Function point duplicate checking method, device and equipment based on large language model application

Abstract

The present invention relates to a method, device and equipment for checking duplicate function points based on the application of a large language model. The method uses the overlap and similarity of two function points in the operation dimension and the object attribute dimension in the system or software development as the standard to judge whether the two function points are repeated. During the process, the reflection mechanism is used to evaluate the current results and a more complete duplicate checking result is obtained through in-depth analysis. The dual dimensions of operation and object attribute are combined to judge whether two function points are repeated, which solves the problem that traditional technology cannot effectively judge whether two function points are repeated. Because the Reflexion framework is referenced and an agent based on the ReAct mechanism is created, it can reflect on itself while generating judgments, analyze and evaluate its own output in real time, and ensure that the accuracy and adaptability of the judgment are effectively enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of software data processing, and relates to a function point duplication checking method, device and equipment based on large language model application. Background Art

[0002] With the vigorous development of the global software industry, the scale of the global software market has ushered in unprecedented growth, which is not only due to the acceleration of digital transformation of enterprises, but also due to the widespread application of technologies such as cloud computing, artificial intelligence and big data. In this context, the management and optimization of function points are increasingly becoming a key technical issue in the software development process. Function points are not only an important technical indicator for measuring software functions and complexity, but also help teams accurately assess project scale, workload and resource requirements.

[0003] However, in the actual development process, repeated definition and implementation of function points often lead to waste of computing resources, time delays and confusion in project management. Especially in large projects, repeated function development is frequent, which further increases the difficulty and time cost of project maintenance. According to research, in software development, due to the redundancy of function points, project costs increase by an average of 20% to 30%. Traditional function point duplication checking methods mainly rely on manual review and basic text comparison. These methods are not only time-consuming and labor-intensive, but also difficult to handle massive documents and complex functional descriptions. Due to the lack of effective technical tools, teams often cannot identify and eliminate redundant functions in a timely manner when the project scales up, thus affecting the overall quality and progress of the project. As the complexity of the project increases, communication between team members becomes more difficult, which in turn increases the risk of repeated function development.

[0004] In recent years, with the continuous advancement of artificial intelligence and big data technologies, natural language processing (NLP) and machine learning algorithms have gradually been applied to the technical field of function point duplicate checking. The application of these technologies can effectively improve the recognition efficiency of function points, especially the application of large language models (LLM) in semantic understanding and context analysis, which can greatly improve the accuracy of function point duplicate checking. However, these traditional technologies still have problems of insufficient generalization ability and low accuracy, especially when there are differences in function descriptions between different teams and project stages, traditional technologies are often unable to adapt. In addition, the demand for function point duplicate checking is not limited to the field of software development, but is also widely used in multiple application scenarios such as demand management, product design, innovation, and business process optimization. These fields also emphasize the accurate identification and management of function points to improve resource utilization, avoid duplicate development, and enhance the overall efficiency of the project. Summary of the invention

[0005] In response to the problems existing in the above-mentioned traditional technologies, the present invention proposes a function point duplication checking method based on the application of a large language model, a function point duplication checking device based on the application of a large language model, and a computer device, which can effectively improve the accuracy and efficiency of function point duplication checking.

[0006] In order to achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0007] On the one hand, a method for checking duplicate function points based on a large language model application is provided, comprising the steps of:

[0008] Call the large language model Qwen to receive two functional points to be checked for duplication in the system or software development;

[0009] The large language model Qwen is used to extract the operation set and object attribute set corresponding to the two function points respectively; the operation is "the action performed on the object described in the function point", and the object attribute is "the set of objects operated described in the function point";

[0010] The created ReAct-based agent is used to analyze the qualitative matching relationship between the operation set and the object attribute set corresponding to the two function points according to the set analysis rules, and the judgment result and corresponding reasons of whether the two function points are repeated are obtained;

[0011] After classifying the two function points using the fine-tuned text classification model, a reference result is obtained as to whether the two function points are repeated;

[0012] Compare the judgment results with the reference results using an agent based on the ReAct mechanism;

[0013] When it is determined that the judgment result is inconsistent with the reference result, the agent based on the ReAct mechanism is guided to enter the reflection stage and reflect in an iterative manner until the iteration end condition is reached to end the iteration process and obtain the duplicate checking result; wherein, the iteration end condition includes that the judgment result is consistent with the reference result, or the number of iterations reaches the set maximum number of iterations, and the duplicate checking result is the judgment result and corresponding reason of whether two functional points are repeated when the judgment result is consistent with the reference result or after a round of iteration corresponding to the maximum number of iterations.

[0014] On the other hand, a function point duplication checking device based on a large language model application is also provided, comprising:

[0015] Function receiving module, used to call the large language model Qwen receiving system or two function points to be checked for duplication in software development;

[0016] The set extraction module is used to use the large language model Qwen to extract the operation set and object attribute set corresponding to the two function points from the two function points respectively; the operation is "the action performed on the object described in the function point", and the object attribute is "the set of objects to be operated described in the function point";

[0017] The qualitative analysis module is used to use the created ReAct-based agent to analyze the qualitative matching relationship between the operation set and the object attribute set corresponding to the two function points according to the set analysis rules, and obtain the judgment result and corresponding reasons of whether the two function points are repeated;

[0018] A reference acquisition module is used to classify two function points using a fine-tuned text classification model to obtain a reference result of whether the two function points are repeated;

[0019] A result comparison module, used to compare the judgment result with the reference result using a proxy based on the ReAct mechanism;

[0020] The reflection and iteration module is used to guide the agent based on the ReAct mechanism to enter the reflection stage and reflect in an iterative manner when it is determined that the judgment result is inconsistent with the reference result, and to terminate the iteration process and obtain the duplicate checking result when the iteration end condition is reached; wherein, the iteration end condition includes that the judgment result is consistent with the reference result, or the number of iterations reaches the set maximum number of iterations, and the duplicate checking result is the judgment result and corresponding reason of whether two functional points are repeated after a round of iteration corresponding to the maximum number of iterations when the judgment result is consistent with the reference result or.

[0021] On the other hand, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned function point duplication checking method based on the large language model application are implemented.

