Interpretable evaluation method and system for machine-generated manuscripts based on graph neural network

Abstract

The present invention discloses a machine-generated manuscript interpretable evaluation method and system based on graph neural network, the method comprising: obtaining machine-generated multimodal news articles, encoding the content of the multimodal news articles, obtaining the overall coding vector of the news articles and the coding vector of the elements of the news articles; inputting the overall coding vector of the news articles and the coding vector of the elements of the news articles into a hierarchical graph neural network, constructing a heterogeneous graph of news articles, updating the overall coding vector of the news articles and the coding vector of the elements of the news articles through the convolution of the heterogeneous graph of news articles based on a multi-level attention mechanism, splicing the updated coding vectors into a news article representation matrix, and outputting interpretable evaluation label prediction results through a fully connected layer. The present invention designs multiple interpretable evaluation labels, and uses a graph neural network to aggregate the content and evaluation information of news articles, conducts a multi-level and fine-grained evaluation of the quality of the manuscripts, and provides interpretable evaluation labels for multimodal news articles.

Description

Technical Field

[0001] The present invention relates to the technical field of manuscript interpretability evaluation, and in particular to a machine-generated manuscript interpretability evaluation method and system based on graph neural network. Background Art

[0002] Manuscript interpretable evaluation can intuitively feedback the quality of manuscripts and help improve the quality of machine writing. Manuscript interpretable evaluation can alleviate the problems of low efficiency, high cost, and different evaluation results of manual evaluation. At present, manuscript interpretable evaluation has been widely used in various fields. For example, in the automatic essay scoring task, vocabulary and semantic overlap are used to evaluate whether an essay is closely related to the topic, and sentence structure features are used to measure the diversity of sentences; in the Wikipedia quality grading task, it is necessary to consider whether the manuscript organization is correct, whether there is supporting evidence, and whether there are appropriate terms; in the community question and answer quality evaluation, the quality of question and answer is evaluated by using features from different sources such as the answer content itself, the profile of the answerer, the interaction between users, and the use of content. The use of manuscript interpretable evaluation can improve the quality of manuscripts and enhance the user reading experience. As people begin to rely on social platforms to obtain manuscript content, in order to ensure the quality of manuscripts and user experience, academia and industry have conducted extensive research on the methods and standards for manuscript quality evaluation.

[0003] Among the methods for evaluating the interpretability of manuscripts, existing methods are based on technologies such as attention neural networks, bidirectional encoder representation models (BERT), and generative pre-trained models (GUID Partition Table, GPT), which score and evaluate the manuscript as a whole. However, this existing objective evaluation method cannot give an interpretability evaluation for a certain paragraph or a certain position in the text, and there is little interpretable feedback information; moreover, the existing evaluation methods for multi-modal news articles rely heavily on manual labor, have little annotated data, and lack interpretable evaluation labels. Summary of the invention

[0004] In order to address the deficiencies of the above-mentioned prior art, the present invention provides an interpretable evaluation method and system for machine-generated manuscripts based on graph neural networks. According to the common problems of machine-generated news and methods of particular concern in news evaluation, a plurality of interpretable evaluation labels are designed, and graph neural networks are used to aggregate news article content and evaluation information, and a multi-level and fine-grained evaluation of manuscript quality is performed. Interpretable evaluation labels for multi-modal news articles are given, and quality problems in news articles are intuitively fed back, and revision suggestion labels are proposed for low-quality news articles.

[0005] In a first aspect, the present disclosure provides a method for interpretable evaluation of machine-generated manuscripts based on graph neural networks.

[0006] A machine-generated manuscript interpretable evaluation method based on graph neural network, including:

[0007] Obtain machine-generated multimodal news articles, encode the contents of the multimodal news articles, and obtain the overall coding vector of the news articles and the coding vector of the elements of the news articles;

[0008] The overall coding vector of the news article and the coding vector of the elements of the news article are input into the hierarchical graph neural network to construct a heterogeneous graph of news articles. The overall coding vector of the news article and the coding vector of the elements of the news article are updated through the news article heterogeneous graph convolution based on the multi-level attention mechanism. The updated coding vectors are spliced ​​into a news article representation matrix, and the final interpretable evaluation label prediction results are output through the fully connected layer.

