CN117421226A - Defect report reconstruction method and system based on large language model - Google Patents

Abstract

The invention discloses a method and system for defect report reconstruction based on a generative large language model. The method includes the following steps: collecting and preprocessing statement sets and report sets, using the statement sets to establish a classification decision-making model, and using the classification decision-making model. Evaluate the quality of centralized defect reports, reconstruct low-quality defect reports using the established prompt template based on the generative large language model, and output a new defect report. The system includes a data acquisition module, a pre-evaluation data processing module, a defect report evaluation module, an input text creation module, and a defect report reconstruction module. Compared with the existing technology, the present invention has the characteristics of strong practicability and high accuracy.

Description

A method and system for defect report reconstruction based on a generative large language model

Technical field

The invention relates to the field of software maintenance, and in particular to a method and system for defect report reconstruction based on a generative large language model.

Background technique

Software defects refer to problems in software, such as documents and programs, that affect the normal operation of the software. They are usually called bugs. Defect reports are indispensable and necessary information in the software maintenance process. They describe software defect phenomena and a collection of reproduction steps. Defect reports are one of the work results of software testers and reflect the value of software testing. Defect reports can accurately identify the defects in software. Description and positioning to facilitate developers to correct. At the same time, it reflects the current quality status of the project/product to facilitate overall project progress and quality control.

However, existing defect reports are mainly written and analyzed manually. Due to the influence of factors such as the varying experience of defect reporters, the complexity of the software itself and its usage scenarios, and the imperfect functions of the defect tracking system, defect fixers refer to The quality of defect reports is often uneven, which causes great trouble to software maintenance. It is not accurate and efficient. Additionally, relying on manual analysis adds cost and overhead.

Existing defect report quality improvement measures include improvements to the defect reporting mechanism and system as well as improvements to the content of defect reports. The present invention improves the content. Existing automatic improvements to the quality of defect reports themselves mainly use similar defect reports to supplement missing information in defect reports. For example, there are currently some works using a hybrid analysis method that combines dynamic and static methods to improve Android applications or crowdsourced test applications. For example, the document "Ctras: Crowdsourcedtest report aggregation and summarization" uses redundant test reports to generate an enhanced version with richer content. Defect report. Group the redundant defect reports according to the description text and pictures in the report, select the most informative report in the group as the master report, and then use the remaining redundant reports to supplement defect understanding and repair The function information is integrated into the main report to obtain a higher quality crowd testing report. There are also some works such as "Bug report enrichment with application of automated fixerrecommendation" that propose methods of adding additional text sentences to enrich their content. The sentence to be added is taken from the problem description content in the historical report. The author first calculated the values of 6 similarity features such as the candidate sentence and the text, topic, component/product of the current report to be expanded, and weighted the 6 feature values. The final similarity value is obtained by summing, and the top K sentences with the largest value are used to expand the short text report. However, due to different project versions of the same function and different minor problems in the same software function, even if the content of the defect report is generally the same, the reported defects may be completely different. Moreover, the use of similar defect reports relies on the existing data of the project. If it is a new project and problem, the effect of this improvement measure will be greatly reduced. The quality of the established defect report is not high and the practicality is not strong.

Contents of the invention

Purpose of the invention: The purpose of the invention is to provide a method and system for defect report reconstruction based on a generative large language model that is highly accurate and practical.

Technical solution: A method for reconstructing defect reports based on a generative large language model according to the present invention, which method includes the following steps:

S1. Collect and preprocess an initial statement set to form a processed statement set. The initial statement set includes several statements and several tags, and one statement corresponds to at least one tag. Collect a report set, and the report set includes several defects. Report, each defect report includes ID, title, and description text;

S2. Perform data enhancement on the processed statement set, and merge it with the initial statement set to obtain an expanded statement set. The expanded statement set includes a label set; preprocess the defect report to form a statement list, and the statement list and defect report one-to-one correspondence;

S3. Establish a classification decision-making model based on the pre-training model of the attention mechanism and the expanded statement set, input the statement list into the classification decision-making model, and the classification decision-making model allocates and counts labels in the label set for the statement list, and evaluates The quality of the defect report corresponding to the statement list is divided into high-quality defect report and low-quality defect report according to the evaluation results;

S4. Obtain the generative large language model and establish the input text of the generative large language model. The input text includes a prompt template and a report text; wherein the prompt template is obtained by manual writing and optimization of the generative large language model; so The description text of the report is extracted from the description text of the low-quality defect report;

S5. Enter the input text into the generative large language model, reconstruct the low-quality defect report, and output a new defect report.

