Text abstract generation method and device, equipment and storage medium

Abstract

The invention relates to the technical field of artificial intelligence and the field of financial science and technology, and discloses a text abstract generation method, which comprises the following steps: obtaining a business text of a target field, generating a general vocabulary embedding matrix and a target field specific vocabulary embedding matrix, and carrying out weighted fusion to form a fused word embedding matrix. And adding each vocabulary embedding vector in the business text and averaging to generate a sentence embedding matrix. A corresponding topic embedding matrix is generated by defining a plurality of topics. And in combination with the matrix, constructing a multi-granularity heterogeneous graph containing words, sentences and theme nodes, and determining a connection relationship between the nodes. And updating the features of the nodes by using a graph neural network, and generating a text abstract through a decoder. According to the method, the semantic information of the general vocabularies and the field specific vocabularies is combined, and the complex relationship in the text is captured by utilizing the multi-modal semantic similarity, so that the key information extraction capability in the text is enhanced, the effect of generating the personalized text abstract is realized, and the quality of the text abstract is improved.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence technology and financial technology, and in particular to a text summary generation method, device, equipment and storage medium. Background Art

[0002] With the rapid development of the financial industry, especially in the insurance field, the number and complexity of various business documents are gradually increasing. Insurance companies need to process massive amounts of text data, such as insurance contracts, claim terms, customer feedback, etc. How to quickly and accurately extract key information from these huge amounts of data and provide customers with concise and easy-to-understand content has become a major challenge in the industry. At present, text summarization technology is mainly divided into two categories: extractive summarization and generative summarization.

[0003] Extractive summarization technology selects some important sentences or paragraphs from the original text and directly splices them to generate a summary. The advantage of this method is that the content of the summary is consistent with the original text, ensuring the accuracy of the information. However, since extractive summarization lacks a deep understanding of the overall semantic structure, it cannot flexibly generate personalized summaries that meet specific needs. Therefore, the quality of the summary is easily limited by the structure of the original text and the quality of the sentences. The generated content may be lengthy, lack coherence, and fail to accurately grasp the core points of the text.

[0004] Generative summarization uses natural language generation technology to generate summaries by reorganizing and expressing the main content of documents. This type of method can flexibly generate personalized summaries according to different scenarios and needs. However, existing generative summarization methods mainly rely on sequence networks, such as long short-term memory networks (LSTM) and gated recurrent units (GRU). These models work well when processing short texts, but due to their limitations, they have difficulty capturing long-range dependent semantic information. Therefore, when processing long texts or complex documents, the quality of the generated summaries is poor and often fails to accurately convey important information in the text.

[0005] In addition, existing generative summarization models also have shortcomings when processing professional texts in different fields. In particular, for professional terms and domain-specific expressions in the financial and insurance industries, the model cannot effectively distinguish between general semantics and domain-specific semantics. This results in the inability to accurately reflect industry priorities and customer needs when generating summaries of professional domain documents. Summary of the invention

[0006] The main purpose of the present invention is to provide a text summary generation method, device, equipment and storage medium, aiming to solve the technical problems that the prior art is difficult to effectively capture the complex semantic relationships in long texts and lacks the ability to process domain-specific information when generating personalized summaries.

[0007] To achieve the above object, the present invention provides a text summary generation method, comprising:

[0008] Obtaining business texts in a target domain, and generating a general vocabulary embedding matrix and a target domain-specific vocabulary embedding matrix according to the business texts;

[0009] Performing weighted fusion on the general vocabulary embedding matrix and the target domain specific vocabulary embedding matrix according to a weight parameter to generate a fused word embedding matrix;

[0010] Based on the fused word embedding matrix, each word embedding vector of the sentence in the business text is added and averaged to generate a sentence embedding matrix;

[0011] Define multiple topics related to the target domain and generate corresponding topic embedding matrices based on multiple topics;

[0012] Using the fused word embedding matrix, sentence embedding matrix and topic embedding matrix, a multi-granularity heterogeneous graph including word nodes, sentence nodes and topic nodes is constructed, and the connection relationship between the word nodes, sentence nodes and topic nodes is determined;

[0013] Based on the connection relationship between the word nodes, sentence nodes and topic nodes, a graph neural network is used to update the features of the nodes in the multi-granularity heterogeneous graph to generate an updated node feature representation;

[0014] Based on the updated node feature representation, a text summary of the business text is generated through a decoder.

[0015] Furthermore, to achieve the above-mentioned purpose, the present invention provides a text summary generating device, comprising:

[0016] A vocabulary embedding generation module is used to obtain business texts in a target domain and generate a general vocabulary embedding matrix and a target domain specific vocabulary embedding matrix according to the business texts;

[0017] A vocabulary embedding fusion module, used for weighted fusion of the general vocabulary embedding matrix and the target domain specific vocabulary embedding matrix according to a weight parameter to generate a fused word embedding matrix;

[0018] A sentence embedding generation module, used to add and average each word embedding vector of the sentence in the business text based on the fused word embedding matrix to generate a sentence embedding matrix;

[0019] The topic embedding generation module is used to define multiple topics related to the target field and generate corresponding topic embedding matrices based on multiple topics;

[0020] A multi-granularity heterogeneous graph construction module is used to construct a multi-granularity heterogeneous graph including word nodes, sentence nodes and topic nodes by using the fused word embedding matrix, sentence embedding matrix and topic embedding matrix, and determine the connection relationship between the word nodes, sentence nodes and topic nodes;

[0021] A feature updating module, for updating the features of nodes in the multi-granularity heterogeneous graph using a graph neural network based on the connection relationship between the word nodes, sentence nodes and topic nodes, and generating an updated node feature representation;

[0022] A text summary generation module is used to generate a text summary of the business text through a decoder based on the updated node feature representation.

[0023] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer device, which includes a memory, a processor, and a text summary generation program stored in the memory and executable on the processor, and the text summary generation program implements the steps of the text summary generation method as described above when executed by the processor.

[0024] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, on which a text summary generation program is stored, and when the text summary generation program is executed by a processor, the steps of the text summary generation method described above are implemented.

[0025] Beneficial effects: The present invention relates to the fields of artificial intelligence technology and financial technology, and discloses a method for generating a text summary. By acquiring the business text of the target field, a general vocabulary embedding matrix and a target field-specific vocabulary embedding matrix are generated, and the two are weightedly fused based on weight parameters to form a fused word embedding matrix. Each vocabulary embedding vector in the business text is added and averaged to generate a sentence embedding matrix. By defining multiple topics, a corresponding topic embedding matrix is ​​generated. In combination with the above-mentioned matrix, a multi-granularity heterogeneous graph containing words, sentences, and topic nodes is constructed to determine the connection relationship between the nodes. The features of the nodes are updated using a graph neural network, and a text summary is generated through a decoder. The present invention combines the semantic information of general vocabulary and domain-specific vocabulary, and uses multimodal semantic similarity to capture the complex relationships in the text, thereby enhancing the ability to extract key information from the text, achieving the effect of generating personalized text summaries, improving the quality of text summaries, and improving customers' reading experience and satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0027] Figure 1A schematic diagram of an application environment of a text summary generation method in an embodiment of the present invention;

[0028] Figure 2 A flowchart of an embodiment of a text summary generation method of the present invention;

[0029] Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the text summary generating device of the present invention;

[0030] Figure 4 A schematic diagram of the structure of a computer device in one embodiment of the present invention;

[0031] Figure 5 FIG. 4 is another schematic diagram of the structure of a computer device in one embodiment of the present invention. DETAILED DESCRIPTION

[0032] 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.

[0033] The text summary generation method provided by the embodiment of the present invention can be applied in Figure 1 In an application environment, the user terminal communicates with the server terminal through a network. The server terminal can obtain the business text of the target field through the user terminal, generate a general vocabulary embedding matrix and a target field specific vocabulary embedding matrix, and perform weighted fusion of the two based on the weight parameter to form a fusion word embedding matrix. Each vocabulary embedding vector in the business text is added and averaged to generate a sentence embedding matrix. By defining multiple topics, a corresponding topic embedding matrix is ​​generated. In combination with the above matrix, a multi-granularity heterogeneous graph containing words, sentences, and topic nodes is constructed to determine the connection relationship between the nodes. The nodes are updated with features using a graph neural network, and a text summary is generated through a decoder. The present invention combines the semantic information of general vocabulary and domain-specific vocabulary, and uses multimodal semantic similarity to capture the complex relationship in the text, thereby enhancing the ability to extract key information from the text, achieving the effect of generating personalized text summaries, improving the quality of text summaries, and improving the reading experience and satisfaction of customers. Among them, the user terminal can be but is not limited to various personal computers, laptops, smart phones, tablet computers, and portable wearable devices. The server terminal can be implemented with an independent server or a server cluster consisting of multiple servers. The present invention is described in detail below through specific embodiments.

