A multi-modal data analysis report automatic generation method

Abstract

This invention relates to the field of data processing and document generation technology, specifically to an automatic method for generating multimodal data analysis reports, comprising the following steps: S1: Collecting raw data; storing multimodal features uniformly in a standardized feature database; performing cross-modal matching on different modal features in the standardized feature database to generate an analysis node set; sorting and filtering key target nodes, and generating analysis insight paragraphs based on these nodes; S3: Filling the analysis insight paragraphs into the corresponding positions of the report skeleton, adjusting the content style according to audience type parameters, generating the final multimodal data analysis report, and outputting it. This invention, through a unified multimodal feature extraction, semantic association matching, and style adaptive generation mechanism, achieves the automatic generation of intelligent foreign trade analysis reports with clear structure, focused content, and appropriate expression under data-driven conditions.

Description

Technical Field

[0001] This invention relates to the field of data processing and document generation technology, and in particular to a method for automatically generating multimodal data analysis reports. Background Technology

[0002] With the increasing digitalization of global trade, foreign trade-related enterprises and management departments have accumulated massive amounts of heterogeneous data, including structured transaction records, unstructured customer reviews, illustrated analysis reports, and multi-source business data from cross-border e-commerce platforms, international logistics systems, and enterprise ERP systems. This data is diverse in form, frequently updated, and contains rich value in business trends, risk warnings, and market insights. However, traditional methods of generating data analysis reports often rely on manual compilation and writing, which is inefficient, subjective, and difficult to adapt to rapidly changing business needs. At the same time, the lack of structured correlation methods between multimodal data leads to fragmented information, making it difficult to form a comprehensive analysis report from a unified perspective, thus affecting the timeliness and accuracy of data-driven decision-making.

[0003] Existing technologies still have significant shortcomings in addressing how to quickly extract key features from multimodal foreign trade data, automatically organize content structure, and generate intelligent analysis reports with styles adapted to different business focuses and target audiences. Specific problems include: first, the lack of a unified feature extraction mechanism for numerical, textual, and mixed data, resulting in ineffective integration of multimodal information; second, the lack of an automatic matching and filling mechanism between analysis content and report structure, making it difficult to generate structured and focused reports; and third, the fixed report expression style, failing to flexibly adjust the content presentation according to different decision-making levels. Therefore, there is an urgent need for an automatic multimodal data analysis report generation method to achieve structured data expression, intelligent content arrangement, and adaptive language style, thereby improving analysis efficiency and application value. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides a method for automatically generating multimodal data analysis reports.

[0005] A method for automatically generating multimodal data analysis reports includes the following steps:

[0006] S1: Collect raw data from multiple preset foreign trade data source interfaces, and classify the data into numerical data, text data and mixed data according to the data content type; extract the core business indicators from the numerical data as numerical features, analyze the key entities and sentiment in the text data to form text features, and analyze the mixed data to obtain structured numerical-text feature pairs; and store the multimodal features in a standardized feature database to form a unified data foundation for subsequent analysis.

[0007] S2: Call the preset business association rule library, perform cross-modal matching on different modal features in the standardized feature database, and generate an analysis node set with initial association strength; combine the externally input report focus parameters and audience type parameters, filter the target node set from the analysis node set, and generate a matching report skeleton structure; then perform an evolution calculation of the association strength on the target node set, sort and filter key target nodes, and generate analysis insight paragraphs based on this.

[0008] S3: Fill the analysis and insight paragraphs into the corresponding positions of the report skeleton, adjust the content style according to the audience type parameters, generate the final multimodal data analysis report and output it.

[0009] Optionally, the foreign trade data sources include customs statistics platforms, cross-border e-commerce platforms, international logistics tracking systems, and enterprise resource planning systems, and collect raw data streams according to preset cycles or trigger events; the data content type classification includes classifying purely structured tables and time series data streams as numerical data; classifying product descriptions, customer reviews, and news information as text data; and classifying documents with embedded data reports and analytical articles with attached numerical summaries as mixed data.

[0010] Optionally, for numerical data, core business indicators, including import and export volume, order volume, year-on-year growth rate, and market share, are extracted using a predefined indicator mapping table to form a numerical feature vector. For textual data, key entities are extracted using named entity recognition technology, including company name, product category, and country / region. The sentiment tendency of the text is determined based on a sentiment analysis model, and the results are combined to form a text feature vector. For mixed data, information segmentation technology is used to first separate the numerical part from the textual description part. Then, numerical features and textual features are extracted separately using the same method. The numerical features and textual features are paired according to the context of the original text to construct structured numerical-textual feature pairs.