Further, in step S1, the process of preprocessing the initial statement set to form a statement set is as follows: 1) Delete duplicate statements and corresponding labels; 2) Clear illegal characters and emphasized characters in the statements, and retain the corresponding labels; 3 )Delete statements and corresponding labels whose number of characters is less than 5 or greater than 1000.

Further, in step S2, the process of preprocessing the defect report to form a statement list is as follows: 1) Delete non-text data in the defect report; 2) Clear illegal characters and emphasized characters in the defect report; 3) Call NLTK's sent_tokenize The function divides the title and description text of the defect report into several statements with a character count of less than or equal to 1000, and summarizes them into a statement list.

Further, in step S2, the data enhancement includes random replacement operation, random insertion operation and random exchange position operation.

Further, in step S3, establishing the classification decision model includes the following steps:

S31. Data preparation: use part of the expanded statement set as a training set and the other part as a verification set;

S32. Establish an initial model: obtain a pre-training model based on the attention mechanism, set the task to a multi-label classification task, and set a loss function for the multi-label classification task to learn the knowledge in the expanded statement set, and classify the labels based on the knowledge. Statements in the defect list that are assigned to the defect report.

S33. Train the initial model to obtain the classification decision-making model: use the training set to train the initial model, and use the verification set to verify. Stop training after meeting the indicator constraints to obtain the classification decision-making model.

Further, the indicators include accuracy, precision, recall and F1 score.

Further, in step S1, the label includes an empty label; in step S2, the label set includes empty labels, OB, EB, and S2R.

Further, in step S3, the process of evaluating the quality of the defect report includes:

1) Receive the statement list corresponding to the defect report;

2) Assign a label to each statement in the statement list;

3) Count and determine whether the statement list includes OB, EB, and S2R at the same time. If so, evaluate the defect report as a high-quality report; if not, evaluate the defect report as a low-quality defect report.

Further, it also includes steps to verify and reconstruct new defect reports:

Preprocess the new defect report, form a corresponding statement list and input it into the classification decision model, evaluate the quality of the new defect report, if it is evaluated as a low-quality defect report, extract the description text of the low-quality defect report as Report text, repeat steps S4-S5 until it is evaluated as a high-quality defect report.

A defect report reconstruction system based on a generative large language model according to the present invention, the system includes:

The data collection module is used to collect and preprocess an initial statement set to form a processed statement set. The initial statement set includes several statements and several types of labels, and one statement corresponds to at least one label; collects a report set, and the report set Includes several defect reports, each of which includes ID, title, and description text.

The pre-evaluation data processing module is used to perform data enhancement on the processed statement set and merge it with the initial statement set to obtain an expanded statement set, where the expanded statement set includes a label set; preprocess the defect report to form a statement list, so The statement list corresponds to the defect report one-to-one.

The defect report evaluation module is used to establish a classification decision-making model based on the pre-training model of the attention mechanism and the expanded statement set, input the statement list into the classification decision-making model, and the classification decision-making model allocates and counts the labels for the statement list. Centralize tags to evaluate the quality of defect reports corresponding to the statement list, and divide them into high-quality defect reports and low-quality defect reports based on the evaluation results.

The input text creation module is used to obtain the generative large language model and establish the input text of the generative large language model. The input text includes a prompt template and a report text; wherein the prompt template is manually written and generated by the generative large language model. Obtained through model optimization; the report text is extracted from the description text of low-quality defect reports.

A defect report reconstruction module is used to input the input text into the generative large language model, reconstruct the low-quality defect report, and output a new defect report.

Beneficial effects: The present invention has the following significant effects: 1. High accuracy: using the pre-training model based on the attention mechanism and the expanded sentence set after data enhancement to train a classification decision-making model with high accuracy; at the same time, using the generative large language model Reconstruct low-quality defect reports and use the massive knowledge and prompt templates inside the generative large language model to guide the generation of a high-quality new defect report; 2. Highly practical: different from traditional manual analysis and use of similar content to expand the original For low-quality defect reports, this invention uses the massive domain knowledge inside the generative large language model to automatically write a high-quality defect report that satisfies software maintenance, reduces the cost of manual understanding of low-quality defect reports, and at the same time can provide support for downstream automated software engineering. The task provides a high-quality data foundation, with wider practical application areas and higher efficiency.

