跳转至

Human-Computer Interaction(CHI/CSCW)学科论文的语言风格分析

——基于 Zotero「各学科 Classic Papers / Human Computer Interaction」子库 94 篇原文的语料库分析(2003–2025)


0. 语料与方法

0.1 语料来源

Zotero 库「各学科 Classic Papers → Human Computer Interaction」子库共 96 个条目,94 篇入语料(2 篇 PDF 为纯中文翻译版剔除:Capel & Brereton 2023 HCAI 地图、Stack Overflow vs ChatGPT 一文;另剔除水印/目录段落)。94 篇、81.8 万词,四学科中规模第二

维度 数值
论文数 / 词数 94 篇 / 817,868 词
主要来源 CHI(约 70 篇)、CSCW/PACMHCI、UbiComp/TOCHI 等 ACM 系
年份跨度 2003–2025(近六成集中在 2015–2023)
篇均长度 8,701 词
引用样式 数字方括号 [n]:90.1 次/万词,94/94 篇全部使用;author-date 仅 0.5 次/万词

内容构成:经典实证 HCI(打字、中断、邮件、多任务)+ 批判/理论 HCI(Feminist HCI、Postcolonial Computing、Critical Race Theory for HCI、probes 系列)+ 2018 年后的 AI/HCAI 浪潮(data cascades、explainability、GAI 协作)。这条"实证—批判—AI"的内部光谱本身就是报告的重要背景:HCI 语料是四学科中内部风格方差最大的

0.2 分析方法

同 Dourish 管线:版面块段落提取 → 清洗 → 词频/n-gram/keyness(vs 其余三学科)→ 修辞动块统计 → 精读归类 → 引文脚本核验。


1. 总体节奏:中庸句长 + 高频段落实换

指标 HCI 四学科排位
平均句长 23.7 词 第 2
句长 P50 / P90 20 / 39
长句占比(>30 词) 23.0% 第 2
短句占比(<10 词) 16.0% 第 2 低
段落中位长度 91 词 最短(SOC 139、BDS 165;部分系 CHI 双栏版式切分,但短段落也是真实文风)

HCI 的节奏感来自段落切换频率而非句子内部:引言立 scene → 相关工作铺陈 → 方法逐小节 → 发现逐条 → discussion 升华,标准 CHI 模板把文章切成很碎的节。句长居中:它比 SE 讲究铺陈(引用堆叠、从句),但没有 BDS 的理论长句负担。

2. 人称与语态:最热闹的"我们"

指标 HCI BDS SOC SE
we /万词 91.5 37.8 52.2 103.0
our /万词 34.4 13.4 18.1 40.5
I /万词 16.8* 11.6 15.8 8.8
be+V-ed 被动式 /万词 72.1 83.5 64.7 82.2

I 的高频有相当部分来自访谈引语*——HCI 大量直接引用参与者原话("I always assumed that I wasn't really that close to [her]" 直接进了标题)。

  • "This paper" 自指 82/94 篇("this article" 仅 5 篇)——会议文体指纹。
  • "we" + 动词的搭配表勾勒出全部研究动作:we found (218) > we used (134) > we conducted (100) > we present (85) > we describe (64) ≈ we argue (64) > we observe (48) > we report (47)。注意 we argue 与 we found 几乎并驾齐驱——实证与论证双轮驱动是 HCI 区别于纯 SE 的特征。

3. 词汇场与引用谱系

3.1 概念词汇场(keyness vs 其余三学科)

HCI 独有词场极为清晰:human-AI / human-centered (127/M) / XAI (202/M) / high-stakes / LLM / prompt / sensemaking / data cascades / folk theories / trust & reliance / cognitive load;老一批经典则贡献 probes (319/M,几乎独有)、affordances (3.5/万词,四学科最高)、seamful、inspirational bits、empathy、experience。批判支线的词场与 BDS 高度重叠:feminism (117/M)、intersectionality、critical design、misgendering、postcolonial——"critique/critical" 出现于 69/94 篇,密度 6.5/万词与 BDS (7.0) 几乎持平,这解释了为什么 Dourish 会被两边同时引用。

3.2 引用谱系(人名 keyness)

Gaver、Sengers、Bardzell & Bardzell、Stolterman、Dourish、Suchman 谱系(Harrison/Tatar/ Sengers 的 meta 常客)、Irani、Ehsan、Bansal(HITL-AI)、Weisz、Liao、Sambasivan。HCI 的谱系是"设计研究 + 批判理论 + 认知科学"三线并存:引 Hutchinson 与引 Heidegger 的可能是同一篇文章。