[0011] Optionally, the numerical feature vectors, text feature vectors, and numerical-text feature pairs are encoded with unified feature identifiers, data timestamps, and data source identifiers, and stored in a standardized feature database.

[0012] Optionally, the preset business association rule base includes multiple cross-modal mapping rules defined in the form of condition-association; by traversing the features in the standardized feature database, different modal features that satisfy the same rule conditions are associated to generate initial analysis nodes, the initial analysis nodes include the associated feature data and the corresponding initial association strength; all initial analysis nodes are summarized to form an analysis node set.

[0013] Optionally, S2 includes matching the report focus parameters with the topic tags of the rules in the business association rule base, and prioritizing the selection of analysis nodes under relevant tags to form a preliminary target set; at the same time, according to the association strength threshold or node number requirement set by the audience type parameter, the final target node set is selected from the preliminary target set.

[0014] Optionally, S2 further includes calling and adapting the corresponding report chapter template from a preset template library based on the report focus parameters and audience type parameters, and generating a report skeleton structure including a title, core argument hierarchy, and data placeholders.

[0015] Optionally, for each node in the target node set, its association strength is iteratively updated based on the changes in its associated features within the latest data period to complete the evolution calculation; all target nodes are sorted according to the association strength after the evolution calculation, and several top-ranked nodes are selected as key target nodes; for each key target node, its associated feature data is substituted into a pre-set insight language template to generate an analysis insight paragraph describing the relationship between features and the business meaning.

[0016] Optionally, S3 includes automatically matching the analysis insight paragraph with the chapter position required by the predefined topic tags and data types in the report skeleton structure, based on the business theme to which the corresponding target node belongs and the type of analysis conclusion, and filling the analysis insight paragraph into the corresponding position.

[0017] Optionally, S3 further includes calling a preset style rule library to style the populated report content according to the audience type parameter; the style rule library includes the configuration of language style, level of detail in chart presentation, and priority of conclusions corresponding to different audience types;

[0018] The content style includes converting technical terms into expressions suitable for the target audience, adjusting the proportion of detailed data and trend conclusions in the analysis and insight paragraphs according to the configuration, and automatically generating or adapting corresponding charts, summaries, and execution outlines; all content that has been styled is compiled according to the report skeleton structure format to generate the final multimodal data analysis report document, and output in a specified format through a preset interface.

[0019] The beneficial effects of this invention are:

[0020] This invention, by constructing a unified multimodal feature extraction process, can extract core feature vectors with business semantics from structured numerical data, unstructured text data, and mixed document data, and standardize and store them in a unified database, establishing a clear and callable data foundation for subsequent cross-modal analysis. Based on this, and combined with a pre-set business association rule library, semantic association matching between multimodal features is achieved, automatically generating analysis nodes with quantifiable intensity scores. Furthermore, target node selection and structural template matching are performed based on report focus parameters and audience type parameters, thereby realizing the automatic construction of analysis structures tailored to business scenarios.

[0021] This invention adapts the language and adjusts the chart presentation of the analysis content through a style rule library. It can flexibly configure the level of detail of data, terminology, and presentation of conclusions according to different audience types, achieving dynamic alignment between report content and user cognitive preferences. Combined with the evolution and update mechanism of analysis and insight paragraphs and the structured filling method, it can continuously output multimodal data analysis reports that focus on business changes, maintain a consistent style, and highlight key points, effectively improving the analysis efficiency and decision support capabilities of foreign trade enterprises in complex data environments. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the automatic generation method for analysis reports according to an embodiment of the present invention. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0025] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0026] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0027] like Figure 1 As shown, a method for automatically generating multimodal data analysis reports includes the following steps:

[0028] S1: Collect raw data from multiple preset foreign trade data source interfaces, and classify the data into numerical data, text data and mixed data according to the data content type; extract the core business indicators from the numerical data as numerical features, analyze the key entities and sentiment in the text data to form text features, and analyze the mixed data to obtain structured numerical-text feature pairs; and store the multimodal features in a standardized feature database to form a unified data foundation for subsequent analysis.

[0029] S11, Data Source Access and Data Acquisition: Connections are established with multiple data sources containing typical foreign trade business information through pre-defined application programming interfaces (APIs), including the customs statistics platform, cross-border e-commerce platform, international logistics tracking system, and enterprise resource planning (ERP) system. The customs statistics platform provides macro-level business data such as import and export trade volume, trade partner structure, and commodity category segmentation; the cross-border e-commerce platform provides operational data such as order records, user behavior, and product sales; the international logistics tracking system provides time-series logistics data such as cargo shipment, transshipment, customs clearance, and receipt status; and the ERP system provides end-to-end data on procurement, inventory, finance, and customer orders.

