Intelligent network information analysis method and system based on large model

Abstract

The invention discloses a network information intelligent analysis method and system based on a large model, and the method comprises the steps: firstly carrying out the preprocessing of network information, constructing a prompt word template through a multi-step thinking method, and processing the network information; calling the large model to obtain a network information analysis result presented by JSON format data; performing format judgment on the analysis result by using a feedback optimization method, and if the format requirement is not met, repeatedly calling the large model, and updating the network information analysis result; and performing JSON format matching on the analysis result of which the output format of the large model still does not meet the requirement to obtain an analysis result meeting the JSON template. According to the method, emotion analysis can be more accurately carried out on the network information, and the efficiency and accuracy of analysis work are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent analysis, and in particular to a network information intelligent analysis method and system based on a large model. Background Art

[0002] The formation and spread of online information is extremely fast, which can quickly affect market sentiment and investment decisions. Therefore, how to effectively monitor and analyze online information is of great significance to corporate organizations and regulatory agencies.

[0003] Traditional network information analysis methods mainly rely on manual monitoring and simple keyword matching technology, which have problems such as low efficiency and insufficient accuracy. With the development of big data and artificial intelligence technology, methods based on machine learning, especially deep learning, have gradually become a new trend in network information analysis. Among them, large models (such as Transformer, etc.) have shown significant advantages in analysis due to their powerful natural language processing capabilities.

[0004] However, existing large-model-based network information analysis methods still face some challenges. For example, large models usually require a lot of computing resources and have strict requirements on the format of input data, especially the length of input text is often limited. In addition, there are many professional terms in various fields, and ordinary large models may not be able to fully understand the meaning of these terms and their application in specific contexts. Therefore, it is particularly necessary to develop a network information intelligent analysis method and system that can effectively cope with the above challenges and fully utilize the advantages of large models. Summary of the invention

[0005] Purpose of the invention: In view of the above problems, the present invention proposes a method and system for intelligent analysis of network information based on a large model, aiming to achieve efficient and accurate information analysis through advanced data preprocessing technology, optimized model architecture and professional domain knowledge integration, so as to provide strong support for relevant decision-making.

[0006] Technical solution: To achieve the purpose of the present invention, the technical solution adopted by the present invention is: A network information intelligent analysis method based on a large model comprises the following steps: Step 1, preprocessing the network information to obtain preprocessed data information; Step 2, constructing a prompt word template through a multi-step thinking method, processing the data information obtained in step 1 to obtain a processed data block; Step 3: Input the processed data block into the big model to obtain the analysis results presented in JSON format, including validity judgment, sentiment analysis and keyword extraction; Step 4: Use the feedback optimization method to determine the format of the analysis results. If the format requirements are not met, the data block, analysis results and determination results in step 3 are input into the big model again to update the network information analysis results. Step 5: Perform JSON format matching on the analysis results that still do not meet the requirements after the output format of the large model in step 4, and obtain analysis results that meet the JSON template; Step 6: Based on the analysis results of the online information, determine whether the online information is valid or invalid, positive, negative or neutral, as well as the full text summary, keywords and reasons for judgment.

[0007] Preferably, data preprocessing is performed on the network news information, including: Determine the length of text information in online news; If the length of the text exceeds the limit of the number of tokens input by the large model, the text summary compression algorithm is used to preprocess the text so that the text information meets the requirements of the large model input text and obtain the preprocessed data information.

[0008] Preferably, the preprocessing is performed using a text summary compression algorithm, including: Generative summary method based on BART model. BART model compresses the text into summary version by encoding and decoding the input text. First, the encoder is used to perform semantic analysis and information extraction on the input text data, and the decoder is used to generate a compressed text summary.

[0009] Preferably, a prompt word template is constructed by a multi-step thinking method to obtain a processed data block, including: The multi-step thinking method splits the network news information analysis prompt word template into five steps: validity judgment, sentiment analysis, keyword generation, summary summary and judgment reason generation, so that the large model can be analyzed according to the process; (1) Effectiveness analysis: Determine whether the input content is relevant to the subject; If not, the content is judged as invalid information, and the judgment result and judgment reason are directly output; if relevant, continue to judge whether the content is related to the subject's event of concern; If the content is irrelevant to the event of interest or only related to the event of no interest, the content is judged as invalid information and the judgment result and judgment reason are directly output; otherwise, the content is judged as valid information and enters the next process; The subject, events of interest and events of non-interest are configured and modified in the prompt word template and the front-end interface of the network information intelligent analysis system; (2) Sentiment Analysis: For valid information related to the subject, semantic analysis is performed to determine the emotional tendency and evaluate the positive, negative or neutral sentiment of the text; (3) Keyword generation: Extract keywords that express the nature of the event, and determine the specific meaning of the keywords in the sentence based on the context; (4) Summary: Generate a summary, which includes the time, location, object and content of the event; (5) Generation of judgment reasons: Based on the judgment result, the judgment reason is generated.

