Method and system for generating procurement review document abstracts in combination with an ai large model

Abstract

The application provides a procurement review document abstract generation method and system combined with an AI large model, relates to the technical field of large language models, and first acquires a set containing a plurality of procurement review document units with review time marks, extracts content theme features and review logic features of each procurement review document unit, then calls a pre-trained AI large model to perform joint abstract generation processing on the content theme features and the review logic features, generates a preliminary abstract text, then performs semantic coherence optimization processing on the preliminary abstract text to obtain an optimized abstract text with semantic coherence, and finally generates a final procurement review document abstract containing a theme keyword set based on the optimized abstract text and outputs the final procurement review document abstract to a target storage terminal in a preset format, so that automatic, efficient and accurate generation of the procurement review document abstract is realized, and the document processing efficiency and information utilization value are improved.

Description

Technical Field

[0001] This invention relates to the field of large language model technology, and more specifically, to a method and system for generating procurement review document summaries by combining AI large models. Background Technology

[0002] In corporate procurement, procurement review documents are crucial for comprehensively evaluating and recording procurement activities, covering aspects such as procurement needs, supplier information, review criteria, and review results. As corporate procurement becomes increasingly complex and frequent, the number of procurement review documents has increased dramatically. These documents are often lengthy and information-heavy, posing significant challenges to subsequent document management and information retrieval.

[0003] Currently, traditional document summarization methods mainly rely on manual extraction of key information or simple keyword matching rules. These methods are not only inefficient, but also difficult to accurately grasp the core content and logical structure of the document, resulting in inconsistent quality of the generated summaries, which cannot meet the needs of enterprises for efficient and accurate management of procurement review documents. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for generating procurement review document summaries using a large AI model, the method comprising:

[0005] Obtain a set of procurement review documents to be processed. The set of procurement review documents contains multiple procurement review document units with review time stamps, and each procurement review document unit corresponds to an independent procurement review business record.

[0006] The procurement review document set is subjected to document feature extraction processing to obtain the content theme features and review logic features of each procurement review document unit;

[0007] The pre-trained AI model is invoked to perform joint summary generation processing on the content topic features and review logic features to generate preliminary summary text for each procurement review document unit. The preliminary summary text contains a word sequence generated based on the key information of the document.

[0008] The preliminary summary text is subjected to semantic coherence optimization processing to obtain a semantically coherent optimized summary text;

[0009] Based on the optimized summary text, a final procurement review document summary containing a set of thematic keywords is generated, and the final procurement review document summary is output to the target storage terminal in a preset format, which includes standardized layout of title, body text and keyword list.

[0010] In another aspect, embodiments of the present invention also provide a procurement review document summary generation system that combines AI large models, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or code. The processor is used to execute the programs, instructions or code in the machine-readable storage medium to implement the above-described method.

[0011] Based on the above, this embodiment of the invention obtains a set of multiple procurement review document units with review time stamps, and extracts content theme features and review logic features for each procurement review document unit. This enables a comprehensive and in-depth understanding of the core content and logical relationships of the documents. A pre-trained AI model is invoked to perform joint summary generation processing on the features, fully utilizing the powerful language understanding and generation capabilities of the AI ​​model to ensure that the generated preliminary summary text accurately covers the key information of the document. Further semantic coherence optimization is performed on the preliminary summary text, making the final procurement review document summary more fluent and natural in semantics, improving the readability and quality of the summary. Finally, the summary containing the set of theme keywords is output to the target storage terminal according to a preset format, achieving standardized formatting and efficient storage of procurement review document summaries. This facilitates document management and information retrieval for enterprises, significantly improving the processing efficiency and information utilization value of procurement review documents. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the execution flow of the procurement review document summary generation method combining AI large model provided in the embodiments of the present invention.

[0013] Figure 2 This is a schematic diagram of exemplary hardware and software components of the procurement review document summary generation system that combines AI large models provided in this embodiment of the invention. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a method for generating procurement review document summaries using an AI-based large-scale model, as provided in one embodiment of the present invention. The following is a detailed description of this method for generating procurement review document summaries using an AI-based large-scale model.

[0015] Step S110: Obtain a set of procurement review documents to be processed. The set of procurement review documents contains multiple procurement review document units with review time stamps. Each procurement review document unit corresponds to an independent procurement review business record.

[0016] In the procurement management process, to conduct a comprehensive and in-depth analysis of the procurement review work, it is first necessary to obtain a set of procurement review documents to be processed. These procurement review documents can come from the information systems of different departments within the enterprise, such as the project management system of the procurement department, the expense approval system of the finance department, and the contract management system of the legal department. The above-mentioned systems store various documents related to procurement review, including purchase requests, supplier quotations, review reports, contract documents, etc.

[0017] Each procurement review document unit includes a review timestamp, which can be a specific date, time, or time period. These timestamps facilitate subsequent sorting and filtering of documents along a timeline, enabling analysis of the characteristics and trends of procurement reviews across different periods. For example, when analyzing changes in a company's annual procurement costs, procurement review documents from different months or quarters can be filtered for comparison based on the review timestamps. Furthermore, each procurement review document unit corresponds to an independent procurement review business record. For instance, a review of equipment procurement for a new production line, from requirement submission, supplier selection, technical evaluation, business negotiations to the final selection of the winning supplier, comprises all relevant documents and records throughout the entire process, constituting a complete procurement review document unit.

[0018] It is worth noting that the collection of procurement review documents must strictly comply with laws, regulations, and the company's internal data management rules. The procurement review documents should be obtained from legitimate data sources such as the company's internal document management system and electronic document repository to ensure the legality and compliance of the document collection's origin. Simultaneously, the collected procurement review documents should be encrypted using encryption algorithms to prevent leakage during transmission and storage.

[0019] Step S120: Perform document feature extraction processing on the procurement review document set to obtain the content theme features and review logic features of each procurement review document unit.

[0020] After obtaining the collection of procurement review documents, document feature extraction processing is required to extract valuable information from them. Content theme features and review logic features reflect the core content and logical structure of the procurement review documents from different perspectives. Content theme features can reveal the main topics and areas involved in the documents, while review logic features reflect the reasoning and decision-making logic in the review process.

[0021] Step S121: Input the procurement review document unit into the topic model for latent semantic analysis. By iteratively calculating the probability distribution of each word unit in the preset topic space, the content topic features of each procurement review document unit are generated. The content topic features include the probability distribution of topic words and the strength of the association between topics.

[0022] When conducting latent semantic analysis, it is necessary to select an appropriate topic model. A common topic model is the Latent Dirichlet Allocation (LDA) model. The predefined topic space is a set of topic categories predefined based on the professional knowledge and business needs of the procurement review domain. These topic categories can cover various key aspects of the procurement review process, such as "supplier qualification assessment", "procurement price analysis", and "technical solution feasibility".

[0023] After the procurement review document units are input into the topic model, the topic model begins iterative calculations. In each iteration, the topic model reassigns the topic to each unit based on its relevance to various topics. Specifically, the topic model considers factors such as the context of the unit within the document and its co-occurrence frequency with other units to determine which topic the unit is more likely to belong to. Then, it updates the probability distribution of the unit across topics based on the new assignment results. This iterative process repeats until the probability distribution reaches a stable state or the preset maximum number of iterations is reached.

[0024] Through the iterative calculations described above, the final topic word distribution probability is generated. This probability reflects the probability of each topic word appearing in various topics. For example, in the topic of "supplier qualification assessment," topics such as "business license" and "industry certification" might have a higher probability of appearance, while in the topic of "purchase price analysis," these topics might have a lower probability of appearance. Simultaneously, the correlation strength between topics is also calculated. The correlation strength between topics measures the relevance between different topics and can be determined by statistically analyzing the number of shared word units between different topics and the importance of these shared word units in their respective topics. If two topics have a large number of shared word units, and these shared word units are highly important in both topics, then the correlation strength between the two topics is high.

[0025] Step S1211: Use a probabilistic generation model to define the probability distribution relationship between topics and terms, establish the latent semantic space of the topic model, map each procurement review document unit to the latent semantic space, and generate an initial topic vector that reflects the topic affiliation of each term. The dimension of the initial topic vector corresponds to the preset number of topics.

