Method and system for generating intelligent insight report based on AI large model

Abstract

The invention relates to the field of intelligent report generation, in particular to an intelligent insight report generation method and system based on an AI large model, and the method comprises the steps: inputting an insight demand, and generating an insight data package comprising insight contents, associated data and industry labels; extracting a basic keyword set of the insight data packet to form a mixed feature code; after mixed feature coding preprocessing, weight distribution is carried out; and after weight distribution of the AI large model, injecting a high-weight feature vector into a semantic understanding core layer, injecting a low-weight feature vector into a logical reasoning layer, and outputting analysis data to form an intelligent insight report. According to the method, word embedding parameters are optimized according to field semantic characteristics, a parameter verification mechanism is introduced, field text characteristics are adapted by adjusting vector dimensions, context windows and low-frequency vocabulary filtering threshold values, window parameter validity is verified through cosine similarity, word frequency threshold value reasonability is verified through standard deviation, and field text characteristic matching is achieved. And it is ensured that the feature vectors can accurately capture domain-specific semantics.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent report generation, in particular to a method and system for generating intelligent insight reports based on AI large models. BACKGROUND

[0002] With the development of artificial intelligence technology, AI large models are gradually widely used in the field of intelligent report generation, and the core requirement is to convert the natural language requirements of users into structured and professional insight reports.

[0003] However, the existing technology still has many shortcomings in practical application, and it is difficult to meet the requirements of users in different industries and different scenarios for report accuracy, adaptability and logical integrity. The existing method mainly stays at the level of standardizing and removing redundancy, keyword extraction, and does not effectively identify the fuzziness of the requirements itself. And the existing technology integrates domain knowledge in the mode of direct calling of general knowledge base or binding of fixed knowledge graph, which lacks modular and reusable design. When adapting to different industries, it is necessary to rebuild the knowledge system and adjust the analysis rules, which is not only cumbersome and time-consuming.

[0004] And in the process of report reasoning, a single model mode is often used, without hierarchical processing according to the importance of feature vectors. The core conclusion and supplementary reasoning basis are mixed together and output, which not only makes it difficult for users to quickly grasp the report highlights, but also lacks targeted verification of secondary features, making the logic of supplementary basis insufficient, and the report as a whole is not prominent and insufficiently supported, which cannot meet the demand for insight depth in professional scenarios. SUMMARY

[0005] The present application provides a method and system for generating intelligent insight reports based on AI large models to solve the technical problems in the prior art.

[0006] The technical solution of the present application to solve the above technical problems is as follows: a method for generating intelligent insight reports based on AI large models, the method comprising: S101, inputting insight requirements, and generating an insight data package including insight content, associated data and industry labels through meta-linguistic analysis including fuzziness evaluation, meta-linguistic splitting and preliminary scene association; S102, extracting a set of basic keywords of the insight data package, matching and loading a domain cognitive plug-in, and forming a hybrid feature code based on the corresponding extracted domain knowledge graph, analysis logic rules and insight dimension templates in the data package; S103, after the mixed feature code preprocessing, input to the AI large model, and introduce the word embedding mechanism, vector coding is carried out on the mixed feature code, and is converted into an initial field feature vector, and the attention strategy is used as a query vector, the semantic similarity of the initial field feature vector is calculated, and weight distribution is carried out; S104, after the weight distribution of the AI large model, the high weight feature vector is injected into the semantic understanding core layer, and the low weight feature vector is injected into the logical reasoning layer, and the output analysis data is formed into an intelligent insight report.

[0007] In a preferred embodiment, in S101, the input insight demand is analyzed by including fuzzy degree evaluation, meta language parsing and preliminary scene association, including: the user inputs the insight demand in natural language, such as analyzing the core reasons why the sales of new beauty products in this quarter do not meet expectations, and the natural language is standardized, redundant mood words are removed, and grammatical errors are corrected. The standardized processing can be based on the current mature NLP, and the output structured demand text is obtained; The industry key words in the structured demand text are matched to determine the industry to which they belong, and the preliminary scene association industry label is generated; The entity elements in the structured demand text are extracted by NER template, such as category, time, etc., the industry entity element template is called, and the template is pre-stored in the system, and the source of the template can be realized by accessing external network from third party or AI large model, the fuzzy degree level is determined according to the number of missing entity elements in the entity element template, the fuzzy degree level can be classified as low, medium and high in this application, which respectively correspond to no missing entity elements, missing one kind of entity elements, and missing two kinds of entity elements and above, and the fuzzy degree evaluation is completed; Based on the fuzzy degree evaluation, the meta language splitting is carried out, including: the industry entity element template prompts the user to supplement the entity elements in the repeated extraction structured demand text to be less than or equal to one kind of entity elements, until the fuzzy degree evaluation is low or medium, the structured demand text is split into at least two split meta languages including core target meta language, data range meta language and awareness depth meta language, wherein the core target meta language is the core purpose of clear insight analysis, the data range meta language is the scope of the required data for analysis, and the awareness depth meta language is the dimension covered by the analysis, and the awareness depth meta language is the style and output requirement of the clear report; Based on the BERT model, the split meta-words and industry labels are taken as inputs, the split meta-words are expanded based on semantic correlation in the industry, and if the split meta-words contain data range meta-words and awareness depth meta-words, the Pearson correlation coefficient is used to calculate the index correlation, the confounding variables are excluded by matching the causal correlation, and it is ensured that the expanded insights have data support. In the currently commonly used BERT model, the correlation algorithm used involves the fusion of semantic correlation, data correlation and causal correlation, which corresponds to the split meta-words one by one. Therefore, the BERT is used for semantic mining in the present application to generate insight data packets including insight content, associated data and industry labels. The insight content is generated by at least one of the awareness depth meta-words and the core target meta-words and is expanded by semantic correlation. The associated data is generated by the data range meta-words. The industry label is generated by the industry keywords in the structured text.

