Source Code fastText Embeddings

fastText pretrained models for semantic representations of source code in Java, Python, PHP, C, C++ and C#.

Model files

The pretrained models are stored in .bin files (of total approximate size 13 GB) which can be accessed at this link: https://doi.org/10.5281/zenodo.2558730

Instructions on how to use the models

Prerequisites

To load a model you will need to first build the fastText library following the instructions: https://fasttext.cc/docs/en/support.html.

Loading and querying the models

The official fastText documentation provides information on the library's functionalities and examples on how you can use pretrained models.

Semantic similarity queries

Simple semantic similarity queries can be performed through the nearest neighbor nn functionality of fastText. Examples on how to load and query the pretrained Java model for nearest words.

../fastText/fasttext nn java-ftskip-dim100-ws5.bin

Find nearest words to:

"toggleFullScreen"

Query word? toggleFullScreen
doToggleFullscreen 0.999999
FullScreen 0.999998
isNotMoveInScreen 0.999998
sFullScreen 0.999998
OSXFullScreen 0.999998
ToggleFullScreen 0.999998
toggleFullscreenAlbumArt 0.999998
fullScreen 0.999998
LoseScreen 0.999998
LooseScreen 0.999997

"BufferedInputStream"

Query word? BufferedInputStream
MyBufferedInputStream 1
TBufferedInputStream 1
BufferedFileInputStream 0.999999
maybeBufferInputStream 0.999999
LoggerBufferedInputStream 0.999999
BufferedBEROctetStream 0.999999
BufferedDataInputStream 0.999999
InputStreamBuffered 0.999999
byteBufferInputStream 0.999999
DirectBufferInputStream 0.999998

"writeLine"

Query word? writeLine
writeMine 0.999999
writeAsLine 0.999998
writeWChar 0.999998
writeLineBreak 0.999998
writeLocks 0.999998
writeTermTo 0.999998
writeABook 0.999998
writeRowStart 0.999998
writeElided 0.999998
writeRowEnd 0.999998

Analogy queries

Analogies may drawn through the information embedded in word vectors. This can be done with the analogies functionality of fastText. For example, in a general knowledge representation model the analogy "Berlin is to Germany what Paris to France" will be retrieved when given the relevant word triplets in this format: Query triplet (A - B + C)? Berlin Germany France

Example of an analogy query in the pretrained Java model

../fastText/fasttext analogies java-ftskip-dim100-ws5
Query triplet (A - B + C)? OutputStreamReader OutputStream InputStream
inputStreamReader 0.999999
InputStreamCSVReader 0.999998
inputStreamName 0.999998
BitStreamReader 0.999997
SlimStreamReader 0.999997
objectStreamReader 0.999997
flStreamReader 0.999997
inStreamReader 0.999997
getDecompressingInputStream 0.999997
resetStream 0.999997