Nike Brand Perception Analysis

Why it works

NLP models can't read sentences the way we do. Tokenization breaks unstructured text into discrete units, specifically individual words, that become the rows of a tidy data frame. In R's tidytext, unnest_tokens() splits each review into one-word-per-row, converting to lowercase and stripping punctuation in one step. This is the foundation that every downstream operation depends on. Counting, filtering, and scoring all operate on these atomic tokens.

Why it works

Words like "the", "is", "for", and "are" appear constantly but carry zero brand-specific meaning. An anti-join against a stopword list (stop_words from tidytext) removes them in one pass. Without this step, TF-IDF and topic modeling would be dominated by these high-frequency filler words instead of the actual signal. Words like "swoosh", "boost", and "c2c" are what actually differentiate brands, and they only surface once the noise is removed.

Why it works

"Running", "runs", and "runner" all point to the same concept. Stemming collapses these inflected forms into a single root ("run") so they get counted together. Instead of fragmenting signal across four variants, we concentrate it in one stem, which boosts statistical power significantly. The tradeoff is that stems are not always real words ("comput" for "computing"). However, for frequency-based methods like TF-IDF and LDA, precision of the root matters less than consistent aggregation.

Why it works

Lexicon-based sentiment analysis maps each token to a pre-defined score. We use three lexicons in parallel: AFINN (integer scores from negative 5 to positive 5), Bing (binary positive/negative), and NRC (8 emotions). Each approaches sentiment differently. AFINN captures intensity ("amazing" = +4 vs "good" = +1), while NRC reveals the emotional texture behind the polarity. By running all three, we triangulate. If all three agree Nike reviews skew positive, that signal is robust.

Why it works

Every NLP technique in this pipeline implicitly treats words as vectors. TF-IDF creates sparse vectors where each dimension is a document. Word embeddings create dense vectors where proximity equals semantic similarity. The classic example is "king" minus "man" plus "woman" approximately equals "queen". In our corpus, this means sneakerhead terms like "c2c", "deadstock", and "grails" cluster together because they are semantically close even though they share no characters. Understanding this geometry is key to why TF-IDF and LDA work, since they operate on the distances between these vector representations.

Why it works

Term Frequency times Inverse Document Frequency is elegant. TF measures how often a word appears in one brand's reviews, and IDF penalizes words that appear in all brands' reviews. The product surfaces words that are both frequent and distinctive. "Shoe" gets crushed because high TF everywhere times low IDF equals noise. "Swoosh" gets amplified because high TF for Nike times high IDF equals brand signal. This is how we discover that "c2c", "snkrs", and "deadstock" are Nike-specific vocabulary tied to the resale and hype culture that does not exist in Adidas or Under Armour reviews.

Why it works

Latent Dirichlet Allocation assumes every document is a mixture of hidden topics, and every topic is a distribution over words. The algorithm discovers these topics unsupervised. We just tell it how many topics to find (k=4). What emerges reveals what consumers actually talk about: performance, product quality, app experience, and customer service. The power of LDA is that it finds structure humans might miss. Reviews mentioning "return" and "refund" cluster with "wait" and "order" even though no one labeled these as a "service" topic.