[0008] In a preferred embodiment, after obtaining the insight data packet, the S102 integrates the insight content text, the associated data text and the industry label text into a unified extraction text by the TF-IDF algorithm, performs coding standardization, calls an open-source word segmentation tool to perform word segmentation on the extraction text, outputs a word sequence, removes non-semantically valuable words based on a general stop word table, removes low-frequency words with an occurrence frequency less than 2, calculates the TF value of each word in the extraction text, i.e. the number of occurrences of a word in the extraction text divided by the total number of words in the extraction text, sorts the TF values in descending order to form a basic keyword set, and calls an index engine of a plug-in library to perform similarity matching between the basic keyword set and the domain label set of each plug-in in the plug-in library. If the similarity score exceeds a similarity threshold, the plug-in with the maximum similarity score is taken as a domain cognitive plug-in. If the similarity scores of all the plug-ins do not exceed the similarity threshold, the BERT is re-associated in S101, similarity matching is performed in combination with the core target meta-words, a domain cognitive plug-in is loaded, the domain knowledge graph, the analysis logic rule and the insight dimension template in the domain cognitive plug-in are extracted and integrated in format to generate mixed feature coding including the domain knowledge graph structure data, the analysis logic rule text and the insight dimension label. The similarity matching between the basic keyword set and the domain label set of each plug-in in the plug-in library is as follows: The domain basic keyword set and the domain label set are traversed to extract words completely consistent in the domain feature keyword set and the domain label set to form an intersection set. The feature keyword set and the domain label set are merged to form a union set after removing duplicate words. Based on the Jaccard similarity coefficient formula, the similarity score is obtained by taking the ratio of the number of elements in the intersection set to the number of elements in the union set, and the domain cognitive plug-in corresponding to the domain label set with the highest similarity score is loaded.

[0009] In a preferred embodiment, the method further includes, before invoking the indexing engine of the plugin library: A general ontology is constructed using an ontology editor. The general ontology contains entity elements, relations, and attributes. The entity element and relation data are imported from a public industry knowledge base and transformed into triples that include entity elements, relations, and attributes. The data is stored using the open-source graph database Neo4j. Cypher statements are used to implement the addition, deletion, modification, and query of knowledge, ensuring efficient querying of knowledge associations and constructing a domain knowledge graph. By mining the text of historical insight reports, the rules in the historical insight reports are transformed into parsable formula text and linked with entity elements in the domain knowledge graph to form analytical logic rules that include calculation objectives, variables, operators and correction factors. That is, the standardized paradigm of conditions, calculation and output is used as the rules and defined by JSONSchema, including rule ID, triggering conditions, calculation functions, output format, confidence threshold, etc. Using the needs of each entity element as a dimension, and associating each dimension with the entity elements and analysis logic rules of the domain knowledge graph, an insight dimension template containing tags and descriptive text is formed, thus completing the construction of the plugin library.

[0010] Furthermore, in this application, the needs of each entity element are used as a dimension. In some industries, such as channel coverage efficiency, user profile matching degree, product core selling point satisfaction, and competitor benchmarking gap, there are also needs corresponding to each entity element in other industries. The analysis logic rules are to transform the rules into parsable formula text and quantify the insight needs issued by different users.

[0011] In a preferred embodiment, after obtaining the hybrid feature encoding, step S103 inputs it into the large AI model based on the Transformer architecture. The format conversion is completed through preprocessing, including: converting the triples of the domain knowledge graph into triple text to ensure the semantic integrity of entity elements and relationships; converting the formula text of the analysis logic rules into standard formula text; normalizing the tags and descriptive text words of the insight dimension template; and finally forming a unified text sequence consisting of triple text, standardized formula text, and normalized tag text. The entity elements of the domain knowledge graph and the variables of the analysis logic rules are all target words, thus completing the format conversion.

[0012] The triples in the domain knowledge graph are triple texts. Specifically, the structured data of the entity element relationship of the triple, such as the first entity element - relation - attribute - second entity element, is transformed into plain text. The formula text of the analysis logic rules is transformed into standard formula text. For the formula-like features in the analysis logic rules, they are transformed into standardized formula text obtained by the calculation target from the first variable, the operator, the second variable, and the correction factor, ensuring that the logical semantics can be parsed.

