Don't Ask to Ask Bot

Project Overview

A Discord bot that detects 'AskToAsk' questions using machine learning and prompts users with guidelines on how to ask their questions directly.

Concept

Problem

Users often waste time asking for permission, asking if the channel they're in is the appropriate one, or asking if any experts in X are online, etc etc.. which can be inefficient and cumbersome for mods. Examples:

noobie123: Any XNA experts around?
ahabz: Anyone here good at JAVA?  
pandasfan90: Can i ask about X here?  
pookiebear827: Anyone familiar with Y?  

Solution

A bot that identifies these types of questions and encourages users to directly ask their questions.

Sure, here's a revised version with clearer instructions:

How To Use:

To get started with the Don't Ask to Ask Bot, you have two easy options:

1. Invite the Bot to Your Server:

   • Click on this Bot Invitation Link.
   • Follow the prompts to select your server and authorize the bot. No setup required—it's ready to use immediately!

2. Test the Bot in Our Community Server:

   • Join our Discord Community server.to see the bot in action.
   • You can interact with the bot and see how it responds to 'AskToAsk' questions without any installation.

No complicated setup is needed. Just invite the bot to your server or test it in ours!

Features

• Detects "AskToAsk" questions.
• Provides prompts to users to directly ask their questions.
• Will Only respond to newcomers in your server(2 responses max per user to avoid spam)
• Uses machine learning for classification.
• Regex filtering for initial screening.

Tools and Technologies

• Programming Languages: Python
• Libraries:
  • Natural Language Processing: SpaCy, NLTK
  • Data Handling: Pandas, NumPy, SQLite
  • Regular Expressions: re
• Machine Learning Model:
  • Linear Support vector machine
• Text Processing Techniques:
  • Tokenization
  • Lemmatization
  • TF-IDF (Term Frequency-Inverse Document Frequency)

Data

• Data Scraping: The data was scraped from Discord servers using Python's requests library. The process involved:
  • Using the Discord API to access channels.
  • Extracting only questions using a regex pattern that looks for interrogatory keywords.
  • Storing the data in a structured format for preprocessing and labeling.
• Dataset:
  • Questions labeled as "permission" or "direct"
  • Size: Approximately 5,252 questions (2,892 direct, 2,360 permission) from an original 8k questions pool.
• Preprocessing:
  • Lemmatization using SpaCy
  • TF-IDF vectorization

Model Training

• Training:
  • Split data into training and testing sets (80/20 split)
  • Shuffle data to ensure randomness
  • Fit the SVM Classifier on the training data
• Evaluation:
  • Accuracy: Achieved 94% accuracy with bigram (1,4) TF-IDF vectorization
  • Type I (False Positives) and Type II (False Negatives) Errors Analysis with emphasis on low FP to not disturb the natural flow of the discord chats.

Additional Techniques

• Regex Filtering:
  • Initial filtering of questions using regular expressions pattern
  • Non-capturing group terms for exclusion and inclusion of certain keywords

Deployment

• Hosting:
  • The bot is hosted on an AWS EC2 instance.
  • The EC2 instance provides the necessary compute resources and uptime for reliable bot performance.
• Bot:
  • Functions on Discord and monitors specified channels.
  • Detects when users ask permission to ask questions.
  • Sends a message prompting users to directly ask their questions.
• Integration:
  • Continuous monitoring and feedback to improve detection and prompts.

Repository Structure

• Folders: -Research:
  • data: Contains datasets of raw and preprocessed data
  • notebooks/models: Notebooks containing Exploratory Data Analysis and model training and testing
  • scraping:notebooks containing scripts for scraping questions, sanity checking the results and storing them.
• Scripts:
  • Data scraping
  • Data preprocessing
  • Model training and evaluation
  • Bot interaction and deployment scripts

License

This project is licensed under the MIT License.

Future Improvements

• Enhancements:
  • Expand the dataset through scraping more "AskToAsk" questions in order to enhance the accuracy
  • Integrate more NLP rule-based matching techniques for better understanding of context

---

Motivation

I'm interested in learning more about natural language processing and looking to land an apprenticeship in France for a master's degree program. This project was a learning experience to improve my skills in NLP and machine learning, and I hope it can be useful to the community.

Note: inspired by Dontasktoask.com.