[0026] In this step, a probabilistic generative model is used to define the probability distribution relationship between topics and terms. The probabilistic generative model can generate the probability distribution of new terms based on known topic and term information. Through this method, the latent semantic space of the topic model is established. The latent semantic space is an abstract space that quantifies and represents the relationship between topics and terms.

[0027] Mapping each procurement review document unit to this latent semantic space is to associate terms in the document with predefined topics. Specifically, the model analyzes each term in the document and determines its position in the latent semantic space based on its probability distribution with various topics. Thus, each procurement review document unit can be represented by a vector, which is the initial topic vector. The dimensions of the initial topic vector correspond to the number of predefined topics, with each dimension representing a topic, and each element in the vector representing the relative importance of each topic within that document unit. For example, if three topics are predefined, the initial topic vector is a three-dimensional vector, where the three elements represent the degree of association between the document unit and these three topics.

[0028] Step S1212: Calculate the semantic correlation degree between the topic words in the semantic space of the initial topic vector. The semantic correlation degree ranges from 0 to 1. The higher the semantic correlation degree, the higher the co-occurrence probability of the term in the topic.

[0029] After obtaining the initial topic vectors, it is necessary to calculate the semantic relevance between topic terms. Semantic relevance reflects the co-occurrence probability of terms in a topic, that is, the likelihood that two terms will appear simultaneously in the same topic. Various methods can be used to calculate semantic relevance, such as word vector-based similarity calculation methods.

[0030] First, the topic terms are converted into corresponding word vector representations. Word vectors are a method of representing terms by mapping them to a low-dimensional vector space, which can capture the semantic relationships between terms. Then, the similarity between the word vectors of any two topic terms is calculated. Common similarity calculation methods include cosine similarity and Euclidean distance. The similarity value calculated by these methods is the semantic relevance, which ranges from 0 to 1. When the semantic relevance is close to 1, it indicates that the two terms have a high probability of co-occurrence in the topic and they have a strong semantic connection; when the semantic relevance is close to 0, it indicates that the two terms have a low probability of co-occurrence in the topic and they have a weak semantic connection.

[0031] Step S1213: Cluster the initial topic vectors according to the semantic relevance. Use an unsupervised clustering algorithm to aggregate the clustered topic vectors into multiple topic cluster units, each topic cluster unit representing an independent content topic.

[0032] Based on the calculated semantic relevance, the initial topic vectors are clustered. Unsupervised clustering algorithms are clustering methods that do not require prior knowledge of data category labels; they automatically divide data into different categories based on the similarity between data points.

[0033] In this process, initial topic vectors are used as input data. Based on the semantic relevance between topic terms, topic vectors with high similarity are grouped together to form topic clusters. Each topic cluster represents an independent content topic, containing a set of semantically related topic vectors. For example, in a procurement review document, different topic clusters such as "supplier evaluation," "procurement cost," and "technical solution" may be formed. Through clustering, the topics in the procurement review document can be further subdivided and summarized.

[0034] Step S1214: Extract the top K terms from the sorted list of each topic cluster unit, which is sorted in descending order of frequency of occurrence, as topic terms. By statistically analyzing the frequency of occurrence of topic terms in the procurement review document unit and performing standardization processing, a topic term distribution probability is generated. This topic term distribution probability is used to measure the importance of topic terms in the corresponding topic.

[0035] After forming thematic clusters, keyword terms need to be extracted from each cluster. Specifically, the terms in each cluster are sorted in descending order of frequency, forming a sorted list. Then, the top K terms from this list are selected as keyword terms. The value of K can be adjusted based on the actual situation, determining the number of keyword terms selected for each topic.

[0036] Next, we will analyze the frequency of these keywords in the procurement review document units. Frequency refers to the ratio of the number of times a keyword appears in a document to the total number of words in the document. To make the frequencies of different keywords comparable, we need to standardize them. Standardization transforms the frequencies to a uniform range, such as between 0 and 1. Through this process, we generate the keyword distribution probability. The keyword distribution probability reflects the importance of a keyword in its corresponding topic; a higher probability value indicates greater importance of the keyword in the topic.

[0037] Step S1215: Analyze the semantic overlap between different topic clusters. By counting the number of shared terms and calculating the sum of the weights of the shared terms in their respective topics, generate a topic association strength descriptor that includes the number of overlapping words and the area of ​​semantic overlap. This topic association strength descriptor is used to characterize the correlation between topics.

[0038] Analyzing the semantic overlap between different topic clusters aims to understand the relevance between topics. Specifically, this involves counting the number of shared terms across different topic clusters. Shared terms are those appearing in two or more topic clusters. Simultaneously, the sum of the weights of these shared terms within their respective topics is calculated. The weights of shared terms can be determined based on their importance within the topic; for example, the probability distribution of topic terms can be used as weights.

[0039] By statistically analyzing the number of shared terms and calculating the sum of their weights, a descriptor for the strength of inter-topic associations is generated. This descriptor includes two aspects: the number of overlapping terms and the semantic overlap area. The number of overlapping terms directly reflects the amount of shared terms between two topics, while the semantic overlap area comprehensively considers the importance of shared terms within their respective topics. This descriptor provides a more comprehensive characterization of the relevance between topics, aiding in understanding the relationships between various topics in procurement review documents.

[0040] Step S122: Perform logical structure parsing on the procurement review document unit to generate the review logic features.

[0041] Analyzing the logical structure of procurement review document units aims to uncover the reasoning and decision-making logic within the review process. The logical characteristics of the review process reflect the logical relationships between review clauses, helping to understand the rationality and coherence of the review process.

[0042] Step S1221: Identify the title hierarchy information in the procurement review document unit, and construct a hierarchical tree structure of the review clauses using the nested relationship of the hierarchy markers. The root node of the hierarchical tree structure is the entire document, and the child nodes are multi-level clauses in descending order of hierarchy. Each node records the clause text, hierarchy depth, and parent node reference.

[0043] When analyzing the logical structure of a procurement review document unit, the first step is to identify the heading hierarchy information within that unit. Heading hierarchy information is typically represented by different fonts, font sizes, indentation, etc. By utilizing the nested relationships of hierarchy markers, such as numerical or alphanumeric numbering, a hierarchical tree structure of the review clauses can be constructed.

[0044] The root node of the hierarchical tree structure represents the entire procurement review document unit, with child nodes arranged in descending order of hierarchy. For example, the clause corresponding to the first-level heading is a direct child node of the root node, the clause corresponding to the second-level heading is a child node of the first-level heading clause, and so on. Each node records the clause text, hierarchy depth, and parent node reference. The clause text is the specific content of the review clause represented by that node, the hierarchy depth indicates the node's position in the hierarchical tree structure, and the parent node reference determines the relationship between the node and its superior nodes. This hierarchical tree structure effectively displays the logical structure of the procurement review document.

[0045] Step S1222: Traverse each review clause, extract the logical connectors, match the logical connectors based on the pre-built logical connector dictionary to determine the explicit logical relationship type between each review clause, and for candidate review clauses that do not have an explicit logical relationship type, determine the implicit logical relationship type by calculating the word vector similarity of the clause text of the candidate review clauses, and obtain the target review clause sequence with logical relationship.

[0046] After constructing the hierarchical tree structure, it is necessary to traverse each review clause and extract its logical connectors. Logical connectors are words used to express logical relationships between clauses, such as "and," "or," and "but." A dictionary of logical connectors is pre-constructed, containing common logical connectors and their corresponding logical relationship types.

[0047] The extracted logical connectors are matched against a dictionary of logical connectors to determine the explicit logical relationships between the review clauses, such as parallel, alternative, and adversative relationships. However, some review clauses may not have obvious logical connectors; these are candidate review clauses. For candidate review clauses, the implicit logical relationship type is determined by calculating the word vector similarity of their clause text.

[0048] Specifically, the text of the candidate review clauses is converted into word vectors, and then the similarity between the word vectors of any two clause texts is calculated. If the similarity is high, it indicates that there may be an implicit logical relationship between the two clauses, such as a causal relationship or a complementary relationship. Through the above method, a sequence of target review clauses with logical relationships is obtained, which demonstrates the logical relationships between the review clauses.

