Data analysis method based on large-scale language model

Abstract

The invention provides a data analysis method based on a large-scale language model, and belongs to the technical field of data analysis. The method comprises the following steps: 1, selecting a target data source, calling the target data source by utilizing an interface API to obtain an original statement, and preprocessing to obtain a first text sequence; 2, screening important cue words to construct a cue word template according to historical statements of a historical library; 3, extracting all features of the first text sequence by using the cue word template, and screening all the features of the first text sequence to obtain important features; 4, taking the important features as a test set, taking historical statements as a training set, and training the initial language model to obtain a large-scale language model; and 5, configuring the large-scale language model to the production environment, analyzing and processing actual statements of the production environment by the large-scale language model to obtain an execution result, and judging an error between the execution result and an expected result. The data analysis method is completely realized through a natural language, and the analysis accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data analysis, and in particular to a data analysis method based on a large-scale language model. Background Art

[0002] At present, traditional methods in data analysis (such as data collection, chart generation, data prediction) generally use manually written SQL query statements, and make predictions on the acquired data based on pre-trained machine learning models, or generate results by performing chart analysis on the data according to preset templates in the data dashboard. This operation is inefficient, not flexible, requires a certain technical threshold, and consumes more human, material, and financial resources.

[0003] Therefore, the present invention proposes a data analysis method based on a large-scale language model. Summary of the invention

[0004] The present invention provides a data analysis method based on a large-scale language model, which is used for analyzing sentences based on natural language prompt words of users and system prompt words, and executing based on the large-scale language model to judge and feedback the execution results.

[0005] In one aspect, the present invention provides a data analysis method based on a large-scale language model, comprising: Step 1: Select a target data source, use the interface API to call the target data source to obtain the original sentence, and obtain the first text sequence after preprocessing the original sentence; Step 2: Filter important prompt words according to the historical sentences in the history library, and build prompt word templates based on the important prompt words; Step 3: extract all features of the first text sequence using the prompt word template, and screen all features of the first text sequence to obtain important features; Step 4: Use important features as test sets and historical sentences as training sets to train the initial language model and obtain a large-scale language model; Step 5: Configure the large-scale language model to the production environment. The large-scale language model analyzes and processes the actual statements in the production environment to obtain execution results, and determines the error between the execution results and the expected results.

[0006] On the other hand, select the target data source and use the interface API to call to obtain the original statement, including: Determine the text domain to be acquired based on data analysis requirements, and select the corresponding target data source based on the text domain; Build the API interface service of the target data source, and call the target data source based on the interface to obtain the original statement.

[0007] On the other hand, after preprocessing the original sentence, the first text sequence is obtained, including: Restore the data format of the original sentence obtained by the interface to obtain the original format text; Remove abnormal characters and HTML tags from the original format text, and convert the original format text into a standard text based on the standard text type; The standard text is segmented using a natural language processing library to obtain a standard segmentation sequence, and stop words in the standard segmentation sequence are removed to obtain a first text sequence.

[0008] On the other hand, according to the historical sentences in the history library, important prompt words are screened, including: According to the preset time screening conditions, historical sentences in the history library are obtained, all historical sentences are segmented according to the preset word segmentation tool, the historical word segmentation sequence is obtained after the stop words are filtered out, and the importance of any word in the historical word segmentation sequence is obtained; ;in represents the importance of word t, ) represents the preset inverse document weight coefficient of vocabulary t, Indicates the number of occurrences of word t in the historical word segmentation sequence, Indicates that there are N historical word segmentation sequences in total. represents the filter coefficient of vocabulary t; For all vocabulary Sort by values ​​and filter out words that are higher than the preset filter threshold. Words with values ​​higher than the preset screening threshold are considered important prompt words.

[0009] On the other hand, a prompt word template is constructed based on important prompt words, including: Constructing a formatted prompt word template structure, inserting important prompt words into the prompt word template structure according to a preset format, and generating a prompt word template; The storage fields of the database are preset, a prompt word database is constructed, and the prompt word templates are stored in the prompt word database.

