{"id":"e0ee4318-4299-4943-8255-e0be4f6573b3","shortId":"krXXt2","kind":"skill","title":"humanize-chinese","tagline":"Detect and humanize AI-generated Chinese text. 20+ rule detection categories plus statistical features (sentence-length CV, short-sentence fraction, comma density, perplexity, GLTR, DivEye) plus scene-aware LR fusion (rule × 0.2 + LR × 0.8) trained on three scenes: general / acad","description":"# Humanize Chinese AI Text v5.0\n\n检测和改写中文 AI 生成文本的完整工具链。可独立运行（统一 CLI 或独立脚本），也可作为 LLM prompt 指南使用。\n\n**v5.0 亮点：** scene-aware 三路 LR 融合 (general / academic / longform)；HC3 fused 准确率 95%；新增 `--scene novel/auto` 长文本场景；新增 `--style novel` 长篇叙事；`--best-of-n N` (默认 10) 多次 humanize 取最低 LR；段落级信号 (paragraph length CV / 跨段 trigram 重复) + 反制改写。\n\n## CLI Tools\n\n### 统一 CLI（推荐）\n\n```bash\n./humanize detect 文本.txt -v                              # 检测 + 详细\n./humanize detect 章节.txt --scene novel                   # 长文本/小说显式 scene\n./humanize detect 稿件.txt --scene auto                    # 按长度自动切 (≥1500 走 longform LR)\n./humanize rewrite 文本.txt -o 改后.txt                    # 改写（默认 best-of-10）\n./humanize rewrite 文本.txt -o 改后.txt --best-of-n 5      # 自调 best-of-N\n./humanize rewrite 文本.txt -o 改后.txt --quick            # 快速模式（跳统计/best-of）\n./humanize academic 论文.txt -o 改后.txt --compare          # 学术降重 + 双评分对比\n./humanize style 章节.txt --style novel                    # 长篇叙事专属 (剔除 AI prompt artifact + markdown headers)\n./humanize style 文本.txt --style xiaohongshu              # 风格转换\n./humanize compare 文本.txt -a                              # 前后对比\n```\n\n### 独立脚本形式（等价）\n\n所有脚本在 `scripts/` 目录下，纯 Python，无依赖。\n\n```bash\n# 检测 AI 模式（20+ 规则维度 + 8 统计特征，0-100 分）\npython scripts/detect_cn.py text.txt\npython scripts/detect_cn.py text.txt -v          # 详细 + 最可疑句子\npython scripts/detect_cn.py text.txt -s           # 仅评分\npython scripts/detect_cn.py text.txt -j           # JSON 输出\n\n# 改写（默认 best-of-10，scene-aware）\npython scripts/humanize_cn.py text.txt -o clean.txt\npython scripts/humanize_cn.py text.txt --scene social -a   # 社交 + 激进\npython scripts/humanize_cn.py text.txt --quick             # 18× 速度，纯替换\npython scripts/humanize_cn.py text.txt --cilin             # 启用 CiLin 同义词扩展\n\n# 风格转换（先自动 humanize 再套风格）\npython scripts/style_cn.py text.txt --style zhihu -o out.txt\n\n# 前后对比\npython scripts/compare_cn.py text.txt --scene tech -a\n\n# 学术论文 AIGC 降重（10 学术维度 + scene-aware academic LR + 双评分）\npython scripts/academic_cn.py paper.txt -o clean.txt --compare\npython scripts/academic_cn.py paper.txt -o clean.txt -a --compare  # 激进\npython scripts/academic_cn.py paper.txt -o clean.txt --quick       # 快速模式\n```\n\n### 评分标准\n\n| 分数 | 等级 | 含义 |\n|------|------|------|\n| 0-24 | LOW | 基本像人写的 |\n| 25-49 | MEDIUM | 有些 AI 