Nine investigations into how American newspapers narrated catastrophe, racial violence, war, and reform between 1898 and 1933 — each built from one digitized archive and one contextual-measurement method, explained first.
This booklet collects nine studies of the early-twentieth-century American press. They ask a common question in different settings — how did newspapers turn an event into public meaning? — and answer it with a common apparatus. Chapter 1 explains that apparatus: the archive the articles are drawn from, and the method used to measure them. The nine investigations follow, grouped into three parts. Each is a standalone article; this front matter binds them into one reading and one method.
American Stories (dell-research-harvard/AmericanStories on Hugging Face) is article-level text extracted from nearly 20 million Chronicling America newspaper scans held by the Library of Congress. Rather than a single page image, each row is one detected article — a headline, a byline where present, and the OCR'd body text — recovered by learned layout-recognition and optical-character-recognition models built by the Dell Research group. The supported range is 1770–1964 (with gaps in the eighteenth century), stored as one compressed archive per year; a single year runs roughly 6 GB.
An article record carries the newspaper's identity, edition, date, page, headline, byline, and OCR text; the underlying scan records additionally retain Library of Congress metadata and the page's layout regions. That is enough to place a story in time, on a page, and — through the newspaper's LoC entry — in a state.
It is not a census of American opinion. The corpus is the public-domain papers Chronicling America chose to digitize, then read by imperfect models. Coverage varies by state, newspaper, and date; OCR introduces noise; segmentation can split or merge stories; and a newspaper's claims may themselves be false or deliberately misleading. Every investigation is written around those limits rather than against them.
The frames in these articles are measured with GABRIEL, an open-source framework for contextual measurement with language models. For a given study it takes a codebook of constructs — interpretive frames such as "employer responsibility" or "white-supremacy endorsement" — and asks a language model to score, on a 0–100 salience scale, how strongly each article foregrounds each frame, returning a supporting evidence quote for every judgment. Scoring runs on GPT-4o mini at temperature 0 through OpenAI's Batch API, which makes a full event cheap to measure (a 30-article, two-model feasibility pilot produced 120 valid structured judgments for about six cents; a full event runs on the order of ten dollars).
The load-bearing discipline is that relevance is measured separately from framing. Whether an article is even about the event is decided by its own relevance prompt and threshold; it is never inferred from the construct scores. A frame score of 0–100 is a claim about textual emphasis, not about truth or author intent.
Turning a raw year-archive into an investigation follows the same seven stages every time:
Three conventions recur across the nine chapters. A salience score (0–100) is the average visible emphasis of a frame, not its truth. Cluster weighting gives one vote to a story, not one vote to every reprint of it — so the article- vs. cluster-weighted gap itself measures what the wire amplified. And an attention line is always a share of the full detected-article denominator, so it tracks coverage rather than how much of that year happens to be digitized.
The nine studies are grouped into three parts. Order within each part is roughly chronological; the grouping is thematic, and easy to re-sort. Each entry links to the full self-contained article.
Written by Vishal Singh (NYU Stern School of Business). Each investigation is a single self-contained HTML file — data, D3 charts, and typography inlined, with no external requests — so any chapter can be hosted, moved, or read on its own. This front matter is the only page that links out to them. The underlying parquet slices, aggregation SQL, validation gates, and data cards live in the companion hf-stories repository.
Corpus: American Stories (Chronicling America / Library of Congress). Contextual measurement: GABRIEL with GPT-4o mini at temperature 0, relevance screened separately from construct scoring, estimates reprint-cluster weighted. Method detail: What American Stories Can Represent.
American Stories and these articles are released under CC BY 4.0.