[0013] In a preferred embodiment, the hybrid feature encoding is preprocessed and then input into a large AI model. A word embedding mechanism is introduced to perform vector encoding, transforming discrete text semantic units into continuous low-dimensional vectors, including: Choose a pre-trained word embedding tool that is from the same source as Transformer, such as the word embedding module of oBERTa. The word embedding results of this type of tool have the highest compatibility with the self-attention layer and feedforward neural network layer of Transformer. It can directly pass vector features without additional format conversion. Download the matching pre-trained weights from the weight library of the pre-trained word embedding tool and load them into the parameter matrix of the word embedding mechanism. Enable parameter modification in the word embedding tool to complete the configuration of custom parameters; Match the words in the unified text sequence with the words embedded in the pre-trained words. If there are mismatched words, initialize a random low-dimensional vector as the initial embedding vector. Based on optimization parameters for general scenarios, parameter configuration is performed. In the core parameters of the word embedding tool, the vector dimension is set to 300 dimensions. 300 dimensions balance semantic representation accuracy and model computational efficiency, a mainstream configuration in the NLP field. The window parameter is set in the word embedding tool, with the window size set to cover the contextual information of two words before and after the current word in the unified text sequence, used to balance local semantics and global associations. The number of training iterations is set to converge to a stable state. If the current word is at the beginning of the unified text sequence with no words to its left and / or at the end of the unified text sequence with no words to its right, then a filler word is generated. The word embedding mechanism performs word embedding on each word within the window. After generating the initial vector, the semantics of the window are fused into the vector of the current word through a local attention algorithm. For example, the vectors of words within the window are summed by position weights, with the weight increasing the closer the position is to the current word. The minimum word frequency threshold is set to 3 to filter out rare words with low frequency of occurrence, simplifying the vector space. In S102, low-frequency words with a frequency of less than 2 in the insight data package have been removed. When the number of occurrences of a word in a unified text sequence is less than 3, a random low-dimensional vector is uniformly assigned. When the number of occurrences of a word in a unified text sequence is greater than or equal to 3, a dedicated 300-dimensional vector is generated during encoding based on the semantics of the word itself and the context window to obtain the initial domain feature vector.

[0014] In a preferred embodiment, after the parameters are configured, a rationality verification is performed, including: Select the initial vectors of the same word in different windows and calculate the cosine similarity between them. If the cosine similarity is less than 0.5, it indicates that the window parameter configuration is effective. Then, the standard deviation of all initial domain feature vectors is calculated. If the standard deviation is greater than 0.15, it indicates that the minimum word frequency threshold is set reasonably.

[0015] In a preferred embodiment, after obtaining the initial domain feature vector, the initial domain feature vector transformed from the core target metaword in the insight data packet is given the first priority, the initial domain feature vector transformed from the data range metaword in the insight data packet is given the second priority, and the initial domain feature vector transformed from the industry tag in the insight data packet is given the third priority. Weights are assigned to the initial domain feature vectors of the first, second, and third priorities. In some other specific embodiments, the weights are 0.5, 0.3, and 0.1 to form a fused query vector. The fused query vector is then normalized so that the query vector modulus is 1. A two-dimensional matrix is ​​created to initialize the similarity score matrix. The row index is set to the query vector, the column index is set to all initial domain feature vectors, and the matrix elements are the cosine similarity scores between the query vector and each initialized initial domain feature vector, representing the semantic similarity with the initial domain feature vector. The cosine similarity scores are projected to the [0,1] interval through dot product calculation and written into the two-dimensional matrix. The core logic of weight allocation is that the higher the semantic similarity, the greater the contribution of the vector to the generation of the insight report, and the higher the weight. Therefore, the cosine similarity scores are differentiated and weighted by the mathematical calculation of the Softmax function, including numerator calculation, denominator calculation, and basic weight calculation. This completes the weight allocation and generates high-weight vector lists and low-weight vector lists, which are then passed to the corresponding levels of the AI ​​large model.

[0016] This invention also provides a system for generating intelligent insight reports based on AI large models, the system comprising: The insight parsing and resource matching module receives natural language insight requests from users, generates insight data packages through standardization processing, fuzziness evaluation, metaword splitting, and semantic association extension of the BERT model, and builds a domain knowledge graph based on the ontology editor, forming a plugin library by analyzing logic rules and insight dimension templates. It extracts basic keyword sets from the insight data packages, matches them with Jaccard similarity coefficients and loads corresponding domain cognitive plugins, and finally integrates them to generate hybrid feature codes. The preprocessing and vector transformation module is used to optimize the data format and transform the hybrid feature encoding into vectorized data. It converts triples in the hybrid feature encoding into triple text, formula text into standard formula text, and normalizes label and descriptive text to form a unified text sequence. It selects a pre-trained word embedding tool that is the same as Transformer, assigns weights according to priority, and fuses them to generate a normalized query vector.

[0017] The attention weight allocation and hierarchical reasoning module is used to achieve differentiated allocation of vector weights through attention strategies, create a similarity score matrix, calculate the cosine similarity between the query vector and the initial domain feature vector, generate a list of high-weight vectors and a list of low-weight vectors through the Softmax function, input the high-weight vectors into the semantic understanding core layer, call the Transformer self-attention sub-layer to generate core insight conclusions, input the low-weight vectors into the post-logic reasoning layer, and call the logic verification module to generate supplementary reasoning basis.

