[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7F26JZX6EY0DF5XWFDB5214":3,"skill-same:01M7F26JZX6EY0DF5XWFDB5214":46},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":6,"ownerDescription":30,"updatedAt":31,"install":32,"resources":45},"01M7F26JZX6EY0DF5XWFDB5214","news-sentiment-finbert",null,"RSS 新闻情绪识别（FinBERT 中文金融情感）安装与运维 — 情绪管线架构、transformers 安装、FinBERT 权重下载、字典法扩充、全量重算、情绪筛选\u002F条形图\u002F个股资讯标签的使用。在 QuantBot \u002F Claude Code 中排查新闻情绪不生效、重新安装 FinBERT、扩充情绪词、触发新闻情绪重算时使用。触发词：新闻情绪、情绪识别、FinBERT、情绪不生效、情绪都是中性、字典法、新闻重算、news sentiment","lza6\u002FArmor-piercing-system-prompts","lza6","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F128880206?v=4","01M7EB3K7WPX5T28EJ2NVHE6EG","Chinese-tingfeng","01M7EB3K97E1KQ16QWKJMFNGX8",11,"MIT","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F05-技能与代理开发\u002Fnews-sentiment-finbert\u002FSKILL.md",[],[],"2026-11-16 10:05:34",false,"---\nname: news-sentiment-finbert\ndescription: \"RSS 新闻情绪识别（FinBERT 中文金融情感）安装与运维 — 情绪管线架构、transformers 安装、FinBERT 权重下载、字典法扩充、全量重算、情绪筛选\u002F条形图\u002F个股资讯标签的使用。在 QuantBot \u002F Claude Code 中排查新闻情绪不生效、重新安装 FinBERT、扩充情绪词、触发新闻情绪重算时使用。触发词：新闻情绪、情绪识别、FinBERT、情绪不生效、情绪都是中性、字典法、新闻重算、news sentiment\"\n---\n\n> ## ⚙️ 运行环境契约（最高优先级，先于本文其余内容执行）\n>\n> 本技能可能运行在 **QuantBot（QwenPaw 容器）** 或**宿主机\u002F本地 Claude Code**。执行前先探测环境（`which docker`、API 连通性），并遵守以下映射规则：\n>\n> 1. **后端 API 地址**：QwenPaw \u002F 容器网络内一律用 `http:\u002F\u002Fquantmind:8000`（`quantmind` 是 docker 网络别名）；仅宿主机调试用 `http:\u002F\u002F127.0.0.1:8000`。正文中出现的 `127.0.0.1:8000`、`localhost:800x`，在 QwenPaw 环境下自动替换为 `http:\u002F\u002Fquantmind:8000`。\n> 2. **取数脚本执行**：凡 import 了 `pandas \u002F duckdb \u002F psycopg2 \u002F numpy \u002F sqlalchemy` 等重依赖或 `backend` 包的脚本，**必须在 quantmind 容器内执行**（QwenPaw 本地 venv 无这些依赖）：\n>    ```bash\n>    docker cp \u003C脚本路径> quantmind:\u002Ftmp\u002F\u003C脚本名> && docker exec -w \u002Fapp quantmind python3 \u002Ftmp\u002F\u003C脚本名> \u003C参数>\n>    ```\n>    脚本源三选一：宿主机 repo `skills\u002F\u003Cname>\u002Fscripts\u002F`、QwenPaw 工作区 `\u002Fapp\u002Fworking\u002Fworkspaces\u002Fdefault\u002Fskills\u002F\u003Cname>\u002Fscripts\u002F`、挂载目录 `\u002Fquantmind\u002Fskills\u002F\u003Cname>\u002Fscripts\u002F`。纯标准库脚本（无重依赖）可在 QwenPaw 本地直接跑。\n> 3. **报告落盘**：股票报告页可见的 MD\u002FPDF 报告，直接写 `\u002Fdata\u002Freports\u002Ftrading_agents\u002F{市场或类别}\u002F{股票名}\u002F`（QwenPaw 对 `\u002Fapp\u002Fdb` 有写权限，**直接写文件，不要 docker cp**）；过程数据 facts 写 `\u002Fdata\u002Freports\u002F\u003C类别>\u002F`（`\u002Fdata` 可写）。