Conversational business intelligent analysis method based on natural language

Abstract

The invention discloses a dialogue type business intelligent analysis method based on a natural language, and relates to the technical field of business management systems, and the method mainly comprises the steps: recognizing the intention type of a user intention contained in a natural language request through a general large language model; according to the identified specific intention type, calling a corresponding special analysis module to perform semantic analysis and parameter extraction on the natural language request so as to generate an operable analysis instruction; calling a business intelligent system data interface to execute the analysis instruction, and obtaining corresponding structured data from a business intelligent system data warehouse; and generating a comprehensive analysis report which is matched with the identified intention type and the corresponding structured data and comprises visual elements and analysis explanations based on the identified intention type and the corresponding structured data. According to the method, the intention type of the natural language request of the user is understood through the large language model, so that the user with a non-professional background can easily perform complex data exploration, and the use threshold is reduced.

Description

Technical Field

[0001] This invention relates to the field of business management system technology, and more specifically to a conversational business intelligence analysis method based on natural language. Background Technology

[0002] With the development of healthcare informatization, the shortcomings of traditional business intelligence (BI) systems in data analysis and decision support have become increasingly apparent. Existing systems typically require users to have specialized technical backgrounds, such as mastering complex query languages ​​or user interfaces, making them difficult for frontline business and management personnel to use directly, resulting in severe technical dependence and efficiency bottlenecks. Their static and rigid reporting models cannot quickly respond to changing business needs, and the analysis process is lengthy, making real-time insights difficult to achieve. Furthermore, data is often scattered across multiple independent systems, forming information silos and lacking effective automated attribution analysis and trend prediction capabilities, making it difficult to deeply explore the value of the data. Conversational BI tools that have emerged in recent years still have unsatisfactory interactive experiences, limited understanding of natural language, and difficulty in supporting flexible and coherent multi-turn analytical dialogues, severely limiting data availability and the effectiveness of decision-making. Therefore, there is an urgent need for a new data analysis solution that can lower the barrier to entry, improve analytical efficiency, and support intelligent interaction. Summary of the Invention

[0003] To lower the barrier to entry for using medical business intelligence systems and improve analytical efficiency, this invention proposes a natural language-based conversational business intelligence analysis method, comprising the following steps: S1: Receive natural language requests from the current user in the conversation; S2: Identify the intent type of the user intent contained in the natural language request through a general large language model. The intent types include indicator query, attribution analysis, trend prediction, and topic generation. S3: Based on the specific intent type identified, call the corresponding dedicated analysis module to extract and semantically parse the natural language request, including indicators, dimensions, and time range, and generate operable analysis instructions; S4: Call the business intelligence system's data interface to execute analysis commands and obtain the corresponding structured data from the business intelligence system's data warehouse; S5: Based on the identified intent type and the corresponding structured data, generate a comprehensive analysis report that matches it and includes visualization elements and analytical explanations; S6: Return and present a comprehensive analysis report to the user.

[0004] This invention uses a general large language model to understand the intent type of a user's natural language request, realizing the transformation from "humans adapting to machines" to "machines understanding humans," enabling users without a professional background to easily conduct complex data exploration and lowering the barrier to entry.

[0005] Furthermore, in step S3, when the intent type is an indicator query, the dedicated analysis module is a parameter extraction and matching module. The parameter extraction and matching module performs dynamic parameter extraction by using a large language model based on LoRA fine-tuning, and uses a hybrid retrieval scheme combining Trie tree and BM25 algorithm to match indicators, dimensions and time parameters.

[0006] Furthermore, the hybrid retrieval scheme first performs precise matching using a Trie tree. If the matching fails, it uses the BM25 algorithm to calculate similarity and perform fuzzy recommendations. If the confidence level of the fuzzy recommendation result is lower than a preset threshold, it calculates cosine similarity and returns a candidate list for the user to choose from.

