AI-based customer admission auxiliary auditing method and system

Abstract

The embodiment of the invention provides an AI-based customer admission auxiliary auditing method and system, and belongs to the technical field of data processing. The method comprises the steps of obtaining initial structured information and initial text information of a target building company; performing paragraph splitting on the initial text information to obtain a target paragraph title and associated text information corresponding to the target paragraph Performing abstract generation processing on the associated text information to obtain target abstract information; performing statement recognition on the target summary information to obtain a target key statement; determining a first key feature vector by using the target paragraph title, the target summary information and the target key statement; performing key feature identification on the initial structured information to obtain target structured information of the target building company; obtaining a second key feature vector by using the target structured information; determining a target risk identification result according to an artificial intelligence algorithm in combination with the first key feature vector and the second key feature vector; and determining an auxiliary auditing result of the target building company according to the target risk identification result.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to an AI-based method and system for assisting in the review of customer access. Background Art

[0002] When financing in the construction industry, comprehensively reviewing the qualifications, financial status, and business capabilities of the target construction company is a key step to ensure the safety and compliance of financing. Although currently, artificial intelligence (AI) technology can be used for assisted review to improve review efficiency and reduce labor costs, when dealing with unstructured texts (such as contracts, project reports, news, customer evaluations, etc.), due to the lack of in-depth understanding of the context and the ability to accurately extract key information, existing AI technologies may miss important risk points or make incorrect judgments, thus weakening the reliability of the assisted review. These problems ultimately affect the overall review effect and further prevent the full potential of AI in improving review efficiency and quality from being realized. Summary of the Invention

[0003] The main purpose of the embodiments of the present invention is to provide an AI-based method and system for assisting in the review of customer access, aiming to solve the problem in related technologies that the reliability of the assisted review is reduced due to the lack of in-depth understanding of the context and the ability to accurately extract key information.

[0004] In a first aspect, the embodiments of the present invention provide an AI-based method for assisting in the review of customer access, including:

[0005] Obtaining initial structured information and initial text information required for the access review of the target construction company;

[0006] Performing paragraph breakdown on the initial text information to obtain corresponding target paragraph titles and associated text information corresponding to the target paragraph titles;

[0007] Performing summary generation processing on the associated text information to obtain target summary information corresponding to the target paragraph titles;

[0008] Performing key sentence recognition on the target summary information to obtain target key sentences corresponding to the target paragraph titles;

[0009] Using the target paragraph titles, the target summary information, and the target key sentences to determine a first key feature vector corresponding to the target construction company under the initial text information;

[0010] Performing key feature recognition on the initial structured information to obtain target structured information corresponding to the target construction company;

[0011] Obtain the second key feature vector corresponding to the target construction company under the initial structured information by using the target structured information;

[0012] Determine the target risk identification result corresponding to the target construction company according to the artificial intelligence algorithm in combination with the first key feature vector and the second key feature vector;

[0013] Determine the auxiliary review result corresponding to the target construction company according to the target risk identification result.

[0014] In a second aspect, an AI-based customer access auxiliary review system provided by an embodiment of the present invention includes:

[0015] A data acquisition module, configured to obtain the initial structured information and the initial text information required for the target construction company during access review;

[0016] A data disassembling module, configured to disassemble the initial text information into paragraphs to obtain the corresponding target paragraph titles and the associated text information corresponding to the target paragraph titles;

[0017] An abstract generation module, configured to perform abstract generation processing on the associated text information to obtain the target abstract information corresponding to the target paragraph title;

[0018] A statement recognition module, configured to perform key statement recognition on the target abstract information to obtain the target key statements corresponding to the target paragraph title;

[0019] A first vector analysis module, configured to determine the first key feature vector corresponding to the target construction company under the initial text information by using the target paragraph title, the target abstract information, and the target key statements;

[0020] A feature recognition module, configured to perform key feature recognition on the initial structured information to obtain the target structured information corresponding to the target construction company;

[0021] A second vector analysis module, configured to obtain the second key feature vector corresponding to the target construction company under the initial structured information by using the target structured information;

[0022] A risk recognition module, configured to determine the target risk identification result corresponding to the target construction company according to the artificial intelligence algorithm in combination with the first key feature vector and the second key feature vector;

[0023] A result determination module, configured to determine the auxiliary review result corresponding to the target construction company according to the target risk identification result.

[0024] An embodiment of the present invention provides an AI-based method and system for assisting in the admission review of customers. The method includes: obtaining initial structured information and initial text information required for the admission review of a target construction company; disassembling the initial text information into paragraphs to obtain corresponding target paragraph titles and associated text information corresponding to the target paragraph titles; performing summary generation processing on the associated text information to obtain target summary information corresponding to the target paragraph titles; identifying key sentences in the target summary information to obtain target key sentences corresponding to the target paragraph titles; using the target paragraph titles, target summary information, and target key sentences to determine a first key feature vector corresponding to the target construction company under the initial text information; identifying key features in the initial structured information to obtain target structured information corresponding to the target construction company; using the target structured information to obtain a second key feature vector corresponding to the target construction company under the initial structured information; determining a target risk identification result corresponding to the target construction company according to an artificial intelligence algorithm in combination with the first key feature vector and the second key feature vector; and determining an auxiliary review result corresponding to the target construction company according to the target risk identification result. By disassembling paragraphs, generating summaries, and identifying key sentences in the initial text information, the method can quickly identify the core content in the initial text information. Then, after obtaining the key feature target structured information in the initial structured information, combining the core content in the initial text information with the target structured information can more accurately identify the potential risks of the target construction company. This multi-dimensional data analysis method can more comprehensively reflect the actual situation of the company than a single information source, and thus the auxiliary review result obtained based on multi-dimensional data analysis and artificial intelligence algorithms improves the accuracy and reliability of the auxiliary review result. It also solves the problem in the related art that the reliability of the auxiliary review is reduced due to the lack of in-depth understanding of the context and the ability to accurately extract key information. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0026] Figure 1 It is a schematic flowchart of an AI-based method for assisting in the admission review of customers provided by an embodiment of the present invention;

[0027] Figure 2 It is a schematic module structure diagram of an AI-based system for assisting in the admission review of customers provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0029] The flowcharts shown in the accompanying drawings are only illustrative examples, and do not necessarily include all the contents and operations / steps, nor do they necessarily need to be executed in the described order. For example, some operations / steps can be decomposed, combined, or partially merged, so the actual execution order may change according to the actual situation.