[0022] One of the above technical solutions has the following advantages and beneficial effects:

[0023] The above-mentioned method, device and equipment for checking duplicate function points based on the application of a large language model, by using the overlap and similarity of two function points in the operation dimension and the object attribute dimension in the system or software development as the standard, to judge whether the two function points are repeated, during which the reflection mechanism is used to evaluate the current results and obtain a more complete duplicate checking result through in-depth analysis. Compared with traditional technology, the above-mentioned scheme can combine the two dimensions of operation and object attribute to judge whether two function points are repeated, solving the problem that traditional technology cannot effectively judge whether two function points are repeated. Unlike traditional technology, the above-mentioned scheme does not simply check duplicate judgment based on word repetition rate or sentence similarity, but analyzes the qualitative matching relationship between the actions performed by the two function points on the object and the set of objects being operated to judge whether they represent the same function. Operation and object attributes are the inherent properties of a function point, and are also a high-level summary of a function, so these two dimensions can be used to make effective judgments.

[0024] In addition, by referring to the Reflexion framework and creating an agent based on the ReAct mechanism, the accuracy and adaptability of judgments are effectively enhanced. Through the ReAct mechanism, the agent is able to self-reflect while generating judgments, analyzing and evaluating its own output in real time. This dynamic feedback mechanism enables the agent to quickly identify potential errors and adjust its judgment strategy, thereby continuously optimizing the duplicate checking results. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the conventional technology, the drawings required for use in the embodiments or the conventional technology descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0026] Figure 1 A flowchart of a method for checking duplicate function points based on a large language model application in one embodiment;

[0027] Figure 2 A schematic diagram of the application process of a method for checking duplicate function points based on a large language model in one embodiment;

[0028] Figure 3 A schematic diagram of a flow chart of analyzing a qualitative matching relationship between an operation set and an object attribute set corresponding to two function points in one embodiment;

[0029] Figure 4 Schematic diagram of the module structure of a function point duplication checking device based on a large language model application in one embodiment. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and Examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by those skilled in the art of the present invention. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0031] It should be noted that the reference to "embodiment" in this article means that the specific features, structures or characteristics described in conjunction with the embodiment may be included in at least one embodiment of the present invention. The presentation of this phrase at various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It will be appreciated by those skilled in the art that the embodiments described herein may be combined with other embodiments. The term "and / or" used in the specification of the present invention and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0032] The following will describe the implementation of the present invention in detail with reference to the accompanying drawings in the embodiment diagram of the present invention.

[0033] In one embodiment, Figure 1 As shown, a function point duplication checking method based on a large language model application is provided, which may include the following processing steps S10 to S20:

[0034] S10, calling the large language model Qwen to receive two functional points to be checked for duplication in the system or software development;

[0035] S12, using the large language model Qwen, extracting the operation set and object attribute set corresponding to the two function points from the two function points respectively; the operation is "the action performed on the object described in the function point", and the object attribute is "the set of objects to be operated described in the function point";

[0036] S14, using the created agent based on the ReAct mechanism to analyze the qualitative matching relationship between the operation set and the object attribute set corresponding to the two function points according to the set analysis rules, and obtaining the judgment result and corresponding reasons of whether the two function points are repeated;

[0037] S16, after classifying the two function points using the fine-tuned text classification model, a reference result of whether the two function points are repeated is obtained;

[0038] S18, comparing the judgment result with the reference result using an agent based on the ReAct mechanism;

[0039] S20, when it is determined that the judgment result is inconsistent with the reference result, guide the agent based on the ReAct mechanism to enter the reflection stage and reflect in an iterative manner until the iteration end condition is reached to end the iteration process and obtain the duplicate checking result; wherein, the iteration end condition includes that the judgment result is consistent with the reference result, or the number of iterations reaches the set maximum number of iterations, and the duplicate checking result is the judgment result and corresponding reason of whether the two functional points are repeated after a round of iteration corresponding to the maximum number of iterations when the judgment result is consistent with the reference result or.

[0040] It can be understood that an agent based on the ReAct mechanism was created with reference to the existing Reflexion framework (a framework for strengthening large language model (LLM) agents, which guides the agent's learning process through natural language feedback rather than weight updates). The ReAct mechanism is an existing new method that solves a variety of language reasoning and decision-making tasks by combining reasoning and action in the language model. The typical process of the ReAct mechanism includes two parts: thought and action. Thought is generated by the language model and is the basis for action. The rationality of the behavior can be measured based on the thinking of the language model; the decision-making process of the ReAct mechanism is more explainable and credible, making the decision of the language model more reasonable. ‌

[0041] The agent based on the ReAct mechanism analyzes the qualitative matching relationship between the operations and object attributes of two function points in the system or software development, and preliminarily determines whether there is duplication. The model Qwen (a series of open source large language models) and the vector model BGE-M3 (the first semantic vector model that integrates the three technical features of multi-linguality, multi-granularity and multi-functionality) provide tool support. At the same time, the fine-tuned BERT (a pre-trained language representation model) model (that is, the fine-tuned text classification model in this manual) is used to complete the text classification work to give a reference result. If the result of the preliminary judgment is inconsistent with the reference result, the large language model is guided to reflect until the result is consistent or the set upper limit of the number of iterations is reached. This method effectively improves the accuracy and efficiency of function point duplication checking and provides an intelligent solution for enterprise software function management.

[0042] Specifically, in this embodiment, the overlap and similarity of two function points in system or software development in two target dimensions are used as standards to preliminarily determine whether the two function points are repeated. The two target dimensions are the operation dimension and the object attribute dimension. A set of operations (which may be referred to as an operation set) is extracted from the two function points, where the operation can be defined as "the action performed on the object described in the function point". These actions may include, but are not limited to, traditional actions such as adding, deleting, modifying and checking, reflecting the specific implementation and business logic of the function point. Next, a set of object attributes (which may be referred to as an object attribute set) is also extracted from the two function points, where the object attributes can be defined as "the set of objects to be operated described in the function point". The object attributes are the target of the operation, reflecting the application scope and affected objects of the function point.

[0043] Then, the operation set and object attribute set extracted from the two function points are analyzed for qualitative matching relationships, and the judgment results and corresponding reasons {res:reason} of whether the two function points are repeated are obtained. This analysis relies on a series of set analysis rules, which are designed to ensure that the two function points have an inclusion relationship or a highly similar relationship in the operation dimension, and at the same time ensure that the two have a high degree of semantic consistency in the object attribute dimension. Finally, the reference results of the fine-tuned text classification model on whether the two input function points are repeated are obtained, and the judgment results obtained above are compared with the reference results, and the consistency between the two is used as the basis for entering the reflection stage. Reflection is carried out in an iterative manner. In the process of no more than N iterations, the large language model will take into account the guiding role of the reference results and be guided to generate new judgment results and corresponding reasons until the new judgment results are consistent with the reference results or exceed the number of reflections (that is, the maximum number of iterations). The iterative process ends and the duplicate checking results are obtained; N is the maximum number of iterations set, also known as the maximum number of reflections, which is a positive integer.

