Issue-Commit link recovery method and system

Abstract

The invention provides an Issue-Commit link recovery method and system, and relates to the technical field of data processing.The method comprises the steps that when an input text exceeds model limitation, an abstract after semantic compression is generated through a semantic abstract module driven by LLM; inputting the abstract or the original text into a link prediction module formed by a pre-training encoder to obtain a candidate set; and inputting the candidate set and the original text into a decoder-only LLM driven semantic reordering module, carrying out reordering, and outputting a final result. According to the method, the truncation loss of the input text can be reduced through the semantic abstract module driven by the LLM at the input end, and the precision of the final result is improved by reordering through the semantic reordering module driven by the decoder-only LLM at the output end, so that the accuracy of the prediction result output by the Issue-Commit link recovery method in an engineering scene is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for restoring Issue-Commit links. Background Technology

[0002] In defect management and source tracing analysis of large-scale software projects, correctly matching defect reports (Issue) with actual fix submissions (Commit) is a crucial foundational task in engineering and research. Currently, Issue-Commit link recovery methods are typically based on information retrieval and vector retrieval, or on deep semantic representation.

[0003] The general technical approach is as follows: First, a language model pre-trained on a large-scale corpus (such as natural language text, source code, and software commit logs) is used to encode the Issue and Commit texts to obtain context-related semantic representations; then, similarity is calculated based on these semantic representations, or the correspondence between Issue and Commit is further learned through a trained matching model; finally, a candidate commit list (Top-k) is returned and output as the prediction result.

[0004] However, the aforementioned existing technologies have two significant drawbacks in practical engineering scenarios, affecting the link recovery effect and practical value, specifically: 1. Semantic loss caused by text truncation: Modern Transformer-like models and many retrieval / encoding processes impose strict limitations on input length (in units of tokens). To meet these maximum input length constraints, Issue and Commit statements are often truncated in engineering practice—that is, the first N tokens are removed. The problem is that the truncated text often contains information crucial to the "fixing intent" or "semantic clues" (such as the specific filename of the code modification, key statement fragments, key information in the error stack, etc.). This loss of semantic content destroys the integrity of the encoded representation, causing the encoded vector to fail to accurately reflect the true semantic intent of the Issue or Commit, resulting in inaccurate similarity assessment, missed retrieval candidates, or scoring bias, ultimately reducing link prediction performance.

[0005] 2. Candidate ranking bias issue: Existing methods often assume that a task is complete as long as the correct commit is included in the Top-k candidates returned by the model. However, in real-world applications, a more important requirement is to prioritize correct commits (especially Top-1) so that automation tools or engineers can immediately adopt their suggestions. Real-world results show that even when correct commits are retrieved, they often do not rank highly in the candidate list. This is because the scoring mechanism of the initial retrieval (or dual encoder) often cannot distinguish the "best match" at the fine-grained semantic differences. Furthermore, retrieval scores are affected by lexical overlap, historical bias, etc., causing the "best" commit, which is actually useful in engineering, to fail to receive the highest score, thus weakening the usability and reliability of the prediction results in a production environment.

[0006] It is evident that the aforementioned problems with existing technologies ultimately affect the accuracy of the prediction results output by the Issue–Commit link recovery method. Summary of the Invention

[0007] In view of the shortcomings of the prior art, the purpose of this invention is to provide an Issue-Commit link recovery method and system, which can solve the technical problem that the Issue-Commit link recovery method in the prior art outputs inaccurate prediction results.

[0008] A first aspect of this invention proposes a method for restoring Issue-Commit links, comprising: When the input text exceeds the model's limits, a semantically compressed summary is generated through the LLM-driven semantic summarization module. The abstract or original text is input into the link prediction module, which is composed of a pre-trained encoder, to obtain a candidate set; The candidate set is reordered with the original text input by a decoder-only LLM-driven semantic reordering module, and the final result is output.