[0018] The integration and generation module is used to integrate the results of hierarchical reasoning to form a complete and structured intelligent insight report. It summarizes the core insight conclusions output by the semantic understanding core layer and the supplementary reasoning basis output by the subsequent logical reasoning layer, sorts out and analyzes the logical relationships and priorities of the data, organizes the content according to a standardized format, and finally generates a complete intelligent insight report.

[0019] The beneficial effects of this invention are: optimizing word embedding parameters for domain semantic characteristics and introducing a parameter verification mechanism. By adjusting the vector dimension, context window, and low-frequency word filtering threshold, the invention adapts to the characteristics of domain text. At the same time, it verifies the validity of window parameters through cosine similarity and the rationality of word frequency threshold through standard deviation, ensuring that the feature vector can accurately capture domain-specific semantics, providing a high-quality data foundation for subsequent reasoning and reducing reasoning errors caused by semantic representation bias. A hierarchical reasoning mechanism is adopted, which assigns weights to feature vectors according to the priority of metawords. High-weight vectors are injected into the semantic understanding core layer to generate core insight conclusions, while low-weight vectors are injected into the logical reasoning layer to generate supplementary reasoning basis. This allows users to quickly grasp the key points of the report, and the logical verification of secondary features provides supplementary supporting basis, meeting the dual needs of depth and completeness of insight in professional scenarios. Employing a modular design with domain-awareness plugins, each plugin integrates the corresponding domain's knowledge graph, analytical logic rules, and insight dimension templates. Jaccard similarity coefficients are used for rapid plugin matching and adaptation. When switching industries or scenarios, simply loading the corresponding plugin completes the adaptation of domain knowledge and rules, eliminating the need for system restructuring or rebuilding the knowledge system. This significantly reduces the operational costs of cross-domain applications and improves report generation efficiency and industry adaptability. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0023] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0024] This embodiment provides a method for generating intelligent insight reports based on AI large models, the method including: S101. Input insight requirements and generate an insight data package including insight content, related data and industry tags through meta-language parsing that includes fuzziness assessment, meta-language decomposition and preliminary scenario association. In S101, the user inputs insight requirements. Through the meta-language parsing, which includes fuzziness assessment, meta-language decomposition, and preliminary scenario association, the user inputs insight requirements in natural language, such as analyzing the core reasons why the sales of new beauty products in this quarter did not meet expectations. The natural language is standardized to remove redundant interjections and correct grammatical errors. The standardization process is based on the relatively mature NLP. The output is a structured requirement text. The industry is determined by matching industry keywords in the structured requirements text, and preliminary industry tags related to the scenario are generated. Entity elements such as category and time are extracted from the structured requirement text by NER template. Industry entity element templates are called and pre-stored in the system. The source of the templates can be achieved by third parties or AI large models accessing external networks. The ambiguity level is determined according to the number of missing entity elements in the entity element template. The ambiguity level can be classified into low, medium and high in this application, corresponding to no missing entity elements, missing one type of entity element, and missing two or more types of entity elements, respectively, to complete the ambiguity assessment. Based on fuzziness assessment, the metaphrase decomposition is performed, including: prompting users to supplement the structured requirement text with an industry entity element template until the entity elements extracted repeatedly are less than or equal to one type of entity element, until the fuzziness assessment is low or medium level, and decomposing the structured requirement text into decomposition metaphrase containing at least two of the following: core objective metaphrase, data scope metaphrase, and perception depth metaphrase. The core objective metaphrase clarifies the core purpose of the insight analysis, the data scope metaphrase defines the scope of data required for the analysis, the perception depth metaphrase limits the dimensions that the analysis needs to cover, and the perception depth metaphrase clarifies the style and output requirements of the report. Based on the BERT model, this application takes split metawords and industry tags as input, extends the semantic association of the split metawords based on the industry, and if the split metawords contain data range metawords and awareness depth metawords, the Pearson correlation coefficient is used to calculate the correlation of the indicators. Causal association matching excludes confounding variables to ensure that the extended insights are supported by data. In the currently commonly used BERT models, the association algorithm used involves the integration of semantic association, data association and causal association, which correspond one by one with the split metawords. Therefore, this application uses BERT for semantic mining to generate an insight data package including insight content, associated data and industry tags. The insight content is generated by at least one of awareness depth metawords and core target metawords and is generated by semantic association extension. The associated data is generated by data range metawords, and the industry tags are generated by industry keywords in structured text.

