Text abstract generation method and device, and storage medium

Abstract

This application discloses a text summarization method, device, and storage medium, relating to the field of data processing technology. The method vectorizes each paragraph of text in a received document to obtain a paragraph text vector; based on a local entity library, it extracts entity words and relationships between them from each paragraph to construct a local knowledge graph; it determines the embedding vector of the local knowledge graph using a pre-trained graph neural network, and concatenates the embedding vector and the paragraph text vector to obtain a fusion vector; based on the similarity between the fusion vector and the local text vector, it determines the text type corresponding to each paragraph, and clusters all paragraphs in the document based on the text type to obtain clustered text clusters; based on the clustered text clusters corresponding to the document, it generates a document summary, ensuring the completeness of the summary's description of the core topic.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to text summarization methods, devices and storage media. Background Technology

[0002] Government office work and large-scale project management have accumulated a massive amount of electronic documents, such as meeting minutes, project progress reports, and problem summaries. To quickly extract core information from these lengthy, unstructured texts, artificial intelligence-based automatic text summarization technology has been widely applied.

[0003] Currently, mainstream text summarization technologies mostly employ keyword / key sentence extraction or deep learning-based generative methods, using word frequency statistics or attention mechanisms to calculate sentence importance scores and extracting high-scoring sentences as summaries. However, for long-term project management documents such as project meeting minutes, information about the same core topic, such as "the progress of Project A," may be scattered throughout various parts of the document. The aforementioned mainstream solutions cannot centrally and completely summarize information about a particular core point, resulting in insufficient completeness in the summary content.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this application is to provide a text summarization method, device, and storage medium, aiming to solve the technical problem of how to improve the integrity of text summarization.

[0006] To achieve the above objectives, this application proposes a text summarization method, which includes:

[0007] Vectorize each paragraph of text in the received document to obtain a paragraph text vector;

[0008] Based on the local entity database, entity words in each paragraph text and the relationships between these entity words are extracted to construct a local knowledge graph.

[0009] The graph embedding vector of the local knowledge graph is determined by a pre-trained graph neural network, and the graph embedding vector and the paragraph text vector are concatenated to obtain a fusion vector;

[0010] Based on the similarity between the fused vector and the local text vector, the text type corresponding to each paragraph text is determined, and all paragraph texts in the document are clustered based on the text type to obtain clustered text clusters;

[0011] A document summary is generated based on the clustered text clusters corresponding to the document.

[0012] In one embodiment, the step of generating a document summary based on the clustered text clusters corresponding to the document includes:

[0013] Based on the text type corresponding to the paragraph text within the clustered text cluster, a type embedding vector corresponding to the clustered text cluster is determined, wherein the type embedding vector represents the text features that the text type should contain;

[0014] Vectorize each sentence within the clustered text to obtain a sentence vector;

[0015] The attention score of the sentence vector is determined by the dot product of the sentence vector and the transpose of the type embedding vector within the clustered text cluster.

[0016] Based on the attention score, the sentence vectors within each of the clustered text clusters are weighted and summed to obtain the cluster vector corresponding to each of the clustered text clusters;

[0017] Document summaries are generated based on the cluster vectors corresponding to each of the described clustered text clusters.

[0018] In one embodiment, the step of generating a document summary based on the cluster vectors corresponding to each of the clustered text clusters includes:

[0019] Based on the preset cluster vector sorting rules, the arrangement order of the cluster vectors corresponding to each clustered text cluster is determined to obtain the cluster vector sequence;

[0020] The cluster vector sequence is used as input parameters to a pre-trained decoder, which decodes the cluster vectors into the document summary.

[0021] In one embodiment, the step of determining the arrangement order of cluster vectors corresponding to each clustered text cluster based on a preset cluster vector sorting rule to obtain a cluster vector sequence includes:

[0022] Obtain the user identifier associated with the document, and call the corresponding summary preference template based on the user identifier. The summary preference template is used to define the sorting order of text types preferred by the user.

[0023] If the text type corresponding to the clustered text cluster is consistent with the text type in the summary preference template, then the arrangement order of the cluster vectors corresponding to each clustered text cluster is determined according to the text type arrangement order defined in the summary preference template, and the cluster vector sequence is obtained.

[0024] In one embodiment, the step of generating a document summary based on the cluster vectors corresponding to each of the clustered text clusters includes:

[0025] Based on the local entity library, entity words and relationships between each entity word are extracted from the text corresponding to each clustered text cluster to construct a new knowledge graph.

