NLP and Data Science project

NLP and Data Science project#

Work in a small group to create a complete, reproducible text-classification project. The main outcome is not a particular accuracy score. It is a project that another student can open, understand, and run.

You can use the prepared GoEmotions data, or find another text-classification dataset that interests your group.

Start with a research question#

For the supplied data, you could ask:

Can patterns in a short comment predict whether a human annotator labelled it as anger, fear, joy, or sadness?

If you choose another dataset, write your own question before beginning the analysis. It should identify the text used as input and the category being predicted.

Create a project#

Make a new folder outside the workshop repository. Give it a meaningful name and organise it so that code, data, and generated results are not mixed together.

Your project might look like this:

emotion_project/
├── README.md
├── requirements.txt
├── data/
│   └── emotions.csv
├── scripts/
│   └── analyse_emotions.py
└── outputs/
    └── confusion_matrix.png

This is an example rather than a mandatory template. Your group should be able to explain every folder it creates.

Write one runnable script#

Create a Python script that:

  1. locates files with pathlib;

  2. loads and inspects the data with pandas;

  3. defines the text input X and category target y;

  4. creates training and test sets;

  5. places text vectorisation and classification in a pipeline;

  6. fits the pipeline using the training data;

  7. evaluates predictions on the test data;

  8. saves at least one useful result in the outputs folder; and

  9. contains comments explaining the important analytical decisions.

Use functions where they make the workflow easier to understand. Add a few assertions that check assumptions which matter to the analysis. The finished script should run from beginning to end without requiring someone to execute lines manually.

Make the project understandable#

Write a short README containing:

  • the research question;

  • the source and licence of the data;

  • what one row represents;

  • instructions for running the script;

  • the model and text representation used;

  • one result; and

  • one important limitation.

Be precise about the target. An emotion classifier predicts labels from text; it does not directly measure a person’s internal emotional state.

Option: choose your own data#

Your group may instead find a text-classification dataset on Kaggle Datasets. Sentiment, topic, spam, intent, authorship, and emotion datasets can all work.

Choose something manageable:

  • the data can be downloaded as CSV or converted easily to a DataFrame;

  • there is one text column and one clearly documented category column;

  • there are enough examples in each category for a train/test split;

  • the licence permits educational reuse;

  • the content is suitable to share in the classroom; and

  • the dataset does not require extensive cleaning before modelling.

Record the dataset page and licence in your README. Do not spend the whole project searching: if your group has not found a suitable dataset after ten minutes, use the provided GoEmotions subset.

Optional reference#

If your group gets stuck on the basic scikit-learn structure, consult this runnable example. In your cloned repository, it is at book/workshop_scripts/emotion_classification_challenge.py.

Use it as a reference rather than copying it unchanged: your own script and README should reflect your dataset and research question.

Supplied data#

The supplied file contains 1,000 screened GoEmotions comments, balanced across anger, fear, joy, and sadness. It is a small teaching sample and is not representative of emotion or Reddit users generally. Source: GoEmotions, used under CC BY 4.0.