wisaacj / sentiment-analysis

Plan of Action

[ ] Lab work

Complete module labs up to week 4 (or 5)

• AuthorProfiling
• AnalogySolver (opt)
• WordNetPractice (opt)

[x] Feature Generation

Generate at least three feature sets using paths through the following set of choices:

• Tokenization (must)
• Lemmatization OR Stemming
• Lowercase (or not)
• Remove stopwords (or not)
• Remove punctuation (or not)
• N-gram generation (pick 1, 2, 3, ..., n)
• Normalisation methods
  • Frequency Normalisation
  • Tf-Idf
  • PPMI

[ ] Naive Bayes

• Implement Naive Bayes from scratch
• Evaluate it on each of the three feature sets you've extracted (using the development splits)
• Pick the best of the three feature sets you've experimented with and present the test split results
• Evaluate the scikit-learn implementation of Naive Bayes on the same three feature sets and compare the results to your implementation

[ ] SGD-based classification and SVMs

For this task, you can use scikit-learn's implementation of the following models:

Logistic Regression

• Train a Logistic Regression classifier
• Evaluate each model on each of the three feature sets (using the respective development splits)
• Pick the best of the three feature sets you've experimented with and present the test split results

SVM

• Train a SVM classifier
• Evaluate each model on each of the three feature sets (using the respective development splits)
• Pick the best of the three feature sets you've experimented with and present the test split results

Subsequently, you should perform hyperparameter optimisation on your best performing model/feature set.

• Try at least 5 different combinations of hyperparameters using the development split
  • Consider using more comprehensive hyperparameter optimisation techniques like gridsearch or k-fold validation.
• Evaluate the best hyperparameters using the test split

Note that you should use 1-hot embeddings based on your features for input.

[x] BERT

• Follow the given tutorials and train BERT to perform sentiment analysis over the same train/dev/test splits of the feature sets.
• Experiment with both the cased and uncased version of BERT

[ ] ChatGPT

TODO: Come up with a plan.