[0026] The weight of each entity node is determined based on the number of other entity nodes associated with each entity node in the knowledge graph.

[0027] Based on the weights of each entity node, the path of the entity node with the highest weight score is determined by traversing the knowledge graph.

[0028] The entity node path is converted into an entity sequence, and the entity sequence and the text corresponding to each of the clustered text clusters are used as prompt words and input into a pre-trained language model. The language model generates a summary sentence corresponding to each clustered text cluster as the document summary.

[0029] In one embodiment, after the step of generating a document summary based on the clustered text clusters corresponding to the document, the method further includes:

[0030] In response to the deletion operation of the document summary, identify the entity words in the deleted sentence;

[0031] When generating a document summary for the next time, and performing the step of determining the weight of each entity node based on the number of other entity nodes associated with each entity node in the knowledge graph, if the entity node in the knowledge graph is an entity word in the deleted sentence, then the weight of the entity node is reduced.

[0032] In one embodiment, after the step of generating a document summary based on the clustered text clusters corresponding to the document, the method further includes:

[0033] In response to the editing operation of the document summary, identify the entity words in the sentences to be retained after editing;

[0034] When generating a document summary for the next time, and performing the step of determining the weight of each entity node based on the number of other entity nodes associated with each entity node in the knowledge graph, if the entity node in the knowledge graph is an entity word in a sentence that has been edited and retained, then the weight of that entity node is increased.

[0035] In addition, to achieve the above objectives, this application also proposes a text summarization generation device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the text summarization generation method as described above.

[0036] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and which, when executed by a processor, implements the steps of the text summarization method described above.

[0037] This application provides a text summarization method, which involves vectorizing each paragraph of text in a received document to obtain a paragraph text vector; extracting entity words and relationships between them from a local entity library to construct a local knowledge graph; determining the embedding vector of the local knowledge graph using a pre-trained graph neural network; concatenating the embedding vector and the paragraph text vector to obtain a fusion vector; determining the text type of each paragraph based on the similarity between the fusion vector and the local text vector; clustering all paragraphs in the document based on the text type to obtain a clustered text cluster; and generating a document summary based on the clustered text clusters corresponding to the document.

[0038] This method constructs a local knowledge graph to identify and associate the same entities scattered throughout a document, forming a structured representation of the core topic. By combining textual semantic vectors with graph embedding vectors, it captures both the thematic content of paragraphs and reflects the entity relationship structure within those paragraphs. Furthermore, dual clustering using text type recognition and semantic similarity ensures that scattered paragraphs about the same core topic are accurately grouped into the same text cluster. Based on the clustering results, a summary fragment is generated for each text cluster. Since each text cluster fully compiles all relevant information about a specific core topic from the document, it can generate a complete summary covering multiple aspects of the topic, ensuring the completeness of the summary's description of the core topic. Attached Figure Description

[0039] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating the method for generating a text digest for this application is provided in Embodiment 1.

[0042] Figure 2 This is a flowchart illustrating Embodiment 2 of the method for generating a text digest of this application.

[0043] Figure 3 A flowchart illustrating Embodiment 3 of the method for generating a text digest of this application is provided.

[0044] Figure 4 This is a schematic diagram of the device structure of the hardware operating environment involved in the text digest generation method in the embodiments of this application.

[0045] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0046] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of this application and are not intended to limit this application.

[0047] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments. It should be noted that all actions involving the acquisition of signals, information, or data in this application are performed in accordance with the relevant data protection laws and regulations of the country where the application is located, and with authorization from the owner of the corresponding device.

[0048] Government office work and large-scale project management have accumulated a massive amount of electronic documents, such as meeting minutes, project progress reports, and problem summaries. To quickly extract core information from these lengthy, unstructured texts, artificial intelligence-based automatic text summarization technology has been widely applied.

[0049] Currently, mainstream text summarization technologies mostly employ keyword / key sentence extraction or deep learning-based generative methods, using word frequency statistics or attention mechanisms to calculate sentence importance scores and extracting high-scoring sentences as summaries. However, for long-term project management documents such as project meeting minutes, information about the same core topic, such as "the progress of Project A," may be scattered throughout various parts of the document. The aforementioned mainstream solutions cannot centrally and completely summarize information about a particular core point, resulting in insufficient completeness in the summary content.

