A large language model system based on keyword tracing

Abstract

The invention provides a large language model system based on keyword tracing, and relates to the technical field of data processing. The system comprises: a data acquisition module; a text preprocessing module; a keyword screening module; a semantic expansion module, which is used for performing semantic expansion processing on keywords in a preliminary keyword data set to generate an extended keyword data set; a keyword propagation detection module, which is used for generating a keyword propagation path data set according to the origin identifier, occurrence frequency, timestamp and user behavior information of the keywords in the extended keyword data set; a heat calculation module, which is used for generating a heat score according to the keyword propagation path data set; a semantic drift detection module, which is used for generating a semantic drift identifier according to the keyword propagation path data set; an information tracing report generation module, which is used for generating a tracing report according to the propagation path data set, the keyword heat score and the semantic drift identifier; the invention has a good keyword tracing effect.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a large language model system based on keyword tracing. Background Art

[0002] In large language model systems, keyword tracing technology is widely used in information monitoring. Existing technologies usually scan and match a large amount of text data through a predefined set of keywords to identify content related to the information. In this process, the system relies on the frequency of occurrence, contextual relationship and location of keywords to determine the information. However, when dealing with dynamic text content, the keyword tracing method of the existing technology relies on static keyword matching methods and cannot adapt to the problem that the semantics of complex texts change over time.

[0003] Keywords may experience semantic drift in different contexts and time periods, that is, the semantics of the same keyword may change significantly in the initial stage and the subsequent information dissemination process. Traditional keyword tracing technology may not be able to capture these semantic changes, and the tracing results are often biased. Summary of the invention

[0004] The purpose of the present invention is to provide a large language model system based on keyword tracing, aiming to solve the problems mentioned in the background technology.

[0005] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0006] A large language model system based on keyword tracing, the system comprising:

[0007] A data collection module, used to obtain text data from multiple information platforms, wherein the text data includes text content, timestamp and user behavior information;

[0008] A text preprocessing module is used to preprocess the text data to obtain a preprocessed text data set, wherein the preprocessing includes word segmentation, denoising and format standardization of the text data;

[0009] A keyword screening module is used to perform preliminary screening of keywords in the preprocessed text data set according to a preset keyword list to generate a preliminary keyword data set;

[0010] A semantic expansion module, used to perform semantic expansion processing on the keywords in the preliminary keyword data set through context analysis and synonym association processing to generate an expanded keyword data set;

[0011] The keyword propagation origin detection module is used to identify the information platform where the keyword first appears based on the timestamp of the keyword in the extended keyword data set, and determine the origin point of the keyword in combination with user behavior information to generate the keyword origin identifier;

[0012] The keyword propagation path calculation module is used to determine the propagation order of keywords in various information platforms based on the origin identification, occurrence frequency, timestamp and user behavior information of the keywords in the extended keyword data set, and generate a keyword propagation path data set;

[0013] The keyword heat calculation module is used to calculate the spread heat of each keyword based on the keyword propagation path data set and generate a heat score;

[0014] The semantic change tracking module is used to analyze the contextual semantics of keywords in different time periods based on the keyword propagation path dataset, identify the semantic drift of the same keyword during the propagation process, and generate a semantic change trajectory dataset;

[0015] A semantic drift detection module is used to generate a semantic drift identifier of a keyword based on a semantic change trajectory dataset;

[0016] The information tracing report generation module is used to generate an information tracing report based on the keyword propagation path dataset, keyword heat score and semantic drift identification.

[0017] Preferably, the text preprocessing module includes:

[0018] The word segmentation processing module is used to perform word segmentation processing on text data, divide continuous text into independent vocabulary units, and generate a word segmentation data set;

[0019] A denoising processing module is used to remove low-frequency or high-redundancy words, irrelevant symbols and stop words in the word segmentation data set according to the frequency of occurrence of the words in the multi-information platform to generate a text data set;

[0020] The format standardization module is used to unify the format of the timestamps of the text data set and to perform consistent conversion on the special characters and punctuation marks in the text to obtain preprocessed data.

[0021] Preferably, the keyword screening module includes:

[0022] A keyword matching module is used to perform preliminary matching of words in the preprocessed text data set according to a preset keyword list to generate a keyword data set;

[0023] The keyword updating module is used to adjust the first keyword data set according to the real-time words appearing in the information platform and the corresponding frequency changes to obtain the preliminary keyword data set.

[0024] Preferably, the semantic expansion module includes:

[0025] The context extraction module is used to extract the combination of several words before and after each keyword in different texts to build a context window for the keyword;

[0026] The relevance calculation module calculates the relevance between the keyword and other words in its context window, identifies the semantic differences of the keyword in different contexts, and obtains a context semantic data set. The relevance is calculated by the following formula:

[0027] ,

[0028] in, For keywords The relevance in the current context, For keywords With the words in the context The semantic weight of For words The frequency of occurrence in the context, For keywords With words The semantic distance of ,in, and Separate words and keywords In the semantic space dimensional word embedding vector representation, is the dimension of the word embedding vector, For keywords With words The adjacent time difference, For words Total number of;

[0029] The synonym extraction module is used to identify synonyms with similar semantics to the keyword based on the association information in the context semantic data set, and generate a synonym extension data set;

[0030] The synonym library update module screens the synonyms in the synonym extension data set according to a preset frequency threshold and a preset relevance threshold, and adds them to the preset preliminary synonym library to obtain a final synonym library;

[0031] The synonym association module is used to calculate the semantic similarity between each synonym and the preliminary keyword based on the final synonym library, and add the synonyms that meet the preset similarity threshold to the preliminary keyword data set to obtain the extended keyword data set. The semantic similarity is calculated by the following formula:

[0032] ;

[0033] in, For keywords Synonyms The semantic similarity between For keywords The context weight in the current context, Synonyms The number of occurrences in the context, For keywords Synonyms The semantic distance of Synonyms The context length is the total number of words directly or indirectly connected to it in all texts. Synonyms The semantic extension depth of The number of words directly and indirectly connected to other words in the current text, Synonyms The semantic information density of ,in, is the total number of occurrences in all texts, , Synonyms In the context The probability of occurrence of a keyword, is the total number of keywords, Synonyms User activity on the information platform, ,in, Synonyms The information platform and words The number of first positive markers associated, Synonyms The information platform and words The number of direct markers involved, Synonyms The information platform and words The number of related second positive markings, Synonyms The total number of users on the information platform, Synonyms The time decay coefficient, Used to indicate synonyms The impact of usage frequency changes over time.

