A multi-layer dynamic aggregated large language model driven letter text analysis method

Abstract

The application discloses a kind of multilayer dynamic polymerization big language model driven letter text analysis method, it is related to natural language processing technical field, including obtaining multi-format letter data and converting into text sequence with timestamp, text sequence is carried out semantic analysis and structure processing, generate text block set, text block set is carried out multi-task training and theme fusion, generate comprehensive semantic vector, based on comprehensive semantic vector constructs dynamic bridge graph, bridge graph is activated associated text block by semantic and time sequence correlation, respond to user query, the content of activated text block is spliced as context Prompt and input big language model, obtain the answer generated once;It has beneficial effects that: through dynamic bridge graph, semantic and time sequence information are associated, and based on context Prompt, the accurate response of big language model is realized, has the advantages of efficient capture complex semantic association, improve context understanding coherence, realize dynamic information aggregation.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method for analyzing correspondence text driven by a large language model with multi-layer dynamic aggregation. Background Technology

[0002] In the context of rapid development in information technology, enterprises and organizations need to receive and process a large number of structured or unstructured correspondence texts in their daily operations, such as business emails, project reports, approval records, complaints and suggestions. These correspondences come from a wide range of sources, have diverse formats, complex semantic levels, and loose structures. They often contain information links across multiple paragraphs, sentences, or even paragraphs, and change rapidly with business needs.

[0003] Traditional text analysis methods often rely on statistical learning or rule-based natural language processing techniques, typically employing keyword matching, topic modeling, or shallow sentiment analysis. While these methods are suitable for processing single sentences or standard-format texts, they often struggle to accurately capture deep semantic relationships when faced with complex discourse structures, contextual dependencies, non-standard language styles, and flexible expressions in real-world correspondence. In particular, they lack the ability to model implicit connections between context, time sequence, and linguistic context, resulting in fragmented information extraction, inconsistent responses, and poor scalability.

[0004] In recent years, pre-trained large language models such as BERT and GPT have made significant progress in natural language processing tasks, possessing powerful contextual understanding and generation capabilities. However, how to efficiently apply them to loosely structured and lengthy correspondence texts still faces the following key challenges:

[0005] 1) Unclear text structure and uneven information density will cause "redundancy + missing information" when directly input into a large model.

[0006] 2) There are semantic bridging relationships between contexts, but there is a lack of structured expression.

[0007] 3) It is difficult to achieve one-time retrieval and centralized response based on queries.

[0008] Therefore, a method for analyzing correspondence text driven by a large language model with multi-layer dynamic aggregation is proposed. Summary of the Invention

[0009] In view of the above-mentioned prior art, this application is hereby proposed. Embodiments of this application provide a method for analyzing correspondence text driven by a large language model with multi-layered dynamic aggregation, which has the advantages of efficiently capturing complex semantic relationships, improving contextual understanding coherence, and realizing dynamic information aggregation.

[0010] According to one aspect of this application, a method for analyzing correspondence text driven by a large language model with multi-layer dynamic aggregation is provided, comprising: acquiring multi-format correspondence data and converting it into a text sequence with timestamps; performing semantic analysis and structural processing on the text sequence to generate a set of text blocks containing temporal sequence and context; performing multi-task training and topic fusion on the text blocks to generate a comprehensive semantic vector; constructing a dynamic bridging graph based on the comprehensive semantic vector, wherein the bridging graph activates associated text blocks through semantic and temporal correlation; responding to a user query, concatenating the content of the activated text blocks into a context prompt and inputting it into a large language model to obtain a one-time generated answer; receiving new correspondence data and updating the bridging graph according to the new correspondence data.

[0011] Compared with the prior art, the letter text analysis method driven by a large language model with multi-layer dynamic aggregation according to the embodiments of this application can associate semantic and temporal information through dynamic bridging graphs and realize the accurate response of the large language model based on the context prompt. It has the advantages of efficiently capturing complex semantic associations, improving the coherence of contextual understanding, and realizing dynamic information aggregation. Attached Figure Description

[0012] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0013] Figure 1 This is a flowchart of a multi-layered dynamic aggregation method for analyzing correspondence text driven by a large language model, as described in this invention.

[0014] Figure 2 This is a flowchart illustrating the bridging graph update process of a multi-layered dynamic aggregation method for analyzing correspondence text driven by a large language model, as described in this invention. Detailed Implementation

[0015] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0016] Application Overview

[0017] In traditional natural language processing systems, the heterogeneity of multi-format email data makes information integration difficult, and temporal relationships and contextual dependencies are hard to model effectively. Semantic analysis of text sequences is limited to local feature extraction and cannot establish dynamic relationships across text blocks, resulting in both input redundancy and semantic gaps in large language models, leading to information fragmentation and logical breaks in the generated responses. When processing long-spanning business email chains, the system cannot automatically identify the correlation between session identifiers and timestamps, resulting in a lack of continuity in cross-email query responses.

[0018] For example, a bank's customer service system needs to process complaint cases that include PDF contracts, scanned receipts, and email bodies. Scanned documents, after OCR conversion, suffer from date formatting errors, and nested references within the email body result in a loose structure. The existing system treats each email as an independent text block, failing to recognize cross-email discussions of fund flows and contract terms. When a user queries "reasons for approval delays," the system only retrieves the "process adjustment notification" from the most recent email, missing the scanned copy of the "system upgrade report" submitted by the technical department three weeks prior. The generated response fails to address approval queue congestion caused by version incompatibility.

[0019] If these issues are not addressed, semantic fragmentation in cross-modal data will lead to incomplete business decision-making, and the omission of key information will cause compliance risks. The lack of temporal correlation prevents historical messages from effectively participating in current conversation reasoning, resulting in factual contradictions in the content generated by the large language model. Insufficient dynamic context maintenance capabilities will reduce the system's stability in multi-turn interactions, requiring users to repeatedly modify query conditions to obtain valid information, significantly increasing the cost of human-computer collaboration.