[0049] Step S1223: Convert the text of each target review condition in the target review clause sequence with logical relationship into a word embedding vector, calculate the similarity between each word embedding vector, and generate logical connection parameters between clauses. The logical connection parameters between clauses include similarity values ​​and logical relationship weights. The similarity values ​​are standardized to ensure that the value range is consistent, and the logical relationship weights are assigned according to the preset relationship priority.

[0050] After obtaining the logically related sequence of target review clauses, the text of each target review condition in the sequence is converted into a word embedding vector. Word embedding vectors are a method of converting textual information into vector representations, which can capture the semantic information of the text.

[0051] The similarity between each word embedding vector is calculated. This similarity can be determined using various similarity calculation methods, such as cosine similarity or Euclidean distance. Based on the calculated similarity, logical connection parameters between clauses are generated. These parameters include both the similarity score and the logical relationship weight.

[0052] Similarity scores need to be standardized to ensure a consistent range, facilitating subsequent comparisons and analysis. Logical relationship weights are assigned based on preset relationship priorities, with different relationship types having different priorities. For example, causal relationships may have a higher priority than parallel relationships, thus resulting in a relatively higher weight for causal relationships. These logical connection parameters between clauses allow for a more accurate description of the logical relationships between the reviewed clauses.

[0053] Step S1224: Perform a depth-first traversal on the hierarchical tree structure, and generate review logic features containing clause level numbers and logical connection parameters according to the descending order of the review clauses and the logical relationship type. Each element in the review logic features corresponds to a structural feature of a review clause. The review logic features establish a feature correspondence with the content theme features through the index order of document paragraphs.

[0054] The hierarchical tree structure is processed by depth-first traversal. Depth-first traversal is a method for traversing a tree structure. It starts from the root node and traverses along the depth direction of the tree until all nodes have been traversed.

[0055] During the traversal, a clause level number is generated for each review clause according to its descending hierarchical order and logical relationship type. The clause level number effectively represents the position of the review clause within the hierarchical tree structure. Simultaneously, combined with the previously calculated logical connection parameters between clauses, a review logical feature containing both the clause level number and the logical connection parameters is generated.

[0056] Each element in the review logic features corresponds to a structural feature of a review clause. These structural features include clause hierarchy information, logical relationship information, etc. Furthermore, the review logic features establish a feature correspondence with the content theme features through the index order of document paragraphs. This allows the logical structure of the review to be associated with the document's content theme.

[0057] For example, step S12241: Start from the root node of the hierarchical tree structure and perform a depth-first traversal operation, visit each node in turn and record the text content of the review clauses corresponding to the node. During the visit, sort the nodes in descending order according to their hierarchical depth to generate a review clause traversal sequence with a sequential visit order.

[0058] A depth-first traversal is performed starting from the root node of the hierarchical tree structure. The depth-first traversal visits nodes as deeply as possible along the branches of the tree until it cannot continue, then backtracks to the previous node and continues visiting other branches.

[0059] During the traversal, each node is visited sequentially, and the text content of the corresponding review clause is recorded. Simultaneously, the nodes are sorted in descending order of their hierarchy depth. Hierarchy depth represents the node's position in the hierarchical tree structure; the deeper the hierarchy, the larger the node's depth value. This descending sorting generates a review clause traversal sequence with a sequential access order. This review clause traversal sequence effectively demonstrates the access order of the review clauses.

[0060] Step S12242: For each review clause node in the review clause traversal sequence, extract the hierarchical depth information and parent node reference information of the review clause node, and generate a unique corresponding clause hierarchical number in combination with the preset hierarchical numbering rules. The hierarchical numbering rules include the root node hierarchical identifier, the child node hierarchical progression identifier, and the node's order identifier in the same level.

[0061] For each review clause node in the review clause traversal sequence, extract its hierarchy depth information and parent node reference information. The hierarchy depth information indicates the node's position in the hierarchy tree structure, and the parent node reference information is used to determine the node's parent node.

[0062] A unique clause level number is generated based on a pre-defined level numbering rule. This rule includes a root node level identifier, child node level progression identifiers, and the node's order within the same level. The root node level identifier indicates the root node's level; the child node level progression identifier indicates the child node's hierarchical relationship relative to its parent node; and the node's order within the same level distinguishes different nodes within the same level. For example, the root node's level number might be "1," its first child node's level number might be "1.1," its second child node's level number might be "1.2," and so on. This method generates a unique clause level number for each review clause node.

[0063] Step S12243: While generating the clause level number, query the explicit or implicit logical relationship type between the review clause node and its adjacent nodes. The explicit logical relationship type is determined by matching the logical connector dictionary, and the implicit logical relationship type is determined by calculating the word vector similarity of the clause text.

[0064] When generating the clause hierarchy number, it is necessary to query the logical relationship type between the review clause node and its adjacent nodes. Logical relationship types are divided into explicit logical relationship types and implicit logical relationship types.

[0065] Explicit logical relation types are determined through matching with a dictionary of logical connectors. This dictionary contains common logical connectors and their corresponding logical relation types, such as "and" for parallel relationships and "or" for alternative relationships. Explicit logical relation types are determined by searching for logical connectors in the review clause text and matching them with the dictionary.

[0066] For cases where there are no obvious logical connectors, i.e., implicit logical relationships, the type of relationship is determined by calculating the word vector similarity of the clause text. The clause text is converted into word vectors, and the similarity between word vectors of adjacent clause texts is calculated. Based on the similarity, the type of implicit logical relationship, such as causal relationship or complementary relationship, is determined.

[0067] Step S12244: Based on the determined logical relationship type, extract the corresponding logical connection parameters from the pre-established logical relationship parameter mapping table. The logical connection parameters include a relationship type identifier, a relationship strength descriptor, and a relationship scope identifier.

[0068] After determining the logical relationship types between the review clause nodes, the corresponding logical connection parameters can be extracted based on a pre-established logical relationship parameter mapping table. This logical relationship parameter mapping table is pre-built and records various parameter information corresponding to different logical relationship types.

[0069] A relation type identifier is used to explicitly indicate the specific type of logical relationship, such as using specific letters or symbols to represent parallel or causal relationships. A relation strength descriptor describes the strength of this logical relationship. Determining the relation strength descriptor requires considering multiple factors, such as the word vector similarity calculated earlier. A high word vector similarity indicates a close logical connection between the two review clauses, and therefore the relation strength descriptor represents a relatively strong relationship.

[0070] The scope identifier of the relationship is used to define the scope of the clauses involved in the above logical relationship, that is, to clarify which clauses are affected by the logical relationship. By extracting these logical connection parameters from the logical relationship parameter mapping table, the logical relationship between the review clauses can be more accurately characterized.

[0071] Step S12245: Associate and bind the clause hierarchy number of each review clause node with the corresponding logical connection parameter, generate sequentially arranged feature elements according to the order of the review clause traversal sequence, and all feature elements together constitute a review logical feature containing the clause hierarchy number and logical connection parameter. Each feature element in the review logical feature corresponds one-to-one with the review clause node, and establishes a feature position correspondence relationship with the content theme feature through the position index of the node in the hierarchical tree structure.

[0072] Next, the clause hierarchy number of each review clause node is associated and bound with its corresponding logical connection parameter. This process tightly combines the clause hierarchy number, which represents the location information of the review clause, with the logical connection parameter, which reflects the logical relationship, to form a complete feature information unit.

[0073] Then, following the order of the previously generated review clause traversal sequence, these associated and bound feature elements are arranged sequentially. The purpose of this arrangement is to ensure that the order of the feature elements matches the access order of the review clauses, facilitating subsequent processing and analysis.

[0074] All these sequentially arranged feature elements together constitute the review logic feature, which includes clause hierarchy numbers and logical connection parameters. Each feature element in the review logic feature has a one-to-one correspondence with a review clause node, thereby ensuring that the logical information of each review clause is accurately reflected in the feature.

[0075] Simultaneously, by indexing the nodes' positions within the hierarchical tree structure, a correspondence between the review logic features and the content theme features is established. This means that, based on the position of the review clauses within the hierarchical tree structure, their logical features can be associated with their corresponding content theme features, achieving a mutual mapping between the logical structure and the content theme.

