An information processing method and apparatus

Abstract

The application relates to the field of artificial intelligence, in particular to an information processing method and device. The method comprises the following steps: obtaining a first file of a target service, and obtaining a first demand feature of the target service; the first file comprises a text for describing the target service; the first demand feature comprises feature data of n dimensions corresponding to the target service, and n is a positive integer. The first file is input into a document analysis model to obtain a first inference result output by the document analysis model. The first demand feature is input into a feature analysis model to obtain a second inference result output by the feature analysis model. According to the first inference result and the second inference result, a target service execution mode corresponding to the target service is determined; the target service execution mode is one of m service execution modes. The method is used for accurately and efficiently selecting a service execution mode suitable for a service from multiple different service execution modes.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and more particularly to an information processing method and apparatus. Background Technology

[0002] Currently, for organizations such as enterprises, public institutions, and social groups, as the types of business they handle increase, they often need to determine the appropriate business execution method from a variety of different business execution methods.

[0003] Therefore, how to accurately and efficiently select the appropriate business execution method from a variety of different business execution methods is a problem that needs to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides an information processing method and apparatus.

[0005] In a first aspect, this application provides an information processing method, comprising: acquiring a first document of a target business, and acquiring a first requirement feature of the target business; the first document includes text describing the target business; the first requirement feature includes n-dimensional feature data corresponding to the target business, where n is a positive integer. The first document is input into a document analysis model to obtain a first inference result output by the document analysis model; the first inference result includes the relevance of each of m business execution methods to the target business, where m is a positive integer. The first requirement feature is input into a feature analysis model to obtain a second inference result output by the feature analysis model; the second inference result includes the relevance of each of the m business execution methods to the target business. Based on the first inference result and the second inference result, a target business execution method corresponding to the target business is determined; the target business execution method is one of the m business execution methods.

[0006] This application takes into account two main points: First, business documents may contain missing information, or, in other words, when determining the business execution method, there may be other requirement characteristics besides the content in the documents that affect the execution method. Therefore, determining the business execution method requires analyzing both the business documents and the requirement characteristics outside the documents. Second, business documents are typically lengthy texts written in natural language (e.g., documents may include textual information such as the specific content of the business, completion deadlines, and budget amounts), while business requirement characteristics can be individual data points (e.g., business requirement priorities, business development schedules, historical problem escape rates, and problem score ratios). Therefore, the amount of data in requirement characteristics is usually smaller than that in documents. Thus, directly inputting both business documents and requirement characteristics into the model may lead to inaccurate model output, resulting in a mismatch between the determined business execution method and the actual business needs. Therefore, in this embodiment, when determining the execution method of the target business, a document analysis model is used to analyze the first document to obtain a first inference result, and a feature analysis model is used to analyze the first requirement feature to obtain a second inference result. Then, based on the first and second inference results, the target business execution method corresponding to the target business is determined. This avoids inputting both the business documents and requirement features into the model, which could lead to a mismatch between the determined execution method and the business. Thus, it achieves the goal of accurately and efficiently selecting the appropriate execution method from a variety of different execution methods.

[0007] In some implementations, the method further includes: obtaining k demand features whose similarity to the first demand feature satisfies a first similarity condition, and the business execution methods corresponding to each of the k demand features, where k is a positive integer. Using the k demand features and their corresponding business execution methods, a large language model is fine-tuned to obtain a feature analysis model.

[0008] In the above implementation, by using k demand features and the corresponding business execution methods for each of the k demand features to fine-tune the large language model, the second inference result output by the feature analysis model can include feature information that is more suitable for determining the business execution method. This makes the target business execution method determined based on the first and second inference results more in line with actual needs.

[0009] In some implementations, obtaining k demand features whose similarity to the first demand feature satisfies a first similarity condition, and the business execution methods corresponding to the k demand features respectively, includes: inputting the first demand feature into a feature vectorization model to obtain the feature vector corresponding to the first demand feature output by the feature vectorization model; the feature vectorization model is a machine learning model that maps demand features to a vector space; determining the k feature vectors with the highest similarity to the feature vectors corresponding to the first demand feature from a feature vector library; the feature vector library includes feature vectors corresponding to multiple demand features respectively; the multiple demand features respectively correspond to one of the m business execution methods; determining the k demand features corresponding to the k feature vectors, and the business execution methods corresponding to the k demand features respectively.

[0010] Through the above implementation method, k demand features that meet the first similarity condition with the first demand feature can be quickly and accurately identified from multiple demand features, and the business execution methods corresponding to the k demand features can be determined respectively.

[0011] In some implementations, the method further includes: training the initial model using a cosine similarity loss function and a cross-entropy loss function to obtain the feature vectorization model; wherein the cosine similarity loss function is used to improve the cosine similarity between feature vectors of different requirements corresponding to the same business execution method; and the cross-entropy loss function is used to reduce the difference between the inference results output by the feature analysis model and the true distribution.

[0012] Through the above implementation, when using the feature vectorization model for feature vector extraction, on the one hand, the cosine similarity of the feature vectors corresponding to the requirements features with the same business execution method can be made as high as possible; on the other hand, the feature vectors output by the model can be used to select k appropriate requirements features to fine-tune the large language model, so that the difference between the inference results output by the fine-tuned feature analysis model and the true distribution is as small as possible.

[0013] In some implementations, the method further includes: obtaining p demand texts that satisfy a second similarity condition with the first file, and the business execution methods corresponding to the p demand texts, where p is a positive integer; and fine-tuning a large language model using the p demand texts and the business execution methods corresponding to the p demand texts to obtain the document analysis model.

[0014] In the above implementation, by using p requirement texts and the business execution methods corresponding to each p requirement text to fine-tune the large language model, the first inference result output by the document analysis model can include feature information that is more suitable for determining the business execution method. This makes the target business execution method determined based on the first and second inference results more in line with actual needs.

[0015] In some implementations, the method further includes: obtaining p requirement texts that satisfy a second similarity condition with the first file, and the business execution methods corresponding to the p requirement texts respectively; obtaining a business logic document; the business logic document includes normative documents of the organization to which the target business belongs; and fine-tuning a large language model using the p requirement texts, the business execution methods corresponding to the p requirement texts respectively, and the business logic document to obtain the document analysis model.

[0016] In the above implementation, by using p requirement texts, the business execution methods corresponding to the p requirement texts, and the business logic document to fine-tune the large language model, the first inference result output by the document analysis model can include feature information that is more suitable for determining the business execution method. This makes the target business execution method determined based on the first and second inference results more in line with actual needs.

[0017] In some implementations, obtaining p demand texts that satisfy a second similarity condition with the first file, and the business execution methods corresponding to the p demand texts respectively, includes: using a general semantic vector model to determine the feature vector corresponding to the first file; determining the p feature vectors with the highest similarity to the feature vector corresponding to the first file from a text vector library; the text vector library includes feature vectors corresponding to multiple demand texts respectively; the multiple demand texts respectively correspond to one of the m business execution methods; determining the p demand texts corresponding to the p feature vectors, and the business execution methods corresponding to the p demand texts respectively.