[0020] To address the aforementioned challenges, this application first considers how to transform multi-format correspondence data into a unified temporal structure to eliminate heterogeneity interference. It then explores a phased enhancement processing scheme to address semantic fragmentation in text sequences and designs a dynamic association mechanism to capture implicit connections across text blocks. Traditional methods of directly inputting data into a large language model can lead to both redundancy and missing information. This application attempts to construct a temporalized text block set during the preprocessing stage, generate a comprehensive vector through hierarchical semantic fusion, and then establish a bridging graph structure to achieve association activation. To avoid fragmented query responses, this application adopts a context-based stitching strategy based on the bridging graph, centrally inputting activated content into the large model to generate coherent answers. For association changes introduced by new data, an incremental update mechanism for the bridging graph is designed to maintain dynamic adaptability.

[0021] Exemplary methods

[0022] Figure 1The illustration depicts a multi-layered dynamic aggregation method for analyzing correspondence text driven by a large language model, according to an embodiment of this application. The method includes: acquiring multi-format correspondence data and converting it into a timestamped text sequence; performing semantic analysis and structural processing on the text sequence to generate a set of text blocks containing temporal sequence and context; performing multi-task training and topic fusion on the text blocks to generate a comprehensive semantic vector; constructing a dynamic bridging graph based on the comprehensive semantic vector, wherein the bridging graph activates associated text blocks through semantic and temporal correlation; responding to a user query by concatenating the content of the activated text blocks into a context prompt and inputting it into the large language model to obtain a one-time generated answer; receiving new correspondence data and updating the bridging graph based on the new correspondence data.

[0023] The multi-format email data includes plain text files, document files, image files, and scanned documents. Specifically, OCR technology can be used to convert images and scanned documents into text, and all formats are encoded, cleaned, and timestamp information is added to generate a unified text sequence. Its purpose is to solve the problem that traditional methods are difficult to handle multi-source heterogeneous data, ensuring that subsequent processing has temporal correlation.

[0024] Semantic analysis and structural processing refer to segmenting, vectorizing, and semantically enhancing a unified text sequence to generate a set of text blocks containing temporal and contextual information. This can be achieved using pre-trained language models and dependency parsing. Its purpose is to eliminate semantic fragmentation caused by loose structures and improve the ability to model contextual relationships.

[0025] The comprehensive semantic vector refers to the semantic aggregation of text blocks through multi-task training and topic fusion. Specifically, it can be generated by joint optimization of masked language model loss, neighbor block discrimination loss, and cross-block contrast loss. Its role is to integrate multi-level semantic features and solve the problems of redundancy and missing information caused by uneven information density.

[0026] The dynamic bridging graph refers to a semantic and temporal correlation network constructed based on comprehensive semantic vectors. Specifically, node connections can be established by calculating cosine similarity or timestamp difference. Its function is to explicitly express implicit relationships across text blocks, providing structured support for contextual retrieval.

[0027] The context prompt refers to concatenating the text blocks activated by the user query in chronological order into the input of the large language model. This can be achieved using similarity matching and neighbor node activation strategies. Its purpose is to centralize relevant context at once, avoiding the problem of inconsistent responses in traditional methods.

[0028] Updating the bridging graph refers to dynamically adjusting semantic and temporal association edges based on new message data. Specifically, it involves repeatedly generating new text blocks and calculating similarity with existing nodes. Its purpose is to adapt to changes in business needs and ensure the model's continuous scalability.

[0029] The core innovation of this application lies in explicitly modeling the semantic and temporal relevance of the letter text through a dynamic bridging graph, generating context-coherent Prompt input by combining a multi-layer dynamic aggregation mechanism, and achieving a centralized response at one time using a large language model, thus overcoming the shortcomings of traditional methods in information fragmentation and inconsistent response under complex discourse structures.

[0030] The working process and principle of this application are as follows: First, multi-format email data is acquired and converted into a time-stamped text sequence. This step unifies the email data of different formats into a processable text form and adds time information to preserve temporal relationships. Next, semantic analysis and structural processing are performed on the text sequence to generate a set of text blocks containing temporal and contextual information. This step identifies semantic units of the text through semantic analysis and organizes them into text blocks considering contextual relationships. Then, multi-task training and topic fusion are performed on the text blocks to generate a comprehensive semantic vector. This process extracts multi-dimensional features of the text blocks through multi-task learning and fuses topic information to generate a more comprehensive semantic representation. A dynamic bridging graph is constructed based on the comprehensive semantic vector. The bridging graph activates related text blocks through semantic and temporal correlations. The bridging graph structure explicitly represents the semantic and temporal relationships between text blocks, facilitating subsequent association activation. When responding to user queries, the content of the activated text blocks is concatenated into a contextual prompt and input into a large language model to obtain a one-time generated answer. This step activates related text blocks through the bridging graph and organizes them as input to the large language model, thereby generating a coherent answer. Finally, new letter data is received, and the bridging graph is updated based on the new letter data. This ensures that the system can dynamically adapt to newly added letter information and maintain the timeliness of the bridging graph.

[0031] As a preferred embodiment, the specific implementation of this application is as follows: First, the system receives multi-format email data, including emails, PDF documents, and scanned images. For image data, OCR processing is performed to extract text content. All text content is uniformly encoded into UTF-8 format and cleaned, such as removing special characters and standardizing date formats. Timestamp information is added to each text, forming a unified text sequence with timestamps. Next, a pre-trained language model, such as DistilBERT, is used to encode the text sequence to obtain word-level semantic vectors. Then, the text is segmented based on punctuation and semantic similarity to generate preliminary text blocks. Subsequently, dependency parsing and sentiment analysis are performed on each text block, and syntactic and sentiment features are fused to generate enhanced semantic vectors. Finally, the enhanced semantic vectors are input into the pre-trained model to obtain block-level semantics. The process involves clustering block-level semantic vectors using a clustering algorithm to generate topic distribution vectors by extracting keywords. These topic distribution vectors are then concatenated with the block-level semantic vectors to obtain a comprehensive semantic vector. A dynamic bridging graph is constructed based on the similarity and temporal information of the comprehensive semantic vectors. Nodes represent text blocks, and edges represent semantic or temporal relationships. When a user query is received, the query is encoded as a vector and matched with the comprehensive semantic vector for similarity. The most relevant text block is selected as the core block, and the adjacent nodes of the core block are activated through the bridging graph structure. The activated text block content is then concatenated temporally to form a context prompt. This prompt is input into a large language model such as GPT-3 to generate a coherent one-time response. When new email data is input, the above process is repeated to generate new text blocks and comprehensive semantic vectors, and the bridging graph structure is updated based on semantic similarity and temporal relationships.