[0034] See also Figure 2 , Figure 2 This is a flowchart of an embodiment of a text summary generation method provided by the present invention. It should be noted that although a logical sequence is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0035] like Figure 2 As shown, the text summary generation method proposed by the present invention includes the following steps:

[0036] S10, obtaining business texts in the target domain, and generating a general vocabulary embedding matrix and a target domain-specific vocabulary embedding matrix according to the business texts;

[0037] In this embodiment, business texts related to the target domain are obtained. The texts may include documents, paragraphs, sentences, etc., and the content covers domain-specific business information, such as data in the insurance, finance, and other industries.

[0038] The target domain is the application domain or business scenario that the system specifically targets. It usually has specific background knowledge, terminology, and business logic. The target domain can be a specific industry or application scenario such as insurance, finance, medical care, and law. Each target domain has its specific business text and terminology. For example: Insurance domain: terms include insurance policy, claim, premium, protection, etc. Financial domain: terms include investment, return, risk control, capital, etc. Medical domain: terms include diagnosis, treatment, medicine, medical insurance, etc. The target domain contains specific business knowledge and professional terminology, which will affect the text processing, vocabulary embedding, node feature update, etc. in the technical solution. According to the characteristics of the target domain, the system can perform personalized optimization to ensure that the generated text summary is more in line with the actual needs and business logic of the domain. For the target domain, professional corpora (such as insurance contracts, financial reports, etc.) can be used to train specific semantic embedding models to improve the model's understanding of the vocabulary in the domain.

[0039] Business texts can be loaded into the system from various sources such as external databases, file systems, networks, etc. through API interfaces or batch file import systems. For real-time data acquisition, the latest field-related texts can be automatically obtained through web crawlers or data streams.

[0040] Use pre-trained general semantic embedding models (such as BERT, Word2Vec, etc.) to embed each word in the business text and generate semantic representations of general vocabulary. BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained language model based on the Transformer architecture with bidirectional encoding capabilities. It improves the understanding of words in different contexts by modeling from the context of the text at the same time. BERT uses pre-training (masked language model and next sentence prediction) and fine-tuning mechanisms. Word2Vec is a model that converts words into vector representations. Through neural network training, semantically similar words are mapped to similar vectors. It has two main methods: CBOW (predicting the central word based on the context) and Skip-gram (predicting the context based on the central word). Word2Vec can capture the grammatical and semantic relationships of words and is widely used in natural language processing tasks such as text classification and word meaning similarity calculation.

[0041] The business text first goes through preprocessing steps, such as word segmentation, removal of stop words and punctuation, etc. The preprocessed text is input into the pre-trained universal semantic embedding model to generate universal vocabulary embedding vectors for each word. Each word embedding vector is a set of numerical vectors with fixed dimensions (such as 300 dimensions or 768 dimensions). The embedded word vectors are combined into a universal vocabulary embedding matrix with a matrix size of "number of words x word vector dimension".

[0042] For example, the business text is segmented by bert and word embedded to get X BERT ∈R L×d , where X BERT is the word embedding matrix generated by the BERT model, which means that each word or vocabulary in the business text is converted into a corresponding vector representation through the BERT model; R L×d Represents the dimension of a matrix, the size of the matrix is ​​L×d, L is the number of tokens in the text (that is, the number or length of words in the text). When processing text, the BERT model will split the input text into tokens (such as words, phrases, etc.), and d represents the dimension of the word embedding vector of each word (token). That is, the BERT model generates a d-dimensional vector representation for each token to capture the semantic information of the word.

[0043] Use a corpus related to the target domain for training to generate a domain-specific vocabulary embedding matrix that reflects domain-specific terms and semantic relationships.

[0044] First, a specific word embedding model (such as domain-specific Word2Vec or BERT model based on domain training) is trained through the domain corpus. The model can capture domain-specific word usage habits and semantic relationships. The processed business text is processed again by this domain-specific word embedding model to generate a domain-specific word embedding matrix. The domain-specific word embedding matrix is ​​similar to the general word embedding matrix and is also constructed according to the "number of words x word vector dimension".

[0045] For example, we use Word2Vec to pre-train on the professional terminology corpus in the insurance field, and input the tokens after BERT segmentation into the pre-trained Word2Vec to obtain the domain-specific word embedding matrix: X DomainEmbedding ∈R L×d .

[0046] Among them, X DomainEmbedding It is a domain-specific word embedding matrix generated by the Word2Vec model, which means converting the words in the business text into corresponding word embedding representations according to a specific domain (such as the insurance field). L×d Represents the dimension of the matrix, the matrix size is L×d, L represents the number of tokens in the text, that is, the number of words or phrases in the text. d represents the dimension of word embedding, similar to the dimension d in the BERT model, where d is the dimension of the domain-specific embedding representation.

[0047] Example description: In the financial field, assume that a large amount of text data about bank loan business is obtained. First, the loan-related business text is processed through the pre-trained BERT model to generate a general vocabulary embedding matrix; then, the financial domain Word2Vec model trained in a specific field is used to generate a domain-specific vocabulary embedding matrix for the text. These two parts of the embedding matrix are used to accurately capture common language features (such as customer name, contact information, etc.) and specific financial terms (such as loan type, interest rate, term, etc.), and ultimately achieve accurate text summary generation, helping banks generate highly customized loan terms overviews for customers.

[0048] By generating a general vocabulary embedding matrix and a target domain-specific vocabulary embedding matrix respectively, an effective combination of general semantics and domain-specific semantics in the text is achieved, which enhances the model's ability to understand business texts, improves the accuracy and personalization of text summary generation, and meets the needs of different customers and scenarios.

[0049] S20, performing weighted fusion on the general vocabulary embedding matrix and the target domain specific vocabulary embedding matrix according to a weight parameter to generate a fused word embedding matrix;

[0050] In this embodiment, the universal vocabulary embedding matrix and the target domain-specific vocabulary embedding matrix are fused through weighting parameters. Each vocabulary embedding vector balances universal semantics with domain-specific semantics through preset or dynamically adjusted weights.

[0051] Get the universal vocabulary embedding matrix and the target domain-specific vocabulary embedding matrix, the dimensions of the matrices are "number of words x word vector dimension". Define two weight parameters α and β to control the proportion of universal vocabulary embedding and domain-specific vocabulary embedding in the fusion. Perform weighted calculation on the embedding vector of each word, the formula is as follows: F fusion (i) = α·X universal (i) + β·X domain (i); where F fusion (i) represents the fused vocabulary embedding vector, X universal (i) and X domain (i) are the universal and domain-specific embedding vectors of the vocabulary respectively, and α and β are weighting parameters.

[0052] The result of weighted fusion is a fused word embedding matrix with the same dimension of "number of words x word vector dimension", where the embedding vector of each word reflects the comprehensive information of the general semantics and domain-specific semantics of the word. The weight parameters can be adjusted according to the needs of the input text or scene to achieve a balance between general semantics and domain-specific semantics.

[0053] When the system is initialized, the default weight parameters α and β are set, for example, α = 0.5 and β = 0.5, to balance general and domain-specific semantics. For certain specific scenarios or domain texts (such as financial data or medical data), the weight parameters can be automatically adjusted through data-driven methods. The values ​​of α and β are dynamically adjusted by analyzing the proportion of domain-specific vocabulary in the text or the contextual features of the text. For example, for texts with obvious domain characteristics, the value of β is increased and the value of α is reduced.

[0054] For example, the pre-trained domain word embedding model Word2Vec is fused with the general word embedding Bert model embedding structure, and the final domain feature matrix F is obtained by weighted summation. domain ∈R L×d .

[0055] F domain =αX BERT +βX DomainEmbedding

[0056] Among them, F domain Represents the domain feature matrix, which is the final fused word embedding representation obtained by weighted fusion of the general word embedding matrix (generated by BERT) and the domain-specific word embedding matrix (generated by Word2Vec). BERT represents the universal vocabulary embedding matrix generated by the BERT model, that is, the embedding matrix that reflects universal semantic information. DomainEmbeddingRepresents the domain-specific vocabulary embedding matrix generated by the Word2Vec model, that is, the semantic embedding matrix for a specific domain (such as the insurance field).

[0057] α and β are weighting parameters that control the fusion ratio of general embedding and domain embedding. α is the weighting parameter of the general BERT word embedding matrix. β is the weighting parameter of the domain word embedding matrix. The constraint α+β=1 in the formula ensures that the sum of the weights of the two matrices is 1 to balance the two types of information. L×d represents the dimension of the final domain feature matrix, the matrix size is L×d, where L is the number of tokens in the text, indicating the number of words. d is the dimension of the word embedding vector. The word vector of each word after fusion word embedding is represented as X i ∈R 1×d , the size is 1×d, where i represents the i-th word.

[0058] Example description: In the generation of customer analysis reports in the financial field, the system first obtains the general financial vocabulary embedding matrix (such as the embedding representation of general vocabulary such as "loan" and "interest rate") and the domain-specific vocabulary embedding matrix (such as the domain-specific embedding representation of specific terms in customer behavior data such as "loan application approval rate" and "credit score"). By setting weight parameters, it is ensured that in the generated customer report summary, both general financial information is retained and the domain characteristics of customer behavior data are reflected.