[0010] On the other hand, all features of the first text sequence are extracted using the prompt word template, and all features of the first text sequence are screened to obtain important features, including: Using a mutual information algorithm to match the vocabulary of the first text sequence with the vocabulary matching degree of the prompt word template; ;in, represents the matching degree between the word x of the first text sequence and the prompt word y of the prompt word template, n represents that the first text sequence has n words in total, m represents that the prompt word template has m prompt words in total, P(x) represents the marginal probability of x, P(y) represents the marginal probability of y, and P(x,y) represents the joint probability of x and y. represents the i-th word in the first text sequence, represents the marginal probability of the i-th word in the first text sequence; According to the vocabulary matching degree of any vocabulary of the first text sequence and any prompt word template, taking the vocabulary with vocabulary matching degree higher than the preset expectation as the feature of the first text sequence; Using the LDA algorithm to analyze the topic distribution of the first text sequence, and converting the features of the first text sequence into a feature vector; The importance scores of the feature vectors are screened according to a preset topic value threshold, and feature vectors with a value greater than the preset topic value threshold are considered important features, thereby obtaining the important features of the first text sequence.

[0011] On the other hand, the important features are used as the test set and the historical sentences are used as the training set to train the initial language model and obtain a large-scale language model, including: The important features are used as the test set, the historical sentences are used as the training set, and the number of vocabulary categories of the historical sentences is extracted as the model temperature parameter; The initial language model framework is constructed and initialized using the model temperature parameters, and the initial language model is trained using the training set to obtain the first language model; The first language model is evaluated using important features to obtain first language model parameters, the analysis deviation of the first language model is analyzed using a loss function, and the first language model parameters are adjusted based on an optimizer to obtain a large-scale language model and package it to obtain a model package file.

[0012] On the other hand, the large-scale language model is configured in the production environment. The large-scale language model analyzes and processes the actual statements in the production environment to obtain the execution results, and determines the error between the execution results and the expected results, including: Deploy the model package file of the large-scale language model to the production environment, wherein the large-scale language model configuration parameters are parameters set when sending HTTP requests to the large-scale language model through the production environment; Large-scale language models analyze production cases to obtain actual statements, and then execute the actual statements in the production environment according to the running types of the actual statements to obtain execution results; Compare the expected results and execution results of the production case to obtain the error value and error source; Reconfigure and redeploy parameters of large-scale language models based on error values ​​and error sources.

[0013] Compared with the prior art, the present invention has the following beneficial effects: The present invention provides a data analysis method based on a large-scale language model, which is used for analyzing sentences based on natural language prompt words of users and system prompt words, and executing based on the large-scale language model to judge and feedback the execution results. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 It is a flowchart of a data analysis method based on a large-scale language model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] Embodiment 1: like Figure 1 As shown, an embodiment of the present invention provides a data analysis method based on a large-scale language model, including: Step 1: Select a target data source, use the interface API to call the target data source to obtain the original sentence, and obtain the first text sequence after preprocessing the original sentence; Step 2: Filter important prompt words according to the historical sentences in the history library, and build prompt word templates based on the important prompt words; Step 3: extract all features of the first text sequence using the prompt word template, and screen all features of the first text sequence to obtain important features; Step 4: Use important features as test sets and historical sentences as training sets to train the initial language model and obtain a large-scale language model; Step 5: Configure the large-scale language model to the production environment. The large-scale language model analyzes and processes the actual statements in the production environment to obtain execution results, and determines the error between the execution results and the expected results.

[0018] In this embodiment, the target data source refers to a specific data source used to obtain the original statement, which may be a database, a file system, an external API, a Web service, or other data storage system.

[0019] In this embodiment, the interface API is a set of rules and protocols that define the interaction between software components, allowing different software applications to communicate with each other and share data.

[0020] In this embodiment, the original sentence refers to unprocessed initial text data obtained from the target data source.

[0021] In this embodiment, the first text sequence refers to a text data sequence obtained by preprocessing the original sentence.

[0022] In this embodiment, the history library refers to a database or data set that stores past statements or data records.

[0023] In this embodiment, the historical statements refer to past data records or text contents stored in the history library, such as SQL statements, image description statements, data processing statements, etc.

[0024] In this embodiment, the important prompt words refer to keyword phrases that can significantly affect the performance of the model in a specific context.

[0025] In this embodiment, the prompt word template is a structured framework format for understanding and processing input text in a specific task.

[0026] In this embodiment, features refer to key attributes extracted from text data for representing and analyzing data, such as word frequency features, sentence length, sentiment intensity, etc.

