Intelligent semantic comprehension and generation method and system based on ensemble learning

Abstract

The invention discloses an intelligent semantic comprehension and generation method and system based on ensemble learning, and relates to the technical field of natural language processing. An intelligent semantic comprehension and generation system based on ensemble learning comprises an ensemble semantic comprehension module, an ensemble semantic generation module and a historical corpus construction module. According to the method, through a weighted fusion method of an integrated semantic understanding model, the advantages of various base learners are combined, and semantic features in a text can be accurately extracted; in combination with historical corpus information of the user, specific language styles and dynamic semantic preferences of the user are effectively captured, and personalization and context consistency of semantic analysis are enhanced; the integrated semantic generation model is combined with historical corpus information of a user in the generation process to dynamically adjust local and global semantic representations, so that high matching of generated contents and personalized requirements of the user is ensured, and semantic accuracy and logic consistency of a generated text are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to an intelligent semantic understanding and generation method and system based on ensemble learning. Background Art

[0002] Current semantic understanding and generation methods mainly rely on deep learning technology. Typical methods include pre-trained language models based on the Transformer architecture, which learn contextual semantic features through large-scale corpora and are widely used in natural language understanding, text generation and machine translation. However, these single models may show limitations when processing complex semantics, such as insufficient generalization capabilities for diverse tasks or high sensitivity to noisy data. Intelligent semantic understanding and generation methods based on ensemble learning combine the advantages of multiple base learners, and organically combine the prediction results of different models through ensemble technology, significantly improving the understanding and generation capabilities of diverse semantics. At the same time, it introduces the learning of semantic expression habits of different users, extracts personalized semantic features and language expression styles, thereby achieving accurate understanding of different user intentions and personalized semantic generation. Summary of the invention

[0003] The present invention aims to provide an intelligent semantic understanding and generation method and system based on ensemble learning for accurate semantic understanding and generation.

[0004] The invention discloses an intelligent semantic understanding and generation method based on integrated learning, comprising the following steps: obtaining an initial text and a corresponding text user; obtaining user historical corpus information based on the text user; constructing the user historical corpus information based on user historical data; the user historical corpus information includes a user semantic style factor, a user semantic hot spot factor and a user semantic preference factor; constructing the user historical corpus information can accurately capture the user's unique language style, dynamic hot spot interest and long-term semantic preference, and realize dynamic quantification and continuous optimization of user semantic features; performing semantic understanding according to the initial text, user historical corpus information and an integrated semantic understanding model to obtain text semantic extraction information; the integrated semantic understanding model accurately extracts text semantic features through integrated learning and weighted fusion, and effectively captures user semantic style, hot spot preference and context association in combination with user historical corpus information, significantly improving the accuracy, personalization and context consistency of semantic understanding; performing semantic generation according to text semantic extraction information, user historical corpus information and an integrated semantic generation model to obtain text generation text; the integrated semantic generation model dynamically adjusts the generated content in combination with local and global semantic representations through multi-level generation, logic verification and user corpus optimization to ensure the semantic accuracy, logical consistency and personalization of text generation.

[0005] As a preferred technical solution of the present invention, the integrated semantic understanding model includes a sentence segmentation layer, a text initial analysis layer, a user-specific corpus analysis layer and a semantic understanding output layer; the sentence segmentation layer is used to segment the initial text into sentences to obtain short sentences of the text to be analyzed. n , n=1, 2, ..., N; N is the number of text sentences contained in the initial text; the initial text analysis layer is used to analyze the text sentences J n Extract feature information and obtain the short sentence feature J of the text to be analyzed n '; The user-specific corpus analysis layer is used to perform semantic understanding based on the user's historical corpus information and the features of the short sentences in the text to be analyzed Jn', and obtain the semantic extraction features of the short sentences in the text T n ; Among them, the user-specific prediction analysis layer is constructed by an integrated learning method; the semantic understanding output layer is used to extract semantic features of all text short sentences T n Combined to obtain text semantic extraction information.