\n> 4. **MD → PDF 转换（按优先级降级）**：\n>    ① `docker exec -w \u002Fapp quantmind python3 backend\u002Fscripts\u002Fmd_to_pdf_report.py \u003C输入.md> \u003C输出.pdf>`（研报级排版，首选）；\n>    ② docker 不可用时，**改用 QwenPaw 内置 `pdf` 技能**把 MD 转成 PDF；\n>    ③ 两者都不可用则只交付 MD，并明确告知用户 PDF 未能生成及原因。\n> 5. 本文中的 `~\u002F.claude`、`cp -r ... ~\u002F.claude\u002Fskills` 等说明仅适用于本地 Claude Code 维护者，**QuantBot 不要执行**。\n\n# 新闻情绪识别（FinBERT）安装与运维技能\n\nQuantMind 的 RSS 新闻情绪识别管线：Huntly 抓取 → celery 每分钟 enrich → 股票\u002F行业\u002F事件标签 + 情绪分 → `news_article_enrichment` 表 → 前端 RSS 面板\u002F个股资讯 Tab 展示利好\u002F利空。\n\n本技能覆盖**情绪层**的安装、验证、重算、排查、词库增强。\n\n## 1. 架构与数据流\n\n```\nHuntly (RSS 抓取, 381 源, 41.5 万文章)\n  → celery-beat 每 60s: news-enrich-recent (run_enrichment_batch)\n  → NewsMatcher (Aho-Corasick + finance_lexicon 5 万词: 股票\u002F行业\u002F事件\u002F情感)\n  → sentiment.score() (两级融合)\n      字典法 dict_score (finance_lexicon sentiment_pos\u002Fneg 词权重差)\n      + FinBERT (bardsai\u002Ffinance-sentiment-zh-base, CPU, conf≥0.55 才融合)\n  → 写入 news_article_enrichment 表 (model_version 带 +finbert 后缀表示 FinBERT 生效)\n  → 前端:\n      \u002Frss-news  NewsPanel: 情绪筛选\u002F利好利空强度排序\u002F统计栏情绪分布条形图\n      个股终端「个股资讯」Tab: 每篇利好\u002F利空标签 (stock_news 关联 enrichment)\n```\n\n**情绪词库**：finance_lexicon 已从最初的 230 词扩充到 **51,887 词**（sentiment_pos 27,227 \u002F sentiment_neg 24,089 \u002F event 571），来源见第 6 节。\n\n## 6. 情绪词库增强（从 230 词 → 5 万词）\n\n情绪词分两级来源，统一存 finance_lexicon：\n\n### 6.1 现成情感词典（通用，可复现导入）\n| 来源 | 词数 | 说明 |\n|---|---|---|\n| **DLUT 情感本体库** (github.com\u002Fyizhanmiao\u002FDLUT-Emotionontology) | 22,001 | 中文情感词汇本体，强度 1-9，极性 1正\u002F2负 |\n| **pysenti** (github.com\u002Fshibing624\u002Fpysenti) | 9,870 | 内置情感词典，连续分数 -7~+7 |\n| **NTUSD 台大词典** (github.com\u002Fntunlplab\u002FNTUSD) | 20,485 | 繁体，Big5 编码，需转简体 |\n\n### 6.2 RSS 标题提炼（金融语境，最精准）\n用 41.5 万条 Huntly 标题 + enrichment 已有情绪标签，做**共现统计**（对数似然比 LLR 筛选）：\n- bearish 标题高频词 → 负向金融情绪词（下跌\u002F暴跌\u002F跌破\u002F违规\u002F立案\u002F警示）\n- bullish 标题高频词 → 正向金融情绪词（涨超\u002F涨停\u002F新高\u002F买入\u002F上调）\n- 产出 5,705 词，贴合 A股\u002F监管\u002F业绩语境，通用词典覆盖不到的（如\"暂停开户\"\"被重锤\"）都在这\n\n### 6.3 复现导入\n```bash\n# 词表已固化: backend\u002Fscripts\u002Fdata\u002Ffinance_sentiment_lexicon.tsv (51,213 词)\n# 导入脚本: backend\u002Fscripts\u002Fimport_sentiment_lexicon.py\n# 本地\npython3 backend\u002Fscripts\u002Fimport_sentiment_lexicon.py\n# 容器\ndocker exec quantmind python3 \u002Fapp\u002Fbackend\u002Fscripts\u002Fimport_sentiment_lexicon.py\n```\n\n导入后重载 matcher：\n```bash\ncurl -s -X POST -H \"$AUTH\" \"$BASE\u002Fapi\u002Fv1\u002Fnews\u002Fenrichment\u002Frun\"  # 触发重算\n```\n\n### 6.4 新增情绪词的完整流程\n1. 