CommitBERT.pdf

CommitBERT: Commit Message Generation
Using Pre-Trained Programming Language Model
Tae-Hwan Jung
Kyung Hee University
nlkey2022@gmail.com
Abstract
In version control using Git, the commit mes-
sage is a document that summarizes source
code changes in natural language. A good
commit message clearly shows the source
code changes, so this enhances collaboration
between developers. To write a good commit
message, the message should briefly summa-
rize the source code changes, which takes a lot
of time and effort. Therefore, a lot of research
has been studied to automatically generate a
commit message when a code modification is
given. However, in most of the studies so far,
there was no curated dataset for code modifica-
tions (additions and deletions) and correspond-
ing commit messages in various programming
languages. The model also had difficulty learn-
ing the contextual representation between code
modification and natural language.
To solve these problems, we propose the fol-
lowing two methods: (1) We collect code mod-
ification and corresponding commit messages
in Github for six languages (Python, PHP, Go,
Java, JavaScript, and Ruby) and release a well-
organized 345K pair dataset. (2) In order to re-
solve the large gap in contextual representation
between programming language (PL) and nat-
ural language (NL), we use CodeBERT (Feng
et al., 2020), a pre-trained language model
(PLM) for programming code, as an initial
model. Using two methods leads to successful
results in the commit message generation task.
Also, this is the first research attempt in fine-
tuning commit generation using various pro-
gramming languages and code PLM. Training
code, dataset, and pretrained weights are avail-
able at https://github.com/graykode/commit-
autosuggestions.
1 Introduction
Commit message is the smallest unit that summa-
rizes source code changes in natural language. Fig-
ure 1 shows the git diff representing code modifi-
cation and the corresponding commit message. A
Message : fix deprecated ref to tokenizer.max len
Figure 1: The figure above shows an example of com-
mit message and git diff in Github. In the Git process,
git diff uses unified format (unidiff 2
): A line marked
in red or green means a modified line, and green high-
lights in ’+’ lines are the added code, whereas red high-
lights in ’-’ lines are the deleted code.
good commit message allows developers to visual-
ize the commit history at a glance, so many teams
try to do high quality commits by creating rules
for commit messages. For example, Conventional
Commits 1 is one of the commit rules to use a verb
of a specified type for the first word like ’Add’ or
’Fix’ and limit the length of the character. It is very
tricky to follow all these rules and write a good
quality commit message, so many developers ig-
nore it due to lack of time and motivation. So it
would be very efficient if the commit message is
automatically written when a code modification is
given.
Similar to text summarization, many studies
have been conducted by taking code modification
X = (x1, ..., xn) as encoder input and commit
message Y = (y1, ..., ym) as decoder input based
on the NMT (Neural machine translation) model.
(Jiang et al., 2017; Loyola et al., 2017; van Hal
et al., 2019) However, taking the code modifica-
tion without distinguishing between the added and
1
https://conventionalcommits.org
2
https://en.wikipedia.org/wiki/Diff
arXiv:2105.14242v1
[cs.CL]
29
May
2021