[0059] By weighted fusion of the general word embedding matrix and the target domain-specific word embedding matrix, the comprehensive expression of general semantics and domain-specific semantics in the text is achieved, ensuring the information fusion effect of different semantic levels. It enhances the understanding of text content and improves the accuracy of extracting domain-specific information, thus providing a more accurate basic semantic representation for downstream text summary generation and improving the quality of personalized summary generation.

[0060] S30, based on the fused word embedding matrix, adding and averaging each word embedding vector of the sentence in the business text to generate a sentence embedding matrix;

[0061] In this embodiment, for each sentence in the business text, the embedding vector of each word in the sentence is extracted based on the fused word embedding matrix and added. The embedding vector represents the semantic information of each word, and by adding them together, the vocabulary information of the entire sentence can be aggregated.

[0062] For each sentence in the business text, extract the embedding vector of each word in the sentence. For example, assuming that the sentence has n words and the embedding vector of each word is a d-dimensional vector, the embedding vector of each word in the sentence can be expressed as X1, X2,..., Xn, and each Xi is a d-dimensional vector. Add the embedding vectors of each word in the sentence to form an initial aggregate vector: Xsum=X1+X2+...+Xn. After adding the vocabulary embedding vectors of each sentence, average them to generate the sentence embedding representation of the sentence. By taking the average value, the contribution of each word in the sentence can be balanced, so that the semantic information of the entire sentence can be expressed in a comprehensive way. The aggregate vector Xsum is averaged to generate the embedding vector of the sentence. Specifically:

[0063]

[0064] Among them, n represents the number of words in the sentence, Xavg is the average embedding vector of the sentence, and the dimension is still d.

[0065] Repeat the above steps for all sentences in the business text, and each sentence will generate an average sentence embedding vector. The embedding vectors of all sentences are combined into a sentence embedding matrix with a matrix size of "number of sentences x sentence embedding dimension (d)".

[0066] For example, the word vector X after word segmentation and embedding of the current sentence i Add and average to get the summary vector v of the current sentence j ∈R 1×d . , where v j Represents the summary vector of the current sentence, that is, the vector obtained by adding the word vectors Xi of all words in the current sentence and taking the average. This vector summarizes the semantic information of all words in the sentence; R 1×d Represents the dimension of the sentence embedding vector. The size of the sentence embedding vector is 1×d, where 1 means that this is a sentence vector (because it is a single vector), and d represents the dimension of the word embedding, that is, the embedding vector of each sentence also has d dimensions. j means that the current sentence is the jth sentence in the document. Each sentence has a sentence embedding representation, which is calculated by taking the weighted average of the embedding representations of all the words in the sentence.

[0067] By adding and averaging the embedding vectors of each word in the sentence, the semantic information of each word is comprehensively represented as sentence-level semantic information. While retaining the semantic features of each word, a sentence embedding matrix with overall semantic representation is generated, which enhances the model's understanding of sentence-level semantic relations, thereby providing a more accurate sentence representation for subsequent summary generation.

[0068] S40, defining multiple topics related to the target domain, and generating corresponding topic embedding matrices based on the multiple topics;

[0069] In this embodiment, according to the business needs of the target domain, multiple topics related to the domain are identified and defined. Each topic represents a core concept or key point in the target domain. These topics will be used to capture semantic information in the text and help generate text summaries that better meet the needs of specific domains.

[0070] By analyzing business texts or existing documents in the field, we can extract the core keywords or concepts in the field. These keywords can represent different topics in the business texts. For example, in the insurance field, we can define topics such as "health insurance", "auto insurance", "claims process", and "insurance terms".

[0071] Use unsupervised learning methods (such as topic model LDA) or domain expert experience to determine multiple topics that match the text content. Each topic is defined by these domain keywords to ensure that these topics can cover the main business needs and concerns in the target field.

[0072] For each defined topic, a corresponding topic embedding matrix is ​​generated to represent the feature information of the topic in the semantic space. The topic embedding matrix will be used to construct subsequent multi-granularity heterogeneous graphs to help the model understand the information related to different topics in the text.

[0073] Each topic is represented by a set of words or phrases, and the words in each topic are embedded using a pre-trained embedding model (such as Word2Vec, BERT, etc.) to generate an embedding vector for each topic. The topic embedding vector is obtained by weighted summation or average processing of the embedding vectors of the topic words, which is similar to the process of generating sentence embeddings. Specifically, assuming that a topic is composed of the keywords "health insurance", "claims", and "medical care", the topic embedding is generated by the following formula:

[0074]

[0075] Among them, Ttopic is the embedding vector of the topic, and X represents the vocabulary embedding vector of each keyword.

[0076] Finally, the embedding vectors of all topics are combined into a topic embedding matrix, and the matrix size is "number of topics x embedding dimension".

[0077] Example description: In the financial field, the system defines multiple topics, such as "loan", "interest rate", "credit score", "risk assessment", etc. The embedding vectors of these topics are generated by weighted average of topic words (such as "loan", "interest", "credit", etc.). The generated topic embedding matrix is ​​used to construct a multi-granularity heterogeneous graph, which helps the model distinguish and extract key information related to each financial topic when generating text summaries. For example, when generating a credit score report, the system will focus on extracting content related to the topics of "credit score" and "risk assessment", and generate a personalized text summary in combination with other business information.

[0078] By defining multiple topics related to the target domain and generating corresponding topic embedding matrices, more accurate semantic information support can be provided for the text summary generation process. Through these topic embeddings, the model can better understand the domain-specific semantic features in the text, realize the personalized customization of text summaries, enhance the ability to extract domain focus, and improve the accuracy and pertinence of summary generation.

[0079] S50, constructing a multi-granularity heterogeneous graph including word nodes, sentence nodes and topic nodes by using the fused word embedding matrix, sentence embedding matrix and topic embedding matrix, and determining the connection relationship between the word nodes, sentence nodes and topic nodes;

[0080] In this embodiment, the word embedding matrix, sentence embedding matrix and topic embedding matrix are fused to create nodes for words, sentences and topics in the text respectively. Each node represents a basic unit in the text, and the heterogeneous graph captures the multi-level semantic structure in the text through these nodes.

[0081] Get the embedding vector of each word from the fused word embedding matrix to generate a word node. The feature vector of the word node contains the general semantics and domain-specific semantics of the word.

[0082] The embedding vector of each sentence is extracted from the sentence embedding matrix to generate a sentence node. The feature vector of the sentence node represents the semantic information of the entire sentence and integrates the semantics of each word in the sentence.

[0083] The embedding vector of each topic is obtained from the topic embedding matrix to generate a topic node. The feature vector of the topic node represents the semantic information related to the topic in a specific domain.

[0084] Finally, the nodes in the heterogeneous graph include three types: word nodes, sentence nodes, and topic nodes. Each node type represents a different semantic level, from fine-grained vocabulary to high-level topics.

[0085] Heterogeneous graphs not only contain multiple nodes, but also need to determine the connection relationship between different nodes. The connection relationship is used to represent the semantic relevance or contextual dependency between nodes.

[0086] Word-sentence node connection: By analyzing whether the word appears in a specific sentence, the connection between the word node and the sentence node is constructed. If a word appears in a sentence, an edge is established between the word node and the corresponding sentence node.

[0087] For example, based on the occurrence of words in a sentence:

[0088]

[0089] A word-sentence (i,j) represents the connection relationship matrix between words and sentences, which specifically indicates whether there is a connection between the i-th word and the j-th sentence; 1 means that if the i-th word appears in the j-th sentence, a connection is established between the word and the sentence, that is, the value of the corresponding position in the matrix is ​​1. 0 means that if the i-th word does not appear in the j-th sentence, there is no connection, and the value of the corresponding position in the matrix is ​​0.

[0090] Sentence-topic node connection: Based on the similarity between sentence embedding and topic embedding, the connection between sentence nodes and topic nodes is established. By calculating the semantic similarity (such as cosine similarity) between sentence embedding vectors and topic embedding vectors, if the similarity exceeds a certain threshold, a connection is established between the sentence node and the relevant topic node.

[0091] For example, based on the similarity between sentences and topics, the cosine similarity between vectors is calculated:

[0092]

[0093] A sentence-topic (j,z) represents the connection relationship matrix between sentence nodes and topic nodes, specifically describing the connection relationship between the jth sentence and the zth topic; 1 means that if the cosine similarity between sentence j and topic z is greater than 0.8, a connection is established between the sentence node and the topic node, and the value of the corresponding position in the matrix is ​​1. 0 means that if the cosine similarity between sentence j and topic z is less than or equal to 0.8, there is no connection, and the value of the corresponding position in the matrix is ​​0. cos(v j ,T z ) represents the sentence vector v j and the topic vector T z Cosine similarity between . Cosine similarity is used to measure the similarity between two vectors. The closer the value is to 1, the more similar they are. j represents the embedding vector of sentence j; T z Represents the embedding vector of topic z.

[0094] Word-topic node connection: The edge between word and topic is set to 0, A word-topic(i,z)=0, there is no connection between word node i and topic node z, so the value of the corresponding position in the matrix is ​​set to 0.

[0095] Word-word node connection: Establish connections between word nodes based on multimodal semantic similarity (such as contextual co-occurrence, syntactic dependency, etc.). Through multimodal similarity methods, it is determined whether two words have a close semantic association in the context, and connections are established between word nodes accordingly.