[0006] As a preferred technical solution of the present invention, the specific steps of constructing a user-specific corpus analysis layer include: the user-specific corpus analysis layer includes M semantic analysis layers, a hierarchical user analysis layer and a comprehensive semantic analysis layer; wherein the M semantic analysis layers are trained by different base learners; in the M semantic analysis layers C m In the analysis, we treat the short sentence features J n 'Extract semantic features to obtain the semantic features of text short sentences J nm '; Where m = 1, 2, ..., M; At the semantic analysis layer C M In the example, for all the text short sentence semantic features J nm 'Perform feature fusion to obtain the fused sentence semantic feature G n ; In the hierarchical user analysis layer, there are style feature analysis sublayer, semantic preference analysis sublayer and context association analysis sublayer; the style feature analysis sublayer is used to fusion short sentence semantic feature G n Combine the user's historical corpus information to weight the user's semantic style factor and obtain the semantic feature F of the style text short sentence n ; The semantic preference analysis sublayer is used to integrate the semantic features of short sentences G n Combine the user's historical corpus information to weight the user's semantic hotspot factor and obtain the semantic feature R of the hotspot text short sentence n ; The contextual association analysis sublayer is used to integrate the semantic features of short sentences G n-1 , integrating the semantic features of short sentences G n and the fusion sentence semantic features G n+1 Combined with the user's historical corpus information, the context semantic analysis is weighted to obtain the context text sentence semantic feature H n ; In the comprehensive semantic analysis layer, the semantic features of the style text short sentences Fn , semantic features of hot text short sentences R n and the context text sentence semantic features H n Perform multi-level weighted fusion to obtain the text sentence semantic extraction feature T n .

[0007] As a preferred technical solution of the present invention, the specific steps of constructing user history corpus information include: collecting user history data sets; the user history data sets contain user history data D i ; User history data includes but is not limited to conversation records, user published content, search records, click behavior data and user feedback information; Set the time node i of user history data, i=1, 2, ...; i represents user history data D i The time node corresponding to the collection; the user's historical data D i Analyze and extract the unique language style features of the text user to obtain the text user language style features W i ; For all text user language style features W i Perform feature quantification to obtain the user semantic style factor; i Perform topic modeling to obtain user hot topic features E i ; Using the formula Generate user hot topic features E i The time decay weight of , we get the weighted user hot topic feature E i '; For all weighted user hot topic features E i 'Perform feature quantification to obtain the user semantic hotspot factor; where A i Represents the user's hot topic feature E i The time decay weight of user historical data D is: I represents the current time node of constructing user historical corpus information; α represents the time decay coefficient; i Generate a long-term preference distribution vector and obtain the user preference distribution vector B i ; Use the sliding window method to capture all user preference distribution vectors B i The user's semantic preference factor is obtained by dynamically changing the user's historical corpus information at preset fixed time periods.

[0008] As a preferred technical solution of the present invention, the integrated semantic generation model includes a semantic generation layer, a sampling verification layer and a result output layer; the semantic generation layer is used to generate text based on text semantic extraction information and a large language model to obtain a text to be verified; the sampling verification layer is used to perform logical verification on the text to be verified to obtain a semantic anomaly assessment result; if the semantic anomaly assessment result reaches a preset generation standard, the text to be verified is used as the text generation text of the initial text and output; otherwise, the user's historical corpus information and the text to be verified are input into the semantic generation layer for text regeneration until the semantic anomaly assessment result reaches the preset generation standard; the result output layer is used to output the text generation text.

[0009] As a preferred technical solution of the present invention, the specific steps of generating text in the semantic generation layer include: extracting semantic features T of text short sentences in the text semantic extraction information n Perform condition recognition to obtain text generation condition features; extract features T based on text generation condition features and semantics of all text short sentences n Perform semantic representation with the language big model to obtain local generated semantic representation and global generated semantic representation; add position encoding based on the local generated semantic representation and the global generated semantic representation to obtain the representation of the text to be corrected; perform text correction on the representation of the text to be corrected based on the user semantic preference factor in the user's historical corpus information to obtain the text to be verified.

[0010] As a preferred technical solution of the present invention, the specific steps of performing logic verification in the sampling verification layer include: performing text preprocessing on the text to be corrected to obtain several groups of verification text units; using the formula Perform word meaning anomaly check to obtain word meaning anomaly score; wherein S represents the word meaning anomaly score, P1 represents the first group of verification text units in any two groups of verification text units, and P2 represents the second group of verification text units in any two groups of verification text units; wherein, represents the probability distribution of the first set of verification text units, The second group verifies the probability distribution of text units; represents the relative entropy between two sets of verification text units; It is used to measure the confidence of the semantics of the semantic units in the second group of verification text units; o is the semantic unit in P2; Represents the probability of semantic unit o in the probability distribution of the second group of verification text units; Sigmoid function is used to normalize the result score; Based on the above steps, traverse the word sense anomaly scores of all verification text units, extract the minimum value of the word sense anomaly scores of all verification text units, and obtain the semantic anomaly assessment result.