[0034] Preferably, the keyword propagation origin detection module includes:

[0035] A first appearance detection module, used to determine the origin point of the first appearance of the keyword in multiple information platforms according to the timestamp of each keyword in the extended keyword data set, and generate an initial propagation time record of the keyword, wherein the origin point includes time and information platform;

[0036] The communication origin identification module is used to calculate the user's communication potential based on the user's activity and interaction frequency, and determine the origin identification of the keyword in combination with the initial communication time record of the keyword;

[0037] The communication potential evaluation module is used to evaluate the communication potential of a keyword at the origin point, and the communication potential is calculated by the following formula:

[0038] ,

[0039] in, For keywords potential for dissemination, For keywords User activity at the origin point, For keywords Direct labeling density in the origin point, For keywords The mean semantic distance between the source node and other similar words, For keywords The time difference between the first occurrence and the current time.

[0040] Preferably, the keyword The direct labeling density in the origin point is calculated by the following formula:

[0041] ,in, is the number of interactions related to the keyword within a specified time. is the number of unique users who participated in direct tagging during the specified time.

[0042] Preferably, the keyword propagation path calculation module includes:

[0043] The path identification module is used to generate a keyword propagation path dataset based on the origin identification of the keyword and by analyzing the frequency of occurrence, timestamps and user behaviors of the keyword in different time periods;

[0044] The propagation path weight calculation module is used to calculate the propagation path weight of keywords in different propagation nodes according to the order of keyword appearance, time interval and user behavior characteristics in the propagation path data set. The propagation path weight is calculated by the following formula:

[0045] ,

[0046] in, For keywords The propagation path weight, For keywords In the The appearance time of the propagation node, For the The user activity of the propagation nodes, For the propagation node The number of hops to the next propagation node or the second forward label interval time, For the The number of interactions among users of each propagation node, is the normalization factor, is the total number of propagation nodes.

[0047] Preferably, the keyword heat calculation module includes:

[0048] The heat evaluation module is used to calculate the heat score of the keyword based on the propagation path weight in the keyword propagation path data set and the user's interactive behavior. The heat score is calculated using the following formula:

[0049] ,

[0050] in, For keywords The popularity rating of For keywords In the propagation node The path weight on For keywords In the propagation node The number of direct tags by the user, For the propagation node On and Keywords The number of occurrences of related positive sentiment words, For the propagation node On and Keywords The number of occurrences of related negative sentiment words, For keywords In the propagation node The number of first positive tags of the user;

[0051] The trend track generation module is used to generate the heat change trend track of the keyword in different time periods according to the heat score, and the heat change trend track is used to describe the heat growth and decay trend of the keyword.

[0052] Preferably, the semantic change tracking module includes:

[0053] The semantic drift detection module is used to analyze the semantic changes of keywords in different time periods and propagation paths based on the propagation path data set, identify the semantic drift of keywords in the multi-node propagation process, and generate semantic drift markers;

[0054] The semantic change trajectory generation module is used to generate the semantic change trajectory of the keyword according to the semantic drift mark of the keyword in the propagation path.

[0055] Preferably, the semantic drift detection module includes:

[0056] The drift degree calculation module is used to analyze the semantic drift degree of the semantic shift of the keyword according to the semantic change trajectory and generate a semantic drift identifier. The semantic drift degree is calculated by the following formula:

[0057] ,

[0058] in, For keywords The degree of semantic drift, For keywords In the The semantic distance between the beginning and the end of a time period, ,in, For keywords In time period The word embedding vector The value of the dimension, For keywords In the previous period The word embedding vector The value of the dimension, For keywords In the The number of occurrences in a time period, For keywords In the The direct labeling density of each time period, For keywords In the The time interval of the time period, For keywords In the User activity in a certain time period, is the total number of time periods.

[0059] The above solution of the present invention includes at least the following beneficial effects:

[0060] First, the system obtains text data containing timestamps and user behavior information from multiple information platforms through the data collection module, ensuring the extensiveness and real-time nature of the data. Compared with the traditional data collection method of a single information platform, this multi-source data collection can capture the information trends of events more comprehensively. Secondly, the text preprocessing module converts the raw data into standardized text suitable for keyword analysis through word segmentation, denoising and format standardization, greatly improving the clarity of the data and the efficiency of subsequent processing.

[0061] In the keyword screening module, the system not only relies on preliminary keyword matching, but also implements a flexible keyword screening mechanism by dynamically adjusting the keyword list. Combined with the context analysis and synonym expansion of the semantic expansion module, the system can cope with the semantic drift of keywords in different contexts. Different from the traditional static keyword matching method, the semantic expansion module of the present invention can dynamically expand the semantic boundaries of keywords and effectively capture the hidden information in the semantic changes, thereby significantly improving the accuracy and extensiveness of keyword tracing.

[0062] In addition, the keyword propagation origin detection module accurately locates the starting node of keyword propagation based on timestamps and user behavior characteristics, laying a solid foundation for subsequent propagation path analysis. Through the propagation path calculation module, the system can identify the breadth and depth of keyword propagation and provide comprehensive information propagation path analysis. On this basis, the system's heat calculation module combines the user's interactive behavior and propagation path weight to calculate the keyword's propagation heat, and then identify hot keywords that users pay more attention to.

[0063] The present invention also dynamically captures the semantic changes of keywords during the dissemination process through the semantic change tracking module and the semantic drift detection module, and generates semantic change trajectories and semantic drift identifiers. These functions enable the system to monitor the semantic evolution of keywords at different nodes, providing users with a deep information analysis perspective. Compared with the prior art, this system not only has a stronger ability to cope with semantic drift, but also ensures the timeliness and accuracy of keyword tracing through real-time analysis and dynamic updates.

