Target object analysis report generation method and device based on semantic compression, and medium

Abstract

This invention relates to the field of text processing and analysis technology, and in particular to a method, device, and medium for generating target object analysis reports based on semantic compression. The method includes: obtaining several historical descriptive texts of a target object from several preset data sources; obtaining several candidate text modules through semantic segmentation; calculating the indicator coverage rate of each candidate text module to a preset risk indicator set based on the current analysis rule set; selecting the candidate text module combination that maximizes the weighted indicator coverage rate under a preset token budget constraint; extracting several information fragments from the combination to form intermediate related texts; updating the intermediate related texts when the current rule set is updated; and generating a target object analysis report by combining the updated rule set and the generation timestamps of each historical descriptive text. This invention significantly reduces the computational cost and token consumption in rule update scenarios, while improving the quality and accuracy of target object analysis reports.

Description

Technical Field

[0001] This invention relates to the field of text processing and analysis technology, and in particular to a method, device and medium for generating target object analysis reports based on semantic compression. Background Technology

[0002] In fields such as financial risk control, corporate credit investigation, and public opinion monitoring, it is often necessary to generate analytical reports based on multi-source data of the target object. Traditional methods mainly rely on two technologies: one is rule-based keyword matching, which extracts information fragments related to the required indicators from the text; the other is direct summary generation based on large language models, which outputs analytical conclusions after inputting the full text into the model.

[0003] However, rule-based keyword matching struggles to handle the diverse expressions in natural language, easily leading to information omissions due to synonyms, sentence structure variations, and other factors. Furthermore, it cannot prioritize information based on its value. Large language models typically charge input and output tokens; directly inputting complete historical texts into the model incurs high computational costs when the text volume is large or the analysis frequency is high. Moreover, analysis rules are often dynamically updated based on business needs and changes in industry standards, requiring traditional methods to reprocess all historical texts, resulting in redundant computation and significant token waste. Therefore, how to reduce computational costs and token consumption in rule-updating scenarios while maintaining analysis quality, and simultaneously improve the accuracy of information extraction, has become a pressing issue in this field. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention provides a method, device, and medium for generating target object analysis reports based on semantic compression. This ensures the generation of high-information-density structured text within a limited token budget, significantly reducing computational costs and token consumption in rule update scenarios while effectively improving the quality and accuracy of target object analysis reports.

[0005] According to a first aspect of the present invention, a method for generating a target object analysis report based on semantic compression is provided, comprising the following steps: S1, obtain a set of historical description text information of the target object from several preset data sources; the set of historical description text information includes several historical description texts and the generation timestamp of each historical description text.

[0006] S2, for each historical description text, several candidate text modules are obtained through semantic segmentation, and corresponding keywords are extracted from each candidate text module based on the current analysis rule set to calculate the indicator coverage rate of each candidate text module to the preset risk indicator set; the current analysis rule set includes the preset risk indicator set and the evaluation rules for each preset risk indicator.

[0007] S3, under the preset token budget constraint, select the candidate text module combination that maximizes the coverage of weighted indicators, and extract several information fragments from them as the semantically compressed intermediate related text corresponding to the candidate text module combination.

[0008] S4. Establish an inverted index for information fragments and candidate text modules. When the current rule set is updated, locate the affected candidate text modules through the inverted index to extract incremental information fragments, or remove the corresponding information fragments from the intermediate related texts, and update the intermediate related texts and inverted indexes.

[0009] S5, based on the updated rule set, the updated intermediate related texts and their corresponding historical description texts, generates a target object analysis report for the preset risk indicator set through a preset large language model.

[0010] According to a second aspect of the present invention, a non-transitory computer-readable storage medium is provided, wherein at least one instruction or at least one program is stored therein, the at least one instruction or the at least one program being loaded and executed by a processor to implement the above-described method for generating target object analysis reports based on semantic compression.

[0011] According to a third aspect of the present invention, an electronic device is provided, including a processor and the aforementioned non-transitory computer-readable storage medium.