the deleted part as model input makes it difficult
to understand the context of modification in the
NMT model. In addition, previous studies tend to
train from scratch when training a model, but this
method does not show good performance because
it creates a large gap in the contextual represen-
tation between programming language (PL) and
natural language (NL). To overcome the problems
in previous studies and train a better commit mes-
sage generation model, our approach follows two
stages:
(1) Collecting and processing data with the pair
of the added and deleted parts of the code X =
((add1, del1), ..., (addn, deln)). To take this pair
dataset into the Transformer-based NMT model
(Vaswani et al., 2017), we use the BERT (Devlin
et al., 2018) fine-tuning method about two sentence-
pair consist of added and deleted parts. This shows
a better BLEU-4 score (Papineni et al., 2002) than
previous works using raw git diff. Similar to Code-
SearchNet (Husain et al., 2019), our data is also
collected for six languages (Python, PHP, Go, Java,
JavaScript, and Ruby) from Github to show good
performance in various languages. We finally re-
leased 345K code modification and commit mes-
sage pair data.
(2) To solve a large gap about contextual repre-
sentation between programming language (PL) and
natural language (NL), we use CodeBERT (Feng
et al., 2020), a language model well-trained in the
code domain as the initial weight. Using Code-
BERT as the initial weight shows that the BLEU-4
score for commit message generation is better than
when using random initialization and RoBERTa
(Liu et al., 2019). Additionally, when we pre-train
the Code-to-NL task to document the source code
in CodeSearchNet and use the initial weight of
commit generation, the contextual representation
between PL and NL is further reduced.
2 Related Work
Commit message generation has been studied in
various ways. Jiang and McMillan (2017) collect
2M commits from the Mauczka et al. (2015) and
top 1K Java projects in Github. Among the commit
messages, only those that keep the format of ”Verb
+ Object” are filtered, grouped into verb types with
similar characteristics, and then the classification
model is trained with the naive Bayes classifier.
Jiang et al. (2017) use the commit data collected
by Jiang and McMillan (2017) to generate the
commit message using an attention-based RNN
encoder-decoder NMT model. They filter again in
a ”verb/direct-object pattern” from 2M data and
finally used the 26K commit message data. Loy-
ola et al. (2017) uses an NMT model similar to
Jiang et al. (2017), but uses git diff and commit
pairs collected from 1∼3 repositories of Python,
Java, JavaScript, and C++ as training data. Liu et al.
(2018) propose a retrieval model using Jiang et al.
(2017)’s 26K commit as training data. Code modi-
fication is represented by bags of words vector, and
the message with the highest cosine similarity is
retrieved. Xu et al. (2019) collect only ’.java’ file
format from Jiang and McMillan (2017) and use
509K dataset as training data for NMT. Also, to mit-
igate the problem of Out-of-Vocabulary (OOV) of
code domain input, they use generation distribution
or copying distribution similar to pointer-generator
networks (See et al., 2017). van Hal et al. (2019)
also argues that the Jiang and McMillan (2017)
entire data is noise and proposes a pre-processing
method that filters the better commit messages.
Liu et al. (2020) argue that it is challenging to
represent the information required for source code
input in the NMT model with a fixed-length. In
order to alleviate this, it is suggested that only the
added and deleted parts of the code modification
be abbreviated as abstract syntax tree (AST) and
applied to the Bi-LSTM model.
Nieb et al. presented a large gap between the
contextual representation between the source code
and the natural language when generating com-
mit messages. Previous studies have used RNN or
LSTM model, they use the transformer model, and
similarly to other studies, they use Liu et al. (2018)
as the training data. To reduce this gap, they try to
reduce the two-loss that predict the next code line
(Explicit Code Changes) and the randomly masked
word in the binary file.
3 Background
3.1 Git Process
Git is a version management system that manages
version history and helps collaboration efficiently.
Git tracks all files in the project in the Working di-
rectory, Staging area, and Repository. The working
directory shows the files in their current state. Af-
ter modifying the file, developers move the files to
the staging area using the add command to record
the modified contents and write a commit mes-
sage through the commit command. Therefore,