[0076] Step S130: Call the pre-trained AI big model to perform joint summary generation processing on the content topic features and review logic features to generate preliminary summary text for each procurement review document unit. The preliminary summary text contains a word sequence generated based on the key information of the document.

[0077] After obtaining the content theme features and review logic features of each procurement review document unit, a pre-trained AI model is needed for joint summary generation. This pre-trained AI model, trained on a large amount of data, possesses powerful language understanding and generation capabilities.

[0078] Combining content theme features and review logic features into the AI ​​model is beneficial because these two features reflect the core information of procurement review documents from different perspectives. Content theme features highlight the main topics and key content of the document, while review logic features demonstrate the logical structure of the review process and the relationships between clauses. By using these two features together, the AI ​​model can more comprehensively and accurately understand the core content of the document, thereby generating more targeted preliminary summary text.

[0079] The preliminary summary text is a sequence of words generated based on the key information of the document. These word sequences are organized and arranged to summarize the main content and key points of the procurement review document. The generated preliminary summary text retains the important information in the document while removing redundant and irrelevant content.

[0080] Step S131: Input the content theme features and review logic features that establish feature correspondence into the input layer of the AI ​​large model, perform standardized mapping, and then generate a joint input feature vector containing the corresponding features through a concatenation operation.

[0081] First, since the content theme features and review logic features establish a feature correspondence through the index order of document paragraphs, these are input into the input layer of the AI ​​large model. At the input layer, these two types of features need to undergo standardized mapping. The purpose of standardized mapping is to transform features from different sources and scales into a unified representation space, making them comparable and consistent.

[0082] The standardization mapping process may involve normalizing features or performing other transformations to ensure that the range and distribution of feature values ​​meet the input requirements of large AI models. After standardization mapping, the content theme features and review logic features are concatenated. This concatenation operation links these features together in a predetermined order to form a new vector, the joint input feature vector. This joint input feature vector integrates information from both the content theme and review logic aspects.

[0083] Step S132: Perform context semantic modeling on the joint input feature vector through the multi-layer Transformer encoding layer of the AI ​​large model to generate a context feature matrix containing position encoding information.

[0084] After the joint input feature vectors enter the multi-layer Transformer encoding layer of a large AI model, contextual semantic modeling can be performed. The Transformer encoding layer is an important component of large AI models, possessing a powerful ability to capture contextual information in sequences.

[0085] Step S1321: Perform position encoding processing on the joint input feature vector, and generate a position encoding vector containing the order information of the review clauses using a fixed position encoding method.

[0086] Before performing contextual semantic modeling, positional encoding of the joint input feature vector is necessary. Because the order of words is crucial for semantic expression in natural language, positional encoding is needed to preserve the order information of the review clauses. A fixed positional encoding method is used to generate a positional encoding vector based on the position of the review clause corresponding to each element in the joint input feature vector. This positional encoding vector assigns a position-related encoding value to each element, enabling the model to distinguish elements in different positions and thus better understand the order and contextual relationships of the review clauses.

[0087] Step S1322: Perform element-wise fusion processing on the location encoding vector and the joint input feature vector to obtain an input feature vector containing location information.

[0088] The generated positional encoding vector is then fused element-wise with the joint input feature vector. Element-wise fusion involves performing some operation, commonly addition, on corresponding elements in the positional encoding vector and the joint input feature vector. Through this fusion process, positional information is incorporated into the joint input feature vector, resulting in an input feature vector that includes positional information. Therefore, when processing the input feature vector, the model can not only acquire information about the features themselves but also utilize positional information to better understand the order and contextual relationships between elements.

[0089] Step S1323: In each coding block of the Transformer coding layer, multi-head self-attention calculation is performed on the input feature vector, dividing the input feature vector into multiple heads, each head independently calculating attention weights, and generating a multi-head attention output vector, which contains semantic representations of different subspaces.

[0090] In each encoding block of the Transformer encoding layer, multi-head self-attention computation can be performed on the input feature vector containing positional information. The multi-head self-attention mechanism is one of the core mechanisms of the Transformer model, which enables the model to pay attention to the relationships between elements at different positions in the sequence when processing the input sequence.

[0091] The specific approach involves dividing the input feature vector into multiple heads, each of which independently calculates attention weights. These attention weights represent the degree of correlation between each element in the input sequence and other elements. The attention weights calculated by each head reflect the semantic representation of a different subspace; that is, different heads focus on different aspects of the input sequence. Through parallel computation of multiple heads, the semantic information of the input sequence can be captured from multiple perspectives, generating a multi-head attention output vector. This multi-head attention output vector contains semantic representations of different subspaces, enabling the model to more comprehensively understand the contextual information of the input sequence.

[0092] Step S1324: Perform feedforward neural network processing on the multi-head attention output vector, and perform feature transformation through a nonlinear activation function to generate an intermediate feature vector after nonlinear transformation.

[0093] After obtaining the multi-head attention output vector, it can be processed by a feedforward neural network. A feedforward neural network is a network structure composed of multiple neuron layers, which can perform non-linear transformations on the input vector. During this process, non-linear activation functions can be used to perform feature transformations on the multi-head attention output vector. Non-linear activation functions introduce non-linear factors, enabling the model to learn more complex patterns and relationships.

[0094] By using a feedforward neural network and a nonlinear activation function, the multi-head attention output vector is transformed into an intermediate feature vector after nonlinear transformation. This intermediate feature vector contains semantic information after further processing and transformation.

[0095] Step S1325: The intermediate feature vector is standardized by layer standardization operation. The standardized feature vector is used as the input of the next coding block. After iterative processing of multiple coding blocks, a context feature matrix containing multiple layers of semantic information is finally generated.

[0096] To ensure the stability and convergence of model training, layer standardization is required on the intermediate feature vectors after nonlinear transformation. Layer standardization is a method of standardizing feature vectors, which adjusts each element of the feature vector to keep the mean and variance of the feature vector within a set range.

[0097] After layer normalization, the normalized feature vector is used as the input for the next encoding block. In each multi-layer Transformer encoding layer, the multi-head self-attention calculation, feedforward neural network processing, and layer normalization operations are repeated for each encoding block. Through iterative processing of multiple encoding blocks, the model can continuously perform deep semantic modeling of the input features, ultimately generating a context feature matrix containing multi-layer semantic information. This context feature matrix integrates semantic information at different levels, enabling a more comprehensive reflection of the contextual relationships and semantic content of the input sequence.

[0098] Step S133: Calculate the weight of each feature unit in the context feature matrix and generate a weight distribution matrix, which represents the importance of different feature units in semantic modeling.

[0099] After obtaining the context feature matrix containing multi-layered semantic information, it is necessary to calculate the weights of each feature unit. The weight calculation aims to determine the importance of different feature units in the semantic modeling process. Various methods can be used to calculate the weights, such as attention-based methods. The weights can be determined by calculating the degree of association between each feature unit and other feature units, or based on factors such as the feature unit's position and eigenvalues ​​in the context matrix.

[0100] The calculated weights of each feature unit are arranged in a predetermined order to generate a weight distribution matrix. Each element in the weight distribution matrix corresponds to a feature unit in the context feature matrix, and the value of the element represents the weight of that feature unit. This weight distribution matrix can intuitively demonstrate the importance of different feature units in semantic modeling; the larger the weight, the more important the feature unit in semantic modeling.

[0101] Step S134: Perform weighted aggregation processing on the context feature matrix according to the weight distribution matrix to obtain the key feature vector focusing on key review information.

[0102] Based on the generated weight distribution matrix, the context feature matrix is ​​subjected to weighted aggregation. Weighted aggregation involves multiplying each feature unit in the context feature matrix by its corresponding weight, and then performing some aggregation operation, commonly summation. This weighted aggregation process highlights feature units with higher weights, i.e., key review information.

[0103] The final result is a key feature vector that focuses on crucial review information. This key feature vector is the result of a weighted aggregation of the context feature matrix, which retains the most important review information in the context while removing some relatively unimportant information.

[0104] Step S135: Call the decoding layer of the AI ​​large model to perform sequence generation processing on the key feature vectors, and use an autoregressive generation method to generate preliminary summary text word by word.