[0044] Therefore, this embodiment fine-tunes the text classification model and combines the reflection mechanism of the Reflexion framework to solve the deficiencies in the above-mentioned traditional methods, discover and solve the problem of functional duplication that enterprises may encounter in system or software development, and improve the accuracy and efficiency of function point duplication checking. By comparing the fine-tuned text classification model with the judgment results of the mainstream large language model, a new solution is provided for the intelligent identification and management of function points, which can not only promote the progress and innovation of the software development industry, but also bring more efficient technical tools and means to various fields of function management.

[0045] The above-mentioned method for checking duplicate function points based on the application of a large language model uses the overlap and similarity of two function points in the operation dimension and the object attribute dimension in the system or software development as the standard to judge whether the two function points are repeated. During this period, the reflection mechanism is used to evaluate the current results and a more complete duplicate checking result is obtained through in-depth analysis. Compared with traditional technologies, the above-mentioned scheme can combine the dual dimensions of operation and object attribute to judge whether two function points are repeated, solving the problem that traditional technologies cannot effectively judge whether two function points are repeated. Unlike traditional technologies, the above-mentioned scheme does not simply check duplicate judgments based on word repetition rate or sentence similarity, but analyzes the qualitative matching relationship between the actions (operations) performed by the two function points on the object and the set of objects (object attributes) operated to judge whether they represent the same function. Operations and object attributes are the inherent properties of a function point and are also a high-level summary of a function. Therefore, these two dimensions can be used to make effective judgments.

[0046] In addition, by referring to the Reflexion framework and creating an agent based on the ReAct mechanism, the accuracy and adaptability of judgments are effectively enhanced. Through the ReAct mechanism, the agent is able to self-reflect while generating judgments, analyzing and evaluating its own output in real time. This dynamic feedback mechanism enables the agent to quickly identify potential errors and adjust its judgment strategy, thereby continuously optimizing the duplicate checking results.

[0047] In one embodiment, further, Figure 2 As shown, in this embodiment, the large language model Qwen (such as the large language model Qwen2) can be called to receive the function point F (such as function point 1 or function point 2), and the operation set {O1, O2, O3, ...} can be extracted from F and returned. Regarding the above step S12: in the process of using the large language model Qwen to extract the operation sets corresponding to the two function points from the two function points respectively, the operation extraction rules that the large language model Qwen needs to follow when extracting the operation include:

[0048] (1) The operation in a function refers to the action performed on the object described in the function point.

[0049] (2) The object attributes in a function refer to the set of objects being operated as described in the function point.

[0050] (3) The subject object in the function refers to the object being operated described in the function point.

[0051] (4) The details of the action cannot be modified, explained or analyzed.

[0052] (5) The separated actions are output in the form of 1.\n2.\n... and start with the sequence number. Only the actions are output.

[0053] For example, the operation set extracted from the function point "Enable old classmates to register, sign in, query and locate offline class information through mobile terminals. Offline class information includes course category, course name, course location, lecturer name, course length, total course length, class sections and total class sections, etc." is {register, sign in, query, locate}.

[0054] In one embodiment, the large language model Qwen2 can be called to receive function point F (such as function point 1 or function point 2), and the object attribute set {P1, P2, P3, ...} can be extracted from F and returned. In the process of using the large language model Qwen to extract the operation sets corresponding to the two function points from the two function points respectively, the object attribute extraction rules that the large language model Qwen needs to follow when extracting object attributes include:

[0055] (1) The operation in a function refers to the action performed on the object described in the function point.

[0056] (2) The object attributes in a function refer to the set of objects being operated as described in the function point.

[0057] (3) The subject object in the function refers to the object being operated described in the function point.

[0058] (4) The details of the object’s attributes cannot be modified, interpreted, or parsed.

[0059] (5) The separated object attributes are output in the form of a sequence number 1.\n2.\n.... and start with the sequence number. Only nouns are output.

[0060] For example, the object attribute set extracted from the function point "Enable old classmates to register, sign in, query and locate offline classroom information through mobile terminals. Offline classroom information includes course category, course name, course location, lecturer name, course length, total course length, class sections and total class sections, etc." is {course category, course name, course location, lecturer name, course length, total course length, class sections, total class sections}.

[0061] In one embodiment, the set analysis rules include a first analysis rule and a second analysis rule. Figure 3 As shown, the process of analyzing the qualitative matching relationship between the operation set and the object attribute set corresponding to the two function points by using the created agent based on the ReAct mechanism according to the set analysis rules may specifically include the following steps:

[0062] S141, using the word embedding model BGE-M3 through the defined first tool to respectively obtain the embedding representations of the object attribute sets of the two function points and calculate the cosine similarity of the object attribute sets of the two function points;

[0063] S143, calling the large language model Qwen by the defined second tool to follow the first analysis rule to determine whether there is an inclusion relationship between the operation sets of the two function points;

[0064] S145, using the word embedding model BGE-M3 through the defined third tool to respectively obtain the embedding representations of the operation sets of the two function points and calculate the cosine similarity of the operation sets of the two function points;

[0065] S147, using the agent based on the ReAct mechanism to access the first tool, the second tool and the third tool respectively, perform qualitative relationship matching on the operations and object attributes of the two function points according to the processing logic of the second analysis rule, and decide to generate a preliminary judgment result or go to the next analysis rule.

[0066] It can be understood that when analyzing the qualitative matching relationship of the operation set and the object attribute set extracted from the two function points and obtaining the judgment result and the corresponding reason {res:reason} of whether the two function points are repeated, the following processing steps may be specifically included:

[0067] Define a tool dedicated to calculating the similarity between sets of object attributes T 1, T 1Use the word embedding model BGE-M3 to obtain the object attribute set of function point 1 {P 11 ,P 12 ,P 13 ,…} and the object attribute set of function point 2 {P 21 ,P 22 ,P 23 ,…}, and use this to calculate the cosine similarity of the two object attribute sets sim 1. If sim 1> the set similarity threshold (for example, 80% or higher), it is determined that function point 1 and function point 2 have a high similarity in the object attribute dimension, otherwise it is determined that function point 1 and function point 2 have a low similarity in the object attribute dimension.