[0018] Using the above implementation method, p requirement texts that meet the first similarity condition with the first document can be quickly and accurately identified from multiple requirement texts, and the business execution method corresponding to each of the p requirement texts can be determined.

[0019] In some implementations, determining the feature vector corresponding to the first file using a general semantic vector model includes: inputting multiple text segments included in the first file into the general semantic vector model to obtain feature vectors corresponding to the multiple text segments respectively; determining the p feature vectors with the highest similarity to the feature vectors corresponding to the first file from the text vector library includes: determining a first similarity set based on the feature vectors corresponding to the multiple text segments and the text vector library; the first similarity set includes: the similarity between the feature vectors corresponding to the multiple text segments and each feature vector in the text vector library; determining p feature vectors based on the first similarity set; the p feature vectors are the p feature vectors in the text vector library corresponding to the highest q similarity in the first similarity set, where q≥p.

[0020] Using the above implementation method, p requirement texts that meet the first similarity condition with the first document can be quickly and accurately identified from multiple requirement texts, and the business execution method corresponding to each of the p requirement texts can be determined.

[0021] In some implementations, obtaining the first requirement feature of the target business includes: obtaining features to be processed; the features to be processed include the original feature data of the n dimensions corresponding to the target business; eliminating the dimensional differences of the original feature data of the n dimensions to obtain intermediate feature data of the n dimensions; and converting the intermediate feature data of the n dimensions into numerical data to obtain the first requirement feature.

[0022] The above implementation method enables the feature data of each dimension in the first requirement feature to be unified in terms of scale and to be unified as numerical data. This makes it easier for the feature vectorization model to extract the feature vector of the first requirement feature and for the feature analysis model to extract the second inference result of the first document.

[0023] In some implementations, m = 2; there are m types of business execution methods, including: assigning business to a preset object set and not assigning business to a preset object set.

[0024] The above implementation method allows for a quick and accurate determination of whether to assign a business function to a preset object set.

[0025] In a second aspect, an information processing apparatus is provided, comprising: an acquisition unit, configured to acquire a first file of a target business and acquire a first requirement feature of the target business; the first file includes text describing the target business; the first requirement feature includes feature data of n dimensions corresponding to the target business, where n is a positive integer; an inference unit, configured to input the first file into a document analysis model to obtain a first inference result output by the document analysis model; the first inference result includes the degree of relevance between each of the m business execution methods and the target business, where m is a positive integer; the inference unit is further configured to input the first requirement feature into a feature analysis model to obtain a second inference result output by the feature analysis model; the second inference result includes the degree of relevance between each of the m business execution methods and the target business; and a determination unit, configured to determine a target business execution method corresponding to the target business based on the first inference result and the second inference result; the target business execution method is one of the m business execution methods.

[0026] In some implementations, the acquisition unit is further configured to acquire k demand features whose similarity to the first demand feature satisfies the first similarity condition, and the business execution methods corresponding to the k demand features respectively, where k is a positive integer; the inference unit is further configured to use the k demand features and the business execution methods corresponding to the k demand features respectively to fine-tune the large language model to obtain the feature analysis model.

[0027] In some implementations, the acquisition unit is further configured to acquire k demand features whose similarity to the first demand feature satisfies a first similarity condition, and the business execution methods corresponding to the k demand features respectively, including: the acquisition unit is further configured to input the first demand feature into a feature vectorization model to obtain the feature vector corresponding to the first demand feature output by the feature vectorization model; the feature vectorization model is a machine learning model that maps demand features to a vector space; the acquisition unit is further configured to determine the k feature vectors with the highest similarity to the feature vector corresponding to the first demand feature from a feature vector library; the feature vector library includes feature vectors corresponding to multiple demand features respectively; the multiple demand features respectively correspond to one of the m business execution methods; the acquisition unit is further configured to determine the k demand features corresponding to the k feature vectors, and the business execution methods corresponding to the k demand features respectively.

[0028] In some implementations, the inference unit is further used to train the initial model using a cosine similarity loss function and a cross-entropy loss function to obtain the feature vectorization model; wherein, the cosine similarity loss function is used to improve the cosine similarity between feature vectors of different demand features corresponding to the same business execution method; and the cross-entropy loss function is used to reduce the difference between the inference result output by the feature analysis model and the true distribution.

[0029] In some implementations, the acquisition unit is further configured to acquire p demand texts that satisfy the second similarity condition with the first file, and the business execution methods corresponding to the p demand texts respectively, where p is a positive integer; the inference unit is further configured to use the p demand texts and the business execution methods corresponding to the p demand texts respectively to fine-tune the large language model to obtain the document analysis model.

[0030] In some implementations, the acquisition unit is further configured to acquire p requirement texts that satisfy a second similarity condition with the first file, and the business execution methods corresponding to the p requirement texts respectively; the acquisition unit is further configured to acquire a business logic document; the business logic document includes normative documents of the organization to which the target business belongs; the reasoning unit is further configured to fine-tune the large language model using the p requirement texts, the business execution methods corresponding to the p requirement texts respectively, and the business logic document to obtain the document analysis model.

[0031] In some implementations, the acquisition unit is further configured to acquire p demand texts whose similarity to the first file satisfies a second similarity condition, and the business execution methods corresponding to the p demand texts respectively. This includes: the acquisition unit is further configured to determine the feature vector corresponding to the first file using a general semantic vector model; the acquisition unit is further configured to determine the p feature vectors with the highest similarity to the feature vector corresponding to the first file from a text vector library; the text vector library includes feature vectors corresponding to multiple demand texts respectively; the multiple demand texts respectively correspond to one of the m business execution methods; the acquisition unit is further configured to determine the p demand texts corresponding to the p feature vectors, and the business execution methods corresponding to the p demand texts respectively.

[0032] In some implementations, the acquisition unit is further configured to determine the feature vector corresponding to the first file using a general semantic vector model, including: the acquisition unit is further configured to input the multiple text segments included in the first file into the general semantic vector model respectively to obtain the feature vectors corresponding to the multiple text segments respectively; the acquisition unit is further configured to determine the p feature vectors with the highest similarity to the feature vectors corresponding to the first file from the text vector library, including: the acquisition unit is further configured to determine a first similarity set based on the feature vectors corresponding to the multiple text segments respectively and the text vector library; the first similarity set includes: the similarity between the feature vectors corresponding to the multiple text segments respectively and each feature vector in the text vector library; the acquisition unit is further configured to determine p feature vectors based on the first similarity set; the p feature vectors are the p feature vectors in the text vector library corresponding to the highest q similarities in the first similarity set, where q≥p.

