01 Intro - Cheat Sheet

Definitions

Concept - Text Mining: extraction of implicit, previously unknown and potentially useful information from large amounts of textual resources (s.4). ~90% of data is unstructured (s.3).
Search vs discovery (s.10): IR = goal-oriented search on text; text mining = opportunistic discovery on text; data mining = discovery on structured data.
Concept - Text Mining Process (s.11): preprocessing → feature generation → feature selection → data mining/pattern discovery → interpretation/evaluation, with feedback.
Concept - Stopword Removal: ~400-500 words, 20-30% of word count (s.15). Concept - Stemming cuts term vector 40-50% (s.17).
Concept - Information Extraction subtasks: NER + disambiguation, relationship extraction, fact extraction (s.9).
Concept - Word Embeddings: dense real vectors (50-300 dims), distributional similarity, Word2Vec skip-gram (s.28).
Concept - Pruning: drop too frequent/infrequent words, percentual or absolute (s.31).

$t f_{ij} = \frac{occurrences of i in j}{words in j}$ (s.25); $i d f_{i} = lo g \frac{N}{d f _{i}}$ ; $w_{ij} = t f_{ij} \times i d f_{i}$ (s.26) (Concept - TF-IDF)
Jaccard: $\frac{M _{11}}{M _{01} + M _{10} + M _{11}}$ (s.35) (Concept - Jaccard Coefficient)
Cosine: $\frac{d _{1} ∙ d _{2}}{∥ d _{1} ∥∥ d _{2} ∥}$ (s.37) (Concept - Cosine Similarity)

	A	B
Stemming vs Concept - Lemmatization	rule-based, fast, may give non-words (amazing → amaz)	dictionary + POS, slower, valid words (amaze) (s.19)
Jaccard vs cosine	binary term occurrence, ignores 0-0	TF / TF-IDF weights (s.35-37)
Clustering vs classification	no labels, needs similarity	labeled training set (s.34, 41)
Bag of words vs embeddings	sparse, no synonyms, loses order	dense, related words close (s.23, 28)
Term occurrence vs term frequency	raw count	count / doc length (s.25)

Worked values: Jaccard 0.44/0.27/0.18 (s.36); cosine example 0.315 (s.37); cosine+TF-IDF 0.13/0.05/0.00 (s.39).

Define text mining; why is it important? (s.3-4)
Difference between IR and text mining (s.10).
List the text mining process steps (s.11).
Why remove stopwords? (s.15) Compare stemming and lemmatization (s.19).
Explain bag of words and TF-IDF (s.23-26); embeddings vs one-hot (s.28).
Compute Jaccard / cosine similarity (s.36-37); why cosine + TF-IDF (s.38-39).
Document classification methods: naive Bayes, SVM, RNN/LSTM, KNN, RF (s.41); sentiment analysis data and tricks (Concept - Sentiment Lexicon, s.43-45).
Why is text mining hard? Synonyms, subtle relations, high dimensionality, little labelled data, ambiguity (s.46, Concept - Text Mining Challenges).