Writing like Dourish: a 21-text corpus analysis of a critical HCI voice (2004–2026)¶
🌐 Language / 语言：中文 · English
Provenance（来源与元数据）
| id | marginalia-002 |
| title | Writing like Dourish: a 21-text corpus analysis of a critical HCI voice (2004–2026) |
| date | 2026-08-15 |
| published | 2026-08-15 |
| kind | analysis |
| issue | 3 |
Writing like Dourish¶
Can a scholarly voice be measured? Twenty-one Dourish texts and ~408k words later: yes — a signature you can count, stable for two decades.
The musing¶
Style advice is usually vibes. I wanted the opposite: take one of critical HCI's most distinctive prose voices — Paul Dourish — and measure it from primary texts, with a pipeline where every quoted sentence is machine-verified against the corpus.
Corpus¶
- 20 texts pulled from my local Zotero (author search over ~5.6k journal articles; bilingual editions excluded; all with PDF attachments), 2004–2026, including two MIT Press monographs; ~324k words after cleaning.
- +1: The Stuff of Bits (2017), extracted from EPUB at the author's implicit invitation of completeness — 584 paragraphs, ~83.5k words.
- Control: 25 papers from other fields, same era, ~393k words — the baseline for keyness.
Method¶
PyMuPDF layout-block extraction → reference stripping → paragraph corpus (paras.json) → word frequencies and 2–5-gram document frequencies → keyness by log-ratio against the control corpus → 60+ rhetorical-pattern regexes → paragraph TF-IDF clustering into 8 recurring motifs (M1–M8) → verify_quotes.py re-finding every quote cited in the report (ligatures/smart-quote normalized) and marking PASS/FAIL. A separate module scores The Stuff of Bits signatures (sob_stats.py: ~30 signature densities vs. baseline; sob_sim.py: paragraph-level TF-IDF similarity).
The measurable signature¶
- Lexicon: practice ×1,041, data ×1,022 lead the corpus.
- Negation-restatement: "not simply/just/only/merely … but" 298× across 18 texts, "rather than" 267× across 19 — roughly one every 350 words.
- "ways in which" ×252 (14 texts); quantified-plural constructions ×697; "that is / in other words" ≈ ×172.
- Rhetorical questions: about one every 200 words in solo theory texts.
- The same syntactic template — not a property of X, but an achievement of interaction — travels from context (2004) through emotion and data to beauty, two decades apart; sentence-level self-recycling reaches 0.97 TF-IDF similarity.
- Stuff of Bits signatures: "that is" 2.45× baseline; materiality 14.98×; everyday 0.12×.
Two specimens (both 2004, verified in-corpus):
"what I want to do here is to reconsider context, not as a representational problem but as an interactional problem."
"Embodiment is not a property of systems, technologies, or artifacts; it is a property of interaction."
What it produced¶
- The report —
Dourish_语言特征分析报告.md(~40KB), every quote verified. - A reusable pipeline — scripts + JSON intermediates + 12 motif quote files.
- A
dourish-styleskill — polishing workflow with Register A/B and recipes P1–P10; red line: never copy Dourish verbatim, only reproduce the mechanism. Published as an artifact: artifact.en.md. - A first experiment — my own OSS position paper revised in that register (with a change log).
Provenance¶
| field | value |
|---|---|
| Data | Zotero (zotero.sqlite, API :23119) → zotero_copy.sqlite (201MB); dourish_analysis/; skill at ~/.zcode/skills/dourish-style/ |
| Initial prompt | "Search my Zotero for Paul Dourish's papers and analyze his language/style features — quantitatively, and verifiably." |
| Time | analysis 2026-08-15, 20:56 → 21:35 · note published 2026-08-15 |
| Agent / model | ZCode CLI · GLM (Zhipu) |
| Issue | #3 |
🌐 阅读中文版