the commit message may contain two or more file
changes.
3.2 Text Summarization based on
Encoder-Decoder Model
With the advent of sequence to sequence learning
(Seq2Seq) (Sutskever et al., 2014), various tasks
between the source and the target domain are being
solved. Text summarization is one of these tasks,
showing good performance through the Seq2Seq
model with a more advanced encoder and decoder.
The encoder and decoder models are trained by
maximizing the conditional log-likelihood below
based on source input X = (x1, ..., xn) and target
input Y = (y1, ..., ym).
p(Y |X; θ) = log
T
X
t=0
p(yt|y<t, X; θ)
where T is the length of the target input, y0 is
the start token, yT is the end token and θ is the
parameter of the model.
In the Transformer (Vaswani et al., 2017) model,
the source input is vectorized into a hidden state
through self-attention as the number of encoder
layers. After that, the target input also learns the
generation distribution through self-attention and
attention to the hidden state of the encoder. It shows
better summarization results than the existing RNN-
based model (Nallapati et al., 2016).
To improve performance, most machine transla-
tions use beam search. It keeps the search area by
K most likely tokens at each step and searches the
next step to generate better text. Generation stops
when the predicted yt is an end token or reaches
the maximum target length.
3.3 CodeSearchNet
CodeSearchNet (Husain et al., 2019) is a dataset to
search code function snippets in natural language.
It is a paired dataset of code function snippets
for six programming languages (Python, PHP, Go,
Java, JavaScript and Ruby) and a docstring summa-
rizing these functions in natural language. A total
of 6M pair datasets is collected from projects with
a re-distribution license. Using the CodeSearch-
Net corpus, retrieval of the code corresponding to
the query composed of natural language can be
resolved. Also, it is possible to resolve the prob-
lem of documenting the code by summarizing it in
natural language (Code-to-NL).
3.4 CodeBERT
Recent NLP studies have shown state-of-the-art
in various tasks through transfer learning consist-
ing of pre-training and fine-tuning (Peters et al.,
2018). In particular, BERT (Devlin et al., 2018) is
a pre-trained language model by predicting masked
words from randomly masked sequence input and
uses only encoder based on Transformer (Vaswani
et al., 2017). It shows good perfomances in various
datasets and is now extending out of the natural
language domain to the voice, video, and code do-
mains.
CodeBERT is a pre-trained language model in
the code domain to learn the relationship between
programming language (PL) and natural language
(NL). In order to learn the representation between
different domains, they refer to the learning method
of ELECTRA (Clark et al., 2020) which is consists
of Generator-Discriminator. NL and Code Genera-
tor predict words from code tokens and comment
tokens masked at a specific rate. Finally, NL-Code
Discriminator is CodeBERT after trained through
binary classification that predicts whether it is re-
placed or original.
CodeBERT shows good results for all tasks in
the code domain. Specially, it shows a higher score
than other pre-trained models in the code to natu-
ral language(Code-to-NL) and code retrieval task
from NL using CodeSearchNet Corpus. In addition,
CodeBERT uses the Byte Pair Encoding (BPE) to-
kenizer (Sennrich et al., 2015) used in RoBERTa,
and does not generate unk tokens in code domain
input.
4 Dataset
We collect a 345K code modification dataset and
commit message pairs from 52K repositories of six
programming languages (Python, PHP, Go, Java,
JavaScript, and Ruby) on Github. When using raw
git diff as model input, it is difficult to distinguish
between added and deleted parts, so unlike Jiang
and McMillan (2017), our dataset focuses only on
the added and deleted lines in git diff. The de-
tailed data collection and pre-processing method
are shown as a pseudo-code in Algorithm 1:
To collect only the code that is a re-distributable
license, we have listed the Github repository name
in the CodeSearchNet dataset. After that, all the
repositories are cloned through multi-threading.
Detailed descriptions of functions that collect the
commit hashes in a repository and the code modifi-