[0009] In a further technical solution, the construction and training process of the hierarchical graph neural network includes:

[0010] Using multiple multimodal news articles as training sets, the contents of the news articles in the training sets are encoded to obtain the overall coding vector of the news articles and the coding vector of the elements of the news articles;

[0011] Label the news articles in the training set with interpretable evaluation labels, and encode the labeled interpretable evaluation labels to obtain the machine evaluation label encoding vector;

[0012] Based on the overall coding vector of the news article, the coding vector of the elements of the news article, and the coding vector of the machine evaluation label, a hierarchical graph neural network based on a multi-level attention mechanism is constructed and trained. The coding vector is input into the hierarchical graph neural network, and the updated coding vector of the news article content and the coding vector of the elements of the news article are output;

[0013] Based on the updated encoding vector, the news article representation vector is constructed by vector concatenation, and the news article representation vector is passed through a fully connected layer to output the final interpretable evaluation label prediction result;

[0014] Based on the annotated interpretable evaluation labels and the errors in label prediction results, a cross entropy loss function is constructed, and the network parameters are updated using the back propagation algorithm until the hierarchical graph neural network training is completed.

[0015] In a second aspect, the present disclosure provides a machine-generated manuscript explainable evaluation system based on graph neural networks.

[0016] A machine-generated manuscript explainable evaluation system based on graph neural network, including:

[0017] The article acquisition module is used to obtain machine-generated multimodal news articles;

[0018] A manuscript encoding module is used to encode the content of multimodal news articles and obtain the overall coding vector of the news article and the coding vector of the elements of the news article;

[0019] The interpretable evaluation label prediction module is used to input the overall coding vector of the news article and the coding vector of the elements of the news article into the hierarchical graph neural network, construct a heterogeneous graph of the news article, update the overall coding vector of the news article and the coding vector of the elements of the news article through the news article heterogeneous graph convolution based on the multi-level attention mechanism, splice the updated coding vectors into a news article representation matrix, and output the final interpretable evaluation label prediction results through the fully connected layer.

[0020] In a third aspect, the present disclosure further provides an electronic device, comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the steps of the method described in the first aspect are completed.

[0021] In a fourth aspect, the present disclosure further provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the method described in the first aspect.

[0022] One or more of the above technical solutions have the following beneficial effects:

[0023] 1. The present invention provides an interpretable evaluation method and system for machine-generated manuscripts based on graph neural networks. According to common problems in machine-generated news and methods of particular concern in news evaluation, multiple interpretable evaluation labels are designed, and graph neural networks are used to aggregate news article content and evaluation information, and multi-level and fine-grained evaluations of manuscript quality are performed. Interpretable evaluation labels for multi-modal news articles are given, quality issues in news articles are intuitively fed back, and revision suggestion labels are proposed for low-quality news articles.

[0024] 2. This paper applies graph convolutional networks to interpretable evaluation of news articles, proposes a new interpretable evaluation training framework for news articles, and also proposes ten interpretable evaluation labels. Compared with existing interpretable evaluation methods for news articles, this paper breaks through the limitation of the fuzzy standards of traditional subjective evaluation methods and alleviates the problem that news article quality evaluation is artificial and difficult to quantify.

[0025] 3. The present invention can be applied to multimodal manuscripts, which can help improve the quality of manuscript writing, improve the efficiency of news creation, accelerate news dissemination, and play an important role in better realizing the supervision of news content. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0027] Figure 1 A schematic diagram of a flow chart of a method for interpretable evaluation of machine-generated manuscripts based on a graph neural network according to an embodiment of the present invention;

[0028] Figure 2 Schematic diagram of an interpretable evaluation framework for news articles in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0030] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. It should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0031] Embodiment 1

[0032] In order to solve the problems of limited interpretable feedback information in traditional objective evaluation methods, heavy reliance on manual evaluation of multi-modal news articles, and limited annotated data, this embodiment provides a machine-generated interpretable evaluation method for articles based on graph neural networks, such as Figure 1 As shown, the specific steps include:

[0033] Obtain machine-generated multimodal news articles, encode the contents of the multimodal news articles, and obtain the overall coding vector of the news articles and the coding vector of the elements of the news articles;

[0034] The overall coding vector of the news article and the coding vector of the elements of the news article are input into the hierarchical graph neural network to construct a heterogeneous graph of news articles. The overall coding vector of the news article and the coding vector of the elements of the news article are updated through the news article heterogeneous graph convolution based on the multi-level attention mechanism. The updated coding vectors are spliced ​​into a news article representation matrix, and the final interpretable evaluation label prediction results are output through the fully connected layer.

