Text processing method and device and computing equipment

Abstract

The embodiment of the invention provides a text processing method and device and computing equipment, and the method comprises the steps: obtaining a business text, and carrying out text content retrieval on the business text to obtain a first retrieval result which represents text content used for determining the business risk of a target business corresponding to the business text in the business text; processing the first retrieval result based on a retrieval model to obtain a second retrieval result, the second retrieval result being a subset of the first retrieval result, and the correlation between the second retrieval result and the business risk being greater than the correlation between the first retrieval result and the business risk; a text generation model is called to process the second retrieval result, an output text is generated, and the output text is at least used for representing the business risk level of the target business. The method can improve the accuracy of the output text.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of data processing technology, and in particular to a text processing method, apparatus, and computing device. Background Art

[0002] With social progress and technological innovation, the text processing process is undergoing a profound transformation from traditional manual processing to artificial intelligence automated processing, and related businesses can be automatically reviewed based on the processing results.

[0003] In related technologies, AI-powered automated text processing uses Retrieval-Augmented Generation (RAG) technology to retrieve key information from business documents, generate search results, and then analyze these results to generate output text. However, this method suffers from inaccurate output text information, which can affect business review results. Summary of the Invention

[0004] The embodiments of the present application provide a text processing method, apparatus, and computing device for solving the technical problem of inaccurate text output during business audit in related technologies.

[0005] In a first aspect, an embodiment of the present application provides a text processing method, comprising:

[0006] Acquire a business text, and perform a text content search on the business text to obtain a first search result, where the first search result represents text content in the business text that is used to determine a business risk of a target business corresponding to the business text;

[0007] Processing the first search result based on the search model to obtain a second search result, wherein the second search result is a subset of the first search result, and the correlation between the second search result and the business risk is greater than the correlation between the first search result and the business risk;

[0008] A text generation model is called to process the second search result to generate an output text, where the output text is at least used to characterize the business risk level of the target business.

[0009] In the above scheme, a first search result can be generated by performing text content retrieval on the business text. The first search result can include more information related to business risks, which can avoid affecting the accuracy of the output text due to omission of important information; the first search result is further processed by the retrieval model to obtain a second search result, and the second search result is analyzed to generate an output text. This can enable the second search result to include important information without introducing too much redundant data, thereby improving the accuracy of the output text and improving the accuracy of business audits.

[0010] In a possible implementation, performing a text content search on the business text to obtain a first search result includes:

[0011] Segmenting the business text to obtain a segmentation result, wherein the segmentation result includes multiple text blocks;

[0012] The segmentation result is compared with a preset statement corresponding to the target business to obtain a first search result.

[0013] In the above solution, by dividing the business text into multiple text blocks, more accurate matching queries can be achieved, noise can be reduced, the accuracy of the first search result can be improved, and the efficiency of text processing can be accelerated.

[0014] In a possible implementation, comparing the segmentation result with a preset statement corresponding to the target business to obtain a first search result includes:

[0015] Performing vectorization processing on the multiple text blocks and the preset sentence to obtain segmentation vectors corresponding to each text block and a target vector corresponding to the preset sentence;

[0016] respectively determining the similarity between each of the segmentation vectors and the target vector;

[0017] The first search result is obtained according to the similarity.

[0018] In the above scheme, by vectorizing the text blocks, the vectorized data can better capture semantic information and improve the accuracy of retrieval. At the same time, the vectorized data is more convenient for mathematical operations, such as facilitating the rapid calculation of similarity.

[0019] In a possible implementation, the retrieval model includes a density-based noise application spatial clustering model, and processing the first retrieval result based on the retrieval model to obtain the second retrieval result includes:

[0020] Determining a cluster radius and a minimum number of points corresponding to the density-based noise application spatial clustering model;

[0021] performing clustering processing on the first search result according to the cluster radius and the minimum number of points to obtain a clustering result, wherein the clustering result includes at least one cluster, and the cluster includes multiple segmentation vectors;

[0022] The segmentation vector in the cluster is determined as the second search result.

[0023] In the above scheme, similar segmentation vectors can be clustered together according to the density-based noise application spatial clustering model to identify noise points. By removing noise points, the accuracy of the retrieval results can be further improved. At the same time, since the density-based noise application spatial clustering model does not require the pre-specification of the number of clusters and can better handle clusters of different densities and shapes, the applicability of text retrieval can be improved.

[0024] In a possible implementation, clustering the first search results to obtain clustering results includes:

[0025] Marking each segmentation vector in the first search result according to the cluster radius and the minimum number of points to obtain a marking result, wherein the marking result includes core points, boundary points, and noise points, and the core point is a segmentation vector that contains at least the minimum number of segmentation vectors within the cluster radius;

[0026] For any core point, all segmentation vectors within the clustering radius neighborhood of the core point are divided into initial clusters;

[0027] According to the initial cluster and the cluster radius, all segmentation vectors within the cluster radius neighborhood of each core point in the initial cluster are divided into the initial cluster to obtain the cluster cluster;

[0028] The segmentation vectors that are not classified into any cluster are determined as noise points to obtain the clustering result.

[0029] In the above scheme, by marking each segmentation vector as a core point, a boundary point, and a noise point, outlier vectors in the first search result (such as paragraphs with semantic drift) can be automatically identified, and outliers can be accurately removed, thereby improving the accuracy of the second search result.

[0030] In a possible implementation, the retrieval model includes a large language model, and processing the first retrieval result based on the retrieval model to obtain the second retrieval result includes:

[0031] determining a first prompt word according to the text blocks corresponding to the segmentation vectors in the first search result and the target business, wherein the first prompt word is used to instruct the large language model to generate the second search result;

[0032] Determining, based on the first prompt word, a semantic relevance between the text block corresponding to each segmentation vector in the first search result and the target business, wherein the semantic relevance indicates a degree of semantic association between the target business and the text block;

[0033] A second search result is determined based on the semantic relevance.

[0034] In the above solution, the large language model has natural language processing capabilities such as semantic understanding and context analysis. By determining the first prompt word, it can help the large language model focus on the target business more accurately, thereby improving the correlation between each segmentation vector in the second search result and the business risk.

