Description:Complete Specification:

The following specification describes and ascertains the nature of this invention and the manner in which it is to be performed.
Field of the invention
[0001] The invention is related to a control unit for provided an efficient code summary and a method thereof.

Background of the invention

[0002] Understanding legacy codes in a big code repository is a challenge in the domain of software engineering. It has been estimated by Liang et.al., 2018 that only 15% of the methods are documented (in the Java GitHub repository). This makes it difficult and time-consuming for developers to comprehend the underlying functionality. For programmers new to a particular programming language, it will be helpful if code snippets are labelled with the corresponding intent that describes what purpose it achieves (MacNeil, 2023). Automating the task of code documentation through comments and explanations can therefore prove beneficial in many ways.

[0003] A technical paper by M. Geng, S.Wang, D. Dong, H.Wang, G. Li, Z. Jin, X. Mao, X. Liao, “Large language models are few-shot summarizers: multi-intent comment generation via in-context learning”, Accepted at the 46th International Conference on Software Engineering (ICSE 2024), discloses a similar concept.

Brief description of the accompanying drawings
[0004] Figure 1 illustrates a control unit for provided an efficient code summary according to one embodiment of the invention; and
[0005] Figure 2 illustrates a flowchart for a method of providing an efficient code summary by a control unit according to the present invention.

Detailed description of the embodiments
[0006] Figure 1 illustrates a control unit for provided an efficient code summary in accordance with one embodiment of the invention. The control unit 10 comprising a code database 12 storing multiple code snippets 14 and their corresponding code summaries and a retriever module 15 provides a related information upon comparing at least one stored code snippet 14(a) and an inputted code snippet 16. The control unit 10 generates a code summary that is relevant to the inputted code snippet 16 by a Large Language module (LLM) 18, based on the comparison between the at least one stored code snippet 14(a) and the inputted code snippet 16.

[0007]Further the construction of the control unit and the components of the control unit is explained in detail. The control unit 10 is chosen from a group of control units comprising a microcontroller, a microprocessor, a digital circuit, an integrated chip, and the like. The code database 12 comprises multiple code snippets 14 and their corresponding summaries. For instance, the code snippets 14 that are stored in the code database 12 are referred as (d1, d2, d3….dn).The control unit 10 performs at least one function using multiple hardware units like a retriever module 15 that comprises three sub modules , a NER (please give full form of this) naming module 15(a), a similarity score generation module 15(b), and a ranking module 15(c). The NER naming module 15(a) segregates each of the words present in the code snippet 14(a) into multiple categories, wherein the multiple categories are a library, a class, a data structure, a function and the like.

[0008]The named entities are the different keywords related to methodology like function, library, algorithm, data type etc. Thus, given any code snippet (14(a) or 16) , this module 15(a) extracts the relevant entities that are present in the code. The NER naming module 15(a) categories the words present in both stored code snippet 14(a) and the input code snippet 16.The similarity score generation module 15(b) generates multiple similarity matrix and their corresponding similarity score upon comparing each of the categorized stored code snippet 14(a) with the categorized input code snippet 16. This module 15(b) determines the percent of two code snippets (14(a) & 16) that can be similar to each other. For instance, if the “q” is a code snippet 16 that is inputted, and “d” is one of the code snippet 14(a) from the code data base 12. Then the NER naming module 15(a) will first segregates each of the words present in the code snippets (14(a) &16) d and q and obtains their entities. Then the control unit 10 using the similarity score generation module 15(b), then calculates a similarity between the corresponding entities of d & q (14(a) & 16) to obtain a similarity score ?????????? (??, ??) .

[0009]The ranking module 15(c) ranks each of the generated similarity matrix after the above disclosed comparison. The control unit 10 selects the relevant stored code snippets 14(a) based on the ranking for generation of code summary of the input code snippet 16. The NER similarity-based ranking of the legacy codes to obtain ???????? –, the control unit 10 applies the similarity score generation method by the second sub-module 15(b) (i.e.., the similarity score generation module 15(b)) between the query code snippet(q) 16 and all code snippets (d1, d2, d3, ….dn) (14(a) in the legacy code base 12 to obtain their ?????????? scores. The control unit 10 then ranks the code snippets 14(a) from the code database 12 in decreasing order of similarity and pick the top k code snippets 14(a) and their corresponding summary. The control unit 10 with the help of LLM 18 generates the code summary based on the input code snippet 16 and relevant stored code snippets 14(a) received from the retriever module 15. The relevant stored code snippets 14(a) are taken from the code data base 12 that have the highest similarity score.

[0010]Figure 2 illustrates a method of providing an efficient code summary by a control unit, in accordance with the present invention. In step S1, multiple code snippets 14 and their corresponding code summaries are stored in a code database 12. In step S2, a related information is retrieved by a retriever module 15 upon comparing at least one stored code snippet 14(a) and an inputted code snippet 16. In step S3, a code summary is generated that is relevant to the inputted code snippet 16 by a LLM 18 , based on the comparison between the at least one stored code snippet 14(a) and the inputted code snippet 16.

[0011]The above method is explained in detail. The present invention addresses the broader goal of legacy code understanding by summarizing code snippets. The input given here, is a line of code / method. The output will be the summary that explains what the code is performing. The present disclosed methodology employed in the control unit 10 leverages the documented code from the historical/legacy code data base 12. The present invention employs the Large Language Model (LLM) 18 to achieve the task. The LLMs 18have advanced the state-of-the-art in many Generative AI (artificial intelligence) task, including several use-cases in the software engineering sector.

[0012]The LLM 18 in an unsupervised way of using prompting and uses any one of the two kinds of prompting techniques like a zero-shot technique, wherein the code snippet 16 is directly given and whose summary will be generated. Another kind of prompting technique is a few-shot technique, wherein some code snippets (14(a)) in the form of is given along with the input code snippet 16, so that the LLM 18 can better understand the task & generate coherent responses.

9. A method for providing an efficient code summary by a control unit (10), said method comprising :
- storing multiple code snippets (14) and their corresponding code summaries in a code data base (12);
- retrieving a related information by a retriever module (15) upon comparing at least one stored code snippet (14(a)) and an inputted code snippet (16);
- generating a code summary that is relevant to said inputted code snippet (16) by a large language module ( LLM ) (18), based on the comparison between the at least one stored code snippet (14(a)) and said inputted code snippet (16).