[0105] Key feature vectors, focusing on crucial review information, are input into the decoding layer of the AI ​​model. The main task of the decoding layer is to generate preliminary summary text based on the input key feature vectors. In this process, an autoregressive generation method is used to generate text word by word.

[0106] Autoregressive generation refers to a method where, during text generation, a large AI model predicts the next word based on previously generated words and input key feature vectors. Specifically, starting from a given symbol, the model continuously predicts the most likely next word, adds it to the generated sequence, and then continues predicting the next word based on the new sequence and key feature vectors. This process is repeated until a complete preliminary summary text is generated. This autoregressive generation method can generate a lexical sequence that conforms to grammatical and semantic logic, forming a preliminary summary text.

[0107] Step S140: Perform semantic coherence optimization processing on the preliminary summary text to obtain a semantically coherent optimized summary text.

[0108] While the preliminary summary text encapsulates the key information of the procurement review document, it may lack semantic coherence. Therefore, semantic coherence optimization is needed to improve the quality and readability of the summary.

[0109] Step S141: Segment the preliminary summary text to obtain multiple summary sentence units.

[0110] First, the preliminary abstract text is segmented into sentences. Sentence segmentation involves dividing the preliminary abstract text into multiple independent abstract sentence units based on the grammatical rules and punctuation of natural language. Each abstract sentence unit expresses relatively complete semantic information. Through this process, subsequent analysis and optimization of the abstract text can be facilitated more easily.

[0111] Step S142: Calculate the semantic similarity between adjacent summary sentence units, generate a semantic similarity matrix, and construct a sentence coherence graph based on the semantic similarity matrix. The nodes in the sentence coherence graph are summary sentence units, the edges are the semantic similarity between summary sentence units, and the weight of the edge is the corresponding similarity value. During the construction process, edges with semantic similarity higher than a preset threshold are retained.

[0112] After obtaining multiple summary sentence units, it is necessary to calculate the semantic similarity between adjacent summary sentence units. Semantic similarity reflects the degree of semantic similarity between two sentences. The following steps can be used to calculate semantic similarity.

[0113] Step S1421: Perform word embedding processing on each summary sentence unit. Use a pre-trained word embedding model to convert each word in the sentence into a low-dimensional dense vector, generating a set of sentence word embedding vectors.

[0114] Word embedding processing is performed on each summary sentence unit using a pre-trained word embedding model. This model, trained on a large-scale corpus, converts each word in the sentence into a low-dimensional dense vector. This transformation converts the text information into vector form, facilitating subsequent computation and processing. Performing this transformation on each word in each summary sentence unit generates a sentence word embedding vector set containing the word embedding vectors of all words in the sentence.

[0115] Step S1422: The sentence word embedding vector set is reduced in dimensionality using the average pooling method, and the average value of all sentence word embedding vectors is calculated to generate the sentence semantic vector.

[0116] To further process the sentence word embedding vector set, average pooling can be used for dimensionality reduction. Average pooling calculates the average of the corresponding elements of all vectors in the sentence word embedding vector set. Through this calculation, multiple word embedding vectors are merged into a sentence semantic vector. The sentence semantic vector is a generalized representation of the semantic information of the entire sentence, reducing the dimensionality of the data while retaining the main semantic information of the sentence.

[0117] Step S1423: Calculate the semantic similarity value of the semantic vectors of any two adjacent sentences, and arrange the semantic similarity values ​​according to the order of the summary sentence units to generate a semantic similarity matrix containing row index and column index. The row index and column index in the semantic similarity matrix correspond to the sequential number of the summary sentence units. The elements in the semantic similarity matrix are the standardized results of the semantic similarity values ​​of the corresponding summary sentence unit pairs.

[0118] Calculate the semantic similarity between any two adjacent sentence semantic vectors. Various similarity calculation methods, such as cosine similarity, can be used to calculate the similarity between the two vectors. Arrange the calculated semantic similarity values ​​according to the order of the summary sentence units to generate a semantic similarity matrix. The row and column indices of this semantic similarity matrix correspond to the sequential numbers of the summary sentence units, and each element in the matrix represents the semantic similarity value of the corresponding summary sentence unit pair. To make different similarity values ​​comparable, these similarity values ​​can be standardized to a uniform range.

[0119] A sentence coherence graph is constructed based on the generated semantic similarity matrix. The sentence coherence graph is a graph structure where nodes are summary sentence units, edges represent the semantic similarity between summary sentence units, and the weight of each edge is its corresponding similarity value. During construction, edges with semantic similarity higher than a preset threshold can be retained. The preset threshold is a pre-defined similarity standard; only when the semantic similarity between two summary sentence units exceeds this threshold will their edges be retained in the sentence coherence graph. This filters out edges with weak semantic connections, making the sentence coherence graph more accurately reflect the main semantic relationships between summary sentence units.

[0120] Step S143: Use a graph neural network model to learn node features of the sentence coherence graph, update the feature representation of the nodes through graph convolution operation, and generate a coherence feature vector for each summary sentence unit. This coherence feature vector reflects the semantic coherence of the sentence in the overall summary.

[0121] Graph neural network models are used to learn node features in sentence coherence graphs. Graph neural networks are neural network models specifically designed for processing graph-structured data, which can learn feature representations of nodes using information from nodes and edges in sentence coherence graphs.

[0122] In graph neural network models, graph convolution operations can be performed. Graph convolution updates the feature representation of a node based on information from its neighboring nodes. Specifically, for each node (i.e., the summary sentence unit) in the sentence coherence graph, information from its neighboring nodes and edge weights (i.e., semantic similarity) can be considered, and the feature representation of that node can be updated through calculation and transformation.

[0123] After multiple graph convolution operations, each summary sentence unit generates a coherence feature vector. This coherence feature vector reflects the semantic coherence of the summary sentence unit within the overall summary. The larger the value of the coherence feature vector, the stronger the semantic connection between the summary sentence unit and other summary sentence units, and the better the coherence within the overall summary.

[0124] Step S144: Adjust the order of the summary sentence units according to the coherence feature vector, arrange sentences with semantic coherence greater than the set coherence threshold adjacently, detect and delete semantically repetitive summary sentence units, and generate semantically coherent optimized summary text.

[0125] The order of summary sentence units can be adjusted based on the generated coherence feature vector. For example, a coherence threshold can be set, and sentences with semantic coherence greater than the threshold can be arranged adjacently, thereby making the summary text more semantically coherent and in line with the expression habits of natural language.

[0126] Meanwhile, it is possible to detect and delete semantically repetitive abstract sentence units. Semantically repetitive abstract sentence units make the abstract text appear redundant and reduce readability. By comparing methods such as the semantic similarity of abstract sentence units, find and delete the semantically repetitive abstract sentence units. After the above adjustments and processing, an optimized abstract text with semantic coherence is finally generated. This optimized abstract text is more fluent and coherent semantically and can better summarize the main content of the procurement review document.

[0127] Step S150: Generate a final procurement review document abstract containing a set of topic keywords based on the optimized abstract text, and output the final procurement review document abstract to the target storage terminal in a preset format, where the preset format includes a standardized layout of a title, a body text, and a keyword list.

[0128] After obtaining the optimized abstract text with semantic coherence, it is necessary to further generate a final procurement review document abstract containing a set of topic keywords. The set of topic keywords can highlight the core theme and key information of the procurement review document, enabling readers to quickly grasp the core points of the document.

[0129] Step S151: Perform keyword extraction processing on the optimized abstract text to generate a candidate set of topic keywords, where the terms in the candidate set of topic keywords include nouns, noun phrases, and professional terms in the procurement review field.

[0130] In order to extract the candidate set of topic keywords from the optimized abstract text, a series of processing is required. First, the optimized abstract text can be subjected to词性 analysis to identify the nouns and noun phrases therein. Nouns and noun phrases often can summarize the main objects and concepts in the text and are important sources of keywords.

[0131] Meanwhile, special attention can be paid to the professional terms in the procurement review field. These professional terms are the core vocabulary in the procurement review document and accurately express the specific concepts and business content in this field. By identifying and extracting the professional terms, it can be ensured that the candidate set of topic keywords covers the key information in the procurement review field.