[0064] This system overcomes the limitations of static keyword matching in existing technologies through the collaborative work of multiple modules. The system combines dynamic data collection, flexible keyword screening, accurate propagation path analysis and semantic change detection to provide a more comprehensive and efficient information monitoring and keyword tracing tool. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is an architecture diagram of a large language model system based on keyword tracing provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0066] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0067] like Figure 1 As shown, an embodiment of the present invention proposes a large language model system based on keyword tracing, and the system includes:

[0068] A data collection module, used to obtain text data from multiple information platforms, wherein the text data includes text content, timestamp and user behavior information;

[0069] A text preprocessing module is used to preprocess the text data to obtain a preprocessed text data set, wherein the preprocessing includes word segmentation, denoising and format standardization of the text data;

[0070] A keyword screening module is used to perform preliminary screening of keywords in the preprocessed text data set according to a preset keyword list to generate a preliminary keyword data set;

[0071] A semantic expansion module, used to perform semantic expansion processing on the keywords in the preliminary keyword data set through context analysis and synonym association processing to generate an expanded keyword data set;

[0072] The keyword propagation origin detection module is used to identify the information platform where the keyword first appears based on the timestamp of the keyword in the extended keyword data set, and determine the origin point of the keyword in combination with user behavior information to generate the keyword origin identifier;

[0073] The keyword propagation path calculation module is used to determine the propagation order of keywords in various information platforms based on the origin identification, occurrence frequency, timestamp and user behavior information of the keywords in the extended keyword data set, and generate a keyword propagation path data set;

[0074] The keyword heat calculation module is used to calculate the spread heat of each keyword based on the keyword propagation path data set and generate a heat score;

[0075] The semantic change tracking module is used to analyze the contextual semantics of keywords in different time periods based on the keyword propagation path dataset, identify the semantic drift of the same keyword during the propagation process, and generate a semantic change trajectory dataset;

[0076] A semantic drift detection module is used to generate a semantic drift identifier of a keyword based on a semantic change trajectory dataset;

[0077] The information tracing report generation module is used to generate an information tracing report based on the keyword propagation path dataset, keyword heat score and semantic drift identification.

[0078] In an embodiment of the present invention, a scheme for acquiring text data from multiple information platforms is provided. The system collects text data through a data acquisition module, and includes timestamps and user behavior information, and can process multi-source heterogeneous data in real time to ensure the timeliness and extensiveness of information. Compared with the prior art, the system has higher data acquisition accuracy, and combines multi-dimensional data, such as user behavior, information platform attributes, etc., which greatly improves the integrity and accuracy of information.

[0079] Through the text preprocessing module, the system can effectively perform word segmentation, denoising and format standardization on the original text, significantly improving the efficiency of subsequent processing steps. In the prior art, the preprocessing steps are usually relatively basic, while the present invention has made significant improvements in word segmentation accuracy, noise processing and timestamp standardization, ensuring the high quality of processed data, thereby improving the accuracy of keyword recognition.

[0080] The keyword screening module in the system performs preliminary screening based on the preset keyword list to generate a preliminary keyword data set. By combining context analysis and synonym association processing, the system can dynamically expand the semantic boundaries of keywords to form an extended keyword data set, thereby effectively dealing with the problem of keyword semantic drift in different scenarios. In contrast, traditional technologies usually rely on static keyword libraries and are difficult to cope with semantic changes, while the present invention can improve the accuracy of keyword recognition through real-time semantic expansion.

[0081] The keyword propagation origin detection module can accurately identify the keyword propagation origin point and generate the keyword origin identifier based on the information platform and user behavior information where the keyword first appeared, combined with the timestamp. This innovation greatly improves the timeliness and accuracy of the system in traceability analysis.

[0082] Through the keyword propagation path calculation module and the keyword heat calculation module, this system can accurately calculate the propagation path and propagation heat of each keyword at each propagation node, generate relevant propagation path and heat data, and help users identify the propagation trend of hot information. Compared with the existing technology, this system can obtain more comprehensive information analysis results through the combination of path calculation and heat analysis.

[0083] The system also includes a semantic change tracking module and a semantic drift detection module, which are used to analyze the semantic changes of keywords and generate semantic change trajectories of keywords, providing refined semantic variation analysis functions. By introducing semantic change trajectories, the system can better capture the semantic evolution of keywords during the dissemination process, and thus provide more valuable information tracing reports.

[0084] Through the mutual cooperation of the above modules, this system provides more efficient and accurate information tracking and keyword tracing functions compared with the existing technology, especially with significant advantages in dealing with semantic drift and multi-source data processing.

[0085] Specifically, the information traceability report generation module generates a complete information traceability report for users, including the keyword propagation origin, path, heat change and semantic drift, by comprehensively analyzing the propagation path data, heat score and semantic drift identification. This report provides a panoramic information analysis from the origin of keywords to the propagation path, heat fluctuation and semantic change. In the prior art, information reports are usually presented in a single dimension, such as heat or propagation path, which makes it difficult to provide comprehensive analysis support. The information traceability report of the present invention integrates data from multiple dimensions, making the report content more comprehensive and insightful.

[0086] Through this module, users can obtain detailed information about the origin node of a keyword, its propagation path, its popularity and sentiment feedback at different nodes, etc. This module can not only help users understand the spread of keywords between nodes, but also identify key nodes and time periods in information dissemination through heat change analysis.

[0087] In addition, the information tracing report can also help users identify semantic drift through semantic change information and understand the evolution of the semantics of keywords during the dissemination process. Compared with traditional information reports, the information tracing report generated by this system is real-time, comprehensive and accurate, providing users with in-depth analysis support, allowing users to make more accurate judgments on the overall situation of information and take targeted countermeasures.

[0088] In a preferred embodiment of the present invention, the text preprocessing module includes:

[0089] The word segmentation processing module is used to perform word segmentation processing on text data, divide continuous text into independent vocabulary units, and generate a word segmentation data set;

[0090] A denoising processing module is used to remove low-frequency or high-redundancy words, irrelevant symbols and stop words in the word segmentation data set according to the frequency of occurrence of the words in the multi-information platform to generate a text data set;

[0091] The format standardization module is used to unify the format of the timestamps of the text data set and to perform consistent conversion on the special characters and punctuation marks in the text to obtain preprocessed data.