[0010] Preferably, the format of the analysis result is determined by using a feedback optimization method, including: First, the default return result JSON format is as follows: is_effective: determines whether the article is effective, 1 for effective and 0 for ineffective; abstract: the extracted article abstract. Invalid content does not need to be extracted and is left blank. type: The sentiment tendency of the article analyzed, including: positive / negative / neutral. Invalid content does not need to be extracted and is left blank. hint_words: Extract no more than 5 keywords from the text. All keywords are concatenated together to form a string as the value of hint_words. Invalid content does not need to be extracted and is left blank. Reason: the reason for judging whether the article is effective and has emotional tendency; Compare the analysis results returned by the large model with the above JSON format to determine whether the analysis results conform to the predetermined JSON output format, check whether there are any fields missing, and finally generate a judgment result; If it is detected that the analysis result does not meet the predetermined format requirements, the data block, analysis result and judgment result in step 3 are input into the large model again for data generation; Set the upper limit of the number of times a single network information message can call the large model.

[0011] Preferably, JSON format matching includes: For the analysis results that do not conform to the JSON format that are ultimately returned by the large model, keyword matching is used to find each field value in the preset return result JSON template; then the field value is filled into the JSON format template. If the corresponding value cannot be found, it is set to empty, and finally the analysis results that conform to the JSON format are output.

[0012] The present invention also provides a network information intelligent analysis system based on a large model, which is used to implement the network information intelligent analysis method based on a large model, including: a data preprocessing module, a prompt word engineering module, a large model calling module, a feedback result optimization module, and an analysis result output module; Data preprocessing module: Use the generative summary algorithm based on the BART model to generate summaries for texts that exceed the input length of the large model; Prompt word engineering module: used to obtain processed data blocks through multi-step thinking prompt word templates for pre-processed data information. By customizing the prompt word template, that is, modifying the configuration of the subject, events of interest and events of no concern, it is adapted to event analysis in different fields. Large model calling module: used to input the data blocks processed by the prompt word engineering module into the large model to obtain analysis results, including validity judgment of network information, positive and negative sentiment analysis, keyword extraction, summary generation and judgment reason generation; Feedback optimization module: It is used to first check the content of the analysis results of the large model to check whether there are some missing values ​​in the results, and then determine whether the format meets the JSON format requirements. If so, the analysis results in JSON format are directly output; If not, call the big model again to generate the analysis result, set the upper limit of the number of times a single message calls the big model, and if the final result still does not conform to the JSON format, adjust its format through the JSON format matching module and output the analysis result; Analysis result output module: includes a JSON format matching module, which uses the field matching method to find the value of each field in the template for the results generated by the final large model that do not conform to the JSON format; if there is no value for a certain field in the result output by the large model, the value of the field is set to empty; finally, the corrected JSON format analysis result is returned.

[0013] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following significant advantages: (1) Improving analysis accuracy: By utilizing the advanced semantic understanding capabilities of the large model, the present invention can comprehensively analyze the content of the article, not just simple keyword matching, thereby greatly improving the accuracy of effectiveness judgment; (2) Enhanced sentiment recognition capabilities: Traditional methods rely on a preset negative vocabulary, which is prone to misjudgment or omission. The method of the present invention can better understand the context, accurately capture the sentiment tendency of the article, and reduce the misjudgment rate; (3) Strong adaptability: The large model has strong generalization ability and can adapt to text content in different fields and is suitable for a variety of application scenarios. The method proposed in this invention supports the configuration of subjects, events of interest and events of no interest, which further enhances its applicability and flexibility. (4) Reducing labor costs: By improving the accuracy of network information analysis, the present invention can significantly reduce reliance on manual review, saving time and human resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a flow chart of the method of the present invention.

[0015] Figure 2 It is a flow chart of the prompt word engineering module of the method of the present invention.

[0016] Figure 3 It is a flow chart of the feedback optimization module of the method of the present invention.

