Chattermill NLP Challenge

At Chattermill we do a lot of cool things with text. One of them is aspect analysis, where we extract mentioned attributes of a product from customer comments (for example speed of delivery for a courier service). This is typically a hard problem to solve. That's where you come in.

The goal

1. We would like you to use your favourite classification algorithm to predict the aspects from the given feedback data.
2. We then also would like to see this model served in a local inference server.

The emphasis is on overall code quality rather than fine-tuning of the algorithm for superior accuracy.

Data

There are 2 comma-separated datasets of customer reviews, train.csv and test.csv. Each comment in train.csv has some multilabel ground-truth combination of aspects, while test.csv provides no ground-truth labels.

Inference Server

• Create an inference server with an endpoint to serve inference requests to the model.
• You may use a framework of your choice such as FastAPI or Flask.
• You may provide a docker file to run the inference server, or otherwise provide clear instructions for running it.

Expected Output

We expect your code to train a classification algorithm on train.csv that infers on test.csv and exports its aspect predictions to a file called test_answers.csv. Everything else is for you to decide. You might want to add some exploratory data analysis or visualisations.

Please submit your solution (including the completed test_answers.csv file) as a zip file using the Greenhouse link provided by our talent team.

Add a README.md file explaining your process. It can be in a format of a short blog post, report or a tutorial. Any cool insight you find along the way, or cool tool you use will be a bonus. The file should contain full instructions on what we need to do to rerun your code.

Notes

• You should be able to complete the project within a few hours, not days.
• Feel free to work in whatever language / framework you are most comfortable in, but we prefer Python for ML tasks.
• Feel free to use any machine learning library like Keras, TensorFlow, PyTorch
• You are free to use word vectors or pre-trained models to complete this task if you wish, however please consider your memory consumption and training times (anything that takes over an hour to train on a GPU is going to be too long)
• Please focus on making the code clear and modular, rather than on the complexity of the solution. Accuracy level itself will not be a success factor as much as our understanding of what you are doing and why. Provide your solution as if it were a research task.
• You may use Jupyter notebooks to code and present your solution if you prefer

Bonus Points

• Clear, readable code
• Good naming of methods
• Being able to concisely explain what you have done and why certain decisions have been made

Get in touch

• If something isn't clear, feel free to submit an issue to this repo or drop us an email at [email protected]