[0032] Through the above-described scheme, this application achieves unified processing and semantic analysis of multi-format correspondence data, overcoming the limitations of traditional methods in handling complex discourse structures and contextual dependencies. The dynamic bridging graph structure effectively captures the semantic and temporal relationships between text blocks, improving the completeness and coherence of information extraction. The contextual prompt generation strategy based on activated text blocks fully utilizes the understanding and generation capabilities of large language models, while avoiding information redundancy and missing information caused by direct input. Furthermore, the dynamic update mechanism of the bridging graph ensures that the system can continuously adapt to newly added correspondence data, maintaining the timeliness of the analysis results. This method significantly improves the accuracy of information extraction and the coherence of responses when processing loosely structured and lengthy correspondence texts, providing more efficient and reliable technical support for the daily correspondence processing of enterprises and organizations.

[0033] In some of the solutions described above in this application, when acquiring multi-format letter data and converting it into a unified text sequence, non-text format files suffer from high recognition error rates, chaotic character encoding, and missing temporal information, making it difficult to accurately capture semantic relationships in subsequent processing steps.

[0034] This application further proposes a method for acquiring multi-format correspondence data and converting it into a timestamped text sequence, including: acquiring multi-format correspondence data, which includes plain text files, document files, image files, and scanned copies; performing grayscale conversion, noise reduction, and text region detection on image files and scanned copies, and generating editable text using OCR technology; and unifying and cleaning the plain text files, document files, and editable text, including correcting misrecognized characters by OCR, unifying date formats, and adding timestamp information, to generate a unified text sequence with timestamps.

[0035] In the image format processing, grayscale conversion and noise reduction techniques are used to reduce optical interference, morphological filtering is used to eliminate noise, and a text region localization algorithm based on connected component analysis is used to extract effective text regions. The OCR process uses a deep learning model to achieve character-level recognition, and a character confidence threshold is set at the output layer to filter low-confidence results. Encoding uniformly adopts the UTF-8 standard to convert text in different encoding formats. In the cleaning stage, regular expressions are used to match and replace date formats, and a confusion matrix is ​​used to replace and correct error-prone characters. Timestamp information is generated by parsing the creation time field in the file metadata or extracting time information according to file naming rules.

[0036] Specifically, when processing scanned documents, the RGB image is first converted to grayscale, and non-local mean filtering is used to eliminate scanning noise. Then, an edge detection algorithm is used to determine the boundaries of the text region. The OCR engine performs line-by-line recognition within the defined region, marking the position coordinates and confidence value of each character when outputting the text. For characters with a confidence value lower than a set threshold, dynamic correction is performed based on the semantic correlation between adjacent characters. During the encoding conversion stage, the character encoding type of the original file is detected and uniformly converted to UTF-8 format. Dates are uniformly recognized using regular expressions for various formats such as "YYYY-MM-DD" and "YYYY / MM / DD", and uniformly converted to the "YYYYMMDD" format. Timestamp appending generates time stamps accurate to the second based on the modification time recorded by the file system. When metadata is missing, the timestamp is completed using date information embedded in the filename. The resulting text sequence not only preserves the original semantic information but also establishes cross-document temporal correlations, providing a data foundation for subsequent time-series analysis.

[0037] In some of the solutions described above in this application, directly performing multi-task training or bridging graph construction on the original text sequence will lead to problems such as insufficient semantic understanding and loss of temporal sequence and context.

[0038] This application further proposes to perform word segmentation on the text sequence to generate a sub-word sequence, input the sub-word sequence into a pre-trained language model to generate sub-word vectors, perform sentence segmentation on the text sequence based on punctuation marks to obtain several candidate sentences, perform semantic enhancement processing on the candidate sentences to generate enhanced semantic vectors, input the enhanced semantic vectors into a pre-trained language model to generate sentence-level semantic vectors, and merge or split the candidate sentences according to the semantic similarity threshold and length rules of the sentence-level semantic vectors to obtain a set of text blocks.

[0039] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0040] The text sequence is segmented to generate a sub-word sequence. Specifically, the text sequence is input into the DistilBERT tokenizer, and the tokenizer's features are set to segment the text.

[0041] The sub-word sequence is input into a pre-trained language model to generate sub-word vectors. The pre-trained language model is the DistilBERT model.

[0042] The text sequence is segmented into sentences based on punctuation marks, yielding several candidate sentences. Furthermore, regular expressions are used to match punctuation marks such as periods, question marks, and exclamation marks as the segmentation criteria.

[0043] The candidate sentences undergo semantic enhancement to generate enhanced semantic vectors. Specifically, Stanza is first used for dependency parsing to obtain dependency type vectors; then, the dependency type vectors are concatenated with the sub-word vectors and input into a two-layer fully connected neural network to obtain syntactic enhancement vectors; next, a lightweight BiLSTM model is used to perform sentiment analysis on the candidate sentences to obtain sentiment distribution vectors; finally, the syntactic enhancement vectors and sentiment distribution vectors are concatenated to obtain enhanced semantic vectors.

[0044] The enhanced semantic vectors are input into a pre-trained language model to generate sentence-level semantic vectors. Then, the DistilBERT model is used to encode the enhanced semantic vectors to obtain sentence-level semantic vectors.

[0045] Candidate sentences are merged or split into text blocks based on semantic similarity thresholds and length rules derived from sentence-level semantic vectors. For example, a semantic similarity threshold of 0.8, a maximum length threshold of 512 characters, and a minimum length threshold of 50 characters are set. Cosine similarity is calculated for adjacent candidate sentences; if the cosine similarity exceeds 0.8, they are merged; if the combined length exceeds 512 characters, they are split; if the split length is less than 50 characters, they are merged with adjacent blocks.