[0050] In view of the above problems, this application proposes a text summarization method, which involves vectorizing each paragraph of text in the received document to obtain paragraph text vectors; extracting entity words and relationships between them from a local entity database to construct a local knowledge graph; determining the embedding vector of the local knowledge graph using a pre-trained graph neural network; concatenating the embedding vector and the paragraph text vector to obtain a fusion vector; determining the text type of each paragraph based on the similarity between the fusion vector and the local text vector; clustering all paragraphs of text in the document based on the text type to obtain clustered text clusters; and generating a document summary based on the clustered text clusters corresponding to the document.

[0051] This method constructs a local knowledge graph to identify and associate the same entities scattered throughout a document, forming a structured representation of the core topic. By combining textual semantic vectors with graph embedding vectors, it captures both the thematic content of paragraphs and reflects the entity relationship structure within those paragraphs. Furthermore, dual clustering using text type recognition and semantic similarity ensures that scattered paragraphs about the same core topic are accurately grouped into the same text cluster. Based on the clustering results, a summary fragment is generated for each text cluster. Since each text cluster fully compiles all relevant information about a specific core topic from the document, it can generate a complete summary covering multiple aspects of the topic, ensuring the completeness of the summary's description of the core topic.

[0052] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or text summarization generation system capable of performing the above functions. The following description uses a text summarization generation system as an example to illustrate this embodiment and the subsequent embodiments.

[0053] Based on this, the first embodiment of this application provides a text summarization method, referring to... Figure 1 In this embodiment, the text summarization generation method includes steps S10 to S50:

[0054] Step S10: Vectorize each paragraph of text in the received document to obtain paragraph text vectors.

[0055] For example, the text in the received document is segmented based on line breaks or specific punctuation marks to obtain individual paragraph texts. Each paragraph text undergoes preprocessing, including but not limited to unified character encoding, removal of irrelevant characters and preset types of stop words, and word segmentation. The preprocessed paragraph texts are then input into a pre-trained text embedding model, such as the Transformer-based pre-trained model BERT (Bidirectional Encoder Representations from Transformers), Sentence-BERT, etc. The text embedding model treats each paragraph text as a whole and outputs a fixed-length, high-dimensional paragraph text vector. The paragraph text vector encodes the overall semantic information of the paragraph text. For example, a paragraph describing "Project A's schedule is delayed" will have its paragraph text vector close to other vectors describing "delay" and "schedule issues" in the vector space.

[0056] Step S20: Based on the local entity library, extract the entity words in each paragraph text and the relationships between the entity words to construct a local knowledge graph.

[0057] It should be noted that the local entity library is a predefined graph database or relational database containing common entity terms and their attributes in the project management domain. It stores the unique ID, standard name, type, attributes of known standard entity terms, and the relationships between standard entity terms.

[0058] Optionally, entity words in the paragraph text are identified and extracted using a NER (Named Entity Recognition) model. These entity words are then matched with standard entity words in a local entity database. The relationships between entity words in the paragraph text are determined based on the associations between the matched standard entity words in the local entity database.

[0059] For example, each paragraph of text is segmented into a sequence of words based on semantic units using a word segmenter. This sequence of words is then input into the embedding layer of a pre-trained word embedding model, such as Word2Vec or BERT. The word embedding model queries its internal word vector table, converting each word in the sequence into a word vector. Simultaneously, a positional encoding vector is generated for each position in the word sequence. The word vectors are then added to the positional encoding vectors to generate a sequence of word context vectors corresponding to each paragraph of text.

[0060] Next, using a NER model such as BERT followed by a linear classification layer, the aforementioned word context vector sequence is read. A multi-layer Transformer encoder is then used to calculate a deep feature vector that incorporates global context information for each word in the sequence. These deep feature vectors are then input into a CRF (Conditional Random Field) layer or a simple classification head to classify each word as an entity and assign it an entity type label. This results in a list of entity words for each paragraph, where each item includes the original text segment to which it belongs, the entity type, and its start and end positions in the original text.

[0061] For each entity word in the entity word list, its original text fragment and entity type are used as the query key to perform semantic retrieval in the local entity database to obtain a candidate matching entity set. The semantic similarity between the entity word in the entity word list and each standard entity word in the candidate matching entity set is calculated in the vector space. The standard entity word with the highest semantic similarity in the candidate matching entity set is associated with that entity word in the entity word list. Based on the association relationships of the standard entity words associated with each entity word in the entity word list in the local entity database, the relationships between the entity words in each paragraph are determined.