[0035] In the machine-generated manuscript interpretability evaluation method based on graph neural network described in this embodiment, a hierarchical graph neural network based on a multi-level attention mechanism is constructed and trained, and the graph neural network is used to predict the interpretability evaluation label of the machine-generated manuscript. In view of the construction and training of the above-mentioned graph neural network, this embodiment builds a news article interpretability evaluation framework, such as Figure 2 As shown in the figure, it mainly includes four aspects: news article content coding, news article evaluation coding, news article collaborative representation and explainable evaluation prediction. Among them, news article content coding realizes the encoding of news article multimodal content, including news article overall coding and news article element coding; news article evaluation coding realizes the encoding of machine evaluation labels; news article collaborative representation is used to construct a heterogeneous graph integrating machine evaluation labels, and aggregates article element information based on a graph neural network with a multi-level attention mechanism; explainable evaluation label prediction realizes the prediction output of the final article evaluation label.

[0036] In this embodiment, the definition of news articles can explain the input and output of the evaluation framework: D = {d 1 ,d 2 ,...,d |D|} represents the collection of news release documents, let d i ={T,C} represents a news article document, where T = {t 1 ,t 2 ,...,t |T|} represents a document paragraph, C = {c 1 ,c 2 ,...,c |C|} represents the image contained in the document. Given any news article document d i ={T,C}, through the above news article interpretable evaluation framework, the interpretable evaluation label of the news article is predicted. This process can be formally defined as:

[0037] P(y j |d i )~f(d i ) (1)

[0038] Among them, P(Y|d i ) represents the probability that the i-th news article document is predicted to be the j-th interpretable label, f(d i ) is a function for modeling this probability, and this embodiment implements the construction of this function through the above-mentioned framework or network.

[0039] Combination Figure 2 The news release shown can explain the evaluation framework and introduce in detail the construction and training process of the hierarchical graph neural network described in this embodiment.

[0040] First, step S1, news article content encoding. Specifically, multiple multimodal news articles are used as a training set, and the news article content in the training set is encoded to obtain the news article overall encoding vector and the news article element encoding vector.

[0041] The overall coding of news releases includes the following coding steps:

[0042] Step S1.1: In order to obtain the representation information of the multimodal news article, this embodiment is based on the multimodal pre-training model of UNIMO (unified modality pre-training framework), and encodes the text modality information and image modality information in the news article in the same semantic space to obtain the text representation matrix R 1 And the picture representation matrix R 2 ;

[0043] Step S1.2: In order to further extract the contextual cross-modal information of the news release, the UNIMO-encoded text and image representation matrices are aggregated through the attention mechanism, and the aggregated features are filtered using the multiplication filtering mechanism to finally obtain the cross-modal contextual representation matrix of the text and image.

[0044] First, we represent the matrix R by text 1 And the picture representation matrix R 2 Multiply each other to get cross-modal information, that is, get text cross-modal information M 1 and image cross-modal information M 2 , as shown in the following formula (2)(3):

[0045]

[0046]

[0047] Secondly, in order to obtain contextual dependencies, based on the text cross-modal information M 1 and image cross-modal information M 2 , use the softmax function to calculate the attention distribution N of text and pictures 1 and N 2 , the calculation process is shown in the following formula (4)(5):

[0048] N 1 =SoftMax(M 1 , dim=1) (4)

[0049] N 2 =SoftMax(M 2 , dim=1) (5)

[0050] Subsequently, in the text we represent the matrix R 1 And the picture representation matrix R2 Apply the attention mechanism to calculate the attention representation information O on the modality 1 and O 2 , the calculation process is shown in the following formula (6):

[0051] O 1 =N 1 ·R 2 (6)

[0052] O 2 =N 2 ·R 1 (7)

[0053] Next, a multiplication filtering mechanism is introduced to focus on the important components of multimodality and discourse. The calculation process is shown in the following formula. The final filtered features are calculated to obtain the cross-modal context representation matrix A 1 and A 2 .

[0054] A 1 =O 1 ☉R 1 (8)

[0055] A 2 =O 2 ☉R 2 (9)

[0056] Step S1.3: Transform the cross-modal context matrix A of text and image 1 and A 2 After concatenation, it is used as the output of the news article content encoding module and combined with the text representation matrix R 1 And the picture representation matrix R 2 The splicing is used as the overall coding vector of the news release to complete the overall coding of the news release.