[0092] In the embodiment of the present invention, through the word segmentation processing module, the system can divide the continuous text into independent vocabulary units and generate a word segmentation data set. This function significantly improves the efficiency of subsequent data analysis. Compared with the basic word segmentation method in the prior art, this system optimizes the word segmentation algorithm according to the characteristics of the multi-information platform to ensure the accuracy of the word segmentation results, which is particularly suitable for the processing of complex texts.

[0093] The denoising module purifies the text data by removing low-frequency words, irrelevant symbols and stop words, making it more suitable for subsequent keyword recognition. Traditional denoising technology is usually too simple and easily leads to information loss. The present invention achieves more sophisticated denoising through frequency analysis of data from different information platforms, ensuring the integrity of text information.

[0094] The format standardization module ensures the standardization of data formats by unifying the timestamp format of text data and converting special characters consistently. This module not only improves data consistency, but also reduces errors in the system when processing multi-source data, and improves the accuracy of overall data processing.

[0095] Through the combination of these modules, the text preprocessing module achieves efficient data cleaning and standardization processing. Compared with the existing technology, it provides a more complete data preprocessing process to ensure the efficiency and accuracy of subsequent processing steps.

[0096] In a preferred embodiment of the present invention, the keyword screening module includes:

[0097] A keyword matching module is used to perform preliminary matching of words in the preprocessed text data set according to a preset keyword list to generate a keyword data set;

[0098] The keyword updating module is used to adjust the first keyword data set according to the real-time words appearing in the information platform and the corresponding frequency changes to obtain the preliminary keyword data set.

[0099] In the embodiment of the present invention, the keyword screening module is used to screen out words that match the preset keyword list from the purified text data set and generate a keyword data set. Traditional technologies usually rely on a static keyword library, while the keyword matching module in the present invention can flexibly adjust the keyword matching rules according to actual needs, thereby improving the dynamic nature of keyword recognition.

[0100] Through the keyword update module, the system can dynamically adjust the keyword library according to the new entries and their frequency changes that appear in real time. The introduction of this module effectively solves the problem of delayed keyword library updates in the prior art, ensuring that the system can cope with the rapidly changing information environment. In traditional systems, keyword libraries usually need to be updated manually and cannot cope with real-time requirements, while the automatic update mechanism of the present invention greatly improves the real-time and adaptability of the keyword library.

[0101] The keyword screening module of this system can not only improve the accuracy of keyword recognition, but also ensure the flexibility of the system through dynamic adjustment and matching of keywords, so as to adapt to various complex data environments.

[0102] Specifically, the keyword matching module performs a preliminary screening of the cleansed text dataset using a preset keyword list to identify keywords in the text that are related to specific events or information. The purpose of this step is to quickly screen out potentially relevant keywords for further processing. The keyword list is usually constructed based on a priori hot words and can be generated through machine learning or statistical methods based on historical data.

[0103] The system can use a database or hash table to store a preset keyword list and scan the text data by matching word by word. In order to improve the matching efficiency, a fast search algorithm based on Bloom filter can be used to filter out words that do not meet the conditions. During the matching process, the system will mark the keywords that meet the conditions and generate a preliminary keyword data set.

[0104] Specifically, the keyword library update module dynamically updates the keyword list in real time based on the current information data to ensure that the system remains sensitive to new hot words. This step screens out potential hot words by analyzing the frequency changes and relevance of words in the information platform and adds them to the keyword library so that the system can identify new keywords in subsequent analysis.

[0105] The keyword library update module can be implemented through data flow analysis technology to include high-frequency new words or highly related words that appear in real-time data into the keyword library. The frequency of words in the latest data can be counted using a sliding window or time-based batch mode. For new high-frequency words, a frequency threshold can be set to automatically add them to the keyword library to ensure that the system reflects the development trend of keywords in real time. Through this dynamic update mechanism, the system's keyword library can maintain flexibility and timeliness, ensuring the accuracy of information monitoring.

[0106] In a preferred embodiment of the present invention, the semantic expansion module includes:

[0107] The context extraction module is used to extract the combination of several words before and after each keyword in different texts to build a context window for the keyword;

[0108] The relevance calculation module calculates the relevance between the keyword and other words in its context window, identifies the semantic differences of the keyword in different contexts, and obtains a context semantic data set. The relevance is calculated by the following formula:

[0109] ,

[0110] in, For keywords The relevance in the current context, For keywords With the words in the context The semantic weight of For words The frequency of occurrence in the context, For keywords With words The semantic distance of ,in, and Separate words and keywords In the semantic space dimensional word embedding vector representation, is the dimension of the word embedding vector, For keywords With words The adjacent time difference, For words Total number of;

[0111] The synonym extraction module is used to identify synonyms with similar semantics to the keyword based on the association information in the context semantic data set, and generate a synonym extension data set;

[0112] The synonym library update module screens the synonyms in the synonym extension data set according to a preset frequency threshold and a preset relevance threshold, and adds them to the preset preliminary synonym library to obtain a final synonym library;

[0113] The synonym association module is used to calculate the semantic similarity between each synonym and the preliminary keyword based on the final synonym library, and add the synonyms that meet the preset similarity threshold to the preliminary keyword data set to obtain the extended keyword data set. The semantic similarity is calculated by the following formula:

[0114] ;

[0115] in, For keywords Synonyms The semantic similarity between For keywords The context weight in the current context, Synonyms The number of occurrences in the context, For keywords Synonyms The semantic distance of Synonyms The context length is the total number of words directly or indirectly connected to it in all texts. Synonyms The semantic extension depth of The number of words directly and indirectly connected to other words in the current text, Synonyms The semantic information density of ,in, is the total number of occurrences in all texts, , Synonyms In the context The probability of occurrence of a keyword, is the total number of keywords, Synonyms User activity on the information platform, ,in, Synonyms The information platform and words The number of first positive markers associated, Synonyms The information platform and words The number of direct markers involved, Synonyms The information platform and words The number of related second positive markings, Synonyms The total number of users on the information platform, Synonyms The time decay coefficient, Used to indicate synonyms The impact of usage frequency changes over time.