[0046] Through the above technical solution, this application achieves automated semantic analysis and structured processing of complex correspondence text. By employing multi-level semantic representation and flexible text block segmentation rules, it effectively captures the semantic relationships between sentences while maintaining appropriate granularity. This method can better preserve contextual information, providing structured input for subsequent multi-task learning and semantic retrieval, thereby improving the accuracy and efficiency of correspondence text analysis.

[0047] In some of the solutions mentioned above in this application, the candidate sentences are obtained by segmenting the text sequence into sentences and then directly merged or split. However, the candidate sentences may be too long or semantically incoherent, resulting in low processing efficiency. At the same time, insufficient grammatical independence leads to poor sentence segmentation accuracy, affecting the quality of the text block set.

[0048] This application further proposes to perform character-level checks on candidate sentences before merging or splitting them. If the number of characters exceeds a preset character threshold or the semantics are not coherent, the sentences are further segmented into clauses. The grammatical independence of the clauses is verified by dependency parsing. If the preset independence conditions are met, the sentence segmentation position is adjusted.

[0049] Among them, character-level testing counts the number of characters in candidate sentences and compares them with a preset threshold. If the count exceeds the threshold, a segmentation operation is triggered. Semantic coherence is judged by detecting whether there are missing logical connectives or topical abrupt changes in the sentence. Dependency parsing extracts the subject-verb-object structure and modification relationship of clauses. If a clause contains complete grammatical components, it is determined to be independent. Adjusting the segmentation position redefines the start and end points of the sentence based on the boundaries of grammatical components.

[0050] Specifically, character-level verification uses a sliding window algorithm to scan candidate sentences character by character. When the cumulative number of characters exceeds a threshold, a temporary segmentation point is generated, and the algorithm checks for semantic breaks before and after the segmentation point. If a break exists, the segmentation point is moved to the nearest punctuation mark or conjunction. Dependency parsing analyzes the dependency tree structure of clauses. If a clause contains a core predicate and the subject-predicate relationship is complete, it is determined to have grammatical independence. When adjusting the segmentation position, if the clause meets the independence condition, the original candidate sentence's segmentation point is corrected to the beginning or end position of the clause. For example, if a candidate sentence has 250 characters, exceeding the preset 200-character threshold, the sliding window detects a comma at character 180 with semantic coherence before and after it, and segments it into two clauses. Further dependency analysis reveals that the second clause lacks a subject, so the segmentation point is adjusted to the period at character 150 to ensure the clause's grammatical integrity. The resulting text block set has higher semantic coherence and structural regularity, providing reliable input for subsequent multi-task training.

[0051] In some of the solutions mentioned above in this application, a scheme is proposed to merge or split candidate sentences to obtain a set of text blocks. However, before merging or splitting, the candidate sentences may have too many characters or be semantically incoherent, which may cause the length of the merged text block to exceed the processing capacity or the sub-clauses after splitting to be grammatically independent, affecting the accuracy of subsequent semantic vector generation and associated text blocks.

[0052] This application further proposes to perform character-level checks on each candidate sentence before merging or splitting. If the number of characters exceeds a preset character threshold or the semantics are not coherent, the sentence is further segmented into clauses. The grammatical independence of the clauses is verified by dependency parsing. If the preset independence conditions are met, the sentence segmentation position is adjusted.

[0053] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0054] Each candidate sentence undergoes a character-level check. If the number of characters exceeds a preset character threshold or the semantics are incoherent, it is further segmented into clauses. Specifically, the preset character threshold is set to 100 characters. For candidate sentences exceeding 100 characters, a sliding window method is used for segmentation. The sliding window size is set to 50 characters, with a step size of 25 characters. A semantic coherence score is calculated for the text segment within each window, and the positions with scores below the threshold are used as segmentation points.

[0055] Furthermore, dependency parsing is used to verify the grammatical independence of clauses. If a preset independence condition is met, the sentence segmentation position is adjusted. Stanza is used for dependency parsing. A dependency tree is generated for each clause, and the integrity score of the subtree is calculated. If the score is higher than a preset threshold, the clause is considered grammatically independent. For clauses that do not meet the independence condition, they are merged with adjacent clauses or further segmented until all clauses meet the grammatical independence requirement.

[0056] Through the aforementioned technical solutions, this application can effectively handle long and complex sentences, improving the semantic integrity and independence of text blocks. This improves the accuracy of subsequent semantic analysis and information extraction, laying the foundation for constructing high-quality text block sets. Furthermore, by verifying grammatical independence, it avoids semantic fragmentation that might result from mechanical segmentation, ensuring the grammatical structural integrity of the text blocks. This refined processing based on linguistic knowledge significantly enhances the robustness and interpretability of text analysis.

[0057] In some of the schemes described above in this application, when merging or splitting candidate sentences, the merging text blocks may have problems such as grammatical logic confusion or inconsistent emotional expression because they rely only on the similarity of sentence-level semantic vectors and character length rules, without fully combining the grammatical structure and sentiment information of the sentences. This may affect the accuracy of subsequent semantic clustering.

[0058] This application further proposes specific methods for semantic enhancement of candidate sentences.

[0059] The semantic enhancement process involves generating dependency type vectors through dependency parsing, concatenating them with sub-word vectors, and then inputting the concatenated vectors into a pre-defined neural network to generate syntactic enhancement vectors. Simultaneously, a pre-defined sentiment classification model generates sentiment distribution vectors, which are then concatenated with the syntactic enhancement vectors to obtain the enhanced semantic vector. Specifically, dependency parsing identifies the grammatical relationships between sentence components and outputs dependency type vectors to represent the grammatical structure; sub-word vectors are derived from the word segmentation results of a pre-trained language model; and the concatenation of these two vectors is further processed by a neural network to extract syntactic and semantic features. The sentiment classification model predicts the probability distribution of candidate sentences based on a pre-defined sentiment labeling system, generating a sentiment distribution vector. These two vectors reflect grammatical and sentiment information respectively, and the concatenation operation integrates multi-dimensional semantic features.

[0060] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0061] Semantic enhancement processing includes the following steps:

[0062] First, dependency parsing is performed on the candidate sentences to generate dependency type vectors. Specifically, Stanza is used to perform dependency parsing on the candidate sentences to obtain the dependency relations between words, and these dependency relations are encoded into vector representations.