从新数据源提取词（如再跑 RSS 统计）\n2. 写入 `backend\u002Fscripts\u002Fdata\u002Ffinance_sentiment_lexicon.tsv`（`term\\tpos|neg\\tweight`）\n3. 跑 `import_sentiment_lexicon.py` 导入\n4. 触发 `\u002Fenrichment\u002Frun` 或 rebuild 让新词生效\n\n## 2. 情绪不生效 \u002F 全是中性 的排查（最常见）\n\n**症状**：RSS 面板里利好\u002F利空占比极低（\u003C5%），几乎全 neutral，置信度恒 0.3。\n\n**根因**：FinBERT 模型没装上（容器缺 `transformers`），回退纯字典法，而字典法词太少打不出分。\n\n**判断方法**：查 `model_version` 有没有 `+finbert` 后缀。\n```bash\ndocker exec quantmind python3 -c \"\nimport asyncio\nfrom backend.shared.database_manager_v2 import get_session\nfrom sqlalchemy import text\nasync def main():\n    async with get_session() as s:\n        r = await s.execute(text(\\\"SELECT sentiment_label, COUNT(*) FROM news_article_enrichment WHERE model_version LIKE '%finbert%' GROUP BY sentiment_label ORDER BY 2 DESC\\\"))\n        print('+finbert 分布:', r.fetchall())\n        r = await s.execute(text(\\\"SELECT COUNT(*) FROM news_article_enrichment WHERE model_version NOT LIKE '%finbert%'\\\"))\n        print('无 finbert 条数:', r.fetchone()[0])\nasyncio.run(main())\n\"\n```\n- 有 `+finbert` 且有 bullish\u002Fbearish 分布 → FinBERT 正常\n- 全 neutral \u002F 无 `+finbert` → FinBERT 没生效，按第 3 节处理\n\n## 3. 安装 \u002F 修复 FinBERT\n\n### 3.1 检查 transformers 是否在容器里\n```bash\ndocker exec quantmind python3 -c \"import transformers; print(transformers.__version__)\"\n# celery worker 也要有（enrich\u002Frebuild 在 celery 里跑）\ndocker exec quantmind-celery python3 -c \"import transformers; print(transformers.__version__)\"\n```\n\n### 3.2 缺的话手动装（临时修复，镜像重建后固化）\n```bash\ndocker exec quantmind pip install --no-cache-dir transformers\ndocker exec quantmind-celery pip install --no-cache-dir transformers\n```\n\n### 3.3 下载\u002F同步 FinBERT 权重（~100MB，离线可用）\n```bash\n# 触发下载（首次）\ndocker exec quantmind python3 -c \"\nfrom backend.services.api.news import sentiment as s\ns._model_ready=False; s._model_failed=False\ns._try_load(); print('ready=', s._model_ready)\n\"\n# 权重缓存位置: \u002Froot\u002F.cache\u002Fhuggingface\u002Fhub\u002Fmodels--bardsai--finance-sentiment-zh-base\n# celery worker 复用（主容器导出 → tar 管道 → celery 导入）\ndocker exec quantmind tar -C \u002Froot\u002F.cache\u002Fhuggingface\u002Fhub -cf - models--bardsai--finance-sentiment-zh-base \\\n  | docker exec -i quantmind-celery sh -c 'mkdir -p \u002Froot\u002F.cache\u002Fhuggingface\u002Fhub && tar -C \u002Froot\u002F.cache\u002Fhuggingface\u002Fhub -xf -'\n```\n\n### 3.4 验证 FinBERT 可用\n```bash\ndocker exec quantmind-celery python3 -c \"\nimport os\nos.environ['HF_HUB_OFFLINE']='1'\nfrom transformers import pipeline\np = pipeline('sentiment-analysis', model='bardsai\u002Ffinance-sentiment-zh-base', device=-1)\nprint(p('公司业绩暴雷，股价暴跌')[0])   # 期望 label=negative, score>0.9\n\"\n```\n\n### 3.5 重启 celery worker 让新代码\u002F新依赖生效\n```bash\ndocker restart quantmind-celery\nsleep 8\n# 确认后台加载成功\ndocker exec quantmind-celery python3 -c \"\nfrom backend.services.api.news import sentiment as s\ns._ensure_loading()\nimport time\nfor _ in range(40):\n    if s.is_available(): break\n    time.sleep(1)\nprint('FinBERT ready:', s.is_available())\n\"\n```\n\n## 4. 触发全量重算（历史情绪重打）\n\n`model_version` 变了（比如补上 `+finbert` 后）需要重算历史才能让旧文章带上情绪。force=true 全量重算 41 万条 Huntly 文章，约 3-4 小时后台跑。\n\n```bash\nTOKEN=$(curl -s -X POST $BASE\u002Fapi\u002Fv1\u002Fauth\u002Flogin -H \"Content-Type: application\u002Fjson\" \\\n  -d '{\"username\":\"admin\",\"password\":\"admin123\",\"tenant_id\":\"default\"}' \\\n  | python3 -c \"import sys,json; print(json.load(sys.stdin).get('access_token',''))\")\n\n# 全量重算（force=true 覆盖已 enrich 的）\ncurl -s -X POST -H \"Authorization: Bearer $TOKEN\" \\\n  \"$BASE\u002Fapi\u002Fv1\u002Fnews\u002Fenrichment\u002Frebuild-all?force=true\"\n# 或增量（跳过 model_version 已是最新的）\ncurl -s -X POST -H \"Authorization: Bearer $TOKEN\" \\\n  \"$BASE\u002Fapi\u002Fv1\u002Fnews\u002Fenrichment\u002Frebuild-all\"\n\n# 查进度\ncurl -s -H \"Authorization: Bearer $TOKEN\" \"$BASE\u002Fapi\u002Fv1\u002Fnews\u002Fenrichment\u002Frebuild-progress\"\n```\n\n> 注意：rebuild 在 celery worker 的后台线程跑，**重启 celery 会中断 rebuild**。重算期间别重启 celery。\n\n## 5. 字典法增强（FinBERT 不可用时的兜底）\n\n情绪是两级融合，FinBERT 置信度 \u003C0.55 或不可用时用字典法。字典词在 `finance_lexicon` 表（kind=sentiment_pos\u002Fneg，weight 为强度）。\n\n扩充方式：\n1. 改仓库 `backend\u002Fscripts\u002Fseed_a_share_stocks.py` 的 `_BUILTIN_SENTIMENT_POS\u002FNEG` 常量（已含 2026-08 扩充的 ~80 个词）\n2. 