[0011] An intelligent semantic understanding and generation system based on ensemble learning, comprising: an integrated semantic understanding module, including a semantic understanding unit; the semantic understanding unit is used to obtain an initial text and a corresponding text user; based on the text user, the user's historical corpus information is obtained; semantic understanding is performed according to the initial text, the user's historical corpus information and the integrated semantic understanding model to obtain text semantic extraction information; the integrated semantic understanding model accurately extracts text semantic features through ensemble learning and weighted fusion, and effectively captures the user's semantic style, hot spot preferences and context associations in combination with the user's historical corpus information, significantly improving the accuracy, personalization and context consistency of semantic understanding; an integrated semantic generation module, including a semantic generation unit; the semantic generation unit is used to extract information according to text semantics, user historical corpus information, and the user's historical corpus information. The semantic generation model is used to generate text based on the user corpus information and the integrated semantic generation model to obtain the generated text; the integrated semantic generation model dynamically adjusts the generated content through multi-level generation, logic verification and user corpus optimization, combining local and global semantic representation to ensure the semantic accuracy, logical consistency and personalization of text generation; the historical corpus construction module includes a corpus construction unit; the corpus construction unit is used to construct user historical corpus information; the user historical corpus information is constructed based on the user's historical data; the user historical corpus information includes user semantic style factors, user semantic hot spot factors and user semantic preference factors; constructing user historical corpus information can accurately capture the user's unique language style, dynamic hot spot interests and long-term semantic preferences, and realize the dynamic quantification and continuous optimization of user semantic features.

[0012] The present invention has the following advantages: 1. The present invention combines the advantages of multiple base learners through a weighted fusion method of an integrated semantic understanding model, and can accurately extract semantic features in the text; combined with user historical corpus information, it effectively captures the user's unique language style and dynamic semantic preferences, and enhances the personalization and contextual consistency of semantic analysis; the integrated semantic generation model combines user historical corpus information during the generation process, and dynamically adjusts local and global semantic representations to ensure that the generated content is highly matched with the user's personalized needs; the multi-level generation and logic verification mechanism improves the semantic accuracy and logical consistency of the generated text, and meets the semantic expression habits of different users.