[0012] The present invention has at least the following beneficial effects: This invention provides a method for generating target object analysis reports based on semantic compression. First, it obtains several historical descriptive texts of the target object from several preset data sources. These texts are then semantically segmented to obtain several candidate text modules. Based on the current analysis rule set, the indicator coverage rate of each candidate text module to a preset risk indicator set is calculated. Semantic segmentation divides long texts into candidate text modules, ensuring thematic consistency within the same candidate text module. Furthermore, the indicator coverage calculation quantifies the information value of each candidate text module. Next, under a preset token budget constraint, a combination of candidate text modules with the highest weighted indicator coverage rate is selected. Several information fragments are extracted from these modules and combined into intermediate related texts, achieving an optimal balance between information coverage quality and compression efficiency. Different strategies are used to update the intermediate related texts when the current rule set is updated, avoiding redundant processing of the entire text and significantly reducing computational costs and token consumption in rule update scenarios. Finally, a target object analysis report is generated based on the updated intermediate related texts, the updated rule set, and the generation timestamps of each historical descriptive text. In summary, this invention significantly reduces computational costs and token consumption in rule update scenarios while improving the quality and accuracy of target object analysis reports. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart of a method for generating target object analysis reports based on semantic compression, provided in an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] This invention provides a method for generating target object analysis reports based on semantic compression, such as... Figure 1 As shown, the method includes the following steps: S1, obtain a set of historical description text information of the target object from several preset data sources; the set of historical description text information includes several historical description texts and the generation timestamp of each historical description text; it can be understood that the generation timestamp of the historical description text is the time when the text was published.

[0017] In one embodiment, when the target is a company, the preset data source includes, but is not limited to, internal corporate databases, publicly available financial reports, news media reports, regulatory agency announcements, and social media platforms. Different types of data sources can be pre-configured with data source levels based on their authority and credibility.

[0018] S2: For each historical descriptive text, several candidate text modules are obtained through semantic segmentation. Based on the current analysis rule set, corresponding keywords are extracted from each candidate text module to calculate the indicator coverage rate of each candidate text module to the preset risk indicator set. In specific implementation, if no keywords are extracted from the candidate text modules corresponding to the historical descriptive text, that historical descriptive text is not considered.

[0019] Specifically, the current set of analysis rules includes a set of preset risk indicators and evaluation rules for each preset risk indicator. For example, the set of preset risk indicators includes, but is not limited to, capacity utilization rate and debt-to-equity ratio. The evaluation rule for capacity utilization rate could be: if the capacity utilization rate is below 60%, the company is deemed to have insufficient capacity risk.

[0020] Specifically, for each historical description text, several candidate text modules are obtained through semantic segmentation, including the following steps: S201 divides the historical description text into several initial paragraphs according to the natural paragraph boundaries.

[0021] S202, calculate the semantic similarity between adjacent initial paragraphs. When the semantic similarity exceeds a preset similarity threshold, merge the adjacent initial paragraphs into the same candidate text module; otherwise, divide the adjacent initial paragraphs into different candidate text modules. Each candidate text module contains at least one initial paragraph. In specific implementation, when merging multiple consecutive initial paragraphs pairwise, multiple consecutive initial paragraphs are merged into the same candidate text module, that is, an initial paragraph exists in only one candidate text module.

[0022] Specifically, cosine similarity is used to calculate the semantic similarity between adjacent initial paragraphs; those skilled in the art know the specific algorithm of cosine similarity and can set the similarity threshold according to actual needs, so it will not be elaborated here.

[0023] As described above, by semantic segmentation, long texts are divided into candidate text modules, which enables the reasonable division of historical descriptive texts according to topic or event boundaries. This ensures that the content within the same candidate text module has thematic consistency, avoids the mixing of irrelevant information, reduces the granular complexity of subsequent information extraction, and lays a structured data foundation for the accurate calculation of subsequent indicator coverage and the selection of text modules.

[0024] Specifically, the step of extracting corresponding keywords from each candidate text module based on the current analysis rule set to calculate the indicator coverage rate of each candidate text module to the preset risk indicator set includes: S210: For any candidate text module, based on the preset risk indicators in the current analysis rule set, extract keywords from the candidate text module that are the same as or semantically consistent with any preset risk indicator, to obtain a keyword set. For example, the debt-to-asset ratio and the debt-to-asset ratio are semantically consistent keywords.