[0062] After obtaining the entity words in each paragraph of text and the relationships between them, the entity words are used as entity nodes, and a local knowledge graph is constructed based on the relationships between them.

[0063] Step S30: Determine the graph embedding vector of the local knowledge graph through a pre-trained graph neural network, and concatenate the graph embedding vector and the paragraph text vector to obtain a fusion vector.

[0064] The local knowledge graph is input into a pre-trained graph neural network. The graph neural network allows each entity node in the local knowledge graph to aggregate information from its neighboring entity nodes, thereby updating its own feature vector. This ensures that each entity node's feature vector contains the contextual information of its graph structure. Global pooling is then performed on the feature vectors of all entity nodes after the above processing to obtain a fixed-length graph embedding vector.

[0065] For the same paragraph text, the corresponding paragraph text vector and the graph embedding vector are connected in dimension to obtain a fusion vector. This fusion vector contains both the semantic information of the paragraph text and the fine-grained entity relationship information, forming a richer paragraph representation.

[0066] Step S40: Based on the similarity between the fused vector and the local text vector, determine the text type corresponding to each paragraph text, and cluster all paragraph texts in the document based on the text type to obtain a clustered text cluster.

[0067] It should be noted that the local text vector is a baseline vector representing different text types, pre-constructed by the text summarization generation system during the training or initialization phase. For example, a large set of labeled paragraph texts is obtained, where each paragraph text is manually labeled with its text type, such as progress reports, risk issues, resolutions, technical solutions, and meeting notices. For each paragraph text in the set, steps S10 to S30 are performed to obtain a fused vector corresponding to each paragraph text as a local text vector, and each local text vector corresponds to a known text type label.

[0068] After determining the similarity between the fused vector and the local text vectors, the text type corresponding to the local text vector with the highest similarity to the fused vector is taken as the text type of the paragraph text corresponding to that fused vector. All paragraph texts with the same text type are then clustered into a single text cluster. In this way, all paragraph text in the document is divided into several text clusters. Each text cluster represents a set of paragraphs with the same text type and revolving around a specific topic.

[0069] Step S50: Generate a document summary based on the clustered text clusters corresponding to the document.

[0070] For each text cluster, since the paragraphs within discuss the same topic and are of the same text type, traditional extraction methods, such as scoring based on word frequency or sentence importance, can be used. One or several of the most representative sentences are selected from all paragraphs in that cluster, and then ordered according to the importance of the topic represented by each cluster or the pre-defined priority of the text type itself, generating a document summary. This summary not only contains the key information of the document but also summarizes and centralizes content about the same core topic scattered across different parts, thus significantly improving the information completeness and structure of the summary.

[0071] Based on the above embodiments of this application, in the second embodiment of this application, the content that is the same as or similar to that in the first embodiment can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The above step S50 includes steps S51 to S55:

[0072] Step S51: Based on the text type corresponding to the paragraph text within the clustered text cluster, determine the type embedding vector corresponding to the clustered text cluster. The type embedding vector represents the text features that the text type should contain.

[0073] Optionally, the text summarization system can maintain a text type embedding matrix, which is a trainable parameter table where each row corresponds to a predefined text type, and each row's data is the type embedding vector corresponding to that text type. Based on the text type label of the current text cluster, the system queries the text type embedding matrix for the specific row corresponding to that text type, and uses that row's data, i.e., the type embedding vector corresponding to that text type, as the type embedding vector for the clustered text. This type embedding vector is then used as a query vector in subsequent steps to filter out the information that best matches the characteristics of that type from the sentences within the clustered text.

[0074] Step S52: Vectorize each sentence within the clustered text cluster to obtain a sentence vector.

[0075] Optionally, using the same method as in step S10, each sentence within the clustered text cluster is encoded to obtain a sentence vector, forming a list of sentence vectors [S1, S2, ..., Sn] corresponding to the clustered text cluster, where Si represents the vector representation of the i-th sentence in the clustered text cluster.

[0076] Step S53: Determine the attention score of the sentence vector based on the dot product of each sentence vector within the clustered text cluster and the transpose of the type embedding vector.