[0116] In the embodiment of the present invention, the semantic expansion module constructs a context window of the keyword by analyzing the context of the keyword, and calculates the semantic relevance of the keyword based on the context information. Compared with the existing static keyword library, the present invention can not only identify the direct semantics of the keyword through context analysis, but also expand its potential semantic relevance, thereby improving the accuracy of keyword matching.

[0117] Through the synonym extraction module, the system can identify words with similar semantics to keywords, and dynamically update the synonym library according to the preset relevance and frequency thresholds to generate an extended keyword data set. Traditional keyword systems have difficulty identifying subtle differences between synonyms, but the present invention solves the problem of semantic misjudgment in synonym expansion through the relevance information of the contextual semantic data set.

[0118] The calculation formula of semantic similarity further enhances the flexibility and accuracy of the system. This formula takes into account multiple factors such as the contextual weight, frequency of occurrence, depth of semantic extension and information density of keywords and synonyms, effectively improving the reliability of synonym matching.

[0119] Through the above innovations, the semantic expansion module realizes dynamic expansion of keywords and response to semantic drift, significantly improving the accuracy of keyword recognition and expansion.

[0120] Specifically, the synonym extraction module can analyze the context of each keyword and identify the changes in its meaning in different contexts. The core of this step is to build a context window for the keyword, that is, a combination of several words before and after, and capture the semantic drift of the keyword in different scenarios by analyzing the semantic association between the keyword and the words in the context.

[0121] In the specific implementation, the context of keywords can be vectorized through natural language processing (NLP) technology. Through word vector models such as Word2Vec, GloVe, etc., the system can convert keywords and context words into points in the vector space, and quantify the semantic relevance between keywords and their context by calculating cosine similarity or Euclidean distance. The size of the context window can be adjusted dynamically to meet the needs of different scenarios for contextual information. For example, in short texts on information platforms, the context window can be set to 5-10 words, while in long news texts, the window range can be expanded to capture more semantic information.

[0122] Specifically, the synonym association module can identify words with similar semantics to the preliminary keywords through the synonym library and include them in the extended keyword data set. This process enables the system to cover more expressions, and even if users use different words or expressions to directly mark the same event, the system can effectively capture it.

[0123] The synonym association module can be combined with existing synonym libraries, such as WordNet, SentiWordNet, etc., to perform semantic expansion. In the specific implementation, the system first performs synonym matching on the preliminary keywords, and includes semantically similar words into the expanded keyword data set. In addition, the context matching algorithm can be introduced to perform secondary verification of the synonym relationship through the semantic information of the context to improve the accuracy of the expansion.

[0124] In a preferred embodiment of the present invention, the keyword propagation origin detection module includes:

[0125] A first appearance detection module, used to determine the origin point of the first appearance of the keyword in multiple information platforms according to the timestamp of each keyword in the extended keyword data set, and generate an initial propagation time record of the keyword, wherein the origin point includes time and information platform;

[0126] The communication origin identification module is used to calculate the user's communication potential based on the user's activity and interaction frequency, and determine the origin identification of the keyword in combination with the initial communication time record of the keyword;

[0127] The communication potential evaluation module is used to evaluate the communication potential of a keyword at the origin point, and the communication potential is calculated by the following formula:

[0128] ,

[0129] in, For keywords potential for dissemination, For keywords User activity at the origin point, For keywords Direct labeling density in the origin point, For keywords The mean semantic distance between the source node and other similar words, For keywords The time difference between the first occurrence and the current time;

[0130] Keywords The direct labeling density in the origin point is calculated by the following formula:

[0131] ,in, is the number of interactions related to the keyword within a specified time. is the number of unique users who participated in direct tagging during the specified time.

[0132] In the embodiment of the present invention, the keyword propagation origin detection module can identify the origin point of the first appearance of the keyword based on the timestamp information in the extended keyword data, and generate the keyword origin identifier in combination with factors such as user activity and interaction frequency. This function is particularly important in information tracking, and can help users accurately determine the original source of hot information, thereby improving the efficiency of information monitoring.

[0133] Through the communication potential evaluation module, the system can quantitatively evaluate the communication potential of keywords at the origin point. The evaluation formula of communication potential combines factors such as user activity, direct tag density, semantic distance and time interval, and can comprehensively evaluate the communication ability of keywords at specific nodes. This innovative function solves the problem of difficulty in quantifying communication potential in existing technologies, and is particularly important for the early identification and tracking of hot information.

[0134] Compared with the existing information tracing technology, the propagation origin detection module of the present invention significantly improves the traceability capability of the keyword propagation path, and provides important data support for subsequent keyword development trend analysis and information prediction.

[0135] In a preferred embodiment of the present invention, the keyword propagation path calculation module includes:

[0136] The path identification module is used to generate a keyword propagation path dataset based on the origin identification of the keyword and by analyzing the frequency of occurrence, timestamps and user behaviors of the keyword in different time periods;

[0137] The propagation path weight calculation module is used to calculate the propagation path weight of keywords in different propagation nodes according to the order of keyword appearance, time interval and user behavior characteristics in the propagation path data set. The propagation path weight is calculated by the following formula:

[0138] ,

[0139] in, For keywords The propagation path weight, For keywords In the The appearance time of the propagation node, For the The user activity of the propagation nodes, For the propagation node The number of hops to the next propagation node or the second forward label interval time, For the The number of interactions among users of each propagation node, is the normalization factor, is the total number of propagation nodes.

[0140] In the embodiment of the present invention, the keyword propagation path calculation module can identify the order of keyword propagation between different information platforms based on the keyword's origin identification and timestamp information, and generate a propagation path data set. By analyzing the frequency and time interval of the keyword on different propagation nodes, the system can quantify the breadth and depth of keyword propagation, providing key data for information monitoring and analysis.

[0141] The propagation path weight calculation module further calculates the weight of the keyword between each propagation node through the propagation path data set. The weight calculation formula comprehensively considers factors such as the appearance time of the keyword, user activity, propagation distance and interaction frequency, and can accurately reflect the propagation effect of the keyword. Compared with the traditional propagation analysis method, the propagation path weight calculation of the present invention not only considers the breadth of propagation, but also introduces the evaluation of propagation depth, making the propagation analysis more comprehensive and accurate.