[0063] Secondly, the dependency type vector and the sub-word vector are concatenated and input into a pre-defined neural network to generate syntactic enhancement vectors. Further, a multilayer perceptron is used as the pre-defined neural network, with the input layer dimension being the sum of the dependency type vector dimension and the sub-word vector dimension, the hidden layer using the ReLU activation function, and the output layer dimension being the same as the input layer dimension.

[0064] Next, a pre-defined sentiment classification model is used to perform sentiment analysis on the candidate sentences, generating sentiment distribution vectors. Then, a lightweight BiLSTM model is used to input the candidate sentences into the model to obtain the probability distributions of positive, neutral, and negative sentiments.

[0065] Finally, the syntactic enhancement vector and the sentiment distribution vector are concatenated to obtain the enhanced semantic vector. Specifically, the two vectors are concatenated in the last dimension to form the final enhanced semantic vector.

[0066] Through the above technical solution, this application achieves multi-dimensional semantic enhancement of candidate sentences. By capturing the structural information of sentences through dependency parsing and combining it with the semantic information provided by sub-word vectors, a richer syntactic representation is generated. Simultaneously, sentiment analysis results are incorporated to further supplement the emotional semantic information of the sentences. This multi-faceted semantic enhancement method can more comprehensively characterize the semantic features of candidate sentences, providing richer and more accurate input for subsequent text block generation and semantic vector calculation, thereby improving the performance and accuracy of the entire email text analysis system.

[0067] In some of the solutions mentioned above in this application, when merging candidate sentences based on semantic similarity, only semantic relevance is considered, and the length of the merged text block is not constrained. This may result in the generated text block being too long or too short, thereby affecting the efficiency and accuracy of subsequent semantic vector aggregation.

[0068] This application further proposes merging or splitting including: calculating the cosine similarity of the sentence-level semantic vectors of candidate sentences; if the similarity exceeds a preset first similarity threshold, merging them into the same text block; if the number of characters in the merged text block exceeds a preset maximum length threshold, splitting it into sub-blocks; if the number of characters in the split sub-blocks is less than a preset minimum length threshold, merging it with adjacent blocks.

[0069] The merging operation is triggered based on a semantic similarity threshold to ensure semantic consistency of related sentences. When splitting sub-blocks, a maximum length threshold is used to avoid information redundancy, and a minimum length threshold is used to prevent information fragmentation. Merging adjacent blocks further optimizes the sub-block length and maintains contextual coherence. The preset maximum length threshold can be set to 500 characters, the minimum length threshold can be set to 100 characters, and the first similarity threshold can be set to 0.85.

[0070] Specifically, after calculating the cosine similarity of candidate sentences, sentences exceeding a threshold are merged to form an initial text block. When the number of characters in the merged block exceeds 500, it is split into multiple sub-blocks based on sentence boundaries. If the number of characters in a sub-block is less than 100, it is merged again with adjacent sub-blocks to ensure that the sub-block length is within a reasonable range. For example, if the merged text block has 600 characters, it needs to be split into two sub-blocks. If one of the sub-blocks has 80 characters, it is merged with an adjacent sub-block to form a final sub-block with 180 characters. This process dynamically adjusts the text block length, preserving semantic relevance while avoiding excessively long or short text blocks that could reduce subsequent processing efficiency or cause semantic information loss.

[0071] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0072] The cosine similarity of the sentence-level semantic vectors of candidate sentences is calculated. If the similarity exceeds a preset first similarity threshold, they are merged into the same text block. Specifically, the first similarity threshold can be set to 0.85. For any two candidate sentences A and B, the cosine similarity of their sentence-level semantic vectors is calculated. If the similarity is greater than 0.85, A and B are merged into the same text block.

[0073] If the number of characters in the merged text block exceeds a preset maximum length threshold, it is split into sub-blocks. Furthermore, the maximum length threshold can be set to 500 characters. When the merged text block exceeds 500 characters, it is truncated from the 500th character, and the excess portion is treated as a new sub-block.

[0074] If the number of characters in a split sub-block is less than a preset minimum length threshold, it is merged with an adjacent block. The minimum length threshold can be set to 100 characters. When the number of characters in a split sub-block is less than 100, the sub-block is merged with an adjacent text block. This avoids creating excessively short text blocks.

[0075] Through the above technical solutions, this application achieves intelligent merging and splitting of candidate sentences. By controlling the merging granularity through a cosine similarity threshold, semantically irrelevant sentences are prevented from being incorrectly merged. Simultaneously, by controlling the text block size through maximum and minimum length thresholds, both processing efficiency degradation caused by excessively long text blocks and semantic fragmentation caused by excessively short text blocks are avoided. This flexible text block segmentation method improves the accuracy and efficiency of subsequent semantic analysis and information retrieval.

[0076] In some of the solutions mentioned above in this application, although the enhanced semantic vectors of words within a text block can reflect local semantic information, they lack effective aggregation of global relationships between multiple words within the block, resulting in fragmented block-level semantic expression. At the same time, existing methods do not integrate topic distribution with block-level semantics, making it difficult to accurately capture the topic relevance across text blocks during the subsequent construction of dynamic bridging graphs, affecting the integrity and relevance expression of the comprehensive semantic vector.

[0077] This application further proposes to aggregate the enhanced semantic vectors of all words within a text block to generate a block-level input vector; input the block-level input vector into a preset block encoder, and generate a block-level semantic vector through joint optimization of masked language model loss, neighbor block discrimination loss, and cross-block comparison loss; perform topic clustering on the block-level semantic vector based on a clustering algorithm, and generate a topic distribution vector through a keyword extraction method; and concatenate the topic distribution vector with the block-level semantic vector to generate a comprehensive semantic vector.

[0078] In this process, the block-level input vector is aggregated using average pooling or max pooling to form an initial representation reflecting the global semantics within the block. The block encoder employs a multi-layer Transformer structure. The masking language model loss optimizes the model's robustness to local semantic loss by randomly masking part of the input vector and predicting the original value. The neighboring block discrimination loss distinguishes between adjacent and non-adjacent blocks through contrastive learning, strengthening the temporal contextual relevance. The cross-block contrastive loss improves semantic discriminativeness by bringing semantically similar blocks closer and pushing away irrelevant blocks. Topic clustering uses K-means or hierarchical clustering algorithms to perform unsupervised grouping of the block-level semantic vectors. High-frequency words or key phrases from each cluster are extracted to generate topic distribution vectors. The topic distribution vectors are weighted by word frequency-inverse document frequency to calculate the weight of each topic, forming a multi-dimensional probability distribution. The concatenation operation fuses the topic distribution vectors and block-level semantic vectors using vector concatenation or weighted summation to generate a comprehensive semantic vector containing both topic and semantic information.