[0030] It should be understood that the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an", and "the" are intended to include the plural forms.

[0031] The embodiments of the present invention provide an AI-based customer access auxiliary review method and system. Among them, the AI-based customer access auxiliary review method can be applied to a terminal device, which can be an electronic device such as a tablet computer, a notebook computer, a desktop computer, a personal digital assistant, and a wearable device. The terminal device can be a server or a server cluster.

[0032] The following will describe in detail some embodiments of the present invention with reference to the accompanying drawings. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0033] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of an AI-based customer access auxiliary review method provided by an embodiment of the present invention.

[0034] As Figure 1 shown, the AI-based customer access auxiliary review method includes steps S101 to S109.

[0035] Step S101: Obtain the initial structured information and initial text information required for the target construction company during the access review.

[0036] Exemplarily, the initial structured information and initial text information required for the access review of the target construction company are obtained from the database. The initial structured information is standardized and formatted data, such as company basic information, financial statements, qualification certificates, etc. The initial text information is unstructured text content, such as application descriptions, project reports, business descriptions, historical records, contract terms, etc. submitted by the company.

[0037] In some embodiments, the obtaining of the initial structured information and initial text information required for the access review of the target construction company includes: determining the identity information, and when the access permission corresponding to the identity information meets the preset permission, obtaining the initial encrypted text required for the access review of the target construction company; obtaining the decryption key according to the identity information, and decrypting the initial encrypted text according to the decryption key to obtain the initial structured information and the initial text information.

[0038] Exemplarily, when obtaining data from the database, it is necessary to first input the identity information to verify the identity information, and then when the verification is passed, obtain the access permission corresponding to the identity information, so that when the access permission is greater than or equal to the preset permission, obtain the initial encrypted text required for the access review of the target construction company. When the identity information verification fails or the access permission is less than the preset permission, the initial encrypted text required for the access review of the target construction company cannot be obtained.

[0039] Exemplarily, obtain the decryption key by obtaining the identity information from the database, and then decrypt the initial encrypted text according to the decryption key to obtain the initial structured information and the initial text information.

[0040] Step S102: Decompose the initial text information into paragraphs to obtain the corresponding target paragraph titles and the associated text information corresponding to the target paragraph titles.

[0041] Exemplarily, the paragraphs in the initial text information are separated by line breaks. Therefore, different paragraphs can be identified according to consecutive line breaks. In some cases, there may be no obvious line breaks between paragraphs, and in this case, punctuation marks such as full stops and semicolons can be used to assist in judging the paragraph boundaries, so as to segment the initial text information according to the punctuation marks or line breaks to obtain the paragraph information corresponding to the initial text information.

[0042] Exemplarily, recognition rules are written for the font size and style of the initial text information in combination with the text position, so as to obtain the corresponding target paragraph title in the initial text information according to the recognition rules. Furthermore, the paragraph information between the current target paragraph title and the next target paragraph title of the current target paragraph title is determined as the associated text information corresponding to the current target paragraph title. That is, each target paragraph title is associated with the paragraph information immediately following it until the next target paragraph title appears. Thus, the initial text information can be effectively disassembled into paragraphs, and the target paragraph title and its corresponding associated text information can be extracted, providing a basis for subsequent processing such as abstract generation and key sentence recognition.

[0043] Step S103: Perform abstract generation processing on the associated text information to obtain the target abstract information corresponding to the target paragraph title.

[0044] Exemplarily, the associated text information is segmented into sentences, and each sentence is tokenized. Furthermore, common stop words such as "de", "shi", "zai", etc. are removed to reduce noise. Then, part-of-speech tagging is performed on the words, and named entities such as person names, place names, and organization names are identified, which usually have relatively high importance. Next, the word frequency or TF-IDF value is calculated, and the words or phrases with higher frequencies are selected as keywords. Furthermore, the sentences containing these keywords are selected as the target abstract information. Or an encoder-decoder architecture, such as a model based on RNN, LSTM, or Transformer, or a pre-trained model such as BART or PEGASUS is used and fine-tuned to adapt to the abstract generation task of this application, so as to obtain the target abstract information corresponding to the target paragraph title.

[0045] In some embodiments, obtaining the target summary information corresponding to the target paragraph title by performing summary generation processing on the associated text information includes: obtaining an initial text vector by performing vector representation on the associated text information using the text representation layer of the summary generation model; obtaining a dependency syntactic analysis result by performing dependency syntactic analysis on the associated text information using the syntactic analysis layer of the summary generation model; obtaining a target text vector by performing syntactic enhancement on the initial text vector according to the syntactic analysis result using the text enhancement layer of the summary generation model; obtaining data distribution information corresponding to the associated text information by performing vocabulary distribution analysis on the target text vector using the attention distribution layer of the summary generation model; generating an initial summary information corresponding to the associated text information by using the fusion pointer generation network layer of the summary generation model according to the target text vector and the data distribution information; obtaining a target score corresponding to the initial summary information by performing quality calculation on the initial summary information using the summary quality evaluation layer of the summary generation model; and obtaining the target summary information corresponding to the target paragraph title by adjusting the initial summary information according to the target score.

[0046] Exemplarily, the summary generation model includes a text representation layer, a syntactic analysis layer, a text enhancement layer, an attention distribution layer, a fusion pointer generation network layer, and a summary quality evaluation layer.

[0047] Exemplarily, before obtaining the target summary information corresponding to the target paragraph title by performing summary generation processing on the associated text information using the summary generation model, multiple text information and the summary information corresponding to each text information are first obtained, so as to use the text information and the summary information as training data to perform model training on the summary generation model, and then obtain a summary generation model that meets the main quality requirements of model generation.

[0048] Exemplarily, the text representation layer of the summary generation model can be BERT, RoBERTa, DistilBERT, etc., and then the associated text information is input into the text representation layer to obtain the initial text vector corresponding to the associated text information.

[0049] Exemplarily, dependency syntactic analysis is performed on the associated text information using the syntactic analysis layer of the summary generation model to obtain a dependency syntactic analysis result. For example, the syntactic analysis layer is such as Stanford CoreNLP, Spacy, AllenNLP, etc., and then the associated text information is input into the syntactic analysis layer to obtain the dependency syntactic analysis result corresponding to the associated text information. The dependency syntactic analysis result includes the dependency relationship and syntactic role of each word.

