tonyeh/chameleon-dataset

--- license: cc-by-4.0 dataset_info: configs: - config_name: combined data_files: - split: train path: chameleon_profiles_combined.csv - config_name: seance data_files: - split: train path: chameleon_profiles_seance.csv - config_name: langextract data_files: - split: train path: chameleon_profiles_langextract.csv --- # Chameleon: A Dataset of Contextual Psychological Profiles ## Overview Chameleon is a dataset of **5,001 contextual psychological profiles** from **1,667 Reddit users**, each measured across multiple subreddit contexts. Unlike existing persona datasets that treat psychological profiles as fixed user attributes, Chameleon measures the same users across multiple contexts, enabling principled decomposition of psychological variance into stable traits and contextual states. The dataset accompanies the paper: > **Beyond Fixed Psychological Personas: State Beats Trait, but Language Models are State-Blind** > Tamunotonye Harry, Ivoline C. Ngong, Chima Nweke, Yuanyuan Feng, Joseph Near > *Findings of the Association for Computational Linguistics: ACL 2026* --- ## Key Finding Using Latent State-Trait (LST) theory and intraclass correlation coefficients, we find that **72–74% of psychological variance is within-person (state)** while only 26–28% is between-person (trait). Context shapes expressed psychology 2–3× more than stable individual differences. --- ## Files | File | Description | |------|-------------| | `chameleon_profiles_combined.csv` | Fused SEANCE + LangExtract profiles (z-normalized mean). **Recommended for most use cases.** | | `chameleon_profiles_seance.csv` | SEANCE-derived profiles (lexicon-based extraction) | | `chameleon_profiles_langextract.csv` | LangExtract-derived profiles (GPT-4o semantic extraction) | Each file contains **5,001 rows × 29 columns**: `post_id`, `user_id`, `subreddit`, and 26 psychological scale scores. --- ## Schema ### Base Columns | Column | Description | |--------|-------------| | `post_id` | Reddit post ID (maps to Webis-TLDR-17) | | `user_id` | Pseudonymized author identifier (e.g., user_0000) | | `subreddit` | Subreddit community (psychological context) | ### Psychological Scale Columns (26 total) #### Big Five Inventory (BFI-44) — Scale: 1–5 | Column | Construct | |--------|-----------| | `bfi_extraversion` | Sociability, assertiveness | | `bfi_agreeableness` | Cooperation, trust | | `bfi_conscientiousness` | Organization, discipline | | `bfi_neuroticism` | Emotional instability, anxiety | | `bfi_openness` | Intellectual curiosity, creativity | #### Schwartz Value Survey (SVS-57) — Scale: −1 to 7 | Column | Construct | |--------|-----------| | `svs_power` | Status, dominance | | `svs_achievement` | Success, competence | | `svs_hedonism` | Pleasure, enjoyment | | `svs_stimulation` | Excitement, novelty | | `svs_self_direction` | Independence, autonomy | | `svs_universalism` | Social justice, tolerance | | `svs_benevolence` | Caring for close others | | `svs_tradition` | Cultural/religious customs | | `svs_conformity` | Rule-following | | `svs_security` | Safety, stability | #### Self-Determination Theory (SDT) — Scale: 1–7 | Column | Construct | |--------|-----------| | `sdt_intrinsic_motivation` | Internal drive, curiosity | | `sdt_extrinsic_motivation` | External rewards | | `sdt_competence` | Feeling capable, effective | | `sdt_autonomy` | Sense of choice, self-direction | | `sdt_relatedness` | Social connection, belonging | #### Domain-Specific Risk-Taking (DOSPERT-40) — Scale: 1–7 | Column | Construct | |--------|-----------| | `dospert_investment` | Financial risk-taking | | `dospert_gambling` | Gambling propensity | | `dospert_health_safety` | Health risk-taking | | `dospert_recreational` | Physical risk-taking | | `dospert_ethical` | Ethical boundary-pushing | | `dospert_social` | Social risk-taking | --- ## Reproducing the Sample from Webis-TLDR-17 The post text is not included in this dataset for privacy reasons (see Ethics section). To reproduce the exact sample: **Source corpus:** [Webis-TLDR-17](https://huggingface.co/datasets/webis/tldr-17) **Filtering criteria:** - Users who posted across at least 3 distinct subreddits - Exactly 3 posts randomly sampled per user from 3 different subreddits - Minimum post length: 50 words - English language content only - Random seed: 42 **Sample size:** 1,667 users × 3 posts = 5,001 posts across 645 unique subreddits Match on `post_id` (Chameleon) = `id` field in Webis-TLDR-17. --- ## Extraction Pipeline Profiles were extracted using a two-stage pipeline: **Stage 1 — Feature Extraction:** - *SEANCE* (Crossley et al., 2017): rule-based lexicon matching, 250+ linguistic indices - *LangExtract* (2025): GPT-4o semantic pattern extraction **Stage 2 — Scale Assessment:** - GPT-4o prompted to respond to validated scale items as if it were the post's author, conditioned on extracted features **Stage 3 — Normalization and Fusion:** - Both methods z-normalized per dimension, then averaged to produce combined profiles --- ## Dataset Statistics | Characteristic | Value | |----------------|-------| | Total posts | 5,001 | | Unique users | 1,667 | | Posts per user | 3 (by design) | | Unique subreddits | 645 | | Subreddits with n ≥ 10 posts | 41 | | Psychological frameworks | 4 | | Psychological dimensions | 26 | | Extraction methods | 2 | **Top subreddits by post count:** AskReddit (1,558), relationships (923), relationship_advice (268), offmychest (198), depression (129), dating_advice (76), self (53), personalfinance (49), SuicideWatch (43), AdviceAnimals (42) --- ## Ethics This dataset constitutes secondary analysis of publicly available data from the Webis-TLDR-17 corpus (CC-BY-4.0), which contains Reddit posts made between 2006 and 2016. - Usernames are pseudonymized — original Reddit usernames are not included - Raw post text is not included to minimize privacy risks - Profiles reflect psychological states expressed in text, not stable traits of individuals - This research does not meet the federal definition of human subjects research under the Common Rule (45 CFR 46.102) --- ## Citation ```bibtex @inproceedings{harry2026chameleon, title = {Beyond Fixed Psychological Personas: State Beats Trait, but Language Models are State-Blind}, author = {Harry, Tamunotonye and Ngong, Ivoline C. and Nweke, Chima and Feng, Yuanyuan and Near, Joseph}, booktitle = {Findings of the Association for Computational Linguistics: ACL 2026}, year = {2026}, address = {San Diego, California} } ``` --- ## License This dataset is released under [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/).license: CC-BY-4.0 dataset_info: 配置项： - config_name: combined data_files: - split: 训练集 path: chameleon_profiles_combined.csv - config_name: seance data_files: - split: 训练集 path: chameleon_profiles_seance.csv - config_name: langextract data_files: - split: 训练集 path: chameleon_profiles_langextract.csv # Chameleon：情境化心理档案数据集 ## 概述 Chameleon 是一个包含**5001份情境化心理档案**的数据集，数据来自**1667位Reddit用户**，每位用户均在多个子论坛（subreddit）情境下被测量。与现有将心理档案视为固定用户属性的人格数据集不同，Chameleon 在多种情境下对同一用户进行测量，从而能够通过严谨的方法将心理差异分解为稳定特质与情境状态。 本数据集配套论文如下： > **《超越固定心理人格：状态胜于特质，但大语言模型（Large Language Model）不识状态》** > 作者：Tamunotonye Harry、Ivoline C. Ngong、Chima Nweke、Yuanyuan Feng、Joseph Near > 发表于：*《计算语言学协会研究发现：ACL 2026》* ## 核心发现 本研究采用潜在状态-特质（Latent State-Trait, LST）理论与组内相关系数开展分析，结果显示**72%~74%的心理差异来自个体内部（即状态层面）**，仅26%~28%来自个体间（即特质层面）。