[0035] In a possible implementation, the retrieval model includes a density-based noise application spatial clustering model and a large language model, and processing the first retrieval result based on the retrieval model to obtain the second retrieval result includes:

[0036] Processing the first search result based on the density-based noise application spatial clustering model to obtain a density-based noise application spatial clustering search result;

[0037] Processing the first search result based on the large language model to obtain a large language search result;

[0038] The density-based noise application spatial clustering retrieval result and the large language retrieval result are merged and deduplicated to obtain a second retrieval result.

[0039] In this solution, a density-based noise spatial clustering model can be used to remove noise points from the first search result, filtering out irrelevant or redundant information. A large language model can then be used to gain a deeper understanding of the text's meaning and context, providing more accurate semantic judgments. Combining these two models leverages the advantages of both: removing noise while enhancing semantic relevance, resulting in more comprehensive and accurate search results.

[0040] In a possible implementation, the retrieval model includes a density-based noise application spatial clustering model and a large language model, and processing the first retrieval result based on the retrieval model to obtain the second retrieval result includes:

[0041] Processing the first search result based on the density-based noise application spatial clustering model to obtain a density-based noise application spatial clustering search result, and processing the density-based noise application spatial clustering search result based on the large language model to obtain the second search result;

[0042] Alternatively, the first search result is processed based on the large language model to obtain a large language search result, and the large language search result is processed based on the density-based noise application spatial clustering model to obtain the second search result.

[0043] In the above scheme, both orders for determining the second search results can combine the noise processing capabilities of the density-based noise application spatial clustering model and the semantic understanding capabilities of the large language model to improve the accuracy of the search results. Furthermore, different processing orders will lead to different intermediate results, which in turn affect the second search results. The appropriate processing order can be selected based on specific business scenarios and data characteristics, improving the applicability of the text processing method.

[0044] In a second aspect, an embodiment of the present application provides a text processing device, comprising:

[0045] An acquisition module, configured to acquire a business text and perform a text content search on the business text to obtain a first search result, wherein the first search result represents text content in the business text used to determine a business risk of a target business corresponding to the business text;

[0046] a first processing module, configured to process the first search result based on a search model to obtain a second search result, wherein the second search result is a subset of the first search result, and a correlation between the second search result and the business risk is greater than a correlation between the first search result and the business risk;

[0047] The second processing module is configured to call a text generation model to process the second search result and generate an output text, wherein the output text is at least used to characterize the business risk level of the target business.

[0048] In the above scheme, a first search result can be generated by performing text content retrieval on the business text. The first search result can include more information related to business risks, which can avoid affecting the accuracy of the output text due to omission of important information; the first search result is further processed by the retrieval model to obtain a second search result, and the second search result is analyzed to generate an output text. This can enable the second search result to include important information without introducing too much redundant data, thereby improving the accuracy of the output text and improving the accuracy of business audits.

[0049] In a possible implementation, the acquisition module is specifically configured to:

[0050] Segmenting the business text to obtain a segmentation result, wherein the segmentation result includes multiple text blocks;

[0051] The segmentation result is compared with a preset statement corresponding to the target business to obtain a first search result.

[0052] In the above solution, by dividing the business text into multiple text blocks, more accurate matching queries can be achieved, noise can be reduced, the accuracy of the first search result can be improved, and the efficiency of text processing can be accelerated.

[0053] In a possible implementation, the acquisition module is specifically configured to:

[0054] Performing vectorization processing on the multiple text blocks and the preset sentence to obtain segmentation vectors corresponding to each text block and a target vector corresponding to the preset sentence;

[0055] respectively determining the similarity between each of the segmentation vectors and the target vector;

[0056] The first search result is obtained according to the similarity.

[0057] In the above scheme, by vectorizing the text blocks, the vectorized data can better capture semantic information and improve the accuracy of retrieval. At the same time, the vectorized data is more convenient for mathematical operations, such as facilitating the rapid calculation of similarity.

[0058] In a possible implementation, the first processing module is specifically configured to:

[0059] Determining a cluster radius and a minimum number of points corresponding to the density-based noise application spatial clustering model;

[0060] performing clustering processing on the first search result according to the cluster radius and the minimum number of points to obtain a clustering result, wherein the clustering result includes at least one cluster, and the cluster includes multiple segmentation vectors;

[0061] The segmentation vector in the cluster is determined as the second search result.

[0062] In the above scheme, similar segmentation vectors can be clustered together according to the density-based noise application spatial clustering model to identify noise points. By removing noise points, the accuracy of the retrieval results can be further improved. At the same time, since the density-based noise application spatial clustering model does not require the pre-specification of the number of clusters and can better handle clusters of different densities and shapes, the applicability of text retrieval can be improved.

[0063] In a possible implementation, the first processing module is specifically configured to:

[0064] Marking each segmentation vector in the first search result according to the cluster radius and the minimum number of points to obtain a marking result, wherein the marking result includes core points, boundary points, and noise points, and the core point is a segmentation vector that contains at least the minimum number of segmentation vectors within the cluster radius;

[0065] For any core point, all segmentation vectors within the clustering radius neighborhood of the core point are divided into initial clusters;

[0066] According to the initial cluster and the cluster radius, all segmentation vectors within the cluster radius neighborhood of each core point in the initial cluster are divided into the initial cluster to obtain the cluster cluster;

[0067] The segmentation vectors that are not classified into any cluster are determined as noise points to obtain the clustering result.

[0068] In the above scheme, by marking each segmentation vector as a core point, a boundary point, and a noise point, outlier vectors in the first search result (such as paragraphs with semantic drift) can be automatically identified, and outliers can be accurately removed, thereby improving the accuracy of the second search result.

[0069] In a possible implementation, the first processing module is specifically configured to:

[0070] determining a first prompt word according to the text blocks corresponding to the segmentation vectors in the first search result and the target business, wherein the first prompt word is used to instruct the large language model to generate the second search result;

[0071] Determining, based on the first prompt word, a semantic relevance between the text block corresponding to each segmentation vector in the first search result and the target business, wherein the semantic relevance indicates a degree of semantic association between the target business and the text block;

[0072] A second search result is determined based on the semantic relevance.