[0077] For example, for each sentence vector Si in the list of sentence vectors corresponding to the above-mentioned clustered text clusters, the dot product of the sentence vector and the transpose of the type embedding vector is determined. In the case of vector dimension matching, this is equivalent to calculating the cosine similarity of the two vectors, obtaining a scalar score. This scalar score reflects the degree of matching between the semantic features of the sentence vector Si and the text type features corresponding to the type embedding vector. The scalar scores corresponding to all sentence vectors are normalized, transforming them into a probability distribution to ensure that the sum of the scalar scores of all sentences is 1. The normalized scalar scores corresponding to each sentence vector are then used as the attention scores for each sentence vector.

[0078] Step S54: Based on the attention score, the sentence vectors within each clustered text cluster are weighted and summed to obtain the cluster vector corresponding to each clustered text cluster.

[0079] Each sentence vector Si is multiplied by its corresponding attention score using a scalar multiplication operation to obtain a weighted sentence vector. All weighted sentence vectors are then summed element-wise; the sum is the cluster vector for this text cluster. This cluster vector not only contains the core semantic content of the clustered text cluster but also highlights sentence information highly relevant to the text type of the clustered text cluster through the attention mechanism.

[0080] Step S55: Generate document summaries based on the cluster vectors corresponding to each of the clustered text clusters.

[0081] After obtaining the cluster vectors highly correlated with the text type for each clustered text cluster, all cluster vectors are input into a pre-trained text summarization generation model, such as the decoder of a Transformer-based generative model. The decoder operates in an autoregressive manner. At each time step, the decoder predicts the next most likely word or subword based on the generated summary prefix and the input cluster vectors as context. Since each cluster vector is already a condensed essence of its corresponding clustered text cluster, the decoder can effectively extract key information from it and generate coherent summary sentences, thus generating a concise summary text for each clustered text cluster. The summary text fragments generated by the decoder for each clustered text cluster are combined and concatenated to form the final document summary.

[0082] Optionally, step S55 above includes steps S551 to S552:

[0083] Step S551: Based on the preset cluster vector sorting rules, determine the arrangement order of the cluster vectors corresponding to each clustered text cluster to obtain the cluster vector sequence.

[0084] Optionally, the preset cluster vector sorting rule can be based on sorting the cluster vectors from largest to smallest, or based on the preset importance order of the text types corresponding to the cluster vectors, or based on the order in which sentences in the corresponding cluster text clusters first appear in the document, etc. According to the above preset cluster vector sorting rule, all cluster vectors are sorted to form an ordered input sequence, namely the aforementioned cluster vector sequence, so that the encoder can concatenate the summary text fragments generated for each cluster text cluster according to the arrangement order of the cluster vectors.

[0085] Optionally, step S551 above further includes steps S5511 to S5512:

[0086] Step S5511: Obtain the user identifier associated with the document, and call the corresponding summary preference template according to the user identifier. The summary preference template is used to define the text type arrangement order of user preferences.

[0087] For example, after receiving a document for which a summary is to be generated, the text summarization system obtains the user identifier associated with the document. This user identifier can be derived from the document's metadata, such as the author, department, login information for the current session, or actively selected by the user. Based on this user identifier, the system queries a pre-set user preference configuration library, which stores preset summary preference templates for different users or user groups, such as "Senior Management," "Project Manager," and "Quality Auditor." Upon successful query, the text summarization system loads the summary preference template bound to the user identifier. The summary preference template is essentially a configuration file or data structure, whose core content defines the text type order of the user role's preferences. For example, a template for a "Senior Executive" user might have the text type order as: "Decision Resolution," "Core Risk," "Key Progress," and "Follow-up Plan"; while a template for a "Technical Lead" user might have the order: "Technical Issues," "Solutions," "Implementation Schedule," and "Resource Requirements."

[0088] Step S5512: If the text type corresponding to the clustered text cluster is consistent with the text type in the summary preference template, then the arrangement order of the cluster vectors corresponding to each clustered text cluster is determined according to the text type arrangement order defined in the summary preference template, and the cluster vector sequence is obtained.

[0089] The text summarization system iterates through all clustered text groups, reads the text type of each cluster, and matches the text type of each cluster with the text type order list defined in the summary preference template. Based on the matching results, a priority value is assigned to each cluster. This priority value is proportional to the index position of the text type of the cluster in the text type order of the summary preference template. For example, if the text type order of the summary preference template is "resolution," "risk," and "progress," then a cluster with the text type "risk" can have a priority value of 1 (assuming counting starts from 0), and a cluster with the text type "progress" can have a priority value of 2.