[0009] Optionally, a semantically compressed summary is generated using an LLM-driven semantic summarization module, specifically including: Construct a summary Prompt template; Set the summarization inference parameters for the LLM-driven semantic summarization module; wherein, the summarization inference parameters include temperature and summarization length range; The LLM-driven semantic summarization module performs semantic compression on the input text based on the summarization inference parameters and the summarization prompt template to obtain a compressed summary of the input text.

[0010] Optionally, the temperature value can be 0.

[0011] Optionally, the digest length ranges from 16 to 256 tokens.

[0012] Optionally, the step of reordering the candidate set with the original text input decoder-only LLM-driven semantic reordering module and outputting the final result specifically includes: Construct system instructions, reorder Prompts, and output formats; Set the reordering inference parameters for the decoder-only LLM-driven semantic reordering module; The semantic reordering module driven by the decoder-only LLM reorders the input candidate set based on the system instructions, the reordering prompt, the output format, the reordering inference parameters, and the original text, and outputs the final result.

[0013] Optionally, the semantic reordering module driven by the decoder-only LLM reorders the input candidate set based on the system instructions, the reordering prompt, the output format, the reordering inference parameters, and the original text, and outputs the final result, specifically including: The system instructions, the reordering prompt, the output format, the reordering inference parameters, the original text, and the candidate set are input into the decoder-only LLM-driven semantic reordering module, so that the decoder-only LLM-driven semantic reordering module compares all candidate commit texts in the candidate set at once and outputs the final result.

[0014] Optionally, the semantic reordering module driven by the decoder-only LLM reorders the input candidate set based on the system instructions, the reordering prompt, the output format, the reordering inference parameters, and the original text, and outputs the final result, specifically including: The system instructions, the reordering prompt, the output format, the reordering inference parameters, the original text, and the candidate set are input into the decoder-only LLM-driven semantic reordering module, so that the decoder-only LLM-driven semantic reordering module analyzes each candidate commit text in the candidate set pairwise for the Issue, and obtains the ranking score of each candidate commit text. The candidate set is aggregated and sorted based on the order of each candidate commit text to obtain the final result.

[0015] Optionally, the candidate commit text in the candidate set can range from 5 to 20 tokens.

[0016] A second aspect of the present invention provides an Issue-Commit link recovery system, comprising: a processor and a memory; The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the Issue-Commit link recovery method as described in the first aspect.

[0017] A third aspect of the present invention provides a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the Issue-Commit link recovery method as described in the first aspect.

[0018] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this embodiment of the invention, the input text truncation loss is reduced by an LLM-driven semantic summarization module at the input end, and the final result accuracy is improved by a decoder-only LLM-driven semantic reordering module at the output end. This significantly improves the accuracy of the prediction results output by the Issue-Commit link recovery method in engineering scenarios. Attached Figure Description

[0019] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0020] Figure 1 This is a flowchart illustrating an Issue-Commit link recovery method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an Issue-Commit link recovery system provided in an embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] The Issue-Commit link recovery method provided by the present invention will be described in detail below with reference to the accompanying drawings, through specific embodiments and application scenarios.

[0023] Reference manual attached Figure 1 The diagram illustrates a flowchart of an Issue-Commit link recovery method provided by an embodiment of the present invention.

[0024] This invention provides a method for restoring Issue-Commit links, which may include the following steps: S101: When the input text exceeds the model limit, a semantically compressed summary is generated through the LLM-driven semantic summarization module.

[0025] In this embodiment of the application, the input text may be the original software engineering text to be processed, which includes the title and description of the Issue, and the commit message of the Commit.

[0026] Basic cleaning methods for raw software engineering text can include: unifying character encoding, removing invisible characters, and removing obvious hyperlinks or explicit problem markers (such as "#1234") to avoid data leakage; alternatively, the original semantic structure can be preserved, and stop-word removal or stemming can be avoided on the input of the LLM-driven semantic summarization module at this stage to prevent damage to the contextual understanding ability of the natural language model.

