of the logically deep evidence, the following calculation is obtained: (5) (6) Among them, and are respectively regarded as the original and regularized relevance measure between the j th word in the logical evidence and the news sequence; S43: Obtain logical evidence of language guidance through the following formula Namely: (7) S44: First, combine with the encoded logical evidence , and input them into the fully connected layer to ensure a consistent representation between the two pieces of evidence: (8) (9) Among them, is a new synthetic representation for logical evidence, is the summation between elements; S45: Predict the probability distribution and perform minimum cross-entropy training based on the true label y: p = softmax( )(10) (11) wherein, y is the true label.

3. A computer system, characterized in that Including: One or more processors, a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in claim 1.

4. A computer-readable storage medium, characterized in that Stored with computer-executable instructions, the instructions are used to implement the method described in claim 1 when executed.

5. A computer program product, characterized in that Including computer-executable instructions, the instructions are used to implement the method described in claim 1 when executed.

Images