Algorithm 1 Code modification parser from the
list of repositories.
1: procedure REPOPARSER(Repos)
2: for Repo in Repos do
3: commits = get commits(Repo)
4: for commit in commits do
5: mods = get modifications(commit)
6: for mod in mods do
7: if filtering(mod, commit) then
8: break
9: end if
10: Save (mod.add, mod.del) to dataset.
11: end for
12: end for
13: end for
14: end procedure
0K 20K 40K 60K 80K 100K 120K 140K
Frequency
add
fix
use
update
remove
make
change
move
allow
improve
implement
create
upgrade
Verb
Types
37.7%
22.1%
6.9%
6.2%
6.2%
5.1%
4.6%
3.6%
3.2%
1.8%
1.7%
0.8%
0.2%
Figure 2: Commit message verb type and frequency
statistics. Only ’upgrade’ is not included in the high
frequency, but is included in a similar way to ’update’.
This refers to the verb group in Jiang and McMillan
(2017).
cations in a commit hash are as follows:
• get_commits is a function that gets the
commit history from the repository. At this
time, the commits of the master branch are
filtered, excluding merge commits. Commits
with code modifications corresponding to 6
the program language(.py, .php, .js, .java, .go,
.ruby) extensions are collected. To implement
this, we use the open-source pydriller (Spadini
et al., 2018).
• get_modifications is a function that
gets the line modified in the commit. Through
this function, it is possible to collect only the
added or deleted parts, not all git diffs.
While collecting the pair dataset, we find that
the relationship between some code modifications
Number of Pair Dataset
Train Validation Test
Number of
Repositories
Python 81517 10318 10258 12361
PHP 64458 8079 8100 16143
JavaScript 50561 6296 6252 11294
Ruby 29842 3772 3680 4581
Java 28069 3549 3552 4123
Go 21945 2699 2812 3960
Total : 345759 52462
Table 1: Dataset Statistics for each language collected
from 52K repositories of six programming languages.
and the corresponding commit message is obscure
and very abstract. Also, we check that some code
modification or commit message is a meaning-
less dummy file. To filter these, we create the
filtering function and the rules as follows.
1. To collect commit messages with various for-
mat distributions, we limit the collection of
up to 50 commits in one repository.
2. We filter commits whose number of files
changed is one or two per commit message.
3. Commit message with issue number is re-
moved because detailed information is abbre-
viated.
4. Similar to Jiang and McMillan (2017), the
non-English commit messages are removed.
5. Since some commit messages are very long,
the first line is fetched.
6. If the token of code through tree-sitter3, a
parser generator tool, exceeds 32 characters, it
is excluded. This removes unnecessary things
like changes to binary files in code diff.
7. By referring to the Jiang and McMillan (2017)
and Conventional Commits(§ 1) rules, the
commit message that begins with a verb is
collected. We use spaCy4 for Pos tagging.
8. We filter commit messages with 13 verb types,
which are the most frequent. Figure 2 shows
the collected verb types and their ratio for the
entire dataset.
As a result, we collect 345K code modification
and commit message pair datasets from 52K Github
repositories and split commit data into 80-10-10
train/validation/test sets. This results are shown in
Table 1.
3
https://tree-sitter.github.io/tree-sitter
4
https://spacy.io