情境对个体外显心理特征的影响程度是稳定个体差异的2~3倍。 ## 数据文件 | 文件名 | 描述 | |--------|------| | `chameleon_profiles_combined.csv` | SEANCE与LangExtract融合的心理档案（经Z标准化均值处理）。**推荐多数场景使用**。 | | `chameleon_profiles_seance.csv` | 基于SEANCE提取的心理档案（基于词典的规则式提取） | | `chameleon_profiles_langextract.csv` | 基于LangExtract提取的心理档案（基于GPT-4o的语义提取） | 每个文件均包含**5001行 × 29列**数据，字段包括：`post_id`、`user_id`、`subreddit`，以及26项心理量表得分。 ## 数据结构 ### 基础字段 | 字段名 | 说明 | |--------|------| | `post_id` | Reddit帖子ID（对应Webis-TLDR-17数据集的id字段） | | `user_id` | 匿名化的作者标识符（例如：user_0000） | | `subreddit` | 所属子论坛社区（即心理测量的情境背景） | ### 心理量表字段（共26项） #### 大五人格量表（Big Five Inventory, BFI-44）—— 计分范围：1~5分 | 字段名 | 构念定义 | |--------|----------| | `bfi_extraversion` | 社交性、决断性 | | `bfi_agreeableness` | 合作性、信任感 | | `bfi_conscientiousness` | 组织性、自律性 | | `bfi_neuroticism` | 情绪不稳定性、焦虑倾向 | | `bfi_openness` | 求知欲、创造力 | #### 施瓦茨价值观量表（Schwartz Value Survey, SVS-57）—— 计分范围：-1~7分 | 字段名 | 构念定义 | |--------|----------| | `svs_power` | 地位、支配欲 | | `svs_achievement` | 成功、能力 | | `svs_hedonism` | 享乐、愉悦 | | `svs_stimulation` | 刺激、新奇 | | `svs_self_direction` | 独立、自主 | | `svs_universalism` | 社会正义、包容 | | `svs_benevolence` | 关爱亲近他人 | | `svs_tradition` | 文化/宗教习俗 | | `svs_conformity` | 守规性 | | `svs_security` | 安全、稳定 | #### 自我决定理论（Self-Determination Theory, SDT）—— 计分范围：1~7分 | 字段名 | 构念定义 | |--------|----------| | `sdt_intrinsic_motivation` | 内在驱动力、好奇心 | | `sdt_extrinsic_motivation` | 外部奖励驱动 | | `sdt_competence` | 能力感、效能感 | | `sdt_autonomy` | 选择权、自主感 | | `sdt_relatedness` | 社会联结、归属感 | #### 领域特异性风险倾向量表（Domain-Specific Risk-Taking, DOSPERT-40）—— 计分范围：1~7分 | 字段名 | 构念定义 | |--------|----------| | `dospert_investment` | 金融风险倾向 | | `dospert_gambling` | 赌博倾向 | | `dospert_health_safety` | 健康与安全风险倾向 | | `dospert_recreational` | 身体活动风险倾向 | | `dospert_ethical` | 伦理边界突破倾向 | | `dospert_social` | 社交风险倾向 | ## 复现Webis-TLDR-17数据集的样本 出于隐私保护考虑，本数据集未包含帖子原文（详见伦理声明章节）。若需复现完全一致的样本，请按以下步骤操作： **源语料库：** [Webis-TLDR-17](https://huggingface.co/datasets/webis/tldr-17) **筛选标准：** - 至少在3个不同子论坛发布过帖子的用户 - 为每位用户从3个不同子论坛中随机采样恰好3篇帖子 - 单篇帖子最小长度：50词 - 仅保留英文内容 - 随机种子：42 **样本规模：** 1667位用户 × 3篇帖子 = 5001篇帖子，覆盖645个独特子论坛 匹配方式：Chameleon数据集的`post_id`字段对应Webis-TLDR-17数据集的`id`字段。 ## 档案提取流程 本数据集的心理档案采用三阶段流程提取： **阶段1 — 特征提取：** - *SEANCE*（Crossley等人，2017）：基于规则的词典匹配方法，可生成250+项语言学指标 - *LangExtract*（2025）：基于GPT-4o的语义模式提取方法 **阶段2 — 量表评估：** 以提取的文本特征为条件，提示GPT-4o模拟帖子作者的身份，回答经过验证的量表题目 **阶段3 — 标准化与融合：** 对两种方法的结果按维度进行Z标准化，再取均值得到融合版心理档案 ## 数据集统计信息 | 统计指标 | 数值 | |----------|------| | 总帖子数 | 5001 | | 独特用户数 | 1667 | | 每位用户的帖子数 | 3（设计设定） | | 独特子论坛数 | 645 | | 帖子数≥10的子论坛数量 | 41 | | 采用的心理学理论框架数 | 4 | | 心理维度总数 | 26 | | 档案提取方法数 | 2 | **帖子数排名前十的子论坛：** AskReddit（1558篇）、relationships（923篇）、relationship_advice（268篇）、offmychest（198篇）、depression（129篇）、dating_advice（76篇）、self（53篇）、personalfinance（49篇）、SuicideWatch（43篇）、AdviceAnimals（42篇） ## 伦理声明 本数据集是对Webis-TLDR-17语料库（采用CC-BY-4.0许可证）中公开数据的二次分析，该语料库包含2006年至2016年间发布的Reddit帖子。 - 用户名已进行匿名化处理，未保留原始Reddit用户名 - 未包含原始帖子文本，以最大程度降低隐私风险 - 本数据集的心理档案反映的是用户在文本中外显的心理状态，而非个体的稳定特质 - 根据美国联邦《共同规则》（45 CFR 46.102），本研究不属于人类受试者研究范畴 ## 引用格式 bibtex @inproceedings{harry2026chameleon, title = {Beyond Fixed Psychological Personas: State Beats Trait, but Language Models are State-Blind}, author = {Harry, Tamunotonye and Ngong, Ivoline C. and Nweke, Chima and Feng, Yuanyuan and Near, Joseph}, booktitle = {Findings of the Association for Computational Linguistics: ACL 2026}, year = {2026}, address = {San Diego, California} } ## 许可证 本数据集采用[CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)许可证发布。