[0057] Regarding the coding of news article elements, the coding method of the elements is slightly different from the overall coding of news articles. When coding the elements of news articles, the news articles are first split into five parts, namely the title, introduction, main text, and ending, as well as the news picture part. Each part is independently regarded as a news article element, and each part is unimodal (text or picture). Since the UNIMO pre-training model can adapt to both unimodal and multimodal tasks, in this embodiment, a unimodal pre-training model based on UNIMO (unified modality pre-training framework) is used to encode the text modal information or picture modal information of each news article element in the semantic space, and obtain the representation matrix of each part, that is, obtain the coding vector of each news article element.

[0058] Step S2: News article evaluation encoding: Specifically, the news articles in the training set are labeled with interpretable evaluation labels, and the labeled interpretable evaluation labels are encoded to obtain a machine evaluation label encoding vector.

[0059] The task of the news article evaluation coding part is to complete the coding of the machine evaluation label. In this embodiment, the subjective and objective score consistency modeling is first performed, and the multimodal news articles are machine evaluated through this model (the evaluation results include the names of objective evaluation indicators and their corresponding scores). The obtained machine evaluation has a unified subjective and objective score, so that the subsequent annotated evaluation labels are accurate. Based on the machine evaluation results, the news articles are annotated with interpretable evaluation labels. In fact, the machine evaluation labels are combined with the objective evaluation indicator names and their scores, and converted into text descriptions. For example, when the hierarchy clarity index score is low, the corresponding machine evaluation label is "poor hierarchy clarity".

[0060] In this embodiment, considering that news articles have distinct characteristics compared to other types of writing styles, it advocates the use of plain language to accurately and efficiently convey information, and has high requirements in terms of clarifying the theme and expression logic. In addition, the problems that are prone to occur in machine-generated texts also have obvious characteristics, such as plot repetition, logical conflicts, and incoherence between sentences. Based on the above two points, this embodiment specifically proposes ten manuscript quality evaluation labels, which are used for manual revision opinion labels during data set construction and interpretable evaluation labels ultimately output by the network model. The ten manuscript quality evaluation labels proposed in this embodiment are shown in Table 1 below, including title inconsistent with content, unclear theme, unclear hierarchy, logical errors, unsmooth expression, incomplete information, unsuccinct expression, non-compliant content, poor image quality, and inconsistency between images and texts.

[0061] Table 1 News release quality evaluation labels

[0062]

[0063] Using the pre-trained model to encode the above machine evaluation labels can better utilize the semantic understanding ability of the pre-trained model. The machine evaluation labels are still encoded using the pre-trained model based on UNIMO, as shown in the following formula:

[0064] H i =UNIMO(h i ),i=1,...,n (10)

[0065] Among them, H i Encode the machine evaluation label vector, h i is the text description of the machine evaluation label, such as "the title does not match the content", etc., and i is the number of machine evaluation labels.

[0066] Step S3, collaborative representation of news articles. Specifically, a hierarchical graph neural network based on a multi-level attention mechanism is constructed, and the hierarchical graph neural network is trained using the overall coding vector of the news article, the coding vector of the news article elements, and the coding vector of the machine evaluation label. The above coding vectors are input into the hierarchical graph neural network, and the updated news article content coding vector and news article element coding vector are output. The updated coding vector is the collaborative representation of the news article. Step S3 specifically includes:

[0067] Step S3.1: Use the news article overall coding vector, news article element coding vector and machine evaluation label coding vector as news article overall nodes, news article element nodes and machine evaluation label nodes to construct a news article heterogeneous graph. Specifically, the strategy or rule for constructing a heterogeneous graph integrating machine evaluation labels is as follows:

[0068] In the graph representation G = (υ, ε), υ is a node set and ε is an edge set. For the heterogeneous graph G to be constructed in this embodiment, the node set can be represented as υ = {V, X, Y}, where V represents the overall node set of news articles, X represents the node set of news article elements, and Y represents the node set of machine evaluation labels; the edge set ε can be represented as Among them, (v i ,v j ) i,j∈[1,N] is the edge between the news article overall node and the news article overall node, is the edge between news article element nodes, (y i ,y j ) i,j∈[1,k] is the edge between machine evaluation label nodes and machine evaluation label nodes, (v i ,y j ) i∈[1,N],j∈[1,k] is the edge between the news article node and the machine evaluation label node, It is the edge between the news article element node and the machine evaluation label node. The following describes the determination of the weights of different edges.