[0132] During the extraction process, some preprocessing operations are also carried out, such as removing stop words and general terms. Stop words are those words that frequently appear in natural language but contribute little to the core semantics of the text, such as "of", "is", "in", etc.; general terms are words that are widely used in multiple fields and may not accurately reflect the specific theme of the procurement review document. By removing these words, noise interference can be reduced and the accuracy of keyword extraction can be improved.

[0133] Step S1511: Identify the nouns and noun phrases in the optimized abstract text and use them as candidate keyword units.

[0134] When processing optimized abstract text, part-of-speech tagging (POS) can be used to identify nouns and noun phrases. POS is a natural language processing technique that assigns a part-of-speech tag to each word based on its grammatical function and semantic features, such as noun, verb, or adjective. Through POS tagging, nouns and noun phrases in the optimized abstract text can be accurately identified and used as candidate keywords. These candidate keywords are words or phrases that have been initially selected as potential subject keywords.

[0135] Step S1512: Remove stop words and generic terms from the keyword candidate units to obtain a preliminary keyword candidate set.

[0136] After obtaining the keyword candidate units, stop words and generic terms need to be removed. A stop word list and a generic term list can be pre-established, and the keyword candidate units can be compared against these lists. If a keyword candidate unit appears in either the stop word list or the generic term list, it is removed from the candidate units. Through this process, a preliminary keyword candidate set can be obtained, with the vocabulary in this preliminary keyword candidate set more focused on the core content of the procurement review document.

[0137] Step S1513: Calculate the frequency of each preliminary keyword candidate in the procurement review document set, and generate the candidate keyword frequency distribution. The keyword frequency distribution is obtained by counting the number of times the keyword appears in all documents and standardizing the data.

[0138] To further filter out valuable keywords, it's necessary to calculate the frequency of each preliminary keyword candidate within the procurement review document set. Specifically, this involves counting the number of times each preliminary keyword candidate appears in the entire procurement review document set. Then, to ensure comparability of the frequencies of different keywords, these occurrences can be standardized. Standardization transforms the frequencies to a uniform range, such as between 0 and 1. This process generates a candidate keyword frequency distribution. The candidate keyword frequency distribution reflects the prevalence of each preliminary keyword candidate within the procurement review document set.

[0139] Step S1514: Based on the frequency distribution of the candidate keywords and the preset keyword importance scoring criteria, select keywords that have a defined meaning in the procurement review field as candidate topic keywords.

[0140] The initial candidate keyword set is screened based on the frequency distribution of the generated candidate keywords and the preset keyword importance scoring criteria. The keyword importance scoring criteria are preset based on professional knowledge and business needs in the procurement review field, and take into account factors such as the frequency of keyword occurrence and its importance in the field.

[0141] If a preliminary keyword candidate appears frequently and has a clear and significant meaning in the procurement review field, it is likely to be selected as a candidate topic keyword. Through the above screening process, more representative and important keywords can be selected from the preliminary keyword candidate set.

[0142] Step S1515: Perform synonym merging on the candidate topic keywords to generate a candidate set of topic keywords.

[0143] After obtaining candidate keywords, they can be processed by merging synonyms. This is because natural language contains many synonyms or near-synonyms that express the same or similar concepts. Without this merging process, redundant information can exist in the candidate keyword set, affecting subsequent analysis and application.

[0144] The specific method of synonym merging is to identify synonyms or near-synonyms among candidate topic keywords using techniques such as semantic similarity calculation, and then merge them into a single keyword. For example, if "supplier reputation" and "supplier credit" are considered synonyms, they are merged into one keyword. Through synonym merging, a final set of candidate topic keywords is generated, which contains more refined and accurate keywords.

[0145] Step S152: Calculate the term frequency and inverse document frequency of each topic keyword in the optimized summary text of the candidate topic keyword set. The term frequency is obtained by dividing the number of occurrences of the topic keyword by the total number of words in the optimized summary text. The inverse document frequency is calculated by using a logarithmic function to determine the rarity of the topic keyword in the procurement review document set. The product of the term frequency and the inverse document frequency of the topic keyword is used as the importance score of the topic keyword.

[0146] After obtaining the candidate set of topic keywords, it is necessary to calculate the importance score for each topic keyword. The importance score is obtained by calculating the product of term frequency and inverse document frequency.

[0147] Term frequency (TF) calculation involves counting the number of times a keyword appears in the optimized summary text, and then dividing that number by the total number of words in the summary text. TF reflects how often a keyword appears in the optimized summary text; the higher the frequency, the more important the keyword is in the text.

[0148] Inverse document frequency (IVF) measures the rarity of a keyword within a procurement review document set. It is calculated using a logarithmic function to analyze the occurrence of a keyword in the procurement review document set. A low number of documents containing a keyword indicates rarity and a high IVF; conversely, a high number of documents containing a keyword indicates a low IVF.

[0149] The importance score of a topic keyword is obtained by multiplying its term frequency by its inverse document frequency. This importance score comprehensively considers both the frequency of the keyword in the optimized summary text and its rarity within the entire procurement review document set, enabling a more accurate assessment of keyword importance.

[0150] Step S153: Sort the candidate set of topic keywords according to the importance score, and select target topic keywords with an importance score higher than a preset threshold to form a target topic keyword set.

[0151] Based on the calculated importance scores, the candidate set of topic keywords is sorted. The sorting can be done from highest to lowest importance score, ensuring that more important keywords appear first.

[0152] Then, a preset threshold is set to filter out target topic keywords with importance scores higher than this threshold. The preset threshold is pre-set based on actual needs and experience, and it is used to control the quantity and quality of target topic keywords. Through filtering, the most important keywords can be selected from the candidate set of topic keywords, forming a target topic keyword set. The keywords in this target topic keyword set can most accurately summarize the core theme of the optimized abstract text.

[0153] Step S154: The set of subject keywords is merged with the optimized summary text, and the merged summary text is formatted according to the preset document format to generate a final procurement review document summary containing a title, body text and a list of subject keywords. The title is generated based on the combination of core keywords in the set of subject keywords, the body text retains the semantic coherence of the optimized summary text, and the list of subject keywords is arranged from high to low according to the importance score.

[0154] The target topic keyword set is integrated with the optimized abstract text. This integration can be achieved by inserting the target topic keywords into appropriate positions within the optimized abstract text, or by highlighting these keywords within the abstract text, thereby enhancing the abstract text's expression of the core topic.

[0155] The merged abstract text is formatted according to a preset document format. The preset document format is a standardized layout that includes a title, body text, and a list of subject keywords.

[0156] The title is generated by combining core keywords from the set of topic keywords. Core keywords are those keywords that best represent the core theme of the procurement review document within the target topic keyword set. By combining and adjusting these core keywords, a title that accurately summarizes the document's theme can be generated.

[0157] The main body will retain the semantic coherence of the optimized abstract text, ensuring that the abstract is logically clear and fluently expressed. At the same time, appropriate formatting adjustments can be made to the main body, such as paragraph breaks and font settings, to improve readability.

[0158] The list of key words will be arranged from highest to lowest importance. This arrangement allows readers to quickly understand the core keywords of the procurement review document and their significance. Through this formatting process, a final summary of the procurement review document, including the title, body text, and list of key words, will be generated.

[0159] Step S155: Output the final procurement review document summary to the target storage terminal according to a preset format, wherein the preset format includes standardized layout of title, body text and keyword list.

[0160] After generating the final procurement review document summary, it is output to the target storage terminal according to the preset format. The target storage terminal can be an enterprise's internal document management system, a cloud storage platform, or other designated storage devices.

[0161] During the output process, the preset format is strictly followed to ensure that the title, body text, and keyword list are formatted correctly. This ensures that the final procurement review document summary stored on the target storage terminal is presented in a standardized and regulated format, facilitating subsequent review, analysis, and use. Whether it's internal management personnel, procurement staff, or other relevant personnel, they can quickly obtain high-quality procurement review document summaries through the target storage terminal, enabling more efficient decision-making and management.

[0162] The following details the construction and training of large AI models, as well as their integration with procurement review scenarios.