[0033] In some implementations, the acquisition unit is used to acquire the first requirement feature of the target service, including: an acquisition unit for acquiring features to be processed; the features to be processed include the original feature data of the n dimensions corresponding to the target service; the acquisition unit is further used to eliminate the dimensional differences of the original feature data of the n dimensions to obtain intermediate feature data of the n dimensions; the acquisition unit is further used to convert the intermediate feature data of the n dimensions into numerical data to obtain the first requirement feature.

[0034] In some implementations, m = 2; the m types of business execution methods include: assigning the business to a preset object set and not assigning the business to a preset object set.

[0035] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor, wherein the memory is used to store a computer program and the processor is used to cause the electronic device to implement the information processing method described in any of the above implementations when executing the computer program.

[0036] Fourthly, embodiments of this application provide a computer-readable storage medium that, when executed by a computing device, causes the computing device to implement the information processing method described in any of the above implementations.

[0037] Fifthly, embodiments of this application provide a computer program product that, when run on a computer, enables the computer to implement the information processing method described in any of the above implementations.

[0038] The technical solution provided in this application has the following advantages compared with the prior art:

[0039] This application takes into account two main points: First, business documents may contain missing information, or, in other words, when determining the business execution method, there may be other requirement characteristics besides the content in the documents that affect the execution method. Therefore, determining the business execution method requires analyzing both the business documents and the requirement characteristics outside the documents. Second, business documents are typically lengthy texts written in natural language (e.g., documents may include textual information such as the specific content of the business, completion deadlines, and budget amounts), while business requirement characteristics can be individual data points (e.g., business requirement priorities, business development schedules, historical problem escape rates, and problem score ratios). Therefore, the amount of data in requirement characteristics is usually smaller than that in documents. Thus, directly inputting both business documents and requirement characteristics into the model may lead to inaccurate model output, resulting in a mismatch between the determined business execution method and the actual business needs. Therefore, in this embodiment, when determining the execution method of the target business, a document analysis model is used to analyze the first document to obtain a first inference result, and a feature analysis model is used to analyze the first requirement feature to obtain a second inference result. Then, based on the first and second inference results, the target business execution method corresponding to the target business is determined. This avoids inputting both the business documents and requirement features into the model, which could lead to a mismatch between the determined execution method and the business. Thus, it achieves the goal of accurately and efficiently selecting the appropriate execution method from a variety of different execution methods. Attached Figure Description

[0040] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is one of the flowcharts illustrating an information processing method provided in an embodiment of this application;

[0043] Figure 2 A second schematic flowchart illustrating an information processing method provided in an embodiment of this application;

[0044] Figure 3 The third schematic flowchart of an information processing method provided in this application embodiment;

[0045] Figure 4 The fourth schematic flowchart of an information processing method provided in this application embodiment;

[0046] Figure 5 Fifth of a flowchart illustrating an information processing method provided in this application embodiment;

[0047] Figure 6 A flowchart illustrating an information processing method provided in this application is shown in Figure 6.

[0048] Figure 7 The seventh flowchart illustrating an information processing method provided in this application embodiment;

[0049] Figure 8 This is the eighth flowchart illustrating an information processing method provided in an embodiment of this application.

[0050] Figure 9 A flowchart illustrating an information processing method provided in this application embodiment is shown in Figure 9.

[0051] Figure 10 This is a flowchart illustrating an information processing method provided in an embodiment of this application.

[0052] Figure 11 This is one of the structural schematic diagrams of an information processing device provided in an embodiment of this application;

[0053] Figure 12 This is a second schematic diagram of the structure of an information processing device provided in an embodiment of this application. Detailed Implementation

[0054] To better understand the above-mentioned objectives, features, and advantages of this application, the solution of this application will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0055] Many specific details are set forth in the following description in order to provide a full understanding of this application, but this application may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of this application, and not all embodiments.

[0056] Currently, determining the execution method for a business task typically requires manual review of the relevant documents and assessment of the appropriate execution method. For example, when distributing business processes, manual review of the documents and assessment of whether to assign the task to a predefined set of objects are usually required. This predefined set of objects can be a collection of multiple business execution entities (e.g., multiple business departments or multiple companies). While this method can select a suitable execution method for the business task, its accuracy is easily affected by human factors, and it is also inefficient.

[0057] To accurately and efficiently select the appropriate business execution method from various options, this embodiment considers the following: Large Language Models (LLMs) can be used to analyze the business documents and determine the appropriate execution method based on the analysis results. However, since the documents may contain complex language structures and implicit semantic relationships, simply combining natural language processing techniques to determine the appropriate execution method will result in poor performance.

[0058] Furthermore, to avoid the aforementioned performance issues, the embodiments of this application take into account:

[0059] On the one hand, there may be missing information in the business documents, or in other words, when determining the business execution method, there may be other requirement characteristics besides the content in the documents that may affect the business execution method. Therefore, when determining the business execution method, it is necessary to analyze not only the business documents, but also the requirement characteristics outside the documents.

[0060] On the other hand, business documents are typically lengthy texts written in natural language (e.g., documents may include textual information such as the specific content of the business, completion deadlines, and budget amounts), while business requirements can be individual data points (e.g., business requirement priorities, development schedules, etc.). Therefore, the amount of data in requirements is usually smaller than that in documents. Consequently, directly inputting both business documents and requirements into the model may result in the analysis results failing to effectively reflect the aspects of business requirements.

[0061] Therefore, in the embodiments of this application, as Figure 1As shown, when determining the execution method of a target business, on the one hand, the target business's documents can be input into a model for analyzing documents (referred to as the document analysis model in this embodiment) to obtain a first inference result, where the first inference result includes the degree of relevance between each of the m execution methods and the target business; on the other hand, the target business's demand characteristics can be input into a model for analyzing demand characteristics (referred to as the feature analysis model in this embodiment) to obtain a second inference result, where the second inference result includes the degree of relevance between each of the m execution methods and the target business. Furthermore, based on the first and second inference results, the target business execution method corresponding to the target business can be determined.

[0062] Based on the above considerations, this application provides an information processing method. The executing entity of the method provided in this application can be an information processing device. When the information processing device is running, it can be used to execute all or part of the steps in the method provided in this application.

[0063] like Figure 2 As shown, the method may include:

[0064] S201. The information processing device obtains the first document of the target service and the first requirement characteristics of the target service.

[0065] The first file includes text describing the target business.

[0066] For example, the first document can be a document containing information such as the specific details of the business, the completion deadline, and the cost. In practical applications, the first document can be a file in various formats that are easy for business personnel to read. For example, the first document can be a document format, a table format, or a slide format.

[0067] The first requirement feature includes: n dimensions of feature data corresponding to the target business, where n is a positive integer.