重跑 seed：\n```bash\ndocker exec quantmind python3 \u002Fapp\u002Fbackend\u002Fscripts\u002Fseed_a_share_stocks.py\n```\n3. 或直接 SQL 插入（幂等，先查后插）：\n```bash\ndocker exec quantmind python3 -c \"\nimport asyncio\nfrom backend.shared.database_manager_v2 import get_session\nfrom sqlalchemy import text\nasync def main():\n    async with get_session() as s:\n        await s.execute(text(\\\"INSERT INTO finance_lexicon (term, kind, weight, enabled) VALUES ('涨停','sentiment_pos',1.0,true)\\\"))\n        await s.commit()\nasyncio.run(main())\n\"\n```\n4. 词表改动立即生效（NewsMatcher 每 600s 重载 \u002F 可触发 `\u002Fenrichment\u002Frun` 重算新文章）\n\n## 6. 情绪数据怎么用（前端已就绪）\n\n| 位置 | 说明 |\n|---|---|\n| `\u002Frss-news` NewsPanel | 情绪筛选（利好\u002F利空\u002F中性）、利好\u002F利空强度排序、统计栏情绪分布条形图（红利好\u002F绿利空\u002F灰中性） |\n| 个股终端 → 个股资讯 Tab | 每篇标题前利好\u002F利空标签（stock_news 按 huntly_page_id join enrichment） |\n| `\u002Fnews\u002Farticles` API | `sentiment=bullish\u002Fbearish\u002Fneutral` 过滤 + `sort=sentiment_bullish\u002Fsentiment_bearish` |\n| `\u002Fnews\u002Fenrichment\u002Fstats` | 当前筛选下的情绪分布计数（前端统计栏数据源） |\n\n## 7. 相关代码文件\n\n- `backend\u002Fservices\u002Fapi\u002Fnews\u002Fsentiment.py` — FinBERT 懒加载\u002F后台加载\u002F失败重试（`_RETRY_AFTER` 冷却 300s）\n- `backend\u002Fservices\u002Fapi\u002Fnews\u002Fenricher.py` — enrich 管线，0.6字典+0.4FinBERT 融合\n- `backend\u002Fservices\u002Fapi\u002Fnews\u002Fmatcher.py` — Aho-Corasick 匹配 + 字典分\n- `backend\u002Fscripts\u002Fseed_a_share_stocks.py` — finance_lexicon 内置词（含情绪词）\n- `backend\u002Fscripts\u002Fimport_sentiment_lexicon.py` — **5 万情绪词批量导入脚本（可复现）**\n- `backend\u002Fscripts\u002Fdata\u002Ffinance_sentiment_lexicon.tsv` — **合并情绪词表 51,213 词（RSS提炼+DLUT+pysenti+NTUSD）**\n- `backend\u002Fservices\u002Fapi\u002Frouters\u002Fnews.py` — \u002Fnews\u002F* 路由（articles\u002Fenrichment\u002Fstats\u002Fsources）\n- `backend\u002Fservices\u002Fapi\u002Frouters\u002Fstock_terminal.py` — \u002Fstock-terminal\u002Fnews 个股资讯（带情绪标签）\n- `electron\u002Fsrc\u002Ffeatures\u002Fnews\u002Fcomponents\u002FNewsPanel.tsx` — RSS 面板（情绪分布条形图）\n- `electron\u002Fsrc\u002Ffeatures\u002Fstock-terminal\u002Fcomponents\u002Ftabs\u002FNewsTab.tsx` — 个股资讯情绪标签\n- `docker\u002FDockerfile.oss` + `requirements\u002Fai.txt` — transformers 依赖 + FinBERT 权重预下载\n\n## 8. 