[0068] Define a tool specifically for determining whether there is a containment relationship between operation sets T 2. In T 2 calls the large language model Qwen2 to use its natural language processing capabilities to operate on the set of functions in point 1 {O 11 ,O 12 ,O 13 ,…} and the operation set of function point 2 {O 21 ,O 22 ,O 23,…} for analysis and judgment. The first analysis rules that the large language model needs to follow when judging whether there is an inclusion relationship between operation sets are:

[0069] (1) An existence inclusion relationship means that the operation of one function point completely exists in the operation of another function point.

[0070] (2) If the two operation sets are the same, it is also judged that there is an inclusion relationship.

[0071] (3) The absence of an inclusion relationship means that the operation of one function point does not completely exist or does not exist at all in the operation of another function point.

[0072] (4) If the semantics of two operations are highly similar, they can be considered as the same operation when making judgments.

[0073] (5) The judgment should be made from two perspectives: whether function point 1 contains function point 2, and whether function point 2 contains function point 1.

[0074] (6) If only function point 1 does not contain function point 2, and function point 2 does not contain function point 1, then the output is “there is no inclusion relationship”. Otherwise, if function point 1 contains function point 2, or function point 2 contains function point 1, then the output is “there is an inclusion relationship”.

[0075] Define a tool dedicated to calculating the similarity between sets of operations T 3. T 3 Use the word embedding model BGE-M3 to get the operation set of function point 1 {O 11 ,O 12 ,O 13 ,…} and the operation set of function point 2 {O 21 ,O 22 ,O 23 ,…}, and use this to calculate the cosine similarity of the two operation sets sim 2. If sim 2>80%, it is determined that function point 1 and function point 2 have a high similarity in the operation dimension; otherwise, it is determined that function point 1 and function point 2 have a low similarity in the operation dimension.

[0076] Build an agent based on the large language model Gemma2 and grant it access to tools T 1. T 2 and T 3, so that it can call the appropriate tool as needed, perform qualitative relationship matching on the operation and object attributes according to the preset order and the processing logic of the second analysis rule, and decide whether to generate a preliminary judgment result or go to the next analysis rule. The specific second analysis rule can be as follows:

[0077] (1) The two repeated functional points are essentially the same functional point, just expressed in a different way.

[0078] (2) The operation in a function refers to the action performed on the object described in the function point.

[0079] (3) The object attributes in a function refer to the set of objects being operated as described in the function point.

[0080] (4) The subject object in the function refers to the object being operated described in the function point.

[0081] (5) Comprehensively judging whether there is duplication only from the two dimensions of the operation and object attributes of the function point, ignoring the tendency and specificity of the operation and object attributes, and ignoring the differences between the subject objects.

[0082] (6) If the object attributes of the similarity of two function points have a similarity above the set similarity threshold, then the function points are judged to have a high similarity in the object attribute dimension. Conversely, if the similarity is below the set similarity threshold, then it is directly judged as non-duplicate.

[0083] (7) If there is a containment relationship between the operations of Function 1 and Function 2 when the similarity of the object attribute dimension is high, they are judged to be duplicates.

[0084] (8) If the similarity of the object attribute dimension is low and there is no inclusion relationship between the operations of Function 1 and Function 2, they are judged as non-duplicate.

[0085] (9) When rules (7) and (8) are unable to make a judgment, if the proportion of common operations of two function points reaches a set proportion threshold (for example, 80% or higher), then they are judged to be duplicates (i.e., there are only a small proportion of additional operations); otherwise, they are judged to be non-duplicates.

[0086] In one embodiment, regarding the above step S20, when it is determined that the judgment result is inconsistent with the reference result, guiding the agent based on the ReAct mechanism to enter the reflection phase and reflect in an iterative manner, each round of iteration may specifically include the following processing steps:

[0087] Record the judgment result of whether two functional points are repeated in the current iteration and the historical execution information when the corresponding reasons are generated;

[0088] When it is determined that the judgment result is inconsistent with the reference result, it is assumed that the judgment result is an incorrect judgment, and the large language model Gemma2 is instructed to generate a new, more correct and different judgment result from the opposite angle of the judgment result, and explain the corresponding reasons for the new judgment result, and record the corresponding reasons as reflection information;

[0089] The historical execution information and reflection information are added to the historical information and then passed to the agent based on the ReAct mechanism. The agent based on the ReAct mechanism combines the updated historical information to enter the second round of iteration reflection, and obtains the judgment results and corresponding reasons of whether the two function points generated in the second round of iteration are repeated.

[0090] It can be understood that when entering the reflection phase, the agent is guided to think and give a deeper analysis. Reflection is performed in an iterative manner. In the process of no more than N iterations (in this example, N=4 can be taken as an example for illustration), the large language model will take into account the guiding role of the reference result (which can be recorded as key) and be guided to generate new judgment results and corresponding reasons.

[0091] For example, first fine-tune the BERT model chinese-macbert-base (a pre-trained language model optimized for Chinese language characteristics) on the existing 373 vertical data, so that it can classify the input function points 1 and 2, and determine whether the two function points are "repeated" or "non-repeated", and thereby obtain the reference result key. The specific implementation process of fine-tuning the BERT model chinese-macbert-base can be: first annotate the existing 373 vertical data, distinguish whether the two function points in each data are "repeated" or "non-repeated", and then obtain the fine-tuning of the pre-trained model chinese-macbert-base, and add a fully connected layer (i.e., classifier layer) after its original structure to achieve the binary classification task. Fine-tuning methods can be roughly divided into two categories, one is full parameter fine-tuning (Full Fine-Tuning), and the other is parameter-efficient fine-tuning (Parameter-Efficient Fine-Tuning). In the process of fine-tuning using labeled data, the classification performance results of the LoRA (Low-Rank Adaptation of Large Language Models) method and the Freeze method in full parameter fine-tuning and parameter-efficient fine-tuning on the data were compared, and the pre-trained model fine-tuned using the Freeze method was determined as the text classification model. Among them, when implementing the Freeze method to fine-tune the model, the embedding layer and the first 10 encoder layers of the model are frozen, and only the remaining 2 encoder layers, one pooler layer, and the parameters of the added fully connected layer are adjusted. This is because the previous layers have learned general semantic features through pre-training, while the latter layers can better capture high-level semantics and domain-specific knowledge, so freezing the parameters of the previous layers through the Freeze method and only fine-tuning the latter layers can be better applicable to the problems to be solved in this application.