[0079] Specifically, the enhanced semantic vectors of each word within a text block are aggregated into a block-level input vector through pooling operations, and then fed into the block encoder for multi-task joint training. The masking language model loss randomly masks some dimensions of the input vector, forcing the model to recover the masked content based on contextual information, thereby enhancing the model's adaptability to local semantic gaps. The neighboring block discrimination loss labels text blocks with adjacent timestamps as positive samples and non-adjacent blocks as negative samples, optimizing the block encoder's ability to capture temporal proximity relationships through contrastive learning. The cross-block contrastive loss calculates the cosine similarity between block-level semantic vectors, treating block pairs with similarity above a preset threshold as positive samples and those below as negative samples. This contrastive loss function adjusts the vector space distribution, improving the clustering of semantically similar blocks. After multi-task joint optimization, the block-level semantic vector simultaneously incorporates local semantic robustness, temporal correlation, and semantic discriminative features. Building upon this foundation, clustering algorithms are employed to group block-level semantic vectors into thematic groups. For instance, K-means is used to divide the semantic vectors into a predetermined number of clusters, and the top N words with the highest TF-IDF values ​​in each cluster are extracted as thematic keywords, forming a distribution vector reflecting the core themes of each cluster. Finally, the thematic distribution vector and the block-level semantic vector are concatenated dimensionally, allowing the comprehensive semantic vector to simultaneously encode the semantic features of the text block and the probability distribution of its associated themes. For example, when the block-level semantic vector has a dimension of 768 and the thematic distribution vector has a dimension of 50, the comprehensive semantic vector expands to 818 dimensions, with the first 768 dimensions carrying semantic information and the last 50 dimensions recording thematic weights. This approach enables the dynamic bridging graph construction process to match associated blocks through semantic similarity and strengthen cross-block connections based on thematic consistency, improving the accuracy and completeness of contextual concatenation during subsequent query responses.

[0080] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0081] The enhanced semantic vectors of words within a text block are aggregated to generate a block-level input vector. Specifically, an attention mechanism is used to weight and sum the word vectors to obtain a fixed-dimensional block-level representation.

[0082] The block-level input vector is fed into a predefined block encoder, and block-level semantic vectors are generated by jointly optimizing the masked language model loss, the neighbor block discrimination loss, and the cross-block comparison loss. The block encoder adopts a Transformer structure. The masked language model task predicts randomly masked words, the neighbor block discrimination task distinguishes between adjacent and non-adjacent text blocks, and the cross-block comparison task maximizes the mutual information of different block representations within the same document.

[0083] The block-level semantic vectors are clustered using a clustering algorithm, and topic distribution vectors are generated through keyword extraction. Furthermore, the K-means algorithm is used for clustering, and the TF-IDF method is used to extract keywords for each category, generating a topic-word distribution matrix.

[0084] The topic distribution vector is concatenated with the block-level semantic vector to generate a comprehensive semantic vector. Thus, the comprehensive semantic vector contains both block-level semantic information and topic distribution information.

[0085] Through the above technical solutions, this application achieves multi-task joint training and topic fusion of text blocks, effectively capturing intra-block word relationships, inter-block contextual associations, and global topic distribution. The integrated semantic vectors fuse multi-level semantic information, improving the accuracy of subsequent text analysis and retrieval. Simultaneously, multi-task training enhances the model's generalization ability, enabling it to adapt to different types and styles of email text. Topic fusion helps to understand text content from a macro perspective, providing support for subsequent text organization and summary generation.

[0086] In some of the solutions mentioned above in this application, a dynamic bridging graph needs to be constructed after generating the comprehensive semantic vector. However, the existing methods do not specify the activation rules for semantic association and temporal association, which makes the bridging graph unable to effectively reflect the multi-dimensional dynamic relationship between text blocks, affecting the accuracy of subsequent query responses.

[0087] This application further proposes to construct a dynamic bridging graph, including: if the cosine similarity of the combined semantic vectors of two text blocks exceeds a preset second similarity threshold, then add a semantic association edge; if the timestamp difference between two text blocks is less than a preset number of days or they share a session identifier, then add a temporal association edge.

[0088] The generation of semantic association edges is based on the quantitative calculation of semantic similarity between text blocks. The second similarity threshold is set in the range of 0.85-0.95. The quantitative determination of semantic association across text blocks is achieved by calculating the cosine value of the comprehensive semantic vector. The generation of temporal association edges adopts a dual-path condition triggering mechanism. Path one sets the timestamp difference threshold to 3-7 calendar days, and path two checks the consistency of session identifiers. The two conditions have a logical OR relationship. Semantic association edges and temporal association edges are stored using different weight coefficients, where the weight of semantic edges is the cosine similarity value, and the weight of temporal edges is the reciprocal of the time difference.

[0089] Specifically, during the bridging graph construction process, the system traverses the comprehensive semantic vectors of all text blocks and calculates the cosine similarity of each pair of vectors using a sliding window approach. When a pair of similarities is detected to exceed a second preset threshold, a weighted semantic edge is automatically generated and stored in the graph database. Simultaneously, the system reads the timestamp metadata of each text block. If the time difference between two blocks is less than a preset number of days or they belong to the same session chain, a temporal edge is established in the graph structure. The superposition of semantic and temporal edges forms a hybrid association network, where semantic edges support cross-time period topic association mining, and temporal edges ensure session continuity. When new email data is input, the system only needs to incrementally calculate the association conditions between the new block and existing blocks, dynamically expanding the bridging graph structure. Through this dual association mechanism, the bridging graph can simultaneously capture the semantic relevance and temporal continuity of text blocks, providing multi-dimensional association path support for subsequent query activation.

