NLP & Topic Modeling

A scholar sorting a scattered heap of letters into a few ribbon-bound stacks by theme

Your text-analysis journey has been a ladder of bigger questions. Counting words asked what appears most? Sentiment asked how does it feel? Today you climb to the biggest one yet: what is this whole pile of text even about? Topic modeling is the deepest "distant reading" move in the course — it surfaces themes running through hundreds of documents that no one could find reading one comment at a time. But it works by a move you should be suspicious of, and getting clear on that move is the whole lesson.

You already do this — instantly

Drop into a conversation halfway through and you figure out what it's about in seconds. Somebody keeps saying budget and deadline — it's a work project. Election, campaign, vote — politics. You read the repeated words, the keywords, the tone, and you infer the subject without anyone announcing it. And when you can't tell, you do the one thing that matters most: you ask — "wait, the group project or the internship?"

That instant, unconscious sorting is exactly the skill we're about to hand to a machine. Watching it try is the fastest way to see what's easy about reading and what's quietly hard.

The computer gets only the words

Here's the gap. You bring tone of voice, facial expression, cultural memory, and the prior conversation to every sentence — and you can always ask "what do you mean?" A computer gets just the words on the screen. No tone, no context, no cultural knowledge unless someone programmed it in, and no way to ask a clarifying question. Closing that gap is the whole field of Natural Language Processing — the autocorrect, spam filters, and translation you already use daily.

Language fights back, because it was built for humans, not machines. "I saw her duck" — the pet or the motion? "That's sick" — disgust or delight? "Oh, great, another meeting" — the words say positive, the meaning is the opposite. Ambiguity, idiom, and sarcasm are easy for you and brutal for a machine, because they live in the context the machine can't see. Keep that blind spot in mind; it comes back at the end.

Organizing a library without reading the books

So how does a machine find themes in text it can't actually understand? Imagine sorting a library with no time to read a single book. You'd look at which words appear together on each cover and back, and infer shelves: history, romance, science fiction. You never read for meaning — you just notice what co-occurs.

That's topic modeling. We'll use LDA (Latent Dirichlet Allocation) from the Gensim library, and it rests on two assumptions:

Each topic is a cluster of words that tend to show up together.
Each document is a mixture of those topics — not filed under one, but "70% this, 30% that."

LDA works backward from the finished text to guess what clusters must have produced it — like reverse-engineering a recipe from the plated dish. tomato + basil + mozzarella → an "Italian" topic. It finds the pattern by raw co-occurrence. That is all it ever does.

The machine clusters; you name

Here is the move to be suspicious of — and the most important sentence in this lecture. LDA hands you back lists of words. It never hands you a name.

Topic 3: money, dream, green, light, decay, gold. The algorithm does not know this is "the American Dream."
You know that. Naming it is your move, not the machine's.

The computer discovered that those words travel together; it has no idea what they mean, that they're about ambition or disappointment or Gatsby. Interpretation is the humanist's job — and it's not neutral. What you name a cluster depends on what you already know, what you're looking for, and where you stand. Two readers can label the same word-list differently and both be defensible. The statistics are the machine's; the meaning is yours. Own that, and don't dress your interpretation up as something the algorithm "found."

There is no correct number of topics

Before LDA runs, you tell it how many topics to look for — num_topics. People assume the algorithm will reveal the right number. It won't. Ask the same comments for two topics and unrelated themes get mashed together; ask for ten and one real theme splinters into fragments.

Neither answer is "wrong." Choosing the number of topics is a research decision you justify — by how interpretable and useful the result is, not by a truth the model uncovered. It's the course's recurring lesson wearing new clothes: a knob you turned is doing work that's easy to mistake for a fact the data handed you. Make the choice visible, and say why you made it.

When one debate fights back

A toy corpus — fifteen constructed comments written in the style of sports, music, and food forums — topic-models beautifully: three obvious subjects, three clean word-lists you can name in a second. That clarity is a setup. Now point the same machine at eleven short comments written to sound like our Texas Ten Commandments debate (also constructed; you meet the real 123-row corpus in HW4), where every voice is arguing about one law, in overlapping vocabulary.

The topics come back muddy. LDA still hands you three tidy word-lists — it always will — but here each one is just more of the same argument, not three distinct themes. There weren't separate subjects to find, so the model imposed structure anyway. (With only eleven documents, size is part of the problem too; HW4 lets you see whether the real corpus does any better.) That's the warning: the algorithm produces clusters whether or not real themes exist, so part of your job is judging whether the output means anything at all. (And because LDA is stochastic, rerun it or run it on another machine and the groupings shift — which is why we pin random_state just to talk about the same result twice.)

What the model can't hear

The bird's-eye view is real power — you can see across hundreds of documents at once, a scale no close reading reaches. But hold it to the same questions you've asked all term:

Whose text got counted? Topic models only see what was digitized and collected. The empty seats aren't in the file — the same lesson the data-as-artifact work pressed on you.
Who decides what it means? The algorithm produces word clusters; humans name them. Whose categories, whose authority? You've met this question before as the logic of classification.
What does it miss? Irony, sarcasm, coded language, the context that flips a meaning — exactly the things that made language hard at the start of this lecture. Computational reading has blind spots, and they're not random.

LDA assumes topics are stable categories. Culture is fluid and contextual. So use the model for what it's genuinely good at — a map of a pile too big to read — and never confuse the map for the territory. The machine finds the pattern; you decide whether the pattern is a finding or just order imposed on noise.

Going deeper (optional): David M. Blei, "Topic Modeling and Digital Humanities" (Journal of Digital Humanities 2.1, 2012); Ted Underwood, "Topic Modeling Made Just Simple Enough"; Benjamin M. Schmidt, "Words Alone: Dismantling Topic Models in the Humanities" (Journal of Digital Humanities 2.1, 2012; the skeptic's case); Catherine D'Ignazio & Lauren Klein, Data Feminism (whose data, whose categories).