[0025] S220, calculate the ratio of the number of preset risk indicators covered by the obtained keyword set to the number of preset risk indicators in the preset risk indicator set, as the indicator coverage rate of the candidate text module itself to the preset risk indicator set; it can be understood that when the keywords and preset risk indicators are semantically consistent, they are also treated as coverage.

[0026] As mentioned above, by extracting keywords that are semantically consistent with the preset risk indicators, it is possible to accurately and comprehensively identify and extract key information related to risk analysis from candidate text modules. Furthermore, the calculation of indicator coverage enables a quantitative assessment of the information value of each candidate text module, providing an objective basis for subsequent candidate text module selection.

[0027] S3, under the preset token budget constraint, select the candidate text module combination that maximizes the coverage of weighted indicators, and extract several information fragments from them as the semantically compressed intermediate related text corresponding to the candidate text module combination.

[0028] In one specific embodiment, candidate text module combinations that maximize weighted index coverage are selected as follows: S310: For each historical description text, obtain its corresponding candidate text modules and determine the number of tokens for each candidate text module. The number of tokens is calculated using the token calculation rules of the GPT-4 model architecture.

[0029] S320, iterate through all candidate text module combinations that satisfy the preset token budget constraint, and calculate the weighted index coverage rate corresponding to each candidate text module combination; the weighted index coverage rate meets the following conditions: , of which S M To satisfy the weighted metric coverage of candidate text module combinations M that meet the preset token budget constraint, W m G represents the preset module weights of candidate text modules m in M. m T is the number of indicators that are covered only by candidate text modules m in M, η is the preset penalty coefficient, and T m is the token length of candidate text module m.

[0030] Furthermore, W m The accuracy of the indicator analysis for candidate text module m.

[0031] Specifically, the accuracy of the indicator analysis is calculated as follows: for any candidate text module, the ratio of the number of indicators correctly extracted from the candidate text module to the total number of indicators to be extracted from the preset risk indicator set is used as the accuracy of the indicator analysis for that candidate text module; where, correct extraction means that the name of the risk indicator in the extracted information fragment is consistent with the name of any preset risk indicator and the indicator value meets the preset format or numerical range requirements; the number of indicators to be extracted refers to the number of preset risk indicators that actually appear in the historical description text where the candidate text module is located.

[0032] S330: Select the candidate text module combination with the largest weighted index coverage value as the candidate text module combination with the maximum weighted index coverage.

[0033] As described above, by traversing all candidate text module combinations that meet the token budget constraint and combining the number of exclusive contribution indicators with the module weight for scoring, the global optimal selection of module combinations is achieved. The introduction of the number of exclusive contribution indicators ensures that each indicator is only scored by the module that uniquely covers it. The token penalty effectively controls the output length within the limited budget, achieving a quantifiable optimal balance between information coverage quality and compression efficiency.

[0034] Specifically, the intermediate related text is determined through the following steps: S301, calculate the compression ratio of each intermediate associated text; the compression ratio is represented by the ratio of the number of tokens in the intermediate associated text to the number of tokens in the corresponding historical description text.

[0035] S302, for each historical description text, when the compression ratio of the intermediate associated text corresponding to the historical description text is greater than a preset compression threshold, iteratively execute the reduction of the preset upper limit of the preset number range of information fragments in the information fragment set and / or the preset character number threshold corresponding to the information fragment by a preset adjustment step size to obtain new intermediate associated text, until the compression ratio of the new intermediate associated text is not greater than the preset compression threshold. Those skilled in the art can set the preset compression threshold according to actual needs, which will not be elaborated here.

[0036] Specifically, the information fragment includes at least one preset risk indicator and the corresponding indicator value.

[0037] Furthermore, a corresponding preset compression threshold is set for each historical description text. The preset compression threshold meets the following conditions: Where ε is the preset compression threshold, B is the preset total compression budget, e is the base of the natural logarithm, λ is the preset attenuation coefficient, and t i and t j Let t represent the unit time difference between the generation timestamp of the i-th historical descriptive text and the current time, and n be the total number of historical descriptive texts. This formula implements a time-aware compression budget allocation; that is, the newer the text, the higher the t value. i The smaller the value, the higher the allocated compression threshold, and the less compression is needed to retain more details.