[0096] For example, in order to calculate more accurate word similarity, the following methods are combined:

[0097] Word vector similarity: Calculate the cosine similarity between word vectors.

[0098]

[0099] Among them, Sim 词向量,ij Represents the cosine similarity between word i and word j, by calculating the word vector X i and X j The angle between them is used to measure their similarity. i and X j Represent the word vectors of word i and word j respectively.

[0100] Contextual similarity: The similarity between words is calculated by the co-occurrence frequency in the context window.

[0101]

[0102] Among them, Sim 上下文,ij Indicates the frequency of word i and word j appearing in the context window. The window size is set to 5 to calculate the context similarity between words. The window size is set to 5, which means that a window of size 5 is used when calculating the context similarity of words.

[0103] Syntactic similarity: The similarity between words is calculated using the path distance in the syntactic dependency graph.

[0104]

[0105] Among them, Sim 句法,ij Represents the distance between word i and word j in the syntactic dependency graph. Path length represents the shortest path between two words in the syntactic dependency graph. The shorter the path, the higher the syntactic similarity between the two words. Path length refers to the shortest path between word i and word j in the syntactic dependency graph. The syntactic dependency here can be constructed using the zh_core_web_sm model of the SpaCy library to process the dependency of Chinese text.

[0106] Final similarity: The above multimodal similarities are weighted and fused to obtain a more accurate similarity between words.

[0107] Simi j =α·Sim 词向量,ij +β·Sim 上下文,ij +γ·Sim 句法,ij

[0108] The final similarity between word i and word j is obtained by weighted fusion of word vector similarity, context similarity and syntactic similarity; α, β and γ are weighting coefficients, which respectively control the contribution of each similarity in the final similarity.

[0109] For any two words i and j, if their similarity Sim ij If it exceeds the preset threshold θ=0.8, an edge is established between them.

[0110]

[0111] Sentence-sentence node connection: Based on the similarity between sentence embedding vectors, connections between sentence nodes are constructed. If two sentences have high semantic similarity, edges are established between their sentence nodes to reflect their semantic similarity relationship.

[0112] For example, the rule for establishing connections between sentences can be

[0113]

[0114] A sentence-sentence (i,j) represents the connection between sentence i and sentence j; if the embedding vector v of sentence i and sentence j i and v j The cosine similarity between i ,v j ) is greater than 0.8, then a connection is established between the sentence nodes, indicating that A sentence-sentence (i,j)=1. Otherwise, no connection is established, indicating that A sentence-sentence (i,j)=0.

[0115] Topic-topic node connection: The edge between topics is set to 0, A topic-topic (i,j)=0, indicating that in the multi-granularity heterogeneous graph, no connection is established between topics and topic nodes.

[0116] By using the fusion word embedding matrix, sentence embedding matrix and topic embedding matrix to construct a multi-granular heterogeneous graph containing word nodes, sentence nodes and topic nodes, a comprehensive understanding of different semantic levels in the text is achieved. By determining the connection relationship between nodes, the heterogeneous graph can effectively capture the semantic dependencies, contextual information and multimodal semantic features in the text. The model's multi-dimensional semantic analysis capabilities for text are enhanced, providing more accurate semantic support for downstream summary generation and other text processing tasks.

[0117] S60, based on the connection relationship between the word nodes, sentence nodes and topic nodes, using a graph neural network to update the features of the nodes in the multi-granularity heterogeneous graph to generate an updated node feature representation;

[0118] In this embodiment, a multi-granularity heterogeneous graph that has been constructed is used, and the graph includes word nodes, sentence nodes, and topic nodes. Each edge in the graph represents the semantic relationship between two nodes, and the connection relationship can be determined based on factors such as word co-occurrence, sentence similarity, and topic similarity. By analyzing each node in the multi-granularity heterogeneous graph, the connection information between the node and the adjacent nodes is obtained. For each node, its adjacent nodes are allocated according to the connection relationship of the edge and the similarity between the nodes. For example, the connection between two word nodes may be based on multimodal semantic similarity; the connection between a sentence node and a topic node is based on semantic matching. A graph neural network (GNN) is used to update the features of each node in the heterogeneous graph. GNN updates the features of the target node by aggregating the feature information of adjacent nodes, so that it not only contains its own semantic information, but also contains contextual information related to adjacent nodes. The initial features of word nodes, sentence nodes, and topic nodes are set to their word embedding, sentence embedding, and topic embedding vectors, respectively. The initial feature vector of each node represents its position in the semantic space. Aggregate adjacent node features: Aggregate the adjacent node features of each node through the aggregation function in the graph neural network. Assuming that the neighboring node of node i is j, the feature update of node i can be expressed as: h i (l+1) =AGGREGATE(h j (l) |j∈N(i))); where h i (l+1)is the feature representation of node i at the l+1 layer, N(i) represents the set of adjacent nodes of node i, AGGREGATE is a function for feature aggregation, and common aggregation functions include sum, average, and maximum. After each layer of feature update is completed, an updated node feature representation is generated. Through multi-layer graph neural network (GNN) iterative update, the features of each node gradually fuse the semantic information from the adjacent nodes to obtain a new node feature representation. These updated node feature representations will be used for subsequent text summary generation. In each layer of GNN, the node features of the previous layer are used to update in combination with the features of the adjacent nodes. After multiple layers of update, an updated node feature representation is generated. The final feature representation of each node contains not only the information of the node itself, but also the feature information of its adjacent nodes. After aggregating the adjacent node features at each layer, a nonlinear activation function (such as ReLU) is applied to increase the expressiveness of the model and generate an updated feature vector.

[0119] By using graph neural networks to update the features of nodes in multi-granular heterogeneous graphs based on the connection relationship between nodes, dynamic fusion of node features is achieved, so that the final feature representation of each node not only contains its own semantic information, but also reflects the semantic association with adjacent nodes. This enhances the model's ability to capture the complex semantic structure of text, provides a more comprehensive node feature representation for subsequent text summary generation, and improves the accuracy and context relevance of the summary.

[0120] S70, generating a text summary of the business text through a decoder based on the updated node feature representation.

[0121] In this embodiment, the node feature representation updated by the graph neural network (GNN) is used as the input of the decoder. The updated node features contain the semantic information of each node (word node, sentence node, topic node) and their semantic associations with other nodes. These feature representations will be passed to the decoder for generating text summaries. The node feature representation updated by GNN has integrated contextual information and semantic associations. The feature representation of each node has already contained information from its neighboring nodes through the previous feature update step. These updated node feature representations are used as the input of the decoder to obtain complete contextual information and ensure that the generated summary can accurately capture the important content in the business text. The decoder is used to convert the input feature information into a text summary in natural language. This process is similar to sequence generation in a text generation model, and the decoder gradually generates each word in the summary. Each time decoding is performed, the decoder generates the next word based on the output of the previous time step and the current node feature representation. At the beginning of the decoding process, the decoder is first initialized, including setting the initial symbol of the decoder (such as the start tag <start>) and initialize the decoder's hidden state. The hidden state is usually initialized based on the updated node features to ensure that the decoder has access to rich contextual information from the beginning. The decoder gradually generates summaries through a recurrent neural network (such as LSTM or GRU) or a Transformer-based architecture. The specific process is as follows: Input the decoder's initial symbol and hidden state to generate the first word. The first generated word is used as input and combined with the current node feature to generate the second word. This process continues until the end symbol ( <end>). During the generation process, the decoder usually uses the attention mechanism to dynamically assign weights to each node feature to ensure that at each time step, the decoder can focus on node features related to the current context. For example, when generating sentences about claims, the decoder will focus on topic nodes and sentence nodes related to claims. When the decoder generates an end symbol, the text summary is completed. The vocabulary sequence output by the decoder is combined into the final text summary. By using the updated node feature representation as the input of the decoder, it can be ensured that the text summary generation process not only contains the content in the original business text, but also captures complex semantic associations through the connection relationship between nodes. The process of gradually generating a text summary through the decoder can ensure that the generated summary content has good semantic coherence and accuracy, thereby improving the overall quality of the summary, which is suitable for a variety of business scenarios, especially scenarios where personalized summaries need to be generated based on the focus of the business field.

[0122] The present invention relates to the fields of artificial intelligence technology and financial technology, and discloses a method for generating a text summary. By acquiring the business text of the target field, a general vocabulary embedding matrix and a target field-specific vocabulary embedding matrix are generated, and the two are weighted and fused based on a weight parameter to form a fused word embedding matrix. Each vocabulary embedding vector in the business text is added and averaged to generate a sentence embedding matrix. By defining multiple topics, a corresponding topic embedding matrix is ​​generated. In combination with the above matrix, a multi-granularity heterogeneous graph containing words, sentences, and topic nodes is constructed to determine the connection relationship between the nodes. The nodes are updated with features using a graph neural network, and a text summary is generated through a decoder. The present invention combines the semantic information of general vocabulary and field-specific vocabulary, and uses multimodal semantic similarity to capture the complex relationship in the text, thereby enhancing the ability to extract key information from the text, achieving the effect of generating personalized text summaries, improving the quality of text summaries, and improving customers' reading experience and satisfaction.