[0007] Furthermore, in step S3, when the intent type is attribution analysis, the dedicated analysis module is an attribution analysis module based on the SHAP model, used to quantify the contribution of each dimension to the fluctuation of the specified indicator.

[0008] Furthermore, the execution process of the attribution analysis module includes: Calculate the SHAP value of all drill-down dimensions for the fluctuation of a specified indicator, and sort them by absolute value to select the top K dimensions with the largest contribution. For each of the first K selected dimensions, the key dimension value causing the indicator fluctuation is determined by calculating and sorting the SHAP value of all specific values ​​of the dimension.

[0009] Furthermore, in step S3, when the intent type is trend prediction, the dedicated analysis module is a time series prediction module, which performs prediction using a NexusTrendNet model that integrates trend items, seasonal items, and external feature items.

[0010] Furthermore, the external characteristics include one or more of holidays, policy adjustments, and seasonal factors.

[0011] Furthermore, in step S3, when the intent type is topic generation, the dedicated analysis module is a general large language model, which summarizes and refines the current or historical dialogue content to generate a structured analysis topic and analysis framework.

[0012] Furthermore, in step S3, if complete analysis parameters cannot be obtained through semantic parsing and parameter extraction, the step further includes: initiating a multi-round dialogue interaction through a general large language model to guide the user to supplement necessary information until a natural language request that satisfies the information required for complete analysis parameters is formed.

[0013] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The natural language-based conversational business intelligence analysis method proposed in this invention understands the intent type of the user's natural language request through a general large language model, realizing the transformation from "human adapting to machine" to "machine understanding human", enabling users without professional background to easily conduct complex data exploration and reducing the threshold for use; (2) In the index query, the large language model based on LoRA fine-tuning is used for parameter extraction, and combined with the Trie and BM25 hybrid retrieval scheme, which not only ensures the accurate matching of professional terms, but also has the ability to tolerate and recommend colloquial and vague expressions, thus improving the robustness of the query and the user experience. (3) By integrating the SHAP model for automated and quantitative intelligent attribution analysis, the key dimensions and specific dimensions of indicator fluctuations can be quickly located, transforming the analysis work that originally relied on human experience into objective data interpretation, thereby improving the efficiency and scientific nature of problem diagnosis. (4) The NexusTrendNet model, which incorporates external features, is used to provide users with accurate and interpretable prediction results. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of a conversational business intelligence analysis method based on natural language. Detailed Implementation

[0015] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0016] Currently, although Business Intelligence (BI) systems are widely used in the field of medical data analysis, traditional technical solutions still have significant limitations in practical applications, hindering the effective release of data value. These systems typically require users to master specialized query languages ​​or complex interfaces, excluding frontline users and managers, resulting in low analytical efficiency and high dependence on technical teams. Their inherent static reporting model is slow to respond and cannot adapt to the rapidly changing analytical needs of business scenarios, causing decision support to often lag behind management needs. Furthermore, the silo effect created by data fragmentation makes cross-system analysis extremely difficult, and advanced analytical functions, such as automated attribution of indicator fluctuations and accurate prediction of future trends, are generally lacking. Although some new conversational tools attempt to introduce natural language interaction to improve the experience, their limited understanding capabilities and rigid dialogue make it difficult to support in-depth, context-coherent multi-turn analytical exploration, ultimately making it difficult to transform massive amounts of data into tangible business insights. Therefore, such as... Figure 1 The present invention proposes a conversational business intelligence analysis method based on natural language, comprising the following steps: S1: Receive natural language requests from the current user in the conversation; S2: Identify the intent type of the user intent contained in the natural language request through a general large language model. The intent types include indicator query, attribution analysis, trend prediction, and topic generation. S3: Based on the specific intent type identified, call the corresponding dedicated analysis module to extract and semantically parse the natural language request, including indicators, dimensions, and time range, and generate operable analysis instructions; S4: Call the business intelligence system's data interface to execute analysis commands and obtain the corresponding structured data from the business intelligence system's data warehouse; S5: Based on the identified intent type and the corresponding structured data, generate a comprehensive analysis report that matches it and includes visualization elements and analytical explanations; S6: Return and present a comprehensive analysis report to the user.