[0038] As described above, this scheme iteratively adjusts the extraction parameters of information fragments, gradually converging the compression ratio of the intermediate related texts to within a preset compression threshold range, thus achieving dynamic control over compression accuracy. This adaptive compression method optimizes the temporal distribution of information density within a limited total token budget, thereby enhancing the information value of the analysis report.

[0039] In one embodiment, before generating the intermediate associated text, the method further includes: S10, determine whether there are cases in the extracted information segments where the preset risk indicators are consistent but the corresponding indicator values ​​conflict. In specific implementation, when the indicator value is a range, determine whether there is a conflict in the threshold corresponding to the range.

[0040] S20, if a conflict exists, calculate the time difference between the generation timestamps of the historical description text corresponding to each conflicting information fragment. If the time difference between the two most recent information fragments is greater than a preset time threshold, then retain the information fragment with the latest generation timestamp as a valid information fragment. Those skilled in the art can set the preset time threshold according to actual needs, for example, 3 days.

[0041] S30. If the time difference between the two latest information fragments is not greater than the preset time threshold, the comprehensive credibility score of the information fragment is calculated by weighting the preset data source level of the historical description text corresponding to the information fragment, the indicator analysis accuracy of the candidate text module corresponding to the information fragment, and the preset weights corresponding to the data source level and the analysis indicator accuracy, and the information fragment with the highest comprehensive credibility score is taken as the valid information fragment.

[0042] The above approach prioritizes time difference judgment to ensure that new information with significant timeliness can quickly cover old information. For conflicting information with similar timeliness, a weighted evaluation is performed by combining data source level and extraction accuracy, taking into account both authority and accuracy. Through a two-level conflict resolution method, the problem of inconsistency in information with multiple different values ​​corresponding to the same indicator is effectively solved, significantly improving the reliability of risk indicator information and providing an accurate data foundation for the subsequent generation of high-quality analysis reports.

[0043] S4. Establish an inverted index for information fragments and candidate text modules. When the current rule set is updated, locate the affected candidate text modules through the inverted index to extract incremental information fragments, or remove the corresponding information fragments from the intermediate associated text, and update the intermediate associated text and the inverted index. This can be understood as follows: For preset risk indicators with added or modified indicator values, find the candidate text modules containing the preset risk indicators with added or modified indicator values ​​in the inverted index, and use them as the affected candidate text modules; for preset risk indicators with deleted values, directly remove the information fragments corresponding to the preset risk indicators from the intermediate associated text.

[0044] Preferably, when storing intermediate associated text and inverted index, the version identifier of the current rule set is also stored.

[0045] As described above, different processing strategies can be adopted according to the changes in the rule set. When adding or modifying indicators, only the affected modules are incrementally extracted. When deleting indicators, the corresponding information fragments are directly extracted, avoiding the repeated processing of the entire text. This significantly reduces the computational cost and token consumption in rule update scenarios, while ensuring the consistency between the inverted index and intermediate related text with the latest rule set.

[0046] Furthermore, if preset risk indicators are added and / or modified during the current rule set update, this also includes: S401, calculate the indicator difference rate based on the current rule set and the updated rule set.

[0047] Specifically, the indicator difference rate meets the following condition: P=(C1+C2) / C0, where C0 is the number of preset risk indicators in the current rule set, C1 is the number of risk indicators added to the updated rule set relative to the current rule set, and C2 is the number of risk indicators whose evaluation rules have changed.

[0048] S402, when the index difference rate is less than the preset difference threshold, locate the affected candidate text module through the inverted index and perform incremental information fragment extraction; otherwise, return to execute S2 to S3 to rebuild the inverted index of information fragments and candidate text modules.

[0049] As described above, when the index difference rate is less than the threshold, the inverted index is used to locate the affected module for incremental extraction, avoiding full duplicate processing. When the index difference rate is not less than the threshold, it automatically falls back to the full processing mode, ensuring that the index and data are fully matched with the latest rule set. This achieves the best balance between efficiency and data quality, significantly reducing the computational cost and token consumption in rule update scenarios.

[0050] S5, based on the updated rule set, the updated intermediate related texts and their corresponding historical description texts, generates a target object analysis report for the preset risk indicator set through a preset large language model.