[0025] S102. Extract the basic keyword set of the insight data package, match and load the domain cognition plugin, and extract the domain knowledge graph, analysis logic rules and insight dimension templates corresponding to the data package to form a hybrid feature code. After acquiring the insight data package, S102 integrates the insight content text, related data text, and industry tag text into a unified text to be extracted using the TF-IDF algorithm. It then performs encoding standardization, calls an open-source word segmentation tool to segment the text to be extracted, outputs a word sequence, removes words with no semantic value based on a general stop word list, and removes low-frequency words with a frequency of less than 2. It calculates the TF value of each word in the text to be extracted, which is the number of times a word appears in the text to be extracted divided by the total number of words in the text to be extracted. The words are then sorted in descending order by TF value to form a basic keyword set. The indexing engine of the plugin library is invoked to perform similarity matching between the basic keyword set and the domain tag set of each plugin in the plugin library. If the similarity exceeds the threshold, the plugin with the highest similarity is taken as the domain cognition plugin. If none of them exceed the similarity threshold, the process returns to S101 for BERT re-association and similarity matching is performed in conjunction with the core target metaword. The domain cognition plugin is loaded, and the domain knowledge graph, analysis logic rules, and insight dimension templates within the domain cognition plugin are extracted and formatted to generate a hybrid feature code containing domain knowledge graph structure data, analysis logic rule text, and insight dimension tags. The basic keyword set is matched with the domain tag sets of each plugin in the plugin library based on similarity. Specifically: Traverse the domain basic keyword set and domain tag set, extract words that are completely consistent with the domain feature word set and the domain tag set, and form the intersection set; Merge the feature word set and the domain label set, and remove duplicate words to form the union set; Based on the Jaccard similarity coefficient formula, the similarity score is obtained by dividing the similarity score by the ratio of the number of elements in the intersection set to the number of elements in the union set, and the domain cognition plugin corresponding to the domain tag set with the highest similarity score is loaded.

[0026] Before invoking the plugin library's indexing engine, the method also includes: A general ontology is constructed using an ontology editor. The general ontology contains entity elements, relations, and attributes. The entity element and relation data are imported from a public industry knowledge base and transformed into triples that include entity elements, relations, and attributes. The data is stored using the open-source graph database Neo4j. Cypher statements are used to implement the addition, deletion, modification, and query of knowledge, ensuring efficient querying of knowledge associations and constructing a domain knowledge graph. By mining the text of historical insight reports, the rules in the historical insight reports are transformed into parsable formula text and linked with entity elements in the domain knowledge graph to form analytical logic rules that include calculation objectives, variables, operators and correction factors. That is, the standardized paradigm of conditions, calculation and output is used as the rules and defined by JSONSchema, including rule ID, triggering conditions, calculation functions, output format, confidence threshold, etc. Using the needs of each entity element as a dimension, and associating each dimension with the entity elements and analysis logic rules of the domain knowledge graph, an insight dimension template containing tags and descriptive text is formed, thus completing the construction of the plugin library.

[0027] Furthermore, in this application, the needs of each entity element are used as a dimension. In some industries, such as channel coverage efficiency, user profile matching degree, product core selling point satisfaction, and competitor benchmarking gap, there are also needs corresponding to each entity element in other industries. The analysis logic rules are to transform the rules into parsable formula text and quantify the insight needs issued by different users.

[0028] S103. After preprocessing the hybrid feature encoding, it is input into the AI ​​large model, and a word embedding mechanism is introduced to perform vector encoding on the hybrid feature encoding, which is transformed into an initial domain feature vector. Then, the attention strategy uses the initial domain feature vector as the query vector to calculate the semantic similarity with the initial domain feature vector and perform weight allocation. After obtaining the hybrid feature encoding, S103 inputs it into the large AI model based on the Transformer architecture. The preprocessing completes the format conversion, including: converting the triples of the domain knowledge graph into triple text to ensure the semantic integrity of entity elements and relationships; converting the formula text of the analysis logic rules into standard formula text; and normalizing the tags and descriptive text words of the insight dimension template. Finally, a unified text sequence is formed, consisting of triple text, standardized formula text, and normalized tag text. The entity elements of the domain knowledge graph and the variables of the analysis logic rules are all target words, thus completing the format conversion.

[0029] The triples in the domain knowledge graph are triple texts. Specifically, the structured data of the entity element relationship of the triple, such as the first entity element - relation - attribute - second entity element, is transformed into plain text. The formula text of the analysis logic rules is transformed into standard formula text. For the formula-like features in the analysis logic rules, they are transformed into standardized formula text obtained by the calculation target from the first variable, the operator, the second variable, and the correction factor, ensuring that the logical semantics can be parsed.

[0030] After preprocessing the hybrid feature encoding, it is input into the large AI model, where a word embedding mechanism is introduced for vector encoding. This transforms discrete text semantic units into continuous low-dimensional vectors, including: Choose a pre-trained word embedding tool that is from the same source as Transformer, such as the word embedding module of oBERTa. The word embedding results of this type of tool have the highest compatibility with the self-attention layer and feedforward neural network layer of Transformer. It can directly pass vector features without additional format conversion. Download the matching pre-trained weights from the weight library of the pre-trained word embedding tool and load them into the parameter matrix of the word embedding mechanism. Enable parameter modification in the word embedding tool to complete the configuration of custom parameters; Match the words in the unified text sequence with the words embedded in the pre-trained words. If there are mismatched words, initialize a random low-dimensional vector as the initial embedding vector. Based on optimization parameters for general scenarios, parameter configuration is performed. In the core parameters of the word embedding tool, the vector dimension is set to 300 dimensions. 300 dimensions balance semantic representation accuracy and model computational efficiency, a mainstream configuration in the NLP field. The window parameter is set in the word embedding tool, with the window size set to cover the contextual information of two words before and after the current word in the unified text sequence, used to balance local semantics and global associations. The number of training iterations is set to converge to a stable state. If the current word is at the beginning of the unified text sequence with no words to its left and / or at the end of the unified text sequence with no words to its right, then a filler word is generated. The word embedding mechanism performs word embedding on each word within the window. After generating the initial vector, the semantics of the window are fused into the vector of the current word through a local attention algorithm. For example, the vectors of words within the window are summed by position weights, with the weight increasing the closer the position is to the current word. The minimum word frequency threshold is set to 3 to filter out rare words with low frequency of occurrence, simplifying the vector space. In S102, low-frequency words with a frequency of less than 2 in the insight data package have been removed. When the number of occurrences of a word in a unified text sequence is less than 3, a random low-dimensional vector is uniformly assigned. When the number of occurrences of a word in a unified text sequence is greater than or equal to 3, a dedicated 300-dimensional vector is generated during encoding based on the semantics of the word itself and the context window to obtain the initial domain feature vector.

