Neural-Code-Search-Evaluation-Dataset

Neural-Code-Search-Evaluation-Dataset presents an evaluation dataset consisting of natural language query and code snippet pairs, with the hope that future work in this area can use this dataset as a common benchmark. We also provide the results of two code search models (NCS, UNIF) from recent work.

The full paper is available at Neural Code Search Evaluation Dataset.

Dataset contents

All the dataset contents are in the data directory.

GitHub Repositories

The most popular Android repositories on GitHub (ranked by the number of stars) is used to create the search corpus. For each repository that we indexed, we provide the link, specific to the commit that was used. In total, there are 24,549 repositories. This is located in data/android_repositories_download_links.txt. We also provide a Python script (download.py) that downloads these GitHub repositories.

Example:

https://github.com/00-00-00/ably-chat/archive/9bb2e36acc24f1cd684ef5d1b98d837055ba9cc8.zip
https://github.com/01sadra/Detoxiom/archive/c3fffd36989b0cd93bd09cbaa35123b9d605f989.zip
https://github.com/0411ameya/MPG_update/archive/27ac5531ca2c2f123e0cb854ebcb4d0441e2bc98.zip
...

Search Corpus

The search corpus is indexed using all method bodies parsed from the 24,549 GitHub repositories. In total, there are 4,716,814 methods in this corpus. The code search model will find relevant code snippets (i.e. method bodies) from this corpus given a natural language query. In this data release, we will provide the following information for each method in the corpus:

id: Each method in the corpus has a unique numeric identifier. This ID number will also be referenced in our evaluation dataset.
filepath: The file path is in the format of :owner/:repo/relative-file-path-to-the-repo
method_name
start_line: Starting line number of the method in the file.
end_line: Ending line number of the method in the file.
url: GitHub link to the method body with commit ID and line numbers encoded.

This is located in two parts (due to GitHub file size constraints): data/search_corpus_1.tar.gz and data/search_corpus_2.tar.gz.

Example:

{
  "id": 4716813,
  "filepath": "Mindgames/VideoStreamServer/playersdk/src/main/java/com/kaltura/playersdk/PlayerViewController.java",
  "method_name": "notifyKPlayerEvent",
  "start_line": 506,
  "end_line": 566,
  "url":  "https://github.com/Mindgames/VideoStreamServer/blob/b7c73d2bcd296b3a24f83cf67d6a5998c7a1af6b/playersdk/src/main/java/com/kaltura/playersdk/PlayerViewController.java\#L506-L566"
}

Evaluation Dataset

The evaluation dataset is composed of 287 Stack Overflow question and answer pairs, for which we release the following information:

stackoverflow_id: Stack Overflow post ID.
question: Title fo the Stack Overflow post.
question_url: URL of the Stack Overflow post.
answer: Code snippet answer to the question.

The questions were collected from a data dump publicly relased by Stack Exchange here. This is located in data/287_android_questions.json.

Example:

{
  "stackoverflow_id": 1109022,
  "question": "Close/hide the Android Soft Keyboard",
  "question_url": "https://stackoverflow.com/questions/1109022/close-hide-the-android-soft-keyboard",
  "question_author": "Vidar Vestnes",
  "question_author_url": "https://stackoverflow.com/users/133858",
  "answer": "// Check if no view has focus:\nView 
        view = this.getCurrentFocus(); \nif view != null) {InputMethodManager 
        imm = (InputMethodManager) getSystemService(Context.INPUT_METHOD_SERVICE);       
        imm.hideSoftInputFromWindow(view.getWindowToken(), 0);}",
  "answer_url": "https://stackoverflow.com/a/1109108",
  "answer_author": "Reto Meier",
  "answer_author_url": "https://stackoverflow.com/users/822",
  "examples": [1841045, 1800067, 1271795],
  "examples_url": [
    "https://github.com/alextselegidis/easyappointmentsandroid-client/blob/39f1e8...",
    "https://github.com/zelloptt/zello-android-clientsdk/blob/87b45b6...",
    "https://github.com/systers/conference-android/blob/a67982abf54e0...",
  ]
}

NCS / UNIF Score Sheet

We provide the evaluation results for two code search models of our creation, each with two variations:

NCS: an unsupervised model which uses word embedding derived directly from the search corpus.
NCS_postrank: an extension of the base NCS model that performs a post-pass ranking, as explained in here.
UNIF_android, UNIF_{stackoverflow}: supervised extensions of the NCS model that uses a bag-of-words-based neural network with attention. The supervision is learned using GitHub-Android-Train and StackOverflow-AndroidTrain datasets, respectively, as described here UNIF.

We provide the rank of the first correct answer (FRank) for each question in our evaluation dataset. The score sheet is saved in a comma-delimited csv file in data/score_sheet.csv.

Example:

No.,StackOverflow ID,NCS FRank,NCS_postrank FRank, UNIF_android FRank,UNIF_stackoverflow FRank
1,1109022,NF,1,1,1
2,4616095,17,1,31,19
3,3004515,2,1,5,2
4,1560788,1,4,5,1
5,3423754,5,1,22,10
6,1397361,NF,3,2,1

License

Neural-Code-Search-Evaluation-Dataset is CC-BY-NC 4.0 (Attr Non-Commercial Inter.) (e.g., FAIR) licensed, as found in the LICENSE file.

Tree sitter based method extractor for one line Java file, in case for some invalid code sample.

Guys, I've found that some Java file in search corpus only has 1-1 form permalink in original github link e.g: https://github.com/cymcsg/UltimateAndroid/blob/678afdda49d1e7c91a36830946a85e0fda541971/deprecated/UltimateAndroidGradle/ultimateandroiduianimation/src/main/java/com/marshalchen/common/uimodule/imageprocessing/FastImageProcessingPipeline.java#L1-L1.

This will cause url duplicate in search corpus and method snippets missing in eval set.

I've found a simple solution for these issue which on the basis of tree-sitter, here's the code:

from tree_sitter import Language, Parser
def get_java_methods(code:str):
    ''' 
         code: java code string
    '''
    def dfs(node, method_list):
        node_childs = node.children
        if node_childs == []: return
        for i in range(len(node_childs)):
            if node_childs[i].type == 'class_declaration':
                dfs(node_childs[i], method_list)
            elif node_childs[i].type == 'class_body':
                method_nodes = node_childs[i].children
                for j in method_nodes:
                    if j.type == 'method_declaration':
                        method_list.append(j)
                return

    JAVA_LANGUAGE = Language('Your tree sitter pretrained model file path/java-language.so', 'java')
    parser = Parser()
    parser.set_language(JAVA_LANGUAGE)
    tree = parser.parse(bytes(code, 'utf8'))
    method_nodes=[]
    dfs(tree.root_node,method_nodes)

    code_list = code.split('\n')
    methods_list = []
    for n in method_nodes:
            method = '\n'.join(code_list[n.start_point[0]:n.end_point[0]+1])
            methods_list.append(method)
    return methods_list,method_nodes

While the method_nodes return value will carry original span(the starting line number and end line number) of java method.
So you can extract the complete method body.

Inspired by blog:https://blog.csdn.net/weixin_43646592/article/details/120639861 great thanks.

facebookresearch / neural-code-search-evaluation-dataset Goto Github PK