[0142] By calculating the propagation path weights, this module can provide data support for subsequent heat analysis and trend prediction, and significantly improve the system's ability to identify propagation trends.

[0143] Specifically, the core goal of the path identification module is to track the spread of keywords between different propagation nodes based on their propagation origins. By detecting the frequency and timestamp of keywords at each node, the system can identify the order of keyword propagation in the entire propagation network and the connection relationship between nodes. The result of path identification is a complete propagation path data set, reflecting the keyword propagation chain starting from the origin node.

[0144] In the specific implementation, the path identification module will parse the origin identifier of the keyword and track the subsequent nodes one by one, recording the propagation time, propagation times and propagation methods at each node. This information can help the system construct the keyword propagation path and further analyze its propagation speed, propagation depth and expansion range between nodes. Compared with simple time series, this module not only identifies the time order of propagation, but also captures the importance of propagation nodes and the propagation intensity of keywords at each node.

[0145] Among them, data recording and tracking: the path identification module first records the first appearance time, propagation times and user interaction data of the keyword at each propagation node through a database or real-time data processing system, such as the first positive tag, direct tag, second positive tag, etc. In the application scenario of the information platform, the propagation path of the information can be identified by analyzing the user's second positive tag chain. Each propagation will generate a node record with a timestamp to establish the propagation timeline of the keyword.

[0146] Among them, the propagation chain is constructed: the system hierarchically processes the propagation path of keywords according to the time sequence and relevance of the propagation nodes. The propagation path of keywords can be represented by a graph structure, where nodes represent the propagation position of information on different information platforms or between users, and edges represent the order and direction of propagation. In order to enhance the visualization of the propagation path, the system can attach information such as the propagation intensity and propagation time interval of different nodes to the path map to provide users with a detailed propagation map.

[0147] Propagation order and depth identification: The module uses timestamp data to accurately identify the order and depth of keyword propagation in different nodes. Through time series analysis, the system can classify the propagation time of keywords in different nodes and analyze their propagation speed and diffusion between nodes. For example, the system can identify the retention time of keywords in each node to determine which nodes may be the key sources of information diffusion.

[0148] In actual operation, it is assumed that the keyword is first released on an information platform, and then the second positive tagging is extended to a wider range through multiple information platform users. The path identification module will first record the first appearance time of the keyword on the information platform and the related interactions, and then track the second positive tagging chain of the keyword on different information platforms. For each second positive tagging, the module will record the second positive tagging time, the information of the second positive tagger, and the interaction data of the second positive tagging content, such as the first positive tagging and direct tagging. With this information, the system can construct a complete propagation path from the information platform to the information platform users in the graph structure, and mark the propagation strength and propagation time differences of each node.

[0149] In a preferred embodiment of the present invention, the keyword heat calculation module includes:

[0150] The heat evaluation module is used to calculate the heat score of the keyword based on the propagation path weight in the keyword propagation path data set and the user's interactive behavior. The heat score is calculated using the following formula:

[0151] ,

[0152] in, For keywords The popularity rating of For keywords In the propagation node The path weight on For keywords In the propagation node The number of direct tags by the user, For the propagation node On and Keywords The number of occurrences of related positive sentiment words, For the propagation node On and Keywords The number of occurrences of related negative sentiment words, For keywords In the propagation node The number of first positive tags of the user;

[0153] The trend track generation module is used to generate the heat change trend track of the keyword in different time periods according to the heat score, and the heat change trend track is used to describe the heat growth and decay trend of the keyword.

[0154] In an embodiment of the present invention, the keyword heat calculation module provides a quantitative method for keyword heat scoring based on the propagation path data set and user interaction behavior. The module can comprehensively analyze the user's first positive labeling, direct labeling, second positive labeling and other behaviors, and calculate the propagation heat of each keyword in combination with the propagation path weight. Compared with the prior art, the heat calculation of this system not only takes into account the current user interaction frequency, but also incorporates the propagation path weight into the calculation of the heat score, so that the heat score can more comprehensively reflect the actual propagation effect of the keyword.

[0155] In addition, the system can also generate trend tracks based on the changes in popularity at different nodes to show how keyword popularity changes over different periods of time. This trend track can help analysts quickly identify the life cycle of hot information and the development of its influence, thereby improving the ability to monitor information events in real time.

[0156] Compared with the traditional heat analysis method that simply relies on the frequency of user interactions, the heat scoring model of this system improves the accuracy of heat analysis by integrating multi-dimensional factors, providing users with more in-depth and insightful information analysis support.

[0157] In a preferred embodiment of the present invention, the semantic change tracking module includes:

[0158] The semantic drift detection module is used to analyze the semantic changes of keywords in different time periods and propagation paths based on the propagation path data set, identify the semantic drift of keywords in the multi-node propagation process, and generate semantic drift markers;

[0159] The semantic change trajectory generation module is used to generate the semantic change trajectory of the keyword according to the semantic drift mark of the keyword in the propagation path.

[0160] In an embodiment of the present invention, a semantic change tracking module is used to analyze the semantic changes of keywords in the propagation path and generate a semantic change trajectory of keywords. The module can detect the semantic drift of keywords at different propagation nodes through context analysis. Compared with the problem that the semantic drift of keywords is difficult to effectively monitor in traditional systems, the present invention introduces a semantic change tracking mechanism, which enables the system to track the semantic change trend of keywords in the propagation process in real time.

[0161] The semantic change trajectory can record the semantic evolution of keywords in the propagation path, helping users understand the changes in the meaning of keywords at different propagation nodes. The process of generating this trajectory comprehensively considers multi-dimensional factors such as the contextual relevance, emotional tendency, and contextual changes of keywords, effectively solving the problem of semantic drift that is difficult to capture in existing technologies. Especially in the information monitoring of hot information and sensitive information, the semantics of keywords may gradually change as direct labeling progresses, and the semantic change trajectory of this system can help users discover these changes in a timely manner, thereby more accurately identifying potential information risks.