痕迹 |\n| 50-74 | HIGH | 大概率 AI 生成 |\n| 75-100 | VERY HIGH | 几乎确定是 AI |\n\n### 参数速查\n\n| 参数 | 说明 |\n|------|------|\n| `-v` | 详细模式，显示可疑句子 |\n| `-s` | 仅评分 |\n| `-j` | JSON 输出 |\n| `-o` | 输出文件 |\n| `-a` | 激进模式 |\n| `--seed N` | 固定随机种子 |\n| `--scene` | general / academic / novel / auto（detect_cn）—— auto 按 ≥1500 字切 longform LR |\n| `--style` | casual / zhihu / xiaohongshu / wechat / academic / literary / weibo / **novel** |\n| `--best-of-n N` | humanize N 次取 LR 最低（默认 10） |\n| `--compare` | 前后对比（学术双评分） |\n| `--quick` | 快速模式（跳过统计优化 + best-of，18× 速度） |\n| `--cilin` | 启用 CiLin 同义词扩展（humanize，38873 词，含碰撞 blacklist） |\n| `--no-humanize` | style 转换前不先去 AI 词 |\n| `--rule-only` | detect 只用规则层（跳 LR 融合） |\n\n### 工作流\n\n```bash\n# 1. 检测\n./humanize detect document.txt -v\n# 2. 改写 + 对比\n./humanize compare document.txt -a -o clean.txt\n# 3. 验证\n./humanize detect clean.txt -s\n# 4. 可选：转风格\n./humanize style clean.txt --style zhihu -o final.txt\n```\n\n### HC3-Chinese 基准测试\n\n阈值基于 [HC3-Chinese](https://github.com/Hello-SimpleAI/chatgpt-comparison-detection) 300+300 人类/AI 样本的 Cohen's d 校准，scene-aware LR 在 500+500 训练：\n\n- 句长变异系数 CV: d = 1.22（最强单信号）\n- 短句占比 (< 10 字): d = 1.21\n- 段落长度 CV: d = -1.49（v5 长文本新信号）\n- 段内句长 CV: d = -2.08（v5 长文本最强信号）\n- 跨段 trigram 重复: d = +1.13（v5 长文本新信号）\n- 困惑度: d = 0.47\n- GLTR top-10 bucket: d = 0.44\n- DivEye skew / kurt: d = 0.41 / 0.29\n- 逗号密度: d = -0.47\n\n**v5.0 100 样本回归测试 (HC3 fused)**：95% 正确分离率 / 55.0 分差距 / +40.5 平均降幅 / 100% 段落保留 / 0 grammar defect。\n\n**长文本 170 样本回归 (longform benchmark)**：gap 51.4 / avg delta +25.1 / 段留 98.8%。By genre：academic +13.8 / news +18.8 / blog +36.3 / novel +19.5 / review +31.8。\n\n**Hero 样本**（best-of-10, seed=42）：academic 100→35 (-65) / 通用 100→35 (-65) / 小红书 100→41 (-59) / 长篇博客 96→41 (-55) / 工作汇报 96→13 (-83)。\n\n---\n\n## LLM 直接使用指南\n\n当用户要求\"去 AI 味\"、\"降 AIGC\"、\"人性化文本\"、\"改成人话\"时，如果无法运行 CLI 工具，按以下流程手动处理。\n\n### 第一步：检测 AI 写作模式\n\n扫描文本中的以下模式，按严重程度分类：\n\n#### 🔴 高危模式（一眼就能看出是 AI）\n\n**三段式套路：**\n- 首先…其次…最后\n- 一方面…另一方面\n- 第一…第二…第三\n\n**机械连接词：**\n值得注意的是、综上所述、不难发现、总而言之、与此同时、由此可见、不仅如此、换句话说、更重要的是、不可否认、显而易见、不言而喻、归根结底\n\n**空洞宏大词：**\n赋能、闭环、数字化转型、协同增效、降本增效、深度融合、全方位、多维度、系统性、高质量发展、新质生产力\n\n#### 🟠 中危模式\n\n**AI 高频词：** 助力、彰显、凸显、底层逻辑、抓手、触达、沉淀、复盘、迭代、破圈、颠覆\n\n**填充废话：** 值得一提的是、众所周知、毫无疑问、具体来说、简而言之\n\n**模板句式：**\n- 随着…的不断发展\n- 在当今…时代\n- 在…的背景下\n- 作为…的重要组成部分\n- 这不仅…更是…\n\n**平衡论述套话：** 虽然…但是…同时、既有…也有…更有\n\n#### 🟡 低危模式\n\n- 犹豫语过多（在一定程度上、某种程度上 出现 >5 次）\n- 列举成瘾（动辄①②③④⑤）\n- 标点滥用（大量分号、破折号）\n- 修辞堆砌（排比对偶过多）\n\n#### ⚪ 风格信号\n\n- 段落长度高度一致\n- 句子长度单调\n- 情感表达平淡\n- 开头方式重复\n- 信息熵低（用词可预测）\n\n### 第二步：改写策略\n\n按以下顺序处理：\n\n**1. 砍掉三段式**\n把\"首先…其次…最后\"打散，用自然过渡代替。不是每个论点都要编号。\n\n**2. 替换 AI 套话**\n- 综上所述 → 总之 / 说到底 / （直接删掉）\n- 值得注意的是 → （直接删掉，后面的话自己能说清楚）\n- 赋能 → 帮助 / 支持 / 提升\n- 数字化转型 → 信息化改造 / 技术升级\n- 不难发现 → 可以看到 / （删掉）\n- 助力 → 帮 / 推动\n\n**3. 