[0073] In the above solution, the large language model has natural language processing capabilities such as semantic understanding and context analysis. By determining the first prompt word, it can help the large language model focus on the target business more accurately, thereby improving the correlation between each segmentation vector in the second search result and the business risk.

[0074] In a possible implementation, the first processing module is specifically configured to:

[0075] Processing the first search result based on the density-based noise application spatial clustering model to obtain a density-based noise application spatial clustering search result;

[0076] Processing the first search result based on the large language model to obtain a large language search result;

[0077] The density-based noise application spatial clustering retrieval result and the large language retrieval result are merged and deduplicated to obtain a second retrieval result.

[0078] In this solution, a density-based noise spatial clustering model can be used to remove noise points from the first search result, filtering out irrelevant or redundant information. A large language model can then be used to gain a deeper understanding of the text's meaning and context, providing more accurate semantic judgments. Combining these two models leverages the advantages of both: removing noise while enhancing semantic relevance, resulting in more comprehensive and accurate search results.

[0079] In a possible implementation, the first processing module is specifically configured to:

[0080] Processing the first search result based on the density-based noise application spatial clustering model to obtain a density-based noise application spatial clustering search result, and processing the density-based noise application spatial clustering search result based on the large language model to obtain the second search result;

[0081] Alternatively, the first search result is processed based on the large language model to obtain a large language search result, and the large language search result is processed based on the density-based noise application spatial clustering model to obtain the second search result.

[0082] In the above scheme, both orders for determining the second search results can combine the noise processing capabilities of the density-based noise application spatial clustering model and the semantic understanding capabilities of the large language model to improve the accuracy of the search results. Furthermore, different processing orders will lead to different intermediate results, which in turn affect the second search results. The appropriate processing order can be selected based on specific business scenarios and data characteristics, improving the applicability of the text processing method.

[0083] In a third aspect, an embodiment of the present application provides a computing device, including: a memory, a processor;

[0084] The memory stores computer-executable instructions;

[0085] The processor executes the computer-executable instructions stored in the memory, so that the processor executes the above first aspect and / or various possible implementations of the first aspect.

[0086] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, they are used to implement the first aspect above and / or various possible implementation methods of the first aspect.

[0087] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the above first aspect and / or various possible implementation methods of the first aspect.

[0088] The text processing method, apparatus and computing device provided in the embodiment of the present application obtain a business text and perform a text content search on the business text to obtain a first search result, wherein the first search result represents the text content in the business text used to determine the business risk of the target business corresponding to the business text; the first search result is processed based on a search model to obtain a second search result, wherein the second search result is a subset of the first search result, and the correlation between the second search result and the business risk is greater than the correlation between the first search result and the business risk; the text generation model is called to process the second search result to generate an output text, and the output text is at least used to represent the business risk level of the target business. In the above method, the first search result can be generated by performing a text content search on the business text, and the first search result can include more information related to the business risk, which can avoid affecting the accuracy of the output text due to the omission of important information; the first search result is further processed by the search model to obtain a second search result, and the second search result is analyzed to generate an output text. The information retrieval capability provided by the search model can be used to make the second search result include important information without introducing too much redundant data, thereby improving the information quality of the second search result, thereby improving the accuracy of the output text and improving the accuracy of the business review. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0090] Figure 1 A schematic diagram of an application scenario provided in an embodiment of the present application;

[0091] Figure 2 A flowchart of a text processing method provided in an embodiment of the present application;

[0092] Figure 3 A schematic diagram of a process for determining a second search result based on a DBSCAN model provided in an embodiment of the present application;

[0093] Figure 4 A schematic diagram of a process for determining a second search result based on an LLM model provided in an embodiment of the present application;

[0094] Figure 5 Schematic diagram of the process of determining the second search result provided in the embodiment of the present application Figure 1 ;

[0095] Figure 6Schematic diagram of the process of determining the second search result provided in the embodiment of the present application Figure 2 ;

[0096] Figure 7 Schematic diagram of the process of determining the second search result provided in the embodiment of the present application Figure 3 ;

[0097] Figure 8 A schematic diagram of the structure of a text processing device provided in an embodiment of the present application;

[0098] Figure 9 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0099] Exemplary embodiments are described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numbers in different figures represent identical or similar elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible implementations consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with certain aspects of the present invention, as detailed in the appended claims.

[0100] For ease of understanding, the following Figure 1 , describes the application scenarios involved in the embodiments of this application.

[0101] Figure 1 This is a schematic diagram of an application scenario provided by the embodiment of this application. Figure 1 , including a client 101 and a server 102. The text processing method provided in the embodiment of the present application can be applied to scenarios where it is necessary to review a business based on a business text, such as loan review, construction project review, house decoration compliance review, fire protection design assessment, etc. For example, during the loan review process of a bank or financial institution, the target business that the applicant needs to handle can be reviewed based on the text related to the loan business, and the feasibility of the target business, the risks involved, etc. can be assessed. A communication connection can be established between the client 11 and the server 12 to realize data transmission between the client 11 and the server 12.

[0102] The client 11 may be any electronic device with a display screen, including but not limited to a smart phone, a tablet computer, a desktop computer, a portable computer, etc. The client 11 may obtain a preset statement corresponding to the target service and send the preset statement to the server 12 .

[0103] The server 12 may receive the preset sentence sent by the client 11 , search the database for text data related to the preset sentence, generate an output text, and transmit the output text to the client 11 .

[0104] With social progress and technological innovation, the text processing process is undergoing a profound transformation from traditional manual processing to artificial intelligence (AI) automated processing.

[0105] In related technologies, AI-powered automated text processing uses search-enhanced generation techniques to retrieve key information from business documents, generate search results, and then analyze these results to generate output text. However, when the search results contain relatively little information, these technologies can easily miss important information, resulting in inaccurate output text. When the search results contain a large amount of information, they can easily introduce excessive redundant data, affecting the analysis process and causing inaccurate text information.

[0106] In response to the above problems, an embodiment of the present application proposes a text processing method, which generates a first search result by performing text content retrieval on the business text. The first search result may include more information related to business risks, which can avoid affecting the accuracy of the output text due to omission of important information; the first search result is further processed by the retrieval model to obtain a second search result, and the second search result is analyzed to generate the output text. By utilizing the information retrieval capability provided by the retrieval model, the second search result can include important information without introducing too much redundant data, thereby improving the information quality of the second search result, thereby improving the accuracy of the output text and improving the accuracy of business audits.

