Skip to content

Latest commit

 

History

History
188 lines (132 loc) · 7.83 KB

README.md

File metadata and controls

188 lines (132 loc) · 7.83 KB
Raw

"Share Love not Hate" - Assessing Hate Speech Detection methods using TF-IDF and POS tagging approach on Twitter (v 1.0.0)

social language language

gif

This is a repository for Hate Detection project under Research Methodology course Spring '21 at Lakehead University, Canada.

Table of Contents

  1. Our Philosophy
  2. About this project
  3. Road-map
  4. Project Status
  5. Usage
  6. Manifest
  7. Dataset
  8. Results
  9. Support

Our Philosophy

Rising cases of online bullying via offensive and hateful comments or tweets online makes hate detection of upmost need in this era. Multiple millenials including celebs, have fallen prey to online bashing afftecting their mental health, increasing anxiety and depression. Our philosophy is to share LOVE not hate 💜

sloth

About this project

  • This project will help identify potential hate texts from non-hateful ones.
  • You can type a message or comment in the input box in UI and hit submit. It will then display either 'This comment is hurtful' or 'GOOD JOB! Please continue spreading love 💌 !' based on harshness of the language used.
  • This feature can be further enhanced to prevent the user from posting such comments and completely avoid the hastle of blocking a user and establish respectful boundaries.
  • Its a plug and play feature, it can be integrated with your website user suggesion or comment box in a jiffy 🏃

sticker

Project Status

☑️ Collected data from HuggingFace library
☑️ Cleaned data, removed punctuations, emojis, special characters like hash tags and Twitter user mentions
☑️ Perform data analysis
☑️ Tokenized with TF-IDF with n-gram
☑️ Tokenized with POS tagging with n-gram
☑️ Task 1 : Multiclass classification between hateful, offensive language and neither

Road-map

  • Task 2 : Include Sentiment classification task to improve classification accuracy
  • Task 3 : Auto convert hateful words into endearing or encouraging antonyms and generate meaningful sentenses

Usage

  • No setup needed, just a google account 😺
  • You can directly open each ipynb Notebook in this repository by copying the notebook to your personal drive to get started on Google Colab!
  • "2_classifier_tfidf_pos_logistic_regression.ipynb" contains code for the best performing model
  • The dataset is pulled in the .ipynb notebook itself so there is no need to explicitly copy data files
  • If you wish to contribute, clone this repository
git clone 'https://github.com/kshitijahande/Hate-Detection.git'

Manifest

  • A list of the top-level files in this project with a description of what each file is.
- README.md                                          ----> This markdown file you are reading.
- 1_classifier_tfidf_svc.ipynb                       ----> This is experiment #1 to achieve Task 1. It uses only TF-IDF with Logistic regression and Linear SVM 
- 2_classifier_tfidf_pos_logistic_regression.ipynb   ----> This is experiment #2 to achieve Task 1. It uses TF-IDF and POS tagging with Logistic regression and Linear SVM 
- 3_classifier_bert.ipynb                            ----> This is a in-progress experiment #3 to extract features using BERT tokenizer
- documentation                                      ----> This folder contains all documentation work
- documentation/report.pdf
- documentation/slides.pdf
- results                                            ----> This folder contains detailed results for experiments #1 and # 2 table

Dataset

  • This dataset is pulled from HuggingFace library, made available by T. Davidson
  • Total rows: 24,783
  • Following columns are accessible:
'count',             --------> the number of CrowdFlower workers voted for class labelling
 'hate_speech_count',            --------> the number of CrowdFlower workers classified tweet as hate speech
 'offensive_language_count',     --------> the number of CrowdFlower workers classified tweet as offensive language
 'neither_count',             --------> the number of CrowdFlower workers classified tweet as neither hate speech nor offensive language
 'class',             --------> Final class label assigned to the tweet      
 'tweet'            --------> Tweet
  • Class column can contain values 0, 1 or 2
"class":[
            0:"hate speech"
            1:"offensive language"
            2:"neither"
]

Results

  • The experiment #2 using TF-IDF and POS tagging with Logistic regression for feature selection and Linear SVM for classification with L2 regularization performed the best, resulting in 0.91 F1-score.

results

Support

social

  • I would highly appreciate and welcome all your contributions and suggestions to improve this work!

cat