[0068] For example, the aforementioned n-dimensional feature data may include feature data of the target business that is not recorded in the first file. For instance, if the first file includes information such as the specific content of the business, the completion deadline, and the budget amount, but does not include: priority information indicating the priority of the target business requirements; scheduling information indicating the development schedule of the business; the problem escape rate indicating the target business (where the problem escape rate can be understood as the proportion of problems that are not discovered and fixed in time and eventually flow into the production environment or release version); and the problem score ratio indicating the target business (where the problem score ratio can be understood as the ratio of the number of problems to the development time). Since the aforementioned priority information, scheduling information, problem escape rate, and problem score ratio have an impact on the choice of business execution method, in this case, the aforementioned four dimensions of feature data—priority information, scheduling information, problem escape rate, and problem score ratio—can be included in the first requirement feature.

[0069] In some implementations, such as Figure 3 As shown, the first requirement characteristics of the target business are obtained, which may specifically include the following S2011-S2013:

[0070] S2011, The information processing device acquires the features to be processed.

[0071] The features to be processed include raw feature data of n dimensions corresponding to the target business.

[0072] For example, taking priority information, scheduling information, problem escape rate, and problem score ratio as four dimensions: the original feature data of the above four dimensions corresponding to the target business are included in Table 1 below:

[0073] Table 1

[0074] Priority information Schedule information Problem escape rate Problem estimation ratio Grade A 1 month 5％ 0.03 (pieces / hour)

[0075] S2012. The information processing device eliminates the dimensional differences of the original feature data of n dimensions and obtains intermediate feature data of n dimensions.

[0076] Eliminating the dimensional differences of the original feature data across n dimensions can specifically include: eliminating dimensional differences between different feature data within the same dimension. Taking "scheduling information" as an example, by eliminating dimensional differences, the feature data of the "scheduling information" dimension can be unified to be in "days".

[0077] Furthermore, eliminating the dimensional differences in the original feature data across n dimensions can specifically include: eliminating the dimensional differences between feature data of different dimensions. For example, taking "scheduling information" and "problem score estimation ratio" as examples, the value of scheduling information may be in the single digit range, while the value of problem score estimation ratio may be in the 0.01 range. If subsequent processing (such as model inference) is performed using "scheduling information" and "problem score estimation ratio" without eliminating dimensional differences, the processing results may have significant errors. Therefore, it is necessary to eliminate the dimensional differences between feature data of different dimensions.

[0078] S2013. The information processing device converts the intermediate feature data of n dimensions into numerical data to obtain the first demand feature.

[0079] For example, taking "priority information" as an example, by converting the intermediate state feature data of n dimensions into numerical data, the "A level" included in the "priority information" can be converted into numerical data, so as to improve the processing capability of the model when performing model inference later.

[0080] S202, The information processing device inputs the first file into the document analysis model and obtains the first inference result output by the document analysis model.

[0081] The first inference result includes the degree of relevance between each of the m business execution methods and the target business. A higher degree of relevance between a business execution method and the target business indicates a greater probability that this method will be selected as the target business execution method.

[0082] For example, a document analysis model can be a large language model that determines the target business execution method from m types of business execution methods based on the input file.

[0083] The first inference result can be the raw, unnormalized predicted value (i.e., logits) output by the document analysis model based on the input file. This raw predicted value can include scores corresponding to m different business execution methods, where each score reflects the relevance of the execution method to the target business.

[0084] S203. The information processing device inputs the first demand feature into the feature analysis model and obtains the second reasoning result output by the feature analysis model.

[0085] The second inference result includes the degree of relevance between each of the m business execution methods and the target business. A higher degree of relevance between a business execution method and the target business indicates a greater probability that this method will be selected as the target business execution method.

[0086] For example, a feature analysis model can be a large language model that determines the target business execution method corresponding to the target business from m types of business execution methods based on the input demand features.

[0087] The second inference result can be the raw, unnormalized predicted value (i.e., logits) output by the feature analysis model based on the input demand features. The raw predicted value can include scores corresponding to m different business execution methods, where each score reflects the relevance of the execution method to the target business.

[0088] S204. The information processing device determines the target service execution method corresponding to the target service based on the first reasoning result and the second reasoning result.

[0089] Among them, the target business execution method is one of m business execution methods.

[0090] For example, the information processing device can add the first inference result to the unnormalized original predicted value in the second inference result, and then perform normalization processing (e.g., using the softmax function) to obtain the probabilities corresponding to each of the m business execution methods. Furthermore, based on the probabilities corresponding to the m business execution methods, the target business execution method can be determined.

[0091] In one implementation, the m types of business execution methods can include: assigning the business to a preset object set and not assigning the business to a preset object set, i.e., m = 2 in this case.

[0092] The first inference result mentioned above can include scores corresponding to two different business execution methods (i.e., assigning the business to a preset object set and not assigning the business to a preset object set). Similarly, the second inference result can also include scores corresponding to these two methods. By adding the scores from the first and second inference results and normalizing them, the probabilities corresponding to the two business execution methods are obtained. Furthermore, based on the probabilities corresponding to the two business execution methods, the target business execution method can be determined.

[0093] Furthermore, in this embodiment, the aforementioned feature analysis model can be obtained by fine-tuning a large language model. The model fine-tuning process is described below. In one implementation, such as... Figure 4 As shown, before inputting the first demand feature into the feature analysis model in S203 to obtain the second inference result output by the feature analysis model, the method may further include the following S205-S206:

[0094] S205. The information processing device acquires k demand features whose similarity to the first demand feature satisfies the first similarity condition, and the business execution methods corresponding to the k demand features respectively.

[0095] For example, a requirement feature library can be established in advance, which includes multiple historical requirement features collected in advance, as well as the business execution methods corresponding to each of the multiple historical requirement features.

[0096] The first similarity condition can be: the k demand features in the demand feature library that have the highest similarity to the first demand feature.

[0097] S206. The information processing device uses k demand features and the business execution methods corresponding to the k demand features to fine-tune the large language model and obtain the feature analysis model.

[0098] For example, the information processing device can utilize k demand features and the corresponding business execution methods for each of the k demand features, and employ the Low-Rank Adaptation of Large Language Models (LoRA) technique to fine-tune the large language model to obtain a feature analysis model. Then, the information processing device can input the first demand feature into the feature analysis model according to the process described in S203 above, and obtain the second inference result output by the feature analysis model.

[0099] Understandably, in practical applications, in one scenario, k requirement features and their corresponding business execution methods can be input into a large language model to fine-tune the model and obtain a feature analysis model. Then, the first requirement feature is input into the feature analysis model to obtain the second inference result output by the model. In another scenario, a few-shot dataset can be constructed, containing k requirement features, their corresponding business execution methods, and the first requirement feature. This few-shot dataset is then input into the large language model, allowing it to be fine-tuned using the few-shot dataset (i.e., fine-tuned using the k requirement features and their corresponding business execution methods). Finally, the fine-tuned feature analysis model outputs the second inference result corresponding to the first requirement feature.

[0100] In addition, in one design, it can be achieved through Figure 5The content of S2051-S2053 involves obtaining k requirement features whose similarity to the first requirement feature satisfies the first similarity condition, and the business execution methods corresponding to each of the k requirement features (i.e., S205):

[0101] S2051. The information processing device inputs the first demand feature into the feature vectorization model to obtain the feature vector corresponding to the first demand feature.