句式重组**\n- 过短的句子合并（\"他很累。他决定休息。\" → \"他累了，干脆歇会儿。\"）\n- 过长的句子拆开（在\"但是\"\"不过\"\"同时\"等转折处断开）\n- 打破均匀节奏（长短句交替，不要每句差不多长）\n\n**4. 减少重复用词**\n同一个词出现 3 次以上就换同义词。比如\"进行\"可以换成\"做\"\"搞\"\"开展\"\"着手\"。\n\n**5. 注入人味**\n- 加一两句口语化表达（场景允许的话）\n- 用具体的例子代替抽象概括\n- 偶尔加个反问或感叹\n- 不要每段都总分总结构\n\n**6. 段落节奏**\n打破每段差不多长的格局。有的段落 2 句话，有的 5 句话，像人写东西时自然的长短变化。\n\n### 第三步：学术论文特殊处理\n\n当文本是学术论文时，改写规则不同——不能口语化，要保持学术严谨性：\n\n**学术专用检测维度：**\n1. AI 学术措辞（\"本文旨在\"\"具有重要意义\"\"进行了深入分析\"）\n2. 被动句式过度（\"被广泛应用\"\"被认为是\"）\n3. 段落结构过于整齐（每段总-分-总）\n4. 连接词密度异常\n5. 同义表达匮乏（\"研究\"出现 8 次）\n6. 引用整合度低（每个引用都是\"XX（2020）指出…\"）\n7. 数据论述模板化（\"从表中可以看出\"）\n8. 过度列举（①②③④ 频繁出现）\n9. 结论过于圆满（只说好不说局限）\n10. 语气过于确定（\"必然\"\"毫无疑问\"）\n\n**学术改写策略：**\n\n- **替换 AI 学术套话（保持学术性）：**\n  - 本文旨在 → 本文尝试 / 本研究关注\n  - 具有重要意义 → 值得关注 / 有一定参考价值\n  - 研究表明 → 前人研究发现 / 已有文献显示 / 笔者观察到\n  - 进行了深入分析 → 做了初步探讨 / 展开了讨论\n  - 取得了显著成效 → 产生了一定效果 / 初见成效\n\n- **减少被动句：**\n  - 被广泛应用 → 得到较多运用 / 在多个领域有所应用\n  - 被认为是 → 通常被看作 / 一般认为\n\n- **注入学术犹豫语（hedging）：**\n  在过于绝对的判断前加\"可能\"\"在一定程度上\"\"就目前而言\"\"初步来看\"\n\n- **增强作者主体性：**\n  - 研究表明 → 笔者认为 / 本研究发现\n  - 可以认为 → 笔者倾向于认为\n\n- **补充局限性：**\n  如果结论段没有提到局限，补一句\"当然，本研究也存在一定局限…\"\n\n- **打破结构均匀度：**\n  调整段落长度，避免每段都一样。合并过短的段落，拆分过长的。\n\n### 第四步：验证\n\n改写完成后，用 CLI 工具验证效果：\n\n```bash\n./humanize detect output.txt -s\n```\n\n目标（基于 v5.0 fused 检测器，best-of-10 humanize）：\n- 通用文本降到 35 分以下（LOW 区间）\n- 学术论文降到 35 分以下（学术专用 + 通用评分均低）\n- 长篇博客/小说（≥1500 字）降到 41 分左右（MEDIUM）\n- 真实 ChatGPT 短输出 baseline 通常已在 5-25 分，改写后再降 3-10 分\n- 刻板化 AI 样板文 (论文模板/八股) 可以看到 60-83 分降幅\n\n注：v5.0 fused 评分融合了 LR ensemble (rule × 0.2 + LR × 0.8)，相同文本的分数会比 v3.x rule-only 更准。`--rule-only` 可降级到纯规则视图。\n\n---\n\n## 配置说明\n\n所有检测模式和替换规则在 `scripts/patterns_cn.json`，可自定义：\n- 添加新 AI 词汇\n- 调整权重\n- 增加替换规则\n- 修改正则匹配\n\n## 外部配置字段\n\n```\ncritical_patterns    — 高权重检测（三段式、连接词、空洞词）\nhigh_signal_patterns — 中权重检测（AI 高频词、模板句）\nreplacements         — 替换词库（正则 + 纯文本）\nacademic_patterns    — 学术专用检测与替换\nscoring              — 权重和阈值配置\n```","tags":["voidborne","humanize","chinese","awesome","agent","skills","for","empirical","research","brycewang-stanford","academic-research","agent-skills"],"capabilities":["skill","source-brycewang-stanford","skill-49-voidborne-d-humanize-chinese","topic-academic-research","topic-agent-skills","topic-ai-agent","topic-awesome-list","topic-communication","topic-copaper","topic-economics","topic-education","topic-empirical-research","topic-international-relations","topic-political-science","topic-psychology"],"categories":["Awesome-Agent-Skills-for-Empirical-Research"],"synonyms":[],"warnings":[],"endpointUrl":"https://skills.sh/brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research/49-voidborne-d-humanize-chinese","protocol":"skill","transport":"skills-sh","auth":{"type":"none","details":{"cli":"npx