[0030] After the connection is established, data collection tasks will be triggered according to two types of collection mechanisms: periodic collection and event-triggered collection. Periodic collection is controlled by a scheduler, such as capturing the previous day's order data at 0:00 every day or obtaining incremental updates from the customs platform once an hour. Event-triggered collection relies on a listening mechanism. For example, when the ERP system detects a new order generation event or the logistics platform updates the package receipt status, the collection task corresponding to that data source will be triggered immediately. By combining these two mechanisms, both the comprehensiveness and real-time performance of the data can be guaranteed.

[0031] S12, Data Type Identification and Classification: After collecting the raw data, different types of data are distinguished, including numerical data, text data, and mixed data. Specifically: Numerical data mainly exists in the form of tables, time series, etc., including trade volume, sales volume, inventory changes, etc. This data can be directly quantified. When distinguishing, structured business data represented in tabular formats such as CSV and XLS, and time series data are classified as numerical data. Text data is mostly natural language text, unstructured, and not suitable for direct calculation. The content involves product descriptions, user reviews, policy interpretations, market information, etc., usually from e-commerce review sections, customer service records, or news websites. When distinguishing, unstructured semantic content such as customer interaction and information dissemination is classified as text data. Mixed data contains both structured and unstructured content. For example, a PDF report contains both charts and detailed analysis descriptions; or a Word document contains both numerical conclusions and market interpretations. This type of data needs to be parsed first, and then each part is separated for subsequent processing. When distinguishing, PDF reports with embedded tables and Word analysis documents containing numerical summaries are classified as mixed data.

[0032] S13, Feature Extraction and Structuring: Perform feature extraction processing on the classified data types, specifically including:

[0033] For numerical data, a pre-defined indicator mapping table is used first. This table acts like a rule list, clearly defining which fields represent key business indicators. The criteria for this include common key operating indicators in the foreign trade industry, such as transaction amount, number of transactions, and year-on-year growth rate; high-frequency indicators that repeatedly appear in the company's historical data analysis; and numerical dimensions that have a strong correlation with operational results in actual business scenarios. The extraction process involves extracting these key fields from the table and combining them according to pre-defined weights to form a structured numerical feature vector. The weights are default values ​​but can also be obtained through regression analysis of historical training data. The purpose is to highlight the importance of different indicators to subsequent analysis. Therefore, by calling the pre-defined indicator mapping table, core business indicators can be extracted to form a numerical feature vector. : ,in, For the first Raw values ​​of business indicators, such as import and export volume, order volume, year-on-year growth rate, etc.; For the first The standardized weights corresponding to each indicator; The number of indicators; These are numerical feature vectors.

[0034] Table 1 Indicator Mapping Table

[0035] Display field names (raw data) Business metric name Indicator Code Standard units Weighting coefficient Should it be included in feature extraction? export_amount Export amount EXAMT US$10,000 0.35 yes import_amount Import amount IMAMT US$10,000 0.3 yes order_quantity Order volume ORDQTY one 0.15 yes yoy_growth Year-on-year growth rate YOYG % 0.1 yes market_share market share MKTSHR % 0.1 yes inventory_level Inventory levels INVLVL Item 0 no

[0036] During feature extraction, all fields are first read from the raw data; then, the table is matched based on the displayed field names; only fields marked as "yes" for feature extraction are retained; next, their values ​​are converted to standard units, such as converting RMB to USD, or converting the number of items to a standard package; finally, they are weighted according to weight coefficients to construct a numerical feature vector. For different application scenarios, such as ocean freight exports, cross-border e-commerce, and international trade analysis, multiple versions of the indicator mapping table can be set up and dynamically switched through configuration management.

[0037] For text-based data, named entity recognition (NER) technology is used to understand which key entities are mentioned in the text. The key entity recognition process includes: first, lexical segmentation and syntactic analysis of the text; then, using a pre-trained NER model to extract specific types of noun phrases, such as company names, product categories, and country / region names; each identified entity is added as a structured element to the feature set. Based on this, a sentiment analysis model is used to determine the text's attitude, such as whether customer reviews are positive or negative. Finally, these key entities plus the sentiment value constitute the text feature vector of the text. Named entity recognition models are used to extract key entity sets. This includes company name, product category, country / region, etc.; the sentiment analysis model is used to calculate sentiment tendency values. Positive values ​​represent positive, negative values ​​represent negative, and 0 represents neutral; these two values ​​construct the text feature vector. , .