[0031] After the parameters are configured, a validity check is performed, including: Select the initial vectors of the same word in different windows and calculate the cosine similarity between them. If the cosine similarity is less than 0.5, it indicates that the window parameter configuration is effective. Then, the standard deviation of all initial domain feature vectors is calculated. If the standard deviation is greater than 0.15, it indicates that the minimum word frequency threshold is set reasonably.

[0032] After weight allocation, the S104 AI big model injects high-weight feature vectors into the semantic understanding core layer and low-weight feature vectors into the logical reasoning layer, outputting analytical data to form an intelligent insight report.

[0033] After obtaining the initial domain feature vector, the initial domain feature vector transformed from the core target metaword in the insight data packet is given the first priority, the initial domain feature vector transformed from the data range metaword in the insight data packet is given the second priority, and the initial domain feature vector transformed from the industry tag in the insight data packet is given the third priority. Weights are assigned to the initial domain feature vectors of the first, second, and third priorities. In some other specific implementations, the weights are 0.5, 0.3, and 0.1 to form a fused query vector. The fused query vector is then normalized so that the query vector modulus is 1. A two-dimensional matrix is ​​created to initialize the similarity score matrix. The row index is set to the query vector, the column index is set to all initial domain feature vectors, and the matrix elements are the cosine similarity scores between the query vector and each initialized initial domain feature vector, representing the semantic similarity with the initial domain feature vector. The cosine similarity scores are projected to the [0,1] interval through dot product calculation and written into the two-dimensional matrix. The core logic of weight allocation is that the higher the semantic similarity, the greater the contribution of the vector to the generation of the insight report, and the higher the weight. Therefore, the cosine similarity scores are differentiated and weighted by the mathematical calculation of the Softmax function, including numerator calculation, denominator calculation, and basic weight calculation. This completes the weight allocation and generates high-weight vector lists and low-weight vector lists, which are then passed to the corresponding levels of the AI ​​large model.

[0034] After weight allocation, S104 injects a list of high-weight vectors into the semantic understanding core layer. The core layer calls the Transformer's self-attention sublayer to perform deep semantic parsing on the high-weight vectors and generate core insight conclusions. After injecting a list of low-weight vectors into the logic reasoning layer, the reasoning layer calls the logic verification module to perform supplementary analysis on the low-weight vectors and generate supplementary reasoning basis.

[0035] This invention also provides a system for generating intelligent insight reports based on AI large models, the system comprising: The insight parsing and resource matching module receives natural language insight requests from users, generates insight data packages through standardization processing, fuzziness evaluation, metaword splitting, and semantic association extension of the BERT model, and builds a domain knowledge graph based on the ontology editor, forming a plugin library by analyzing logic rules and insight dimension templates. It extracts basic keyword sets from the insight data packages, matches them with Jaccard similarity coefficients and loads corresponding domain cognitive plugins, and finally integrates them to generate hybrid feature codes. The preprocessing and vector transformation module is used to optimize the data format and transform the hybrid feature encoding into vectorized data. It converts triples in the hybrid feature encoding into triple text, formula text into standard formula text, and normalizes label and descriptive text to form a unified text sequence. It selects a pre-trained word embedding tool that is the same as Transformer, assigns weights according to priority, and fuses them to generate a normalized query vector.

[0036] The attention weight allocation and hierarchical reasoning module is used to achieve differentiated allocation of vector weights through attention strategies, create a similarity score matrix, calculate the cosine similarity between the query vector and the initial domain feature vector, generate a list of high-weight vectors and a list of low-weight vectors through the Softmax function, input the high-weight vectors into the semantic understanding core layer, call the Transformer self-attention sub-layer to generate core insight conclusions, input the low-weight vectors into the post-logic reasoning layer, and call the logic verification module to generate supplementary reasoning basis.

[0037] The integration and generation module is used to integrate the results of hierarchical reasoning to form a complete and structured intelligent insight report. It summarizes the core insight conclusions output by the semantic understanding core layer and the supplementary reasoning basis output by the subsequent logical reasoning layer, sorts out and analyzes the logical relationships and priorities of the data, organizes the content according to a standardized format, and finally generates a complete intelligent insight report.

[0038] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0039] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0040] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxesFigure 1 A device that provides the functions specified in one or more boxes.