Description of the drawings

Figure 1 is a general flow chart of the method of the present invention.

Figure 2 is a specific flow chart for generating a new defect report.

Detailed ways

The present invention will be further elucidated below in conjunction with the accompanying drawings and specific embodiments.

Please refer to Figure 1. The present invention discloses a method for reconstructing defect reports based on a generative large language model. The method includes the following steps:

S1. Collect and preprocess an initial statement set to form a processed statement set. The initial statement set includes several statements and several tags, and one statement corresponds to at least one tag. Collect a report set, and the report set includes several defects. Report, each defect report includes ID, title, and description text.

S2. Perform data enhancement on the processed statement set, and merge it with the initial statement set to obtain an expanded statement set, where the expanded statement set includes a label set. The defect report is preprocessed to form a statement list, and the statement list corresponds to the defect report one-to-one.

S3. Establish a classification decision-making model based on the pre-training model of the attention mechanism and the expanded statement set, input the statement list into the classification decision-making model, and the classification decision-making model allocates and counts the labels in the label set for the statement list, and evaluates the statements The quality of the defect reports corresponding to the list is divided into high-quality defect reports and low-quality defect reports according to the evaluation results.

S4. Obtain the generative large language model and establish the input text of the generative large language model. The input text includes a prompt template and a report text; wherein the prompt template is obtained by manual writing and optimization of the generative large language model; so The description text is extracted from the description text of low-quality defect reports.

S5. Enter the input text into the generative large language model, reconstruct the low-quality defect report, and output a new defect report.

Among them, the method is explained in detail. First, an initial statement set and report set are collected from open source communities and defect tracking platforms (such as JIRA, Bugzilla, Mozilla) and public manually annotated defect report statements and statement label data sets.

In step S1, the process of preprocessing the initial statement set to form the processed statement set is as follows: 1) Delete duplicate statements and corresponding labels; 2) Clear illegal characters and emphasized characters (characters that cannot be encoded by UTF-8, emoji) in the statement emoticons, etc.), and retain the corresponding tags; 3) Delete sentences and corresponding tags whose number of characters is less than 5 or greater than 1,000.

In step S2, the process of preprocessing the defect report to form a statement list is as follows: 1) Delete non-text data (such as hyperlinks, pictures) in the defect report; 2) Clear illegal characters and emphasized characters in the defect report; 3) Call NLTK The sent_tokenize function divides the title and description text of the defect report into several statements with a character count of less than or equal to 1000, and summarizes them into a list of statements. Among them, NLTK (Natural Language Toolkit) is a widely used Python natural language processing tool library, which provides a general and easy-to-use natural language processing toolset. It also provides scalable and reusable modules and algorithms to meet the needs of The needs of different users. The sent_tokenize function implements sentence boundary detection and sentence segmentation through a combination of statistics and heuristic rules. Its core is to train the algorithm to learn the sentence characteristics of a given language.

In step S2, data enhancement includes random replacement operations, random insertion operations and random exchange position operations to solve the problem of data category imbalance and obtain better data. The specific process is as follows:

S21. Use the word_tokenize function in NLTK to tokenize all words in the statement and turn each word in the statement into a token. Among them, word_tokenize implements word boundary detection and word division through a combination of statistics and heuristic rules. Tokenization is a type of de-identification technology that replaces plaintext data with pseudonymous tokens that correspond to it one-to-one. Tokens are used for efficient circulation in subsequent application environments, because tokens correspond to plaintext one-to-one and can replace plaintext transmission, exchange, storage and use in most scenarios.

S22. The random replacement operation is: randomly select n tokens in a sentence, obtain the synonym set corresponding to each selected token from the WordNet database, randomly select a synonym from the synonym set, and replace the corresponding token. Among them, n=0.01*token_num, token_num is the number of tokens contained in a sentence; WordNet is a large-scale English vocabulary database that organizes English words and phrases according to their meanings and establishes the relationship between word meanings. kind of relationship. Synonyms are organized into a synset.

The random insertion operation is: randomly select n tokens in a sentence, obtain the synonym set corresponding to each selected token from the WordNet database, randomly select a synonym from the synonym set, and randomly select a synonym in the sentence. Insert the synonym at the position.