[0123] In one embodiment, the above S50 includes:

[0124] S501, obtaining a word embedding representation from the fused word embedding matrix as a word node;

[0125] S502, obtaining a sentence embedding representation from the sentence embedding matrix as a sentence node;

[0126] S503, obtaining a topic embedding representation from the topic embedding matrix as a topic node;

[0127] S504, constructing a multi-granularity heterogeneous graph including word nodes, sentence nodes, and topic nodes;

[0128] S505, when a word appears in a sentence, establishing a connection between the word node and the sentence node;

[0129] S506, establishing a connection between the sentence node and the topic node based on the semantic similarity between the sentence embedding representation and the topic embedding representation;

[0130] S507, establishing connections between word nodes based on multimodal similarity;

[0131] S508, establishing connections between sentence nodes according to semantic similarities between sentence embedding representations;

[0132] S509: No connection is established between subject nodes.

[0133] In this embodiment, the embedding vector of each word is extracted from the previously generated fusion word embedding matrix, and these embedding vectors are used as feature representations of word nodes. The embedding representation of each word node contains both general semantic information and domain-specific semantic information. From the fusion word embedding matrix, the vector of each word is extracted. For example, for the word "claim" in a certain insurance text, the embedding vector obtained after weighted fusion represents the specific semantics of the word. Each word is created as a word node, and the feature of the node is the corresponding embedding vector. The embedding vector of the sentence is extracted from the previously generated sentence embedding matrix as the feature representation of the sentence node. The embedding representation of each sentence node integrates the semantic information of all words in the sentence and is the overall representation of the sentence semantics. From the sentence embedding matrix, the vector of each sentence is extracted. For example, the sentence "The policy covers medical and hospitalization expenses" will be represented as a feature vector through the embedding matrix as the feature representation of the sentence node. The embedding vector of each topic is extracted from the previously generated topic embedding matrix as the feature representation of the topic node. The feature of each topic node represents the semantic information of a topic in the field, such as "medical insurance", "accident insurance", etc. Through the domain-specific corpus, the embedding vector of each topic-related vocabulary set is extracted. These embedding vectors are weighted to generate semantic representations of topics and used as features of topic nodes. For example, the topic "medical insurance" may generate its feature embedding vectors through keywords such as "health", "hospitalization", and "claims". All generated word nodes, sentence nodes, and topic nodes are used as nodes of the graph to construct a multi-granularity heterogeneous graph. In this graph, nodes of different granularities represent different semantic levels. In the heterogeneous graph, word nodes, sentence nodes, and topic nodes are used as nodes of different levels. Each node constitutes a vertex in the graph according to its embedding vector, representing a different semantic unit of the text. The connection relationship between word nodes and sentence nodes is based on the appearance of the vocabulary in the sentence. If a word appears in a sentence, a connection is established between the word node and the sentence node. For example, if the two words "policy" and "hospitalization" appear in the sentence "The policy covers hospitalization expenses", a connection is established between the word nodes "policy" and "hospitalization" and the sentence node, indicating that these words are related to the sentence. By calculating the semantic similarity between the sentence embedding vector and the topic embedding vector, if the similarity exceeds the set threshold, a connection is established between the sentence node and the topic node. Using semantic similarity algorithms (such as cosine similarity), the semantic similarity between the sentence and the topic is calculated. For example, the sentence "This policy covers medical expenses" has a high similarity with the topic "Medical Insurance", so a connection is established between the sentence node and the topic node "Medical Insurance". The connection between word nodes is based on multimodal similarity, including contextual co-occurrence information, semantic similarity, and syntactic similarity. If two words have a high similarity in multimodal information, a connection is established between them.For example, the two words "hospitalization" and "claim" often appear in the same context in the text or have similar syntactic dependencies, so a connection can be established between the two word nodes to indicate their semantic relevance. If two sentences have high semantic similarity, a connection is established between their sentence nodes. The connection relationship is determined by the similarity of the sentence embedding vectors. For example, the sentences "The policy covers hospitalization expenses" and "The policy provides hospitalization claims" have high semantic similarity, and a connection can be established between their sentence nodes. In a heterogeneous graph, topic nodes are not connected because topic nodes usually represent different fields or concepts, and there is no need to establish a direct semantic association between them. Each topic node is only connected to its related sentence nodes, and no connection is established between topic nodes to ensure a clear hierarchy of the graph structure.

[0134] This embodiment uses the fusion word embedding matrix, sentence embedding matrix and topic embedding matrix to construct a multi-granular heterogeneous graph containing word nodes, sentence nodes and topic nodes, and determines the connection relationship between them, which can capture multi-level semantic information in the text. Through the construction of the heterogeneous graph, the system can identify important semantic nodes in the text and the relationship between them, providing more accurate basic data for subsequent text summary generation, and improving the personalization and semantic integrity of the summary.

[0135] In one embodiment, the above S507 includes:

[0136] S5071, determining semantic similarity between word nodes through vocabulary embedding representation;

[0137] S5072, determining context similarity between word nodes based on context co-occurrence information, wherein the context co-occurrence information is obtained by analyzing the frequency of occurrence of words in the business text or the relationship between words;

[0138] S5073, determining syntactic similarity between word nodes based on dependency syntactic analysis, wherein the dependency syntactic analysis is obtained by parsing sentence structures in the business text;

[0139] S5074, weighted fusion of semantic similarity, contextual similarity and syntactic similarity to generate the total similarity between word nodes;

[0140] S5075, when the total similarity exceeds a preset threshold, a connection is established between the word nodes.

[0141] In this embodiment, the semantic similarity of each word node is calculated based on its embedded representation. The vocabulary embedding representation contains the semantic information of the word, and the distance between two words in the semantic space can be judged by similarity calculation (such as cosine similarity), thereby determining their semantic similarity. Using the embedding vector of each word node, the semantic similarity between two word nodes is calculated. For example, the words "claim" and "compensation" may be close in the semantic space, indicating that their semantic similarity is high. Similarity can be calculated by cosine similarity, and the formula is as follows: Similarity Among them, A and B are the embedding vectors of two words respectively.

[0142] By analyzing the contextual co-occurrence information of words in the text, the contextual similarity between word nodes is calculated. Contextual co-occurrence information refers to whether two words often appear in the same context, which indicates that they may have similar semantic functions when used. By analyzing the frequency of occurrence of words in business texts and their contextual relationships, it is determined whether two words appear in the same context. For example, if "insurance" and "claims" often appear in the same paragraph, they can be considered to have high contextual similarity. The syntactic similarity between word nodes is determined by dependency syntactic analysis. Dependency syntactic analysis is used to parse the grammatical structure in a sentence and determine the syntactic relationship between words. If two words have similar dependency relationships in the syntactic structure, they may also be semantically related. By parsing the syntactic structure in the text, the dependency relationship between words is determined. For example, in the sentence "claims include medical expenses", there may be a dependency relationship between the words "claims" and "medical". If a similar dependency relationship appears in another sentence (such as "compensation includes hospitalization expenses"), it can be considered that "claims" and "compensation" have similar syntactic functions. Taking semantic similarity, contextual similarity and syntactic similarity into consideration, these three similarity information are integrated by weighted fusion to generate the total similarity between word nodes. The information weight of each similarity can be adjusted according to the specific business scenario. The semantic similarity, contextual similarity and syntactic similarity are weighted and fused respectively using the predefined weight parameters α, β and γ, and the formula is as follows: Total similarity = α × semantic similarity + β × contextual similarity + γ × syntactic similarity. Among them, α, β and γ can be adjusted according to the application scenario to ensure that the contribution of different similarity indicators to the total similarity is reasonable. After calculating the total similarity between word nodes, it is determined whether it exceeds the preset similarity threshold. If the total similarity exceeds the threshold, it is considered that the two word nodes are semantically related enough, so a connection is established between them. Set a similarity threshold (for example, 0.8), and when the total similarity of two word nodes exceeds the threshold, a connection is established between them. For example, the total similarity of the words "claim" and "compensation" exceeds the threshold, so they establish an edge in the heterogeneous graph, indicating that they are semantically and contextually related.

[0143] This embodiment can effectively capture the semantic, contextual, and syntactic relationships between words by establishing connections between word nodes based on multimodal similarity. By weighted fusion of multiple similarity information, the system can more accurately represent the correlation between word nodes when constructing a heterogeneous graph, providing more comprehensive semantic support for subsequent text summary generation. This not only improves the accuracy of text summaries, but also enhances the ability of text summaries to capture complex semantic structures.