[0038] The named entity recognition model structure is BERT+CRF, which is: original text → word segmentation → BERT encoding layer → BiLSTM → CRF layer → entity sequence label. The BERT encoding layer is used to transform each word in the sentence into a vector representation containing contextual semantics; the BiLSTM layer is used to further model contextual dependencies, making the model pay more attention to word order and contextual logic; the CRF layer is used to output the optimal label sequence to ensure the rationality of entity boundaries, such as avoiding label mismatches like B-company followed by E-location.

[0039] The sentiment analysis model structure is BERT+Softmax, i.e., original text → word segmentation → BERT encoding layer → average pooling or [CLS] vector → fully connected layer → Softmax → sentiment classification result. The BERT encoding layer converts the entire text into a semantically meaningful vector. The pooling layer or [CLS] vector is used to obtain the overall semantic representation of the sentence. The fully connected layer + Softmax maps the vector to a sentiment probability distribution, for example, 82% positive sentiment; 12% neutral sentiment; 6% negative sentiment. The final output is a positive, neutral, or negative sentiment label, which can be further assigned a sentiment tendency score, such as +1 / -1 / 0 or a normalized score.

[0040] For mixed data, document parsing and information segmentation techniques are used to separate the structured table content from the surrounding text paragraphs. When extracting table areas, all areas with table structures are first scanned, and rows, columns, cells, and their corresponding values ​​and title fields are extracted. The information in each cell is labeled with its location and associated context during processing. When extracting text paragraphs, the text content other than tables is extracted according to paragraph structure, and semantic units are identified, such as a sentence describing a trend or interpreting a data point, forming semantic fragments. Then, corresponding numerical feature vectors and text feature vectors are generated respectively. For combinations of tables and text, associations are established based on paragraph order and citation indicators. If a paragraph appears after a table and contains phrases such as "visible in the table above" or "as can be seen below," it is considered an interpretation of the table. If the table has a title or explanatory row, these parts are extracted as additional explanations and paired with the table content. If a text paragraph explicitly mentions numerical content, such as a 12.5% ​​increase, it is checked whether the value comes from a previous table to establish a backtracking link. After segmentation, the numerical fields in the table are processed into structured index values; the descriptive semantics in the text paragraphs are processed into text features; and the references and contextual relationships established between the two are encapsulated into numerical-text feature pairs. Preserve its original pairing structure.

[0041] S14, Feature Standardization and Storage: The above steps have extracted numerical feature vectors, text feature vectors, and numerical-text feature pairs. However, their formats, sources, and structures are different and cannot be directly used for the next step of analysis and matching. Therefore, these features must be processed. When processing features, three identifier fields will be added to each feature item or feature pair, including:

[0042] Feature identifier (FID) is equivalent to assigning a unique number to each feature data, such as FT20260101_001, which makes it easy to quickly locate the content and source of the feature in subsequent analysis or backtracking;

[0043] Data timestamp This indicates the point in time when the feature was extracted or the time period to which the original data it corresponds to belongs. This is very important for time series analysis, trend judgment, or data updates.

[0044] Data source identifier This is used to indicate which data platform or system the feature comes from, such as an ERP system logistics platform, customs interface, etc., so that multi-source data can maintain a clear traceability path during fusion analysis; all feature data that has undergone the above standardization process will be uniformly stored in a dedicated database, called the standardized feature database.

[0045] S2: Call the preset business association rule library, perform cross-modal matching on different modal features in the standardized feature database, and generate an analysis node set with initial association strength; combine the externally input report focus parameters and audience type parameters, filter the target node set from the analysis node set, and generate a matching report skeleton structure; then perform an evolution calculation of the association strength on the target node set, sort and filter key target nodes, and generate analysis insight paragraphs based on these nodes.

[0046] S21, Cross-modal matching and analysis node generation: Invoke the preset business association rule library, which consists of multiple cross-modal mapping rules defined in the form of conditions-associations. Each rule is used to describe the association relationship of different modal features at the business semantic level. The condition part is used to define which features conform to the association relationship in business semantics. The association part includes the feature category to be associated, the weight coefficient of each feature, and the result generation method. It can be understood as a pairing instruction manual oriented towards business semantics.