[0090] The text summarization system compares the priority values ​​of all clustered text clusters and arranges all cluster vectors in ascending order of priority value, i.e., according to the text type order of the summary preference template, to obtain a cluster vector sequence. For clustered text clusters whose text type is not in the summary preference template, a default low priority can be set, or they can be processed according to default rules such as the order in which sentences in the cluster first appear in the document, or they can be directly deleted.

[0091] Step S552: Use the cluster vector sequence as input parameters of the pre-trained decoder, and decode the cluster vector into the document summary through the decoder.

[0092] The cluster vector sequence is used as the input parameter of the pre-trained decoder. The summary text fragments generated by the decoder for each clustered text cluster are concatenated in the order of the cluster vectors to form the final document summary.

[0093] Understandably, users in different roles and at different levels have drastically different information needs regarding the same project document. For example, senior managers may be most concerned with decision-making and risks in order to maintain strategic control, while frontline implementers may be more focused on specific tasks and schedules. This embodiment, through customized summary preference templates, allows the text summary generation system to automatically prioritize the text types that users care about most, generating tailored summaries that improve the direct relevance and reading efficiency of the generated summary information.

[0094] Based on the above embodiments of this application, in the third embodiment of this application, the same or similar content as the above embodiments can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 The above step S50 includes steps S56 to S59:

[0095] Step S56: Based on the local entity database, extract the entity words in the text corresponding to each clustered text cluster and the relationships between the entity words to construct a new knowledge graph.

[0096] For example, using the same method as step S20 in the first embodiment, entity words in the text corresponding to each cluster of text are identified and extracted through the NER model. These entity words are then matched with standard entity words in the local entity database. The relationships between entity words in the text corresponding to each cluster of text are determined based on the associations between the matched standard entity words in the local entity database. After obtaining the entity words in the text corresponding to each cluster of text and the relationships between them, the entity words are used as nodes, and a new knowledge graph is constructed based on the relationships between them.

[0097] Step S57: Determine the weight of each entity node based on the number of other entity nodes associated with each entity node in the knowledge graph.

[0098] The text summarization system traverses each entity node in the knowledge graph, counts the number of connections between that entity node and other directly connected entity nodes, and uses this connection number as the weight of that entity node. It's understandable that traditional key sentence extraction methods judge importance based on word frequency statistics or the sentence's position in the text. This method is effective within a single sentence or paragraph, but cannot identify cross-paragraph elements. A core entity, such as "Project A," repeatedly mentioned in different sentences and from different perspectives throughout the text might be overlooked because it doesn't appear frequently enough or in a prominent position in any single sentence. This embodiment, however, constructs a knowledge graph of clustered text clusters, transforming the semantic relationships of the text into connections in a graph structure. In this graph, the number of connections for an entity node intuitively reflects how central that entity is in the discussion network of the entire clustered text cluster. A high number of connections means that the entity word has a direct relationship with many other entities within the clustered text cluster, such as people, time, tasks, and problems. For example, the node "Project A" might be associated with multiple other nodes such as "Manager Li" (responsible person), "Q2 milestone" (time), "interface delay" (problem), and "testing phase" (status). Therefore, entities with a high number of connections often represent the center of the entire topic knowledge network, and their weight is also higher.

[0099] Step S58: Based on the weights of each entity node, traverse the knowledge graph to determine the path of the entity node with the highest weight score.

[0100] For example, one or more entity nodes are selected from the knowledge graph as candidate starting points. Starting from these candidate starting points, the knowledge graph is traversed to query all possible entity node paths. During the traversal, the weight values ​​of all nodes traversed along the entity node path are accumulated, serving as the weight score for that entity node path. Among all explored entity node paths, the one with the highest total weight score is selected. This path with the highest total weight score connects a group of core entities with high weight and close relationships. To prevent excessively long paths from containing redundant information, an upper limit can be set for the path length, i.e., the number of entity nodes.

[0101] The starting and path nodes of the entity node path with the highest total weighted score are inevitably core entity nodes with a high number of connections. Paths generated with these core entity nodes as the framework naturally connect various peripheral information most relevant to the core topic. For example, when generating a summary about "Project A's progress," the text summarization system doesn't randomly select sentences mentioning "Project A," but instead finds entity paths with "Project A" as the core node, connecting key information such as "current status," "major obstacles," "responsible person," and "next stage." The resulting summary systematically and structurally presents multiple key aspects of the topic, avoiding information fragmentation and omissions.