[0027] In this embodiment, important features refer to attributes that can have a significant impact on the prediction and classification results of the model after being screened in the text data.

[0028] In this embodiment, the test set is a set of data used to evaluate and verify the performance of the model.

[0029] In this embodiment, the training set is a set of data used to train a large-scale language model.

[0030] In this embodiment, the large-scale language model refers to a model trained using deep learning technology to understand, generate and operate natural language text.

[0031] In this embodiment, the production environment refers to a real, actual running environment in which applications, systems or services are used in daily operations.

[0032] In this embodiment, the execution result refers to the output content after analysis and processing by the large-scale language model.

[0033] In this embodiment, the expected result refers to the target output benefit obtained from the entire process.

[0034] In this embodiment, the actual sentence refers to the natural language text that needs to be analyzed and processed in a specific application scenario, such as "automatically generate a response or suggestion by analyzing the sentence.".

[0035] The working principle and beneficial effects of the above technical solution are: through data source extraction, prompt word screening, feature extraction and screening, model training and evaluation, a closed-loop process is formed to significantly improve text analysis accuracy and model performance, ensuring effectiveness and reliability in practical applications.

[0036] Embodiment 2: On the basis of the above-mentioned embodiment 1, the target data source is selected, and the original statement is obtained by calling the interface API, including: Determine the text domain to be acquired based on data analysis requirements, and select the corresponding target data source based on the text domain; Build the API interface service of the target data source, and call the target data source based on the interface to obtain the original statement.

[0037] In this embodiment, data analysis requirements refer to problems, goals, and expected results that need to be solved in business or research.

[0038] In this embodiment, the text domain refers to the scope of text types or contents that the analysis focuses on, such as social media, customer feedback, product reviews, etc.

[0039] The working principle and beneficial effects of the above technical solution are: by determining the text field, selecting the data source, and building the API interface, the original text data is automatically acquired, which improves the accuracy and efficiency of data acquisition and ensures the real-time and scalability of the analysis.

[0040] Embodiment 3: On the basis of the above-mentioned embodiment 2, the first text sequence is obtained after preprocessing the original sentence, including: Restore the data format of the original sentence obtained by the interface to obtain the original format text; Remove abnormal characters and HTML tags from the original format text, and convert the original format text into a standard text based on the standard text type; The standard text is segmented using a natural language processing library to obtain a standard segmentation sequence, and stop words in the standard segmentation sequence are removed to obtain a first text sequence.

[0041] In this embodiment, the data format refers to the structure of data organization and storage in a computer system, such as json, character, int, etc.

[0042] In this embodiment, the original format text is the initial data obtained from the interface, which is usually organized and stored in a specific data format (such as JSON, XML, etc.).

[0043] In this embodiment, abnormal characters refer to those characters that do not meet the expectations and may interfere with the analysis results, such as line breaks, tab characters, etc.

[0044] In this embodiment, HTML tags are a markup language used to build web pages, including: <html>, <body>, etc.

[0045] In this embodiment, the standard text type refers to a text format that has been preprocessed, normalized, and conforms to a specific structure.

[0046] In this embodiment, standard text refers to text that has been processed and normalized.

[0047] In this embodiment, the natural language processing library is a tool and software package for processing and analyzing natural language text, including operations such as preprocessing, word segmentation, syntactic analysis, sentiment analysis, entity recognition, etc.

[0048] In this embodiment, word segmentation refers to splitting a text string into individual words or sub-words.

[0049] In this embodiment, stop words* refer to common words that are usually considered to make no significant contribution to analysis or modeling in text processing, such as: "de", "shi", "zai", "he", etc.

[0050] In this embodiment, the first text sequence refers to the final form of the text data obtained after a series of text processing steps.

[0051] The working principle and beneficial effects of the above technical solution are: by restoring the original format, processing abnormal characters, converting the format, and performing word segmentation, the text quality and standardization are improved, a concise text sequence is generated, and high-quality basic data is provided for subsequent analysis.

[0052] Embodiment 4: Based on the above Embodiment 1, important prompt words are screened according to the historical statements in the historical library, including: According to the preset time screening conditions, the historical statements in the historical library are obtained, all historical statements are segmented according to the preset word segmentation tool, the stop words are filtered out to obtain the historical segmentation sequence, and the importance degree of any vocabulary in the historical segmentation sequence is obtained; ; where represents the importance degree of vocabulary t, ) represents the preset inverse document weight coefficient of vocabulary t, Indicates the number of occurrences of word t in the historical word segmentation sequence, Indicates that there are N historical word segmentation sequences in total. represents the filter coefficient of vocabulary t; For all vocabulary Sort by values ​​and filter out words that are higher than the preset filter threshold. Words with values ​​higher than the preset screening threshold are important prompt words.