[0013] 2. The present invention divides the initial text through the sentence division layer, reduces the analysis granularity, refines the text processing into multiple short sentences, reduces complexity, and improves the model's recognition ability of detailed semantics; the text initial analysis layer extracts the feature information of text short sentences to provide rich context and basic semantic features for subsequent semantic understanding; the user-specific corpus analysis layer combines the user's historical corpus information with the current text features, and through the integrated learning optimization model, it can capture the user's personalized semantic preferences and dynamic semantic style, and achieve accurate understanding of complex semantics; the introduction of the user's historical corpus information enables the semantic understanding process to adapt to the semantic style, hot spot preferences and long-term semantic preferences of different users; the application of the integrated learning method in the user-specific corpus analysis layer, through multi-model weighted fusion, effectively combines global and local semantic information, and significantly improves the personalized semantic capture capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A schematic diagram of the structure of an intelligent semantic understanding and generation system based on ensemble learning adopted in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0016] Embodiment 1, an intelligent semantic understanding and generation method based on ensemble learning, comprising the following steps: obtaining an initial text and a corresponding text user; obtaining user historical corpus information based on the text user; obtaining the source of the initial text, for example, obtaining the text content input by the user through an application interface, a chat window or an API interface, such scenarios are common in chat robots, question-answering systems or automated writing tools; automatically generating the initial text by a specific task system; generally, each user has a unique ID, and the text data is directly bound to the ID to match the corresponding text user; performing semantic understanding based on the initial text, the user's historical corpus information and the integrated semantic understanding model to obtain text semantic extraction information; the integrated semantic understanding model accurately extracts text semantic features through ensemble learning and weighted fusion, and effectively captures the user's semantic style, hot spot preferences and context associations in combination with the user's historical corpus information, significantly improving the accuracy, personalization and context consistency of semantic understanding; the integrated semantic understanding model includes a sentence division layer, a text initial analysis layer, a user-specific corpus analysis layer and a semantic understanding output layer; the sentence division layer is used to divide the initial text into sentences to obtain the text short sentences to be analyzed. n , n=1, 2, ..., N; N is the number of text sentences contained in the initial text; the initial text analysis layer is used to analyze the text sentences J n Extract feature information and obtain the short sentence feature J of the text to be analyzedn '; The user-specific corpus analysis layer is used to perform semantic understanding based on the user's historical corpus information and the features of the short sentences in the text to be analyzed Jn', and obtain the semantic extraction features of the short sentences in the text T n ; Among them, the user-specific prediction analysis layer is constructed by an integrated learning method; the semantic understanding output layer is used to extract semantic features of all text short sentences T n The initial text is divided into multiple short sentences through the sentence division layer to reduce the analysis granularity and refine the text processing into multiple short sentences, thus reducing the complexity and improving the model's recognition ability of detailed semantics. The initial text analysis layer extracts the feature information of short sentences to provide rich context and basic semantic features for subsequent semantic understanding. The user-specific corpus analysis layer combines the user's historical corpus information with the current text features and optimizes the model through ensemble learning to capture the user's personalized semantic preferences and dynamic semantic style, thus achieving accurate understanding of complex semantics. The introduction of the user's historical corpus information enables the semantic understanding process to adapt to the semantic style, hot spot preferences and long-term semantic preferences of different users. The application of ensemble learning methods in the user-specific corpus analysis layer effectively combines global and local semantic information through multi-model weighted fusion, thus significantly improving the personalized semantic capture ability. The specific steps of constructing the user-specific corpus analysis layer include: the user-specific corpus analysis layer contains M semantic analysis layers, hierarchical user analysis layers and comprehensive semantic analysis layers. Among them, the M semantic analysis layers are trained by different base learners. In the M semantic analysis layers C m In the analysis, we treat the short sentence features J n 'Extract semantic features to obtain the semantic features of text short sentences J nm '; Where m = 1, 2, ..., M; At the semantic analysis layer C M In the example, for all the text short sentence semantic features J nm 'Perform feature fusion to obtain the fused sentence semantic feature G n ; In the hierarchical user analysis layer, there are style feature analysis sublayer, semantic preference analysis sublayer and context association analysis sublayer; the style feature analysis sublayer is used to fusion short sentence semantic feature G n Combine the user's historical corpus information to weight the user's semantic style factor and obtain the semantic feature F of the style text short sentence n ; The semantic preference analysis sublayer is used to integrate the semantic features of short sentences G n Combine the user's historical corpus information to weight the user's semantic hotspot factor and obtain the semantic feature R of the hotspot text short sentence n ; The contextual association analysis sublayer is used to integrate the semantic features of short sentences G n-1 , integrating the semantic features of short sentences G n and the fusion sentence semantic features G n+1Combined with the user's historical corpus information, the context semantic analysis is weighted to obtain the context text sentence semantic feature H n ; In the comprehensive semantic analysis layer, the semantic features of the style text short sentences F n , semantic features of hot text short sentences R n and the context text sentence semantic features H n Perform multi-level weighted fusion to obtain the text sentence semantic extraction feature T n ; Through different base learners (such as deep learning models, rule models or statistical models), multi-perspective semantic analysis of text short sentence features can fully extract the semantic features of the text and avoid the limitations of a single model; the feature fusion step further integrates multiple analysis results to improve the comprehensiveness and accuracy of semantic extraction; the fusion of short sentence semantic features G nThe generation retains the strengths of different learners and forms a multi-level and comprehensive semantic feature expression; based on the user's semantic style factor, it can generate personalized features for each user by custom weighting, so that semantic understanding is closer to the user's language style and expression habits; through the weighting of semantic hotspot factors in the user's corpus, it captures the user's dynamic interests and concerns, and adapts to the personalized needs of users in different scenarios; it integrates the semantic information of the previous short sentence, the current short sentence and the next short sentence, and combines the user's historical corpus for context weighting to ensure that the analysis process can capture semantic coherence and global consistency; it effectively avoids the semantic fault problem that may be caused by independent analysis of short sentences, and improves the model's ability to parse complex contextual relationships; it combines the semantic features of style text short sentences, The semantic features of hot text sentences and context text sentences are multi-level weighted fused to generate the final text sentence semantic extraction features; the comprehensive semantic analysis process realizes the organic combination of semantic style, dynamic hot spots and context association, ensuring that the semantic understanding results have high accuracy and strong applicability; the base learners include: logistic regression is a model commonly used for binary or multi-classification tasks. The parameters are estimated by maximizing the log-likelihood function. The training process includes inputting features and labels, fitting the model to the training data, and finally evaluating the classification accuracy; support vector machines are suitable for high-dimensional data or linearly inseparable data. By selecting a suitable kernel function, the hyperplane that maximizes the interval is found; the training process includes selecting a kernel function, training the model, and Test data to evaluate performance; decision trees are a common method for dealing with nonlinear relationships and rely on recursive partitioning of data sets; during the training process, tree structures are built based on the importance of features and are ultimately used for classification or regression tasks; random forests improve classification or regression performance by building multiple decision trees and voting to fuse their results; the training process includes randomly sampling data and features for each tree, generating multiple weak classifiers, and then obtaining prediction results through majority voting; semantic generation is performed based on text semantic extraction information, user historical corpus information, and an integrated semantic generation model to generate text; the integrated semantic generation model dynamically adjusts the generated content through multi-level generation, logic verification, and user corpus optimization, combining local and global semantic representations to ensure the semantic accuracy of text generation The integrated semantic generation model includes a semantic generation layer, a sampling verification layer and a result output layer; the semantic generation layer is used to generate text based on text semantic extraction information and a large language model to obtain a text to be verified; the sampling verification layer is used to perform logical verification on the text to be verified to obtain a semantic anomaly assessment result; if the semantic anomaly assessment result meets the preset generation standard, the text to be verified is used as the text generation text of the initial text for output; otherwise, the user's historical corpus information and the text to be verified are input into the semantic generation layer for text regeneration until the semantic anomaly assessment result meets the preset generation standard; the preset generation standard is set by professional