[0017] Specifically, the business intelligence system (hereinafter referred to as "the system") first receives natural language requests from the current user through a graphical user interface or voice interface. These requests can be simple queries, such as "check the number of outpatient and emergency room visits last month," or complex analytical needs, such as "why were the number of outpatient and emergency room visits so high on August 5th" or "predict the outpatient volume trend for the next three months." The system supports multi-turn dialogues, and the current request is understood in conjunction with the dialogue context to ensure the continuity and intelligence of the interaction.

[0018] Subsequently, the system uses a large language model to identify the types of user intent contained in the natural language requests. This large language model, as a general-purpose semantic understanding engine, can effectively distinguish user goals without requiring fine-tuning for specific functions. These intent types mainly include, but are not limited to: indicator query, attribution analysis, trend prediction, and topic generation. For example, for the request "Query the number of discharged patients at Hospital A last week," the large language model would identify its intent as "indicator query"; for the request "Analyze why the number of surgeries decreased this week," it would be identified as "attribution analysis"; for the request "Predict the demand for drug inventory next quarter," it would be identified as "trend prediction"; and for the request "Summarize the discussion about outpatient efficiency into a topic," it would be identified as "topic generation."

[0019] Once the user's intent is successfully identified, the system will call the corresponding dedicated analysis module to perform in-depth semantic parsing and parameter extraction on the natural language request, based on the specific intent type identified, and finally generate analysis instructions that can be executed by the underlying business intelligence system.

[0020] When the identified intent is a metric query, the system invokes a parameter extraction and matching module specifically designed for this function. This module first uses a large language model fine-tuned using Low-Rank Adaptation (LoRA) technology for dynamic parameter extraction. The training data for this fine-tuning process is constructed in a specific way: metrics are extracted from the BI data warehouse ( ), dimension ( ), and randomly generate time ranges ( (This is a sentence fragment that can be translated into a sentence or phrase, but the context is unclear.) The formula for generating training sample pairs is: , "Help me find out". "according to" "Perform a query" ",according to "Query" , in, for A round-robin dialogue contains a set of natural language requests. For the natural language request in the i-th round of dialogue, A natural language generation function used to process structured elements (indicators) in a BI data warehouse. Dimensions ,time Combine these elements into a query that matches the business scenario. Output in JSON format. A structured mapping function used to map indicators from input natural language. Dimensions ,time These elements are mapped to a standardized JSON format. Let be the training sample pair for the i-th round of dialogue.

[0021] Here, the present invention introduces a LoRA adapter in the forward propagation process of the large language model, the computation of which is expressed as follows: , , , in, Embedded for Token For word embedding functions, A sequence of natural language requests. For embedding matrix; For the first Transformer layers enhanced with LoRA It is a complete Transformer layer computation process. It takes the hidden state of the previous layer as input, adds a linear transformation matrix injected by LoRA, and then passes through self-attention, feedforward network, residual connection and layer normalization to output the hidden state of the current layer. For the first The original layer weight matrix is ​​kept frozen during LoRA fine-tuning. Let B be the LoRA matrix, which can be used for training. Its function is to map a low-dimensional vector back to a high-dimensional output space, thus restoring the original layer's dimension. Let A be the LoRA matrix, which can be used for training. Its function is to compress the input from a high-dimensional space to a low-dimensional space, thereby reducing the number of parameters. For the first Layer bias vector; For output projection, For the t-th output token, Generate a sequence for history , A natural language request sequence; The original weight matrix for the language modeling head is the matrix in the last layer of the large language model that transforms the hidden states into vocabulary probabilities. It has a large number of parameters and remains unchanged during LoRA fine-tuning without being updated. LoRA matrix for modeling the language head, which maps the low-dimensional representation back to the sum of the original and the new. The same output dimension is the new trainable part added by LoRA. The LoRA matrix for the language modeling head compresses the input hidden states from high dimension to low dimension and is the newly added trainable part of LoRA. Let LoRA be the hidden state of the Lth layer at time step t. Specific adapters are provided for different concepts such as metrics, dimensions, and time. For example, the parameter adapter has the following parameters: , , .in, For the real number field, It is a low-rank dimension, used to compress the input into a subspace much smaller than the original dimension. For input dimensions, For output dimensions, Responsible for compression, It is responsible for mapping back to the output dimension, thus forming a low-rank trainable update of the original weights. Only the LoRA parameters are updated during model training. This freezes the parameters of the original model. In this way, the large language model is efficiently fine-tuned into a precise "metric-dimension-time" parameter extractor.

