• Data quality review: Analyze missing values, duplicates, and relevance.
• Text cleaning: Remove noisy data, irrelevant data, or data outside the study’s context.
• Remove punctuation: Identify and remove irrelevant punctuation, while preserving necessary ones if required.
• Remove numbers: Depending on whether they add value to the study or not.
• Whitespace normalization: Remove double spaces between sentences.
• Convert to lowercase or uppercase.
• Lemmatization: Convert words to their base form, whether verbs or other parts of speech.
• Remove stopwords: Common words that don’t add significant meaning, such as conjunctions or short words.
• Word tokenization: Split the text into individual words or tokens for statistical analysis.
• N-gram tokenization: Capture sets of n words (n-grams) to understand context.
• Bag of words: Represent text as a word count matrix, yielding a matrix with the number of occurrences for each word. —> corpus
• TF-IDF representation: Inverse document frequency - $wordFrequency/totalWords$
• Embedding: Represent words as dense vectors in high-dimensional space (Word2Vec, GloVe, FastText).
• Oversampling: Increase the examples of the minority class, useful in classification problems with imbalanced classes.
• Undersampling: Reduce examples from the majority class, useful when there is an excess of examples from one class.
• Dimensionality reduction: Identify and remove words with very low frequency in the corpus.
• PCA reduction: Reduce the number of dimensions in word vectors.

Resampling

Split the data for evaluation. How to divide datasets into subsets by percentage for training/validation.

• Percentage validation: Take a percentage of the data for test and training (70%-30%).
• Cross-validation: Split the data into several $k$ (3,5,10) subsets and randomly select one $k$ subset for testing while using the rest for training. The accuracy is the mean of the iterations.