5 CommitBERT
We propose the idea of generating a commit
message through the CodeBERT model with the
our dataset (§ 4). To this end, this section de-
scribes how to feed inputs code modification (X =
((add1, del1), ..., (addn, deln))) and commit mes-
sage (Y = (msg1, ..., msgn)) to CodeBERT and
how to use pre-trained weights more efficiently to
reduce the gap in contextual representation between
programming language (PL) and natural language
(NL).
5.1 CodeBERT for Commit Message
Generation
We feed the code modification to the encoder and
a commit message to the decoder input by follow-
ing the NMT model. Especially for code mod-
ification in the encoder, similar inputs are con-
catenated, and different types of inputs are sep-
arated by a sentence separator (sep). Applying this
to our CommitBERT in the same way, added to-
kens (Add = (add1, ..., addn)) and deleted tokens
(Del = (del1, ..., deln)) of similar types are con-
nected to each other, and sentence separators are
inserted between them. Therefore, the conditional-
likelihood is as follows:
p(M|C; θ) = log
T
X
t=0
p(mt|m<t, C; θ),
m<t = (m0, m1, ..., mt−1)
C = concat([cls], Add, [sep], Del, [sep])
where M is commit message tokens, C is code
modification tokens and concat is list concatena-
tion function. [cls] and [sep] are speical tokens,
which are a start token and a sentence separator
token respectively. Other notions are the same as
Section 3.2.
Unlike previous works, all code modifications in
git diff are not used as input and only changed lines
in code modification are used. Since this removes
unnecessary inputs, it shows a significant perfor-
mance improvement in summarizing code modifi-
cations in natural language. Figure 3 shows how
the code modification is actually taken as model
input.
5.2 Initialize Pretrained Weights
To reduce the gap difference between two do-
mains(PL, NL), We use the pretrained CodeBERT
as the initial weight. Furthermore, we determine
that removing deleted tokens from our dataset (§ 4)
is similar to the Code-to-NL task in CodeSearchNet
(Section 3.3). Using this feature, we use the ini-
tial weight after training the Code-to-NL task with
CodeBERT as the initial weight. This method of
training shows better results than only using Code-
BERT weight in commit message generation.
6 Experiment
To verify the proposal in Section 5 in the commit
message generation task, we do two experiments.
(1) Compare the commit message generation re-
sults of using all code modifications as inputs and
using only the added or deleted lines as inputs. (2)
Ablation study several initial model weights to find
the weight with the smallest gap in contextual rep-
resentation between PL and NL.
6.1 Experiment Setup
Our implementation uses CodeXGLUE’s code-text
pipeline library 5. We use the same model archi-
tecture and experimental parameters for the two
experiments below. As a model architecture, the
encoder and decoder use 12 and 3 Transformer lay-
ers. We use 5e-5 as the learning rate and train on
one V100 GPU with a 32 batch size. We also use
256 as the maximum source input length and 128
as the target input length, 10 training epochs, and
10 as the beam size k.
6.2 Compare Model Input Type
To experiment generating a commit message ac-
cording to the input type, only 4135 data is col-
lected from data with code modification in the
‘.java’ files among 26K training data of Loyola et al.
(2017). Then we transform these 4135 data into
two types, respectively, and experiment with train-
ing data for RoBERTa and CodeBERT weights:
(a) entire code modification in git diff and (b) only
changed lines in code modification. Figure 3 shows
these two differences in detail.
Table 3 shows the BLEU-4 values when infer-
ence with the test set after training about these two
types. Both initial weights show worse results than
(b), even though type (a) takes a more extended
input to the model. This shows that lines other than
changed lines as input data disturb training when
generating the commit message.
5
https://github.com/microsoft/CodeXGLUE

