NLP Text Classification Interview Questions

Why Text Classification Fundamentals Still Anchor NLP Interviews

Text classification is the task most NLP interview loops use as a baseline, precisely because it's simple enough to discuss end to end in fifteen minutes while still touching every stage of a real NLP system: preprocessing, feature representation, modeling, and evaluation. Interviewers use it to check whether a candidate can reason about a full pipeline, not just name a model.

This guide covers three concepts that come up together: how a text classification pipeline is structured end to end, the different approaches to sentiment analysis specifically, and why word embeddings replaced one-hot encoding as the default way to represent words. Each section links a practice problem.

The Text Classification Pipeline

"Design an end-to-end pipeline that classifies customer support tickets into categories. Walk me through every stage."

A text classification pipeline has four stages interviewers expect you to name explicitly. Preprocessing turns raw text into a cleaner, more consistent form: tokenization splits text into words or subword units, lowercasing normalizes case so "Refund" and "refund" aren't treated as different tokens, and stopword removal (dropping very common, low-information words) is sometimes applied, though it's increasingly skipped for models that can learn to downweight uninformative tokens on their own.

Feature representation converts the cleaned text into a numeric form a model can consume. Classical approaches use bag-of-words (a vector of word counts) or TF-IDF (word counts reweighted to downweight terms that appear in most documents and upweight terms that are distinctive to a given document). Modern approaches use dense embeddings, either pretrained word embeddings averaged or pooled across a document, or contextual embeddings from a transformer encoder.

Modeling trains a classifier on top of that representation — logistic regression or a linear SVM are strong, fast baselines for bag-of-words or TF-IDF features, while a fine-tuned transformer is the standard choice when embeddings and maximum accuracy matter more than latency or interpretability. Evaluation should go beyond raw accuracy, especially when the classes are imbalanced (which support ticket categories usually are): precision, recall, and F1 per class, plus a confusion matrix, tell you which categories the model actually struggles with, information accuracy alone hides.

The framing interviewers reward: this is a pipeline with real trade-offs at every stage, not a single "use BERT" answer — being able to justify why bag-of-words plus logistic regression might be the right call for a fast, interpretable baseline, versus when a fine-tuned transformer is worth its extra cost, is what separates a strong answer from a name-dropped one. Practice: Text Classification Pipeline.

Sentiment Analysis: Three Approaches, Three Trade-offs

"How would you build a sentiment classifier for product reviews, and what are your options?"

Lexicon-based approaches score text using a predefined dictionary that maps words (and sometimes phrases) to polarity scores, then combine those scores, often with hand-crafted rules for negation and intensifiers, into an overall sentiment score, without any model training. Tools like VADER are common examples. This approach needs no labeled data and runs fast, but it struggles with context-dependent meaning, sarcasm, and phrasing the lexicon wasn't designed to handle.

Classical machine learning approaches convert text into bag-of-words or TF-IDF features and train a standard classifier — logistic regression, an SVM, or similar — on labeled sentiment examples. This captures patterns a fixed lexicon can't, at the cost of needing labeled training data, and it still doesn't model word order or long-range context particularly well since bag-of-words representations discard sequence information.

Deep learning approaches, from LSTM-based classifiers to fine-tuned transformers, capture context and word order directly, and fine-tuned transformer models are the current standard when accuracy matters most, particularly for handling negation, sarcasm, and mixed sentiment within a single piece of text. The trade-off is cost: more labeled data, more compute for training and inference, and less interpretability than a lexicon or a linear classifier over sparse features.

A detail worth having ready: sentiment analysis is sometimes framed as binary (positive versus negative), and sometimes as a finer-grained scale (a one-to-five star prediction, or positive, negative, and neutral). The finer-grained version is a harder problem in practice, since the boundary between adjacent classes, such as three stars versus four stars, is often genuinely ambiguous even to a human labeler, which shows up directly in lower inter-annotator agreement on the training data itself and, in turn, a harder ceiling for any model trained on it.

The answer interviewers want isn't "always use a transformer" — it's the ability to place these three approaches on a spectrum of cost versus context-handling ability and pick the right point on that spectrum for the actual constraints of the problem: how much labeled data exists, what latency budget you have, and how much sentiment nuance the product actually needs to capture. Practice: Sentiment Analysis Approaches.

Word Embeddings vs One-Hot Encoding

"Why not just represent every word as a one-hot vector and let the model figure out the rest?"

A one-hot vector represents a word as a vector the length of the entire vocabulary, with a single 1 at that word's index and 0 everywhere else. This representation has two structural problems. First, dimensionality: the vector length equals vocabulary size, which for a real corpus can be tens or hundreds of thousands of dimensions, almost all of them zero for any given word. Second, and more important, every pair of distinct one-hot vectors is orthogonal by construction — the cosine similarity between "cat" and "dog" is exactly the same as between "cat" and "spreadsheet," which means the representation encodes zero information about which words are related in meaning or usage.

Word embeddings fix both problems at once. Words are mapped to dense, lower-dimensional vectors — commonly in the range of 100 to 300 dimensions for classical word embeddings — that are learned (or fine-tuned) so that words appearing in similar contexts end up with similar vectors. That single property, similar words having similar vectors, is what lets a model trained on one word generalize to related words it may have seen less often, since the model operates on the geometry of the embedding space rather than on an arbitrary, meaningless index.

Interviewers sometimes push on why this matters beyond "smaller vectors": with one-hot encoding, a model has no way to transfer what it learned about "excellent" to a related word like "outstanding" that appeared rarely in training data, because the two words' representations share no structure at all. With embeddings, if the two words end up with similar vectors, the model's learned behavior for one naturally extends to the other. Practice: Word Embeddings vs One-Hot Encoding.

How to Prepare

1. Practice naming all four pipeline stages — preprocessing, representation, modeling, evaluation — and be ready to justify a specific choice at each stage, not just list them.
2. For sentiment analysis, have the three-approach spectrum ready (lexicon, classical ML, deep learning) along with the cost each step up buys you.
3. For embeddings versus one-hot, lead with the orthogonality problem — no two one-hot vectors are more similar than any other pair — since that's the concrete failure mode, not just "embeddings are denser."

For the rest of the concepts this track covers, see our NLP interview questions, which links practice problems across text processing, modeling, and information extraction topics.