[0092] In the first iteration, the judgment result and the corresponding reason {res:reason} of whether the two function points are repeated are obtained according to the steps in the previous text, and the historical execution information H(1) when {res:reason} is generated is recorded. The judgment result res is compared with the reference result key. If the two are consistent, the judgment result and the corresponding reason {res:reason} of the first iteration are the final results. If the two are inconsistent, it is assumed that the judgment result res is an incorrect judgment, and the large language model Gemma2 is instructed to generate a new, more correct and different judgment result from the judgment result res at the opposite angle of the judgment result res, and explain the corresponding reason, which is recorded as reflection information R(1). The historical execution information H(1) is added to the historical information, and the reflection information R(1) is added to the historical information as a comparison of the historical execution information H(1). The updated historical information is transmitted to the agent, and the agent combines the historical information to make a judgment and analysis on whether function points 1 and 2 are repeated again according to the above series of set analysis rules, that is, a new judgment result and the corresponding reason {res:reason} of whether the two function points are repeated are obtained, which is also a judgment-reason pair. The historical information includes pre-information (operations and object attributes extracted from function point 1 and function point 2), historical execution information, and reflection information.

[0093] Similarly, in the i-th iteration, if the newly obtained judgment result res in the i-1th round is inconsistent with the reference result key, then the historical execution information H(i) and reflection information R(i) of the i-th iteration will be added to the historical information. At this time, the historical information has accumulated all the historical execution information and reflection information from the 1st iteration to the i-th iteration. It should be noted that the entire reflection iteration process continues until the new judgment result res is consistent with the reference result key. If the judgment result res newly generated by the agent is still inconsistent with the reference result key after N rounds of iteration, that is, the agent insists on the original judgment, then the final result after N rounds of iteration is the judgment result of whether the two function points generated in the Nth iteration are repeated and the corresponding reason {res: reason}.

[0094] In some embodiments, experiments on 373 pairs of function points using the above-mentioned function point duplication checking method based on the large language model application show that the accuracy of the function point duplication checking method based on the large language model application in judging whether the function points are repeated is 91.42%, the precision rate is 84.03%, the recall rate is 93.08%, and the F1 value is 88.32%, which verifies that its duplication checking effect is very superior.

[0095] It should be understood that although Figures 1 to 3The steps in the method are shown in sequence as indicated by the arrows, but the steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of the steps, and the steps can be executed in other orders. Figures 1 to 3 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequentially, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0096] In one embodiment, Figure 4 As shown, a function point duplicate checking device 100 based on the application of a large language model is provided, which may include a function receiving module 11, a set extraction module 13, a qualitative analysis module 15, a reference acquisition module 17, a result comparison module 19 and a reflection iteration module 21. Among them, the function receiving module 11 is used to call the large language model Qwen to receive two function points to be checked for duplicate in the system or software development. The set extraction module 13 is used to use the large language model Qwen to extract the operation set and object attribute set corresponding to the two function points from the two function points respectively; the operation is "the action performed on the object described in the function point", and the object attribute is "the set of objects operated described in the function point". The qualitative analysis module 15 is used to use the created agent based on the ReAct mechanism to analyze the qualitative matching relationship of the operation set and object attribute set corresponding to the two function points according to the set analysis rules, and obtain the judgment result and corresponding reason of whether the two function points are repeated. The reference acquisition module 17 is used to classify the two function points using the fine-tuned text classification model, and obtain the reference result of whether the two function points are repeated. The result comparison module 19 is used to compare the judgment result with the reference result using an agent based on the ReAct mechanism. The reflection iteration module 21 is used to guide the agent based on the ReAct mechanism to enter the reflection stage and reflect in an iterative manner when it is determined that the judgment result is inconsistent with the reference result, until the iteration end condition is reached to end the iteration process and obtain the duplicate checking result; wherein, the iteration end condition includes that the judgment result is consistent with the reference result, or the number of iterations reaches the set maximum number of iterations, and the duplicate checking result is the judgment result and corresponding reason of whether the two functional points are repeated after a round of iteration corresponding to the maximum number of iterations when the judgment result is consistent with the reference result or.

[0097] The above-mentioned function point duplicate checking device 100 based on the application of a large language model, by using the overlap and similarity of two function points in the operation dimension and the object attribute dimension in the system or software development as the standard, to judge whether the two function points are repeated, during which the current results are evaluated through the reflection mechanism and a more complete duplicate checking result is obtained through in-depth analysis. Compared with the traditional technology, the above-mentioned scheme can combine the two dimensions of operation and object attribute to judge whether the two function points are repeated, solving the problem that the traditional technology cannot effectively judge whether the two function points are repeated. Unlike the traditional technology, the above-mentioned scheme does not simply check and judge based on the word repetition rate or sentence similarity, but analyzes the qualitative matching relationship between the actions performed by the two function points on the object and the set of objects operated to judge whether they represent the same function. Operation and object attributes are the inherent properties of a function point, and are also a high-level summary of a function, so these two dimensions can be used to make effective judgments.

[0098] In addition, by referring to the Reflexion framework and creating an agent based on the ReAct mechanism, the accuracy and adaptability of judgments are effectively enhanced. Through the ReAct mechanism, the agent is able to self-reflect while generating judgments, analyzing and evaluating its own output in real time. This dynamic feedback mechanism enables the agent to quickly identify potential errors and adjust its judgment strategy, thereby continuously optimizing the duplicate checking results.

[0099] In one embodiment, the set analysis rules include a first analysis rule and a second analysis rule, and the qualitative analysis module 15 can be specifically used to: obtain the embedding representation of the object attribute sets of the two function points respectively by using the word embedding model BGE-M3 through the defined first tool and calculate the cosine similarity of the object attribute sets of the two function points; call the large language model Qwen through the defined second tool to follow the first analysis rule to determine whether there is an inclusion relationship between the operation sets of the two function points; obtain the embedding representation of the operation sets of the two function points respectively by using the word embedding model BGE-M3 through the defined third tool and calculate the cosine similarity of the operation sets of the two function points; use the agent based on the ReAct mechanism to access the first tool, the second tool and the third tool respectively, perform qualitative relationship matching on the operations and object attributes of the two function points according to the processing logic of the second analysis rule, and decide to generate a preliminary judgment result or go to the next analysis rule; wherein the first analysis rule includes:

[0100] The existence of an inclusion relationship means that the operation of one function point completely exists in the operation of another function point;

[0101] If the two operation sets are the same, it is also judged that there is an inclusion relationship;

[0102] The absence of an inclusion relationship means that the operation of one function point does not completely exist or does not exist at all in the operation of another function point;

[0103] If two operations have highly similar semantics, they can be considered as the same operation when making judgments;

[0104] It is necessary to judge from two perspectives: whether function point 1 contains function point 2, and whether function point 2 contains function point 1;

[0105] Only when function point 1 does not contain function point 2, and function point 2 does not contain function point 1, the output is "there is no inclusion relationship"; otherwise, if function point 1 contains function point 2, or function point 2 contains function point 1, either one of them is true, then the output is "there is an inclusion relationship".