(a) Code modification in git diff
Encoder
(CodeBERT)
. . .
def add
def (
subtract
[CLS] a , b ) [SEP]
: ( a , b ) [SEP]
:
. . .
Decoder
Fix function name
. . .
(b) CommitBERT input
Figure 3: Illustration of a code modification example in git diff (a) and method of taking it to the input of
CommitBERT (b). (b) shows that all code modification lines in (a) are not used, and only changed lines are as
input. So, in this example, code modification (a) includes return a - b, but not in the model input (b).
Metric Initial Weight Python PHP JavaScript Java Go Ruby
BLEU-4 (Test)
(a) Random 7.95 7.01 8.41 7.60 10.38 7.17
(b) RoBERTa 10.94 9.71 9.50 6.40 10.21 8.95
(c) CodeBERT 12.05 13.06 10.47 8.91 11.19 10.33
(d) CodeBERT + Code-to-NL 12.93 14.30 11.49 9.81 12.76 10.56
PPL (Dev)
(a) Random 144.60 138.39 195.98 275.84 257.29 207.67
(b) RoBERTa 76.02 81.97 103.48 164.32 122.70 104.68
(c) CodeBERT 68.18 63.90 94.62 116.50 109.43 91.50
(d) CodeBERT + Code-to-NL 49.29 47.89 75.53 77.80 64.43 82.82
Table 2: Commit message generation result for 4 initial weights. In (c), CodeBERT is used as the initial weight.
And (d) uses the weight trained on the Code-to-NL task in CodeSearchNet with CodeBERT as the initial weight.
As a result, it shows BLEU-4 for the test set after training and the best PPL for the validation set in the during
training.
Initial Weight Input Type BLEU-4
RoBERTa
(a) All code modification 10.91
(b) Only changed lines (Ours) 12.52
CodeBERT
(a) All code modification 11.77
(b) Only changed lines (Ours) 13.32
Table 3: The result of generating the commit message
for the input type after collecting 4135 data with only
source code change among the data of Loyola et al.
(2017). (a) uses entire git diff(unidiff) as input, and
(b) uses only the changed line according to Section 5.1
as input.
6.3 Ablation study on initial weight
We do an ablation study while changing the ini-
tial weight of the model for 345K datasets in
six programming languages collected in Section
4. As mentioned in 5.2, when the model weight
with high comprehension in the code domain is
used as the initial weight, it is assumed that the
large gap in contextual representation between PL
and NL would be greatly reduced. To prove this,
we train the commit message generation task for
four weights as initial model weights: Random,
RoBERTa6, CodeBERT7, and the weights trained
6
https://huggingface.co/roberta-base
7
https://huggingface.co/microsoft/codebert-base
on the Code-to-NL task(Section 3.3) with Code-
BERT. Except for this initial weight, all training
parameters are the same.
Table 2 shows BLEU-4 for the test set and PPL
for the dev set for each of the four weights after
training. As a result, using weights trained on the
Code-to-NL task with CodeBERT as the initial
weight shows the best results for test BLEU-4 and
dev PPL. It also shows good performance regard-
less of programming language.
7 Conclusion and Future Work
Our work presented a model summarizing code
modifications to solve the difficulty of humans
manually writing commit messages. To this end,
this paper proposed a method of collecting data, a
method of taking it to a model, and a method of
improving performance. As a result, it showed a
successful result in generating a commit message
using our proposed methods. Consequently, our
work can help developers who have difficulty writ-
ing commit messages even with the application.
Although it is possible to generate a high-quality
commit message with a pre-trained model, future
studies to understand the code syntax structure re-
main in our work. As a solution to this, Com-
mitBERT should be converted to AST (Abstract