[0102] Among them, the feature vectorization model can be a machine learning model that maps demand features to a vector space.

[0103] Specifically, a feature vectorization model can include an input layer, a hidden layer, and an output layer.

[0104] The input layer provides high-quality input data (i.e., demand features) for subsequent hidden layers. The hidden layers are used to understand and analyze these demand features. Specifically, hidden layers can be composed of multiple fully connected layers (FC). Each fully connected layer can introduce non-linearity through activation functions (such as Rectified Linear Units, ReLU) to enhance the model's expressive power. The main function of these fully connected layers is to progressively extract more abstract and higher-level feature representations from the input data. Through layer-by-layer abstraction, the feature vectorization model can capture the complex relationships and inherent semantic information between demand features. In other words, the design of the hidden layers aims to improve the model's generalization ability, enabling it to adapt to different types of demand features. Through multiple layers of non-linear transformations, the feature vectorization model can not only identify simple linear relationships but also capture more complex non-linear relationships, thereby improving the ability to understand and analyze demand features. Furthermore, the output layer can be a multi-dimensional (e.g., 1024-dimensional) fully connected layer, which translates the high-level feature representations into a fixed-dimensional vector space.

[0105] In one possible design, the method may also include:

[0106] The initial model is trained using the cosine similarity loss function and the cross-entropy loss function to obtain the feature vectorization model.

[0107] The cosine similarity loss function is used to improve the cosine similarity between feature vectors of different requirements corresponding to the same business execution method. This helps the model learn how to cluster feature vectors corresponding to similar requirements in the vector space.

[0108] The cross-entropy loss function is used to reduce the discrepancy between the inference results output by the feature analysis model and the true distribution. This optimizes the performance of the feature analysis model on classification tasks.

[0109] In practical implementation, the weights of the cosine similarity loss function and the cross-entropy loss function in the total loss can be set according to actual needs. For example, the weights of the cosine similarity loss function and the cross-entropy loss function in the total loss can be adjusted to a ratio of 4:6. In the above possible design, by combining the cosine similarity loss function and the cross-entropy loss function as the loss function in the training process of the feature vectorization model, it is possible to cluster the feature vectors corresponding to the required features of the same category in the vector space, and at the same time, it helps to ensure the classification accuracy of the feature analysis model.

[0110] S2052, The information processing device determines the k feature vectors with the highest similarity to the feature vectors corresponding to the first requirement feature from the feature vector library.

[0111] The feature vector library includes feature vectors corresponding to multiple requirement features. Furthermore, each of these multiple requirement features corresponds to one of m business execution methods. For example, these multiple requirement features could be requirement features corresponding to multiple distributed business processes, and the business execution methods corresponding to these multiple requirement features could specifically be the business execution methods distributed to each business process. Exemplarily, in practical applications, these multiple requirement features can be input into a feature vectorization model for vectorization, thereby obtaining the feature vectors corresponding to each requirement feature, thus completing the construction of the aforementioned feature vector library.

[0112] The information processing device can calculate the similarity between each feature vector in the feature vector library and the feature vector corresponding to the first requirement feature, so as to determine the k feature vectors with the highest similarity to the feature vector corresponding to the first requirement feature.

[0113] S2053. The information processing device determines the k demand features corresponding to the k feature vectors, and the business execution methods corresponding to the k demand features respectively.

[0114] Among them, the k demand features corresponding to the k feature vectors can be understood as the k demand features whose similarity to the first demand feature satisfies the first similarity condition.

[0115] The following sections will introduce the model fine-tuning process using two different implementation methods.

[0116] In the first implementation, such as Figure 6As shown, before inputting the first file into the document analysis model in S202 to obtain the first inference result output by the document analysis model, the method may further include the following S207-S208:

[0117] S207. The information processing device acquires p demand texts whose similarity to the first file satisfies the second similarity condition, and the business execution methods corresponding to the p demand texts respectively, where p is a positive integer.

[0118] For example, a requirement text library can be pre-established, which includes multiple historical requirement texts collected beforehand. Each historical requirement text can specifically include paragraphs or sentences from documents containing distributed business requirements. Each of these historical requirement texts corresponds to one of m business execution methods; for example, the business execution method corresponding to each historical requirement text can be the business execution method that was distributed to the corresponding business.

[0119] The second similarity condition can be specifically the p most similar requirement texts in the requirement text library that are most similar to the first file.

[0120] S208. The information processing device uses p requirement texts and the business execution methods corresponding to each p requirement text to fine-tune the large language model and obtain a document analysis model.

[0121] For example, the information processing device can utilize p requirement texts and the corresponding business execution methods for each of the p requirement texts, and employ LoRA technology for large language models to fine-tune the large language model to obtain a document analysis model. Then, the information processing device can input the first file into the document analysis model according to the process described in S202 above, and obtain the first inference result output by the document analysis model.

[0122] Understandably, in practical applications, in one scenario, p requirement texts and their corresponding business execution methods can be input into a large language model to fine-tune the model and obtain a document analysis model. Then, the first document is input into the document analysis model to obtain the first inference result. In another scenario, a few-shot dataset can be constructed, containing p requirement texts, their corresponding business execution methods, and the first document. This few-shot dataset is then input into the large language model, allowing it to be fine-tuned using the p requirement texts and their corresponding business execution methods. Finally, the fine-tuned document analysis model outputs the first inference result corresponding to the first document.

[0123] In addition, in one design, it can be achieved through Figure 7 From the content of S2071-S2073, obtain p requirement texts that satisfy the second similarity condition with the first file, and the business execution methods corresponding to each of the p requirement texts (i.e., S207):

[0124] S2071. The information processing device uses a general semantic vector model to determine the feature vector corresponding to the first file.

[0125] For example, an information processing device can input a first file into a general semantic vector model to obtain the feature vector corresponding to the first file output by the general semantic vector model.

[0126] S2072, The information processing device determines the p feature vectors with the highest similarity to the feature vectors corresponding to the first file from the text vector library.

[0127] The text vector library includes feature vectors corresponding to multiple requirement texts. These requirement texts each correspond to one of m different business execution methods.

[0128] For example, on the one hand, multiple historical requirement texts from the requirement text library mentioned above can be input into a general semantic vector model to obtain the feature vectors corresponding to each historical requirement text; on the other hand, the feature vectors corresponding to each historical requirement text can be used to construct a text vector library.

[0129] In one possible design, the information processing device can calculate the similarity between each feature vector in the text vector library and the feature vector corresponding to the first file, so as to determine the p feature vectors with the highest similarity to the feature vector corresponding to the first file.

[0130] S2073. The information processing device determines the p demand texts corresponding to the p feature vectors, and the business execution methods corresponding to the p demand texts respectively.

[0131] Among them, the p demand texts corresponding to the p feature vectors can be understood as the p demand texts whose similarity to the first demand feature satisfies the second similarity condition.

