Method and system for translation of codes based on semantic similarity

Patent metadata

Field	Value
Publication number	US-2023034984-A1
Application number	US-202217743511-A
Country	US
Kind code	A1
Filing date	May 13, 2022
Priority date	Jun 29, 2021
Publication date	Feb 2, 2023
Grant date	—

Abstract

Code translation is an evolving field and due to advancements in the infrastructure and compute power. The existing methods for code translation are time and effort intensive. A method and system for translation of codes based on the semantic similarity have been provided. A machine learning model is developed, that understands and encapsulates the semantics of the code in the source side and translates the semantic equivalent code which is more maintainable and efficient compared to one to one translation. The system is configured to group a plurality of statements present in the source code together into blocks of code and comprehend the semantics of the block. The system is also trained to understand syntactically different but semantically similar statements. While understanding the semantics of the block and translating, the unused/duplicate code etc. gets eliminated. The translated code is better architected and native to the target environment.

First claim

What is claimed is:

1 . A processor implemented method for translation of codes based on a semantic similarity, the method comprising:
providing, via a user interface, a source code for translation as an input; parsing, via one or more hardware processors, the source code using a parser; creating, via the one or more hardware processors, a program flow graph using the parsed source code, wherein the program flow graph is used to establish a set of relations between a plurality of statements present in the source code; creating, via the one or more hardware processors, a matrix of the established set of relations between the plurality of statements; splitting, via the one or more hardware processors, the source code into a plurality of blocks using the created matrix, wherein the statements in one block out of the plurality of blocks are closer to each other as compared to other statements irrespective of their physical presence in the source code; vectorizing, via the one or more hardware processors, the plurality of blocks to get a plurality of vectorized blocks; training, via the one or more hardware processors, a model to understand a semantic equivalence between the plurality of blocks of the source code irrespective of a manner in which the plurality of blocks is syntactically coded; identifying, via the one or more hardware processors, semantically similar statements out of the plurality of vectorized blocks using the trained model; selecting, via the user interface, a target language in which the source code needs to be translated; and translating, via the one or more hardware processors, the source code into the target language using a decoder based on the identified semantically similar statements, wherein the decoder is a pretrained machine learning model and configured to choose semantically same, but syntactically different statements.

2 . The method of claim 1 wherein the decoder is configured to ensure a translation of functionally equivalent code in the target language that is native to a target architecture.

3 . The method of claim 1 , wherein the set of relations comprises:
whether two statements amongst the plurality of statements in a program flow are part of the same path, whether there is a definition of a data element in a first statement which is used in the second statement, and whether the first statement contains a definition of the data element that influences the second statement.

4 . The method of claim 1 , further comprising eliminating one or more of unused and duplicate codes from the source code using the set of relations.

5 . The method of claim 1 , wherein the plurality of vectorized blocks are generated based on a plurality of properties of the source code comprising at least one of usage, value associated, scope of usage, data types, data structure, data size, and relations with other data elements.

6 . A system for translation of codes based on a semantic similarity, the system comprises:
a user interface for providing a source code for translation as an input and a target language in which the source code needs to be translated; one or more hardware processors; a memory in communication with the one or more hardware processors, wherein the one or more first hardware processors are configured to execute programmed instructions stored in the one or more first memories, to:
parse the source code using a parser;
create a program flow graph using the parsed source code, wherein the program flow graph is used to establish a set of relations between a plurality of statements present in the source code;
create a matrix of the established set of relations between the plurality of statements;
split the source code into a plurality of blocks using the created matrix, wherein the statements in one block out of the plurality of blocks are closer to each other as compared to other statements irrespective of their physical presence in the source code;
vectorize the plurality of blocks to get a plurality of vectorized blocks;
train a model to understand a semantic equivalence between the plurality of blocks of the source code irrespective of a manner in which the plurality of blocks is syntactically coded;
identify semantically similar statements out of the plurality of vectorized blocks using the trained model; and
translate the source code into the target language using a decoder based on the identified semantically similar statements, wherein the decoder is a pretrained machine learning model and configured to choose semantically same, but syntactically different statements.

7 . The system of claim 6 , wherein the decoder is configured to ensure a translation of functionally equivalent code in the target language that is native to a target architecture.

8 . The system of claim 6 , wherein the set of relations comprises:
whether two statements amongst the plurality of statements in a program flow are part of the same path, whether there is a definition of a data element in a first statement which is used in the second statement, and whether the first statement contains a definition of the data element that influences the second statement.

9 . The system of claim 6 further configured to eliminate one or more of unused and duplicate codes from the source code using the set of relations.

10 . The system of claim 6 , wherein the plurality of vectorized blocks are generated based on a plurality of properties of the source code comprising at least one of usage, value associated, scope of usage, data types, data structure, data size, and relations with other data elements.

11 . One or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause:
providing, via a user interface, a source code for translation as an input; parsing, the source code using a parser; creating, via the one or more hardware processors, a program flow graph using the parsed source code, wherein the program flow graph is used to establish a set of relations between a plurality of statements present in the source code; creating, via the one or more hardware processors, a matrix of the established set of relations between the plurality of statements; splitting, via the one or more hardware processors, the source code into a plurality of blocks using the created matrix, wherein the statements in one block out of the plurality of blocks are closer to each other as compared to other statements irrespective of their physical presence in the source code; vectorizing, via the one or more hardware processors, the plurality of blocks to get a plurality of vectorized blocks; training, via the one or more hardware processors, a model to understand a semantic equivalence between the plurality of blocks of the source code irrespective of a manner in which the plurality of blocks is syntactically coded; identifying, via the one or more hardware processors, semantically similar statements out of the plurality of vectorized blocks using the trained model; selecting, via the user interface, a target language in which the source code needs to be translated; and translating, via the one or more hardware processors, the source code into the target language using a decoder based on the identified semantically similar statements, wherein the decoder is a pretrained machine learning model and configured to choose semantically same, but syntactically different statements.

12 . The one or more non-transitory machine-readable information storage mediums of claim 11 wherein the decoder is configured to ensure a translation of functionally equi

Assignees

• Tata Consultancy Services Ltd

Inventors

• Ravindranath Yogananda
• Mehalingam Tamildurai
• Venkatanarayanan Balakrishnan
• Adithyan Reshinth Gnana
• Banerjee Shrayan
• Thuruvas Senthil Aditya

Classifications

• G06F8/433

  Dependency analysis; Data or control flow analysis · CPC title

• G06F8/51 Primary

  Source to source · CPC title

• G06F8/427

  Parsing · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/09

  Supervised learning · CPC title