[0090] As a preferred embodiment, the solution of this application is implemented as follows: When constructing the dynamic bridging graph, the cosine similarity of the comprehensive semantic vectors of the text blocks is first calculated. When the cosine similarity of the vectors of two text blocks reaches 0.85, semantic association edges are established in the bridging graph. For text blocks with a timestamp difference of less than 5 days, or text blocks with the same session identifier, temporal association edges are further established. Semantic association edges are stored in a bidirectional graph structure, and temporal association edges are recorded in a singly linked list to record the time sequence. Specifically, in the process of establishing association edges, semantic similarity calculation uses the dot product operation after vector normalization, timestamp difference calculation takes the absolute value in days, and session identifier matching uses a string exact comparison method.

[0091] Through the above technical solution, this application effectively solves the problem of simultaneously modeling semantic associations and temporal continuity in letter text analysis. By employing a dual association mechanism of a dynamic bridging graph, it preserves the semantic similarity features between text blocks while capturing the temporal logical relationships within the business scenario. This structure enables the large language model to accurately activate contextual fragments with close semantic and temporal relevance when processing user queries, avoiding the fragmented response problem caused by isolated text block analysis in traditional methods. Simultaneously, the dynamic update mechanism ensures the bridging graph's adaptability to newly added letter data, guaranteeing the real-time nature of the association relationships.

[0092] In some of the solutions mentioned above in this application, the construction of dynamic bridging graphs relies solely on the similarity of comprehensive semantic vectors to establish associated edges, failing to effectively identify continuous relationships in the time dimension or session context. This results in the inability to form effective associations between text blocks on the same topic across time periods or sessions, affecting the integrity of context splicing in subsequent query responses.

[0093] This application further proposes adding a semantic association edge when the cosine similarity of the combined semantic vectors of two text blocks exceeds a preset second similarity threshold, and adding a temporal association edge when the timestamp difference between the two text blocks is less than a preset number of days or when they share a session identifier.

[0094] The semantic association edges are established based on a cosine similarity threshold, which was determined through multiple experiments, for example, set to 0.85. The temporal association edges are determined based on two independent logical branches, with the preset number of days dynamically adjusted according to the business scenario, typically 7 days. The shared session identifier is obtained by analyzing email header information or document metadata. Semantic and temporal association edges use different weight coefficients and are visually distinguished by solid and dashed lines respectively in the bridging graph. Together, they form a composite association network.

[0095] Specifically, in generating the dynamic bridging graph, the comprehensive semantic vector matrix of all text blocks is first traversed, and cosine similarity is calculated in batches through matrix operations. When a text block pair with a similarity exceeding a threshold is detected, a weighted semantic edge is automatically created in the graph, with the weight value linearly mapping the similarity value range. Simultaneously, the timestamp information of all text blocks is traversed, and temporal edges are created for adjacent text blocks with a time interval within a set number of days. Temporal edges are also forcibly created for text blocks with the same session identifier. For example, when processing project progress emails from three consecutive days, if two emails have a subject similarity of 0.87 and a time interval of two days, both semantic and temporal edges are created simultaneously. When a user queries a timeline related to project progress, the bridging graph activates historical and latest emails through a dual association mechanism, ensuring that the context received by the large language model contains complete time-series information and semantic association content.

[0096] Through the above technical solution, this application effectively solves the problem of contextual association breakage in long text processing using large language models. By employing a dynamic activation mechanism, it accurately captures semantic relationships across documents, ensuring that the generated answer covers both the core query content and maintains the complete logical chain of business events. This solution, while maintaining the generation capability of large models, avoids the risk of key information being overwhelmed by directly inputting massive amounts of text, thus improving the system's response accuracy to a level that can support actual business decision-making.

[0097] In some of the solutions described above in this application, the existing bridging graph cannot be dynamically updated according to the new letter data, which means that the entire bridging graph structure needs to be repeatedly generated when processing new letters, resulting in wasted computing resources and response delays.

[0098] This application further proposes to generate new text blocks and new comprehensive semantic vectors by repeating the aforementioned process based on new letter data; calculate the similarity between the new comprehensive semantic vector and the existing comprehensive semantic vector; add semantic association edges if the similarity exceeds a preset second similarity threshold; and add time-series edges if the timestamps of the new text block and the existing text block are less than a preset number of days or share a session identifier.

[0099] Specifically, the same encoder and clustering algorithm as the original data processing are used when generating new text blocks to ensure vector space consistency. Cosine similarity algorithm is used to calculate similarity, with a preset second similarity threshold set between 0.85 and 0.95. The preset number of days for adding time-series edges is 3 or 7 days, with the specific value adjusted according to the business scenario. Shared session identifiers are implemented by parsing the session ID field in the email header metadata.

[0100] Specifically, upon receiving a new message, the system generates a new comprehensive semantic vector through OCR conversion, semantic analysis, and topic fusion. This vector is then compared with existing vectors in the bridging graph for batch similarity. If the similarity exceeds a threshold, semantic association edges are automatically established. Simultaneously, the system extracts the timestamp and session identifier of the new text block and performs temporal matching with nodes in the bridging graph: text blocks with a time difference within a preset number of days establish temporal association edges, and text blocks sharing the same session ID establish strong temporal associations. This process employs an incremental update mechanism, performing only local calculations on new data and associated nodes to avoid full graph reconstruction. Through the dual association of semantic and temporal edges, the newly added text block is accurately embedded into the original graph structure, ensuring that subsequent query responses can utilize the contextual relationships formed by the latest data.

[0101] As a preferred embodiment, the solution of this application is implemented as follows: When the system receives a new PDF email containing a customer complaint, it converts it into a timestamped text sequence using OCR recognition technology. After semantic analysis and structural processing, this text sequence generates a text block containing the complaint date and product model description, and a comprehensive semantic vector containing product fault characteristics. The similarity between this comprehensive semantic vector and the text block vectors in historical work order records is calculated using a cosine similarity algorithm. The text block with a repair record of the same product from three months prior reaches a similarity value of 0.87. A semantic association edge is established based on a preset similarity threshold of 0.85. Simultaneously, it is detected that the time difference between the new text block and the historical text block is 92 days, exceeding the preset 90-day association period; therefore, no temporal association edge is established. However, a session identifier sharing the same customer number is detected, and a temporal association edge is established accordingly to complete the bridging graph update.