[0132] In one possible design, considering that business documents may be lengthy, the first document can be divided into multiple text segments, and the feature vector of each segment can be determined. Then, using the feature vectors of each text segment, p feature vectors can be determined. Specifically, such as... Figure 8 As shown, S2071 may specifically include the following S2071a:

[0133] S2071a The information processing device inputs the multiple text segments included in the first document into the general semantic vector model to obtain the feature vectors corresponding to the multiple text segments respectively.

[0134] For example, the first file can be divided into multiple text segments, with each natural paragraph constituting one text segment. These segments can then be input into a general semantic vector model to obtain feature vectors for each text segment. Alternatively, the first file can be divided into multiple text segments, with each sentence constituting one text segment. These segments can then be input into a general semantic vector model to obtain feature vectors for each text segment.

[0135] In addition, S2072 may specifically include the following S2072a-S2072b:

[0136] S2072a The information processing device determines the first similarity set based on the feature vectors corresponding to multiple text segments and the text vector library.

[0137] The first similarity set includes the similarity between the feature vectors corresponding to multiple text paragraphs and the feature vectors in the text vector library.

[0138] For example, if the first file is divided into 5 text segments, the feature vectors corresponding to the 5 text segments respectively include feature vector x1, feature vector x2, feature vector x3, feature vector x4, and feature vector x5; in addition, the text vector library includes 5 feature vectors: feature vector y1, feature vector y2, feature vector y3, feature vector y4, and feature vector y5. Then, by calculating the similarity between the feature vectors corresponding to the 5 text segments and the 5 feature vectors in the text vector library in turn, the 25 similarity scores included in the first similarity set can be obtained.

[0139] S2072b: The information processing device determines p feature vectors based on the first similarity set.

[0140] Here, the p feature vectors are the p feature vectors in the text vector library corresponding to the q highest similarity in the first similarity set, where q≥p.

[0141] For example, continuing to divide the first file into 5 text segments, each segment has a corresponding feature vector x1, x2, x3, x4, and x5. The text vector library also contains 5 feature vectors: y1, y2, y3, y4, and y5. After calculating the similarity between the feature vectors corresponding to the 5 text segments and the 5 feature vectors in the text vector library, resulting in 25 similarity scores in the first similarity set, these 25 similarity scores are sorted from largest to smallest. The results are assumed to be as follows:

[0142] S(x1,y1)→S(x2,y2)→S(x3,y1)→S(x3,y2)→S(x2,y5)→S(x4,y3)…

[0143] Here, S(A,B) represents the similarity between objects A and B. Furthermore, for simplicity, only the top 6 similarity scores are shown in the above ranking.

[0144] In the example above, if p is 3, then the p feature vectors include: feature vector y1, feature vector y2, and feature vector y5, which are the three feature vectors in the text vector library corresponding to the five highest similarities in the first similarity set.

[0145] In the second implementation, such as Figure 9 As shown, before inputting the first file into the document analysis model in S202 to obtain the first inference result output by the document analysis model, the method may further include the following S209-S211:

[0146] S209. The information processing device acquires p demand texts whose similarity to the first file satisfies the second similarity condition, and the business execution methods corresponding to the p demand texts respectively, where p is a positive integer.

[0147] The specific implementation process of S209 can be referred to the content corresponding to S207 above, and the repeated parts will not be repeated here.

[0148] S210, The information processing device obtains the business logic document.

[0149] S211. The information processing device uses p requirement texts, the business execution methods corresponding to the p requirement texts, and the business logic document to fine-tune the large language model and obtain a document analysis model.

[0150] For example, the information processing device can utilize p requirement texts, the business execution methods corresponding to each of the p requirement texts, and the business logic document, and employ LoRA technology for large language models to fine-tune the large language model to obtain a document analysis model. Then, the information processing device can input the first file into the document analysis model according to the process described in S202 above, and obtain the first inference result output by the document analysis model.

[0151] Understandably, in practical applications, in one scenario, p requirement texts, their corresponding business execution methods, and business logic documents can be input into a large language model for fine-tuning to obtain a document analysis model. Then, the first file is input into the document analysis model to obtain the first inference result. In another scenario, a few-shot dataset can be constructed, including p requirement texts, their corresponding business execution methods, business logic documents, and the first file. This few-shot dataset is then input into the large language model, allowing it to be fine-tuned using the p requirement texts, their corresponding business execution methods, and business logic documents. Finally, the fine-tuned document analysis model outputs the first inference result corresponding to the first file.

[0152] The following describes the implementation process of the method provided in this application embodiment by taking the application of the method provided in this application embodiment to determine whether to assign a service to a preset object set as an example. Figure 10 As shown, the method may include:

[0153] S301. Obtain the first requirement characteristics and first document of the target business.

[0154] The specific implementation process of S301 can be referred to the corresponding content of S201 above.

[0155] On the one hand, after obtaining the first requirement characteristics and the first file of the target business, the logits output by the LLM file (which can be understood as the first inference result mentioned above) can be obtained through the following content S302-S306:

[0156] S302. Divide the first file into multiple text segments.

[0157] S303. Input multiple text paragraphs into the general semantic vector model to obtain the feature vectors corresponding to the multiple text paragraphs respectively.

[0158] The specific implementation process of S302 and S303 can be referred to the implementation process of obtaining the feature vectors corresponding to multiple text segments in S2071a above.

[0159] S304. Determine the p feature vectors with the highest similarity to the feature vectors corresponding to the first file from the text vector library.

[0160] The specific implementation process of S303 can be found in the corresponding content of S2072-S2073 above.

[0161] S305. Construct a few-shot dataset. This few-shot dataset includes p requirement texts, the corresponding business execution methods, business logic documents, and a first file for each of the p requirement texts.

[0162] S306. Input the few-shot dataset into the large language model so that the large language model is first fine-tuned according to the few-shot dataset (that is, fine-tuned using the p requirement texts, the business execution methods and business logic documents corresponding to the p requirement texts included in the few-shot dataset), and then output the logits corresponding to the first file using the fine-tuned file LLM (which can be understood as the document analysis model mentioned above).

[0163] The logits corresponding to the first file may include the original predicted values ​​assigned to the preset object set and those not assigned to the preset object set.

[0164] On the other hand, after obtaining the first requirement characteristics and the first document of the target business, the logits output by the requirement characteristic LLM can be obtained through the following content S307-S309 (which can be understood as the second inference result mentioned above).

[0165] S307. Determine the k feature vectors with the highest similarity to the feature vectors corresponding to the first requirement feature from the feature vector library.

[0166] The specific implementation process of S307 can be found in the corresponding content of S2052-S2053 above.

[0167] S308. Construct the few-shot dataset.

[0168] The few-shot dataset includes: k requirement features, the business execution methods corresponding to the k requirement features, and the first requirement feature.