skills add brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research","source_repo":"https://github.com/brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research","install_from":"skills.sh"}},"qualityScore":"0.700","qualityRationale":"deterministic score 0.70 from registry signals: · indexed on github topic:agent-skills · 1032 github stars · SKILL.md body (6,686 chars)","verified":false,"liveness":"unknown","lastLivenessCheck":null,"agentReviews":{"count":0,"score_avg":null,"cost_usd_avg":null,"success_rate":null,"latency_p50_ms":null,"narrative_summary":null,"summary_updated_at":null},"enrichmentModel":"deterministic:skill-github:v1","enrichmentVersion":1,"enrichedAt":"2026-05-18T18:53:15.877Z","embedding":null,"createdAt":"2026-05-05T00:53:21.045Z","updatedAt":"2026-05-18T18:53:15.877Z","lastSeenAt":"2026-05-18T18:53:15.877Z","tsv":"'+1.13':541 '+13.8':593 '+18.8':595 '+19.5':599 '+25.1':587 '+300':499 '+31.8':601 '+36.3':597 '+40.5':571 '+500':513 '-0.47':561 '-1.49':528 '-10':549,979 '-100':233,362 '-2.08':534 '-24':346 '-25':975 '-49':350 '-55':625 '-59':621 '-65':613,617 '-74':356 '-83':629,988 '/ai':501 '/best-of':179 '/hello-simpleai/chatgpt-comparison-detection)':497 '/humanize':112,119,128,139,152,169,180,190,203,210,458,465,473,480,937 '0':232,345,575 '0.2':39,997 '0.29':558 '0.41':557 '0.44':552 '0.47':546 '0.8':41,999 '1':456,751,836 '1.21':524 '1.22':518 '10':93,151,260,312,418,521,607,875,949 '100':563,573,611,615,619 '13':628 '1500':135,394,963 '170':579 '18':281,428 '2':462,760,823,842 '20':12,228 '2020':864 '25':349 '3':471,784,803,846,978 '300':498 '35':612,616,952,957 '38873':435 '4':477,800,852 '41':620,624,966 '42':609 '5':163,732,812,826,854,974 '50':355 '500':512 '51.4':584 '55.0':569 '6':819,860 '60':987 '7':866 '75':361 '8':230,858,869 '9':872 '95':78,567 '96':623,627 '98.8':589 'acad':47 'academ':73,181,317,387,403,592,610,1038 'ai':8,50,54,198,226,353,359,366,444,634,647,653,690,762,837,881,982,1015,1031 'ai-gener':7 'aigc':310,637 'artifact':200 'auto':133,389,392 'avg':585 'awar':35,68,263,316,509 'baselin':972 'bash':111,224,455,936 'benchmark':582 'best':88,149,160,166,258,408,426,605,947 'best-of':148,257,425,604,946 'best-of-n':87,159,165,407 'blacklist':438 'blog':596 'bucket':550 'casual':399 'categori':15 'chatgpt':970 'chines':3,10,49,489,494 'cilin':287,289,430,432 'clean.txt':268,324,330,338,470,475,482 'cli':58,106,109,642,934 'cn':391 'cohen':503 'comma':27 'compar':187,211,325,332,419,466 'critic':1021 'cv':22,101,516,526,532 'd':505,517,523,527,533,540,545,551,556,560 'defect':577 'delta':586 'densiti':28 'detect':4,14,113,120,129,390,449,459,474,938 'divey':31,553 'document.txt':460,467 'ensembl':995 