[0102] Through the above technical solutions, this application realizes the efficient integration capability of the letter analysis system for new data. By automatically identifying and establishing semantic relationships and business session continuity between text blocks through a dynamic update mechanism, it solves the problem of relationship breakage in traditional methods when processing incremental data, ensures the collaborative analysis capability of historical data and new data, and thus improves the contextual integrity of user query response and the effectiveness of business decision support.

[0103] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for analyzing letter text based on a large language model driven by multi-layer dynamic aggregation, characterized in that, The method comprises the following steps: acquiring multi-format document data and converting it into a timestamped text sequence; performing semantic analysis and structural processing on the text sequence to generate a set of text blocks containing timing and context; performing multi-task training and theme fusion on the set of text blocks to generate a comprehensive semantic vector; constructing a dynamic bridging graph based on the comprehensive semantic vector, which activates associated text blocks through semantic and timing correlation; in response to a user query, concatenating the contents of the activated text blocks into a contextual Prompt and inputting it into a large language model to obtain a one-time generated answer; receiving new document data and updating the bridging graph based on the new document data; wherein the generation of the set of text blocks containing timing and context comprises: performing word segmentation processing on the text sequence to generate a subword sequence; inputting the subword sequence into a pre-trained language model to generate a subword vector; performing sentence segmentation on the text sequence based on punctuation to obtain a number of candidate sentences; performing semantic enhancement processing on the candidate sentences to generate enhanced semantic vectors; inputting the enhanced semantic vectors into a pre-trained language model to generate sentence-level semantic vectors; merging or splitting the candidate sentences according to the semantic similarity threshold and length rules of the sentence-level semantic vectors to obtain a set of text blocks.

2. The method of claim 1, wherein the method is based on a multi-layer dynamic aggregation of large language model driven document text analysis. The acquisition of multi-format document data and the conversion into a timestamped text sequence comprises: acquiring multi-format document data, including plain text files, document files, image files, and scanned documents; performing grayscale, noise reduction, and text region detection on the image files and scanned documents, and generating editable text through OCR technology; performing encoding unification and cleaning on the plain text files, document files, and editable text, including correcting OCR misrecognized characters, unifying date formats, and adding timestamp information, to generate a timestamped unified text sequence.

3. The method of claim 1, wherein the method is based on a multi-layer dynamic aggregation of large language model driven document text analysis. Before merging or splitting the candidate sentences, the method further comprises: performing character-level inspection on each candidate sentence, and if the number of characters exceeds a predetermined character threshold or the semantics are not coherent, further splitting the sentence into sub-sentences; verifying the grammatical independence of the sub-sentences through dependency syntax analysis, and adjusting the sentence segmentation position if the predetermined independence condition is met.

4. The method of claim 3, wherein the method is based on a multi-layer dynamic aggregation of large language model driven document text analysis. The semantic enhancement processing comprises: performing dependency syntax analysis on the candidate sentences to generate dependency type vectors; concatenating the dependency type vectors and subword vectors and inputting them into a pre-set neural network to generate syntax-enhanced vectors; performing sentiment analysis on the candidate sentences through a pre-set sentiment classification model to generate sentiment distribution vectors; concatenating the syntax-enhanced vectors and sentiment distribution vectors to obtain enhanced semantic vectors.

5. The method of claim 4, wherein the method is based on a multi-layer dynamic aggregation of large language model driven document text analysis. The merging or splitting comprises: performing cosine similarity calculation on the sentence-level semantic vectors of the candidate sentences, and if the similarity exceeds a predetermined first similarity threshold, merging them into the same text block; if the number of characters in the merged text block exceeds a predetermined maximum length threshold, splitting it into sub-blocks; if the number of characters in the split sub-blocks is less than a predetermined minimum length threshold, merging them with adjacent blocks.

6. The method of claim 5, wherein the method is based on a multi-layer dynamic aggregation of large language model driven document text analysis. The multi-task training and theme fusion comprises: aggregating the enhanced semantic vectors of all words in the text block to generate a block-level input vector; The block-level input vector is input into a preset block encoder, and a block-level semantic vector is generated through joint optimization of mask language model loss, neighbor block discrimination loss and cross-block comparison loss. The block-level semantic vectors are clustered based on a clustering algorithm, and a topic distribution vector is generated by a keyword extraction method. The topic distribution vector is concatenated with the block-level semantic vector to generate a comprehensive semantic vector.

7. The method of claim 1, wherein the method is a multi-layer dynamic aggregation based large language model driven document text analysis method. The construction of the dynamic bridging graph includes: If the cosine similarity of the combined semantic vectors of two text blocks exceeds the preset second similarity threshold, then a semantic association edge is added. If the difference between the timestamps of two text blocks is less than a preset number of days or a shared session identifier, then add a time-series association edge.

8. The method of claim 1, wherein the method is a multi-layer dynamic aggregation based large language model driven document text analysis method. The response to the user query concatenates the activated text block content into a contextual Prompt input to the large language model, obtaining a one-time generated answer including: Encode user queries into query vectors; The query vector and the comprehensive semantic vector are matched by cosine similarity, and the top few text blocks with the highest cosine similarity are selected as core associated blocks; The core associated block is marked as activated, and its neighboring node text blocks are activated through the semantic and temporal associated edges in the bridging graph. Concatenate all activated text blocks in chronological order to generate a context prompt; Input the context prompt into the large language model to obtain a one-time generated answer.

9. The method of claim 1, wherein the method is a multi-layer dynamic aggregation based large language model driven document text analysis method, characterized in that, The updating of the bridging graph includes: Based on the new letter data, the conversion to a timestamped text sequence, the semantic analysis and structural processing, and the multi-task training and topic fusion are re-executed to generate new text blocks and new comprehensive semantic vectors; Calculate the similarity between the new integrated semantic vector and the existing integrated semantic vector; If the similarity exceeds a preset second similarity threshold, a semantic association edge is added; If the timestamp of the new text block is less than the preset number of days or shares a session identifier with the existing text block, then add a time-series edge.

Images