[0050] Exemplarily, the text enhancement layer of the abstract generation model performs syntactic enhancement on the initial text vector according to the syntactic analysis result to obtain the target text vector. According to the syntactic analysis result, an enhancement strategy is designed. For example, vectors representing core vocabulary and key dependency relationships can be enhanced, and thus the enhancement strategy is applied to the initial text vector to generate the target text vector. The specific method can be weighted average, feature splicing, etc.

[0051] Exemplarily, the attention distribution layer of the abstract generation model performs vocabulary distribution analysis on the target text vector to obtain the data distribution information corresponding to the associated text information. For example, an attention mechanism is selected, such as self-attention, bidirectional attention, etc. The target text vector is input into the attention mechanism, and then the vocabulary distribution information is extracted from the output of the attention mechanism, which is usually a probability distribution indicating the importance of each word, so as to determine the data distribution information corresponding to the associated text information according to the vocabulary distribution information.

[0052] Exemplarily, the fusion pointer generation network layer of the abstract generation model generates the initial abstract information corresponding to the associated text information according to the target text vector and the data distribution information. For example, the fusion pointer generation network layer is a Pointer-Generator network, which combines the pointer mechanism and the generation mechanism and can select important words from the input text and generate new sentences. Thus, the target text vector and the data distribution information are input into the generation network, and then the initial abstract information is obtained from the generation network.

[0053] Exemplarily, the abstract quality evaluation layer of the abstract generation model calculates the quality of the initial abstract information to obtain the target score corresponding to the initial abstract information. For example, using abstract quality evaluation metrics such as ROUGE, BLEU, METEOR, etc., the initial abstract information is compared with the initial text information to calculate the quality score of the initial abstract information, and then the quality score is determined as the target score corresponding to the initial abstract information.

[0054] Exemplarily, the initial abstract information is adjusted according to the target score to obtain the target abstract information corresponding to the target paragraph title. For example, if the target score is lower than the preset score, the fusion pointer generation network layer is re-run to try to generate different abstracts until the abstract with the highest quality or the abstract with a target score greater than the preset score is selected as the target abstract information corresponding to the target paragraph title.

[0055] In some embodiments, obtaining the target score corresponding to the initial summary information by calculating the quality of the initial summary information using the summary quality evaluation layer of the summary generation model includes: obtaining the target keywords corresponding to the associated text information and the target relationships between the target keywords, and obtaining the first text information corresponding to the target keywords and the target relationships in the associated text information; using the summary quality evaluation layer to perform information analysis on the initial summary information to obtain the summary keywords corresponding to the initial summary information and the summary relationships between the summary keywords, and obtaining the second text information corresponding to the summary keywords and the summary relationships in the initial summary information; using the summary quality evaluation layer to determine the first loss value corresponding to the initial summary information according to the target keywords, the target relationships, the summary keywords, and the summary relationships; using the summary quality evaluation layer to fuse the first loss value according to the first text information and the second text information to obtain the second loss value corresponding to the initial summary information; using the summary quality evaluation layer to perform score conversion according to the second loss value to obtain the target score corresponding to the initial summary information.

[0056] Exemplarily, the summary quality evaluation layer uses keyword extraction techniques (such as TF-IDF, TextRank, RAKE, etc.) to obtain the target keywords corresponding to the associated text information, and then uses relationship extraction techniques (such as dependency syntactic analysis, named entity relationship extraction, etc.) to extract the target relationships between the target keywords from the associated text information, thereby recording the context text in which the target keywords and the target relationships appear in the associated text information, that is, the first text information.

[0057] Exemplarily, the summary quality evaluation layer uses keyword extraction techniques (such as TF-IDF, TextRank, RAKE, etc.) to perform information analysis on the initial summary information to obtain the summary keywords corresponding to the initial summary information, and then uses relationship extraction techniques (such as dependency syntactic analysis, named entity relationship extraction, etc.) to obtain the summary relationships between the summary keywords from the initial summary information, thereby recording the context text in which the summary keywords and the summary relationships appear in the initial summary information, that is, the second text information.

[0058] Exemplarily, the abstract quality assessment layer uses methods such as Jaccard similarity and cosine similarity to compare the first similarity value between the target keyword and the abstract keyword, and then determines the mapping relationship between the target keyword and the abstract keyword according to the first similarity value, so as to obtain the mapping word corresponding to the target keyword in the abstract keyword. Furthermore, the first association relationship between the mapping words corresponding to the target keyword in the abstract keyword is obtained from the target relationship, and then the second association relationship corresponding to the mapping word in the abstract relationship is obtained. Thus, the second similarity between the first association relationship and the second association relationship is calculated using methods such as edit distance and structural similarity, and then the first loss value corresponding to the initial abstract information is determined according to the second similarity.

[0059] Exemplarily, methods such as cosine similarity and BERT are used to compare the third similarity value between the first text information and the second text information, and then the third similarity value and the first loss value are fused using a weighted average method to determine the second loss value corresponding to the initial abstract information.

[0060] Exemplarily, the abstract quality assessment layer performs score conversion according to the second loss value to obtain the target score corresponding to the initial abstract information. For example, the second loss value is converted into the target score. Simple linear conversion or non-linear conversion methods can be used. The target score can be normalized to a fixed range (such as 0 to 1) for easy evaluation and comparison.

[0061] Specifically, through the extraction of keywords and relationships, the core information and structure of the text can be captured more accurately, providing a basis for subsequent comparison and evaluation. This application not only compares keywords, but also compares the relationships between keywords and context information to ensure the integrity and accuracy of the abstract. Furthermore, through the loss calculation in multiple dimensions (keywords, relationships, context), the quality of the abstract can be evaluated more comprehensively.