[0144] In one embodiment, the above S60 includes:

[0145] S601, extracting features of sentence nodes from the sentence embedding matrix as initial feature representations of the sentence nodes;

[0146] S602, extracting features of word nodes from the fused word embedding matrix as initial feature representations of the word nodes;

[0147] S603, extracting the features of the topic node from the topic embedding matrix as the initial feature representation of the topic node;

[0148] S604, constructing a three-layer graph neural network to update the features of each node layer by layer;

[0149] S605, in each layer of the graph neural network, based on the connection relationship between the nodes, the features of the adjacent nodes are used to perform weighted update on the features of the target node;

[0150] S606, determining the weights between the target node and the adjacent nodes through the attention mechanism, performing weighted summation on the features of the adjacent nodes based on the weights, and using the result of the weighted summation as a feature update input of the target node;

[0151] S607, performing nonlinear processing on the weighted summation result through an activation function to generate a new feature representation of the target node;

[0152] S608, after each layer of the graph neural network completes the feature update, the updated node features are passed to the next layer of the graph neural network for updating;

[0153] S609, when the update operation of the three-layer graph neural network is completed, the final updated node feature representation is generated.

[0154] In this embodiment, the sentence embedding matrix is ​​a whole sentence representation obtained by adding and averaging the word embeddings in each sentence. In this step, the features of each sentence are extracted from the sentence embedding matrix and input into the graph neural network as the initial features of the sentence node. By accessing the sentence embedding matrix, the embedding vector of the sentence node is extracted. For example, the sentence "The customer needs medical insurance" will be represented as a vector, which is input into the graph neural network as the initial feature of the sentence node.

[0155] The fused word embedding matrix contains the common and domain-specific semantic information of words. In this step, the feature representation of each word is extracted from the matrix and used as the initial feature of the word node. The embedding vector of each word is extracted from the fused word embedding matrix. For example, the embedding vector of the word "security" will be used as the initial feature of the word node. The topic embedding matrix is ​​generated by analyzing the professional corpus in the field, and the embedding vector of each topic node represents the semantic information of the topic. In this step, the topic embedding vector is extracted as the initial feature of the topic node and input into the graph neural network. The feature vectors of related topics such as "medical security" in the field are extracted from the topic embedding matrix as the initial feature representation of the topic node. The graph neural network (GNN) consists of three layers, and each layer updates the node features based on the connection relationship between the nodes. The three-layer network structure helps to fuse the information of each node layer by layer and realize the gradual refinement of the node features. Construct a GNN with three layers, where the input of each layer is the node features updated in the previous layer. Through this hierarchical structure, the feature relationship between the node and its adjacent nodes can be effectively captured. The features of each node are based on the feature representation of its adjacent nodes when they are updated. The feature update of nodes in each layer depends on the features of its neighboring nodes to ensure that its features incorporate more context and semantic information. In each layer of GNN, the feature representation of the neighboring nodes is used to perform weighted updates on the target node through the connection relationship between the nodes. For example, when the feature of the sentence node "This policy provides medical insurance" is updated, the features of the word nodes "policy", "medical care", and "insurance" connected to it will be referenced. The attention mechanism is used to dynamically adjust the degree of influence of neighboring nodes on the feature update of the target node. The weight of each neighboring node's contribution to the target node's features is calculated by the attention mechanism to ensure the accuracy of feature updates. The weight between the target node and its neighboring nodes is calculated using the attention mechanism, and the features of the neighboring nodes are weighted summed based on the weight. In this way, the features of important nodes will have a greater impact on the update of the target node features. After the weighted summation, the results are processed using a nonlinear activation function (such as ReLU) to generate new target node features. Nonlinear activation functions can enhance the model's ability to express complex patterns. The feature representation after weighted summation is nonlinearly processed to generate an updated node feature representation. For example, for the feature of the word node "security", the feature update result after referring to the adjacent nodes "medical" and "insurance" generates new features through the activation function. After each layer of feature update is completed, the updated node features will be passed to the next layer of GNN for further update. The update process of each layer continues to integrate more contextual information based on the existing features. The updated node features will be passed on at each layer. After the first layer is processed, the generated features will be input to the second layer, and so on, to ensure that the node features can be gradually refined.After the three-layer graph neural network update operation is completed, the generated final node feature representation will integrate the information of all adjacent nodes and contain more comprehensive and accurate semantic information. After the three-layer update, the node feature representation is refined and can be used as input for subsequent steps (such as text summary generation). For example, the final feature representation of the sentence node "This policy provides medical insurance" will fuse multiple layers of information from word nodes and topic nodes.

[0156] For example, the initial feature vector of a node is the word, sentence or topic vector corresponding to each node.

[0157] For example, the initial feature vector of a sentence node is: The initial feature vector of the word node is: The initial feature vector of the topic node is:

[0158] Graph Attention Network:

[0159]

[0160] Among them, the initial feature vector of the sentence node The initial feature of each sentence node is extracted from the sentence embedding matrix v j , represents the embedding vector of the jth sentence; the initial feature vector of the word node The initial features of each word node are extracted from the fused word embedding matrix X i , represents the embedding vector of the i-th word; the initial feature vector of the topic node The initial features of each topic node are extracted from the topic embedding matrix T z , represents the embedding vector of the zth topic. j (l) Represents the feature representation of node j in the lth layer, with a dimension of 1*d. represents the feature representation of node i in the (l+1)th layer, N represents all nodes, and W (l) ∈R d*d is a learnable weight matrix, all nodes in layer l share W (l) Weight matrix. α ij is the attention coefficient:

[0161]

[0162] is a numerical value. σ is the Relu activation function. a∈R in the attention coefficient 1*2d , attention coefficient α ij The formula indicates that after node i is transformed by weight matrix w, it is concatenated with the transformation result of node j and then passes through a∈R 1*2d After the weight vector dot product, the value obtained is subjected to the softmax operation. The updated word vectors of all word nodes are extracted to form a sequence Encoderoutput∈R L*d As the output of the encoder, it is passed to the decoder for further processing.

[0163] This embodiment can effectively capture the relationship between different semantic levels and contexts in the text by using the graph neural network to update node features layer by layer. Through the feature update and fusion mechanism of the multi-layer network, each node feature can fully integrate the information of its adjacent nodes to generate a more accurate and comprehensive semantic representation. It can improve the accuracy of subsequent text summary generation and is particularly suitable for complex semantic capture and summary generation tasks of domain-specific text.

[0164] In one embodiment, the above S10 includes:

[0165] S101, obtaining a business text in a target field, wherein the business text includes a plurality of documents or paragraphs;

[0166] S102, preprocessing the business text to remove stop words, punctuation marks and special symbols, and generating a text sequence after word segmentation;

[0167] S103, inputting the text sequence into a pre-trained universal semantic embedding model to generate a universal vocabulary embedding matrix;

[0168] S104, inputting the text sequence into a domain semantic embedding model trained based on a professional corpus of a target domain to generate a target domain specific vocabulary embedding matrix.

[0169] In this embodiment, the business texts in the target domain can come from multiple documents, paragraphs or reports. These text data contain terminology and semantic information in a specific domain, such as insurance contracts, claim terms, customer reports, etc. These texts will become the basic data for the subsequent text summary generation process. The business texts collected from the target domain may be composed of multiple independent documents, such as customer insurance contracts, medical insurance policies, etc. in the insurance field. All these documents are merged to prepare for subsequent processing steps.

[0170] Preprocessing is a necessary step in text processing, mainly including removing stop words (such as common but semantically meaningless words like "de", "le", "he", etc.), deleting punctuation marks and special symbols (such as commas, periods, parentheses, etc.), so as to ensure that the semantics of the subsequent processed text is clearer. Through word segmentation, the text sequence will be split into individual words or phrases for further analysis. Use natural language processing (NLP) tools to preprocess business texts, remove unnecessary parts, and extract effective vocabulary. For example, the sentence "The customer has purchased medical insurance" will be segmented into "customer", "purchase", "medical", "insurance", and stop words and punctuation marks will be removed.

[0171] The general semantic embedding model is pre-trained with a large amount of general text data (such as encyclopedia articles, news texts, etc.), and it can capture the lexical semantic information in the general domain. Each vocabulary in the text sequence will be converted into an embedding vector through this model, and the embedding vector represents the position of the vocabulary in the semantic space. For example, input the text sequence "The customer purchased medical insurance" into the pre-trained general semantic embedding model (such as the BERT model) to generate the embedding vector of each vocabulary, and the generated vector matrix represents the general vocabulary embedding of the entire sentence.

[0172] The domain semantic embedding model is trained with a specific domain's professional corpus, and it can capture the semantic information of specific terms in that domain. For example, in the insurance domain, the embedding vector of the vocabulary "claim settlement" in the general semantic embedding model may be similar to that of "compensation", but in the insurance domain semantic embedding model, it may have a higher correlation with "claim", "policy". Input the preprocessed text sequence into the semantic embedding model of the target domain. For example, input the insurance contract text into the domain-specific Word2Vec model. This model is trained based on the corpus of the insurance industry, can capture the semantics of domain-specific vocabulary, and generate the corresponding embedding matrix.

[0173] In this embodiment, by inputting the text sequence into the general semantic embedding model and the domain semantic embedding model, generating the general vocabulary embedding matrix and the domain vocabulary embedding matrix, it can capture both the general semantic information in the text and the professional semantic information within a specific domain. This dual-embedding method not only improves the accuracy of lexical semantic understanding but also can better combine the domain-specific knowledge background when generating text summaries, ensuring that the generated summaries are more accurate and conform to the domain semantic characteristics.