[0047] Table 2. Examples of typical business association rules

[0048] Rule Number condition Modal feature participants Feature weight allocation illustrate R001 Same product category & same country Numerical values: Export amount, order volume; Text: User comment sentiment value Export value: 0.5; Order volume: 0.3; Emotional value: 0.2 Used to determine the business trend of a certain type of product in a certain market. R002 Consistent regions & overlapping time periods Numerical values: Year-on-year growth rate; Text: Media reports; Sentiment Growth rate 0.6, sentiment score 0.4 Used to determine the correlation between business activity and market sentiment in a specific region. R003 Company name is consistent Numerical values: Return rate; Mixed data: Platform complaint data Return rate: 0.4%; Complaint rate: 0.6%. Used to identify the relationship between enterprise service quality and user satisfaction.

[0049] The system iterates through the multimodal feature data in the standardized feature database, matches and associates different modal features that satisfy the same rule conditions, and generates an initial analysis node based on each successfully associated set of features; the specific operations are as follows:

[0050] Extract feature metadata, such as product category, timestamp, and country of origin;

[0051] The feature is compared with the condition part of each rule in the rule base to determine whether the feature meets the premise of a certain rule;

[0052] Find multimodal feature combinations that meet a certain rule, such as the sentiment feature of a text, the export value of a product, and the order volume of the product all meeting rule R001;

[0053] These combinations are packaged into an initial analysis node, and its initial association strength is calculated using the following formula based on the feature weights set in the rules: ,in, For the first Standardized feature values ​​of the associated features; For the business association rule, the first Each feature has a pre-set weight coefficient; The number of features associated with the current node; The initial analysis focuses on the association strength of the nodes; all nodes that meet the generated rules are collected to form an analysis node set, awaiting subsequent filtering, sorting, and report generation operations.

[0054] S22, Target Node Filtering: Receives externally inputted report focus parameters and audience type parameters; matches the report focus parameters with the topic tags corresponding to each rule in the business-related rule base, filtering out analysis nodes consistent with the topic tags to form a preliminary target node set; topic tags are a mechanism for classifying analysis nodes, used to link each business rule with a specific analysis topic. This can be understood as each rule being manually assigned one or more topic keywords during its construction, such as market trends, customer feedback, price fluctuations, etc. The matching process is as follows: External users or the system input report focus parameters, such as analyzing export growth or monitoring customer satisfaction; based on keyword extraction technology, this focus parameter is semantically matched with the topic tags attached to each rule in the rule base; all analysis nodes whose rule tags are consistent with the semantics of the report focus are filtered out, forming a preliminary target node set; this ensures that subsequent report content will focus on analyzing issues that users care about, rather than general discussions.

[0055] After initially identifying nodes relevant to the topic, further filtering is required. This is because there may be dozens of nodes under the same topic, but the report cannot use them all. Therefore, the initial target node set needs to be filtered according to the preset association strength threshold or target node quantity constraint in the audience type parameter to obtain the final target node set. The association strength threshold includes setting a minimum credibility threshold; only nodes with an association strength greater than this value are considered to have sufficient explanatory value. The association strength threshold has two values: if the association strength ranges between [0, 1], the threshold can be set to 0.6; if the strength is a weighted score with a maximum of 100, the threshold is set to 60. The node quantity constraint is used to limit the number of final nodes to prevent the analysis content from becoming lengthy and unfocused. When the report is aimed at end customers or general operations personnel, the structure is simple and intuitive, retaining only the top N nodes. The value is set to output a maximum of 5 nodes or only display the first three key conclusions. The purpose of this sub-step is to filter out the most relevant and representative analysis nodes to the report's objectives through topic tags and personalized constraint mechanisms, ensuring both content focus and the value density of the output information.

[0056] S23, Report skeleton structure generation: Based on the report focus parameters and audience type parameters, a matching report chapter template is retrieved from the preset report template library; the report template's structure includes:

[0057] Chapter heading structure: The report uses first- and second-level headings to reflect logical hierarchy, such as overview, trend analysis, recommendations, etc.

[0058] Core Argument Hierarchy: The key points structure under each chapter, such as trend description + cause analysis + prediction + suggestions, is used to guide the order of content generation;

[0059] Feature data placeholders: Slots reserved in the template for filling feature data, conclusions, charts, or ratios extracted from the analysis nodes;

[0060] Tone and format tags: Define the writing style of this template, such as decision-oriented, explanation-oriented, or data-intensive, for use by the subsequent language generation module.