[0027] In this embodiment of the application, a tokenizer can be used to decompose the original software engineering text into sub-word tokens and calculate the token length L_input.

[0028] Set the maximum token threshold MaxToken (preferred embodiment: MaxToken is 35 tokens to be compatible with the input limit of the backbone encoder). If L_input ≤ MaxToken, then the original software engineering text (or the lightly formatted original software engineering text) is directly marked as "passed" and used as the input for the subsequent S102; If L_input > MaxToken, then proceed with the digest compression process.

[0029] As an optional implementation, S101 may generate a semantically compressed summary through an LLM-driven semantic summarization module in the following ways: Construct a summary Prompt template; Set the summarization inference parameters for the LLM-driven semantic summarization module; wherein, the summarization inference parameters include temperature and summarization length range; The LLM-driven semantic summarization module performs semantic compression on the input text based on the summarization inference parameters and the summarization prompt template to obtain a compressed summary of the input text.

[0030] This implementation method involves constructing a summary prompt template, setting inference parameters including temperature and summary length range, and using an LLM-driven semantic summarization module to perform semantic compression on the input text. This method can accurately guide the direction of semantic summary generation, the temperature parameter can flexibly adjust the creativity and determinism of the generated content, and the reasonable setting of the summary length range ensures that the result is concise and the information is complete. Ultimately, it efficiently generates compressed summaries that meet the requirements, improving the efficiency and quality of information processing.

[0031] In this embodiment, the temperature value is 0. The summary length ranges from 16 to 256 tokens. The summary prompt template can contain system instructions and task definitions, explicitly requiring the LLM to generate a summary that "preserves the core intent and remediation actions while having a limited length." (Preferred embodiment: summary length is approximately 35–45 tokens). Temperature=0 ensures stable output; sampling can be optionally used to generate multiple candidate summaries for posterior fusion.

[0032] In this embodiment of the application, a decoder-only LLM (a large language model with only decoder, preferably DeepSeek-v3, GPT-like model or equivalent model) can be used to perform semantic compression on the input text to obtain a semantically compressed summary S_text.

[0033] Optionally, the compressed summary can also be cleaned by standardizing character encoding, removing invisible characters, and removing obvious hyperlinks or explicit problem markers.

[0034] The tokenizer is used to decompose the summary into sub-tokens, and the output is passed to S102 in a standardized data structure.

[0035] S102: Input the summary or original text into the link prediction module composed of a pre-trained encoder to obtain a candidate set.

[0036] In this embodiment, the link prediction module is a backbone prediction engine compatible with the present invention (referencing the existing lssue-Commit link recovery method based on pre-trained models), which serves as a robust and efficient candidate generator in the present invention.

[0037] In this embodiment of the application, the training method for the link prediction module can be: 1. Data Construction and Negative Sample Strategy: 1.1 Collect training sets from the project repository, construct positive samples (issue, commit) (obtained by explicit citation or manual annotation), and construct negative samples (using a similarity-aware selection strategy).

[0038] 1.2 Use semantically summarized text (by calling the LLM-driven semantic summarization module) or use the original text directly (depending on the training scheme).

[0039] 2. Knowledge distillation and student model training (refer to the existing lssue-Commit link recovery method based on pre-trained models): 2.1 Select the teacher model (CodeBERT) as the teacher network; construct the student model; 2.2 Knowledge distillation in the intermediate layer: Minimize the MSE (mean squared error) of the hidden representations of the i-th layer of the teacher model and the j-th layer of the student model to pass on the hierarchical semantic representation; 2.3 Fine-tuning the student model using multiple tasks: The main task is similarity regression / binary classification (maximizing the vector similarity of true link pairs and minimizing negative pairs). The auxiliary tasks include predicting related code files in commits (issue-code relevance) and contrastive loss based on one-to-many commits. The total loss is as follows: in, This represents the main link prediction loss based on similarity regression; This represents the comparative loss for a commit operation in a one-to-many link; This represents the auxiliary task loss used for the correlation of problem codes; It is the task weight.