[0106] The second analysis rule includes:

[0107] The two repeated function points are essentially the same function point, just expressed in a different way;

[0108] The operation in a function refers to the action performed on the object described in the function point;

[0109] The object attributes in a function refer to the set of objects being operated as described in the function point;

[0110] The subject object in the function refers to the object being operated described in the function point;

[0111] Only the operation and object attributes of the function points are used to comprehensively judge whether there is duplication, ignoring the tendency and pertinence of the operation and object attributes, and ignoring the differences between the subject objects;

[0112] If the object attributes of the similarity of two function points have a similarity above the set similarity threshold, the function points are judged to have a high similarity in the object attribute dimension; conversely, if it is lower than the set similarity threshold, it is directly judged as not repeated;

[0113] If the similarity of the object attribute dimension is high, and there is an inclusion relationship between the operations of function 1 and function 2, they are judged as duplicates;

[0114] If the similarity of the object attribute dimension is low, and there is no inclusion relationship between the operations of function 1 and function 2, then they are judged as non-repetitive;

[0115] When a judgment cannot be made due to high or low similarity in the object attribute dimension, if the common operation ratio of the two function points exceeds the set ratio threshold, then they are judged to be duplicates, otherwise they are judged to be non-duplicates.

[0116] In one embodiment, when the reflection iteration module 21 determines that the judgment result is inconsistent with the reference result, it guides the agent based on the ReAct mechanism to enter the reflection stage and iteratively reflect in each round of iteration. It can be used to: record the judgment result of whether two functional points are repeated in the current round of iteration and the historical execution information when the corresponding reason is generated; when it is determined that the judgment result is inconsistent with the reference result, it is assumed that the judgment result is an incorrect judgment, and the large language model Gemma2 is instructed to generate a new, more correct and different judgment result from the opposite angle of the judgment result, and explain the corresponding reason for the new judgment result, and record the corresponding reason as reflection information; after adding the historical execution information and the reflection information to the historical information, it is transmitted to the agent based on the ReAct mechanism, and the agent based on the ReAct mechanism combines the updated historical information to enter the second round of reflection, and obtains the judgment result and corresponding reason of whether the two functional points generated in the second round of iteration are repeated.

[0117] In one embodiment, the operation extraction rules that the large language model Qwen needs to follow when extracting operations include:

[0118] The operation in a function refers to the action performed on the object described in the function point;

[0119] The object attributes in a function refer to the set of objects being operated as described in the function point;

[0120] The subject object in the function refers to the object being operated described in the function point;

[0121] The details of the action cannot be modified, explained or analyzed;

[0122] The separated actions are output in the form of serial numbers and start with the serial numbers. Only the actions are output.

[0123] It can be understood that the specific limitations of the function point duplication checking device 100 based on the large language model application can refer to the corresponding limitations of the function point duplication checking method based on the large language model application mentioned above, and will not be repeated here.

[0124] Each module in the function point duplicate checking device 100 based on the large language model application can be fully or partially implemented by software, hardware and their combination. The above modules can be embedded in or independent of a device with data processing functions in the form of hardware, or can be stored in the memory of the aforementioned device in the form of software, so that the processor can call and execute the operations corresponding to the above modules. The aforementioned device can be, but is not limited to, various types of computing terminals already available in the art.

[0125] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the following processing steps when executing the computer program: calling a large language model Qwen to receive two function points to be checked for duplicates in a system or software development; using the large language model Qwen to extract operation sets and object attribute sets corresponding to the two function points respectively from the two function points; the operation is "the action performed on the object described in the function point", and the object attribute is "the set of objects operated described in the function point"; using the created agent based on the ReAct mechanism to analyze the qualitative matching relationship of the operation set and the object attribute set corresponding to the two function points according to the set analysis rules, and obtain the two function points. The judgment result and corresponding reason of whether the two function points are repeated; after classifying the two function points by using the fine-tuned text classification model, a reference result of whether the two function points are repeated is obtained; the judgment result and the reference result are compared by using an agent based on the ReAct mechanism; when it is determined that the judgment result is inconsistent with the reference result, the agent based on the ReAct mechanism is guided to enter the reflection stage and reflect in an iterative manner until the iteration end condition is met and the iteration process is terminated and the duplicate checking result is obtained; wherein, the iteration end condition includes that the judgment result is consistent with the reference result, or the number of iterations reaches the set maximum number of iterations, and the duplicate checking result is the judgment result and the corresponding reason of whether the two function points are repeated when the judgment result is consistent with the reference result or after one round of iteration corresponding to the maximum number of iterations.

[0126] In one embodiment, when the processor executes the computer program, it can also implement the steps or sub-steps added in each embodiment of the above-mentioned function point duplication checking method based on the large language model application.

[0127] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following processing steps are implemented: calling a large language model Qwen to receive two function points to be checked for duplicates in a system or software development; using the large language model Qwen to extract operation sets and object attribute sets corresponding to the two function points respectively from the two function points; the operation is "the action performed on the object described in the function point", and the object attribute is "the set of objects operated described in the function point"; using the created agent based on the ReAct mechanism to analyze the qualitative matching relationship of the operation set and the object attribute set corresponding to the two function points according to the set analysis rules, and obtain whether the two function points are duplicates. The judgment result and corresponding reason of duplication are obtained; after classifying the two function points by using the fine-tuned text classification model, a reference result of whether the two function points are repeated is obtained; the judgment result and the reference result are compared by using the agent based on the ReAct mechanism; when it is determined that the judgment result is inconsistent with the reference result, the agent based on the ReAct mechanism is guided to enter the reflection stage and reflect in an iterative manner until the iteration end condition is met and the iteration process is terminated and the duplication check result is obtained; wherein, the iteration end condition includes that the judgment result is consistent with the reference result, or the number of iterations reaches the set maximum number of iterations, and the duplication check result is the judgment result and the corresponding reason of whether the two function points are repeated when the judgment result is consistent with the reference result or after a round of iteration corresponding to the maximum number of iterations.