[0062] In some embodiments, the determining, by the summary quality evaluation layer, of the first loss value corresponding to the initial summary information according to the target keyword, the target relationship, the summary keyword, and the summary relationship includes: using the summary quality evaluation layer to perform data comparison on the target keyword, the target relationship, the summary keyword, and the summary relationship to obtain a comparison result; when the comparison result is that the target keyword and the summary keyword are similar, and the target relationship and the summary relationship are similar, determining that the target characterization value corresponding to the target keyword, the target relationship, the summary keyword, and the summary relationship is a first value; when the comparison result is that the target keyword and the summary keyword are similar, but the target relationship and the summary relationship are not similar, determining that the target characterization value corresponding to the target keyword, the target relationship, the summary keyword, and the summary relationship is a second value; when the comparison result is that the target keyword and the summary keyword are not similar, but the target relationship and the summary relationship are similar, determining that the target characterization value corresponding to the target keyword, the target relationship, the summary keyword, and the summary relationship is a third value; when the comparison result is that the target keyword and the summary keyword are not similar, and the target relationship and the summary relationship are not similar, determining that the target characterization value corresponding to the target keyword, the target relationship, the summary keyword, and the summary relationship is a fourth value; using the summary quality evaluation layer to fuse the first value, the second value, the third value, and the fourth value to determine the first loss value corresponding to the initial summary information; wherein, the first loss value is obtained according to the following formula:

[0063]

[0064] Wherein, represents the first loss value corresponding to the i-th initial summary information, represents the first value, m represents the quantity corresponding to all the first values determined under all the comparison results, represents the second value, n represents the quantity corresponding to all the second values determined under all the comparison results, represents the third value, k represents the quantity corresponding to all the third values determined under all the comparison results, represents the fourth value, h represents the quantity corresponding to all the fourth values determined under all the comparison results, and lg represents the logarithmic function with base 10.

[0065] Exemplarily, methods such as Jaccard similarity and cosine similarity are used to compare the first similarity value between the target keyword and the abstract keyword. Then, based on the first similarity value, the mapping relationship between the target keyword and the abstract keyword is determined, so as to obtain the mapping word corresponding to the target keyword in the abstract keyword. Furthermore, the first association relationship between the mapping words corresponding to the target keyword in the abstract keyword is obtained from the target relationship. Then, the second association relationship corresponding to the mapping word in the abstract relationship is obtained, and thus the second similarity value between the first association relationship and the second association relationship is calculated using methods such as edit distance and structural similarity.

[0066] Exemplarily, a first similarity threshold between the target keyword and the abstract keyword and a second similarity threshold between the target relationship and the abstract relationship are set. Then, when the first similarity value is greater than or equal to the first similarity threshold, it is determined that the comparison result is that the target keyword and the abstract keyword are similar; when the second similarity value is greater than or equal to the second similarity threshold, it is determined that the comparison result is that the target relationship and the abstract relationship are similar. Furthermore, the target characterization value among the target keyword, the abstract keyword, the target relationship, and the abstract relationship involved under the first similarity value and the second similarity value is the first numerical value. For example, the first numerical value is set to 1.

[0067] Exemplarily, when the first similarity value is greater than or equal to the first similarity threshold, it is determined that the comparison result is that the target keyword and the abstract keyword are similar; when the second similarity value is less than the second similarity threshold, it is determined that the comparison result is that the target relationship and the abstract relationship are not similar. Furthermore, the target characterization value among the target keyword, the abstract keyword, the target relationship, and the abstract relationship involved under the first similarity value and the second similarity value is the second numerical value. For example, the second numerical value is set to 0.1.

[0068] Exemplarily, when the first similarity value is less than the first similarity threshold, it is determined that the comparison result is that the target keyword and the abstract keyword are not similar; when the second similarity value is greater than or equal to the second similarity threshold, it is determined that the comparison result is that the target relationship and the abstract relationship are similar. Furthermore, the target characterization value among the target keyword, the abstract keyword, the target relationship, and the abstract relationship involved under the first similarity value and the second similarity value is the third numerical value. For example, the third numerical value is set to 0.01.

[0069] Exemplarily, when the first similarity value is less than the first similarity threshold, it is determined that the comparison result is that the target keyword and the abstract keyword are not similar; when the second similarity value is less than the second similarity threshold, it is determined that the comparison result is that the target relationship and the abstract relationship are not similar. Furthermore, the target characterization value among the target keyword, the abstract keyword, the target relationship, and the abstract relationship involved under the first similarity value and the second similarity value is the fourth numerical value. For example, the fourth numerical value is set to 0.

[0070] It should be noted that in this application, the first value is greater than the second value, the second value is greater than the third value, and the third value is greater than the fourth value.

[0071] Exemplarily, the abstract quality assessment layer fuses the first value, the second value, the third value, and the fourth value according to the following formula, and further determines the first loss value corresponding to the initial abstract information:

[0072]

[0073] Where, represents the first loss value corresponding to the i-th initial abstract information, represents the first value, m represents the quantity corresponding to all the first values determined under all comparison results, represents the second value, n represents the quantity corresponding to all the second values determined under all comparison results, represents the third value, k represents the quantity corresponding to all the third values determined under all comparison results, represents the fourth value, h represents the quantity corresponding to all the fourth values determined under all comparison results, and lg represents the logarithmic function with base 10.

[0074] Exemplarily, by fusing four different target characterization values, the similarity of keywords and relationships is comprehensively considered, providing a more comprehensive evaluation. The logarithmic function can amplify small differences, enabling fine distinctions to be made even when the loss values are close. Furthermore, by taking the logarithm, the influence of extreme values can be reduced, making the loss values more stable and reliable.

[0075] Step S104: Identify the key sentences of the target abstract information to obtain the target key sentences corresponding to the target paragraph title.

[0076] Exemplarily, identify sentences with specific structures, such as compound sentences, complex sentences, or sentences containing important grammatical structures (such as conditional sentences, causal sentences) to obtain the initial key sentences. Then, mark the sentences directly related to the target paragraph title or containing key information from the initial key sentences, and further determine the sentence as the target key sentence corresponding to the target paragraph title.

[0077] In some embodiments, identifying key sentences from the target summary information to obtain the target key sentences corresponding to the target paragraph title includes: segmenting the target summary information into target segmented sentences, and calculating a first similarity between the target paragraph title and the target segmented sentences; extracting sentence keywords from the target segmented sentences, and calculating a second similarity between the sentence keywords and the target paragraph title; determining the position information corresponding to the target segmented sentences, and determining a first important representation value corresponding to the target segmented sentences according to the position information; determining a preset summary word, and determining a second important representation value corresponding to the target segmented sentences according to the preset summary word; fusing the first similarity, the second similarity, the first important representation value, and the second important representation value to determine a sentence score corresponding to the target segmented sentences; and determining the target key sentences corresponding to the target paragraph title from the target segmented sentences according to the sentence score.

[0078] Exemplarily, the target summary information is segmented into individual sentences according to punctuation marks (such as full stops, question marks, exclamation marks, etc.) to form target segmented sentences. And ensure that the segmented target segmented sentences are complete and do not contain unreasonable sentence breaks.