[0061] Example template structure for managers analyzing export trends:

[0062] [First-level heading]: Quarterly Analysis Report on Export Business

[0063] Chapter 1: Overall Overview

[0064] Data time range (#time_range#)

[0065] Total export amount (#total_export_amount#)

[0066] Year-on-year growth rate (#yoy_growth_rate#)

[0067] [Chapter 2]: Key Market Trend Analysis

[0068] Market region (#market_region#)

[0069] Export product categories (#product_category#)

[0070] Key indicator trend chart (#trend_chart_url#)

[0071] Chapter 3: Risks and Recommendations

[0072] Potential volatility indicator (#risk_indicator_desc#)

[0073] Recommendations (#recommendation#)

[0074] Note: #...# represents placeholders and will be replaced by actual data or analysis paragraphs in subsequent steps. Each chapter is organized around the characteristics reflected in the target node.

[0075] After matching, the parameters of the called chapter template are adapted to generate a report skeleton structure containing chapter titles, core argument hierarchy, and feature data placeholders, which is used to carry the subsequent generated analysis and insight content.

[0076] S24, Association Strength Evolution Calculation and Analytical Insight Generation: In the previous step, a report skeleton was generated using a template, and a set of target analysis nodes highly matching the report theme were selected. However, there are still differences in quality among the nodes because the data is dynamic. The purpose of this step is to re-evaluate the analytical value of each node based on the latest data; identify the most critical nodes and focus on them in the report; and automatically generate analysis paragraphs using natural language templates to clearly explain the data. Each analysis node initially has an initial association strength derived from S21, but this is only a static value based on historical data and rule settings. In real-world business, the actual performance of features may change, so it is necessary to dynamically re-evaluate these nodes based on the latest data in each reporting period, such as weekly or monthly. This process is called association strength evolution calculation.

[0077] The evolved correlation strength The calculation method is as follows: ,in, This represents the initial association strength of the analyzed node; For the first The amount of change of each associated feature within the current data period; Used to characterize the The evolution weighting coefficients for the degree of influence of changes in each feature; The number of associated features participating in the evolution calculation; The strength of the association after evolution calculation is as follows: Extract the feature value changes within the current data period, such as the export volume of a certain commodity increasing by 15% compared to the previous month; according to the preset change impact weights, perform a weighted sum of these changes; add the change score to the initial strength to obtain the evolved strength; the final result reflects whether this node is worth including in the report under the current business dynamics; each time this calculation is performed, it is equivalent to giving each node a latest value score.

[0078] Based on the evolved correlation strength The system sorts the analysis nodes in the target node set and selects the top-ranked nodes based on their correlation strength as key target nodes. Then, it pre-sets insight language templates for each analysis scenario; these templates are sentence frameworks for writing reports, complete with placeholders. For each key target node, the system fills the template with the associated numerical, sentiment, and entity class features. It automatically selects different sentence structures based on the trend direction of the features (increase / decrease) and outputs complete insight statements describing what happened, why, and what it might mean.

[0079] Example: Let the key target nodes be: Product Category: Electronic Components; Market Region: Latin America; Monthly Export Growth Rate: +22%; User Sentiment Index Improvement: +0.35. The generated insight paragraph would be: In the past month, exports of electronic components to the Latin American market have increased significantly by +22%. Simultaneously, the sentiment of user reviews has shifted from neutral to significantly positive by +0.35. This may reflect a recovery in demand for related products and increased customer satisfaction in the market. It is recommended to continuously monitor price fluctuations and changes in public opinion. The essence of this process is to dynamically evaluate the value of each analytical node based on the latest data and automatically generate high-quality analytical paragraphs using natural language. This makes the report more than just a template; it possesses intelligent analytical capabilities that highlight key points, provide logical explanations, and reveal trends, truly achieving data-driven decision support.

[0080] S3: Fill the analysis and insight paragraphs into the corresponding positions of the report skeleton, adjust the content style according to the audience type parameters, generate the final multimodal data analysis report and output it;

[0081] S31, Filling and Positioning of Insight Paragraphs: Previous steps have generated multiple insight paragraphs, each representing a business conclusion related to a key objective node. The current task is to automatically place these paragraphs in the most appropriate positions within the report framework, ensuring a logical final report structure and clear content distribution. This involves retrieving the insight paragraphs generated in step S2 and identifying the key objective node corresponding to each paragraph; matching the business theme tags and analysis conclusion types associated with the key objective nodes with the predefined chapter structures in the report skeleton, where each chapter structure is pre-associated with theme tags and data type requirements; business theme tags provide a semantic classification of the business content in the report, clarifying which business area the paragraph refers to. These tags are set during the generation of the analysis nodes and carried over to the insight paragraphs. Common business theme tags include export trends, customer sentiment, market share, channel performance, product category fluctuations, and public opinion risks; each chapter in the report skeleton is associated with one or more theme tags, indicating which type of analysis content is allowed or recommended for that chapter. Data type requirements include structured numerical analysis, sentiment-based text analysis, multimodal analysis, and summary descriptions; these are used to determine whether the feature type and output method of the current paragraph meet the requirements of the chapter, thus deciding whether to include it. If a match is successful, the corresponding analysis insight paragraph is automatically populated into the chapter position within the report's skeleton structure that matches its theme tags and data type requirements, completing a one-to-one mapping between content and structure.