[0069] 1) Edge weights between news article nodes and news article nodes

[0070] For the news article overall node, the news article overall coding vector is used as the node representation. The news article overall coding vector is the initial node embedding of the node. The weight calculation method of the edge formed by two news article overall nodes is shown in the following formula (11), that is, there is no edge connecting different news article overall nodes.

[0071]

[0072] 2) Edge weights between news article element nodes

[0073] For news article element nodes, the news article element encoding vector is used as the node representation, and the news article element encoding vector is the initial node embedding of the node. When two news article elements come from the same news article, the edge weight between them is 1, and in other cases the edge weight is 0. The calculation method is shown in the following formula (12).

[0074]

[0075] 3) Edge weights between machine evaluation label nodes and machine evaluation label nodes

[0076] For a machine evaluation label node, the machine evaluation label encoding vector is used as the node representation, and the machine evaluation label encoding vector is the initial node embedding of the node. The edge weight calculation method formed by two machine evaluation label nodes is shown in the following formula (13), that is, there is no edge connecting different machine evaluation label nodes.

[0077]

[0078] 4) News article overall node and machine evaluation label node

[0079] Machine evaluation labels play an important role in predicting explainable evaluation labels. Therefore, an edge consisting of the news article overall node and the machine evaluation label node is set. The weight calculation method is as follows: when the machine evaluation label corresponding to the news article overall node is the same as the machine label node, the weight value is 1, and in other cases the edge weight value is 0, as shown in formula (14).

[0080]

[0081] Among them, f(v i ) represents the overall node v of the i-th news article i The corresponding machine evaluation label.

[0082] 5) News article element nodes and machine evaluation label nodes

[0083] Slightly different from the previous four edge constructions, when determining the edge weights between news article element nodes and machine evaluation label nodes, expert experience is used as prior knowledge to establish a connection between the machine evaluation label and the relevant news article elements. For example, a news article is evaluated by the subjective and objective scoring consistency model as a low-quality machine evaluation label with a blurred image. Therefore, this label is irrelevant to the news article elements related to the text, and only an edge is constructed between the news article element node related to the image and the machine evaluation label node. The details are shown in the following formula (15).

[0084]

[0085] According to the above heterogeneous graph construction rules integrating machine evaluation labels, a news article heterogeneous graph is constructed, which includes the overall heterogeneous graph of news articles and the heterogeneous graph of news article elements. The above construction of the heterogeneous graph is actually equivalent to constructing a hierarchical graph neural network, which includes the overall news article evaluation layer and the news article element evaluation layer, wherein the overall news article evaluation layer includes the overall news article nodes and the machine evaluation label nodes, and the news article element evaluation layer includes the news article element nodes and the machine evaluation label nodes, and the machine evaluation label nodes in the two layers share the same code.

[0086] Step S3.2: Based on the heterogeneous graph constructed above, a multi-level attention mechanism (HGAT) for heterogeneous graphs is introduced to calculate the heterogeneous graph attention at the type level and the node level respectively.

[0087] Among them, the calculation method of the type-level heterogeneous graph attention weight includes: the type-level attention mechanism uses the node type as a unit to participate in the attention calculation, that is, to assign weights to different node types. Taking node v in a heterogeneous graph network as an example, the node embedding representation (i.e., node feature representation) of node v is denoted as h v First, we need to calculate the embedding representation of the nodes adjacent to it and whose node type is τ That is, sum the node feature vectors (i.e., node feature representations) that meet the conditions, where v' represents all nodes except node v, and h v' A represents the node feature representation of node v'; vv' Indicates whether the node v and the node v' are 1. The matrix A is the matrix calculated by formula (15). When the value of the matrix A is 1, the eigenvector of the node v' is considered in the summation calculation. Indicates A vv' Then, we use the feature representation h of node v to represent v The node embedding representation h with type τ τ , the type-level attention score is calculated according to formula (16):

[0088]

[0089] Among them, σ(·) is the activation function, is the attention vector of node type τ, and || is the concatenation operation.