[0041] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0042] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0043] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0044] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for generating intelligent insight reports based on AI large-scale models, characterized in that, The method includes: S101. Input insight requirements and generate an insight data package including insight content, related data and industry tags through meta-language parsing that includes fuzziness assessment, meta-language decomposition and preliminary scenario association. S102. Extract the basic keyword set of the insight data package, match and load the domain cognition plugin, and extract the domain knowledge graph, analysis logic rules and insight dimension templates corresponding to the data package to form a hybrid feature code. S103. After preprocessing the hybrid feature encoding, it is input into the AI ​​large model, and a word embedding mechanism is introduced to perform vector encoding on the hybrid feature encoding, which is transformed into an initial domain feature vector. Then, the attention strategy uses the initial domain feature vector as the query vector to calculate the semantic similarity with the initial domain feature vector and perform weight allocation. After weight allocation, the S104 AI big model injects high-weight feature vectors into the semantic understanding core layer and low-weight feature vectors into the logical reasoning layer, outputting analytical data to form an intelligent insight report.

2. The method for generating intelligent insight reports based on AI large models according to claim 1, characterized in that, In S101, the input insight requirements are processed through a meta-language parsing process that includes fuzziness assessment, meta-language decomposition, and preliminary scene association. This process includes: the user inputs the insight requirements in natural language, the natural language is standardized, and a structured requirement text is output. The industry is determined by matching industry keywords in the structured requirements text, and preliminary industry tags related to the scenario are generated. Extract entity elements from the structured requirements text, call the industry entity element template, determine the ambiguity level according to the number of missing entity elements in the entity element template, and complete the ambiguity assessment. Based on fuzziness assessment, the metaphrase splitting is performed, including: prompting users with industry entity element templates to supplement the entity elements in the repeatedly extracted structured requirement text to be less than or equal to one type of entity element, and splitting the structured requirement text into split metaphrase that contains at least two of the following: core target metaphrase, data range metaphrase, and perception depth metaphrase. Based on the BERT model, using split metawords and industry tags as input, the split metawords are semantically expanded based on industry to generate an insight data package that includes insight content, related data, and industry tags.

3. The method for generating intelligent insight reports based on AI large models according to claim 1, characterized in that, After acquiring the insight data package, S102 integrates the insight content text, related data text, and industry tag text into a unified text to be extracted using the TF-IDF algorithm, and performs encoding standardization. It then calls an open-source word segmentation tool to segment the text to be extracted, outputs a word sequence, removes low-frequency words with a frequency of less than 2, calculates the TF value of each word in the text to be extracted, and sorts them in descending order by TF value to form a basic keyword set. The indexing engine of the plugin library is called to perform similarity matching between the basic keyword set and the domain tag set of each plugin in the plugin library. The domain cognition plugin is loaded, and the domain knowledge graph, analysis logic rules and insight dimension templates in the domain cognition plugin are extracted and formatted to generate a hybrid feature code containing domain knowledge graph structure data, analysis logic rule text and insight dimension tags. The basic keyword set is matched with the domain tag sets of each plugin in the plugin library based on similarity. Specifically: Traverse the domain basic keyword set and domain tag set, extract words that are completely consistent with the domain feature word set and the domain tag set, and form the intersection set; Merge the feature word set and the domain label set, and remove duplicate words to form the union set; Based on the Jaccard similarity coefficient formula, the similarity score is obtained by dividing the similarity score by the ratio of the number of elements in the intersection set to the number of elements in the union set, and the domain cognition plugin corresponding to the domain tag set with the highest similarity score is loaded.

4. The method for generating intelligent insight reports based on AI large models according to claim 3, characterized in that, The method further includes the following steps before invoking the plugin library's indexing engine: A general ontology is constructed using an ontology editor. The general ontology contains entity elements, relations, and attributes. The entity element and relation data are imported from a public industry knowledge base and transformed into triples that include entity elements, relations, and attributes, which are then used to construct a domain knowledge graph. By mining the text of historical insight reports, the rules in the historical insight reports are transformed into parsable formula text and linked with entity elements in the domain knowledge graph to form analytical logic rules that include calculation objectives, variables, operators and correction factors. Using the needs of each entity element as a dimension, and associating each dimension with the entity elements and analysis logic rules of the domain knowledge graph, an insight dimension template containing tags and descriptive text is formed, thus completing the construction of the plugin library.

5. The method for generating intelligent insight reports based on AI large models according to claim 1, characterized in that, After obtaining the hybrid feature encoding, S103 inputs it into the large AI model based on the Transformer architecture. The format conversion is completed through preprocessing, including: converting the triples of the domain knowledge graph into triple text, converting the formula text of the analysis logic rules into standard formula text, and normalizing the tags and descriptive text words of the insight dimension template. Finally, a unified text sequence consisting of triple text, standardized formula text, and normalized tag text is formed, thus completing the format conversion.