[0079] Exemplarily, a method based on keyword matching, lexical overlap, or semantic similarity is used to compare the target paragraph title and the target segmented sentences one by one, and then calculate the high-frequency lexical overlap degree related to the paragraph title in the target segmented sentences, or use synonym expansion to measure semantic similarity, so as to quantify the comparison result into a numerical value representing the similarity degree between the target paragraph title and each target segmented sentence, and then obtain the first similarity.

[0080] Exemplarily, a keyword recognition algorithm is used to extract the core keywords of each target segmented sentence. Commonly used keywords can be nouns, verbs, or high-frequency words. Then, the extracted keywords are compared with the target paragraph title by using keyword matching degree or semantic similarity (such as similarity based on synonyms or word vectors) to calculate the second similarity between the keywords and the paragraph title.

[0081] Exemplarily, determine the position of each target segmented sentence in the target summary information (for example, the relative position of the sentence in the paragraph, whether it is the first sentence, the middle sentence, or the last sentence), and then assign different weights according to the position information. For example: the first sentence usually contains the main idea of the paragraph and has a higher weight. The middle sentence and the last sentence are assigned decreasing weights according to their roles in the paragraph, so as to obtain the first important representation value corresponding to the target segmented sentences.

[0082] Exemplarily, a set of commonly used preset summary words such as "in summary", "in conclusion", "therefore", etc. are determined. These words are usually used to introduce key summary sentences. If the target segmentation sentence contains a preset summary word, a higher weight is assigned; if not, a lower weight is assigned, so as to obtain the second important representation value corresponding to the target segmentation sentence.

[0083] Exemplarily, the first similarity, the second similarity, the first important representation value, and the second important representation value are weighted and summed according to certain weights to obtain the sentence score of each target segmentation sentence. Then, the sentence scores of all target segmentation sentences are sorted, and several sentences with the highest scores are selected as the target key sentences. If multiple key sentences are needed, a score threshold can be set, and the sentences with scores higher than the threshold are selected. If only one key sentence is needed, the sentence with the highest score is selected.

[0084] Specifically, by fusing multi-dimensional indicators such as similarity, position information, and summary words, the importance of each sentence can be evaluated more comprehensively. Furthermore, the target key sentences corresponding to the target paragraph title can be efficiently extracted from the target summary information, and their accuracy and representativeness can be ensured.

[0085] Step S105: Determine the first key feature vector corresponding to the target construction company under the initial text information by using the target paragraph title, the target summary information, and the target key sentences.

[0086] Exemplarily, the text representation model is used to obtain the corresponding representation vectors of the target paragraph title, the target summary information, and the target key sentences through sentence representation respectively, and then the representation vectors are fused to obtain the first key feature vector corresponding to the target construction company under the initial text information.

[0087] Step S106: Perform key feature recognition on the initial structured information to obtain the target structured information corresponding to the target construction company.

[0088] Exemplarily, corresponding preset key features are set according to experts or historical experience, and then the feature distance between the initial structured information and the preset key features is calculated. Thus, when the feature distance is less than or equal to the preset distance, the initial structured information corresponding to the feature distance is determined as the target structured information corresponding to the target construction company; when the feature distance is greater than the preset distance, the initial structured information corresponding to the feature distance is deleted.

[0089] In some embodiments, obtaining the target structured information corresponding to the target construction company by performing key feature recognition on the initial structured information includes: calculating similarity information for each first sub-structured information in the initial structured information to obtain the corresponding similarity quantity of the first sub-structured information; determining the feature entropy corresponding to the first sub-structured information according to the similarity quantity, and determining the feature distribution value corresponding to the first sub-structured information according to the feature entropy; calculating the distribution difference value between any two of the first sub-structured information according to the feature distribution value; clustering the first sub-structured information according to the distribution difference value to obtain a target clustering result; calculating the target distance between the second sub-structured information in each cluster in the target clustering result and the remaining structured information in the cluster; obtaining the target structured information corresponding to the target construction company from the initial structured information; wherein, the distribution difference value is obtained according to the following formula:

[0090]

[0091] wherein, represents the distribution difference value between the i-th first sub-structured information and the j-th first sub-structured information; represents the feature distribution value corresponding to the i-th first sub-structured information; represents the feature distribution value corresponding to the j-th first sub-structured information; represents the feature distribution value corresponding to the k-th first sub-structured information; num represents the total quantity of the first sub-structured information.

[0092] Exemplarily, based on content matching, keyword overlap, feature similarity (such as numerical similarity or matching degree of classification features), etc., calculate the similarity between each first sub-structured information in the initial structured information and other sub-structured information, and then count the number of similarities between each first sub-structured information and other sub-structured information as its corresponding similarity quantity.

[0093] Exemplarily, calculate the feature entropy of each first sub-structured information using the information entropy formula according to the similarity quantity, then sum all the feature entropies to obtain the target entropy, and thus divide each feature entropy by the target entropy to obtain the feature distribution value of each first sub-structured information. The feature distribution value reflects the degree of concentration or dispersion of information.

[0094] Exemplarily, for any two pieces of first sub-structured information, calculate the difference between their feature distribution values, so as to obtain the distribution difference value between each pair of first sub-structured information, which is used for subsequent clustering analysis. Among them, the distribution difference value is obtained according to the following formula:

[0095]

[0096] Among them, represents the distribution difference value between the i-th piece of first sub-structured information and the j-th piece of first sub-structured information; represents the feature distribution value corresponding to the i-th piece of first sub-structured information; represents the feature distribution value corresponding to the j-th piece of first sub-structured information; represents the feature distribution value corresponding to the k-th piece of first sub-structured information; num represents the total number corresponding to the first sub-structured information.

[0097] Exemplarily, use hierarchical clustering, K-means clustering or other suitable clustering algorithms to cluster the first sub-structured information using the distribution difference value. The basis for clustering is the distribution difference value. The smaller the difference value, the more likely it is to be assigned to the same cluster, so as to obtain the target clustering result. Each cluster in the target clustering result contains a group of similar first sub-structured information.

[0098] Exemplarily, for each clustering cluster in the target clustering result, use methods such as Euclidean distance, Manhattan distance, cosine distance, etc. to calculate the target distance between the second sub-structured information in each clustering cluster and the remaining structured information in the clustering cluster, and then sum the target distances between the second sub-structured information and each remaining structured information in the clustering cluster to obtain the distance sum value between the second sub-structured information and all the remaining structured information. Thus, sort the second sub-structured information in the clustering cluster from high to low according to the distance sum value, and then determine the second sub-structured information corresponding to the larger distance sum value in the sorting result as the cluster-structured information corresponding to the clustering cluster. Thus, after summarizing all the cluster-structured information, determine the target structured information corresponding to the target construction company.