[0102] Step S59: Convert the entity node path into an entity sequence, and input the entity sequence and the text corresponding to each of the clustered text clusters as prompt words into the pre-trained language model. Generate summary sentences corresponding to each clustered text cluster through the language model, which serve as the document summary.

[0103] The entity nodes in the path with the highest total weight score are converted into corresponding entity name strings in order, forming an entity sequence. This entity sequence is then combined with the text corresponding to the corresponding text cluster to construct a structured cue word. This cue word is input into a pre-trained language model. Based on its internal parameters, the instructions in the cue word, entity cues, and text context, the language model understands the core elements to be summarized and generates a fluent, concise text that covers the key entity information as the summary sentence for that text cluster. The summary sentences of all text clusters are combined to form the final document summary.

[0104] This embodiment provides a generative summarization method based on knowledge graph core path guidance. Its core logic is to first construct a global knowledge graph of topic-clustered text clusters, then locate the most core entity node paths through graph analysis, and finally use these entity node paths along with the original text as prompts to guide a large language model to generate summary sentences that are both faithful to the core facts of the original text and coherent and natural.

[0105] Based on the third embodiment of this application, in the fourth embodiment of this application, as an optional implementation, after the step of generating a document summary based on the clustered text cluster corresponding to the document, in response to the deletion operation of the document summary, the entity words in the deleted sentence are determined; when the next document summary is generated, and the step of determining the weight of each entity node based on the number of other entity nodes associated with each entity node in the knowledge graph is executed, if the entity node in the knowledge graph is an entity word in the deleted sentence, the weight of the entity node is reduced.

[0106] For example, after the text summarization system generates and presents a document summary, the user performs a deletion operation, removing one or more sentences from the summary. The text summarization system captures this deletion operation and treats it as a negative feedback signal, indicating that the specific information contained in the deleted sentence is not recognized or needed by the user in the current context. Next, the text summarization system performs text analysis on the deleted sentences, using entity recognition technology to extract entity words from them. These entity words are considered the content elements that led to the user's rejection of the sentence. When the user triggers a text summarization task for the same document or a document of the same type again, the text summarization system, in the process of re-executing the determination of initial weights based on the connection count of each entity node in the knowledge graph, introduces a weight decay factor to reduce the weight of the entity node corresponding to the entity word in the deleted sentence. For example, the text summarization system checks each entity node in the current knowledge graph; if it is an entity word from a deleted sentence in the previous document summary, it multiplies the initial weight of that entity node based on the connection count by a coefficient less than 1, or directly subtracts a fixed value.

[0107] Entity nodes with reduced weights become less competitive in the subsequent step of determining the path to the entity node with the highest weight score, and their paths are less likely to be selected as the skeleton for generating the summary. Therefore, the new summary generated based on the new path and prompt words will tend to weaken or no longer contain content strongly related to the entity words in these deleted sentences.

[0108] As another optional implementation, after the step of generating a document summary based on the clustered text clusters corresponding to the document, in response to the editing operation of the document summary, the entity words in the sentences retained after editing are determined; when the next document summary is generated, and the step of determining the weight of each entity node based on the number of other entity nodes associated with each entity node in the knowledge graph is executed, if the entity node in the knowledge graph is an entity word in the sentence retained after editing, then the weight of the entity node is increased.

[0109] For example, after the text summarization system generates and presents a document summary, the user performs an editing operation to modify the content of the document summary. The text summarization system recognizes the sentences that the user ultimately chooses to keep as a positive feedback signal, indicating that the information contained in these sentences is recognized by the user as relevant, important, or well-expressed. The text summarization system performs entity recognition on the sentences that have been edited and retained, extracting entity words. These entities are considered key elements constituting a high-quality summary recognized by the user. When the user triggers the text summarization task again, a weight gain factor is introduced when the text summarization system recalculates the weights of each entity node in the knowledge graph. For example, the text summarization system checks entity nodes in the knowledge graph; if it is an entity word from a sentence that was edited and retained in the previous summary, then the initial weight of that entity node, calculated based on its connection number, is multiplied by a coefficient greater than 1, or increased by a fixed value.

[0110] Entity nodes with increased weight become more important in the subsequent step of determining the path of the entity node with the highest weight score, and the path containing it is more likely to be selected as the skeleton for generating the summary. Therefore, the newly generated summary will tend to strengthen and include more information related to the entity words of these edited and preserved sentences, thus aligning with user preferences and judgments.