[0022] After extracting the initial parameters, the system does not use them directly. Instead, it employs a hybrid retrieval scheme combining Trie trees and the BM25 algorithm to perform precise matching or fuzzy recommendations on the extracted parameters (especially indicator and dimension names). Specifically, the system first constructs a Trie tree containing all standard indicator and dimension names. It then performs a fast and precise match on the words in the user's query using the Trie tree. If a match fails (e.g., the user used "number of outpatient and emergency visits" instead of the standard term "number of outpatient and emergency visits"), the BM25 algorithm is activated to calculate the similarity between the result and all standard terms for fuzzy recommendations. If the highest confidence score in the fuzzy recommendations is below a preset threshold, the system further provides a list of N candidate results with the highest similarity scores through cosine similarity calculation, guiding the user to make a selection via an interactive interface. After the user makes a selection, the system automatically optimizes the original query, semantically fusing the selected content with the original question to form a standardized query, thereby improving the accuracy of the query.

[0023] When the identified intent is attribution analysis, such as a user asking "Why is the indicator value abnormal on a certain day?", the dedicated analysis module, specifically the attribution analysis module based on the SHAP (SHapley Additive exPlanations) model, is invoked. The core purpose of this module is to quantify the contribution of each dimension to the fluctuation of the specified indicator. Its execution process is a two-stage drill-down and sorting process: In the first stage, the SHAP value of all drill-down dimensions N (e.g., the relevant drill-down dimensions under the "Outpatient and Emergency Visits" indicator are shown in Table 1) is calculated with respect to the fluctuation of the specified indicator. The calculation formula is based on the Shapley value in cooperative game theory: , in, The contribution value for dimension i. Let be the i-th candidate analysis dimension, which is the attribute used to divide or analyze the indicators. N is the set of all feasible dimensions, and S is a subset of N that does not contain dimension i. The SHAP function is a metric function used to evaluate the predicted value or contribution of a metric or target variable within a given set of dimensions. After calculation, the SHAP values ​​of all dimensions are sorted in descending order of their absolute values, and the top K dimensions with the largest contributions (e.g., K=5) are selected.

[0024] Table 1: Drill-down dimensions for "Outpatient and Emergency Visits" In the second stage, for the K key dimensions selected above, for each selected dimension, all its specific values ​​are used as new analysis objects, and the SHAP value of each value's total contribution to that dimension is calculated. The calculation formula is as follows: , in, This represents the contribution of value j to the dimension, where D represents the set of all values ​​under that dimension, and P is a subset of set D that does not contain value j. Finally, the absolute values ​​of the SHAP (Shape-Up Points) for all values ​​under each key dimension are sorted in descending order to pinpoint the key dimension value causing the fluctuation (e.g., discovering that the department "Respiratory Medicine" or the disease name "Influenza" caused the abnormal number of outpatient and emergency room visits). The entire process transforms the interpretable output of machine learning into an attribution report that users can directly understand.