and technical personnel according to actual conditions; the result output layer is used to output text generation text;The specific steps of generating text in the semantic generation layer include: extracting semantic features T of text short sentences in the text semantic extraction information; n Perform condition recognition to obtain text generation condition features; extract features T based on text generation condition features and semantics of all text short sentences n The semantic representation is performed with the language big model to obtain the local generated semantic representation and the global generated semantic representation; position encoding is added based on the local generated semantic representation and the global generated semantic representation to obtain the representation of the text to be corrected; the text to be corrected is corrected in combination with the user semantic preference factor in the user's historical corpus information to obtain the text to be verified; the semantic generation layer generates text based on the text semantic extraction information and the language big model, and accurately captures the generation conditions through conditional recognition of the semantic features of the text short sentences, effectively improving the semantic matching degree of the generated content; the local generated semantic representation is combined with the global generated semantic representation, so that the generated text content can pay attention to local details while maintaining the coherence of global semantics; the sampling verification layer performs logical verification on the generated text, and ensures that the generated content is logically rigorous and coherent through the semantic anomaly assessment mechanism; for text that does not meet the preset generation standard, the model will dynamically regenerate it in combination with the user's historical corpus information The generated content is iteratively optimized to avoid the degradation of user experience due to logical errors; the introduction of user semantic preference factors enables the generated content to reflect the user's language style, hot spots of interest and dynamic needs; in the text correction process, the user's personalized characteristics are integrated into the generated text to make the output results more in line with user expectations, improve the personalization of text generation and user satisfaction; through the multi-level architecture design of the semantic generation layer, sampling verification layer and result output layer, and the combination of local and global semantic representation, this method significantly improves the semantic accuracy, logical consistency and personalized performance of text generation; combined with the dynamic optimization strategy of user historical corpus information, the generated content is more in line with user needs, and the robustness and scalability of the generation process are enhanced; the specific steps of performing logical verification in the sampling verification layer include: performing text preprocessing on the text to be corrected to obtain several groups of verification text units; using the formula for any two groups of verification text units Perform word meaning anomaly check to obtain word meaning anomaly score; wherein S represents the word meaning anomaly score, P1 represents the first group of verification text units in any two groups of verification text units, and P2 represents the second group of verification text units in any two groups of verification text units; wherein, represents the probability distribution of the first set of verification text units, The second group verifies the probability distribution of text units; represents the relative entropy between two sets of verification text units; It is used to measure the confidence of the semantics of the semantic units in the second group of verification text units; o is the semantic unit in P2; Represents the probability of semantic unit o in the probability distribution of the second group of verification text units; the Sigmoid function is used to normalize the result score; based on the above steps, traverse the word meaning anomaly scores of all verification text units, extract the minimum value of the word meaning anomaly scores of all verification text units, and obtain the semantic anomaly evaluation result; use the formula to calculate the semantic anomaly score, and combine the word probability distribution and relative entropy measurement methods to effectively evaluate the degree of text semantic anomaly; introduce the comparison of the two groups of verification text units into logical verification, which can accurately capture semantic deviations and text anomalies and improve the accuracy of verification; by comparing the semantic distribution characteristics of the verification units, ensure that the generated text semantics are globally consistent; the Sigmoid function smoothes the score output to avoid abnormal scores being too sensitive or distorted, which helps to maintain logical coherence in the generated text; for text content whose scores do not meet the preset standards, the system can dynamically generate alternative texts based on user corpus information to ensure that the generated content meets the semantic standards and logic; this iterative verification mechanism not only improves the quality of generated text, but also enhances the robustness of the model; through the above steps and the semantic anomaly scoring formula, this method can significantly improve the logical consistency and semantic accuracy of the generated text, and at the same time enhance the robustness and adaptability of the model by dynamically optimizing the generated content. This process provides reliable technical support for the high-quality output of intelligent text generation, ensuring the effectiveness and availability of generated content in multi-scenario applications; user historical corpus information is constructed based on user historical data; user historical corpus information contains user semantic style factors, user semantic hot spot factors and user semantic preference factors; constructing user historical corpus information can accurately capture the user's unique language style, dynamic hot spot interests and long-term semantic preferences, and realize dynamic quantification and continuous optimization of user semantic features; the specific steps of constructing user historical corpus information include: collecting user historical data sets; user historical data sets contain user historical data D i ; User history data includes but is not limited to conversation records, user published content, search records, click behavior data and user feedback information; Set the time node i of user history data, i=1, 2, ...; i represents user history data D i The method of collecting user historical data sets includes the following steps and approaches: collecting all interaction records between users and the system, such as chatbot conversations, customer service records, etc.; extracting content posted by users on the platform, such as social media posts, comments, speeches on question-and-answer platforms, etc.; ensuring that the user's language expression and keywords are recorded; obtaining the user's search keywords and query history, including search engines, site searches, and semantic search behaviors; combining search frequency and time series data to analyze user interest hotspots; collecting user click behaviors on web pages or applications, including clicked links, browsed pages, and dwell time, and analyzing the types of content users pay attention to and their interactive habits; the collected data is public and legal data; for user historical data Di Analyze and extract the unique language style features of the text user to obtain the text user language style features W i ; For all text user language style features W i Perform feature quantification to obtain the user semantic style factor; i Perform topic modeling to obtain user hot topic features E i ; Using the formula Generate user hot topic features E i The time decay weight of , we get the weighted user hot topic feature E i '; For all weighted user hot topic features E i 'Perform feature quantification to obtain the user semantic hotspot factor; where A i Represents the user's hot topic feature E i The time decay weight is, I represents the current time node of constructing the user's historical corpus information; α represents the time decay coefficient; the time decay coefficient is set by professional technicians according to the actual situation; i Generate a long-term preference distribution vector and obtain the user preference distribution vector B i ; Use the sliding window method to capture all user preference distribution vectors B idynamic changes to obtain the user semantic preference factor; wherein, the user historical corpus information is updated at preset fixed time periods; the preset fixed time period is set by professional and technical personnel according to actual conditions; by analyzing the conversation records, published content, etc. in the user's historical data, the user's unique language style features are extracted to generate the user semantic style factor; the language style features are quantified so that the model can recognize the user's word usage habits and expressions, so as to better reflect the user's language style in the generated content; topic modeling identifies hot topics in the user's historical data, and combines the time decay weight formula to enable the model to give priority to recent user interests; the time decay weight effectively reduces the impact of expired information while retaining the significance of the user's recent interest topics; the weighted user semantic hot spot factor helps the model dynamically adapt to the user's current focus when generating content, and enhances the timeliness and relevance of semantics; using the long-term preference distribution vector, the dynamic changes are captured through the sliding window method, which can fully reflect the user's preference at different times. The generation of user semantic preference factors integrates short-term and long-term corpus information, so that the model can capture both short-term user behavior and long-term semantic preferences; a fixed time period is set to update the user's historical corpus information to ensure that the user's dynamic behavior can be reflected in the model in a timely manner; for example, in an intelligent customer service system, the intelligent semantic understanding and generation method can be used to provide a personalized user interaction experience; by collecting the user's historical conversation records, feedback information, and search behavior, the user's language style, hot topics, and long-term preferences can be extracted to build a user corpus; when a user initiates a consultation, the semantic understanding module is used to accurately parse the user's questions, and the user corpus information and the semantic generation model are combined to generate personalized answers that fit the user's language style; at the same time, by dynamically adjusting hot topics and context associations, the reply logic is ensured to be consistent and highly relevant to the user's current needs, thereby achieving a fast response and high-satisfaction intelligent service experience. This method is widely applicable to scenarios such as e-commerce customer service, technical support, and intelligent question and answer.