引用格式上 ACM 数字引用 [44] 高密度内嵌,造成 HCI 正文里每句话都可能挂 1–3 个方括号——"文献堆叠句"成为其视觉签名:

"A substantial body of research in human-computer interaction has focused on interruption and attention management on the computer [1, 3, 4, 6, 8, 9, 13, 14, 16, 18]." —— Jin & Dabbish (2009), Self-interruption on the computer, Introduction

4. 高频短语与签名句式

模式 HCI(次/万词) 排位
hedge: may / might / suggest 16.6 / 7.6 / 8.8 三项全部四学科第一
not simply...but 1.5 第 2(BDS 2.9)
rather than 3.5 第 2
存在句 there is/are 10.0 与 SOC/SE 持平
p< 检验报告 5.4 第一(SE 1.5、SOC 1.4)
  • 模糊限制语全学科最密(may/might/suggest 三项第一):HCI 的知识主张永远留有余地——"may help"、"suggests that",因为它的证据是 12–30 人的用户研究,而它的话语对象是要据此做设计的社群。
  • "In this paper, we present..." 是全库最高频的论文开场公式("In this paper" 193 次于 71 篇;"we present" 34/94 篇)。摘要第一句直接报菜名:

"We present the notion of 'bridging concepts' as a particular form of intermediary knowledge in HCI research, residing between theory and practice. We argue that bridging concepts address the challenge of facilitating exchange between theory and practice in HCI..." —— Dalsgaard & Dindler (2011), Between theory and practice: bridging concepts in HCI research, Abstract

"We present a qualitative study of mobile communication via WeChat in Southern China, focusing on the rapid proliferation of emoji and stickers and the lessening dependence on text." —— Zhou, Hentschel & Kumar (2017), Goodbye text, hello emoji, Abstract

5. 修辞动块:CHI 论文的仪式清单

5.1 "revisit" 姿态:对经典温柔造反

HCI 实证文爱用 revisit / revisit-the-present-understanding 框架,把新数据放进老问题:

"This paper revisits the present understanding of typing, which originates mostly from studies of trained typists using the ten-finger touch typing system. Our goal is to characterise the majority of present-day users who are untrained and employ diverse, self-taught techniques." —— Feit, Weir & Oulasvirta (2016), How We Type, Abstract

"We are concerned that the current understanding mostly originates from an era when typing was much more homogenous than today." —— 同文, Introduction

5.2 参与者与伦理的仪式化细节

HCI 对"人"的登记是四学科中最细的:招募渠道、报酬金额、母语、利手都有:

"We recruited participants through a combination of developer communities, distribution lists, professional networks, and personal contacts, using snowball and purposive sampling [89] that was iterative until saturation." —— Sambasivan et al. (2021), "everyone wants to do the model work, not the data work", §Method

"Each participant received a thank you gift in the form of a gift card, with amounts localised in consultation with regional experts (100 USD for the US, 27 USD for India, 35 USD for East and West African countries)." —— 同文(注意:连报酬的"区域公平性"都要交代——这是 CHI 伦理审查文化的语言痕迹)

量化上:participants/subjects 合计 24.7 次/万词,四学科第一;"we recruited / IRB / informed consent / compensation" 动块 14+ 篇高频出现。SE 论文也做访谈,但极少这样逐项登记"人"的属性。

5.3 三角验证叙事(triangulation narration)

"先访谈、再问卷、再部署"的顺序句法:

"Second, we conducted a survey to triangulate our interview and field study results. Within the survey, we asked participants in a short text response if there were any pain points we had missed: no new pain points were identified through these responses." —— Chattopadhyay et al. (2020), What's Wrong with Computational Notebooks?

5.4 "Design implications" 作为贡献出口

实证发现必须折算成对设计者的建议,这是 HCI 独有的贡献货币(design implications 13/94 篇、implications for design 5/94 篇,其他学科近乎为零):

"Yet one of the most valorized outcomes of scientific research in HCI is its implications for design. But design is an intervention, an intentional effort to create change." —— Bardzell & Bardzell (2011), Towards a feminist HCI methodology(批判支线则反过来审视这个格式本身)

5.5 开场钩子:新闻语 / 参与者语录 / 反问

  • 新闻钩子:WeChat 论文以 Emojicon 新闻稿开篇("Emoji-mania is in full force...")再引出研究问题;
  • 参与者语录直接做标题:"everyone wants to do the model work, not the data work""I always assumed that I wasn't really that close to [her]""'It's Reducing a Human Being to a Percentage'"
  • 问题驱动开场:"We need to better understand the phenomenon of self-interruption, because it is a driver of multitasking behavior and contributor to fragmented attention."(Dabbish et al. 2011)