[0169] S309. Input the few-shot dataset into the large language model so that the large language model is first fine-tuned according to the few-shot dataset (that is, fine-tuned using the k requirement features included in the few-shot dataset and the business execution methods corresponding to the k requirement features respectively), and then outputs the logits corresponding to the first requirement feature using the fine-tuned requirement feature LLM (which can be understood as the feature analysis model mentioned above).

[0170] The logits corresponding to the first requirement feature may include the original predicted values ​​assigned to the preset object set and those not assigned to the preset object set.

[0171] S310. Add the logits corresponding to the first requirement feature output by the LLM file and the requirement feature LLM and normalize them to obtain the probabilities corresponding to being assigned to the preset object set and not being assigned to the preset object set.

[0172] S311. Output the results based on the probabilities assigned to the preset object set and not assigned to the preset object set, respectively.

[0173] The result can be either assigned to a preset object set or not assigned to any item in the preset object set.

[0174] Based on the same inventive concept, as an implementation of the above method, this application also provides an information processing device. This embodiment corresponds to the aforementioned method embodiment. For ease of reading, this embodiment will not repeat the details of the aforementioned method embodiment one by one, but it should be clear that the information processing device in this embodiment can correspondingly implement all the contents of the aforementioned method embodiment.

[0175] Specifically, such as Figure 11 As shown, the information processing device 40 may include: an acquisition unit 401, a reasoning unit 402, and a determination unit 403. Wherein:

[0176] The acquisition unit 401 is used to acquire a first file of the target business and acquire a first requirement feature of the target business; the first file includes text describing the target business; the first requirement feature includes n-dimensional feature data corresponding to the target business, where n is a positive integer;

[0177] Inference unit 402 is used to input the first file into the document analysis model and obtain the first inference result output by the document analysis model; the first inference result includes the degree of correlation between each of the m business execution methods and the target business, where m is a positive integer;

[0178] The reasoning unit 402 is further configured to input the first demand feature into the feature analysis model to obtain a second reasoning result output by the feature analysis model; the second reasoning result includes the degree of correlation between each of the m business execution methods and the target business;

[0179] The determining unit 403 is used to determine the target service execution method corresponding to the target service based on the first reasoning result and the second reasoning result; the target service execution method is one of the m service execution methods.

[0180] In some implementations, the acquisition unit 401 is further configured to acquire k demand features whose similarity to the first demand feature satisfies the first similarity condition, and the business execution methods corresponding to the k demand features respectively, where k is a positive integer;

[0181] The reasoning unit 402 is also used to fine-tune the large language model using the k requirement features and the business execution methods corresponding to the k requirement features respectively, so as to obtain the feature analysis model.

[0182] In some implementations, the acquisition unit 401 is further configured to acquire k demand features whose similarity to the first demand feature satisfies a first similarity condition, and the business execution methods corresponding to the k demand features respectively, including:

[0183] The acquisition unit 401 is further configured to input the first demand feature into the feature vectorization model to obtain the feature vector corresponding to the first demand feature output by the feature vectorization model; the feature vectorization model is a machine learning model that maps demand features to a vector space.

[0184] The acquisition unit 401 is further configured to determine the k feature vectors with the highest similarity to the feature vectors corresponding to the first requirement feature from the feature vector library; the feature vector library includes feature vectors corresponding to multiple requirement features respectively; the multiple requirement features respectively correspond to one of the m business execution methods;

[0185] The acquisition unit 401 is also used to determine the k demand features corresponding to the k feature vectors, and the business execution methods corresponding to the k demand features respectively.

[0186] In some implementations, the inference unit 402 is further used to train the initial model using a cosine similarity loss function and a cross-entropy loss function to obtain the feature vectorization model; wherein, the cosine similarity loss function is used to improve the cosine similarity between feature vectors of different demand features corresponding to the same business execution method; and the cross-entropy loss function is used to reduce the difference between the inference result output by the feature analysis model and the true distribution.

[0187] In some implementations, the acquisition unit 401 is further configured to acquire p demand texts that satisfy the second similarity condition with the first file, and the business execution methods corresponding to the p demand texts respectively, where p is a positive integer;

[0188] The reasoning unit 401 is also used to fine-tune the large language model using the p requirement texts and the business execution methods corresponding to the p requirement texts respectively, so as to obtain the document analysis model.

[0189] In some implementations, the acquisition unit 401 is further configured to acquire p demand texts that satisfy the second similarity condition with the first file, and the business execution methods corresponding to the p demand texts respectively;

[0190] The acquisition unit 401 is also used to acquire a business logic document; the business logic document includes the normative documents of the organization to which the target business belongs;

[0191] The reasoning unit 402 is also used to fine-tune the large language model using the p requirement texts, the business execution methods corresponding to the p requirement texts respectively, and the business logic document to obtain the document analysis model.

[0192] In some implementations, the acquisition unit 401 is further configured to acquire p demand texts whose similarity to the first file satisfies a second similarity condition, and the business execution methods corresponding to the p demand texts respectively, including:

[0193] The acquisition unit 401 is also used to determine the feature vector corresponding to the first file using a general semantic vector model;

[0194] The acquisition unit 401 is further configured to determine the p feature vectors with the highest similarity to the feature vectors corresponding to the first file from the text vector library; the text vector library includes feature vectors corresponding to multiple requirement texts; the multiple requirement texts correspond to one of the m business execution methods.

[0195] The acquisition unit 401 is also used to determine the p requirement texts corresponding to the p feature vectors, and the business execution methods corresponding to the p requirement texts respectively.

[0196] In some implementations, the acquisition unit 401 is further used to determine the feature vector corresponding to the first file using a general semantic vector model, including:

[0197] The acquisition unit 401 is further configured to input the multiple text paragraphs included in the first file into a general semantic vector model to obtain feature vectors corresponding to the multiple text paragraphs respectively;

[0198] The acquisition unit 401 is further configured to determine, from the text vector library, the p feature vectors with the highest similarity to the feature vector corresponding to the first file, including:

[0199] The acquisition unit 401 is further configured to determine a first similarity set based on the feature vectors corresponding to the plurality of text segments and a text vector library; the first similarity set includes: the similarity between the feature vectors corresponding to the plurality of text segments and each feature vector in the text vector library;

[0200] The acquisition unit 401 is further configured to determine p feature vectors based on the first similarity set; the p feature vectors are the p feature vectors in the text vector library corresponding to the highest q similarities in the first similarity set, where q≥p.

[0201] In some implementations, the acquisition unit 401 is used to acquire the first requirement feature of the target service, including:

[0202] The acquisition unit 401 is used to acquire features to be processed; the features to be processed include the original feature data of the n dimensions corresponding to the target service;

[0203] The acquisition unit 401 is also used to eliminate the dimensional differences of the original feature data of the n dimensions to obtain the intermediate feature data of the n dimensions.

[0204] The acquisition unit 401 is also used to convert the intermediate state feature data of the n dimensions into numerical data to obtain the first required feature.