[0017] Example 2, an intelligent semantic understanding and generation system based on ensemble learning, see Figure 1As shown, it includes: an integrated semantic understanding module, including a semantic understanding unit; the semantic understanding unit is used to obtain the initial text and the corresponding text user; the user's historical corpus information is obtained based on the text user; semantic understanding is performed according to the initial text, the user's historical corpus information and the integrated semantic understanding model to obtain text semantic extraction information; the integrated semantic understanding model accurately extracts text semantic features through integrated learning and weighted fusion, and effectively captures the user's semantic style, hot spot preferences and context associations in combination with the user's historical corpus information, significantly improving the accuracy, personalization and context consistency of semantic understanding; the integrated semantic generation module includes a semantic generation unit; the semantic generation unit is used to extract text semantic information, user historical corpus information and integrated semantic generation information according to the text semantic extraction information, user historical corpus information and integrated semantic generation information. The model is used for semantic generation to obtain text generation text; the integrated semantic generation model dynamically adjusts the generated content through multi-level generation, logic verification and user corpus optimization, combining local and global semantic representation to ensure the semantic accuracy, logical consistency and personalization of text generation; the historical corpus construction module includes a corpus construction unit; the corpus construction unit is used to construct user historical corpus information; user historical corpus information is constructed based on user historical data; user historical corpus information includes user semantic style factors, user semantic hot spot factors and user semantic preference factors; constructing user historical corpus information can accurately capture the user's unique language style, dynamic hot spot interests and long-term semantic preferences, and realize dynamic quantification and continuous optimization of user semantic features.