[0025] When the identified intent is trend prediction, the time series prediction module is invoked. The core of this module is a prediction model called NexusTrendNet, which effectively integrates trend, seasonal, and external feature components. The model's prediction formula can be expressed as: , in, Represents the future The predicted value at each time step, It is a neural network that integrates trend, seasonality, and external features. Its goal is to decompose time series data into three main components: the trend term (…). ), capturing long-term trends; seasonal items ( ), capturing periodic fluctuations (such as daily, weekly, or monthly cycles); external characteristic items ( ), and handle the impact of holidays, policies or seasonal events on forecasts. It's the past A historical observation sequence at each time step, Predicting the future The external feature vectors were manually entered at specific time points. This model can be further decomposed into: , in, It is a trend term learned by a neural network, capturing long-term changes; It is a seasonal item, capturing cyclical fluctuations; This is a key component for processing external features, including holidays, policy adjustments, and seasonal factors, used to learn the nonlinear impact of these external features on the predicted values. Furthermore, the model calculates the uncertainty of the prediction and outputs confidence intervals. , in, It is a confidence interval. It is the quantile of the standard normal distribution. It is the standard deviation of the prediction. This makes the prediction result not just a single value, but a probability range, thus providing a risk reference for decision-making.

[0026] When the identified intent is topic generation, the system directly utilizes a large language model as a dedicated analysis module. This module automatically generates a structured analysis topic and framework by deeply summarizing and refining the content of the current conversation or historical dialogue records. For example, after a series of questions and answers with the system regarding outpatient efficiency, bed turnover rate, and average length of stay, a user can input a natural language request to "generate a topic from the above dialogue." The system will then automatically generate a topic titled "Comprehensive Analysis of Hospital Operational Efficiency" and structurally save the analysis framework and key conclusions.

[0027] Furthermore, during the semantic parsing and parameter extraction process, if the extracted parameters are found to be incomplete (for example, a user queries "query outpatient and emergency room visits" but does not specify a time range), the system will not directly report an error. Instead, it will initiate multi-round dialogue interaction through the large language model to guide the user to gradually supplement the necessary information until a natural language request with complete parameters is formed, thereby ensuring the smooth execution of the process.

[0028] The system can then call the data interface of the business intelligence (BI) system to execute the previously generated actionable analysis commands. These commands have been converted into a standardized format that the underlying BI system can parse, and the corresponding high-quality structured data is retrieved from the business intelligence system's data warehouse (BI data warehouse).

[0029] Subsequently, based on the initially identified intent type and the subsequently obtained structured data, the system generates a comprehensive analysis report that matches it. This report is not simply a list of data, but rather integrates visualization elements (such as charts and graphs) and textual analysis and explanations generated by a large language model. For example, for indicator queries, the report may include a bar chart and a textual description of the data trend; for attribution analysis, the report will list the dimensions that contribute the most and their values, and generate an attribution interpretation; for trend prediction, the report will display the prediction curve and confidence interval, along with textual explanations of future trends and potential risks.

[0030] Finally, the system returns and presents the generated comprehensive analysis report to the user through the user interface, completing a full intelligent analysis interaction loop. Users can also further interact with the report through the interface, such as drilling down to view detailed data, adjusting chart types, or saving it to a topic library for future reuse.

[0031] In summary, the natural language-based conversational business intelligence analysis method proposed in this invention fundamentally solves the pain points of traditional BI tools, namely high threshold, low efficiency, and weak analysis, by constructing a conversational business intelligence analysis system with intent recognition as the scheduling core and integrating multiple dedicated intelligent analysis modules.

[0032] Specifically, the system first uses a large language model to accurately understand the intent type of a user's natural language request, realizing the transformation from "humans adapting to machines" to "machines understanding humans." This allows users without a technical background to easily explore complex data, greatly lowering the barrier to entry. In the core indicator query function, a large language model based on LoRA fine-tuning is used for parameter extraction, combined with a hybrid Trie and BM25 retrieval scheme. This ensures accurate matching of technical terms while also providing tolerance and recommendation capabilities for colloquial and ambiguous expressions.