6. The method for generating intelligent insight reports based on AI large models according to claim 1, characterized in that, After preprocessing the hybrid feature encoding, it is input into the large AI model, where a word embedding mechanism is introduced for vector encoding, including: We select a pre-trained word embedding tool that is of the same origin as Transformer, download matching pre-trained weights from the weight library of the pre-trained word embedding tool, and load them into the parameter matrix of the word embedding mechanism. Enable parameter modification in the word embedding tool to complete the configuration of custom parameters; Match the words in the unified text sequence with the words embedded in the pre-trained words. If there are mismatched words, initialize a random low-dimensional vector as the initial embedding vector. Based on the optimization parameters for general scenarios, parameter configuration is performed. The vector dimension parameter is set to 300 dimensions in the core parameters of the word embedding tool. The window parameter is also set to cover the contextual information of two words before and after the current word in the unified text sequence. If the current word is at the beginning of the unified text sequence with no words to its left and / or at the end with no words to its right, a filler word is generated. After generating an initial vector for each word within the window, the word embedding mechanism uses a local attention algorithm to fuse the window semantics into the current word's vector. The minimum word frequency threshold is set to 3. If the number of occurrences of a word in the unified text sequence is less than 3, a random low-dimensional vector is uniformly assigned. If the number of occurrences of a word in the unified text sequence is greater than or equal to 3, a 300-dimensional vector is generated during encoding based on the word's own semantics and the context window, thus obtaining the initial domain feature vector.

7. The method for generating intelligent insight reports based on AI large models according to claim 6, characterized in that, After the parameters are configured, a validity check is performed, including: Select the initial vectors of the same word in different windows and calculate the cosine similarity between them. If the cosine similarity is less than 0.5, it indicates that the window parameter configuration is effective. Then, the standard deviation of all initial domain feature vectors is calculated. If the standard deviation is greater than 0.15, it indicates that the minimum word frequency threshold is set reasonably.

8. The method for generating intelligent insight reports based on AI large models according to claim 7, characterized in that, After obtaining the initial domain feature vector, the initial domain feature vector transformed from the core target metaword in the insight data packet is given the first priority, the initial domain feature vector transformed from the data range metaword in the insight data packet is given the second priority, and the initial domain feature vector transformed from the industry tag in the insight data packet is given the third priority. Weights are assigned to the initial domain feature vectors of the first, second, and third priorities to form a fused query vector. The fused query vector is then normalized to make the query vector modulus 1. Create a two-dimensional matrix to initialize the similarity score matrix, and set the row index to the query vector, the column index to all initial domain feature vectors, and the matrix elements to the cosine similarity scores between the query vector and each initialized initial domain feature vector. Calculate the cosine similarity scores by dot product and project them into the [0,1] interval, and write the cosine similarity scores into the two-dimensional matrix. Finally, perform differential assignment of cosine similarity scores through mathematical calculation using the Softmax function to complete the weight allocation, generating high-weight vector lists and low-weight vector lists respectively, and passing them to the corresponding levels of the AI ​​large model.

9. The method for generating intelligent insight reports based on AI large models according to claim 8, characterized in that, In step S104, after weight allocation, the list of high-weight vectors is injected into the semantic understanding core layer. The core layer calls the self-attention sublayer of Transformer to perform deep semantic parsing on the high-weight vectors and generate core insight conclusions. After the list of low-weight vectors is injected into the logic reasoning layer, the reasoning layer calls the logic verification module to perform supplementary analysis on the low-weight vectors and generate supplementary reasoning basis.

10. A system for generating intelligent insight reports based on AI large-scale models, applied to the method for generating intelligent insight reports based on AI large-scale models as described in any one of claims 1-9, characterized in that, The system includes: The insight parsing and resource matching module receives natural language insight requests from users, generates insight data packages through standardization processing, fuzziness evaluation, metaword splitting, and semantic association extension of the BERT model, and builds a domain knowledge graph based on the ontology editor, forming a plugin library by analyzing logic rules and insight dimension templates. It extracts basic keyword sets from the insight data packages, matches them with Jaccard similarity coefficients and loads corresponding domain cognitive plugins, and finally integrates them to generate hybrid feature codes. The preprocessing and vector transformation module is used to optimize the data format and transform the hybrid feature encoding into vectorized data. It transforms the triples in the hybrid feature encoding into triple text, the formula text into standard formula text, and normalizes the label and descriptive text to form a unified text sequence. It selects a pre-trained word embedding tool that is the same as Transformer, assigns weights according to priority, and fuses them to generate a normalized query vector. Attention weight allocation and hierarchical reasoning module: It is used to achieve differentiated allocation of vector weights through attention strategy, create a similarity score matrix, calculate the cosine similarity between the query vector and the initial domain feature vector, generate a list of high-weight vectors and a list of low-weight vectors through the Softmax function, input the high-weight vectors into the semantic understanding core layer, call the Transformer self-attention sub-layer to generate core insight conclusions, input the low-weight vectors into the post-logic reasoning layer, and call the logic verification module to generate supplementary reasoning basis; The integration and generation module is used to integrate the results of hierarchical reasoning to form a complete and structured intelligent insight report. It summarizes the core insight conclusions output by the semantic understanding core layer and the supplementary reasoning basis output by the subsequent logical reasoning layer, sorts out and analyzes the logical relationships and priorities of the data, organizes the content according to a standardized format, and finally generates a complete intelligent insight report.

Images