[0174] In one embodiment, the above S20 includes:

[0175] S201, initialize the general weight and the domain weight based on the overall semantic characteristics of the business text;

[0176] S202, in the weighted fusion process, adjusting the general weight and the domain weight according to the context of the business text;

[0177] S203, based on the general weight and the domain weight, performing hierarchical weighted fusion on the general vocabulary embedding matrix and the target domain specific vocabulary embedding matrix, and processing the vocabulary embedding representation of each layer in turn;

[0178] S204, after processing and fusing different semantic levels layer by layer, a final fused word embedding matrix is ​​generated.

[0179] In this embodiment, before the fusion of the general vocabulary embedding matrix and the domain vocabulary embedding matrix, it is necessary to set the general weight and domain weight according to the overall semantic characteristics of the business text. The general weight controls the contribution of the general semantic embedding in the fusion, and the domain weight controls the contribution of the domain-specific embedding. The system sets the initial weight parameters by analyzing the overall semantics of the business text (such as the proportion of general and domain-specific information). For example, if the business text has a large proportion of domain-specific vocabulary, the domain weight may be high, while the general weight is relatively low.

[0180] In the weighted fusion process of the vocabulary embedding matrix, the general weight and domain weight need to be dynamically adjusted according to the context changes of the text. This dynamic adjustment ensures that the weighted fusion results in different context scenarios can more flexibly reflect semantic changes. When processing text, the system adjusts the weight parameters in real time according to the changes in context. For example, when a paragraph contains more professional terms, the domain weight may increase, while the general weight decreases, thereby increasing the influence of domain vocabulary embedding.

[0181] Weighted fusion is not completed all at once, but is carried out step by step in layers. The word embedding representations of each layer will be fused based on the general weights and domain weights. By processing layer by layer, information at different semantic levels can be effectively fused. During the fusion process, the system will first process the low-level semantic information of the vocabulary, and apply the general weights and domain weights for preliminary fusion. Then, the system processes the higher-level semantic information layer by layer to ensure that the embedding representations at each level can be fused.

[0182] After completing the weighted fusion of all levels, the final fused word embedding matrix will be generated. This matrix contains both general semantic information and domain-specific semantic information, and can provide more accurate vocabulary representation for subsequent tasks. When all levels of weighted fusion processing are completed, the system will generate a fused vocabulary embedding matrix, which contains the fused vocabulary embedding representation. It will serve as input for subsequent steps (such as text summary generation) to ensure that domain knowledge and general semantics can be combined when generating summaries.

[0183] This embodiment can consider both general semantics and domain-specific semantics when generating vocabulary embedding representations by weighted fusion of the general vocabulary embedding matrix and the domain vocabulary embedding matrix. This hierarchical weighted fusion method enables the fused vocabulary embedding matrix to more accurately reflect the semantic features of the text, providing more accurate and rich semantic support for subsequent text summary generation. The layer-by-layer processing method further ensures that semantic information at different levels can be effectively captured to avoid semantic information loss.

[0184] In one embodiment, the above S70 includes:

[0185] S701, inputting the updated node feature representation into a decoder for processing input feature information when the decoder is initialized;

[0186] S702, initializing the decoder, including setting an initial symbol of the decoder and initializing a hidden state of the decoder based on the updated node feature representation;

[0187] S703, in each time step, the decoder generates the next word based on the word and node feature representation generated in the previous time step;

[0188] S704, adjusting the weight of the node features during the text generation process, and performing weighted processing on the node features using an attention mechanism;

[0189] S705, combining the generated vocabulary sequences to form a complete text summary and outputting it.

[0190] In this embodiment, the node feature representation updated in the graph neural network will be used as the input of the decoder. The decoder needs to be initialized based on these feature information to provide a basis for subsequent text generation. The updated node feature representation contains the semantic information of sentence nodes, word nodes, and topic nodes. The decoder extracts the feature vectors of sentence, word, topic and other nodes from the multi-granular heterogeneous graph node features generated in the previous step. These feature vectors are input into the decoder as input features during initialization.

[0191] During the decoding process, the decoder state needs to be initialized. This is done by using the decoder's initial symbols (e.g. <start>Symbol) starts text generation, and initializes the decoder's hidden state based on the node feature representation to ensure that the decoder can correctly capture the context and semantic information of the text. <start>”) is input into the decoder, and the initial hidden state of the decoder is set based on the node feature representation. This hidden state will be used to store the contextual information needed in the text generation process.

[0192] The decoder uses a recursive method to generate text summaries step by step. In each time step, the word generated in the previous time step is used as input, and combined with the current node feature representation, the next word is generated until the end symbol is generated. During the decoding process, the decoder uses the last generated word as input and combines the node feature representation to generate the next word. For example, if the previous word is "medical", the decoder will generate the next word based on the node features (such as nodes related to "medical").

[0193] During the text generation process, the decoder dynamically adjusts the weights of node features through the attention mechanism. This mechanism enables the decoder to focus on the node features that are most important for generating text at a certain moment, ensuring that the generated text is more accurate. In each generation step, the attention mechanism calculates the weights of each node feature to ensure that at different stages of text generation, the decoder can dynamically adjust the node features it focuses on based on the text context. For example, when generating "guarantee", the system may increase the attention weight of the "insurance" or "claims" node.

[0194] When the decoder has generated all the words, it combines them to form a complete text summary. The generated text summary will serve as a summary of the business text, conveying the main information in a concise and clear manner. The decoder stops generating when it generates the end symbol, and combines all the generated word sequences into a complete sentence or paragraph as the final text summary. The generated summary may be "This policy provides comprehensive medical protection services."

[0195] This embodiment uses node feature representation to initialize the decoder, and combines the attention mechanism to dynamically adjust the weight of the node feature. The decoder can generate a high-quality summary that conforms to the semantic logic of the text. During the decoding process, the system can adjust the weight of the node feature based on the generated context to ensure that the generated text summary not only accurately reflects the core information of the business text, but also highlights the key content based on the semantics of the text. The final generated text summary is concise and clear, and is suitable for domain-specific summary generation scenarios.

[0196] In one embodiment, a text summary generation device is provided, and the text summary generation device corresponds to the text summary generation method in the above embodiment. Figure 3 , Figure 3 This is a functional module diagram of a preferred embodiment of the text summary generation device of the present invention. Vocabulary embedding generation module 10, vocabulary embedding fusion module 20, sentence embedding generation module 30, topic embedding generation module 40, multi-granularity heterogeneous graph construction module 50, feature update module 60 and text summary generation module 70. The functional modules are described in detail as follows:

[0197] A vocabulary embedding generation module 10 is used to obtain business texts in a target domain and generate a general vocabulary embedding matrix and a target domain specific vocabulary embedding matrix according to the business texts;

[0198] A vocabulary embedding fusion module 20, configured to perform weighted fusion of the general vocabulary embedding matrix and the target domain specific vocabulary embedding matrix according to a weight parameter to generate a fused word embedding matrix;

[0199] A sentence embedding generation module 30, configured to add and average each word embedding vector of a sentence in the business text based on the fused word embedding matrix to generate a sentence embedding matrix;

[0200] A topic embedding generation module 40, used to define multiple topics related to the target domain and generate corresponding topic embedding matrices based on the multiple topics;

[0201] A multi-granularity heterogeneous graph construction module 50 is used to construct a multi-granularity heterogeneous graph including word nodes, sentence nodes and topic nodes by using the fused word embedding matrix, sentence embedding matrix and topic embedding matrix, and determine the connection relationship between the word nodes, sentence nodes and topic nodes;

[0202] A feature updating module 60 is used to update the features of the nodes in the multi-granularity heterogeneous graph using a graph neural network based on the connection relationship between the word nodes, sentence nodes and topic nodes, and generate an updated node feature representation;

[0203] The text summary generation module 70 is used to generate a text summary of the business text through a decoder based on the updated node feature representation.

[0204] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external user terminal through a network connection. When the computer program is executed by the processor, it realizes the functions or steps of a text summary generation method server side.

[0205] In one embodiment, a computer device is provided. The computer device may be a user terminal, and its internal structure diagram may be as follows: Figure 5 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a user-side method for generating a text summary

[0206] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the following steps when executing the computer program:

[0207] Obtaining business texts in a target domain, and generating a general vocabulary embedding matrix and a target domain-specific vocabulary embedding matrix according to the business texts;

[0208] Performing weighted fusion on the general vocabulary embedding matrix and the target domain specific vocabulary embedding matrix according to a weight parameter to generate a fused word embedding matrix;

[0209] Based on the fused word embedding matrix, each word embedding vector of the sentence in the business text is added and averaged to generate a sentence embedding matrix;

[0210] Define multiple topics related to the target domain and generate corresponding topic embedding matrices based on multiple topics;

[0211] Using the fused word embedding matrix, sentence embedding matrix and topic embedding matrix, a multi-granularity heterogeneous graph including word nodes, sentence nodes and topic nodes is constructed, and the connection relationship between the word nodes, sentence nodes and topic nodes is determined;

[0212] Based on the connection relationship between the word nodes, sentence nodes and topic nodes, a graph neural network is used to update the features of the nodes in the multi-granularity heterogeneous graph to generate an updated node feature representation;

[0213] Based on the updated node feature representation, a text summary of the business text is generated through a decoder.