[0053] In this embodiment, the preset time screening condition refers to the data range and time period selected when analyzing historical data.

[0054] In this embodiment, the historical word segmentation sequence is a vocabulary sequence extracted from the historical text and obtained after word segmentation processing.

[0055] In this embodiment, importance is an indicator used to measure the importance of a word relative to other words in text analysis.

[0056] In this embodiment, the screening threshold is preset The working principle and beneficial effects of the above technical solution are: through historical sentence segmentation and importance calculation, key prompt words are screened out, the efficiency of information extraction and the level of intelligent analysis are improved, and decision-making and data processing are supported.

[0057] Embodiment 5: On the basis of the above-mentioned embodiment 4, a prompt word template is constructed based on the important prompt words, including: Constructing a formatted prompt word template structure, inserting important prompt words into the prompt word template structure according to a preset format, and generating a prompt word template; The storage fields of the database are preset, a prompt word database is constructed, and the prompt word templates are stored in the prompt word database.

[0058] In this embodiment, the preset format refers to a data structure or layout format that has been defined in a specific scenario or application.

[0059] In this embodiment, the prompt word database is a database for storing prompt word templates, and the fields include information such as task type, data source, and important prompt words.

[0060] The working principle and beneficial effects of the above technical solution are: constructing a formatted prompt word template, inserting important prompt words and storing them in the database, ensuring that the information is structured, easy to query, and highly consistent, and supporting subsequent intelligent applications and data analysis.

[0061] Embodiment 6: On the basis of the above-mentioned embodiment 1, all features of the first text sequence are extracted using the prompt word template, and all features of the first text sequence are screened to obtain important features, including: Using a mutual information algorithm to match the vocabulary of the first text sequence with the vocabulary matching degree of the prompt word template; ;in, represents the matching degree between the word x of the first text sequence and the prompt word y of the prompt word template, n represents that the first text sequence has n words in total, m represents that the prompt word template has m prompt words in total, P(x) represents the marginal probability of x, P(y) represents the marginal probability of y, and P(x,y) represents the joint probability of x and y. represents the i-th word in the first text sequence, represents the marginal probability of the i-th word in the first text sequence; According to the vocabulary matching degree of any vocabulary of the first text sequence and any prompt word template, taking the vocabulary with vocabulary matching degree higher than the preset expectation as the feature of the first text sequence; Using the LDA algorithm to analyze the topic distribution of the first text sequence, and converting the features of the first text sequence into a feature vector; The importance scores of the feature vectors are screened according to a preset topic value threshold, and feature vectors with a value greater than the preset topic value threshold are considered important features, thereby obtaining the important features of the first text sequence.

[0062] In this embodiment, the mutual information algorithm is a statistic used to quantify the mutual dependence between two random variables, and measures the degree to which the information of one variable can reduce the uncertainty of another variable.

[0063] In this embodiment, the vocabulary matching degree is an indicator used to measure the relevance between two text sequences.

[0064] In this embodiment, the preset expectation refers to a set of pre-set standard thresholds.

[0065] In this embodiment, the marginal probability refers to the probability of a certain random variable alone.

[0066] In this embodiment, the preset topic value threshold is a standard value for filtering the importance score of a feature.

[0067] In this embodiment, the feature vector is a way of numerically representing a data object (such as text, image, audio, etc.).

[0068] In this embodiment, the importance score is used to evaluate the influence of certain features (such as vocabulary or feature vectors) on a specific task (such as classification, clustering or topic modeling).

[0069] The working principle and beneficial effects of the above technical solution are: using mutual information and LDA algorithm to extract text features and analyze topics, screen important features, and improve the accuracy and intelligence of text analysis.

