Natural Language Processing Project

Planning the project

Goals

The goal of this project is to try to predict the programming language used in a repo based solely on the content of the readme file attached to that repo.

Deliverables will include:

• This repo containing:
  • A well-documented Jupyter Notebook detailing the process to create a predictive model
  • This Readme.md detailing project planning and exection
  • Final model created to predict programming language
• A full slideshow containing well-labelled visualizations

Where did the data come from?

• The data was acquired from over 400 github repositories that focused on technology
• The resulting dataset can be accessed via 'repos.csv', located within this repository

Data Dictionary

The variables are the following:

Feature	Description	Data Type
repo	The name of the github repository	Object
language	The programming language used in the repo	Object
read_me contents	The text that makes up the README file	Object

Project Steps

Acquire:

• I got the data using Zach Gulde's acquire.py
• I converted it to a dataframe
• I then wrote it to a csv for ease-of-use
• Lastly, I converted the csv back to a dataframe to begin the prep process

Prep:

After prepping my data, I dropped from an initial 407 observatins to 293 observations

• I dropped my null values: 70 in 'language" and 4 in 'readme_contents'
• I dropped any language that had less than 10 occurances
• I used my prepare.py to clean, tokenize, and lemmatize the 'readme_contents' column
• My prepare.py created new columns: "clean_lemmatized" and "clean_clean_lemmatized"

Explore:

• Created a dictionary to hold subsets by language
• Identified and dropped many "noise words": words that either appeared in too many READMEs or were too generic to be helpful
• I made bigrams to compare the top languages

Modeling:

• I split my data
• I then vectorized my data so that I could use classification models
• I established a baseline of 31% - the mode value
• I ran a Linear Regression model and a Decision Tree model
• I decided to oversample my data using SMOTE
• I decided to reduce my data to the top 5 languages

Conclusion and Next Steps:

• Because the initial baseline was so low, even a weak-performing model could beat it
• Each model was overfit
• As small training datasets are prone to over-fitting, I'd get a larger dataset to sample
• I'd like to work with SMOTE more and get the hang of it