[0090] Afterwards, the type-level attention score a τ Input into the softmax function to calculate the attention weights a of different node categories τ, the calculation method is shown in the following formula (17):

[0091]

[0092] The calculation method of the attention weight of the node-level heterogeneous graph includes: the node-level attention is based on the type-level attention, and weights are assigned to all nodes of different types adjacent to the central node. For a central node v of type τ in the heterogeneous graph, let its neighbor node set be N v , then the node v and the node v'(v'∈N v ) can be calculated according to the following formula (18):

[0093] b vv' =σ(V T α τ' [h v ||h v' ]) (18)

[0094] Among them, σ(·) is the activation function, V is the node-level attention vector, and α τ' is the type-level attention weight of node v and node v'.

[0095] Afterwards, the type-level attention score b vv‘ Input into the softmax function to calculate the node-level attention weight β vv' , the calculation method is shown in the following formula (19):

[0096]

[0097] Step S3.3: Based on the attention weights at the type level and the node level, combined with the node embedding representation of each node in the heterogeneous graph, a multi-layer attention mechanism operation is performed to update the node embedding representation of each node in the heterogeneous graph, that is, to update the news release content encoding vector and the news release element encoding vector, and finally output the updated news release content encoding vector and news release element encoding vector.

[0098] Specifically, the multi-level attention mechanism is applied to heterogeneous graph convolution, and the calculation method is shown in the following formula:

[0099]

[0100] Among them, H (l+1) is the feature matrix of the l+1th layer of HGAT, is the node feature matrix of type τ at the lth layer of HGAT, is the weight matrix used by the node of type τ in the lth layer of HGAT. The weight matrix is ​​the parameter to be learned and is continuously updated during the training process. τ is the attention matrix, β vv' is the attention matrix B τ The element of the vth row and v'th column in , σ(·) is the activation function. The feature matrix H calculated by the above formula (20) (l+1) Represents the updated encoding vector.

[0101] Step S4: Based on the updated news article content encoding vector and news article element encoding vector, a news article representation vector is constructed by vector splicing, and the news article representation vector is passed through a fully connected layer to output the final interpretable evaluation label prediction result.

[0102] Furthermore, based on the annotated interpretable evaluation labels and the errors in label prediction results, a cross entropy loss function is constructed, and the network parameters are updated using the back propagation algorithm until the multi-level graph neural network training is completed.

[0103] In this embodiment, the prediction of the explainable evaluation label requires the use of the overall coding vector of the news article and the coding vector of the news article elements updated by the hierarchical graph neural network. First, the element coding vectors belonging to the same news article (the coding vector is the updated coding vector) need to be average pooled, as shown in the following formula (21):

[0104]

[0105] Among them, e' i is a coding vector of a manuscript element in the same news article, n is the number of news article elements, and e' is the coding vector of the news article element after the average pooling of all the coding vectors of the manuscript elements in the same news article.

[0106] Then, the calculated news article element encoding vector e' is concatenated with the news article overall encoding vector e to obtain the news article representation vector e c , as shown in the following formula (22):

[0107] e c =concat(e,e') (22)

[0108] The representation of the news article overall node and news article element node in the hierarchical graph neural network is calculated through the above (22) and then concatenated into the news article representation matrix E c , input a fully connected layer with 10 output nodes, and use the softmax layer to output the prediction matrix Z of the news article's interpretable evaluation labels, as shown in the following formula (23):

[0109] Z = softmax(WE c +b) (23)

[0110] Among them, W is the learnable weight matrix and b is the bias term.

[0111] In this embodiment, the cross entropy loss function is used to obtain the error between the label prediction result and the annotated interpretable evaluation label, as shown in the following formula (24), and the back propagation algorithm is used to update the parameters of the network model until the multi-level graph neural network training is completed.

[0112]

[0113] Among them, N is the number of news articles, C is the number of explainable evaluation label categories, and Y is the explainable evaluation label matrix.

[0114] In this embodiment, based on the above-mentioned trained hierarchical graph neural network, the obtained machine-generated multimodal news articles are encoded, and the news article encoding vector is input into the trained hierarchical graph neural network to output the final interpretable evaluation label prediction result, which can intuitively feedback the quality problems in the news articles and propose revision suggestion labels for low-quality news articles, help improve the quality of writing, improve the efficiency of news creation, accelerate news dissemination, and play an important role in better realizing news content supervision.