[0070] Embodiment 7: On the basis of the above-mentioned embodiment 1, the important features are used as the test set and the historical sentences are used as the training set to train the initial language model to obtain a large-scale language model, including: The important features are used as the test set, the historical sentences are used as the training set, and the number of vocabulary categories of the historical sentences is extracted as the model temperature parameter; The initial language model framework is constructed and initialized using the model temperature parameters, and the initial language model is trained using the training set to obtain the first language model; The first language model is evaluated using important features to obtain first language model parameters, the analysis deviation of the first language model is analyzed using a loss function, and the first language model parameters are adjusted based on an optimizer to obtain a large-scale language model and package it to obtain a model package file.

[0071] In this embodiment, the vocabulary category refers to a set of vocabulary classified according to rules.

[0072] In this embodiment, the model temperature parameter is an important hyperparameter for controlling the randomness and diversity of the model output.

[0073] In this embodiment, initialization refers to setting the initial parameter state of the model before training begins.

[0074] In this embodiment, the initial language model refers to a basic model obtained in the early stage of building and training the language model.

[0075] In this embodiment, the first language model is a model that has been preliminarily trained and optimized.

[0076] In this embodiment, the first language model parameter refers to a set of adjustable values ​​obtained during the training and evaluation process, including: weight parameters, bias parameters, temperature parameters, etc.

[0077] In this embodiment, the evaluation is to substitute important features into the first language model, perform cross-validation, and obtain the optimal parameter configuration of the model.

[0078] In this embodiment, the loss function is a function used to measure the gap between the model prediction value and the true value.

[0079] In this embodiment, the analysis deviation is the prediction error of the model during the training and testing process.

[0080] In this embodiment, the model packaging file refers to saving the trained model and its related information (such as parameters, configuration, metadata, etc.) in a packaged format.

[0081] The working principle and beneficial effects of the above technical solution are: by extracting features and model temperature parameters, building and optimizing the language model, improving training efficiency and accuracy, and finally generating an easy-to-use large-scale language model package file.

[0082] Embodiment 8: On the basis of the above-mentioned embodiment 1, a large-scale language model is configured in a production environment, and the large-scale language model analyzes and processes actual statements in the production environment to obtain execution results, and determines the error between the execution results and the expected results, including: Deploy the model package file of the large-scale language model to the production environment, wherein the large-scale language model configuration parameters are parameters set when sending HTTP requests to the large-scale language model through the production environment; Large-scale language models analyze production cases to obtain actual statements, and then execute the actual statements in the production environment according to the running types of the actual statements to obtain execution results; Compare the expected results and execution results of the production case to obtain the error value and error source; Reconfigure and redeploy parameters of large-scale language models based on error values ​​and error sources.

[0083] In this embodiment, the HTTP request is a request sent via HTTP (Hypertext Transfer Protocol) for exchanging data between a client and a server.

[0084] In this embodiment, the production case refers to an instance in which the model is used to perform tasks or process specific data in an actual application environment.

[0085] In this embodiment, the actual statement refers to a specific instruction obtained by analyzing a production case.

[0086] In this embodiment, the operation type refers to the specific task performed by the large-scale language model when executing an actual sentence.

[0087] In this embodiment, business comparison refers to the process of comparing actual execution results with expected results at the business level.

[0088] In this embodiment, the error source refers to various factors that cause the difference between the actual execution result and the expected result, such as: input quality, model parameters, data representativeness, etc.

[0089] In this embodiment, the error value is an indicator for measuring the gap between the actual output of the model and the expected output.

[0090] The working principle and beneficial effects of the above technical solution are: dynamically configure large-scale language models through HTTP requests, execute actual statements and analyze errors, continuously optimize model parameters, and improve accuracy and adaptability in production environments.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data analysis method based on a large-scale language model, characterized in that: include: Step 1: Select a target data source, use the interface API to call the target data source to obtain the original sentence, and obtain the first text sequence after preprocessing the original sentence; Step 2: Filter important prompt words according to the historical sentences in the history library, and build prompt word templates based on the important prompt words; Step 3: extract all features of the first text sequence using the prompt word template, and screen all features of the first text sequence to obtain important features; Step 4: Use important features as test sets and historical sentences as training sets to train the initial language model and obtain a large-scale language model; Step 5: Configure the large-scale language model to the production environment. The large-scale language model analyzes and processes the actual statements in the production environment to obtain execution results, and determines the error between the execution results and the expected results.

2. The data analysis method based on a large-scale language model according to claim 1, characterized in that: Select the target data source and use the interface API to obtain the original statement, including: Determine the text domain to be acquired based on data analysis requirements, and select the corresponding target data source based on the text domain; Build the API interface service of the target data source, and call the target data source based on the interface to obtain the original statement.