[0214] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0215] Obtaining business texts in a target domain, and generating a general vocabulary embedding matrix and a target domain-specific vocabulary embedding matrix according to the business texts;

[0216] Performing weighted fusion on the general vocabulary embedding matrix and the target domain specific vocabulary embedding matrix according to a weight parameter to generate a fused word embedding matrix;

[0217] Based on the fused word embedding matrix, each word embedding vector of the sentence in the business text is added and averaged to generate a sentence embedding matrix;

[0218] Define multiple topics related to the target domain and generate corresponding topic embedding matrices based on multiple topics;

[0219] Using the fused word embedding matrix, sentence embedding matrix and topic embedding matrix, a multi-granularity heterogeneous graph including word nodes, sentence nodes and topic nodes is constructed, and the connection relationship between the word nodes, sentence nodes and topic nodes is determined;

[0220] Based on the connection relationship between the word nodes, sentence nodes and topic nodes, a graph neural network is used to update the features of the nodes in the multi-granularity heterogeneous graph to generate an updated node feature representation;

[0221] Based on the updated node feature representation, a text summary of the business text is generated through a decoder.

[0222] It should be noted that the above functions or steps that can be implemented by the computer-readable storage medium or computer device can refer to the relevant descriptions on the server side and the user side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0223] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0224] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0225] It should be noted that if software tools or components other than those of the Company appear in the embodiments of the present application, they are only used for illustration and do not represent actual use. The above-described embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the above-mentioned embodiments, a person of ordinary skill in the art should understand that the technical solutions described in the above-mentioned embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents; and these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.< / start> < / start> < / end> < / start>

Claims

1. A text summary generation method, characterized in that: The following steps are involved: Obtaining business texts in a target domain, and generating a general vocabulary embedding matrix and a target domain-specific vocabulary embedding matrix according to the business texts; Performing weighted fusion on the general vocabulary embedding matrix and the target domain specific vocabulary embedding matrix according to a weight parameter to generate a fused word embedding matrix; Based on the fused word embedding matrix, each word embedding vector of the sentence in the business text is added and averaged to generate a sentence embedding matrix; Define multiple topics related to the target domain and generate corresponding topic embedding matrices based on multiple topics; Using the fused word embedding matrix, sentence embedding matrix and topic embedding matrix, a multi-granularity heterogeneous graph including word nodes, sentence nodes and topic nodes is constructed, and the connection relationship between the word nodes, sentence nodes and topic nodes is determined; Based on the connection relationship between the word nodes, sentence nodes and topic nodes, a graph neural network is used to update the features of the nodes in the multi-granularity heterogeneous graph to generate an updated node feature representation; Based on the updated node feature representation, a text summary of the business text is generated through a decoder.

2. The text summary generation method according to claim 1, characterized in that: Using the fused word embedding matrix, sentence embedding matrix and topic embedding matrix, a multi-granularity heterogeneous graph including word nodes, sentence nodes and topic nodes is constructed, and the connection relationship between the word nodes, sentence nodes and topic nodes is determined, including: Obtain word embedding representations from the fused word embedding matrix as word nodes; Obtaining a sentence embedding representation from the sentence embedding matrix as a sentence node; Obtaining a topic embedding representation from the topic embedding matrix as a topic node; Construct a multi-granularity heterogeneous graph containing word nodes, sentence nodes, and topic nodes; When a word appears in a sentence, a connection is established between the word node and the sentence node; Based on the semantic similarity between sentence embedding representation and topic embedding representation, a connection is established between sentence nodes and topic nodes. Establish connections between word nodes based on multimodal similarities; Establish connections between sentence nodes based on the semantic similarity between sentence embedding representations; No connections are established between topic nodes.

3. The text summary generation method according to claim 2, characterized in that: Based on multimodal similarity, connections are established between word nodes, including: Determine the semantic similarity between word nodes through vocabulary embedding representation; Determining context similarity between word nodes based on context co-occurrence information, wherein the context co-occurrence information is obtained by analyzing the frequency of occurrence of words in the business text or the relationship between words; Determining syntactic similarity between word nodes based on dependency syntactic analysis, wherein the dependency syntactic analysis is obtained by parsing sentence structures in the business text; The semantic similarity, contextual similarity and syntactic similarity are weightedly integrated to generate the total similarity between word nodes; When the total similarity exceeds a preset threshold, a connection is established between the word nodes.

4. The text summary generation method as claimed in claim 1, characterized in that: Based on the connection relationship between nodes, the graph neural network is used to update the features of nodes in the multi-granularity heterogeneous graph to generate updated node feature representations, including: Extract the features of sentence nodes from the sentence embedding matrix as the initial feature representation of the sentence nodes; Extract the features of word nodes from the fused word embedding matrix as the initial feature representation of the word nodes; Extract the features of the topic node from the topic embedding matrix as the initial feature representation of the topic node; Construct a three-layer graph neural network to update the features of each node layer by layer; In each layer of the graph neural network, based on the connection relationship between nodes, the features of the adjacent nodes are used to perform weighted updates on the features of the target node; Determine the weights between the target node and the adjacent nodes through the attention mechanism, perform weighted summation on the features of the adjacent nodes based on the weights, and use the result of the weighted summation as the feature update input of the target node; The weighted summation result is processed nonlinearly through the activation function to generate a new feature representation of the target node; After each layer of the graph neural network completes the feature update, the updated node features are passed to the next layer of the graph neural network for updating; When the update operation of the three-layer graph neural network is completed, the final updated node feature representation is generated.

5. The text summary generation method according to claim 1, characterized in that: Obtain business texts in the target domain, and generate a general vocabulary embedding matrix and a target domain-specific vocabulary embedding matrix according to the business texts, including: Acquire a business text in a target field, where the business text includes multiple documents or paragraphs; Preprocessing the business text to remove stop words, punctuation marks and special symbols, and generating a text sequence after word segmentation; Inputting the text sequence into a pre-trained universal semantic embedding model to generate a universal vocabulary embedding matrix; The text sequence is input into a domain semantic embedding model trained based on a professional corpus of the target domain to generate a target domain specific vocabulary embedding matrix.

6. The text summary generation method according to claim 1, characterized in that: The general vocabulary embedding matrix and the target domain specific vocabulary embedding matrix are weightedly fused according to a weight parameter to generate a fused word embedding matrix, including: Initialize the general weight and domain weight based on the overall semantic features of the business text; In the weighted fusion process, the general weight and the domain weight are adjusted according to the context of the business text; Based on the universal weight and the domain weight, performing hierarchical weighted fusion on the universal vocabulary embedding matrix and the target domain specific vocabulary embedding matrix, and processing the vocabulary embedding representation of each layer in turn; After processing and fusing different semantic levels layer by layer, the final fused word embedding matrix is ​​generated.

7. The text summary generation method according to claim 1, characterized in that: Based on the updated node feature representation, generating a text summary of the business text through a decoder, including: The updated node feature representation is input into the decoder for processing the input feature information when the decoder is initialized; Initializing the decoder, including setting the initial symbol of the decoder and initializing the hidden state of the decoder based on the updated node feature representation; At each time step, the decoder generates the next word based on the word and node feature representation generated at the previous time step; Adjust the weights of node features during text generation and use the attention mechanism to weight node features; Combine the generated word sequences to form a complete text summary and output it.

8. A text summary generation device, characterized in that: The text summary generating device comprises: A vocabulary embedding generation module is used to obtain business texts in a target domain and generate a general vocabulary embedding matrix and a target domain specific vocabulary embedding matrix according to the business texts; A vocabulary embedding fusion module, used for weighted fusion of the general vocabulary embedding matrix and the target domain specific vocabulary embedding matrix according to a weight parameter to generate a fused word embedding matrix; A sentence embedding generation module, used to add and average each word embedding vector of the sentence in the business text based on the fused word embedding matrix to generate a sentence embedding matrix; The topic embedding generation module is used to define multiple topics related to the target field and generate corresponding topic embedding matrices based on multiple topics; A multi-granularity heterogeneous graph construction module is used to construct a multi-granularity heterogeneous graph including word nodes, sentence nodes and topic nodes by using the fused word embedding matrix, sentence embedding matrix and topic embedding matrix, and determine the connection relationship between the word nodes, sentence nodes and topic nodes; A feature updating module, for updating the features of nodes in the multi-granularity heterogeneous graph using a graph neural network based on the connection relationship between the word nodes, sentence nodes and topic nodes, and generating an updated node feature representation; A text summary generation module is used to generate a text summary of the business text through a decoder based on the updated node feature representation.

9. A computer device, characterized in that: The computer device includes a memory, a processor, and a text summary generation program stored in the memory and executable on the processor. When the text summary generation program is executed by the processor, the steps of the text summary generation method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The storage medium stores a text summary generation program, and when the text summary generation program is executed by the processor, the steps of the text summary generation method according to any one of claims 1 to 7 are implemented.

Images