We have collected user reviews on Glassdoor.com for various companies. The goal of this challenge is to build a classifier based on our psychological model (referred to as the O'Reilly Model).
This model uses 6 dimensions to represent company culture:
- Adaptability
- Collaboration
- Customer orientation
- Detail orientation
- Focus on Principles (or Integrity)
- Results orientation
The challenge is to build a text classifier that predicts to which dimension a review is the most related to, based on its content. We have labelled a dataset, so that you can take a supervised learning approach if needed.
The labelled dataset is a pickled DataFrame with 2 columns: the review's text and the label of the main dimension of that review.
The unlabelled dataset contains one .json file per company and contains their Glassdoor reviews. Feel free to manipulate and restructure this dataset to whatever works best for you.
- Use Python as programming language
- Achieve an accuracy of 90% of trained classifier
- Process the unlabelled data using the classifier and save the output in a file, so that we can analyze it
- Explain your approach: your code should be documented and you should be able to explain your decisions
- Discuss the next steps to further improve the classifier (eg: work with multiple languages)
- Explain how would make this classifier available in production
Submit your challenge either by sharing your fork of this repository on Github or by sending your local repository as a compressed archive via email/gist (if you want the challenge to remain private).
The deliverables should be easy to run and visualize. You should provide either a Jupyter Notebook or a runnable project (with clean environment and dependencies management, we expect the use of virtualenv/pipenv or Docker)
We expect you to spend about 2-4 hours on this challenge. If you find yourself spending more time on it, you can send your results as it should be enough to evaluate.
Tips:
- Make sure to remain focused and not get side-tracked, or the challenge will take more than 2-4 hours to complete.
- We are also not expecting production ready code, so you can leave out some aspects you would otherwise consider important (however, documenting these decisions is always a plus).