Software Engineering classics (314 papers, 3.3M words)¶

Provenance（来源与元数据）

id	marginalia-005
title	Four academic voices, measured: the language styles of Big Data & Society / HCI / Sociology / Software Engineering classics (314 papers, 3.3M words)
date	2026-08-15
published	2026-08-15
kind	analysis
issue	11

Four academic voices¶

After measuring one author's voice (002: Dourish), measure a discipline: same Zotero library, same pipeline, four fields of classic papers. Can stylistic differences be counted? Yes — and each discipline turns out to run its own conceptual machine.

Musing¶

Writing advice is usually given per author ("write like X"), but academic writing is shaped by disciplinary ritual: submit to a venue, inherit its syntax and move-set. I wanted to test how measurable this is — not the impressionistic "SE papers are dry, sociology papers are winding", but densities, rankings, and machine-verifiable example sentences. This is also the sequel to 002: Dourish's signature construction runs at 9.2 per 10k words. What does the average paper in his (or neighbouring) fields look like?

Corpus¶

Four sub-collections of my Zotero "Classic Papers by Discipline" library — 347 items, 327 with PDFs; after cleaning (watermarks out, Chinese-translation blocks out, hopeless OCR out, monographs out): 314 papers / 3.3M words.

Discipline	Papers	Words	Home venues	Years
Big Data & Society	27	211,858	Big Data & Society, New Media & Society	2004–2021
Human-Computer Interaction	94	817,868	CHI, CSCW (ACM)	2003–2025
Sociology	79	1,105,574	AJS, ASR, BJS	1975–2026
Software Engineering	114	1,166,765	ICSE, FSE, TSE, MSR	1987–2023

Method¶

Same pipeline as 002: PyMuPDF block-level paragraph extraction → drop references/watermarks/CJK blocks → word frequencies and 2–5-gram document frequencies → log-odds keyness (discipline vs the pooled other three) → 60+ rhetorical-move regexes (RQ lists, threats-to-validity, we+verb collocations, the epistemology/power lexicon…) → title statistics → close reading. All 39 quotes were machine-verified by verify_quotes.py (tolerant of ligature loss, line-broken words, and footnote digits).

The four voices¶

Big Data & Society: the long-sentence public intellectual. Mean sentence 26.8 words; densest semicolons (46/10k); zero contractions across the whole library; the not X but Y / rather than refutation pattern ranks first — restate the mainstream understanding, then flip it; highest passive rate (agency suspended — precisely how the critiqued mechanisms operate); self-refers as "This article", opens with manifesto declarations ("Data are a form of power."), closes with an exhortation to the discipline; titles love the "The X of Y" concept lever.

HCI: the excited workshop host. "We present…" straight out of the gate; the richest we-verb repertoire (found/conducted/present/describe/argue); all three hedging markers rank first (may/might/suggest) — permanently honest about small samples; participant registration down to compensation amounts and handedness; interview quotes and p-values side by side; findings must be cashed out as design implications; the most playful titles of the four (first person, parody, participants' own words as titles).

Sociology: the theoretical statistician in dialogue with both Weber and the GSS. Longest papers (14k words); em-dashes, contractions, and scare quotes all rank first — the loosest prose texture; lowest passivity (the researcher is always actively operating on data); the evidence syntax is "consistent with / net of / Model 1 → Model 3"; juxtaposed epigraph openings (a Weber passage against a campaign attack ad); and a journal ritual found nowhere else — AJS abstracts forced into third-person "The authors" while the body says "we" throughout.

Software Engineering: the engineer who writes in lists. Median sentence just 16 words; em-dashes all but banned; we/our both rank first but "I" ranks last; verbs are all tool-verbs (use/found/present/describe); the noun-field is countable objects (bugs/commits/repositories); the heaviest fixed liturgy — "To understand X" opener → bulleted contributions → numbered RQs → a Threats-to-Validity confessional in three parts (construct/internal/external) → future work (70% of papers, guaranteed).

Three findings I like most¶

One pronoun betrays your genre: journal papers self-refer as "this article" (BDS 21/27, SOC 53/79), conference papers as "this paper" (HCI 82/94, SE 103/114). One glance at the self-reference tells you which register you're in.
Dourish is still an outlier: even BDS, the closest discipline, averages only ~1/3 of his personal density of the not-simply construction (2.9 vs 9.2 per 10k). "Writing like the discipline" and "writing like Dourish" are different calibration targets.
Style is a spectrum, not a border: critical HCI (Bardzell/Irani/Keyes/Dourish line) forms a BDS-style enclave inside the HCI library (critique density 6.5 vs BDS 7.0); Zeller and Hindle are the rhetorical defectors inside SE; computational sociology is drifting toward SE's empirical liturgy.

Cheat sheet¶

	Syntax signature	Person	Evidence	Ending
BDS	not X but Y / rather than / increasingly	This article + I	author-date sparring	exhortation
HCI	We present / may-suggest / quotes+stats	we (the workshop)	N+pay+p-values+quotes	design implications
SOC	consistent with / net of / in other words	we (abstract: the authors)	model ladders vs hypotheses	limitations
SE	To understand X / a set of / 16-word sentences	we+our (the team)	dataset scale + tables	Threats → Future Work

Deliverables & limits¶

Five full reports (Chinese, published verbatim): BDS · HCI · Sociology · SE · cross-discipline comparison
The reusable pipeline stays local at ZCodeProject/discipline_style_analysis/ (resolve → extract → metrics → keyness → clean → quote-verify); paras.json keys are discipline_ZoteroKey.
Limits: the sub-collections are a hand-curated "my classics" list, not a random sample; SE lacks 19 papers that exist only as IEEE web snapshots; SOC includes 1970s–80s classics, so part of the difference is era, not field; paragraph-level statistics are affected by single/double-column layouts (sentence-level metrics are not).

Possible next steps: distill the four portraits into four polishing skills (mirroring 002's dourish-style), or turn the keyness word lists into a "discipline camouflage detector" — give it a passage, it tells you which field it most sounds like.

🌐 阅读中文版