[0040] For each input Issue, the prediction module calculates the similarity score of all candidate submissions in the test set and outputs a candidate set C_rerank based on the score.

[0041] S103: Reorder the candidate set with the original text input using a decoder-only LLM-driven semantic reordering module and output the final result.

[0042] In this embodiment, the top K_rerank candidate Commit texts can be selected from the candidate set C_rerank (K_rerank can be adjusted between 5 and 20 tokens, preferably 10 tokens; if the project LAG is large, K_rerank can be increased), as well as the original Issue text (title & description) and the original Commit text (Commit message). This application prefers to use the original text directly instead of the summary text to preserve the full context.

[0043] Prepare the necessary string format (e.g., a list of commit_ids, a list of commit_messages) for each candidate commit text, and construct an LLM reordering Prompt.

[0044] As an optional implementation, S103 may reorder the candidate set with the original text input decoder-only LLM-driven semantic reordering module and output the final result in the following ways: Construct system instructions, reorder prompts, and output formats; Set the reordering inference parameters for the decoder-only LLM-driven semantic reordering module; The semantic reordering module driven by the decoder-only LLM reorders the input candidate set based on the system instructions, the reordering prompt, the output format, the reordering inference parameters, and the original text, and outputs the final result.

[0045] This implementation method provides clear guidance and a standardized framework for semantic reordering by constructing system instructions, reordering prompts, and output formats, ensuring that the reordering process has a clear objective and precise direction. Setting reordering inference parameters allows for flexible adjustment of the model's running state to adapt to different scenario requirements. Utilizing a decoder-only LLM-driven semantic reordering module, the candidate set is reordered by integrating information from multiple sources, effectively uncovering deep semantic relationships in the text, improving the rationality of the ranking, and ultimately outputting high-quality results that better conform to semantic logic and practical needs, thus enhancing information processing efficiency.

[0046] In this embodiment of the application, the output format can be arranged in descending order of candidate numbers separated by commas, and only the number sequence is output without any additional text, so as to ensure the feasibility of parsing; Recommended template: Take the issue's title, description, and 10 candidate commit messages as input, and specify the output example format "3,5,1,2,4,6,7,8,9,10").

[0047] In this embodiment of the application, the reordering inference parameters can be set as follows: temperature=0 (deterministic), the maximum token threshold max_tokens is sufficient to return the sort, and the number of candidate commit messages top_p is set as needed.

[0048] In this embodiment of the application, in order to ensure parsing robustness, Prompt must strictly define the output format and perform output compliance checks (if the output does not conform to the rules, the original search sorting or LLM is used to re-call the fallback strategy).

[0049] In this embodiment of the application, the sorting method for reordering the input candidate set can be implemented in two equivalent ways: 1. Listwise reordering: Allows the LLM to compare all candidates at once and output a total order (preferred implementation). 2. Pairwise scoring: Call the LLM on each candidate pairwise to get the score or sorting votes, and then aggregate and sort the candidates (used to control the output format or parallelize the LLM calls).

[0050] Optionally, the semantic reordering module driven by the decoder-only LLM reorders the input candidate set based on the system instructions, the reordering prompt, the output format, the reordering inference parameters, and the original text, and the final result is output in the following manner: The system instructions, the reordering prompt, the output format, the reordering inference parameters, the original text, and the candidate set are input into the decoder-only LLM-driven semantic reordering module, so that the decoder-only LLM-driven semantic reordering module compares all candidate commit texts in the candidate set at once and outputs the final result.

[0051] This implementation method involves inputting multiple key information types, such as system commands and reordering prompts, along with the original text and candidate set, into a decoder-only LLM-driven semantic reordering module. This allows the module to comprehensively compare all committed texts in the candidate set at once. This holistic processing avoids the information fragmentation and error accumulation that can occur with step-by-step operations. It can more accurately grasp the semantic differences and relationships between candidate texts, thereby outputting a more relevant and logically coherent final result, effectively improving the accuracy and efficiency of semantic reordering and optimizing the information processing flow.