[0128] In one embodiment, when the computer program is executed by a processor, it can also implement the steps or sub-steps added in each embodiment of the above-mentioned function point duplication checking method based on the large language model application.

[0129] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus dynamic random access memory (RambusDRAM, referred to as RDRAM) and interface dynamic random access memory (DRDRAM).

[0130] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0131] The above embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of protection of the invention. It should be pointed out that, for those of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the attached claims.

Claims

1. A method for checking duplicate function points based on a large language model application, characterized in that: Includes steps: Call the large language model Qwen to receive two functional points to be checked for duplication in the system or software development; The large language model Qwen is used to extract the operation set and object attribute set corresponding to the two function points from the two function points respectively; the operation is "the action performed on the object described in the function point", and the object attribute is "the set of objects to be operated described in the function point"; Using the created agent based on the ReAct mechanism to analyze the qualitative matching relationship of the operation set and the object attribute set corresponding to the two function points according to the set analysis rules, the judgment result and corresponding reasons of whether the two function points are repeated are obtained; After classifying the two function points using the fine-tuned text classification model, a reference result of whether the two function points are repeated is obtained; Comparing the judgment result with the reference result using an agent based on the ReAct mechanism; When it is determined that the judgment result is inconsistent with the reference result, the agent based on the ReAct mechanism is guided to enter the reflection stage and reflect in an iterative manner until the iteration end condition is reached, the iteration process is terminated and the duplicate checking result is obtained; wherein, the iteration end condition includes that the judgment result is consistent with the reference result, or the number of iterations reaches the set maximum number of iterations, and the duplicate checking result is the judgment result and corresponding reason of whether the two functional points are repeated when the judgment result is consistent with the reference result or after a round of iteration corresponding to the maximum number of iterations.

2. The method for checking duplicate function points based on large language model application according to claim 1 is characterized in that: In the process of respectively extracting the operation sets corresponding to the two function points from the two function points using the large language model Qwen, the operation extraction rules that the large language model Qwen needs to follow when extracting operations include: The operation in a function refers to the action performed on the object described in the function point; The object attributes in a function refer to the set of objects being operated as described in the function point; The subject object in the function refers to the object being operated described in the function point; The details of the action cannot be modified, explained or analyzed; The separated actions are output in the form of serial numbers and start with the serial numbers. Only the actions are output.

3. The method for checking duplicate function points based on large language model application according to claim 1 or 2, characterized in that: In the process of respectively extracting the operation sets corresponding to the two function points from the two function points using the large language model Qwen, the object attribute extraction rules that the large language model Qwen needs to follow when extracting object attributes include: The operation in a function refers to the action performed on the object described in the function point; The object attributes in a function refer to the set of objects being operated as described in the function point; The subject object in the function refers to the object being operated described in the function point; The details of object attributes cannot be modified, interpreted, or parsed; The separated object attributes are output in serial number format and start with the serial number, and only nouns are output.

4. The method for checking duplicate function points based on large language model application according to claim 1, characterized in that: The set analysis rules include a first analysis rule and a second analysis rule. The process of analyzing the qualitative matching relationship between the operation set and the object attribute set corresponding to the two function points by using the created agent based on the ReAct mechanism according to the set analysis rules includes: Using the word embedding model BGE-M3 through the defined first tool, respectively obtain the embedding representation of the object attribute sets of the two function points and calculate the cosine similarity of the object attribute sets of the two function points; Calling the large language model Qwen by a defined second tool to follow the first analysis rule to respectively determine whether there is an inclusion relationship between the operation sets of the two function points; Using the word embedding model BGE-M3 through the defined third tool, respectively obtain the embedding representation of the operation sets of the two function points and calculate the cosine similarity of the operation sets of the two function points; The first tool, the second tool and the third tool are accessed respectively by using a proxy based on the ReAct mechanism, and the operations and object attributes of the two function points are qualitatively matched according to the processing logic of the second analysis rule, and it is decided to generate a preliminary judgment result or go to the next analysis rule; wherein, The first analysis rule includes: The existence of an inclusion relationship means that the operation of one function point completely exists in the operation of another function point; If the two operation sets are the same, it is also judged that there is an inclusion relationship; The absence of an inclusion relationship means that the operation of one function point does not completely exist or does not exist at all in the operation of another function point; If two operations have highly similar semantics, they can be considered as the same operation when making judgments; It is necessary to judge from two perspectives: whether function point 1 contains function point 2, and whether function point 2 contains function point 1; Only when function point 1 does not contain function point 2, and function point 2 does not contain function point 1, the output is "no inclusion relationship exists"; otherwise, if function point 1 contains function point 2, or function point 2 contains function point 1, the output is "there is an inclusion relationship"; The second analysis rule includes: The two repeated function points are essentially the same function point, just expressed in a different way; The operation in a function refers to the action performed on the object described in the function point; The object attributes in a function refer to the set of objects being operated as described in the function point; The subject object in the function refers to the object being operated described in the function point; Only the operation and object attributes of the function points are used to comprehensively judge whether there is duplication, ignoring the tendency and pertinence of the operation and object attributes, and ignoring the differences between the subject objects; If the object attributes of the similarity of two function points have a similarity above the set similarity threshold, the function points are judged to have a high similarity in the object attribute dimension; conversely, if it is lower than the set similarity threshold, it is directly judged as not repeated; If the similarity of the object attribute dimension is high, and there is an inclusion relationship between the operations of function 1 and function 2, they are judged as duplicates; If the similarity of the object attribute dimension is low, and there is no inclusion relationship between the operations of function 1 and function 2, then they are judged as non-repetitive; When a judgment cannot be made due to high or low similarity in the object attribute dimension, if the common operation ratio of the two function points exceeds the set ratio threshold, then they are judged to be duplicates, otherwise they are judged to be non-duplicates.