The random exchange position operation is: randomly selecting two tokens in one statement and exchanging the positions of the two tokens.

It is worth noting that during the data augmentation process, the labels of the newly generated sentences are consistent with the labels of the original sentence samples. In this embodiment, in step S1, the label includes an empty label. In step S2, the label set includes empty labels, OB, EB, and S2R.

In step S3, a classification decision model is established, which is a process of using the pre-training model BERT based on the attention mechanism and the expanded statement set to assign zero or more labels from the label set OB, EB, and S2R to the statements in the statement list. model, train the classification decision model and verify the effectiveness of the model. Specifically, it includes the following steps:

S31. Data preparation: In order to prevent the trained classification decision-making model from overfitting, part of the expanded statement set is used as a training set, and the other part is used as a verification set. The expanded statement set is obtained by merging the enhanced data set and the initial statement set. This can expand the size of the data set and facilitate the classification decision-making model to learn more useful and comprehensive knowledge.

S32. Establish the initial model: Obtain the pre-training model based on the attention mechanism, set the task to a multi-label classification task, and set the loss function of the multi-label classification task to BCEWithLogitsLoss to learn the knowledge in the expanded statement set, and based on the knowledge Assign labels to statements in the defect list corresponding to the defect report; the formula for BCEWithLogitsLoss is:

Among them, N is the number of samples; C is the number of labels; _xi is the model output of the i-th sample; _yi is the true label of the i-th sample, which is 0 or 1.

Among them, the existing method to complete the multi-label classification task is to train three two-classification models, train a yes or no two-classification model for each label in OB, EB, and S2R, and then compare the results of the three two-classification models. Splicing. The result of this is that multiple models need to be trained, and multi-label samples are input into the two-classification model to assign a single label, which does not meet the characteristics of the expanded statement set and will affect the performance of the model. Using the pre-trained model BERT to complete the multi-label classification task of statements only requires training one model, and all labels in the label set are fully utilized during the training process. The knowledge learned by the model is more in line with the characteristics of the expanded statement set. Commonly used loss functions in multi-label classification tasks are Binary CrossEntropy and BCEWithLogitsLoss loss functions. These loss functions can simultaneously calculate the probability of each label and allow one statement to belong to multiple labels. However, the BCEWithLogitsLoss used in this embodiment calculates the loss directly on the original output of the model without first applying the Sigmoid function. This not only improves calculation efficiency, but also avoids the numerical instability problem caused by the Sigmoid function.

S33. Train the initial model to obtain the classification decision-making model: use the training set to train the initial model, and use the verification set to verify. Stop training after meeting the indicator constraints to obtain the classification decision-making model. Among them, the indicators include accuracy, precision, recall and F1 score. Accuracy is the ratio of the number of correctly predicted labels to the total number of labels. Precision is the ratio of the number of samples correctly predicted as positive examples to the number of all samples predicted as positive examples. Recall is the ratio of the number of samples correctly predicted as positive examples to the number of all samples that are actually positive examples. In many scenarios, not only must the prediction be accurate, but the model must also be able to identify as many positive examples as possible, which requires Evaluate the model's recall, that is, its ability to identify positive samples. The F1 score (F1-score) is the harmonic average of Precision and Recall, which can comprehensively consider the accuracy and recall rate of the model and evaluate the effect of the model more comprehensively.

These evaluation metrics can measure both the performance of the model on each label and the overall performance of the model on multiple labels. And, the accuracy formula is as follows:

The accuracy formula is as follows:

The recall formula is as follows:

The F1 score formula is as follows:

In the formula, TP is the true positive example, that is, the number of positive cases that are predicted to be positive but actually positive; TN is the true negative cases, that is, the number of negative cases that are predicted to be negative but actually negative; FP is the false positive case, that is, the number of positive cases that are predicted to be positive but actually negative The number; FN is the number of false negatives, that is, the number that is predicted to be negative but is actually positive.

In step S3, the process of evaluating defect report quality includes:

1) Receive the statement list corresponding to the defect report;

2) Assign a label to each statement in the statement list;

3) Count and determine whether the statement list includes OB, EB, and S2R at the same time; if included, evaluate the defect report as a high-quality report; if not, evaluate the defect report as a low-quality defect report.