3. A data analysis method based on a large-scale language model according to claim 2, characterized in that: After preprocessing the original sentence, the first text sequence is obtained, including: Restore the data format of the original sentence obtained by the interface to obtain the original format text; Remove abnormal characters and HTML tags from the original format text, and convert the original format text into a standard text based on the standard text type; The standard text is segmented using a natural language processing library to obtain a standard segmentation sequence, and stop words in the standard segmentation sequence are removed to obtain a first text sequence.

4. The data analysis method based on a large-scale language model according to claim 1, characterized in that: According to the historical sentences in the history library, filter out important prompt words, including: According to the preset time screening conditions, historical sentences in the history library are obtained, all historical sentences are segmented according to the preset word segmentation tool, the historical word segmentation sequence is obtained after the stop words are filtered out, and the importance of any word in the historical word segmentation sequence is obtained; ;in represents the importance of word t, ) represents the preset inverse document weight coefficient of vocabulary t, Indicates the number of occurrences of word t in the historical word segmentation sequence, Indicates that there are N historical word segmentation sequences in total. represents the filter coefficient of vocabulary t; For all words Sort by values ​​and filter out words that are higher than the preset filter threshold. Words with values ​​higher than the preset screening threshold are considered important prompt words.

5. The data analysis method based on a large-scale language model according to claim 4, characterized in that: Build a prompt word template based on important prompt words, including: Constructing a formatted prompt word template structure, inserting important prompt words into the prompt word template structure according to a preset format, and generating a prompt word template; The storage fields of the database are preset, a prompt word database is constructed, and the prompt word templates are stored in the prompt word database.

6. The data analysis method based on a large-scale language model according to claim 1, characterized in that: All features of the first text sequence are extracted using the prompt word template, and all features of the first text sequence are screened to obtain important features, including: Using a mutual information algorithm to match the vocabulary of the first text sequence with the vocabulary matching degree of the prompt word template; ;in, represents the matching degree between the word x of the first text sequence and the prompt word y of the prompt word template, n represents that the first text sequence has n words in total, m represents that the prompt word template has m prompt words in total, P(x) represents the marginal probability of x, P(y) represents the marginal probability of y, and P(x,y) represents the joint probability of x and y. represents the i-th word in the first text sequence, represents the marginal probability of the i-th word in the first text sequence; According to the vocabulary matching degree of any vocabulary of the first text sequence and any prompt word template, taking the vocabulary with a vocabulary matching degree higher than a preset expectation as the feature of the first text sequence; Using the LDA algorithm to analyze the topic distribution of the first text sequence, and converting the features of the first text sequence into a feature vector; The importance scores of the feature vectors are screened according to a preset topic value threshold, and feature vectors with importance scores greater than the preset topic value threshold are considered important features, thereby obtaining the important features of the first text sequence.

7. The data analysis method based on a large-scale language model according to claim 1, characterized in that: The important features are used as the test set and the historical sentences are used as the training set to train the initial language model and obtain a large-scale language model, including: The important features are used as the test set, the historical sentences are used as the training set, and the number of vocabulary categories of the historical sentences is extracted as the model temperature parameter; The initial language model framework is constructed and initialized using the model temperature parameters, and the initial language model is trained using the training set to obtain the first language model; The first language model is evaluated using important features to obtain first language model parameters, the analysis deviation of the first language model is analyzed using a loss function, and the first language model parameters are adjusted based on an optimizer to obtain a large-scale language model and package it to obtain a model package file.

8. The data analysis method based on a large-scale language model according to claim 1, characterized in that: The large-scale language model is configured in the production environment. The large-scale language model analyzes and processes the actual statements in the production environment to obtain the execution results, and determines the error between the execution results and the expected results, including: Deploy the model package file of the large-scale language model to the production environment, wherein the large-scale language model configuration parameters are parameters set when sending HTTP requests to the large-scale language model through the production environment; Large-scale language models analyze production cases to obtain actual statements, and then execute the actual statements in the production environment according to the running types of the actual statements to obtain execution results; Compare the expected results and execution results of the production case to obtain the error value and error source; Reconfigure and redeploy parameters of large-scale language models based on error values ​​and error sources.

Images