[0107] The following specific embodiments are used to describe the technical solutions of the embodiments of the present application in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0108] Figure 2 This is a flowchart of a text processing method provided in an embodiment of the present application. The execution subject of the method can be a computing device or a processor set in a computing device. For example, the computing device can be a server. For example, the processor in the computing device can be a CPU in the server. For ease of understanding, the following description is based on the example of the execution subject being a computing device. Figure 2 , the method may include:

[0109] S101: Acquire a business text, and perform a text content search on the business text to obtain a first search result.

[0110] Business documents refer to textual materials related to the target business. For example, if the target business is loan review, business documents may include loan contracts, loan process specification documents, historical loan business materials, etc.; for example, if the target business is construction project review, business documents may include construction contract documents, technical specification documents, construction process records, regulatory documents, etc.; for example, if the target business is home renovation compliance review, business documents may include renovation plan documents, material compliance certificates, approval and filing materials, construction process documents, etc.

[0111] The first search result is text content in the business text used to determine the business risk of the target business. Based on the RAG model, the business text can be searched for text content according to the preset text corresponding to the target business to obtain the first search result. Specifically, the first search result can be determined by segmenting the business text to obtain segmentation results, wherein the segmentation results include multiple text blocks; and comparing the segmentation results with the preset sentences corresponding to the target business to obtain the first search result.

[0112] The segmentation of the business text can be based on fixed-length segmentation, segmentation based on natural paragraphs, segmentation based on semantics, etc., and the embodiments of the present application do not impose specific restrictions on this. For example, when segmenting based on fixed length, the business text can be segmented according to a preset number of characters or words to obtain multiple text blocks of fixed length. For example, when segmenting based on natural paragraphs, the business text can be segmented according to punctuation marks (periods, line breaks, section breaks, etc.) to obtain multiple text blocks of different lengths. For example, when segmenting based on semantics, semantic boundaries can be identified according to a language processing model, and segmentation can be performed according to the integrity of the semantic content to obtain multiple text blocks of different lengths.

[0113] When comparing the segmentation results with the preset sentences corresponding to the target business, multiple text blocks and preset sentences can be vectorized to obtain the segmentation vectors corresponding to each text block and the target vectors corresponding to the preset sentences; and the similarity between each segmentation vector and the target vector is determined respectively.

[0114] Specifically, vectorization of multiple text blocks and pre-set sentences can be achieved using a pre-set model, such as Sentence Transformers or the BERT series. The pre-set model performs pre-processing, positional encoding, and pooling on the input text (i.e., text blocks and pre-set sentences), and outputs vectors corresponding to the input text (i.e., segmentation vectors and target vectors). Each text block corresponds to its own segmentation vector, and the pre-set text corresponds to a target vector.

[0115] The similarity between each segmentation vector and the target vector can be determined by calculating cosine similarity. For example, the cosine similarity between each segmentation vector and the target vector can be calculated to obtain multiple cosine similarity values. Cosine similarity is used to measure the degree of directional proximity between two vectors. The closer the cosine similarity value is to 1, the more similar the two vectors are.

[0116] The first search result may be determined based on the cosine similarity value and a first threshold value, wherein the first threshold value may be used to indicate the cosine similarity value or to indicate the number of segmentation vectors in the first search result.

[0117] For example, when the first threshold is used to indicate the cosine similarity value, the segmentation vectors with cosine similarity values ​​greater than or equal to the first threshold can be determined as the first search result, wherein the number of segmentation vectors in the first search result can be different for different business texts or target businesses.

[0118] For example, when the first threshold is used to indicate the number of segmentation vectors in the first search result, the segmentation vectors can be sorted in descending order of cosine similarity values, and the segmentation vectors before the first threshold in the sorting result are determined as the first search result, wherein the number of segmentation vectors in the first search result is the same for different business texts or target businesses.

[0119] In the embodiment of the present application, the first threshold can be adjusted so that the first search result can include more segmentation vectors related to business risks, thereby ensuring that the first search result does not miss important information.

[0120] S102: Process the first search result based on the search model to obtain a second search result.

[0121] The second search result is a subset of the first search result, and the correlation between the second search result and the business risk is greater than the correlation between the first search result and the business risk.

[0122] The retrieval model is used to perform further retrieval processing on the first retrieval result to obtain a second retrieval result. The retrieval model may include a density-based spatial clustering of applications with noise (DBSCAN) model, a large language model (LLM), etc.

[0123] In one possible implementation, the first search result can be processed based on a DBSCAN model to obtain a second search result. Specifically, the DBSCAN model can connect high-density regions to form clusters based on the density of each segmentation vector in the first search result, while treating low-density regions as noise or boundaries and removing segmentation vectors corresponding to the noise points from the first search result to obtain the second search result.

[0124] In one possible implementation, the first search results can be processed based on an LLM model to obtain a second search result. Specifically, the LLM model can determine, based on the semantic relevance between the text blocks corresponding to each segmentation vector in the first search result and the preset sentence, the text blocks in the first search result that are more relevant to the business risk of the target business as the second search result.

[0125] In one possible implementation, the first search result can be processed based on the DBSCAN model and the LLM model to obtain a second search result. For example, the first search result can be processed using the DBSCAN model to obtain a DBSCAN search result, the first search result can be processed using the LLM model to obtain an LLM search result, and the union of the DBSCAN search result and the LLM search result can be obtained to obtain the second search result. For example, the first search result can be processed using the DBSCAN model to obtain a DBSCAN search result, and the DBSCAN search result can be processed again using the LLM model to obtain the second search result. For example, the first search result can be processed using the LLM model to obtain an LLM search result, and the LLM search result can be processed again using the DBSCAN model to obtain the second search result.