[0163] The construction of a large AI model mainly includes modules such as an input layer, multi-layer Transformer encoding layers, and decoding layers. The input layer receives processed feature data, which in this case are content-theme features and review logic features with established feature correspondences. These two types of features are first standardized and mapped at the input layer, transforming different features into a unified representation space to ensure comparability and consistency. Then, a joint input feature vector is generated through concatenation. During the standardization and mapping process, features can be normalized or transformed in other ways to ensure that the range and distribution of feature values ​​meet the model's input requirements.

[0164] The multi-layer Transformer encoding layer is the core of the model, used for contextual semantic modeling of the joint input feature vector. In the encoding layer, the joint input feature vector is first subjected to positional encoding. A fixed-positional encoding method is used to generate a positional encoded vector containing the order information of the review clauses. This vector is then fused element-wise with the joint input feature vector to obtain an input feature vector containing positional information. This is done because in natural language, the order of words is crucial for semantic expression, and positional encoding preserves the order information of the review clauses, helping the model better understand contextual relationships.

[0165] Within each encoding block, multi-head self-attention computation can be performed. The input feature vector is divided into multiple heads, each independently calculating attention weights. These attention weights reflect the degree of correlation between elements at different positions in the input sequence. Different heads focus on different aspects of the input sequence, thereby capturing the semantic information of the input sequence from multiple perspectives and generating a multi-head attention output vector. Next, the multi-head attention output vector is processed by a feedforward neural network, using a non-linear activation function to perform feature transformation, generating a non-linearly transformed intermediate feature vector. Finally, the intermediate feature vector is standardized through layer normalization, and the processed feature vector is used as the input to the next encoding block. After iterative processing across multiple encoding blocks, a context feature matrix containing multi-layered semantic information is finally generated.

[0166] The main task of the decoding layer is to generate preliminary summary text based on the input key feature vectors. It employs an autoregressive generation method to generate text word by word; that is, the large AI model predicts the next word based on previously generated words and the input key feature vectors, continuously iterating until a complete preliminary summary text is generated.

[0167] In training large-scale AI models, the first step is to prepare a large amount of training data. This training data consists of representative samples selected from a collection of procurement review documents, and these data must be labeled. The labeling process requires the joint participation of professional procurement reviewers and natural language processing experts to ensure the accuracy and consistency of the labeling. The labeled content includes document summary information, key themes, logical relationships, etc., and this labeled information will serve as the target output for training.

[0168] The training process typically employs optimization algorithms such as stochastic gradient descent. In each training iteration, the input data (i.e., content topic features and review logic features) is fed into the AI ​​model, which outputs preliminary summary text. This output is then compared with the labeled target summary, and a loss function is calculated. The loss function measures the difference between the AI ​​model's output and the target output. By continuously adjusting the parameters of the AI ​​model through optimization algorithms, the value of the loss function is gradually reduced, thereby improving the model's performance.

[0169] The parameter settings during training are also crucial. The learning rate is an important parameter that controls the step size of parameter updates in a large AI model. If the learning rate is too large, the model may skip the optimal solution; if the learning rate is too small, the model's convergence speed will be very slow. A common approach is to use a learning rate decay strategy, employing a larger learning rate at the beginning of training and gradually decreasing it as training progresses to improve training efficiency and stability.

[0170] Batch size is also an important parameter, determining the amount of data input to the large AI model during each training iteration. Choosing the right batch size requires considering both computational resources and the performance of the large AI model. Larger batch sizes can leverage the advantages of parallel computing to speed up training, but may cause the model to converge to local optima; smaller batch sizes can increase the model's randomness, helping it escape local optima, but training will be slower.

[0171] The number of training epochs refers to the number of times the model is trained on the entire training dataset. Too few epochs may prevent the large AI model from fully learning the features and patterns in the data; too many epochs may lead to overfitting, meaning the large AI model performs well on the training data but poorly on the test data. Therefore, methods such as cross-validation are needed to select an appropriate number of training epochs.

[0172] When applying AI large-scale models to procurement review scenarios, the input data consists of content theme features and review logic features processed through feature extraction. Content theme features include the probability distribution of keywords and the strength of inter-topic associations, reflecting the core content and main topics of the procurement review document. Review logic features include clause hierarchy numbers and logical connection parameters, embodying the reasoning and decision-making logic of the review process. By standardizing and concatenating these two types of features, a joint input feature vector is generated and input into the model.

[0173] The output data of the AI ​​large-scale model is a preliminary summary text, which is a word sequence generated based on the key information of the document, summarizing the main content and key viewpoints of the procurement review document. Next, the preliminary summary text undergoes semantic coherence optimization. This involves sentence segmentation, semantic similarity calculation, construction of a sentence coherence graph, and node feature learning using a graph neural network model. This process adjusts sentence order, removes semantically repetitive sentences, and generates a semantically coherent optimized summary text. Finally, based on the optimized summary text, a final procurement review document summary containing a set of thematic keywords is generated and output to the target storage terminal according to a preset format.

[0174] Figure 2 The illustration shows exemplary hardware and software components of a procurement review document summary generation system 100 incorporating an AI large model, which can implement the ideas of this application, according to some embodiments of this application. For example, processor 120 can be used in the procurement review document summary generation system 100 incorporating an AI large model and to perform the functions in this application.

[0175] The procurement review document summary generation system 100, which incorporates an AI big data model, can be either a general-purpose server or a special-purpose server; both can be used to implement the procurement review document summary generation method incorporating an AI big data model as described in this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0176] For example, the procurement review document summary generation system 100 incorporating AI big data models may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the procurement review document summary generation system 100 incorporating AI big data models may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The procurement review document summary generation system 100 incorporating AI big data models also includes an I / O interface 150 between the computer and other input / output devices.

[0177] For ease of explanation, only one processor is described in the AI-integrated big data procurement review document summary generation system 100. However, it should be noted that the AI-integrated big data procurement review document summary generation system 100 of this application may also include multiple processors, and therefore the steps performed by one processor described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the AI-integrated big data procurement review document summary generation system 100 performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0178] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned method for generating procurement review document summaries combined with AI large models is implemented.

[0179] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A procurement review document abstract generation method combined with an AI large model, characterized by, The method comprises: acquiring a set of procurement review documents to be processed, the set of procurement review documents comprising a plurality of procurement review document units with review time marks, each procurement review document unit corresponding to an independent procurement review business record; performing document feature extraction processing on the set of procurement review documents to obtain content theme features and review logic features of each procurement review document unit; calling a pre-trained AI large model to perform joint summary generation processing on the content theme features and the review logic features to generate a preliminary summary text of each procurement review document unit, the preliminary summary text comprising a word sequence generated based on document key information; performing semantic coherence optimization processing on the preliminary summary text to obtain a semantically coherent optimized summary text; generating a final procurement review document summary comprising a set of theme keywords based on the optimized summary text, and outputting the final procurement review document summary to a target storage terminal in a preset format, the preset format comprising a standardized layout of a title, a main text, and a keyword list; The document feature extraction processing on the set of procurement review documents to obtain the content theme features and the review logic features of each procurement review document unit comprises: inputting the procurement review document unit into a topic model for latent semantic analysis, calculating the probability distribution of each word unit in a preset theme space through iteration, and generating the content theme features of each procurement review document unit, the content theme features comprising theme word distribution probability and theme interrelation strength; performing logic structure analysis on the procurement review document unit to generate the review logic features; The calling of the pre-trained AI large model to perform joint summary generation processing on the content theme features and the review logic features to generate a preliminary summary text of each procurement review document unit comprises: inputting the content theme features and the review logic features with established feature correspondence into the input layer of the AI large model for standardized mapping and then generating a corresponding joint input feature vector through splicing operation; performing context semantic modeling processing on the joint input feature vector through the multi-layer Transformer encoding layer of the AI large model to generate a context feature matrix comprising position encoding information; calculating the weight of each feature unit in the context feature matrix to generate a weight distribution matrix, the weight distribution matrix representing the importance of different feature units in semantic modeling; performing weighted aggregation processing on the context feature matrix according to the weight distribution matrix to obtain a key feature vector focusing on key review information; calling the decoding layer of the AI large model to perform sequence generation processing on the key feature vector, and generating a preliminary summary text in a word-by-word manner using an autoregressive generation method; The semantic coherence optimization processing on the preliminary summary text to obtain a semantically coherent optimized summary text comprises: performing sentence segmentation processing on the preliminary summary text to obtain a plurality of summary sentence units; computing semantic similarity between adjacent summary sentence units, generating a semantic similarity matrix, and constructing a sentence coherence graph according to the semantic similarity matrix, wherein nodes in the sentence coherence graph are summary sentence units, edges are semantic similarity between summary sentence units, and weights of edges are corresponding similarity values, and edges with a semantic similarity higher than a preset threshold are reserved during the construction process; adopting a graph neural network model to perform node feature learning on the sentence coherence graph, updating feature representations of nodes through graph convolution operations, and generating a coherence feature vector of each summary sentence unit, which reflects a semantic coherence degree of the sentence in the overall summary; adjusting an order of the summary sentence units according to the coherence feature vector, arranging sentences with a semantic coherence degree greater than a set coherence degree threshold adjacently, and detecting and deleting semantically repetitive summary sentence units to generate an optimized summary text with semantic coherence.

