Step 1 · Keyword Search

The first decision in any corpus project is what to leave out. A model can read a few hundred documents attentively; it cannot read three hundred thousand. A keyword search is the crudest tool in text analysis, but used at this stage it does exactly one job well: it defines the subject of the study, leaving everything about how that subject is described to be discovered in Step 2.

Section 1The nkmedia corpus

The running example uses the lab's nkmedia database (version 1, July 2026): every article published online by the Korean Central News Agency (KCNA) and the party daily Rodong Sinmun between 1998 and October 2025. Rodong Sinmun articles that KCNA also carried are kept once, under a separate category.

At 168 million characters the corpus is roughly 40 million tokens: far beyond any model's context window, and beyond anything a person could skim. Even the 2012–2025 portion is about 165,000 articles. Something has to narrow it.

Section 2Why restrict by keyword

There are three ways to shrink a corpus: sample it at random, filter it by metadata (dates, sections, authors), or filter it by content. Random sampling keeps everything in proportion, which is the problem: a topic that occupies 0.3 percent of a newspaper occupies 0.3 percent of the sample. Metadata filters are precise but rarely track a research subject. A content filter does, and a keyword in the title is the most defensible content filter available before you know what you are looking for, because a newspaper's own editors have already judged the article to be about that thing.

The keyword should name the subject of the research, not a hypothesis about it. Searching for 자력갱생 (self-reliance) or 제재 (sanctions) would pre-select the frames the study might later want to measure. Searching for 기업소 (enterprise) selects the organisational unit and says nothing about how it is portrayed. Three further tests, applied in order:

1. 1Is it small enough to read?The target is a corpus a model can read in a handful of passes and a person can spot-check: a few hundred to low thousands of documents, under about a million characters.
2. 2Is it diverse enough to be interesting?Check the distribution across years and sources. A keyword that only fires in two years, or only in one newspaper, will give a dataset that cannot speak to change.
3. 3Does it miss too much?Compare title matches with body matches. If the term appears in the body of ten times as many articles, a title filter is strict; decide whether that strictness is acceptable for the subject.

Section 3Choosing the keyword

Table 1.2 shows the candidates considered for the workshop, with the number of articles each would select. Counts are for the whole corpus (body and title) and for titles in 2012–2025, the Kim Jong Un era, which is the period the running example studies.

The workshop uses 기업소 in the title, 2012–2025. It names an organisational unit, so it is neutral about framing. At 946 articles and 549,000 characters it is readable in full by a model and checkable by a person in an afternoon. It spans every year and both newspapers. And it excludes a lot: 공장 (factory) titles would roughly triple the corpus and 현지지도 (on-site guidance) would pre-select leadership visits, which is one of the things the study wants to measure rather than assume.

Section 4The workshop corpus

The restricted corpus has 946 articles. Table 1.3 shows how they fall across years; the distribution is uneven and that unevenness is itself a finding to carry into Step 2.

The corpus file keeps the nkmedia columns and adds one: doc_id, a short stable identifier (NKE-0001 to NKE-0946, in date order). Every later file (memos, codebook examples, dataset rows, validation samples) refers to articles by this identifier, so that a claim about an article can always be traced back to its text.

Section 5Download and set up your project folder

The corpus is distributed as an Excel workbook, zipped. Download the zip, create a project folder on your laptop, and unzip it there. Everything you produce in later steps goes in the same folder, so the AI assistant you use in Step 2 can see all of it at once.

• corpus_nk_enterprises_2012-2025.zipThe workshop corpus. Contains the .xlsx below.465 KB · zip
• corpus_nk_enterprises_2012-2025.xlsxSheet corpus: 946 rows, one per article. Sheet README: column definitions and provenance.468 KB · xlsx

A suggested layout. Keep the corpus file unchanged; everything else is derived from it.

nk-enterprises/
├── corpus/
│   └── corpus_nk_enterprises_2012-2025.xlsx   # Step 1 – never edited
├── exploration/                                # Step 2 – the model's memos
├── codebook/                                   # Step 3 – codebook_v2.xlsx, codebook.json
├── extraction/                                 # Step 4 – article batches, coded batches, dataset
├── validation/                                 # Step 5 – human samples, reports
└── analysis/                                   # tables and figures