[0205] In some implementations, m = 2; the m types of business execution methods include: assigning the business to a preset object set and not assigning the business to a preset object set.

[0206] The information processing device 40 provided in this embodiment can perform... Figures 2-10 The method provided is similar in principle and technical effect, and will not be elaborated here.

[0207] Based on the same inventive concept, embodiments of this application also provide an electronic device. Figure 12 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application, such as... Figure 12 As shown, the electronic device provided in this embodiment includes a memory 501 and a processor 502. The memory 501 is used to store computer programs, and the processor 502 is used to execute any of the methods provided in the above embodiments when executing the computer programs.

[0208] Based on the same inventive concept, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the computing device to implement the method provided in the above embodiments.

[0209] Based on the same inventive concept, this application also provides a computer program product that, when run on a computer, enables the computing device to implement the method provided in the above embodiments.

[0210] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.

[0211] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0212] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0213] Computer-readable media include both permanent and non-permanent, removable and non-removable storage media. Storage media can store information using any method or technology; the information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0214] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. An information processing method, characterized in that, The method includes: Obtain a first file of the target business and obtain a first requirement feature of the target business; the first file includes text describing the target business; the first requirement feature includes n-dimensional feature data corresponding to the target business, where n is a positive integer; The first file is input into the document analysis model to obtain the first inference result output by the document analysis model; the first inference result includes the degree of correlation between each of the m business execution methods and the target business, where m is a positive integer; The first requirement feature is input into the feature analysis model to obtain the second inference result output by the feature analysis model; the second inference result includes the degree of correlation between each of the m business execution methods and the target business; Based on the first reasoning result and the second reasoning result, the target business execution method corresponding to the target business is determined; the target business execution method is one of the m business execution methods.

2. The method according to claim 1, characterized in that, The method further includes: Obtain k demand features whose similarity to the first demand feature satisfies the first similarity condition, and the business execution methods corresponding to the k demand features respectively, where k is a positive integer; By using the k requirement features and the corresponding business execution methods, the large language model is fine-tuned to obtain the feature analysis model.

3. The method according to claim 2, characterized in that, The step of acquiring k demand features whose similarity to the first demand feature satisfies the first similarity condition, and the business execution methods corresponding to the k demand features respectively, includes: The first demand feature is input into the feature vectorization model to obtain the feature vector corresponding to the first demand feature output by the feature vectorization model; the feature vectorization model is a machine learning model that maps the demand feature to a vector space. The k feature vectors with the highest similarity to the feature vectors corresponding to the first requirement feature are determined from the feature vector library; the feature vector library includes feature vectors corresponding to multiple requirement features; the multiple requirement features correspond to one of the m business execution methods. Determine the k requirement features corresponding to the k feature vectors, and the business execution methods corresponding to the k requirement features respectively.

4. The method according to claim 3, characterized in that, The method further includes: The initial model is trained using the cosine similarity loss function and the cross-entropy loss function to obtain the feature vectorization model. The cosine similarity loss function is used to improve the cosine similarity between feature vectors of different requirements corresponding to the same business execution method. The cross-entropy loss function is used to reduce the difference between the inference results output by the feature analysis model and the true distribution.

5. The method according to claim 1, characterized in that, The method further includes: Obtain p requirement texts that satisfy the second similarity condition with the first file, and the business execution methods corresponding to the p requirement texts respectively, where p is a positive integer; By using the p requirement texts and the corresponding business execution methods for each of the p requirement texts, the large language model is fine-tuned to obtain the document analysis model.

6. The method according to claim 1, characterized in that, The method further includes: Obtain p requirement texts that satisfy the second similarity condition with the first file, and the business execution methods corresponding to the p requirement texts respectively; Obtain the business logic document; the business logic document includes the normative documents of the organization to which the target business belongs; By using the p requirement texts, the business execution methods corresponding to the p requirement texts, and the business logic document, the large language model is fine-tuned to obtain the document analysis model.

7. The method according to claim 5 or 6, characterized in that, The step of obtaining p demand texts that satisfy the second similarity condition with the first file, and the business execution methods corresponding to the p demand texts respectively, includes: The feature vector corresponding to the first file is determined using a general semantic vector model; The p feature vectors with the highest similarity to the feature vectors corresponding to the first file are determined from the text vector library; the text vector library includes feature vectors corresponding to multiple requirement texts; the multiple requirement texts correspond to one of the m business execution methods. Determine the p requirement texts corresponding to the p feature vectors, and the business execution methods corresponding to the p requirement texts respectively.

8. The method according to claim 7, characterized in that, The step of using a general semantic vector model to determine the feature vector corresponding to the first file includes: The first file contains multiple text paragraphs, which are then input into a general semantic vector model to obtain feature vectors corresponding to each of the multiple text paragraphs. The step of determining the p feature vectors with the highest similarity to the feature vectors corresponding to the first file from the text vector library includes: A first similarity set is determined based on the feature vectors corresponding to the multiple text segments and a text vector library; the first similarity set includes the similarity between the feature vectors corresponding to the multiple text segments and each feature vector in the text vector library; Based on the first similarity set, p feature vectors are determined; the p feature vectors are the p feature vectors in the text vector library corresponding to the highest q similarities in the first similarity set, where q≥p.

9. The method according to any one of claims 3 or 4, characterized in that, The first requirement feature for obtaining the target service includes: Obtain the features to be processed; the features to be processed include the original feature data of the n dimensions corresponding to the target business; Eliminate the dimensional differences of the original feature data of the n dimensions to obtain intermediate feature data of the n dimensions; The intermediate state feature data of the n dimensions are converted into numerical data to obtain the first requirement feature.

10. The method according to any one of claims 1-6, characterized in that, Where m = 2; The m types of business execution methods include: assigning the business to a preset object set and not assigning the business to a preset object set.

11. An information processing device, characterized in that, include: The acquisition unit is used to acquire the first file of the target service and acquire the first requirement characteristics of the target service; The first file includes text describing the target business; the first requirement feature includes feature data of n dimensions corresponding to the target business, where n is a positive integer; The inference unit is used to input the first file into the document analysis model and obtain the first inference result output by the document analysis model; the first inference result includes the degree of correlation between each of the m business execution methods and the target business, where m is a positive integer; The reasoning unit is further configured to input the first demand feature into the feature analysis model to obtain a second reasoning result output by the feature analysis model; the second reasoning result includes the degree of correlation between each of the m business execution methods and the target business; The determining unit is configured to determine the target service execution method corresponding to the target service based on the first reasoning result and the second reasoning result; the target service execution method is one of the m service execution methods.

12. An electronic device, characterized in that, include: A memory and a processor, the memory being used to store a computer program and the processor being used to cause the electronic device to implement the information processing method according to any one of claims 1-10 when executing the computer program.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a computing device, causes the computing device to implement the information processing method according to any one of claims 1-10.

14. A computer program product, characterized in that, When the computer program product is run on a computer, the computer enables the computer to implement the information processing method according to any one of claims 1-10.

Images