Before you can build a corpus, you need to answer a critical question: What, exactly, are you studying? You can't just collect “English.” You need to define a target population. This is the specific variety of language your corpus is meant to represent. The more precise your definition, the more powerful your analysis will be.

Think of it like a biologist studying a species. They don't just study “birds”; they study the migratory patterns of the Arctic Tern or the diet of the Galápagos Finch. In linguistics, your target population might be “formal academic writing in biology journals from 2010-2020,” “transcripts of corporate earnings calls,” or “British English tweets about football.” Each of these represents a unique ecosystem of language.

It’s almost always impossible to collect every single text from your target population. Instead, you create a sample that is representative of the whole. The goal is to build a miniature version of your target language variety, capturing its key characteristics in the right proportions.

Two common methods for this are proportional and stratified sampling. Imagine you're studying a news website that publishes 70% news articles, 20% opinion pieces, and 10% feature stories. A proportional sample would mirror these percentages exactly. If you collect 100 texts, 70 would be news, 20 opinion, and 10 features.

Sometimes, though, you need to ensure you have enough data from less common categories for a valid analysis. In our example, 10 feature stories might not be enough. This is where stratified sampling comes in. You might decide to collect an equal number of texts from each category, say 30 news, 30 opinion, and 30 features. This oversamples the rarer categories, giving you more data to work with. Later, you can weight the results during analysis to reflect the real-world proportions.

The choice between these methods depends on your research goals. Do you want a perfect mirror of the language variety, or do you need to zoom in on its less common components?

Representativeness isn't just about which texts you select. It's also about how you sample the language within those texts. Linguist Douglas Biber's 1993 framework highlights this. Just grabbing the first paragraph of every news article would give you a skewed view, since journalistic leads have a very specific structure. A good sample needs to represent the entire text.

To achieve this, you might take several smaller samples from the beginning, middle, and end of each document. These sampling units could be fixed-length chunks of words (e.g., 500-word segments) or structural units like paragraphs. The key is to avoid bias by sampling systematically across the whole text.

There's a constant trade-off between corpus size and quality. A massive, but poorly sampled, corpus is often less useful than a smaller, meticulously balanced one. A representative 1-million-word corpus will yield more reliable findings than a biased 100-million-word corpus.

Raw text collected from the real world is messy. This “noisy” data can corrupt your analysis if you don't remove it. Data cleaning is the process of stripping out irrelevant information to preserve the integrity of your research.

What counts as noise? It can be anything that isn't part of the core language you're studying.

• Metadata: Timestamps, author names, or headlines.
• Formatting Artifacts: Remnants of HTML tags like <div> or <img>, or extra whitespace.
• Boilerplate Text: Navigation menus from a website, copyright notices, or cookie consent banners.
• Transcription Errors: Mistakes from optical character recognition (OCR) or automated speech-to-text services.

With a well-defined target, a strategic sampling method, and a clean dataset, you have a powerful resource for exploring how language really works.