[0051] Specifically, the S5 steps include the following: S501, based on the updated rule set, extracts corresponding information fragments from each of the updated intermediate related texts, performs rule judgment, and then labels the target object with a risk tag. For example, when the evaluation rule of any preset risk indicator is met, the target object is labeled with a risk tag corresponding to that preset risk indicator.

[0052] S502, the updated intermediate related texts are concatenated according to the generation timestamps of the corresponding historical description texts, and duplicate risk indicators appearing during the concatenation process are deduplicated to generate structured text.

[0053] S503 inputs structured text and risk labels into a preset large language model, and generates a target object analysis report that includes risk assessment results corresponding to the risk labels for a preset set of risk indicators.

[0054] The above-mentioned new risk labels based on the updated rule set ensure the accuracy of information data. The collaborative input of risk labels and structured text effectively reduces the input token consumption of the large language model, while enabling the large language model to focus on the latest and most critical risk information, making the generated analysis report more reliable and timely.

[0055] Embodiments of the present invention also provide a non-transitory computer-readable storage medium that can be disposed in an electronic device to store at least one instruction or at least one program related to implementing a method in the method embodiments, wherein the at least one instruction or the at least one program is loaded and executed by the processor to implement the method provided in the above embodiments.

[0056] Embodiments of the present invention also provide an electronic device, including a processor and the aforementioned non-transitory computer-readable storage medium.

[0057] While specific embodiments of the invention have been described in detail by way of example, those skilled in the art should understand that the examples are for illustrative purposes only and not intended to limit the scope of the invention. It should also be understood that various modifications can be made to the embodiments without departing from the scope and spirit of the invention. The scope of the invention is defined by the appended claims.

Claims

1. A method for generating target object analysis reports based on semantic compression, characterized in that, The method includes the following steps: S1, Obtain a set of historical description text information of the target object from several preset data sources; the set of historical description text information includes several historical description texts and the generation timestamp of each historical description text; S2, for each historical description text, several candidate text modules are obtained through semantic segmentation, and corresponding keywords are extracted from each candidate text module based on the current analysis rule set to calculate the indicator coverage rate of each candidate text module to the preset risk indicator set; the current analysis rule set includes the preset risk indicator set and the evaluation rules for each preset risk indicator. S3, under the preset token budget constraint, select the candidate text module combination that maximizes the coverage of weighted indicators, and extract several information fragments from it as the semantically compressed intermediate related text corresponding to the candidate text module combination. S4. Establish an inverted index for information fragments and candidate text modules. When the current rule set is updated, locate the affected candidate text modules through the inverted index to extract incremental information fragments, or remove the corresponding information fragments from the intermediate associated text, and update the intermediate associated text and inverted index. S5, based on the updated rule set, the updated intermediate related texts and their corresponding historical description texts, generates a target object analysis report for the preset risk indicator set through a preset large language model.

2. The method for generating target object analysis reports based on semantic compression according to claim 1, characterized in that, In step S2, for each historical description text, several candidate text modules are obtained through semantic segmentation, including the following steps: S201, divide the historical description text into several initial paragraphs according to the natural paragraph boundaries; S202, calculate the semantic similarity between adjacent initial paragraphs. When the semantic similarity exceeds the preset similarity threshold, merge the adjacent initial paragraphs into the same candidate text module; otherwise, divide the adjacent initial paragraphs into different candidate text modules. Each candidate text module contains at least one initial paragraph.

3. The method for generating target object analysis reports based on semantic compression according to claim 1, characterized in that, The step of extracting corresponding keywords from each candidate text module based on the current analysis rule set to calculate the indicator coverage rate of each candidate text module to the preset risk indicator set includes: S210, For any candidate text module, based on the preset risk indicators in the current analysis rule set, extract keywords from the candidate text module that are the same as or semantically consistent with any preset risk indicator to obtain a keyword set; S220, calculate the ratio of the number of preset risk indicators covered by the obtained keyword set to the number of preset risk indicators in the preset risk indicator set, and use it as the indicator coverage rate of the candidate text module itself on the preset risk indicator set.