[0099] Step S107, obtain the second key feature vector corresponding to the target construction company under the initial structured information by using the target structured information.

[0100] Exemplarily, use a machine learning model or a neural network model to perform vector representation on the target structured information, so as to obtain the second key feature vector corresponding to the target construction company under the initial structured information.

[0101] Step S108: Determine the target risk identification result corresponding to the target construction company according to the artificial intelligence algorithm in combination with the first key feature vector and the second key feature vector.

[0102] Exemplarily, the first key feature vector and the second key feature vector are weighted and combined according to the importance of the features to form a target feature vector. Then, according to the specific requirements of risk identification and data characteristics, a suitable artificial intelligence algorithm such as a machine learning model (such as logistic regression, support vector machine, random forest) and a deep learning model (such as neural network) is selected to identify the risk type of the target feature vector, so as to obtain the target risk identification result corresponding to the target construction company. The target risk identification result is any one of high risk, medium risk, and low risk.

[0103] For example, historical data is collected, including the feature vectors and corresponding risk labels of the target construction company with known risk levels, so as to train the selected artificial intelligence model using the training data set, then adjust the model parameters, and optimize the model performance through methods such as cross-validation, so as to obtain the trained artificial intelligence model. Then, the artificial intelligence model is used to identify the risk type of the target feature vector, so as to obtain the target risk identification result corresponding to the target construction company.

[0104] In some embodiments, the step of determining the target risk identification result corresponding to the target construction company according to the artificial intelligence algorithm in combination with the first key feature vector and the second key feature vector includes: establishing a target association graph corresponding to the target construction company according to the target paragraph title, the target abstract information, the target key sentences, and the target structured information; performing path recognition on the target association graph to obtain a corresponding target association path; using the feature fusion layer of the risk identification model to perform information fusion on the first key feature vector and the second key feature vector according to the target association path to obtain a target key feature vector; using the risk classification layer of the risk identification model to perform risk judgment on the target association path according to the target key feature vector to obtain an initial risk identification result; using the risk fusion layer of the risk identification model to perform risk fusion on the initial risk identification result according to the target association path to obtain the target risk identification result corresponding to the target construction company.

[0105] Exemplarily, the correlation relationships between the target structured information and the target paragraph title, the target abstract information, and the target key sentences are calculated respectively. For example, the similarity between the target structured information and the target paragraph title is w1, the similarity between the target structured information and the target abstract information is w2, and the similarity between the target structured information and the target key sentences is w3. Then, the sum of w1, w2, and w3 is calculated to obtain the target correlation value. Thus, when the target correlation value is greater than the preset value, it is determined that there is a correlation between the second key feature vector corresponding to the target structured information and the first key feature vectors corresponding to the target paragraph title, the target abstract information, and the target key sentences. Furthermore, the second key feature vector and the first key feature vectors are respectively determined as the graph nodes corresponding to the target correlation graph, and the target correlation value is determined as the node weight between the second key feature vector and the first key feature vectors.

[0106] Exemplarily, all the paths existing in the target correlation graph are obtained to obtain the corresponding target correlation path. Then, the feature fusion layer of the risk identification model is used to perform information fusion according to the first key feature vector and the second key feature vector involved in the target correlation path, so as to obtain the target key feature vector corresponding to the target correlation path.

[0107] Exemplarily, the target key feature vector is input into the risk classification layer of the risk identification model. For example, a classification algorithm (such as logistic regression, support vector machine, random forest, or neural network) classifies the target key feature vector, so as to output an initial risk identification result, including a risk level and a risk probability value corresponding to the risk level.

[0108] Exemplarily, the initial risk identification result and the target correlation path are input into the risk fusion layer of the risk identification model. Thus, according to the intensity or importance of the target correlation path, weights are assigned to each path. For example, path A reflects the relationship between the company and financial risks, and path B reflects the relationship between the company and technical risks. The weight of path A may be higher than that of path B. Then, methods such as weighted average or voting mechanism are used to combine the initial risk identification result with the path weights. For example, if the weight of path A is 0.7, the weight of path B is 0.3, and the initial risk identification results are "high risk" (path A) and "low risk" (path B) respectively, the target risk identification result can be calculated through weighted average. If there are multiple target correlation paths, the risk contribution of each path can be analyzed one by one, and then the target risk identification result corresponding to the target construction company can be obtained.

[0109] In some embodiments, the risk fusion layer of the risk identification model performs risk fusion on the initial risk identification results according to the target association path to obtain the target risk identification result corresponding to the target construction company, including: using the risk fusion layer to determine the node weight corresponding to each graph node in the target association path, and determining the target weight corresponding to the target association path according to the node weight; using the risk fusion layer to perform risk fusion on the initial risk identification results according to the target weight to obtain the target risk identification result corresponding to the target construction company.

[0110] Exemplarily, when constructing the target association graph, the target association value is obtained by using the similarity between the target structured information and the target paragraph title, the similarity between the target structured information and the target abstract information, and the similarity between the target structured information and the target key sentence. Thus, when the second key feature vector and the first key feature vector are respectively determined as the graph nodes corresponding to the target association graph, the target association value is determined as the node weight between the second key feature vector and the first key feature vector. Furthermore, the risk fusion layer determines the node weight corresponding to each graph node in the target association path from the target association graph, and then sums up the node weights involved in the target association path to obtain the target weight corresponding to the target association path. The risk fusion layer of the risk identification model uses the target weight to perform weighted averaging on multiple initial risk identification results to obtain the target risk identification result.

[0111] Step S109: Determine the auxiliary review result corresponding to the target construction company according to the target risk identification result.

[0112] Exemplarily, the risk level corresponding to the maximum risk probability value in the target risk identification result is obtained, and then the auxiliary review result corresponding to the target construction company is determined according to this risk level. For example, when the risk level is high risk, the auxiliary review result is that the target construction company has high risk, please review carefully; when the risk level is medium risk, the auxiliary review result is that the target construction company has medium risk, please further investigate and pay close attention; when the risk level is low risk, the auxiliary review result is that the target construction company has low risk.