[0018] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. An intelligent semantic understanding and generation method based on ensemble learning, characterized in that: The following steps are involved: Obtain the initial text and the corresponding text user; obtain user historical corpus information based on the text user; user historical corpus information is constructed based on user historical data; user historical corpus information includes user semantic style factors, user semantic hot spot factors and user semantic preference factors; constructing user historical corpus information can accurately capture the user's unique language style, dynamic hot spot interests and long-term semantic preferences, and realize dynamic quantification and continuous optimization of user semantic features; perform semantic understanding based on the initial text, user historical corpus information and integrated semantic understanding model to obtain text semantic extraction information; the integrated semantic understanding model accurately extracts text semantic features through integrated learning and weighted fusion, and effectively captures user semantic style, hot spot preferences and context associations in combination with user historical corpus information, significantly improving the accuracy, personalization and context consistency of semantic understanding; Semantic generation is performed based on text semantic extraction information, user historical corpus information and integrated semantic generation model to obtain text generation text; the integrated semantic generation model dynamically adjusts the generated content through multi-level generation, logic verification and user corpus optimization, combining local and global semantic representation to ensure the semantic accuracy, logical consistency and personalization of text generation.

2. The intelligent semantic understanding and generation method based on ensemble learning according to claim 1, characterized in that: The integrated semantic understanding model includes a sentence segmentation layer, a text initial analysis layer, a user-specific corpus analysis layer, and a semantic understanding output layer; the sentence segmentation layer is used to segment the initial text into sentences and obtain the short sentences of the text to be analyzed. n , n=1, 2, ..., N; N is the number of text sentences contained in the initial text; the initial text analysis layer is used to analyze the text sentences J n Extract feature information and obtain the short sentence feature J of the text to be analyzed n '; The user-specific corpus analysis layer is used to perform semantic understanding based on the user's historical corpus information and the features of the short sentences in the text to be analyzed Jn', and obtain the semantic extraction features of the short sentences in the text T n ; Among them, the user-specific prediction analysis layer is constructed by an integrated learning method; the semantic understanding output layer is used to extract semantic features of all text short sentences T n Combined to obtain text semantic extraction information.

3. The intelligent semantic understanding and generation method based on ensemble learning according to claim 2, characterized in that: The specific steps of constructing the user-specific corpus analysis layer include: the user-specific corpus analysis layer includes M semantic analysis layers, hierarchical user analysis layer and comprehensive semantic analysis layer; wherein the M semantic analysis layers are trained by different base learners; in the M semantic analysis layers C m In the analysis, we treat the short sentence features J n 'Extract semantic features to obtain the semantic features of text short sentences J nm '; Where m = 1, 2, ..., M; At the semantic analysis layer C M In the example, for all the text short sentence semantic features J nm 'Perform feature fusion to obtain the fused sentence semantic feature G n ; In the hierarchical user analysis layer, there are style feature analysis sublayer, semantic preference analysis sublayer and context association analysis sublayer; the style feature analysis sublayer is used to fusion short sentence semantic feature G n Combine the user's historical corpus information to weight the user's semantic style factor and obtain the semantic feature F of the style text short sentence n ; The semantic preference analysis sublayer is used to integrate the semantic features of short sentences G n Combine the user's historical corpus information to weight the user's semantic hotspot factor and obtain the semantic feature R of the hotspot text short sentence n ; The contextual association analysis sublayer is used to integrate the semantic features of short sentences G n-1 , integrating the semantic features of short sentences G n and fusion sentence semantic features G n+1 Combined with the user's historical corpus information, the context semantic analysis is weighted to obtain the context text sentence semantic feature H n ; In the comprehensive semantic analysis layer, the semantic features of the style text short sentences F n , semantic features of hot text short sentences R n and the context text sentence semantic features H n Perform multi-level weighted fusion to obtain the text short sentence semantic extraction feature T n .

