0wn463 / nlp-issue-labeller

Hi 👋, I'm 0WN463

• ⚡ Fun fact: I use Arch btw
• I converted my notes from my academic days into interactive notebooks for future undergrads and people who're curious to learn, under project Wellspring

Other me's

https://glitch.com/@0WN3D

Languages and Tools:

 

nlp-issue-labeller's Issues

Dockerize our application

Check for issues with multiple tags in dataset

If we find issues that belongs to 2 different tag, do we keep both or drop both?
I'd vote for drop both since it's quite a pain to account for them down the line

Scraped data analysis

Analysis summary (using jupyter notebook)

Overall
-------
Total number of tensorflow issues:19207
Total number of rust issues:18022
Total number of kubernetes issues:18955

Tensorflow issue analysis
-------------------------
Number of feature issues:2461
Number of bug issues:15098
Number of doc issues:1648

Rust issue analysis
-------------------
Number of feature issues:10163
Number of bug issues:7049
Number of doc issues:766

Kubernetes issue analysis
-------------------------
Number of feature issues:4302
Number of bug issues:11695
Number of doc issues:968

Total issue analysis
-------------------------
Number of feature issues:16926
Number of bug issues:33842
Number of doc issues:3382

German repo issue analysis
--------------------------
Total number of corona widget issues:192
Total number of openWB issues:145

French repo issue analysis
--------------------------
Total number of DVF-app issues:104
Total number of Grafikart issues:313
Total number of azure docs issues:247
Total number of bcdlibre issues:36

Readme line 27 needs to be more specific

Dockerize application

Sequence classifier returns empty array

Just testing stuff

Function `has_code_block` does not return expected output

Function `has_code_block` does not return expected output

this is a nonsense issue

i don't know what this issue is

Class standardisation policy

We will classify labels of all scraped issues into our standard set of classes.
Namely, {feature, bug, doc}

Tensorflow:

Class	Labels
feature	type:feature
bug	type:bug, type:build/install, type:performance, type:support
doc	type:docs-feature, type:docs-bug

Rust:

Class	Labels
feature	C-feature-request, C-feature-accepted, C-enhancement
bug	C-bug
doc	T-doc

Kubernetes:

Class	Labels
feature	kind/feature, kind/api-change
bug	kind/bug, kind/failing-test
doc	kind/documentation

Refactor remove_url to replace_url

Update Report

Implement Autolabeller bot

what the heck is this issue?

Finish report

Documentation for codebase

Dockerize our application

Tokenisation

Ideas to tokenise both issue title & description text

text	regex
title	white space
description	white space, \r, \n, \t

For description text, we might also want to skip code blocks and reference links. They can be vectorised separately for information extraction.

Consultation slides & Poster

Consultation Slides:
https://docs.google.com/presentation/d/1rleH5cI1pHe5NWnodWVVc_4bLxFnH8FQ9KvQdbApgAw/edit?usp=sharing (updated)

Poster: (preview only)
https://create.piktochart.com/output/53480512-nlp-poster

Update dependencies for library pip 4.2.2

We urgently need to upgrade, else our application will be broken.

Deadline: Next Friday!!

Update on pipeline and feature issues

Tasks:

• HTTP links (maybe image link as well) are not yet replaced from “text”, which might affect word embedding.
• Indicative word count (bug, feature, doc) can be added into feature vector. here
• Do some selections on the feature vector, remove irrelevant ones that adversely impact accuracy.
• dataframe_generator might want to incorporate xxx_links_extracted.json
• Update code_count, link_count and image_count method in handcrafted_feature_extractor.py (links and code blocks can get directly from JSON data, consider remove image feature)

Model-wise, maybe SGDClassifier can be tried as well.

Feature Engineering

Language-dependent:

• count number of indicative word occurrences
  • feature: feature, features
  • doc: doc, docs, documentation, documentations
  • bug: bug, bugs, fix, fixes
• text-based features:
  • title length
  • description length
  • distinct word count
  • uppercase word count
  • stopword count
• Use word embedding to detect word similarity
  • compare Gensim word2vec model
    • CBOW: gensim.models.Word2Vec(data, min_count = 1, size = 100, window = 5)
    • Skip-gram: gensim.models.Word2Vec(data, min_count = 1, size = 100, window = 5, sg = 1)
  • basic implementation: https://towardsdatascience.com/a-beginners-guide-to-word-embedding-with-gensim-word2vec-model-5970fa56cc92

Language-independent: (markdown)

• common markdown element count of occurrence
  • highlight: bold, italics, section name
  • link: image, reference
  • code: code line (between ``), code block (between ``` ```)
  • special: list, task
  • possible APIs:
    • https://github.com/darul75/markdown-parser
    • https://stackoverflow.com/a/41009148
• element length (e.g. bold word count, code block word count, list item count)
  • implementation: rely on parser output from above
• further information extraction on meaningful entities (e.g. reference URL, code block content)
  • implementation: TODO

Language-independent: (GitHub)

• @github_username count of occurrence
  • implementation: custom tokeniser
• issue \ pr reference e.g. #12 count of occurrence
  • implementation: custom tokeniser

Dockerize our application