[0113] Please refer to Figure 2 , Figure 2An AI-based customer access auxiliary audit system 200 provided by an embodiment of the present application. The AI-based customer access auxiliary audit system 200 includes a data acquisition module 201, a data disassembling module 202, an abstract generating module 203, a statement recognition module 204, a first vector analysis module 205, a feature recognition module 206, a second vector analysis module 207, a risk recognition module 208, and a result determination module 209. Among them, the data acquisition module 201 is used to obtain the initial structured information and initial text information required for the access audit of the target construction company; the data disassembling module 202 is used to disassemble the initial text information into corresponding target paragraph titles and associated text information corresponding to the target paragraph titles; the abstract generating module 203 is used to perform an abstract generating process on the associated text information to obtain target abstract information corresponding to the target paragraph titles; the statement recognition module 204 is used to identify key statements in the target abstract information to obtain target key statements corresponding to the target paragraph titles; the first vector analysis module 205 is used to determine a first key feature vector corresponding to the target construction company under the initial text information by using the target paragraph titles, the target abstract information, and the target key statements; the feature recognition module 206 is used to identify key features in the initial structured information to obtain target structured information corresponding to the target construction company; the second vector analysis module 207 is used to obtain a second key feature vector corresponding to the target construction company under the initial structured information by using the target structured information; the risk recognition module 208 is used to determine a target risk recognition result corresponding to the target construction company according to an artificial intelligence algorithm in combination with the first key feature vector and the second key feature vector; the result determination module 209 is used to determine an auxiliary audit result corresponding to the target construction company according to the target risk recognition result.

[0114] In some embodiments, the AI-based customer access auxiliary audit system 200 can be applied to a terminal device.

[0115] It should be noted that those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the above-described AI-based customer access auxiliary audit system 200 can refer to the corresponding process in the foregoing embodiment of the AI-based customer access auxiliary audit method, and will not be elaborated herein.

[0116] An embodiment of the present invention also provides a storage medium for computer-readable storage. The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of any one of the AI-based customer access auxiliary audit methods provided in the specification of the embodiments of the present invention.

[0117] Among them, the storage medium may be an internal storage unit of the terminal device described in the foregoing embodiments, such as the hard disk or memory of the terminal device. The storage medium may also be an external storage device of the terminal device, such as a plug-in hard disk equipped on the terminal device, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc.

[0118] Those of ordinary skill in the art can understand that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, and their appropriate combinations. In the hardware embodiment, the division between the functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be executed by several physical components in cooperation. Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or may be implemented as hardware, or may be implemented as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or transitory medium). As is well known to those of ordinary skill in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disks (DVDs) or other optical disk storage, magnetic cassettes, tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. In addition, as is well known to those of ordinary skill in the art, a communication medium typically includes computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transmission mechanism, and may include any information delivery medium.

[0119] It should be understood that the term "and / or" used in the specification and appended claims of the present invention refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. It should be noted that herein, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or system comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or system. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or system comprising the element.

[0120] The serial numbers of the above embodiments of the present invention are only for description and do not represent the superiority or inferiority of the embodiments. The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. An AI-based method for assisting in the review of customer access, characterized in that, The method includes: Obtaining the initial structured information and initial text information required for the access review of the target construction company; Performing paragraph disassembling on the initial text information to obtain the corresponding target paragraph titles and the associated text information corresponding to the target paragraph titles; Performing summary generation processing on the associated text information to obtain the target summary information corresponding to the target paragraph titles; Performing key sentence recognition on the target summary information to obtain the target key sentences corresponding to the target paragraph titles; Using the target paragraph titles, the target summary information, and the target key sentences to determine the first key feature vector corresponding to the target construction company under the initial text information; Performing key feature recognition on the initial structured information to obtain the target structured information corresponding to the target construction company; Using the target structured information to obtain the second key feature vector corresponding to the target construction company under the initial structured information; Determining the target risk recognition result corresponding to the target construction company according to the artificial intelligence algorithm by combining the first key feature vector and the second key feature vector; Determining the auxiliary review result corresponding to the target construction company according to the target risk recognition result.

2. The method according to claim 1, characterized in that, The performing summary generation processing on the associated text information to obtain the target summary information corresponding to the target paragraph titles includes: Performing vector representation on the associated text information by using the text representation layer of the summary generation model to obtain the initial text vector; Performing dependency syntactic analysis on the associated text information by using the syntactic analysis layer of the summary generation model to obtain the syntactic analysis result; Performing syntactic enhancement on the initial text vector according to the syntactic analysis result by using the text enhancement layer of the summary generation model to obtain the target text vector; Performing vocabulary distribution analysis on the target text vector by using the attention distribution layer of the summary generation model to obtain the data distribution information corresponding to the associated text information; Generating the initial summary information corresponding to the associated text information by using the fusion pointer generation network layer of the summary generation model according to the target text vector and the data distribution information; Performing quality calculation on the initial summary information by using the summary quality evaluation layer of the summary generation model to obtain the target score corresponding to the initial summary information; Adjusting the initial summary information according to the target score to obtain the target summary information corresponding to the target paragraph titles.

3. The method according to claim 2, characterized in that, The performing quality calculation on the initial summary information by using the summary quality evaluation layer of the summary generation model to obtain the target score corresponding to the initial summary information includes: Obtaining the target keywords corresponding to the associated text information and the target relationships between the target keywords, and obtaining the first text information corresponding to the target keywords and the target relationships in the associated text information; Performing information analysis on the initial summary information by using the summary quality evaluation layer to obtain the summary keywords corresponding to the initial summary information and the summary relationships between the summary keywords, and obtaining the second text information corresponding to the summary keywords and the summary relationships in the initial summary information; Using the abstract quality evaluation layer, determine a first loss value corresponding to the initial abstract information according to the target keyword, the target relationship, the abstract keyword, and the abstract relationship; Using the abstract quality evaluation layer, fuse the first loss value according to the first text information and the second text information to obtain a second loss value corresponding to the initial abstract information; Using the abstract quality evaluation layer, perform score conversion according to the second loss value to obtain the target score corresponding to the initial abstract information.