[0126] Among them, the DBSCAN model realizes data partitioning through density-based clustering. It is good at discovering clusters of arbitrary shapes and identifying noise points. Especially in common scenarios such as contract review, the contract text usually contains a large amount of unstructured information, such as clause descriptions, responsibility division, etc. On the one hand, the DBSCAN model does not rely on the preset cluster shape or number, and can automatically adapt to the diversity of the contract text, clustering related keywords or risk points together, which helps the text generation model to quickly determine relevant content and improve the quality of the output text; on the other hand, the training process of the DBSCAN model does not require the pre-specified number of clusters. In contract review, it can automatically adapt to the characteristics of different contract texts without manual adjustment of parameters, thereby improving the efficiency of generating the second search result.

[0127] S103: Calling a text generation model to process the second search result and generate output text.

[0128] The output text is at least used to characterize the business risk level of the target business.

[0129] A text generation model can be an AI model based on a deep learning architecture. It can be trained with data to achieve deep understanding of natural language and text generation. For example, a text generation model could be a Large Language Model (LLM). This model can be pre-trained using a large amount of data relevant to the target business, ensuring greater reliability of the output text.

[0130] The text generation model can analyze the text content in the text blocks corresponding to each segmentation vector in the second search result, as well as the preset sentences corresponding to the target business, and output relevant content such as the business risks of the target business, the compliance of the target business, the resource efficiency assessment of the target business, and abnormal behavior analysis in text form to achieve automated review of the target business.

[0131] For example, when the target business is loan application review, the output text can include relevant content such as the lender's credit risk, loan process compliance risk, market risk, etc., and give recommended application results based on various risks.

[0132] The text processing method provided in the embodiment of the present application obtains a business text and performs a text content search on the business text to obtain a first search result, the first search result characterizing the text content in the business text used to determine the business risk of the target business corresponding to the business text; processes the first search result based on a search model to obtain a second search result, wherein the second search result is a subset of the first search result, and the correlation between the second search result and the business risk is greater than the correlation between the first search result and the business risk; calls a text generation model to process the second search result to generate an output text, the output text at least used to characterize the business risk level of the target business. In the above method, the first search result can be generated by performing a text content search on the business text, and the first search result can include more information related to the business risk, which can avoid affecting the accuracy of the output text due to the omission of important information; the first search result is further processed by the search model to obtain a second search result, and the second search result is analyzed to generate an output text. By utilizing the information retrieval capability provided by the search model, the second search result can be included in the important information without introducing too much redundant data, thereby improving the information quality of the second search result, thereby improving the accuracy of the output text and improving the accuracy of the business audit.

[0133] Based on the above embodiments, Figures 3 to 7 The process of determining the second search result is described in detail. The second search result can be determined based on a search model. The search model can include a DBSCAN model and an LLM model. The second search result can be determined based on a single model or based on both models.

[0134] Figure 3 Schematic diagram of the process of determining the second search result based on the DBSCAN model provided in the embodiment of the present application. In the embodiment of the present application, the second search result is determined, such as Figure 3 Shown, including:

[0135] S201. Determine the cluster radius and minimum number of points corresponding to the DBSCAN model.

[0136] The clustering radius is a parameter used in the DBSCAN model to define the neighborhood range of the segmentation vector. The smaller the clustering radius value, the narrower the neighborhood range, and the stronger the correlation between the text block corresponding to each segmentation vector in the second search result and the business risk of the target business.

[0137] The minimum number of points is a parameter used in the DBSCAN model to define core points. A point is considered a core point if the number of segmentation vectors within its cluster radius is greater than or equal to the minimum number of points. A larger minimum number indicates a stronger correlation between the text blocks corresponding to the segmentation vectors in the second search results and the business risk of the target business.

[0138] S202: Cluster the first search result according to the cluster radius and the minimum number of points to obtain a cluster result.

[0139] The clustering result includes at least one cluster, and the cluster includes multiple segmentation vectors.

[0140] The clustering result can be determined in the following way: according to the clustering radius and the minimum number of points, each segmentation vector in the first search result is marked to obtain a marking result, which includes core points, boundary points and noise points. The core point is a segmentation vector that contains at least the minimum number of segmentation vectors within the clustering radius; for any core point, all segmentation vectors within the clustering radius neighborhood of the core point are divided into initial clusters; according to the initial cluster and the clustering radius, all segmentation vectors within the clustering radius neighborhood of each core point in the initial cluster are divided into the initial cluster to obtain a cluster cluster; the segmentation vector that is not divided into any cluster is determined as a noise point to obtain a clustering result.

[0141] Specifically, when obtaining the marking results, first traverse each segmentation vector (data point) in the first search result. If the number of data points (including itself) contained in the cluster radius neighborhood of a segmentation vector is greater than or equal to the minimum number of points, then the segmentation vector is marked as a core point; if a segmentation vector is located in the cluster radius neighborhood of a core point, but the number of data points contained in its own neighborhood radius is less than the minimum number of points, then the segmentation vector is marked as a boundary point; if a segmentation vector is neither a core point nor belongs to the cluster radius neighborhood of any core point, then the segmentation vector is marked as a noise point.

[0142] When determining the clustering results, for any segmentation vector marked as a core point, all segmentation vectors (including the core point itself) within the clustering radius neighborhood of the core point are divided into a cluster to obtain an initial cluster. For each core point in the initial cluster, all segmentation vectors within its clustering radius neighborhood are checked again. If these segmentation vectors have not been divided into any clusters, they are divided into the current initial cluster. The above checking and division process is repeated until there are no new undivided vectors in the clustering radius neighborhood of all core points in the initial cluster, and the initial cluster is expanded into a cluster cluster.

[0143] The clustering results include at least one cluster, each of which includes multiple segmentation vectors assigned to the cluster. Optionally, the clustering results may also include noise points. The greater the correlation between the vector blocks corresponding to the segmentation vectors in the first search result and the business risk, the fewer the noise points.

[0144] S203: Determine the segmentation vector in each cluster as the second search result.

[0145] The second search result is the segmentation vector obtained after removing the noise points in the first search result. That is, the second search result removes the segmentation vectors in the first search result that have a low correlation with the business risk corresponding to the target business, and only retains the segmentation vectors that have a high correlation with the business risk of the target business, which can ensure the accuracy of the output text.