'featur':18 'final.txt':486 'fraction':26 'fuse':76,566,944,992 'fusion':37 'gap':583 'general':46,72,386 'generat':9 'genr':591 'github.com':496 'github.com/hello-simpleai/chatgpt-comparison-detection)':495 'gltr':30,547 'grammar':576 'hc3':75,488,493,565 'hc3-chinese':487,492 'header':202 'hedg':908 'hero':602 'high':357,364,1027 'human':2,6,48,95,293,412,434,441,950 'humanize-chines':1 'j':252,375 'json':253,376 'kurt':555 'length':21,100 'literari':404 'llm':61,630 'longform':74,137,396,581 'low':347,954 'lr':36,40,70,97,138,318,397,415,452,510,994,998 'markdown':201 'medium':351,968 'n':90,91,162,168,383,410,411,413 'news':594 'no-human':439 'novel':85,124,195,388,406,598 'novel/auto':81 'o':143,156,173,184,267,300,323,329,337,378,469,485 'out.txt':301 'output.txt':939 'paper.txt':322,328,336 'paragraph':99 'pattern':1022,1029,1039 'perplex':29 'plus':16,32 'prompt':62,199 'python':222,235,238,244,249,264,269,277,284,295,303,320,326,334 'quick':176,280,339,422 'replac':1034 'review':600 'rewrit':140,153,170 'rule':13,38,447,996,1003,1007 'rule-on':446,1002,1006 'scene':34,45,67,80,123,127,132,262,272,306,315,385,508 'scene-awar':33,66,261,314,507 'score':1041 'script':219 'scripts/academic_cn.py':321,327,335 'scripts/compare_cn.py':304 'scripts/detect_cn.py':236,239,245,250 'scripts/humanize_cn.py':265,270,278,285 'scripts/patterns_cn.json':1012 'scripts/style_cn.py':296 'seed':382,608 'sentenc':20,25 'sentence-length':19 'short':24 'short-sent':23 'signal':1028 'skew':554 'skill' 'skill-49-voidborne-d-humanize-chinese' 'social':273 'source-brycewang-stanford' 'statist':17 'style':84,191,194,204,207,298,398,442,481,483 'tech':307 'text':11,51 'text.txt':237,240,246,251,266,271,279,286,297,305 'three':44 'tool':107 'top':548 'topic-academic-research' 'topic-agent-skills' 'topic-ai-agent' 'topic-awesome-list' 'topic-communication' 'topic-copaper' 'topic-economics' 'topic-education' 'topic-empirical-research' 'topic-international-relations' 'topic-political-science' 'topic-psychology' 'train':42 'trigram':103,538 'txt':115,122,131,142,145,155,158,172,175,183,186,193,206,213 'v':116,241,370,461 'v3.x':1001 'v5':529,535,542 'v5.0':52,64,562,943,991 'wechat':402 'weibo':405 'xiaohongshu':208,401 'xx':863 'zhihu':299,400,484 '一方面':658 '一眼就能看出是':652 '一般认为':906 '三段式':1024 '三段式套路':654 '三路':69 '不仅如此':670 '不可否认':673 '不是每个论点都要编号':759 '不能口语化':833 '不要每句差不多长':799 '不要每段都总分总结构':818 '不言而喻':675 '不过':794 '不难发现':666,778 '与此同时':668 '中危模式':689 '中权重检测':1030 '也可作为':60 '也有':725 '产生了一定效果':898 '亮点':65 '人性化文本':638 '人类':500 '仅评分':248,374 '从表中可以看出':868 '他决定休息':788 '他很累':787 '他累了':789 '众所周知':705 '但是':722,793 '低危模式':727 '作为':716 '保持学术性':883 '信息化改造':776 '信息熵低':746 '修改正则匹配':1019 '修辞堆砌':739 '值得一提的是':704 '值得关注':888 '值得注意的是':664,768 '做':808 '做了初步探讨':895 '偶尔加个反问或感叹':817 '像人写东西时自然的长短变化':828 '先自动':292 '全方位':684 '八股':985 '其次':656,755 '具体来说':707 '具有重要意义':840,887 '再套风格':294 '写作模式':648 '准确率':77 '减少被动句':900 '减少重复用词':801 '几乎确定是':365 '凸显':694 '出现':731,857 '分':234,850,976,980 '分以下':953,958 '分左右':967 '分差距':570 '分数':342 '分降幅':989 '列举成瘾':734 '初步来看':913 '初见成效':899 '删掉':780 '刻板化':981 '前人研究发现':891 '前后对比':215,302,420 '剔除':197 '加一两句口语化表达':814 '动辄':735 '助力':692,781 '区间':955 '协同增效':681 '去':633 '参数':368 '参数速查':367 '双评分':319 '双评分对比':189 '反制改写':105 '取得了显著成效':897 '取最低':96 '句子长度单调':743 '句式重组':785 '句话':824,827 '句长变异系数':515 '另一方面':659 '只用规则层':450 '只说好不说局限':874 '可以换成':807 '可以看到':779,986 '可以认为':918 '可独立运行':56 '可能':910 '可自定义':1013 '可选':478 '可降级到纯规则视图':1009 '合并过短的段落':928 '同一个词出现':802 '同义表达匮乏':855 '同义词扩展':290,433 '同时':723,795 '后面的话自己能说清楚':770 '含义':344 '含碰撞':437 '启用':288,431 '味':635 '困惑度':544 '固定随机种子':384 '在':511,714,792 '在一定程度上':729,911 '在多个领域有所应用':903 '在当今':712 '在过于绝对的判断前加':909 '场景允许的话':815 '基于':942 '基准测试':490 '基本像人写的':348 '填充废话':703 '增加替换规则':1018 '增强作者主体性':914 '复盘':699 '外部配置字段':1020 '多次':94 '多维度':685 '大概率':358 '大量分号':737 '套话':763 '如果无法运行':641 '如果结论段没有提到局限':921 '字':522,964 '字切':395 '学术专用':959 '学术专用检测与替换':1040 '学术专用检测维度':835 '学术双评分':421 '学术套话':882 '学术措辞':838 '学术改写策略':879 '学术维度':313 '学术论文':309 '学术论文特殊处理':830 '学术论文降到':956 '学术降重':188 '对比':464 '小红书':618 '小说':962 '小说显式':126 '就目前而言':912 '展开了讨论':896 '工作汇报':626 '工作流':454 '工具':643 '工具验证效果':935 '已有文献显示':892 '帮':782 '帮助':772 '干脆歇会儿':790 '平均降幅':572 '平衡论述套话':720 '底层逻辑':695 '开头方式重复':745 '开展':810 '引用整合度低':861 '归根结底':676 '当文本是学术论文时':831 '当然':923 '当用户要求':632 '彰显':693 '得到较多运用':902 '必然':877 '快速模式':177,340,423 '总':851 '总之':765 '总而言之':667 '情感表达平淡':744 '或独立脚本':59 '所有检测模式和替换规则在':1011 '所有脚本在':218 '打散':757 '打破均匀节奏':797 '打破每段差不多长的格局':821 '打破结构均匀度':925 '扫描文本中的以下模式':649 '技术升级':777 '把':753 '抓手':696 '拆分过长的':929 '指出':865 '指南使用':63 '按':393 '按严重程度分类':650 '按以下流程手动处理':644 '按以下顺序处理':750 '按长度自动切':134 '换句话说':671 '排比对偶过多':740 '推动':783 '推荐':110 '提升':774 '搞':809 '支持':773 '改写':146,255,463 '改写后再降':977 '改写完成后':932 '改写策略':749 '改写规则不同':832 '改后':144,157,174,185 '改成人话':639 '数字化转型':680,775 '数据论述模板化':867 '文本':114,141,154,171,205,212 '新增':79,83 '新质生产力':688 '无依赖':223 '既有':724 '时':640 '时代':713 '显示可疑句子':372 '显而易见':674 '更准':1005 '更是':719 '更有':726 '更重要的是':672 '替换':761,880 '替换词库':1035 '最低':416 '最可疑句子':243 '最后':657,756 '最强单信号':519 '有一定参考价值':889 '有些':352 '有的':825 '有的段落':822 '本文尝试':885 '本文旨在':839,884 '本研究也存在一定局限':924 '本研究关注':886 '本研究发现':917 '机械连接词':663 '权重和阈值配置':1042 '某种程度上':730 '标点滥用':736 '校准':506 '样本':603 '样本回归':580 '样本回归测试':564 '样本的':502 '样板文':983 '检测':117,225,457,646 '检测和改写中文':53 '检测器':945 '模式':227 '模板句':1033 '模板句式':709 '次':733,859 '次以上就换同义词':804 '次取':414 '正则':1036 '正确分离率':568 '段内句长':531 '段留':588 '段落保留':574 '段落级信号':98 '段落结构过于整齐':847 '段落节奏':820 '段落长度':525 '段落长度高度一致':742 '每个引用都是':862 '每段总':849 '每段总-分-总':848 '比如':805 '毫无疑问':706,878 '沉淀':698 '注':990 '注入人味':813 '注入学术犹豫语':907 '深度融合':683 '添加新':1014 '激进':276,333 '激进模式':381 '犹豫语过多':728 '独立脚本形式':216 '生成':360 '生成文本的完整工具链':55 '用':933 '用具体的例子代替抽象概括':816 '用自然过渡代替':758 '用词可预测':747 '由此可见':669 '痕迹':354 '的不断发展':711 '的背景下':715 '的重要组成部分':717 '目录下':220 '目标':941 '直接使用指南':631 '直接删掉':767,769 '相同文本的分数会比':1000 '真实':969 '着手':811 '短句占比':520 '短输出':971 '砍掉三段式':752 '研究':856 '研究表明':890,915 '破圈':701 '破折号':738 '社交':275 '稿件':130 '空洞宏大词':677 '空洞词':1026 '章节':121,192 '笔者倾向于认为':919 '笔者观察到':893 '笔者认为':916 '第一':660 '第一步':645 '第三':662 '第三步':829 '第二':661 '第二步':748 '第四步':930 '等价':217 '等级':343 '等转折处断开':796 '简而言之':708 '系统性':686 '纯':221 '纯文本':1037 '纯替换':283 '结论过于圆满':873 '统一':57,108 '统计特征':231 '综上所述':665,764 '自调':164 '虽然':721 '融合':71,453 '补一句':922 '补充局限性':920 '被动句式过度':843 '被广泛应用':844,901 '被认为是':845,904 '要保持学术严谨性':834 '规则维度':229 '触达':697 '训练':514 '论文':182 '论文模板':984 '评分标准':341 '评分融合了':993 '词':436,445 '词汇':1016 '详细':118,242 '详细模式':371 '语气过于确定':876 '说到底':766 '说明':369 '调整权重':1017 '调整段落长度':926 '赋能':678,771 '走':136 '跨段':102,537 '跳':451 '跳统计':178 '跳过统计优化':424 '转换前不先去':443 '转风格':479 '输出':254,377 '输出文件':379 '过度列举':870 '过短的句子合并':786 '过长的句子拆开':791 '这不仅':718 '进行':806 '进行了深入分析':841,894 '连接词':1025 '连接词密度异常':853 '迭代':700 '逗号密度':559 '通常已在':973 '通常被看作':905 '通用':614 '通用文本降到':951 '通用评分均低':960 '速度':282,429 '避免每段都一样':927 '配置说明':1010 '重复':104,539 '长文本':125,578 '长文本场景':82 '长文本新信号':530,543 '长文本最强信号':536 '长短句交替':798 '长篇博客':622,961 '长篇叙事':86 '长篇叙事专属':196 '闭环':679 '阈值基于':491 '降':636 '降到':965 '降本增效':682 '降重':311 '随着':710 '频繁出现':871 '颠覆':702 '风格信号':741 '风格转换':209,291 '首先':655,754 '验证':472,931 '高危模式':651 '高权重检测':1023 '高质量发展':687 '高频词':691,1032 '默认':92,147,256,417","prices":[{"id":"f997bb4c-abfd-4d41-95de-8aab0e605b50","listingId":"e0ee4318-4299-4943-8255-e0be4f6573b3","amountUsd":"0","unit":"free","nativeCurrency":null,"nativeAmount":null,"chain":null,"payTo":null,"paymentMethod":"skill-free","isPrimary":true,"details":{"org":"brycewang-stanford","category":"Awesome-Agent-Skills-for-Empirical-Research","install_from":"skills.sh"},"createdAt":"2026-05-05T00:53:21.045Z"}],"sources":[{"listingId":"e0ee4318-4299-4943-8255-e0be4f6573b3","source":"github","sourceId":"brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research/49-voidborne-d-humanize-chinese","sourceUrl":"https://github.com/brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research/tree/main/skills/49-voidborne-d-humanize-chinese","isPrimary":false,"firstSeenAt":"2026-05-05T00:53:21.045Z","lastSeenAt":"2026-05-18T18:53:15.877Z"}],"details":{"listingId":"e0ee4318-4299-4943-8255-e0be4f6573b3","quickStartSnippet":null,"exampleRequest":null,"exampleResponse":null,"schema":null,"openapiUrl":null,"