[0162] The semantic change track can also provide support for information prediction. In practical applications, if the semantic change trend of a keyword is significant, the system can automatically mark the keyword as a possible risk word to remind users to pay attention to the dynamic evolution of information. Compared with traditional information monitoring systems, the semantic change tracking module of the present invention provides users with a more forward-looking information monitoring solution by dynamically capturing the semantic change trend of keywords.

[0163] Specifically, the semantic drift detection module is used to detect the semantic changes of keywords in different time periods and generate semantic drift marks in a quantitative way. This module mainly analyzes the semantic change trajectory of keywords and identifies whether there is a significant semantic shift in the process of keyword dissemination, so as to reflect its semantic stability in different contexts.

[0164] In the specific implementation, the system can use the word vector model to quantify the semantics of keywords in different time periods. By calculating the difference in word vectors for the same keyword in different time periods, the system can quantify the semantic drift of the keyword. For example, the word vectors generated by deep learning models such as BERT or GPT are used to calculate the cosine similarity or Euclidean distance difference of the keywords in each time period. If the semantic drift reaches the set threshold, a corresponding drift mark is generated. In this way, the system can remind users when it detects that a semantic drift has occurred in a keyword, ensuring that users can pay attention to the changes in the meaning of keywords in different contexts in real time.

[0165] Specifically, the semantic change trajectory generation module can quantify the magnitude of semantic drift by analyzing the semantic change trajectory of keywords, and generate a quantitative mark of the drift magnitude, so that the system can further identify the semantic stability of keywords. The purpose of this step is to ensure that semantic drift is not just detected as existing or not, but can specifically indicate the degree of semantic drift, thereby providing more detailed information for information monitoring.

[0166] The calculation of the drift amplitude can be quantified by the similarity or distance between word vectors. For example, the degree of semantic drift of the keyword can be obtained based on the Euclidean distance or cosine similarity between the word vector of each time period and the word vector of the previous time period. The system can also set a semantic drift threshold. When the drift amplitude exceeds a specific range, an event that automatically identifies the drift is triggered. In addition, contextual information can be further introduced into the calculation formula to consider the impact of emotional tendency and user activity on keyword semantics to ensure the accuracy of semantic drift detection.

[0167] In a preferred embodiment of the present invention, the semantic drift detection module includes:

[0168] The drift degree calculation module is used to analyze the semantic drift degree of the semantic shift of the keyword according to the semantic change trajectory and generate a semantic drift identifier. The semantic drift degree is calculated by the following formula:

[0169] ,

[0170] in, For keywords The degree of semantic drift, For keywords In the The semantic distance between the beginning and the end of a time period, ,in, For keywords In time period The word embedding vector The value of the dimension, For keywords In the previous period The word embedding vector The value of the dimension, For keywords In the The number of occurrences in a time period, For keywords In the The direct labeling density of each time period, For keywords In the The time interval of the time period, For keywords In the User activity in a certain time period, is the total number of time periods.

[0171] In an embodiment of the present invention, the semantic drift detection module is used to evaluate the semantic stability and variation of keywords and generate semantic drift identifiers of keywords. By analyzing the semantic variation of keywords in different time periods, the system can accurately identify the degree of semantic drift of keywords and use the semantic drift identifiers for information analysis. This module is particularly important in the current information platform and multi-information platform information environment, because hot words often undergo semantic drift during the dissemination process due to changes in context and user usage.

[0172] The semantic drift indicator combines multiple factors such as the semantic drift value and frequency of occurrence of keywords, and can accurately measure the semantic stability of keywords. In actual applications, if the semantic drift indicator of a keyword is large, the system can automatically mark the keyword as a sensitive word, thereby reminding users to pay special attention to the spread of the keyword. For example, in hot information, the semantics of a keyword may shift in a short period of time due to different direct markings by users. The semantic drift detection module can promptly feedback this change to analysts.

[0173] The semantic drift indicator provides a quantitative indicator for information monitoring and risk prediction. Compared with the existing technology, the present invention not only focuses on the semantic change trend of keywords, but also provides a more quantitative semantic stability assessment method by calculating the degree of semantic drift, which greatly improves the system's ability to identify and respond to complex information events.

[0174] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in the field of data processing technology, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A large language model system based on keyword tracing, characterized in that: The system comprises: A data collection module, used to obtain text data from multiple information platforms, wherein the text data includes text content, timestamp and user behavior information; A text preprocessing module is used to preprocess the text data to obtain a preprocessed text data set, wherein the preprocessing includes word segmentation, denoising and format standardization of the text data; A keyword screening module is used to perform preliminary screening of keywords in the preprocessed text data set according to a preset keyword list to generate a preliminary keyword data set; A semantic expansion module, used to perform semantic expansion processing on the keywords in the preliminary keyword data set through context analysis and synonym association processing to generate an expanded keyword data set; The keyword propagation origin detection module is used to identify the information platform where the keyword first appears based on the timestamp of the keyword in the extended keyword data set, and determine the origin point of the keyword in combination with user behavior information to generate the keyword origin identifier; The keyword propagation path calculation module is used to determine the propagation order of keywords in various information platforms based on the origin identification, occurrence frequency, timestamp and user behavior information of the keywords in the extended keyword data set, and generate a keyword propagation path data set; The keyword heat calculation module is used to calculate the spread heat of each keyword based on the keyword propagation path data set and generate a heat score; The semantic change tracking module is used to analyze the contextual semantics of keywords in different time periods based on the keyword propagation path dataset, identify the semantic drift of the same keyword during the propagation process, and generate a semantic change trajectory dataset; A semantic drift detection module is used to generate a semantic drift identifier of a keyword based on a semantic change trajectory dataset; The information tracing report generation module is used to generate an information tracing report based on the keyword propagation path dataset, keyword heat score and semantic drift identification; The keyword propagation origin detection module includes: A first appearance detection module, used to determine the origin point of the first appearance of the keyword in multiple information platforms according to the timestamp of each keyword in the extended keyword data set, and generate an initial propagation time record of the keyword, wherein the origin point includes time and information platform; The communication origin identification module is used to calculate the user's communication potential based on the user's activity and interaction frequency, and determine the origin identification of the keyword in combination with the initial communication time record of the keyword; The communication potential evaluation module is used to evaluate the communication potential of a keyword at the origin point, and the communication potential is calculated by the following formula: ； in, For keywords potential for dissemination, For keywords User activity at the origin point, For keywords Direct labeling density in the origin point, For keywords The mean semantic distance between the source node and other similar words, For keywords The time difference between the first occurrence and the current time; Keywords The direct labeling density in the origin point is calculated by the following formula: ,in, is the number of interactions related to the keyword within a specified time. is the number of unique users who participated in direct tagging during the specified time; The keyword propagation path calculation module includes: The path identification module is used to generate a keyword propagation path dataset based on the origin identification of the keyword and by analyzing the frequency of occurrence, timestamps and user behaviors of the keyword in different time periods; The propagation path weight calculation module is used to calculate the propagation path weight of keywords in different propagation nodes according to the order of keyword appearance, time interval and user behavior characteristics in the propagation path data set. The propagation path weight is calculated by the following formula: ； in, For keywords The propagation path weight, For keywords In the The appearance time of the propagation node, For the The user activity of the propagation nodes, For the propagation node The number of hops to the next propagation node or the second forward label interval time, For the The number of interactions among users of each propagation node, is the normalization factor, is the total number of propagation nodes.