[0033] Furthermore, this invention achieves automated and quantitative intelligent attribution analysis by integrating the SHAP model. This enables rapid identification of key and specific dimensions of indicator fluctuations, transforming analysis work that previously relied on manual experience into objective data interpretation, significantly improving the efficiency and scientific rigor of problem diagnosis. In terms of trend prediction, the NexusTrendNet model, which incorporates external features (such as holidays and policies), is used to obtain accurate and interpretable prediction results, providing decision support for managers.

[0034] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0035] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0036] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0037] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. A conversational business intelligence analysis method based on natural language, characterized in that, Including the following steps: S1: Receive natural language requests from the current user in the conversation; S2: Identify the intent type of the user intent contained in the natural language request through a general large language model. The intent types include indicator query, attribution analysis, trend prediction, and topic generation. S3: Based on the specific intent type identified, call the corresponding dedicated analysis module to extract and semantically parse the natural language request, including indicators, dimensions, and time range, and generate operable analysis instructions; S4: Call the business intelligence system's data interface to execute analysis commands and obtain the corresponding structured data from the business intelligence system's data warehouse; S5: Based on the identified intent type and the corresponding structured data, generate a comprehensive analysis report that matches it and includes visualization elements and analytical explanations; S6: Return and present a comprehensive analysis report to the user.

2. The natural language-based conversational business intelligence analysis method as described in claim 1, characterized in that, In step S3, when the intent type is an indicator query, the dedicated analysis module is a parameter extraction and matching module. The parameter extraction and matching module uses a large language model based on LoRA fine-tuning to extract dynamic parameters and uses a hybrid retrieval scheme combining Trie tree and BM25 algorithm to match indicators, dimensions and time parameters.

3. The natural language-based conversational business intelligence analysis method as described in claim 2, characterized in that, The hybrid retrieval scheme first performs precise matching using a Trie tree. If the matching fails, it uses the BM25 algorithm to calculate similarity and make fuzzy recommendations. If the confidence of the fuzzy recommendation result is lower than a preset threshold, it calculates cosine similarity and returns a candidate list for the user to choose from.

4. The natural language-based conversational business intelligence analysis method as described in claim 1, characterized in that, In step S3, when the intent type is attribution analysis, the dedicated analysis module is an attribution analysis module based on the SHAP model, used to quantify the contribution of each dimension to the fluctuation of the specified indicator.

5. The natural language-based conversational business intelligence analysis method as described in claim 4, characterized in that, The execution process of the attribution analysis module includes: Calculate the SHAP value of all drill-down dimensions for the fluctuation of a specified indicator, and sort them by absolute value to select the top K dimensions with the largest contribution. For each of the first K selected dimensions, the key dimension value causing the indicator fluctuation is determined by calculating and sorting the SHAP value of all specific values ​​of the dimension.

6. The natural language-based conversational business intelligence analysis method as described in claim 1, characterized in that, In step S3, when the intent type is trend prediction, the dedicated analysis module is a time series prediction module, which performs prediction using a NexusTrendNet model that integrates trend items, seasonal items, and external feature items.

7. The natural language-based conversational business intelligence analysis method as described in claim 6, characterized in that, The external characteristics include one or more of the following: holidays, policy adjustments, and seasonal factors.

8. The natural language-based conversational business intelligence analysis method as described in claim 1, characterized in that, In step S3, when the intent type is topic generation, the dedicated analysis module is a general large language model, which summarizes and refines the current or historical dialogue content to generate a structured analysis topic and analysis framework.

9. The natural language-based conversational business intelligence analysis method as described in claim 1, characterized in that, If complete analysis parameters cannot be obtained through semantic parsing and parameter extraction in step S3, the method further includes the step of initiating a multi-round dialogue interaction through a general large language model to guide the user to supplement the necessary information until a natural language request that satisfies the information required for complete analysis parameters is formed.

Images