[0115] Embodiment 2

[0116] This embodiment provides a machine-generated manuscript interpretability evaluation system based on a graph neural network, including:

[0117] The article acquisition module is used to obtain machine-generated multimodal news articles;

[0118] A manuscript encoding module is used to encode the content of multimodal news articles and obtain the overall coding vector of the news article and the coding vector of the elements of the news article;

[0119] The interpretable evaluation label prediction module is used to input the overall coding vector of the news article and the coding vector of the elements of the news article into the hierarchical graph neural network, construct a heterogeneous graph of the news article, update the overall coding vector of the news article and the coding vector of the elements of the news article through the news article heterogeneous graph convolution based on the multi-level attention mechanism, splice the updated coding vectors into a news article representation matrix, and output the final interpretable evaluation label prediction results through the fully connected layer.

[0120] Embodiment 3

[0121] This embodiment provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps in the method for interpretable evaluation of machine-generated manuscripts based on graph neural networks as described above are completed.

[0122] Embodiment 4

[0123] This embodiment also provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps in the machine-generated manuscript interpretable evaluation method based on graph neural network as described above are completed.

[0124] The steps involved in the above embodiments 2 to 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0125] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0126] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A machine-generated manuscript interpretable evaluation method based on graph neural network, characterized by: include: Obtain machine-generated multimodal news articles, encode the contents of the multimodal news articles, and obtain the overall coding vector of the news articles and the coding vector of the elements of the news articles; The overall coding vector of the news article and the coding vector of the elements of the news article are input into the hierarchical graph neural network to construct a news article heterogeneous graph. The overall coding vector of the news article and the coding vector of the elements of the news article are updated through the news article heterogeneous graph convolution based on the multi-level attention mechanism. The updated coding vectors are spliced ​​into a news article representation matrix, and the final interpretable evaluation label prediction results are output through the fully connected layer. The construction and training process of the hierarchical graph neural network includes: Using multiple multimodal news articles as training sets, the contents of the news articles in the training sets are encoded to obtain the overall coding vector of the news articles and the coding vector of the elements of the news articles; Label the news articles in the training set with interpretable evaluation labels, and encode the labeled interpretable evaluation labels to obtain the machine evaluation label encoding vector; Based on the overall coding vector of the news article, the coding vector of the elements of the news article, and the coding vector of the machine evaluation label, a hierarchical graph neural network based on a multi-level attention mechanism is constructed and trained. The coding vector is input into the hierarchical graph neural network, and the updated coding vector of the news article content and the coding vector of the elements of the news article are output; Based on the updated encoding vector, the news article representation vector is constructed by vector concatenation, and the news article representation vector is passed through a fully connected layer to output the final interpretable evaluation label prediction result; Based on the annotated interpretable evaluation labels and the errors of label prediction results, a cross entropy loss function is constructed, and the network parameters are updated using the back propagation algorithm until the hierarchical graph neural network training is completed; The encoding vector is input into the hierarchical graph neural network, and the updated news article content encoding vector and news article element encoding vector are output, including: The news article overall coding vector, news article element coding vector and machine evaluation label coding vector are used as news article overall nodes, news article element nodes and machine evaluation label nodes to construct a news article heterogeneous graph; A multi-level attention mechanism of heterogeneous graph is introduced into the news article heterogeneous graph, and the attention of heterogeneous graph is calculated at the type level and node level respectively to obtain the attention weights at the type level and node level. Based on the attention weights at the type level and the node level, combined with the node embedding representation of each node in the heterogeneous graph, a multi-layer attention mechanism operation is performed to update the node embedding representation of each node in the heterogeneous graph, that is, to update the news article content encoding vector and the news article element encoding vector.

2. The method for evaluating the interpretability of machine-generated manuscripts based on graph neural networks as claimed in claim 1, characterized in that: Encode the content of multimodal news articles to obtain the overall coding vector of the news articles, including: Based on the multimodal pre-training model of UNIMO, the text modal information and image modal information in the news release are encoded in the same semantic space to obtain the text representation matrix and the image representation matrix respectively; The text representation matrix and the image representation matrix are aggregated through the attention mechanism, and the aggregated features are filtered using the multiplication filtering mechanism to finally obtain the cross-modal context representation matrix of the text and the image; After concatenating the cross-modal context representation matrices of text and images, they are concatenated with the text representation matrix and the image representation matrix to obtain the overall encoding vector of the news release.