4. The intelligent semantic understanding and generation method based on ensemble learning according to claim 3 is characterized in that: The specific steps of constructing user history corpus information include: collecting user history data sets; the user history data sets contain user history data D i ; User history data includes but is not limited to conversation records, user published content, search records, click behavior data and user feedback information; Set the time node i of user history data, i=1, 2, ...; i represents user history data D i The time node corresponding to the collection; the user's historical data D i Analyze and extract the unique language style features of the text user to obtain the text user language style features W i ; For all text user language style features W i Perform feature quantification to obtain the user semantic style factor; i Perform topic modeling to obtain user hot topic features E i ; Using the formula Generate user hot topic features E i The time decay weight of , we get the weighted user hot topic feature E i '; For all weighted user hot topic features E i 'Perform feature quantification to obtain the user semantic hotspot factor; where A i Represents the user's hot topic feature E i The time decay weight of user historical data D is: I represents the current time node of constructing user historical corpus information; α represents the time decay coefficient; i Generate a long-term preference distribution vector and obtain the user preference distribution vector B i ; Use the sliding window method to capture all user preference distribution vectors B i The user's semantic preference factor is obtained by dynamically changing the user's historical corpus information at preset fixed time periods.

5. The intelligent semantic understanding and generation method based on ensemble learning according to claim 4 is characterized in that: The integrated semantic generation model includes a semantic generation layer, a sampling verification layer, and a result output layer; the semantic generation layer is used to generate text based on text semantic extraction information and a large language model to obtain the text to be verified; the sampling verification layer is used to perform logical verification on the text to be verified to obtain the semantic anomaly assessment result; If the semantic anomaly assessment result reaches the preset generation standard, the text to be verified is output as the text generation text of the initial text; otherwise, the user's historical corpus information and the text to be verified are input into the semantic generation layer for text regeneration until the semantic anomaly assessment result reaches the preset generation standard; the result output layer is used to output the text generation text.

6. The intelligent semantic understanding and generation method based on ensemble learning according to claim 5, characterized in that: The specific steps of text generation in the semantic generation layer include: extracting semantic features T from the text short sentences in the text semantic extraction information n Perform condition recognition to obtain text generation condition features; extract features T based on text generation condition features and semantics of all text short sentences n Perform semantic representation with the language big model to obtain local generated semantic representation and global generated semantic representation; add position encoding based on the local generated semantic representation and the global generated semantic representation to obtain the representation of the text to be corrected; perform text correction on the representation of the text to be corrected based on the user semantic preference factor in the user's historical corpus information to obtain the text to be verified.

7. The intelligent semantic understanding and generation method based on ensemble learning according to claim 6 is characterized in that: The specific steps of logical verification in the sampling verification layer include: preprocessing the text to be corrected to obtain several groups of verification text units; using the formula Perform word meaning anomaly check to obtain word meaning anomaly score; wherein S represents the word meaning anomaly score, P1 represents the first group of verification text units in any two groups of verification text units, and P2 represents the second group of verification text units in any two groups of verification text units; wherein, represents the probability distribution of the first set of verification text units, The second group verifies the probability distribution of text units; represents the relative entropy between two sets of verification text units; It is used to measure the confidence of the semantics of the semantic units in the second group of verification text units; o is the semantic unit in P2; Represents the probability of semantic unit o in the probability distribution of the second group of verification text units; Sigmoid function is used to normalize the result score; Based on the above steps, traverse the word sense anomaly scores of all verification text units, extract the minimum value of the word sense anomaly scores of all verification text units, and obtain the semantic anomaly assessment result.

8. An intelligent semantic understanding and generation system based on ensemble learning, characterized in that: The system applies an intelligent semantic understanding and generation method based on integrated learning as described in any one of claims 1 to 7 above, including: an integrated semantic understanding module, including a semantic understanding unit; the semantic understanding unit is used to obtain the initial text and the corresponding text user; obtain the user's historical corpus information based on the text user; perform semantic understanding based on the initial text, the user's historical corpus information and the integrated semantic understanding model to obtain text semantic extraction information; the integrated semantic understanding model accurately extracts text semantic features through integrated learning and weighted fusion, and effectively captures the user's semantic style, hot spot preferences and context associations in combination with the user's historical corpus information, significantly improving the accuracy, personalization and context consistency of semantic understanding; an integrated semantic generation module, including a semantic generation unit; the semantic generation unit is used to extract text semantic information based on the text semantics. The semantic extraction information, user historical corpus information and integrated semantic generation model are used for semantic generation to obtain text generation text; the integrated semantic generation model dynamically adjusts the generated content through multi-level generation, logic verification and user corpus optimization, combining local and global semantic representation to ensure the semantic accuracy, logical consistency and personalization of text generation; the historical corpus construction module includes a corpus construction unit; the corpus construction unit is used to construct user historical corpus information; user historical corpus information is constructed based on user historical data; user historical corpus information includes user semantic style factors, user semantic hot spot factors and user semantic preference factors; constructing user historical corpus information can accurately capture the user's unique language style, dynamic hot spot interests and long-term semantic preferences, and realize dynamic quantification and continuous optimization of user semantic features.

Images