Appearance
Balance scale and close reading when your question is about prevalence or comparison across a corpus too large to read in full, and the signal is something you can count or measure. Use close reading alone when the object is a single text's nuance, irony or argument. The honest answer is that most good cultural-analytics projects are a loop: the computational pass proposes patterns, close reading tests and interprets them, and the results refine the next pass. This guide gives the decision signals.
When does scale actually earn its place?
Scale is worth its cost only when three things hold:
- Volume exceeds what a team can read — typically several thousand documents up.
- The question is countable — frequencies, co-occurrences, trends, clusters — not "what does this poem mean".
- Coverage matters — you need to know about the boring 90% of the corpus, not just the canonical exemplars.
If any one fails, a careful manual sample usually beats a model. A 300-letter archive does not need topic modelling; read it.
When should I not reach for computation?
Distant reading actively misleads when the corpus is small, the signal is rhetorical rather than lexical, or the documents are too noisy (heavy OCR error, mixed languages) for the model to find anything real. Running an LDA model on 200 short, error-ridden documents produces topics that are artefacts of the noise. The cost — your time, plus the false confidence of a colourful chart — exceeds the benefit.
The loop: how the two methods feed each other
Treat them as alternating phases, not rival camps:
1. DISTANT → cluster / count / trend across the whole corpus
2. SAMPLE → randomly pull N documents from the pattern of interest
3. CLOSE → read those N; do they mean what the model implies?
4. REVISE → fix tokenisation, stopwords, time bins; re-run step 1The discipline is in step 2: sample before you decide what you believe. Reading only the documents that confirm the cluster's label is how you manufacture findings.
Comparing the trade-offs
| Dimension | Close reading | Distant reading |
|---|---|---|
| Best corpus size | 1-300 documents | 3,000+ documents |
| Captures | Nuance, irony, argument | Prevalence, trend, structure |
| Main risk | Cherry-picking famous texts | Mistaking noise for signal |
| Reproducible | Hard | Yes, if code is shared |
| Time cost | Scales with documents | Front-loaded, then cheap |
How do I avoid cherry-picking when I switch modes?
Pre-commit your close-reading sample. Either pre-register the document IDs before seeing model output, or draw a random sample from the target cluster with a fixed seed:
python
sample = cluster_df.sample(n=15, random_state=7)["doc_id"].tolist()
print(sample) # read exactly these, no swapping for "better" examplesIf you swap out a document because it "doesn't fit", you've left analysis and entered illustration.
A worked decision
Suppose you have 12,000 19th-century newspaper editorials and want to know whether anxiety about "machinery" rose before 1850. That is countable, comparative, and far too large to read — distant reading fits. But the interpretation of any spike (was it fear, pride, or sarcasm?) requires reading a random sample of the spiking documents. Neither method alone answers the question; the loop does.
Key Takeaways
- Reach for scale when the corpus is too large to read, the question is countable, and coverage matters.
- Stay with close reading for single texts, rhetorical nuance, or small/noisy corpora.
- Treat the methods as a loop — distant proposes, close disposes, then you refine.
- Sample documents for close reading before deciding what they mean, to avoid cherry-picking.
- Convergence between methods is validation; divergence is where the interesting findings live.
- Small, error-ridden corpora produce model artefacts — don't compute on them.
- "Distant reading" is a stance, not a synonym for big data.
Frequently Asked Questions
When does distant reading actually add value over close reading?
When the question is comparative or about prevalence across thousands of documents that no person could read in full, and where the signal is countable. For a single text's nuance, close reading wins outright.
How many documents justify a computational pass?
There's no hard line, but below a few hundred documents you can usually read the sample directly and gain more than any model gives you. Above a few thousand, manual coverage becomes infeasible and scale earns its place.
Doesn't computation just confirm what close reading already found?
Sometimes, and that's a feature — convergence is validation. The value appears when scale surfaces a pattern close reading missed, or when it shows a famous "pattern" doesn't generalise.
How do I move between the two without cherry-picking?
Pre-register which documents you'll close-read before you see the model's output, or sample them randomly from the cluster of interest. Reading only the examples that fit your hypothesis is the classic trap.
Can a small team do both?
Yes, by treating them as a loop, not parallel tracks: distant reading proposes, close reading disposes, then you refine the computational pass. Budget the close-reading time explicitly.
Is 'distant reading' the same as 'big data'?
No. Distant reading is a stance — reading patterns across a corpus rather than individual texts — and works at modest scale. Big data is about volume and infrastructure; the two often coincide but aren't identical.