Systems and methods for auto-captioning repositories from source code

Patent metadata

Field	Value
Publication number	US-12039296-B2
Application number	US-202218054684-A
Country	US
Kind code	B2
Filing date	Nov 11, 2022
Priority date	Nov 11, 2022
Publication date	Jul 16, 2024
Grant date	Jul 16, 2024

Abstract

Systems and methods for auto-captioning repositories from source code are disclosed. A method for code repository embedding may include a computer program executed by an electronic device: (1) extracting docstring embeddings, code embeddings, and dependency embeddings from scripts in a repository; (2) applying a machine learning algorithm to each of the docstring embeddings, the code embeddings, and the dependency embeddings, wherein outputs of each of the machine learning algorithms comprises a vector; (3) concatenating the vectors into an output embedding vector; (4) weighting the output embedding vector using an attention mechanism, resulting in a repository representation comprising an abstract vector; (5) generating a plurality of tags for the repository representation or the output embedding vector representing the weights of the tags, using a trained neural network; (6) applying a tags-to-caption transformer to the tags or the output embedding vector, resulting in a caption; and (7) outputting the caption.

First claim

What is claimed is:

1. A method for code repository embedding, comprising:
extracting, by a computer program executed by an electronic device, docstring embeddings, code embeddings, and dependency embeddings from scripts in a repository;
applying, by the computer program, a machine learning algorithm to each of the docstring embeddings, the code embeddings, and the dependency embeddings, wherein outputs of each of the machine learning algorithms comprises a vector;
concatenating, by the computer program, the vectors into an output embedding vector;
weighting, by the computer program, the output embedding vector using an attention mechanism, resulting in a repository representation comprising an abstract vector;
generating, by the computer program, a plurality of tags for the repository representation or the output embedding vector representing the weights of the tags, using a trained neural network;
applying, by the computer program, a tags-to-caption transformer to the tags or the output embedding vector, resulting in a caption; and
outputting, by the computer program, the caption.

2. The method of claim 1 , further comprising:
identifying, by the computer program, the scripts in a code repositories database.

3. The method of claim 1 , wherein the computer program identifies the docstring embeddings, the code embeddings, and the dependency embeddings using a Bi-directional Encoder Representations from Transformers.

4. The method of claim 1 , wherein the machine learning algorithm comprises a gated recurrent unit (GRU) that is part of a Recurrent Neural Network (RNN).

5. The method of claim 1 , wherein the attention weights are trained from a curated dataset collected from a labelled dataset.

6. The method of claim 1 , wherein the repository representation comprises a numerical representation of the repository.

7. A method for code repository embedding, comprising:
extracting, by a computer program executed by an electronic device, docstring embeddings, code embeddings, and dependency embeddings from scripts in a repository;
applying, by the computer program, a machine learning algorithm to each of the docstring embeddings, the code embeddings, and the dependency embeddings, wherein outputs of each of the machine learning algorithms comprises a vector;
concatenating, by the computer program, the vectors into an output embedding vector;
weighting, by the computer program, the output embedding vector using an attention mechanism, resulting in a repository representation comprising an abstract vector;
selecting, by the computer program, a plurality of tags for the repository representation using a reinforcement learning agent;
applying, by the computer program, a tags-to-caption transformer to the tags or the output embedding vector, resulting in a caption; and
outputting, by the computer program, the caption.

8. The method of claim 7 , further comprising:
identifying, by the computer program, the scripts in a code repositories database.

9. The method of claim 7 , wherein the computer program identifies the docstring embeddings, the code embeddings, and the dependency embeddings using a Bi-directional Encoder Representations from Transformers.

10. The method of claim 7 , wherein the machine learning algorithm comprises a gated recurrent unit (GRU) that is part of a Recurrent Neural Network (RNN).

11. The method of claim 7 , wherein the reinforcement learning agent comprises a deep neural network.

12. The method of claim 7 , wherein the attention weights are trained from a curated dataset collected from a labelled dataset.

13. The method of claim 7 , wherein the repository representation comprises a numerical representation of the repository.

14. The method of claim 7 , further comprising:
reward shaping the reinforcement learning agent with feedback on the tag selection.

15. A non-transitory computer readable storage medium, including instructions stored thereon, which when read and executed by one or more computer processors, cause the one or more computer processors to perform steps comprising:
extracting docstring embeddings, code embeddings, and dependency embeddings from scripts in a repository;
applying a machine learning algorithm to each of the docstring embeddings, the code embeddings, and the dependency embeddings, wherein outputs of each of the machine learning algorithms comprises a vector;
concatenating the vectors into an output embedding vector; weighting the output embedding vector using an attention mechanism, resulting in a repository representation comprising an abstract vector;
generating a plurality of tags for the repository representation or the output embedding vector representing the weights of the tags, using a trained neural network;
applying a tags-to-caption transformer to the tags or the output embedding vector, resulting in a caption; and
outputting the caption.

16. The non-transitory computer readable storage medium of claim 15 , wherein the docstring embeddings, the code embeddings, and the dependency embeddings are identified using a Bi-directional Encoder Representations from Transformers.

17. The non-transitory computer readable storage medium of claim 15 , wherein the machine learning algorithm comprises a gated recurrent unit (GRU) that is part of a Recurrent Neural Network (RNN).

18. The non-transitory computer readable storage medium of claim 15 , wherein tags are generated using reinforcement learning agent to select the tags.

19. The non-transitory computer readable storage medium of claim 18 , wherein the reinforcement learning agent comprises a deep neural network.

20. The non-transitory computer readable storage medium of claim 18 , further including instructions stored thereon, which when read and executed by one or more computer processors, cause the one or more computer processors to reward shape the reinforcement learning agent with feedback on the tag selection.

Assignees

• Jpmorgan Chase Bank Na

Inventors

• Satsangi Yash
• Eloul Shaltiel
• Moran Sean

Classifications

• G06F8/311

  Functional or applicative languages; Rewrite languages · CPC title

• G06F9/45512

  Command shells · CPC title

• G06F8/71

  Version control (security arrangements therefor G06F21/57); Configuration management · CPC title

• G06F8/33 Primary

  Intelligent editors · CPC title

• G06F8/30 Primary

  Creation or generation of source code · CPC title