3. The method for evaluating the interpretability of machine-generated manuscripts based on graph neural networks as claimed in claim 1, characterized in that: The graph neural network includes a news article overall evaluation layer and a news article element evaluation layer; The news article overall evaluation layer includes news article overall nodes and machine evaluation label nodes, and the news article element evaluation layer includes news article element nodes and machine evaluation label nodes. The machine evaluation label nodes in the two layers share coding.

4. The method for evaluating the interpretability of machine-generated manuscripts based on graph neural networks as claimed in claim 1, characterized in that: The explainable evaluation labels include title inconsistent with content, unclear theme, unclear structure, logical errors, unsmooth expression, incomplete information, unsuccinct expression, non-compliant content, poor picture quality, and inconsistency between pictures and text.

5. A machine-generated manuscript interpretable evaluation system based on graph neural network, characterized by: include: The article acquisition module is used to obtain machine-generated multimodal news articles; A manuscript encoding module is used to encode the content of multimodal news articles and obtain the overall coding vector of the news article and the coding vector of the elements of the news article; The interpretable evaluation label prediction module is used to input the overall coding vector of the news article and the coding vector of the elements of the news article into the hierarchical graph neural network, construct a heterogeneous graph of news articles, update the overall coding vector of the news article and the coding vector of the elements of the news article through the news article heterogeneous graph convolution based on the multi-level attention mechanism, splice the updated coding vectors into the news article representation matrix, and output the final interpretable evaluation label prediction result through the fully connected layer; The construction and training process of the hierarchical graph neural network includes: Using multiple multimodal news articles as training sets, the contents of the news articles in the training sets are encoded to obtain the overall coding vector of the news articles and the coding vector of the elements of the news articles; Label the news articles in the training set with interpretable evaluation labels, and encode the labeled interpretable evaluation labels to obtain the machine evaluation label encoding vector; Based on the overall coding vector of the news article, the coding vector of the elements of the news article, and the coding vector of the machine evaluation label, a hierarchical graph neural network based on a multi-level attention mechanism is constructed and trained. The coding vector is input into the hierarchical graph neural network, and the updated coding vector of the news article content and the coding vector of the elements of the news article are output; Based on the updated encoding vector, the news article representation vector is constructed by vector concatenation, and the news article representation vector is passed through a fully connected layer to output the final interpretable evaluation label prediction result; Based on the annotated interpretable evaluation labels and the errors of label prediction results, a cross entropy loss function is constructed, and the network parameters are updated using the back propagation algorithm until the hierarchical graph neural network training is completed; The encoding vector is input into the hierarchical graph neural network, and the updated news article content encoding vector and news article element encoding vector are output, including: The news article overall coding vector, news article element coding vector and machine evaluation label coding vector are used as news article overall nodes, news article element nodes and machine evaluation label nodes to construct a news article heterogeneous graph; A multi-level attention mechanism of heterogeneous graph is introduced into the news article heterogeneous graph, and the attention of heterogeneous graph is calculated at the type level and node level respectively to obtain the attention weights at the type level and node level. Based on the attention weights at the type level and the node level, combined with the node embedding representation of each node in the heterogeneous graph, a multi-layer attention mechanism operation is performed to update the node embedding representation of each node in the heterogeneous graph, that is, to update the news article content encoding vector and the news article element encoding vector.

6. The machine-generated manuscript interpretable evaluation system based on graph neural network as claimed in claim 5, characterized in that: Encode the content of multimodal news articles to obtain the overall coding vector of the news articles, including: Based on the multimodal pre-training model of UNIMO, the text modal information and image modal information in the news release are encoded in the same semantic space to obtain the text representation matrix and the image representation matrix respectively; The text representation matrix and the image representation matrix are aggregated through the attention mechanism, and the aggregated features are filtered using the multiplication filtering mechanism to finally obtain the cross-modal context representation matrix of the text and the image; After concatenating the cross-modal context representation matrices of text and images, they are concatenated with the text representation matrix and the image representation matrix to obtain the overall encoding vector of the news release.

7. An electronic device, characterized in that: The method comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps of a method for interpretable evaluation of machine-generated manuscripts based on a graph neural network as described in any one of claims 1 to 4 are completed.

8. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the steps of a machine-generated manuscript interpretable evaluation method based on a graph neural network as described in any one of claims 1-4.

Images