5. The method for checking duplicate function points based on large language model application according to claim 4 is characterized in that: When it is determined that the judgment result is inconsistent with the reference result, the agent based on the ReAct mechanism is guided to enter the reflection phase and reflect in an iterative manner. Each round of iteration includes the following steps: Record the judgment result of whether the two function points are repeated in the current iteration and the historical execution information when the corresponding reasons are generated; When it is determined that the judgment result is inconsistent with the reference result, it is assumed that the judgment result is an erroneous judgment, and the large language model Gemma2 is instructed to generate a new, more correct and different judgment result from the opposite angle of the judgment result, and explain the corresponding reason of the new judgment result, and record the corresponding reason as reflection information; The historical execution information and the reflection information are added to the historical information and then transmitted to the agent based on the ReAct mechanism. The agent based on the ReAct mechanism combines the updated historical information to enter the second round of iterative reflection, and obtains the judgment result and corresponding reason of whether the two function points generated in the second round of iteration are repeated.

6. A function point duplication checking device based on a large language model application, characterized in that: include: Function receiving module, used to call the large language model Qwen receiving system or two function points to be checked for duplication in software development; A set extraction module, used to extract the operation set and object attribute set corresponding to the two function points from the two function points respectively by using the large language model Qwen; the operation is "the action performed on the object described in the function point", and the object attribute is "the set of objects operated described in the function point"; A qualitative analysis module, used to analyze the qualitative matching relationship between the operation set and the object attribute set corresponding to the two function points by using the created agent based on the ReAct mechanism according to the set analysis rules, and obtain the judgment result and corresponding reasons of whether the two function points are repeated; A reference acquisition module, used to classify the two function points using a fine-tuned text classification model to obtain a reference result of whether the two function points are repeated; A result comparison module, used for comparing the judgment result with the reference result by using a proxy based on the ReAct mechanism; A reflection and iteration module is used to guide the agent based on the ReAct mechanism to enter the reflection stage and reflect in an iterative manner when it is determined that the judgment result is inconsistent with the reference result, until the iteration end condition is reached, the iteration process is terminated and the duplicate checking result is obtained; wherein, the iteration end condition includes that the judgment result is consistent with the reference result, or the number of iterations reaches the set maximum number of iterations, and the duplicate checking result is the judgment result and corresponding reason of whether the two functional points are repeated when the judgment result is consistent with the reference result or after a round of iteration corresponding to the maximum number of iterations.

7. The function point duplicate checking device based on large language model application according to claim 6 is characterized in that: The set analysis rules include a first analysis rule and a second analysis rule, and the qualitative analysis module is used to: Using the word embedding model BGE-M3 through the defined first tool, respectively obtain the embedding representation of the object attribute sets of the two function points and calculate the cosine similarity of the object attribute sets of the two function points; Calling the large language model Qwen by a defined second tool to follow the first analysis rule to respectively determine whether there is an inclusion relationship between the operation sets of the two function points; Using the word embedding model BGE-M3 through the defined third tool, respectively obtain the embedding representation of the operation sets of the two function points and calculate the cosine similarity of the operation sets of the two function points; The first tool, the second tool and the third tool are accessed respectively by using a proxy based on the ReAct mechanism, and the operations and object attributes of the two function points are qualitatively matched according to the processing logic of the second analysis rule, and it is decided to generate a preliminary judgment result or go to the next analysis rule; wherein, The first analysis rule includes: The existence of an inclusion relationship means that the operation of one function point completely exists in the operation of another function point; If the two operation sets are the same, it is also judged that there is an inclusion relationship; The absence of an inclusion relationship means that the operation of one function point does not completely exist or does not exist at all in the operation of another function point; If two operations have highly similar semantics, they can be considered as the same operation when making judgments; It is necessary to judge from two perspectives: whether function point 1 contains function point 2, and whether function point 2 contains function point 1; Only when function point 1 does not contain function point 2, and function point 2 does not contain function point 1, the output is "no inclusion relationship exists"; otherwise, if function point 1 contains function point 2, or function point 2 contains function point 1, the output is "there is an inclusion relationship"; The second analysis rule includes: The two repeated function points are essentially the same function point, just expressed in a different way; The operation in a function refers to the action performed on the object described in the function point; The object attributes in a function refer to the set of objects being operated as described in the function point; The subject object in the function refers to the object being operated described in the function point; Only the operation and object attributes of the function points are used to comprehensively judge whether there is duplication, ignoring the tendency and pertinence of the operation and object attributes, and ignoring the differences between the subject objects; If the object attributes of the similarity of two function points have a similarity above the set similarity threshold, the function points are judged to have a high similarity in the object attribute dimension; conversely, if it is lower than the set similarity threshold, it is directly judged as not repeated; If the similarity of the object attribute dimension is high, and there is an inclusion relationship between the operations of function 1 and function 2, they are judged as duplicates; If the similarity of the object attribute dimension is low, and there is no inclusion relationship between the operations of function 1 and function 2, then they are judged as non-repetitive; When a judgment cannot be made due to high or low similarity in the object attribute dimension, if the common operation ratio of the two function points exceeds the set ratio threshold, then they are judged to be duplicates, otherwise they are judged to be non-duplicates.

8. The function point duplicate checking device based on large language model application according to claim 7 is characterized in that: When determining that the judgment result is inconsistent with the reference result, the reflection iteration module guides the agent based on the ReAct mechanism to enter the reflection phase and iterates the reflection in each round of iteration to: Record the judgment result of whether the two function points are repeated in the current iteration and the historical execution information when the corresponding reasons are generated; When it is determined that the judgment result is inconsistent with the reference result, it is assumed that the judgment result is an erroneous judgment, and the large language model Gemma2 is instructed to generate a new, more correct and different judgment result from the opposite angle of the judgment result, and explain the corresponding reason of the new judgment result, and record the corresponding reason as reflection information; The historical execution information and the reflection information are added to the historical information and then transmitted to the agent based on the ReAct mechanism. The agent based on the ReAct mechanism combines the updated historical information to enter the second round of iterative reflection, and obtains the judgment result and corresponding reason of whether the two function points generated in the second round of iteration are repeated.

9. The function point duplicate checking device based on large language model application according to claim 6, characterized in that: The operation extraction rules that the large language model Qwen needs to follow when extracting operations include: The operation in a function refers to the action performed on the object described in the function point; The object attributes in a function refer to the set of objects being operated as described in the function point; The subject object in the function refers to the object being operated described in the function point; The details of the action cannot be modified, explained or analyzed; The separated actions are output in the form of serial numbers and start with the serial numbers. Only the actions are output.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the function point duplication checking method based on the large language model application as described in any one of claims 1 to 5 are implemented.

Images