4. The method according to claim 3, characterized in that, The step of using the abstract quality evaluation layer to determine a first loss value corresponding to the initial abstract information according to the target keyword, the target relationship, the abstract keyword, and the abstract relationship includes: Using the abstract quality evaluation layer to perform data comparison on the target keyword, the target relationship, the abstract keyword, and the abstract relationship to obtain a comparison result; When the comparison result is that the target keyword is similar to the abstract keyword, and the target relationship is similar to the abstract relationship, then determine that the target characterization value corresponding to the target keyword, the target relationship, the abstract keyword, and the abstract relationship is a first numerical value; When the comparison result is that the target keyword is similar to the abstract keyword, but the target relationship is not similar to the abstract relationship, then determine that the target characterization value corresponding to the target keyword, the target relationship, the abstract keyword, and the abstract relationship is a second numerical value; When the comparison result is that the target keyword is not similar to the abstract keyword, but the target relationship is similar to the abstract relationship, then determine that the target characterization value corresponding to the target keyword, the target relationship, the abstract keyword, and the abstract relationship is a third numerical value; When the comparison result is that the target keyword is not similar to the abstract keyword, and the target relationship is not similar to the abstract relationship, then determine that the target characterization value corresponding to the target keyword, the target relationship, the abstract keyword, and the abstract relationship is a fourth numerical value; Using the abstract quality evaluation layer to fuse the first numerical value, the second numerical value, the third numerical value, and the fourth numerical value to determine the first loss value corresponding to the initial abstract information; Wherein, the first loss value is obtained according to the following formula: wherein, represents the first loss value corresponding to the i-th initial summary information, represents the first numerical value, m represents the quantity corresponding to all the first numerical values determined under all the comparison results, represents the second numerical value, n represents the quantity corresponding to all the second numerical values determined under all the comparison results, represents the third numerical value, k represents the quantity corresponding to all the third numerical values determined under all the comparison results, represents the fourth numerical value, h represents the quantity corresponding to all the fourth numerical values determined under all the comparison results, and lg represents the logarithmic function with base 10.

5. The method according to claim 1, wherein The step of performing key sentence recognition on the target abstract information to obtain the target key sentence corresponding to the target paragraph title includes: Performing sentence segmentation on the target abstract information to obtain target segmented sentences, and calculating a first similarity between the target paragraph title and the target segmented sentences; Performing keyword extraction on the target segmented sentences to obtain sentence keywords, and calculating a second similarity between the sentence keywords and the target paragraph title; Determining the position information corresponding to the target segmented sentence, and determining a first important characterization value corresponding to the target segmented sentence according to the position information; Determining a preset summary word, and determining a second important characterization value corresponding to the target segmented sentence according to the preset summary word; Determine the statement score corresponding to the target segmentation statement by integrating the first similarity, the second similarity, the first important characterization value, and the second important characterization value; Determine the target key statement corresponding to the target paragraph title from the target segmentation statement according to the statement score.

6. The method according to claim 1, characterized in that The obtaining of the target structured information corresponding to the target construction company by performing key feature recognition on the initial structured information includes: Calculate the similarity information for each first sub-structured information in the initial structured information to obtain the corresponding similarity quantity of the first sub-structured information; Determine the feature entropy corresponding to the first sub-structured information according to the similarity quantity, and determine the feature distribution value corresponding to the first sub-structured information according to the feature entropy; Calculate the distribution difference value between any two of the first sub-structured information according to the feature distribution value; Cluster the first sub-structured information according to the distribution difference value to obtain a target clustering result; Calculate the target distance between the second sub-structured information in each clustering cluster in the target clustering result and the remaining structured information in the clustering cluster; Obtain the target structured information corresponding to the target construction company from the initial structured information according to the target distance; Among them, the distribution difference value is obtained according to the following formula: Among them, represents the distribution difference value between the i-th and the j-th of the first sub-structured information; represents the feature distribution value corresponding to the i-th of the first sub-structured information; represents the feature distribution value corresponding to the j-th of the first sub-structured information; represents the feature distribution value corresponding to the k-th of the first sub-structured information; num represents the total quantity corresponding to the first sub-structured information.

7. The method according to claim 1, wherein The determination of the target risk recognition result corresponding to the target construction company by combining the first key feature vector and the second key feature vector according to the artificial intelligence algorithm includes: Establish a target association graph corresponding to the target construction company according to the target paragraph title, the target abstract information, the target key statement, and the target structured information; Perform path recognition on the target association graph to obtain a corresponding target association path; Use the feature fusion layer of the risk recognition model to perform information fusion on the first key feature vector and the second key feature vector according to the target association path to obtain a target key feature vector; Use the risk classification layer of the risk recognition model to perform risk judgment on the target association path according to the target key feature vector to obtain an initial risk recognition result; Use the risk fusion layer of the risk recognition model to perform risk fusion on the initial risk recognition result according to the target association path to obtain the target risk recognition result corresponding to the target construction company.

8. The method according to claim 7, wherein The use of the risk fusion layer of the risk recognition model to perform risk fusion on the initial risk recognition result according to the target association path to obtain the target risk recognition result corresponding to the target construction company includes: Use the risk fusion layer to determine the node weight corresponding to each graph node in the target association path, and determine the target weight corresponding to the target association path according to the node weight; Use the risk fusion layer to perform risk fusion on the initial risk recognition result according to the target weight to obtain the target risk recognition result corresponding to the target construction company.

9. The method according to any one of claims 1-8, characterized in that, The obtaining of the initial structured information and the initial text information required for the target construction company during the admission review includes: Identify the identity information, and when the access permission corresponding to the identity information meets the preset permission, obtain the initial encrypted text required for the target construction company during the access review; Obtain the decryption key according to the identity information, and decrypt the initial encrypted text with the decryption key to obtain the initial structured information and the initial text information.

10. An AI-based customer access auxiliary review system, characterized in that, It includes: A data acquisition module, configured to obtain the initial structured information and the initial text information required for the target construction company during the access review; A data disassembling module, configured to disassemble the initial text information into paragraphs to obtain the corresponding target paragraph titles and the associated text information corresponding to the target paragraph titles; An abstract generation module, configured to perform abstract generation processing on the associated text information to obtain the target abstract information corresponding to the target paragraph title; A statement recognition module, configured to perform key statement recognition on the target abstract information to obtain the target key statements corresponding to the target paragraph title; A first vector analysis module, configured to determine the first key feature vector corresponding to the target construction company under the initial text information by using the target paragraph title, the target abstract information, and the target key statements; A feature recognition module, configured to perform key feature recognition on the initial structured information to obtain the target structured information corresponding to the target construction company; A second vector analysis module, configured to obtain the second key feature vector corresponding to the target construction company under the initial structured information by using the target structured information; A risk recognition module, configured to determine the target risk recognition result corresponding to the target construction company according to the artificial intelligence algorithm in combination with the first key feature vector and the second key feature vector; A result determination module, configured to determine the auxiliary review result corresponding to the target construction company according to the target risk recognition result.

Images