2. A large language model system based on keyword tracing according to claim 1, characterized in that: The text preprocessing module comprises: The word segmentation processing module is used to perform word segmentation processing on text data, divide continuous text into independent vocabulary units, and generate a word segmentation data set; A denoising processing module is used to remove low-frequency or high-redundancy words, irrelevant symbols and stop words in the word segmentation data set according to the frequency of occurrence of the words in the multi-information platform to generate a text data set; The format standardization module is used to unify the format of the timestamps of the text data set and to perform consistent conversion on the special characters and punctuation marks in the text to obtain preprocessed data.

3. A large language model system based on keyword tracing according to claim 2, characterized in that: The keyword screening module includes: A keyword matching module is used to perform preliminary matching of words in the preprocessed text data set according to a preset keyword list to generate a keyword data set; The keyword updating module is used to adjust the first keyword data set according to the real-time words appearing in the information platform and the corresponding frequency changes to obtain the preliminary keyword data set.

4. A large language model system based on keyword tracing according to claim 3, characterized in that: The semantic extension module includes: The context extraction module is used to extract the combination of several words before and after each keyword in different texts to build a context window for the keyword; The relevance calculation module calculates the relevance between the keyword and other words in its context window, identifies the semantic differences of the keyword in different contexts, and obtains a context semantic data set. The relevance is calculated by the following formula: ； in, For keywords The relevance in the current context, For keywords With the words in the context The semantic weight of For words The frequency of occurrence in the context, For keywords With words The semantic distance of ,in, and Separate words and keywords In the semantic space dimensional word embedding vector representation, is the dimension of the word embedding vector, For keywords With words The adjacent time difference, For words Total number of; The synonym extraction module is used to identify synonyms with similar semantics to the keyword based on the association information in the context semantic data set, and generate a synonym extension data set; The synonym library update module screens the synonyms in the synonym extension data set according to a preset frequency threshold and a preset relevance threshold, and adds them to the preset preliminary synonym library to obtain a final synonym library; The synonym association module is used to calculate the semantic similarity between each synonym and the preliminary keyword based on the final synonym library, and add the synonyms that meet the preset similarity threshold to the preliminary keyword data set to obtain the extended keyword data set. The semantic similarity is calculated by the following formula: ； in, For keywords Synonyms The semantic similarity between For keywords The context weight in the current context, Synonyms The number of occurrences in the context, For keywords Synonyms The semantic distance of Synonyms The context length is the total number of words directly or indirectly connected to it in all texts. Synonyms The semantic extension depth of The number of words directly and indirectly connected to other words in the current text, Synonyms The semantic information density of ,in, is the total number of occurrences in all texts, , Synonyms In the context The probability of occurrence of a keyword, is the total number of keywords, Synonyms User activity on the information platform, ,in, Synonyms The information platform and words The number of first positive markers associated, Synonyms The information platform and words The number of direct markers involved, Synonyms The information platform and words The number of related second positive markings, Synonyms The total number of users on the information platform, Synonyms The time decay coefficient, Used to indicate synonyms The impact of usage frequency changes over time.

5. The large language model system based on keyword tracing according to claim 4 is characterized in that: The keyword heat calculation module includes: The heat evaluation module is used to calculate the heat score of the keyword based on the propagation path weight in the keyword propagation path data set and the user's interactive behavior. The heat score is calculated using the following formula: ； in, For keywords The popularity rating of For keywords In the propagation node The path weight on For keywords In the propagation node The number of direct tags by the user, For the propagation node On and Keywords The number of occurrences of related positive sentiment words, For the propagation node On and Keywords The number of occurrences of related negative sentiment words, For keywords In the propagation node The number of first positive tags of the user; The trend track generation module is used to generate the heat change trend track of the keyword in different time periods according to the heat score, and the heat change trend track is used to describe the heat growth and decay trend of the keyword.

6. A large language model system based on keyword tracing according to claim 5, characterized in that: The semantic change tracking module includes: The semantic drift detection module is used to analyze the semantic changes of keywords in different time periods and propagation paths based on the propagation path data set, identify the semantic drift of keywords in the multi-node propagation process, and generate semantic drift markers; The semantic change trajectory generation module is used to generate the semantic change trajectory of the keyword according to the semantic drift mark of the keyword in the propagation path.

7. The large language model system based on keyword tracing according to claim 6, characterized in that: The semantic drift detection module includes: The drift degree calculation module is used to analyze the semantic drift degree of the semantic shift of the keyword according to the semantic change trajectory and generate a semantic drift identifier. The semantic drift degree is calculated by the following formula: ； in, For keywords The degree of semantic drift, For keywords In the The semantic distance between the beginning and the end of a time period, ,in, For keywords In time period The word embedding vector The value of the dimension, For keywords In the previous period The word embedding vector The value of the dimension, For keywords In the The number of occurrences in a time period, For keywords In the The direct labeling density of each time period, For keywords In the The time interval of the time period, For keywords In the User activity in a certain time period, is the total number of time periods.

Images