Questions: Computational and Digital Methods in History

Question 1

A historian uses topic modeling on digitized 19th-century American newspapers and finds a sharp spike in a 'finance and panic' topic cluster around 1857. What is the appropriate next step?

• A. Publish the finding as evidence that financial anxiety significantly increased in 1857
• B. Conclude the algorithm has identified a cause of the Panic of 1857
• C. Use the pattern as a hypothesis to investigate with traditional archival research and close reading
• D. Expand the corpus to non-English newspapers to confirm the finding before drawing any conclusions

Answer: (C)

Topic modeling detects statistical co-occurrence patterns — it does not explain why a cluster appeared, what it meant to contemporaries, or whether it reflects a genuine historical shift or an artifact of corpus composition. The spike is a hypothesis-generating finding, not a conclusion. It gives the historian a precise question: why does this discourse cluster peak here, and what does it mean? That question must be answered through archival research, close reading of the flagged texts, and attention to what was happening in 1857 — not through the algorithm alone. Computational methods identify patterns; historians interpret them.

Question 2

A historian uses network analysis on digitized correspondence archives and finds a particular 18th-century intellectual was highly 'central' in the Republic of Letters. What is the most significant methodological concern?

• A. Network centrality measures cannot distinguish between single letters and sustained correspondence
• B. The findings depend on which correspondence collections were digitized, and systematic biases in digitization may distort the network structure
• C. Centrality measures were developed for social media networks and are not appropriate for historical data
• D. The historian's own prior knowledge about the figure may have influenced which archives were selected

Answer: (B)

This is the corpus bias problem operating at scale. If digitization has systematically favored certain languages, institutions, or social positions, then the network reflects not the actual historical network but the network of what was preserved and scanned. An intellectual whose correspondence was archived by a major university may appear central simply because their letters survived and were digitized; a figure who communicated extensively but whose letters are in private or undigitized collections will appear peripheral. The algorithm faithfully analyzes the data it receives — the problem is that the data is not a neutral sample of historical reality.

Question 3

Topic modeling algorithms identify historically meaningful topics because they are specifically designed to detect semantic coherence in historical language.

• T. True
• F. False

Answer: False

Topic modeling algorithms detect statistical co-occurrence — words that appear together frequently across documents. The algorithm has no knowledge of historical meaning, context, or intentionality. What it produces is a list of co-occurring word clusters; the historian must then interpret whether and why these clusters represent meaningful historical topics. This interpretation is where the displaced interpretive labor sits. A cluster might reflect a genuine historical discourse, a genre convention, a scribal abbreviation, or a digitization artifact — the algorithm cannot distinguish among these possibilities.

Question 4

Digitization of historical sources is a neutral process that does not introduce systematic bias into computational historical analysis.

• T. True
• F. False

Answer: False

Digitization is highly selective and reflects existing power structures: printed books have been more digitized than manuscripts, English more than other languages, state and institutional archives more than community or vernacular records. These selection biases are not random — they systematically underrepresent the perspectives of those with less institutional power. When computational methods are applied to biased corpora, they produce systematic biases at scale: a topic model of 'American thought' based on published books will not represent the range of voices that constituted American thought. Transparency about what was included and excluded is an ethical requirement of computational history.

Question 5

Explain what it means to say that computational methods 'displace rather than replace' interpretive judgment in historical research.

Model answer: When reading documents by hand, interpretive choices are visible — the historian chose these sources and made these connections. When an algorithm processes thousands of documents, the interpretive decisions are embedded in technical choices (corpus assembly, algorithm selection, parameter settings) and in how output is interpreted. Interpretive labor shifts to a less visible location, but it does not disappear.

This matters because computational output can feel more authoritative than handmade arguments — the algorithm processed 100,000 documents, after all. But the decision about what corpus to assemble, which algorithm to apply, what parameters to set, and what the resulting patterns mean historically all reflect human judgment at every step. The best computational history makes these choices explicit and visible to readers, showing what was included and excluded and why the patterns are being interpreted as meaningful rather than artifactual. That transparency is what distinguishes rigorous computational history from scientism dressed up in historical garb.