Language Reference / Generated
Python
added figsize to plot methods
added figure size to plot weights
PHP
added default value to fieldtype
Added default values of the fieldtype
JavaScript
Fix missing = in delete uri
Fixed an issue with delete
Java
Fixed the parsing of orders without a cid
Fix bug in exchange order
Go
Use ioutil . Discard for benchmark
Use ioutil . Discard for logging
Ruby
fixing schema validation issues with
CCR export
fixing validation of ccr export
Table 4: The result of generating the commit message
for six languages (Python, PHP, Go, Java, JavaScript,
and Ruby) and the corresponding reference. We used
the (d) model of Table 2.
Syntax Tree) before code modification is taken into
the encoder like (Liu et al., 2020).
Acknowledgments
The author would like to thank Gyuwan Kim,
Dongjun Lee, Mansu Kim and the anonymous re-
viewers for their thoughtful paper review.
References
Kevin Clark, Minh-Thang Luong, Quoc V Le, and
Christopher D Manning. 2020. Electra: Pre-training
text encoders as discriminators rather than genera-
tors. arXiv preprint arXiv:2003.10555.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. 2018. Bert: Pre-training of deep
bidirectional transformers for language understand-
ing. arXiv preprint arXiv:1810.04805.
Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xi-
aocheng Feng, Ming Gong, Linjun Shou, Bing Qin,
Ting Liu, Daxin Jiang, et al. 2020. Codebert: A
pre-trained model for programming and natural lan-
guages. arXiv preprint arXiv:2002.08155.
SRP van Hal, Mathieu Post, and Kasper Wendel. 2019.
Generating commit messages from git diffs. arXiv
preprint arXiv:1911.11690.
Hamel Husain, Ho-Hsiang Wu, Tiferet Gazit, Miltiadis
Allamanis, and Marc Brockschmidt. 2019. Code-
searchnet challenge: Evaluating the state of seman-
tic code search. arXiv preprint arXiv:1909.09436.
Siyuan Jiang, Ameer Armaly, and Collin McMillan.
2017. Automatically generating commit messages
from diffs using neural machine translation. In 2017
32nd IEEE/ACM International Conference on Auto-
mated Software Engineering (ASE), pages 135–146.
IEEE.
Siyuan Jiang and Collin McMillan. 2017. Towards
automatic generation of short summaries of com-
mits. In 2017 IEEE/ACM 25th International Con-
ference on Program Comprehension (ICPC), pages
320–323. IEEE.
Shangqing Liu, Cuiyun Gao, Sen Chen, Nie Lun Yiu,
and Yang Liu. 2020. Atom: Commit message gener-
ation based on abstract syntax tree and hybrid rank-
ing. IEEE Transactions on Software Engineering.
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man-
dar Joshi, Danqi Chen, Omer Levy, Mike Lewis,
Luke Zettlemoyer, and Veselin Stoyanov. 2019.
Roberta: A robustly optimized bert pretraining ap-
proach. arXiv preprint arXiv:1907.11692.
Zhongxin Liu, Xin Xia, Ahmed E Hassan, David Lo,
Zhenchang Xing, and Xinyu Wang. 2018. Neural-
machine-translation-based commit message genera-
tion: how far are we? In Proceedings of the 33rd
ACM/IEEE International Conference on Automated
Software Engineering, pages 373–384.
Pablo Loyola, Edison Marrese-Taylor, and Yutaka Mat-
suo. 2017. A neural architecture for generating natu-
ral language descriptions from source code changes.
arXiv preprint arXiv:1704.04856.
Andreas Mauczka, Florian Brosch, Christian Schanes,
and Thomas Grechenig. 2015. Dataset of developer-
labeled commit messages. In 2015 IEEE/ACM 12th
Working Conference on Mining Software Reposito-
ries, pages 490–493. IEEE.
Ramesh Nallapati, Bowen Zhou, Caglar Gulcehre,
Bing Xiang, et al. 2016. Abstractive text summariza-
tion using sequence-to-sequence rnns and beyond.
Lun Yiu Nieb, Cuiyun Gaoa, Zhicong Zhongc, Wai
Lamb, Yang Liud, and Zenglin Xua. Coregen: Con-
textualized code representation learning for commit
message generation.
Kishore Papineni, Salim Roukos, Todd Ward, and Wei-
Jing Zhu. 2002. Bleu: a method for automatic eval-
uation of machine translation. In Proceedings of the
40th annual meeting of the Association for Compu-
tational Linguistics, pages 311–318.
Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt
Gardner, Christopher Clark, Kenton Lee, and Luke
Zettlemoyer. 2018. Deep contextualized word repre-
sentations. arXiv preprint arXiv:1802.05365.
Abigail See, Peter J Liu, and Christopher D Man-
ning. 2017. Get to the point: Summarization
with pointer-generator networks. arXiv preprint
arXiv:1704.04368.
Rico Sennrich, Barry Haddow, and Alexandra Birch.
2015. Neural machine translation of rare words with
subword units. arXiv preprint arXiv:1508.07909.

Davide Spadini, Maurı́cio Aniche, and Alberto Bac-
chelli. 2018. Pydriller: Python framework for min-
ing software repositories. In Proceedings of the
2018 26th ACM Joint Meeting on European Soft-
ware Engineering Conference and Symposium on
the Foundations of Software Engineering, pages
908–911.
Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014.
Sequence to sequence learning with neural networks.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob
Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz
Kaiser, and Illia Polosukhin. 2017. Attention is all
you need. arXiv preprint arXiv:1706.03762.
Shengbin Xu, Yuan Yao, Feng Xu, Tianxiao Gu, Hang-
hang Tong, and Jian Lu. 2019. Commit message
generation for source code changes. In IJCAI.

CommitBERT.pdf

More Related Content

Similar to CommitBERT.pdf (20)

Recently uploaded (20)

CommitBERT.pdf