While preparing a large text corpus for topic modeling, a practitioner wants a standard preprocessing step that reduces surface-form variety and thus complexity. Which technique fits that goal?
Select an answer to reveal the explanation.
Short Explanation and Infographic
Okay, here's the real-world move. Topic models hate pointless vocabulary explosion—"run," "runs," "running" as three totally different things. Stemming or lemmatization folds those forms toward a common base so your topics stay cleaner. Think of it like sorting mail by last name, not every nickname on the envelope. Exam trap: batch standardization, image augments, resizing photos—those are from other ML neighborhoods. Don't drag CV tricks into a text prep question. For topic modeling complexity reduction, morphologically normalize the words. Got it? Let's keep rolling.
Full explanation below image
Full Explanation
Topic modeling methods such as latent Dirichlet allocation operate on bag-of-words-style counts over a vocabulary. When every inflection and morphological variant is treated as a distinct type, the vocabulary grows and probability mass fragments across related forms. Stemming applies rule-based truncation to produce crude stems; lemmatization uses linguistic knowledge to map tokens to dictionary lemmas. Either approach is a common preprocessing step that reduces data complexity by collapsing related surface forms, improving the coherence and statistical support of topics for many corpora. That complexity reduction is exactly what the scenario asks for when a practitioner prepares text before discovering latent themes.
Additional text-cleaning steps often accompany this choice—lowercasing, tokenization, stop-word removal, and rare-term pruning—but among the options given, stemming or lemmatization is the technique that specifically targets morphological variety. Batch standardization normalizes numeric feature scales across batches or features; it is standard in neural training on continuous inputs, not a primary text-normalization tool for classical topic models. Data augmentation techniques that synthesize new examples (especially image flips, crops, and color jitter) address sample diversity in supervised learning, not vocabulary compression for unsupervised topics on documents. Resizing images is purely a vision pipeline operation and has no role in preparing a text-only topic-modeling corpus.
Practitioners should note trade-offs: aggressive stemming can merge unrelated words and harm interpretability of top words per topic; lemmatization is usually gentler but needs language-appropriate tooling and may be slower. Domain language in biomedical or legal text may require custom lexicons so technical terms are not over-collapsed. For the exam, pair topic modeling plus reduce complexity of text with morphological normalization via stem or lemma. Memory aid: fewer word shapes yield cleaner topics. Keep vision tricks and generic numeric scaling answers for their own domains, and reserve stemming or lemmatization for this NLP preprocessing goal whenever surface-form variety is the stated pain point.