[0052] Optionally, the semantic reordering module driven by the decoder-only LLM reorders the input candidate set based on the system instructions, the reordering prompt, the output format, the reordering inference parameters, and the original text, and the final result is output in the following manner: The system instructions, the reordering prompt, the output format, the reordering inference parameters, the original text, and the candidate set are input into the decoder-only LLM-driven semantic reordering module, so that the decoder-only LLM-driven semantic reordering module analyzes each candidate commit text in the candidate set pairwise for the Issue, and obtains the ranking score of each candidate commit text. The candidate set is aggregated and sorted based on the order of each candidate commit text to obtain the final result.

[0053] This implementation method involves inputting multiple types of information into a decoder-only LLM-driven semantic re-ranking module. The module then analyzes each candidate commit text in the candidate set pairwise, deriving a ranking score. This pairwise analysis delves into the semantic fit between the issue and each candidate commit text, making the ranking scores more targeted and accurate. Subsequent aggregation and ranking based on these scores ensures that the final result fully considers the characteristics and relevance of each candidate text, effectively improving the rationality and accuracy of the re-ranking and providing a strong guarantee for obtaining high-quality, satisfactory results.

[0054] In this embodiment, the candidate commit text in the candidate set ranges from 5 to 20 tokens. The return value of LLM (expected to be a comma-separated sequence of indices or a score for each item) is parsed into a new candidate permutation R_new or a score vector s_re_rank.

[0055] It's also acceptable. Based on the ranking of R_new and the score of s_re_rank, output the final sorting result C_final. Each item contains commit_id, final score s_final (or LLM ranking), and (optionally) supporting evidence (LLM can be prompted to return key sentences or evidence fragments that support the ranking). In addition, manual confirmation / correction can be used as feedback: If the user or automatic verification confirms the correctness of the final ranking, the confirmed samples can be added to the incremental training set for subsequent online fine-tuning of the re-ranking module or the link prediction module (forming closed-loop learning).

[0056] In addition, in resource- or latency-constrained production environments, knowledge distillation of the LLM can be performed, using the sorting labels generated by teacherLLM to train a lightweight listwise sorter for high-concurrency deployment.

[0057] In this embodiment of the application, a list-style prompt (sorting all candidates at once) is preferred to allow LLM to utilize cross-example contrast.

[0058] Optionally, to reduce costs, it is recommended to cache multiple queries for the same issue (reuse LLM results within a certain time window).

[0059] Additionally, when the LLM cannot return a clear sort (e.g., the output is anomalous or inconsistent), a fallback is employed: either returning in descending order by s_retrieval or using a lightweight sorting model.

[0060] Several technical parameters and preferred scope of the preferred embodiments of the present invention Token truncation threshold MaxToken: This patent preferably specifies 35 tokens; optional range 16–512 tokens (depending on encoder capabilities).

[0061] Summary text length: preferably 35–45 tokens; optional 16–256 tokens.

[0062] Re-ranking Top-K (K_rerank): Preferred 10 tokens; optional 5–20 tokens.

[0063] LLM inference parameters: temperature=0 (good for experimental repeatability); max_output_tokens depends on the summary / sorting task.

[0064] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: In this embodiment of the invention, the input text truncation loss is reduced by an LLM-driven semantic summarization module at the input end, and the final result accuracy is improved by a decoder-only LLM-driven semantic reordering module at the output end. This significantly improves the accuracy of the prediction results output by the Issue-Commit link recovery method in engineering scenarios.

[0065] Reference manual attached Figure 2 The diagram shows a structural schematic of an Issue-Commit link recovery system provided in an embodiment of the present invention.