2. The method of claim 1, wherein, inputting the procurement review document units into a topic model for latent semantic analysis, generating content topic features of each procurement review document unit by iteratively calculating a probability distribution of each word unit in a preset topic space, including: adopting a probabilistic generative model to define a probability distribution relationship between topics and word items, establishing a latent semantic space of the topic model, mapping each of the procurement review document units to the latent semantic space, and generating an initial topic vector reflecting topic attribution of each word item, wherein a dimension of the initial topic vector corresponds to a preset number of topics; calculating a semantic correlation degree between topic words in the semantic space of the initial topic vector, wherein the semantic correlation degree has a value range of 0 to 1, and a higher semantic correlation degree indicates a higher co-occurrence probability of a word item in a topic; performing clustering processing on the initial topic vector according to the semantic correlation degree, and aggregating the clustered topic vectors into a plurality of topic cluster units by using an unsupervised clustering algorithm, wherein each topic cluster unit represents an independent content topic; extracting top-K word items in a sorted list in descending order of frequency of occurrence in each topic cluster unit as topic words, generating a topic word distribution probability by statistically implementing frequency of occurrence of the topic words in the procurement review document units and performing standardization processing, and using the topic word distribution probability to measure an importance degree of a topic word in a corresponding topic; analyzing a semantic overlap degree between different topic cluster units, generating a topic interrelation strength descriptor including an overlap word quantity and a semantic overlap area by statistically implementing a quantity of shared word items and calculating a sum of weights of the shared word items in respective topics, and using the topic interrelation strength descriptor to represent a correlation between topics; and, the logical structure analysis on the procurement review document units to generate the review logical features, including: identifying title hierarchical information in the procurement review document units, constructing a hierarchical tree structure of review clauses by using a nested relationship of hierarchical markers, wherein a root node of the hierarchical tree structure is a document as a whole, child nodes are a plurality of clauses in descending order of hierarchical order, and each node records clause text, hierarchical depth, and parent node reference. Traverse each review clause, extract the logical connectives therein, match the logical connectives based on a pre-constructed logical connective dictionary to determine the explicit logical relationship types between the respective review clauses, and for a candidate review clause for which no explicit logical relationship type exists, determine the implicit logical relationship type by calculating the word vector similarity of the clause text of the candidate review clause to obtain a target review clause sequence with logical relationships; Convert the clause text of each target review condition in the target review clause sequence with logical relationships into a word embedding vector, calculate the similarity between the respective word embedding vectors, and generate an inter-clause logical connection parameter, which includes a similarity value and a logical relationship weight, wherein the similarity value is standardized to have a consistent value range, and the logical relationship weight is assigned according to a pre-set relationship priority; Perform a depth-first traversal of the hierarchical tree structure, and generate a review logical feature containing a clause hierarchical number and a logical connection parameter in a hierarchical descending order of the review clauses and a logical relationship type, wherein each element in the review logical feature corresponds to a structural feature of a review clause, and the review logical feature establishes a feature correspondence relationship with the content theme feature through an index order of a document paragraph.

3. The method of claim 1, wherein, The context semantic modeling processing of the joint input feature vector by the multi-layer Transformer encoding layer of the AI large model generates a context feature matrix containing position encoding information, including: Perform position encoding processing on the joint input feature vector, and generate a position encoding vector containing review clause order information using a fixed position encoding method; Perform element-by-element fusion processing on the position encoding vector and the joint input feature vector to obtain an input feature vector containing position information; In each encoding block of the Transformer encoding layer, perform multi-head self-attention calculation processing on the input feature vector, divide the input feature vector into multiple heads, independently calculate the attention weight for each head, and generate a multi-head attention output vector containing semantic representations of different subspaces; Perform feedforward neural network processing on the multi-head attention output vector, and generate an intermediate feature vector after nonlinear transformation by a nonlinear activation function; Perform standardization processing on the intermediate feature vector by layer normalization operation, and use the feature vector after standardization processing as the input of the next encoding block. After iterative processing of multiple encoding blocks, a context feature matrix containing multi-layer semantic information is finally generated.

4. The method of claim 1, wherein, The calculation of the semantic similarity between adjacent summary sentence units generates a semantic similarity matrix, including: Perform word embedding processing on each summary sentence unit, convert each word in the sentence into a low-dimensional dense vector using a pre-trained word embedding model, and generate a sentence word embedding vector set; Perform dimensionality reduction processing on the sentence word embedding vector set using an average pooling method, calculate the element average value of all sentence word embedding vectors, and generate a sentence semantic vector; The semantic similarity values of any two adjacent sentence semantic vectors are calculated, the semantic similarity values are arranged according to the order of the summary sentence units, a semantic similarity matrix containing row indexes and column indexes is generated, the row indexes and the column indexes correspond to the sequential numbers of the summary sentence units, and the elements in the semantic similarity matrix are the standardization results of the semantic similarity values of the corresponding summary sentence unit pairs.

5. The method of claim 1, wherein, The final procurement review document summary containing the theme keyword set is generated based on the optimized summary text, including: The keyword extraction processing is performed on the optimized summary text to generate a theme keyword candidate set, and the terms in the theme keyword candidate set include nouns, noun phrases and procurement review field professional terms. The term frequency and the inverse document frequency of each theme keyword in the theme keyword candidate set in the optimized summary text are calculated, the term frequency is obtained by dividing the number of occurrences of the theme keyword by the total number of words in the optimized summary text, the inverse document frequency is calculated by a logarithmic function to calculate the rarity of the theme keyword in the procurement review document set, and the product of the term frequency and the inverse document frequency of the theme keyword is used as the importance score of the theme keyword. The theme keyword candidate set is sorted according to the importance score, and target theme keywords with importance scores higher than a preset threshold are screened out to form a target theme keyword set. The theme keyword set is fused with the optimized summary text, and the fused summary text is typeset according to a preset document format to generate a final procurement review document summary containing a title, a main text and a theme keyword list, wherein the title is generated according to the core keyword combination in the theme keyword set, the main text retains the semantic coherence of the optimized summary text, and the theme keyword list is arranged from high to low according to the importance score.

6. The method of claim 5, wherein, The keyword extraction processing is performed on the optimized summary text to generate a theme keyword candidate set, including: Nouns and noun phrases in the optimized summary text are identified as keyword candidate units; Stop words and general terms in the keyword candidate units are removed to obtain a preliminary keyword candidate set; The occurrence frequency of each preliminary keyword candidate in the procurement review document set is calculated to generate a candidate keyword frequency distribution, and the keyword frequency distribution is obtained by standardizing the number of occurrences of the keyword in all documents; According to the candidate keyword frequency distribution and a preset keyword importance score standard, keywords with a set meaning in the procurement review field are screened out as candidate theme keywords; The candidate theme keywords are subjected to synonym merging processing to generate a theme keyword candidate set. 7.A system for generating a summary of a procurement review document in combination with an AI large model, characterized by, A processor and a memory are included, the memory and the processor are connected, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to realize the method in any one of claims 1-6.

Images