Write a report (max 1,500 words) on the challenges the dataset presents, the solutions, and your findings, which will be assessed as follows:
1) Discuss the following feature extraction techniques and explain how they work and their advantages and disadvantages
a) Term Frequency-Inverse Document Frequency (TF-IDF) [10%]
b) BERT [10%]
2) Two step Classification:
a) Related/Unrelated classification:
i) Use TF-IDF features to train a standard Machine Learning model (e.g. SVM, Naïve Bayes, Random Forest), and discuss its performance on the testing set to classify whether the article body is related or unrelated to the headline. [15%]
ii) Train one Deep Learning model (e.g., LSTM, RNN, CNN). Explain and justify the architecture of the deep learning model, the hyperparameters used, and the loss function. Discuss the performance on the testing set to classify whether the article body is related or unrelated to the headline. [15%]
iii) Analyse and compare the performance results for the two models.
[10%]
b) Agree/Disagree/Discuss classification:
i) Build a new deep learning model of your choice to classify articles into the remaining three categories (Agree/Disagree/Discuss). The inputs to this model should be only samples that are related to the headline (i.e. you should train and test your model on only these samples). [15%].
ii) Analyse the performance of your model and report the results. [10%]
3) What are the ethical implications of your proposed solutions? What are the potential biases and future misuse cases? [10%]
4) Academic English writing, with good use of technical vocabulary, correct grammar, appropriate document structure and referencing where relevant. [5%]
You should submit your 1,500-word report and also the associated Jupyter notebook used to produce your analysis and graphs.