[0017] Figure 4 It is a framework diagram of the network information intelligent analysis system of the present invention. DETAILED DESCRIPTION

[0018] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0019] The present invention discloses a method for intelligent analysis of network information based on a large model, the process is as follows: Figure 1 As shown, it includes: S110, firstly, preprocessing the input network information to obtain the preprocessed data. Then, the text length of the input information is judged. For the super-long text that exceeds the limit of the number of tokens input to the large model, it is input into the summary compression module; for the text that does not exceed the limit, it is directly input into the prompt word engineering module. S120, extracting and compressing the information of the super-long text to obtain the processed data. S130, inputting the compressed or non-super-long text information into the prompt word module to generate a data format suitable for the input of the large model. S140, inputting the data processed by the prompt word template into the large model to obtain the analysis result. S150, format verification and content review of the analysis result output by the large model. If the format is found to be inconsistent or the content is missing, the network information text content, the large model analysis result and the judgment result are sent back to S130, and then input into the large model after being processed by the prompt word template again. This process is repeated up to three times. S160, if the final analysis result still has format problems in the above process, the result will be reconstructed and output through the JSON template matching code. If the analysis result meets expectations, the final analysis result is directly output.

[0020] First, it is necessary to preprocess the network information. The specific steps of preprocessing are: data cleaning, removing noise information such as HTML tags, advertising links, irrelevant symbols, and retaining pure text content; deduplication, deleting duplicate content to avoid redundant information during the analysis process; text length judgment, for long texts that exceed the input limit of the large model, use text summarization or compression algorithms to process them to ensure that all input data meets the requirements of the large model; format conversion and syntax analysis, convert the format of the original network information data, correct syntax errors, make the data more standardized and standardized, and facilitate subsequent analysis and processing.

[0021] The intelligent analysis method of network information mainly relies on large models to realize its core analysis function. However, due to the limitation of the performance of current large models, they set an upper limit on the length of the input text. In view of the fact that network information on the Internet platform often has super-long texts, after completing the data preprocessing, it is necessary to judge the length of the text. For texts that exceed the word limit of the large model, a summary compression algorithm is required. The summary compression algorithm selected by the present invention is a generative summary method based on the BART model. The BART model can compress super-long texts into shorter summary versions through its encoding-decoding mechanism. Specifically, the method first uses an encoder to perform in-depth semantic analysis and information extraction on the input text information, and then generates a refined text summary through a decoder to ensure that the summary content covers the key information of the original text. This generative summary method can not only efficiently process super-long texts, but also maintain the integrity and accuracy of the information, reduce the need for manual intervention, and thus improve the reliability of subsequent analysis results.

[0022] After preprocessing, the text information needs to go through the prompt word engineering module before it can be input into the big model. In this method, the constructed prompt word engineering module uses a multi-step thinking method to ensure that the format and content of the input data can better guide the big model to perform accurate information analysis.

[0023] like Figure 2 As shown in the figure, the specific functions of the prompt word engineering module include: automatically generating prompt word templates by combining specific keywords of the specified subject. This module automatically generates prompt word templates for the input of the large model by integrating the subject, events of interest and events of non-interest configured by the user. These prompt word templates can not only guide the large model to pay attention to the key information of the specified subject, but also ensure that the format of the input data matches the input requirements of the large model, thereby improving the accuracy and consistency of the analysis.

[0024] The processing logic of the prompt word engineering module uses a multi-step thinking method, as follows: the format of the input and output data is limited. The prompt word template not only specifies the format of the input data, but also clarifies the structure of the output data to ensure the standardization and interpretability of the analysis results. S131, validity analysis, conducts validity judgment on the preprocessed text data and identifies information related to the specified subject. This step aims to filter out irrelevant or low-quality data to ensure the accuracy and reliability of subsequent analysis. S132, sentiment analysis, analyzes the sentiment tendency of text information and determines whether it is positive, negative or neutral. Sentiment analysis is an important part of the analysis process, which can help the subject understand the changes in market sentiment and adjust strategies in time. S133, summary, generates summary based on full-text information. The summary can present the core content of the information concisely and clearly, making it easier for decision makers to quickly understand the situation. S134, keyword generation, extracts keywords that reflect the core content of the text data. The extraction of keywords helps to quickly grasp the focus of the information and provide a basis for subsequent in-depth analysis. S135, judgment reason generation, provides corresponding judgment reasons for the large model analysis results, explains the basis and reasoning process of the results, this step enhances the transparency and credibility of the analysis results, and enables decision makers to better understand and accept the analysis conclusions. S136, format limitation, limits the format of input and output data. The prompt word template not only specifies the format of the input data, but also clarifies the structure of the output data, ensuring the standardization and interpretability of the analysis results.

[0025] After the input text information is processed by the prompt word template, the large model is directly called and the processed information is input into the large model to obtain the analysis result. However, due to the uncontrollable output of the large model and the model capacity problem, the output analysis results may have problems such as missing content and inconsistent output format. Therefore, the output results need to be checked through feedback optimization methods.