4. The method for generating target object analysis reports based on semantic compression according to claim 1, characterized in that, In step S3, the intermediate related text is determined through the following steps: S301, calculate the compression ratio of each intermediate associated text; the compression ratio is represented by the ratio of the number of tokens in the intermediate associated text to the number of tokens in the corresponding historical description text; S302, for each historical description text, when the compression ratio of the intermediate associated text corresponding to the historical description text is greater than the preset compression threshold, iteratively execute the reduction of the upper limit of the preset number range of information fragments in the information fragment set and / or the preset character number threshold corresponding to the information fragment by a preset adjustment step size to obtain new intermediate associated text, until the compression ratio of the new intermediate associated text is not greater than the preset compression threshold; wherein, the information fragment includes at least one preset risk indicator and the indicator value corresponding to the preset risk indicator.

5. The method for generating target object analysis reports based on semantic compression according to claim 1, characterized in that, The candidate text module combination that maximizes the weighted index coverage is selected using the following method: S310, For each historical description text, obtain its corresponding candidate text modules and determine the number of tokens for each candidate text module; S320, iterate through all candidate text module combinations that satisfy the preset token budget constraint, and calculate the weighted index coverage rate corresponding to each candidate text module combination; the weighted index coverage rate meets the following conditions: , of which S M To satisfy the weighted metric coverage of candidate text module combinations M that meet the preset token budget constraint, W m G represents the preset module weights of candidate text modules m in M. m T is the number of indicators that are covered only by candidate text modules m in M, η is the preset penalty coefficient, and T m The length of the token for candidate text module m; S330: Select the candidate text module combination with the largest weighted index coverage value as the candidate text module combination with the maximum weighted index coverage.

6. The method for generating target object analysis reports based on semantic compression according to claim 1, characterized in that, In step S4, if risk indicators are added and / or modified when the current rule set is updated, it also includes: S401, Calculate the indicator difference rate based on the current rule set and the updated rule set; wherein, the indicator difference rate meets the following condition: P=(C1+C2) / C0, where C0 is the number of preset risk indicators in the current rule set, C1 is the number of risk indicators added in the updated rule set relative to the current rule set, and C2 is the number of risk indicators whose evaluation rules have changed; S402, when the index difference rate is less than the preset difference threshold, locate the affected candidate text module through the inverted index and perform incremental information fragment extraction; otherwise, return to execute S2 to S3 to rebuild the inverted index of information fragments and candidate text modules.

7. The method for generating target object analysis reports based on semantic compression according to claim 1, characterized in that, The S5 steps include the following: S501, based on the updated rule set, extracts the corresponding information fragments from each of the updated intermediate related texts, and labels the target object with risk tags after rule judgment; S502, the updated intermediate related texts are concatenated according to the generation timestamps of the corresponding historical description texts, and duplicate risk indicators appearing during the concatenation process are deduplicated to generate structured text. S503 inputs structured text and risk labels into a preset large language model, and generates a target object analysis report that includes risk assessment results corresponding to the risk labels for a preset set of risk indicators.

8. The method for generating target object analysis reports based on semantic compression according to claim 1, characterized in that, The method further includes: S10, Determine whether there is a situation in the extracted information segment where the preset risk indicators are consistent but the corresponding indicator values ​​conflict; S20, if there is a conflict, calculate the time difference between the generation timestamps of the historical description text corresponding to each conflicting information fragment. If the time difference between the two latest information fragments is greater than the preset time threshold, retain the information fragment with the latest generation timestamp as the valid information fragment. S30. If the time difference between the two latest information fragments is not greater than the preset time threshold, the comprehensive credibility score of the information fragment is calculated by weighting the preset data source level of the historical description text corresponding to the information fragment, the indicator analysis accuracy of the candidate text module corresponding to the information fragment, and the preset weights corresponding to the data source level and the analysis indicator accuracy, and the information fragment with the highest comprehensive credibility score is taken as the valid information fragment.

9. A non-transitory computer-readable storage medium, wherein the storage medium stores at least one instruction or at least one program segment, characterized in that, The at least one instruction or the at least one program segment is loaded and executed by the processor to implement the target object analysis report generation method based on semantic compression as described in any one of claims 1-8.

10. An electronic device, characterized in that, Includes a processor and the non-transitory computer-readable storage medium as described in claim 9.

Images