[0111] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the text digest generation method of this application. Any simple transformations based on this technical concept are within the protection scope of this application.

[0112] This application provides a text summarization generation device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the text summarization generation method in Embodiment 1 above.

[0113] The following is for reference. Figure 4 The diagram illustrates a structural schematic suitable for implementing the text summarization generation device of the embodiments of this application. The text summarization generation device in the embodiments of this application may include, but is not limited to, mobile terminals such as laptops and tablets (PADs, Portable Application Description), and fixed terminals such as desktop computers. Figure 4 The text summarization device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0114] like Figure 4As shown, the text summarization device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The random access memory 1004 also stores various programs and data required for the operation of the text summarization device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the text summarizing device to communicate wirelessly or wiredly with other devices to exchange data. Although a text summarizing device with various systems is shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems may be implemented alternatively.

[0115] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0116] The text summarization device provided in this application, employing the text summarization method described in the above embodiments, can solve the technical problem of how to improve the completeness of text summarization generation. Compared with the prior art, the beneficial effects of the text summarization device provided in this application are the same as those of the text summarization method provided in the above embodiments, and other technical features of this text summarization device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0117] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0118] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0119] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the text digest generation method in the above embodiments.

[0120] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, radio frequency (RF), etc., or any suitable combination thereof.

[0121] The aforementioned computer-readable storage medium may be included in the text summarizing device; or it may exist independently and not be assembled into the text summarizing device.

[0122] The aforementioned computer-readable storage medium carries one or more programs that, when executed by a text digest generation device, enable the text digest generation device to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, or as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0123] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0124] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0125] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described text summarization generation method, thereby solving the technical problem of how to improve the integrity of text summarization generation. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the text summarization generation method provided in the above embodiments, and will not be repeated here.

[0126] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the text summarization method described above.

[0127] The computer program product provided in this application can solve the technical problem of how to improve the completeness of text summarization generation. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the text summarization generation method provided in the above embodiments, and will not be repeated here.

[0128] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A text summarization method, characterized in that, The text summarization method includes: Vectorize each paragraph of text in the received document to obtain a paragraph text vector; Based on the local entity database, entity words in each paragraph text and the relationships between these entity words are extracted to construct a local knowledge graph. The graph embedding vector of the local knowledge graph is determined by a pre-trained graph neural network, and the graph embedding vector and the paragraph text vector are concatenated to obtain a fusion vector; Based on the similarity between the fused vector and the local text vector, the text type corresponding to each paragraph text is determined, and all paragraph texts in the document are clustered based on the text type to obtain clustered text clusters; Based on the local entity library, entity words and relationships between each entity word are extracted from the text corresponding to each clustered text cluster to construct a new knowledge graph. The weight of each entity node is determined based on the number of other entity nodes associated with each entity node in the knowledge graph. Based on the weights of each entity node, the path of the entity node with the highest weight score is determined by traversing the knowledge graph. The entity node paths are converted into entity sequences, and the entity sequences and the text corresponding to each of the clustered text clusters are used as prompt words and input into a pre-trained language model. The language model generates summary sentences corresponding to each clustered text cluster as document summaries.

2. The text summarization method as described in claim 1, characterized in that, After the step of generating summary sentences corresponding to each clustered text cluster using the language model as document summaries, the method further includes: In response to the deletion operation of the document summary, identify the entity words in the deleted sentence; When generating a document summary for the next time, and performing the step of determining the weight of each entity node based on the number of other entity nodes associated with each entity node in the knowledge graph, if the entity node in the knowledge graph is an entity word in the deleted sentence, then the weight of the entity node is reduced.

3. The text summarization method as described in claim 1, characterized in that, After the step of generating summary sentences corresponding to each clustered text cluster using the language model as document summaries, the method further includes: In response to the editing operation of the document summary, identify the entity words in the sentences to be retained after editing; When generating a document summary for the next time, and performing the step of determining the weight of each entity node based on the number of other entity nodes associated with each entity node in the knowledge graph, if the entity node in the knowledge graph is an entity word in a sentence that has been edited and retained, then the weight of that entity node is increased.

4. A text summarization generation device, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the text summarization method as described in any one of claims 1 to 3.

5. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the text summarization method as described in any one of claims 1 to 3.

Images