[0026] Feedback optimization process such as Figure 3 As shown, first, the data processed by the prompt word template is input into the big model to generate an analysis result (S140), and then the analysis result output by the big model is obtained (S141), and then the format and content of the analysis result are checked (S150). If the content format of the analysis result conforms to the preset template, the result is directly output (S160). If not, the input of S140, the analysis result of S141 and the judgment result of S150 are input into the big model again to generate a new analysis result. In order to avoid repeated calls due to format problems, the maximum number of calls is set to 3 times. If the content format of the result still does not meet the requirements after the third call, the result is reconstructed by using the JSON format matching code to ensure that the output result conforms to the preset template before outputting (S161).

[0027] The preset return result JSON format is as follows: is_effective: determines whether the article is effective, 1 for effective and 0 for invalid; abstract: extracted article abstract, invalid content does not need to be extracted and is left blank; type: analyzed article sentiment, including: positive / negative / neutral, invalid content does not need to be extracted and is left blank; hint_words: extract no more than 5 keywords from the article, all keywords are concatenated together to form a string as the value of hint_words, invalid content does not need to be extracted and is left blank; reason: determines whether the article is effective and the reason for the sentiment; compare the analysis results returned by the large model with the above JSON format, determine whether the analysis results conform to the predetermined JSON output format, check whether there are any fields missing, and finally generate the judgment result.

[0028] In summary, the network information intelligent analysis method based on the big model of this embodiment is explained, which adopts the data analysis technology based on the big model to timely understand the user's emotional tendency towards the product by analyzing the relevant information of the designated subject on the Internet. Compared with the traditional keyword matching and manual analysis methods, this analysis method greatly improves the efficiency and objectivity of the analysis.

[0029] Figure 4 This is the framework diagram of the network information intelligent analysis system, such as Figure 4 As shown, the network information intelligent analysis system based on the big model according to this embodiment includes: a data preprocessing module 110, which is used to preprocess the input text data and perform summary processing on the text that exceeds the upper limit of the big model acceptance; a prompt word template module 120, which uses a multi-step thinking method to help the big model better answer questions and give fixed format requirements; a big model calling module 130, which inputs the processed data into the big model to obtain the analysis result; a feedback optimization module 140, which is used to check and judge the results output by the big model, and call the analysis results that do not meet the requirements again to regenerate the analysis results; an analysis result output module 150, which directly outputs the analysis results that meet the requirements, and outputs the results that do not meet the requirements after conversion in JSON format. JSON format matching includes: using keyword matching for the analysis results that do not meet the JSON format finally returned by the big model, finding each field value in the preset return result JSON template; then filling the field value into the JSON format template, if the corresponding value cannot be found, it is set to empty, and finally outputting the analysis results that meet the JSON format.

[0030] Here, those skilled in the art can understand that the specific functions and operations of each unit and module in the above-mentioned large model-based network information intelligent analysis system have been referred to above. Figures 1 to 3The method has been introduced in detail in the description of the large model-based network information intelligent analysis method, so its repeated description will be omitted.

[0031] In summary, the network information intelligent analysis system based on the big model is elaborated in detail. The system uses advanced artificial intelligence data analysis technology to process and analyze the information related to the specified subject on the Internet and generate analysis results using the big model. This method not only realizes the intelligence of data analysis, but also significantly improves the efficiency and objectivity of analysis.

Claims

1. A network information intelligent analysis method based on a large model, characterized in that: The following steps are involved: Step 1, preprocessing the network information to obtain preprocessed data information; Step 2, constructing a prompt word template through a multi-step thinking method, processing the data information obtained in step 1 to obtain a processed data block; Step 3: Input the processed data block into the big model to obtain the network information analysis results presented in JSON format, including validity judgment, sentiment analysis and keyword extraction; Step 4: Use the feedback optimization method to determine the format of the analysis results. If the format requirements are not met, the data block, analysis results and determination results in step 3 are input into the big model again to update the network information analysis results. Step 5: Perform JSON format matching on the analysis results that still do not meet the requirements after the output format of the large model in step 4, and obtain analysis results that meet the JSON template; Step 6: Based on the analysis results of the online information, determine whether the online information is valid or invalid, positive, negative or neutral, as well as the full text summary, keywords and reasons for judgment.

2. The method for intelligent analysis of network information based on a large model according to claim 1, characterized in that: Perform data preprocessing on network information, including: Determine the length of text information in online news; If the length of the text exceeds the limit of the number of tokens input by the large model, the text summary compression algorithm is used to preprocess the network information text so that the text information meets the requirements of the large model input text and obtains the preprocessed data information.