[0146] The method for determining the second search result provided in an embodiment of the present application uses the DBSCAN model to mark the segmentation vectors as core points, boundary points, and noise points, clusters similar segmentation vectors together, and can automatically identify outlier vectors in the first search result (such as paragraphs with semantic drift). By removing noise points, the accuracy of the second search result can be improved. At the same time, since the DBSCAN model does not require the pre-specification of the number of clusters and can better handle clusters of different densities and shapes, it can improve the applicability of text retrieval.

[0147] Figure 4 Schematic diagram of the process of determining the second search result based on the LLM model provided in the embodiment of the present application. Figure 4 Shown, including:

[0148] S301: Determine a first prompt word according to the text blocks corresponding to the segmentation vectors in the first search result and the target business.

[0149] The first prompt word is a structured natural language instruction used to instruct the LLM model to generate the second search result. That is, the LLM model understands the search task, the format of the input data, and the expected output form based on the first prompt word. The first prompt word can be preset, that is, during the LLM training process, different prompt words can be set to guide the LLM model to output results close to the expected results. When processing the first search result, an appropriate first prompt word can be determined based on the semantic content of the text block corresponding to the segmentation vector and the preset sentence corresponding to the target business, so that the text block corresponding to each segmentation vector in the second search result is more highly correlated with the business risk of the target business.

[0150] S302: Determine the semantic relevance between the text block corresponding to each segmentation vector in the first search result and the target business according to the first prompt word.

[0151] Semantic relevance indicates the degree of semantic connection between the target business and the text block. The LLM model can analyze the various elements in the prompt word to understand the task objectives, input data, and output data format requirements. Through natural language understanding, it deeply understands the content of each text block, including identifying keywords, extracting entities, and analyzing grammatical structure. It then associates the text block content with the target business in the prompt word, identifies the semantic connection between the two, and determines semantic relevance.

[0152] It can be obtained through semantic matching, similarity calculation, and other methods. For example, when using semantic matching to obtain relevance, the LLM model can use the Prompt Engineering model to design scoring rules based on semantic coverage, intent matching, information redundancy, and other aspects to perform semantic relevance scoring and obtain a semantic relevance value. For example, when using similarity calculation to obtain relevance, the semantic relevance value can be determined by calculating the cosine similarity of the segmentation vectors corresponding to the preset sentence and the text block through models such as Sentence-BERT, and by counting the frequency and position weight of the core words in the preset sentence in the text block.

[0153] S303: Determine a second search result based on semantic relevance.

[0154] The second search result may be determined based on the semantic relevance value and a second threshold value, wherein the second threshold value may be used to indicate the semantic relevance value of each segmentation vector in the second search result, or to indicate the number of segmentation vectors in the second search result.

[0155] For example, when the second threshold is used to indicate the semantic relevance value, the segmentation vectors in the first search result whose semantic relevance values ​​are greater than or equal to the second threshold are determined as the second search results, wherein the number of segmentation vectors in the second search result may be different for different business texts or target businesses.

[0156] For example, when the second threshold is used to indicate the number of segmentation vectors in the second search result, each segmentation vector can be sorted in descending order of cosine similarity value, and the first second threshold segmentation vectors in the sorting result are determined as the second search result, wherein the number of segmentation vectors in the second search result is the same for different business texts or target businesses.

[0157] The method for determining the second search result provided in an embodiment of the present application processes the first search result using an LLM model to obtain the second search result. By leveraging the LLM model's natural language processing capabilities, such as semantic understanding and contextual analysis, and by determining the first prompt word, the LLM model is helped to more accurately focus on the target business, thereby improving the correlation between each segmentation vector in the second search result and the business risk.

[0158] Figure 5 Schematic diagram of the process of determining the second search result provided in the embodiment of the present application Figure 1 .like Figure 5 As shown, it includes: a DBSCAN model and an LLM model. In an embodiment of the present application, the first search result can be processed based on the DBSCAN model to obtain a DBSCAN search result; the first search result can be processed based on the LLM model to obtain an LLM search result; the DBSCAN search result and the LLM search result can be merged and deduplicated to obtain a second search result.

[0159] Among them, the process of using the DBSCAN model to obtain DBSCAN retrieval results is the same as Figure 3 The process of obtaining the second search result shown in is similar to that of obtaining the second search result using the LLM model. Figure 4 The process of obtaining the second search result shown in is similar and will not be repeated here. This application does not limit the order of obtaining the DBSCAN search result and the LLM search result. The DBSCAN model and the LLM model can be executed in parallel, or the DBSCAN model can be executed first to obtain the DBSCAN search result, and then the LLM model can be executed to obtain the LLM search result. Alternatively, the LLM model can be executed first to obtain the LLM search result, and then the DBSCAN model can be executed to obtain the DBSCAN search result.

[0160] After obtaining the DBSCAN search result and the LLM search result, the two search results can be merged and deduplicated, that is, the union of the DBSCAN search result and the LLM search result is obtained to obtain the second search result.

[0161] The method for determining the second search result provided in the embodiments of this application can remove noise points in the first search result through the DBSCAN model, filtering out irrelevant or redundant information, and deeply understand the meaning and context of the text through the LLM model, thereby providing more accurate semantic judgment. By combining the two models, the advantages of both can be fully utilized, namely, removing noise and enhancing semantic relevance, resulting in more comprehensive and accurate search results.

[0162] Figure 6 Schematic diagram of the process of determining the second search result provided in the embodiment of the present application Figure 2 .like Figure 6 As shown, it includes: DBSCAN model and LLM model. In the embodiment of the present application, the first search result can be processed based on the DBSCAN model to obtain the DBSCAN search result; and the DBSCAN search result can be processed based on the LLM model to obtain the second search result.

[0163] The process of using the DBSCAN model to obtain DBSCAN retrieval results is similar to Figure 3 The process of obtaining the second search result using the LLM model is similar to that shown in Figure 4 The process of obtaining the second search result shown in is similar and will not be repeated here.

[0164] Figure 7 Schematic diagram of the process of determining the second search result provided in the embodiment of the present application Figure 3 .like Figure 7 As shown, it includes: DBSCAN model and LLM model. In the embodiment of the present application, the first search result can be processed based on the LLM model to obtain the LLM search result; and the LLM search result can be processed based on the DBSCAN model to obtain the second search result.

