US 11,983,513 B2
Multi-lingual code generation with zero-shot inference
Colin Bruce Clement, Seattle, WA (US); Shuai Lu, Beijing (CN); Neelakantan Sundaresan, Bellevue, WA (US); Alexey Svyatkovskiy, Bellevue, WA (US); and Duyu Tang, Beijing (CN)
Assigned to MICROSOFT TECHNOLOGY LICENSING, LLC., Redmond, CA (US)
Filed by MICROSOFT TECHNOLOGY LICENSING, LLC., Redmond, WA (US)
Filed on May 24, 2023, as Appl. No. 18/201,535.
Application 18/201,535 is a continuation of application No. 17/978,627, filed on Nov. 1, 2022, granted, now 11,693,630.
Application 17/978,627 is a continuation of application No. 17/140,091, filed on Jan. 3, 2021, granted, now 11,513,774, issued on Nov. 29, 2022.
Prior Publication US 2023/0359443 A1, Nov. 9, 2023
This patent is subject to a terminal disclaimer.
Int. Cl. G06F 8/33 (2018.01); G06F 8/41 (2018.01); G06F 18/211 (2023.01); G06F 18/213 (2023.01); G06N 3/088 (2023.01)
CPC G06F 8/33 (2013.01) [G06F 8/44 (2013.01); G06F 18/211 (2023.01); G06F 18/213 (2023.01); G06N 3/088 (2013.01)] 18 Claims
OG exemplary drawing
 
1. A code completion system, comprising:
a processor; and
a memory, wherein the memory comprises instructions that when executed by the processor perform actions that:
obtain a partially-formed source code snippet of a source code program;
extract a file-level context and a local-level context of the partially-formed source code snippet;
search for a source code candidate to complete the partially-formed source code snippet, wherein the search generates a plurality of partial source code candidates at each of a plurality of time steps, wherein the search utilizes a deep learning model to generate a conditional probability distribution given the file-level context and the local-level context, wherein the search uses the conditional probability distribution at each time step to predict a next token of each partial source code candidate conditioned on previously-generated tokens of a respective partial source code candidate; and
upon completion of the search, output at least one partial source code candidate as the source code candidate to complete the partially-formed source code snippet.