3. The method for intelligent analysis of network information based on a large model according to claim 2 is characterized in that: Preprocess using text summary compression algorithm, including: Generative summary method based on BART model. BART model compresses the text into summary version by encoding and decoding the input text. First, the encoder is used to perform semantic analysis and information extraction on the input text data, and the decoder is used to generate a compressed text summary.

4. The method for intelligent analysis of network information based on a large model according to claim 1, characterized in that: The prompt word template is constructed through a multi-step thinking method to obtain the processed data block, including: The multi-step thinking method splits the network information analysis prompt word template into five steps: validity judgment, sentiment analysis, keyword generation, summary summary and judgment reason generation, so that the large model can be analyzed according to the process; (1) Effectiveness analysis: Determine whether the input content is related to the main body of online news information; If not, the content is judged as invalid information, and the judgment result and judgment reason are directly output; if relevant, continue to judge whether the content is related to the subject's event of concern; If the content is irrelevant to the event of interest or only related to the event of no interest, the content is judged as invalid information and the judgment result and judgment reason are directly output; otherwise, the content is judged as valid information and enters the next process; The subject, events of interest and events of non-interest are configured and modified in the prompt word template and the front-end interface of the network information intelligent analysis system; (2) Sentiment Analysis: For valid information related to the subject, semantic analysis is performed to determine the emotional tendency and evaluate the positive, negative or neutral sentiment of the text; (3) Keyword generation: Extract keywords that express the nature of the event, and determine the specific meaning of the keywords in the sentence based on the context; (4) Summary: Generate a summary, which includes the time, location, object and content of the event; (5) Generation of judgment reasons: Based on the judgment result, the judgment reason is generated.

5. The method for intelligent analysis of network information based on a large model according to claim 1, characterized in that: The feedback optimization method is used to determine the format of the analysis results, including: First, the default return result JSON format is as follows: is_effective: determines whether the article is effective, 1 for effective and 0 for ineffective; abstract: the extracted article abstract. Invalid content does not need to be extracted and is left blank. type: The sentiment tendency of the article analyzed, including: positive / negative / neutral. Invalid content does not need to be extracted and is left blank. hint_words: Extract no more than 5 keywords from the text. All keywords are concatenated together to form a string as the value of hint_words. Invalid content does not need to be extracted and is left blank. Reason: the reason for judging whether the article is effective and has emotional tendency; Compare the analysis results returned by the large model with the above JSON format to determine whether the analysis results conform to the predetermined JSON output format, check whether there are any fields missing, and finally generate a judgment result; If it is detected that the analysis result generated by the large model does not meet the predetermined format requirements, the data block, analysis result and judgment result in step 3 are input into the large model again for data generation; Set the upper limit of the number of times a single network information message can call the large model.

6. The method for intelligent analysis of network information based on a large model according to claim 1, characterized in that: JSON format matches include: For the analysis results that do not conform to the JSON format that are ultimately returned by the large model, keyword matching is used to find each field value in the preset return result JSON template; then the field value is filled into the JSON format template. If the corresponding value cannot be found, it is set to empty, and finally the analysis results that conform to the JSON format are output.

7. A network information intelligent analysis system based on a large model, characterized in that: The system is used to implement the large model-based network information intelligent analysis method described in any one of claims 1 to 6, including: a data preprocessing module, a prompt word engineering module, a large model calling module, a feedback result optimization module, and an analysis result output module; Data preprocessing module: Use the generative summary algorithm based on the BART model to generate summaries for texts that exceed the input length of the large model; Prompt word engineering module: used to obtain processed data blocks through multi-step thinking prompt word templates for pre-processed data information. By customizing the prompt word template, that is, modifying the configuration of the subject, events of interest and events of no concern, it is adapted to the analysis of network information events in different fields. Large model calling module: used to input the data blocks processed by the prompt word engineering module into the large model to obtain analysis results, including validity judgment of network information, positive and negative sentiment analysis, keyword extraction, summary generation and judgment reason generation; Feedback optimization module: It is used to first check the content of the analysis results of the large model to check whether there are some missing values ​​in the results, and then determine whether the format meets the JSON format requirements. If so, the analysis results in JSON format are directly output; If not, call the big model again to generate the analysis result, set the upper limit of the number of times a single message calls the big model, and if the final result still does not conform to the JSON format, adjust its format through the JSON format matching module and output the analysis result; Analysis result output module: includes a JSON format matching module, which uses the field matching method to find the value of each field in the template for the results generated by the final large model that do not conform to the JSON format; if there is no value for a certain field in the result output by the large model, the value of the field is set to empty; finally, the corrected JSON format analysis result is returned.

Images