[0165] The process of using the LLM model to obtain LLM search results is similar to Figure 4 The process of obtaining the second search result using the DBSCAN model is similar to that shown in Figure 4 The process of obtaining the second search result shown in is similar and will not be repeated here.

[0166] Figure 6 and Figure 7 The provided method for determining the second search result and its two possible ordering methods both combine the noise processing capabilities of the DBSCAN model with the semantic understanding capabilities of the LLM model, improving the accuracy of the search results. Furthermore, different processing orders lead to different intermediate results, which in turn affect the second search result. Therefore, the appropriate processing order can be selected based on specific business scenarios and data characteristics, improving the applicability of the text processing method.

[0167] Figure 8 This is a structural diagram of a text processing device provided in an embodiment of the present application. Figure 8 As shown, the text processing device 20 includes: an acquisition module 21, a first processing module 22 and a second processing module 23, wherein:

[0168] An acquisition module 21 is configured to acquire a business text and perform a text content search on the business text to obtain a first search result, wherein the first search result represents text content in the business text that is used to determine a business risk of a target business corresponding to the business text;

[0169] a first processing module 22 configured to process the first search result based on a search model to obtain a second search result, wherein the second search result is a subset of the first search result, and the relevance of the second search result to the business risk is greater than the relevance of the first search result to the business risk;

[0170] The second processing module 23 is configured to call a text generation model to process the second search result and generate an output text, wherein the output text is at least used to characterize the business risk level of the target business.

[0171] In a possible implementation, the acquisition module 21 is specifically configured to:

[0172] Segmenting the business text to obtain a segmentation result, wherein the segmentation result includes multiple text blocks;

[0173] The segmentation result is compared with a preset statement corresponding to the target business to obtain a first search result.

[0174] In the above solution, by dividing the business text into multiple text blocks, more accurate matching queries can be achieved, noise can be reduced, the accuracy of the first search result can be improved, and the efficiency of text processing can be accelerated.

[0175] In a possible implementation, the acquisition module 21 is specifically configured to:

[0176] Performing vectorization processing on the multiple text blocks and the preset sentence to obtain a segmentation vector corresponding to each text block and a target vector corresponding to the preset sentence;

[0177] respectively determining the similarity between each of the segmentation vectors and the target vector;

[0178] The first search result is obtained according to the similarity.

[0179] In a possible implementation, the first processing module 22 is specifically configured to:

[0180] Determine the cluster radius and minimum number of points corresponding to the DBSCAN model;

[0181] performing clustering processing on the first search result according to the cluster radius and the minimum number of points to obtain a clustering result, wherein the clustering result includes at least one cluster, and the cluster includes multiple segmentation vectors;

[0182] The segmentation vector in the cluster is determined as the second search result.

[0183] In a possible implementation, the first processing module 22 is specifically configured to:

[0184] Marking each segmentation vector in the first search result according to the cluster radius and the minimum number of points to obtain a marking result, wherein the marking result includes core points, boundary points, and noise points, and the core point is a segmentation vector that contains at least the minimum number of segmentation vectors within the cluster radius;

[0185] For any core point, all segmentation vectors within the clustering radius neighborhood of the core point are divided into initial clusters;

[0186] According to the initial cluster and the cluster radius, all segmentation vectors within the cluster radius neighborhood of each core point in the initial cluster are divided into the initial cluster to obtain the cluster cluster;

[0187] The segmentation vectors that are not classified into any cluster are determined as noise points to obtain the clustering result.

[0188] In a possible implementation, the first processing module 22 is specifically configured to:

[0189] Determining a first prompt word according to the text blocks corresponding to the segmentation vectors in the first search result and the target business, where the first prompt word is used to instruct the LLM model to generate the second search result;

[0190] Determining, based on the first prompt word, a semantic relevance between the text block corresponding to each segmentation vector in the first search result and the target business, wherein the semantic relevance indicates a degree of semantic association between the target business and the text block;

[0191] A second search result is determined based on the semantic relevance.

[0192] In a possible implementation, the first processing module 22 is specifically configured to:

[0193] Processing the first search result based on the DBSCAN model to obtain a DBSCAN search result;

[0194] Processing the first search result based on the LLM model to obtain an LLM search result;

[0195] The DBSCAN search result and the LLM search result are merged and deduplicated to obtain a second search result.

[0196] In a possible implementation, the first processing module 22 is specifically configured to:

[0197] Processing the first search result based on the DBSCAN model to obtain a DBSCAN search result, and processing the DBSCAN search result based on the LLM model to obtain the second search result;

[0198] Alternatively, the first search result is processed based on the LLM model to obtain an LLM search result, and the LLM search result is processed based on the DBSCAN model to obtain the second search result.

[0199] The text processing device provided in this embodiment can execute the method shown in any of the above method embodiments. Its implementation principles and technical effects are similar, and are not described in detail in this embodiment.

[0200] Figure 9 This is a schematic diagram of the structure of a computing device provided in an embodiment of the present application. Figure 9 As shown, the computing device 30 may include: a processor 31 and a memory 32, wherein the processor 31 and the memory 32 can communicate; exemplarily, the processor 31 and the memory 32 communicate via a communication bus 33, the memory 32 is used to store computer execution instructions, and the processor 31 is used to call the computer execution instructions in the memory to execute the text processing method shown in any of the above method embodiments.

[0201] Optionally, the computing device 30 may further include a communication interface, which may include a transmitter and / or a receiver.

[0202] The computing device 30 may be the computing device shown in any of the above method embodiments, and may execute the text processing method shown in any of the above method embodiments.

[0203] Optionally, the processor may be a CPU, a GPU, a baseboard management controller (BMC), other general-purpose processors, a digital signal processor (DSP), or an application-specific integrated circuit (ASIC). The general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of the present application may be directly implemented as being executed by a hardware processor, or may be implemented by a combination of hardware and software modules in the processor.

[0204] An embodiment of the present application provides a computer-readable storage medium having computer-executable instructions stored thereon; the computer-executable instructions are used to implement the text replacement method as described in any of the above embodiments.

[0205] An embodiment of the present application provides a computer program product, which includes a computer program. When the computer program is executed, it enables a computer to perform the above-mentioned text replacement method.