6. 标题风格:四学科最会玩标题的

特征 比例(n=96) 四学科对比
含冒号 55% 次高
带引号 19% 第一
疑问/How-Why 开头 24% 第一(4% 完整问句 + 20% how/why/what 开头)
>10 词 52% 次长
动词/分词开头 14%

样例即可见其文娱精神: - 文学戏仿/俗语改写Why Johnny Can't Prompt(2023);Old wine in new bottles or novel challenges(2011);He says, she says(2007);If not now, when?(2004) - 交错配列Goodbye text, hello emoji(2017);Is seeing believing?(Cosley et al.);Gender Recognition or Gender Reductionism?(2018) - 第一人称自白Am I wasting my time organizing email?I did that!Why do I keep interrupting myself?First I "like" it, then I hide it - 参与者原话"everyone wants to do the model work, not the data work" - 大词自问:What is interaction?(Hornbæk & Oulasvirta 2017);What is "critical" about critical design?What is Human-Centered about Human-Centered AI?

CHI 标题允许出现任何其他工科会议都会毙掉的东西:第一人称、感叹号、双关、参与者脏话级别的口语。标题是 CHI 论文的"第二摘要",兼做情绪动员。

7. 风格画像总结

HCI 的声音:一个兴奋的工坊主持人。 "We" 密度极高(91.5/万词)且动词丰富(found/conducted/present/describe/argue);摘要以 "We present..." 开门见山;证据永远是"我们做了什么"的流水叙事 + 三角验证;对参与者的登记细致到报酬与利手;hedges 三项全学科第一(对小样本永远诚实);p 值与引语并陈,统计和故事各占一半;发现必须折算成 design implications;标题敢用第一人称、引号和俗语。内部同时存在一条批判支线(Bardzell/Irani/Keyes/Dourish 系),其词场与句式(not X but Y、scare quotes)直接继承自 BDS 风格家族。

若要模仿该文体(写作配方): 1. 摘要第一句:"We present/describe a study of X...";正文自指一律 "this paper"; 2. 方法节按"参与者→程序→分析"分小节,逐项交代 N、招募渠道、报酬、伦理审查; 3. 发现用 "We found that... (P1)" + 参与者原话双轨呈现,引语保真不清洗; 4. 每个强主张配一个 hedge(may/suggest),每个统计结论给出 p 值; 5. Discussion 末尾必设 "Design Implications" 小节,把发现翻译给设计者; 6. 标题冒号前放一个钩子(问句/引语/俗语改写),冒号后放研究内容; 7. 引用用 [n] 数字堆叠;相关工作节按主题分组而非逐篇罗列。


附录 A:局限说明

  1. 2 篇仅有中文翻译版 PDF(HCAI 地图 2023、SO vs ChatGPT 2023)已剔除;Zotero 条目仍在库中。
  2. CHI 双栏版式使段落统计偏碎(中位 91 词含切分伪差),句子级指标不受影响。
  3. 子库内部异质性大(经典实证 vs 批判理论 vs HCAI),本文按主流趋势概括,批判支线单独标注。
  4. 水印(ACM 版权页)、卷首作者地址块已过滤;连字丢失已归并。

附录 B:语料构成(94 篇节选代表 + 分布)

  • 经典实证:Adamczyk & Bailey 2004(中断时机);Jin & Dabbish 2009(自我中断);Mark, Gudith & Klocke 2008(中断代价);Feit et al. 2016(How We Type);Whittaker et al. 2011(email refinding);Kittur et al.(Wikipedia×2、Mechanical Turk×2)
  • 批判/理论:Bardzell 2010(Feminist HCI);Bardzell & Bardzell 2011(feminist methodology)+ 2012(critical design);Irani et al. 2010(postcolonial);Irani & Silberman 2013(Turkopticon);Boehner et al. 2007(probes);Hornbæk & Oulasvirta 2017(What is interaction?);Ogbonnaya-Ogburu et al. 2020(CRT for HCI);Keyes 2018(Misgendering Machines)
  • 社计算/GAM:Burke, Kraut & Marlow 2011;Bernstein et al. 2013;Eslami et al. 2015/2016(folk theories);Starbird et al. 2018(disinformation as work)
  • AI/HCAI 浪潮:Amershi et al. 2019(Guidelines);Sambasivan et al. 2021(data cascades);Bansal et al. 2021;Buçinca et al. 2021;Ehsan et al. 2021;Chen et al. 2023;Zamfirescu-Pereira et al. 2023(Why Johnny Can't Prompt) (完整 94 篇键值、年份、词数见 discipline_style_analysis/corpus_meta.json。)