[0066] This invention provides an Issue-Commit link recovery system 20, comprising: a processor 201 and a memory 202; The memory 202 stores programs or instructions that can run on the processor 201. When the program or instructions are executed by the processor 201, they implement the steps of the Issue-Commit link recovery method described above and achieve the same technical effect. To avoid repetition, the present invention will not elaborate further.

[0067] It should be understood that the processor 201 in this embodiment of the invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0068] It should also be understood that the memory 202 in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DR RAM).

[0069] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0070] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0071] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0072] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0073] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0074] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0075] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0076] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0077] This invention provides a readable storage medium comprising: storing a program or instructions on the readable storage medium, wherein when the program or instructions are executed by a processor, the program or instructions implement the steps of the above-described Issue-Commit link recovery method and achieve the same technical effect. To avoid repetition, this invention will not elaborate further.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. A method for restoring Issue-Commit links, characterized in that, include: When the input text exceeds the model's limits, a semantically compressed summary is generated through the LLM-driven semantic summarization module. The abstract or original text is input into the link prediction module, which is composed of a pre-trained encoder, to obtain a candidate set; The candidate set is reordered with the original text input by a decoder-only LLM-driven semantic reordering module, and the final result is output.

2. The Issue-Commit link recovery method according to claim 1, characterized in that, The generation of semantically compressed summaries via the LLM-driven semantic summarization module specifically includes: Construct a summary Prompt template; Set the summarization inference parameters for the LLM-driven semantic summarization module; wherein, the summarization inference parameters include temperature and summarization length range; The LLM-driven semantic summarization module performs semantic compression on the input text based on the summarization inference parameters and the summarization prompt template to obtain a compressed summary of the input text.

3. The Issue-Commit link recovery method according to claim 2, characterized in that, The temperature value is 0.

4. The Issue-Commit link recovery method according to claim 2, characterized in that, The digest length ranges from 16 to 256 tokens.

5. The Issue-Commit link recovery method according to claim 1, characterized in that, The semantic reordering module, which reorders the candidate set with the original text input decoder-only LLM-driven semantic reordering module and outputs the final result, specifically includes: Construct system instructions, reorder Prompts, and output formats; Set the reordering inference parameters for the decoder-only LLM-driven semantic reordering module; The semantic reordering module driven by the decoder-only LLM reorders the input candidate set based on the system instructions, the reordering prompt, the output format, the reordering inference parameters, and the original text, and outputs the final result.

6. The Issue-Commit link recovery method according to claim 5, characterized in that, The semantic reordering module driven by the decoder-only LLM reorders the input candidate set based on the system instructions, the reordering prompt, the output format, the reordering inference parameters, and the original text, and outputs the final result, specifically including: The system instructions, the reordering prompt, the output format, the reordering inference parameters, the original text, and the candidate set are input into the decoder-only LLM-driven semantic reordering module, so that the decoder-only LLM-driven semantic reordering module compares all candidate commit texts in the candidate set at once and outputs the final result.

7. The Issue-Commit link recovery method according to claim 5, characterized in that, The semantic reordering module driven by the decoder-only LLM reorders the input candidate set based on the system instructions, the reordering prompt, the output format, the reordering inference parameters, and the original text, and outputs the final result, specifically including: The system instructions, the reordering prompt, the output format, the reordering inference parameters, the original text, and the candidate set are input into the decoder-only LLM-driven semantic reordering module, so that the decoder-only LLM-driven semantic reordering module analyzes each candidate commit text in the candidate set pairwise for the Issue, and obtains the ranking score of each candidate commit text. The candidate set is aggregated and sorted based on the order of each candidate commit text to obtain the final result.

8. The Issue-Commit link recovery method according to claim 7, characterized in that, The candidate commit text in the candidate set ranges from 5 to 20 tokens.

9. An Issue-Commit link recovery system, characterized in that, include: Processor and memory; The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the Issue-Commit link recovery method as described in any one of claims 1 to 8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the Issue-Commit link recovery method as described in any one of claims 1 to 8.

Images