[0206] All or part of the steps in the above-mentioned method embodiments may be implemented by hardware associated with program instructions. The aforementioned program may be stored in a readable memory. When executed, the program performs the steps of the above-mentioned method embodiments. The aforementioned memory (storage medium) includes: read-only memory (ROM), RAM, flash memory, hard disk, solid-state drive, magnetic tape, floppy disk, optical disk, and any combination thereof.

[0207] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable terminal device to generate a machine, so that the instructions executed by the processing unit of the computer or other programmable terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0208] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable terminal device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0209] These computer program instructions can also be loaded onto a computer or other programmable terminal device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable device to implement the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0210] Obviously, those skilled in the art may make various changes and modifications to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalents, the embodiments of the present application are intended to include such modifications and variations.

[0211] In the embodiments of the present application, the terms "include", "comprise" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element. The term "or" and its variations may refer to "and / or". In the embodiments of the present application, the terms "first", "second", etc. are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. In the embodiments of the present application, "multiple" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects associated before and after are in an "or" relationship.

[0212] Those skilled in the art will readily conceive of other embodiments of the present application after considering the invention disclosed in the specification and practice. The embodiments of the present application are intended to cover any variations, uses, or adaptations of the embodiments of the present application, which follow the general principles of the embodiments of the present application and include common knowledge or customary technical means in the art that are not disclosed in the embodiments of the present application.

Claims

1. A text processing method, characterized in that: include: Acquire a business text, and perform a text content search on the business text to obtain a first search result, where the first search result represents text content in the business text that is used to determine a business risk of a target business corresponding to the business text; Processing the first search result based on the search model to obtain a second search result, wherein the second search result is a subset of the first search result, and the correlation between the second search result and the business risk is greater than the correlation between the first search result and the business risk; A text generation model is called to process the second search result to generate an output text, where the output text is at least used to characterize the business risk level of the target business.

2. The method according to claim 1, characterized in that Performing a text content search on the business text to obtain a first search result, including: Segmenting the business text to obtain a segmentation result, wherein the segmentation result includes multiple text blocks; The segmentation result is compared with a preset statement corresponding to the target business to obtain a first search result.

3. The method according to claim 2, characterized in that Comparing the segmentation result with the preset statement corresponding to the target business to obtain a first search result, including: Performing vectorization processing on the multiple text blocks and the preset sentence to obtain a segmentation vector corresponding to each text block and a target vector corresponding to the preset sentence; respectively determining the similarity between each of the segmentation vectors and the target vector; The first search result is obtained according to the similarity.

4. The method according to claim 1, wherein The retrieval model includes a density-based noise application spatial clustering model, and processing the first retrieval result based on the retrieval model to obtain a second retrieval result includes: Determining a cluster radius and a minimum number of points corresponding to the density-based noise application spatial clustering model; performing clustering processing on the first search result according to the cluster radius and the minimum number of points to obtain a clustering result, wherein the clustering result includes at least one cluster, and the cluster includes multiple segmentation vectors; The segmentation vector in the cluster is determined as the second search result.

5. The method according to claim 4, characterized in that Clustering the first search results to obtain clustering results includes: Marking each segmentation vector in the first search result according to the cluster radius and the minimum number of points to obtain a marking result, wherein the marking result includes core points, boundary points, and noise points, and the core point is a segmentation vector that contains at least the minimum number of segmentation vectors within the cluster radius; For any core point, all segmentation vectors within the clustering radius neighborhood of the core point are divided into initial clusters; According to the initial cluster and the cluster radius, all segmentation vectors within the cluster radius neighborhood of each core point in the initial cluster are divided into the initial cluster to obtain the cluster cluster; The segmentation vectors that are not classified into any cluster are determined as noise points to obtain the clustering result.

6. The method according to claim 1, characterized in that The retrieval model includes a large language model, and processing the first retrieval result based on the retrieval model to obtain a second retrieval result includes: determining a first prompt word according to the text blocks corresponding to the segmentation vectors in the first search result and the target business, wherein the first prompt word is used to instruct the large language model to generate the second search result; Determining, based on the first prompt word, a semantic relevance between the text block corresponding to each segmentation vector in the first search result and the target business, wherein the semantic relevance indicates a degree of semantic association between the target business and the text block; A second search result is determined based on the semantic relevance.

7. The method according to any one of claims 1 to 6, characterized in that The retrieval model includes a density-based noise application space clustering model and a large language model, and the processing of the first retrieval result based on the retrieval model to obtain a second retrieval result includes: Processing the first search result based on the density-based noise application spatial clustering model to obtain a density-based noise application spatial clustering search result; Processing the first search result based on the large language model to obtain a large language search result; The density-based noise application spatial clustering retrieval result and the large language retrieval result are merged and deduplicated to obtain a second retrieval result.

8. The method according to any one of claims 1 to 6, characterized in that The retrieval model includes a density-based noise application space clustering model and a large language model, and the processing of the first retrieval result based on the retrieval model to obtain a second retrieval result includes: Processing the first search result based on the density-based noise application spatial clustering model to obtain a density-based noise application spatial clustering search result, and processing the density-based noise application spatial clustering search result based on the large language model to obtain the second search result; Alternatively, the first search result is processed based on the large language model to obtain a large language search result, and the large language search result is processed based on the density-based noise application spatial clustering model to obtain the second search result.

9. A text processing device, characterized in that: include: An acquisition module, configured to acquire a business text and perform a text content search on the business text to obtain a first search result, wherein the first search result represents text content in the business text used to determine a business risk of a target business corresponding to the business text; a first processing module, configured to process the first search result based on a search model to obtain a second search result, wherein the second search result is a subset of the first search result, and a correlation between the second search result and the business risk is greater than a correlation between the first search result and the business risk; The second processing module is configured to call a text generation model to process the second search result and generate an output text, wherein the output text is at least used to characterize the business risk level of the target business.

10. A computing device, characterized in that The computing device includes a processor connected to a memory; The memory is used to store computer-executable instructions; The processor executes the computer instructions stored in the memory to enable the computing device to implement the method according to any one of claims 1 to 8.

Images