In step S4, the process of establishing a prompt template is as follows:

S41. Manually write the initial prompt template;

The prompt template includes role assignment text, definition and requirements of tags (OB, EB, S2R), and task description. The role assignment text is used to assign roles to the generative large language model, and guide the generative large language model to output knowledge that is more in line with the role. The definition and requirements of the labels are used to help the generative large language model understand the key information in the task. The task description is used to allow the generative large language model to understand the task and constrain the output range of the model. In this embodiment, the role assignment text is: Your role is a senior software engineer, so you have rich experience in the fields of software testing and software maintenance. The definition and requirements of tags (OB, EB, S2R) are specifically, OB (observation behavior): related software behavior, action, output or result. Sentences without information like "The system doesn't work" are not considered OB. EB (Expected Behavior): Sentences containing phrases related to what the software should/want/expect to happen, such as 'should...', 'expect...', 'hope...' should be considered EB. Suggestions or recommendations to fix bugs are not considered EB. S2R (Step to Reproduction): If a sentence potentially contains an action or operation by the user, it should be considered S2R. Things like 'to reproduction', 'steps to reproduction', 'follow these steps', are not considered S2R". In this example, the task description is: You should infer appropriate details from the context and use clear OB/EB /S2R supplements complete bug reports. Where possible, you should improve the wording of existing OB, EB, and S2R statements to make them clearer. For S2R, you should give as clear, accurate, and complete steps as possible .

S42. Input the initial prompt template into the generative large language model for summary and optimization, and output the optimized prompt template. Let the generative large language model summarize and optimize the manually written initial prompt template, so that the model can more comprehensively and accurately understand the tasks and related definitions and requirements described in the prompt template.

S43. Verify the optimization prompt template, ensure that the optimization prompt template meets the task requirements and then output it as a prompt template. The verification method is manual verification.

In addition, the present invention also includes the steps of verifying and reconstructing the new defect report:

The new defect report output by preprocessing step S5 forms a corresponding statement list and is input into the classification decision model to evaluate the quality of the new defect report. If it is evaluated as a low-quality defect report, extract the low-quality defect report. The description text is used as the report text, and steps S4-S5 are repeated until it is evaluated as a high-quality defect report.

The invention also discloses a system for reconstructing defect reports based on a generative large language model. The system includes:

Data collection module: used to collect and preprocess an initial statement set to form a processed statement set. The initial statement set includes several statements and several tags, and one statement corresponds to at least one tag; collects a report set, and the report set Includes several defect reports, each of which includes ID, title, and description text;

Pre-assessment data processing module: used to perform data enhancement on the processed statement set and merge it with the initial statement set to obtain an expanded statement set, where the expanded statement set includes a label set; preprocess the defect report to form a statement list, so The statement list corresponds to the defect report one-to-one;

Defect report evaluation module: establish a classification decision-making model based on the pre-training model of the attention mechanism and the expanded statement set, input the statement list into the classification decision-making model, and the classification decision-making model allocates and counts the statements in the label set for the statement list Label, evaluate the quality of the defect report corresponding to the statement list, and divide it into high-quality defect report and low-quality defect report according to the evaluation results;

Input text creation module: obtains the generative large language model and establishes the input text of the generative large language model. The input text includes a prompt template and a report text; wherein the prompt template is manually written and optimized by the generative large language model. Obtained; the report text is extracted from the description text of the low-quality defect report;

Defect report reconstruction module: used to input the input text into the generative large language model, reconstruct the low-quality defect report, and output a new defect report.

Claims (10)