[0082] S32, Stylistic Adjustment of Report Content: The previous step successfully filled each analysis paragraph into the report structure, but this content is currently in a standard expression state. However, in real-world scenarios, the target audience for different reports varies, potentially including senior management, data analysts, client representatives, and policy personnel. Therefore, the content style needs to be adjusted based on the audience's characteristics. Specifically, the audience type parameter determined in S2 is received first, and a pre-defined style rule library is invoked based on this parameter. This style rule library defines corresponding language style configurations, chart detail configurations, and conclusion presentation priority configurations for different audience types. Based on this style rule library, stylistic adjustments are performed on the report content of the already filled analysis insight paragraphs. Specific adjustments include:

[0083] Replace or restructure the terminology and sentence structures used in the original analysis paragraphs: convert technical terms into easy-to-understand colloquial expressions; adjust sentence structures to reduce complex passive sentences and stacked compound sentences; add logical cue words such as therefore, explanation, and visible; to improve readability, that is, convert the technical terms in the analysis and insight paragraphs into expressions that are suitable for the target audience's comprehension.

[0084] The analysis section of a report typically consists of two parts: a detailed description of data changes and a concluding statement of business trends, reasons, and recommendations. Different content ratios are configured for different audiences, controlling the amount of data presented and the amount of conclusions stated. Therefore, the length ratio of detailed data descriptions to trend conclusions in the analysis and insight section should be adjusted according to the aforementioned configuration rules.

[0085] Different groups of people have different levels of reliance on charts. Decision-makers prefer conclusions plus concise charts; analysts prefer complete charts plus data details; external clients are more suitable for charts with clear color schemes and explanatory labels. Therefore, based on the rule base configuration, the chart style is automatically selected, generated or adjusted, and supplemented with auxiliary structural content such as summary descriptions and execution summaries.

[0086] During the stylization process, the relationship between the proportion of data description and the proportion of conclusion description in the analysis and insight paragraphs can be expressed as follows: ,in, To analyze and understand the proportion of detailed data content in the paragraph; To analyze and understand the proportion of trend conclusions or business interpretations in paragraphs; different audience types correspond to different style rule bases. and Value combinations.

[0087] The example structure of the style rule library is shown in Table 3:

[0088] Audience type Language style Data / Conclusion Rd:Rc Chart Presentation Hierarchy Generate summary Add execution summary? Senior management Decision-oriented 0.3:0.7 Overview yes yes Data Analyst Technology-oriented 0.7:0.3 Full details no no First-line sales / customer service Simple and concise 0.5:0.5 Visual representation yes no Policy regulators Standardized neutrality 0.4:0.6 Comparison + Caption yes yes

[0089] In Table 3 above, language style determines the formality and professionalism of the vocabulary and sentence structure used in the analyzed paragraphs; data / conclusion ratio This is used to control the proportion of detailed data descriptions and trend conclusions in each analysis section; the chart presentation hierarchy is used to select whether to display the full chart, whether to simplify the information, and whether to add labels / explanations; whether to generate a summary / execution summary, and whether to automatically attach a concise summary or high-level guidance. This step S32 is a key step in transforming the analysis results into a report that the audience can read, understand, and be willing to read. By calling the style rule library, the report content not only has data support and analytical depth, but also has linguistic adaptability and structural aesthetics, realizing a complete closed loop from intelligent data analysis to intelligent language expression.

[0090] S33, Report Generation and Output: Compile all the report content after stylization adjustments according to the predefined chapter order and layout specifications in the report skeleton structure to generate a complete multimodal data analysis report document; output the multimodal data analysis report in a specified format through a preset output interface, including an editable document format, a fixed layout document format, or an online visual report format.