agentsTxtUrl":null,"citations":[],"useCases":[],"bestFor":[],"notFor":[],"kindDetails":{"org":"brycewang-stanford","slug":"49-voidborne-d-humanize-chinese","github":{"repo":"brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research","stars":1032,"topics":["academic-research","agent-skills","ai-agent","awesome-list","communication","copaper","economics","education","empirical-research","international-relations","political-science","psychology","public-administration","reproducible-research","skills-library","social-science","sociology"],"license":"other","html_url":"https://github.com/brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research","pushed_at":"2026-05-11T09:41:12Z","description":"🔬 A curated collection of 23,000+ agent skills for empirical research across 8 social science disciplines. | 精选 23,000+ AI Agent 技能库，覆盖8大社会科学学科的实证研究。CoPaper.AI 20分钟完成一篇可复现的规范实证论文，并支持用户上传 Skills。-- Maintained by CoPaper.AI from Stanford REAP.","skill_md_sha":"8d8d10ed16f117f8146c667a50a055b4e48d38cd","skill_md_path":"skills/49-voidborne-d-humanize-chinese/SKILL.md","default_branch":"main","skill_tree_url":"https://github.com/brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research/tree/main/skills/49-voidborne-d-humanize-chinese"},"layout":"multi","source":"github","category":"Awesome-Agent-Skills-for-Empirical-Research","frontmatter":{"name":"humanize-chinese","description":"Detect and humanize AI-generated Chinese text. 20+ rule detection categories plus statistical features (sentence-length CV, short-sentence fraction, comma density, perplexity, GLTR, DivEye) plus scene-aware LR fusion (rule × 0.2 + LR × 0.8) trained on three scenes: general / academic / longform 长文本 (≥1500 字)。Unified CLI: ./humanize {detect,rewrite,academic,style,compare}. 8 style transforms (casual/zhihu/xiaohongshu/wechat/academic/literary/weibo/novel)。 Multi-paragraph rewriting (paragraph length CV、跨段 trigram 重复) plus best-of-N humanize (默认 N=10 取最低 LR)。165 replacement patterns + CiLin 同义词词林 38873 with collision blacklist。 Academic paper AIGC reduction for CNKI/VIP/Wanfang (知网/维普/万方 AIGC 检测降重)。 Pure Python, no dependencies, offline。v5.0.0 — HC3 fused 准确率 95%、学术 hero 100→35 (-65)、 工作汇报 96→13 (-83)、长篇博客 96→41 (-55)。 Use when user says: \"去AI味\", \"降AIGC\", \"人性化文本\", \"humanize chinese\", \"AI检测\", \"AIGC降重\", \"去除AI痕迹\", \"文本改写\", \"论文降重\", \"知网检测\", \"维普检测\", \"AI写作检测\", \"让文字更自然\", \"detect AI text\", \"humanize text\", \"reduce AIGC score\", \"make text human-like\", \"去ai化\", \"改成人话\", \"去机器味\", \"降低AI率\", \"过AIGC检测\", \"长文本改写\", \"小说改写\""},"skills_sh_url":"https://skills.sh/brycewang-stanford/Awesome-Agent-Skills-for-Empirical-Research/49-voidborne-d-humanize-chinese"},"updatedAt":"2026-05-18T18:53:15.877Z"}}