1. A method for reconstructing defect reports based on a generative large language model, characterized in that the method includes the following steps: S1. Collect and preprocess an initial statement set to form a processed statement set. The initial statement set includes several statements and several tags, and one statement corresponds to at least one tag. Collect a report set, and the report set includes several defects. Report, each defect report includes ID, title, and description text; S2. Perform data enhancement on the processed statement set, and merge it with the initial statement set to obtain an expanded statement set. The expanded statement set includes a label set; preprocess the defect report to form a statement list, and the statement list and defect report one-to-one correspondence; S3. Establish a classification decision-making model based on the pre-training model of the attention mechanism and the expanded statement set, input the statement list into the classification decision-making model, and the classification decision-making model allocates and counts labels in the label set for the statement list, and evaluates The quality of the defect report corresponding to the statement list is divided into high-quality defect report and low-quality defect report according to the evaluation results; S4. Obtain the generative large language model and establish the input text of the generative large language model. The input text includes a prompt template and a report text; wherein the prompt template is obtained by manual writing and optimization of the generative large language model; so The description text of the report is extracted from the description text of the low-quality defect report; S5. Enter the input text into the generative large language model, reconstruct the low-quality defect report, and output a new defect report. 2. The method of defect report reconstruction based on a generative large language model according to claim 1, characterized in that, in step S1, the process of preprocessing the initial statement set to form a statement set is as follows: 1) Delete duplicate statements. and corresponding labels; 2) Clear illegal characters and emphasized characters in the statement, and retain the corresponding labels; 3) Delete statements and corresponding labels whose number of characters is less than 5 or greater than 1,000. 3. The method for reconstructing a defect report based on a generative large language model according to claim 1, characterized in that, in step S2, the process of preprocessing the defect report to form a statement list is as follows: 1) Delete the defect report non-text data; 2) Clear illegal characters and emphasized characters in the defect report; 3) Call NLTK's sent_tokenize function to divide the title and description text of the defect report into several statements with a number of characters less than or equal to 1000, and summarize them into statements list. 4. The method of defect report reconstruction based on a generative large language model according to claim 2, characterized in that in step S2, the data enhancement includes a random replacement operation, a random insertion operation and a random exchange position operation. 5. The method of defect report reconstruction based on a generative large language model according to claim 1, characterized in that, in step S3, establishing a classification decision model includes the following steps: S31. Data preparation: use part of the expanded statement set as a training set and the other part as a verification set; S32. Establish an initial model: obtain a pre-training model based on the attention mechanism, set the task to a multi-label classification task, and set a loss function for the multi-label classification task to learn the knowledge in the expanded statement set, and classify the labels based on the knowledge. Statements in the defect list assigned to the defect report; S33. Train the initial model to obtain the classification decision-making model: use the training set to train the initial model, and use the verification set to verify. Stop training after meeting the indicator constraints to obtain the classification decision-making model. 6. The method of defect report reconstruction based on a generative large language model according to claim 5, wherein the indicators include accuracy, precision, recall and F1 score. 7. The method of defect report reconstruction based on a generative large language model according to claim 1, characterized in that, in step S1, the label includes an empty label; in step S2, the label set includes an empty label, OB, EB, S2R. 8. The method of defect report reconstruction based on a generative large language model according to claim 7, characterized in that, in step S3, the process of evaluating the quality of the defect report includes: 1) Receive the statement list corresponding to the defect report; 2) Assign a label to each statement in the statement list; 3) Count and determine whether the statement list includes OB, EB, and S2R at the same time. If so, evaluate the defect report as a high-quality report; if not, evaluate the defect report as a low-quality defect report. 9. The method for reconstructing defect reports based on a generative large language model according to claim 1, further comprising the step of verifying and reconstructing the new defect report: Preprocess the new defect report, form a corresponding statement list and input it into the classification decision model, evaluate the quality of the new defect report, if it is evaluated as a low-quality defect report, extract the description text of the low-quality defect report as Report text, repeat steps S4-S5 until it is evaluated as a high-quality defect report. 10. A system for reconstructing defect reports based on a generative large language model, characterized in that the system includes: The data collection module is used to collect and preprocess an initial statement set to form a processed statement set. The initial statement set includes several statements and several types of labels, and one statement corresponds to at least one label; collects a report set, and the report set Includes several defect reports, each of which includes ID, title, and description text; The pre-evaluation data processing module is used to perform data enhancement on the processed statement set and merge it with the initial statement set to obtain an expanded statement set, where the expanded statement set includes a label set; preprocess the defect report to form a statement list, so The statement list corresponds to the defect report one-to-one; The defect report evaluation module is used to establish a classification decision-making model based on the pre-training model of the attention mechanism and the expanded statement set, input the statement list into the classification decision-making model, and the classification decision-making model allocates and counts the labels for the statement list. Centralize labels to evaluate the quality of defect reports corresponding to the statement list, and divide them into high-quality defect reports and low-quality defect reports based on the evaluation results; The input text creation module is used to obtain the generative large language model and establish the input text of the generative large language model. The input text includes a prompt template and a report text; wherein the prompt template is manually written and generated by the generative large language model. Obtained through model optimization; the report text is extracted from the description text of low-quality defect reports; A defect report reconstruction module is used to input the input text into the generative large language model, reconstruct the low-quality defect report, and output a new defect report.