[0091] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0092] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for automatically generating multimodal data analysis reports, characterized in that, Includes the following steps: S1: Collect raw data from multiple preset foreign trade data source interfaces, and classify the data into numerical data, text data and mixed data according to the data content type; extract the core business indicators from the numerical data as numerical features, analyze the key entities and sentiment in the text data to form text features, and analyze the mixed data to obtain structured numerical-text feature pairs; and store the multimodal features in a standardized feature database to form a unified data foundation for subsequent analysis. S2: Call the preset business association rule library, perform cross-modal matching on different modal features in the standardized feature database, and generate an analysis node set with initial association strength; combine the externally input report focus parameters and audience type parameters, filter the target node set from the analysis node set, and generate a matching report skeleton structure; Subsequently, the association strength of the target node set is calculated, the key target nodes are sorted and selected, and the analysis and insight paragraphs are generated based on these nodes. S3: Fill the analysis and insight paragraphs into the corresponding positions of the report skeleton, adjust the content style according to the audience type parameters, generate the final multimodal data analysis report and output it.

2. The method for automatically generating multimodal data analysis reports according to claim 1, characterized in that, The foreign trade data sources include customs statistics platforms, cross-border e-commerce platforms, international logistics tracking systems, and enterprise resource planning systems, and collect raw data streams according to preset cycles or trigger events; the data content type classification includes classifying pure structured tables and time series data streams as numerical data; Product descriptions, customer reviews, and news information are categorized as text-based data; documents with embedded data reports and analytical articles with accompanying numerical summaries are categorized as mixed-type data.

3. The method for automatically generating multimodal data analysis reports according to claim 2, characterized in that, For numerical data, core business indicators, including import and export volume, order volume, year-on-year growth rate, and market share, are extracted through a predefined indicator mapping table to form a numerical feature vector. For text-based data, named entity recognition technology is used to extract key entities, including company name, product category, and country / region. The sentiment analysis model is used to determine the sentiment tendency of the text and combine them to form a text feature vector. For mixed-type data, information segmentation technology is used to first separate the numerical part and the text description part. Then, numerical features and text features are extracted in the same way. The numerical features and text features are paired according to the context of the original text to construct structured numerical-text feature pairs.

4. The method for automatically generating multimodal data analysis reports according to claim 3, characterized in that, The numerical feature vectors, text feature vectors, and numerical-text feature pairs are encoded with unified feature identifiers, data timestamps, and data source identifiers, and stored in a standardized feature database.

5. The method for automatically generating multimodal data analysis reports according to claim 1, characterized in that, The preset business association rule base includes multiple cross-modal mapping rules defined in the form of condition-association; by traversing the features in the standardized feature database, different modal features that meet the same rule conditions are associated to generate initial analysis nodes. The initial analysis nodes include the associated feature data and the corresponding initial association strength; all initial analysis nodes are summarized to form an analysis node set.

6. The method for automatically generating multimodal data analysis reports according to claim 5, characterized in that, S2 includes matching the report focus parameters with the topic tags of the rules in the business association rule base, and prioritizing the selection of analysis nodes under relevant tags to form a preliminary target set; at the same time, according to the association strength threshold or node number requirement set by the audience type parameter, the final target node set is selected from the preliminary target set.

7. The method for automatically generating multimodal data analysis reports according to claim 6, characterized in that, The S2 further includes calling and adapting the corresponding report chapter template from the preset template library according to the report focus parameters and audience type parameters, and generating a report skeleton structure including a title, core argument hierarchy and data placeholders.

8. The method for automatically generating multimodal data analysis reports according to claim 7, characterized in that, For each node in the target node set, its association strength is iteratively updated based on the changes in its associated features within the latest data period to complete the evolution calculation; all target nodes are sorted according to the association strength after the evolution calculation, and the top-ranked nodes are selected as key target nodes; for each key target node, its associated feature data is substituted into a pre-set insight language template to generate an analysis insight paragraph describing the relationship between features and the business meaning.

9. The method for automatically generating multimodal data analysis reports according to claim 1, characterized in that, S3 includes automatically matching the analysis and insight paragraphs with the predefined topic tags and data type requirements of the report skeleton structure, based on the business theme to which the corresponding target node belongs and the type of analysis conclusion, and filling the analysis and insight paragraphs into the corresponding positions.

10. The method for automatically generating multimodal data analysis reports according to claim 1, characterized in that, S3 further includes calling a preset style rule library to style the populated report content according to the audience type parameter; the style rule library includes the configuration of language style, level of detail in chart presentation, and priority of conclusions corresponding to different audience types; The content style includes converting technical terms into expressions suitable for the target audience, adjusting the proportion of detailed data and trend conclusions in the analysis and insight paragraphs according to the configuration, and automatically generating or adapting corresponding charts, summaries, and execution outlines; all content that has been styled is compiled according to the report skeleton structure format to generate the final multimodal data analysis report document, and output in a specified format through a preset interface.

Images