相关技能\n\n- **[[quantmind-operations]]** — RSS 新闻对接与分析（文章拉取\u002F过滤\u002F富化统计）\n- **[[quantdb-fields]]** — 数据字段口径\n","> ## ⚙️ 运行环境契约（最高优先级，先于本文其余内容执行）\n>\n> 本技能可能运行在 **QuantBot（QwenPaw 容器）** 或**宿主机\u002F本地 Claude Code**。执行前先探测环境（`which docker`、API 连通性），并遵守以下映射规则：\n>\n> 1. **后端 API 地址**：QwenPaw \u002F 容器网络内一律用 `http:\u002F\u002Fquantmind:8000`（`quantmind` 是 docker 网络别名）；仅宿主机调试用 `http:\u002F\u002F127.0.0.1:8000`。正文中出现的 `127.0.0.1:8000`、`localhost:800x`，在 QwenPaw 环境下自动替换为 `http:\u002F\u002Fquantmind:8000`。\n> 2. **取数脚本执行**：凡 import 了 `pandas \u002F duckdb \u002F psycopg2 \u002F numpy \u002F sqlalchemy` 等重依赖或 `backend` 包的脚本，**必须在 quantmind 容器内执行**（QwenPaw 本地 venv 无这些依赖）：\n>    ```bash\n>    docker cp \u003C脚本路径> quantmind:\u002Ftmp\u002F\u003C脚本名> && docker exec -w \u002Fapp quantmind python3 \u002Ftmp\u002F\u003C脚本名> \u003C参数>\n>    ```\n>    脚本源三选一：宿主机 repo `skills\u002F\u003Cname>\u002Fscripts\u002F`、QwenPaw 工作区 `\u002Fapp\u002Fworking\u002Fworkspaces\u002Fdefault\u002Fskills\u002F\u003Cname>\u002Fscripts\u002F`、挂载目录 `\u002Fquantmind\u002Fskills\u002F\u003Cname>\u002Fscripts\u002F`。纯标准库脚本（无重依赖）可在 QwenPaw 本地直接跑。\n> 3. **报告落盘**：股票报告页可见的 MD\u002FPDF 报告，直接写 `\u002Fdata\u002Freports\u002Ftrading_agents\u002F{市场或类别}\u002F{股票名}\u002F`（QwenPaw 对 `\u002Fapp\u002Fdb` 有写权限，**直接写文件，不要 docker cp**）；过程数据 facts 写 `\u002Fdata\u002Freports\u002F\u003C类别>\u002F`（`\u002Fdata` 可写）。\n> 4. **MD → PDF 转换（按优先级降级）**：\n>    ① `docker exec -w \u002Fapp quantmind python3 backend\u002Fscripts\u002Fmd_to_pdf_report.py \u003C输入.md> \u003C输出.pdf>`（研报级排版，首选）；\n>    ② docker 不可用时，**改用 QwenPaw 内置 `pdf` 技能**把 MD 转成 PDF；\n>    ③ 两者都不可用则只交付 MD，并明确告知用户 PDF 未能生成及原因。\n> 5. 本文中的 `~\u002F.claude`、`cp -r ... ~\u002F.claude\u002Fskills` 等说明仅适用于本地 Claude Code 维护者，**QuantBot 不要执行**。\n\n# 新闻情绪识别（FinBERT）安装与运维技能\n\nQuantMind 的 RSS 新闻情绪识别管线：Huntly 抓取 → celery 每分钟 enrich → 股票\u002F行业\u002F事件标签 + 情绪分 → `news_article_enrichment` 表 → 前端 RSS 面板\u002F个股资讯 Tab 展示利好\u002F利空。\n\n本技能覆盖**情绪层**的安装、验证、重算、排查、词库增强。\n\n## 1. 架构与数据流\n\n```\nHuntly (RSS 抓取, 381 源, 41.5 万文章)\n  → celery-beat 每 60s: news-enrich-recent (run_enrichment_batch)\n  → NewsMatcher (Aho-Corasick + finance_lexicon 5 万词: 股票\u002F行业\u002F事件\u002F情感)\n  → sentiment.score() (两级融合)\n      字典法 dict_score (finance_lexicon sentiment_pos\u002Fneg 词权重差)\n      + FinBERT (bardsai\u002Ffinance-sentiment-zh-base, CPU, conf≥0.55 才融合)\n  → 写入 news_article_enrichment 表 (model_version 带 +finbert 后缀表示 FinBERT 生效)\n  → 前端:\n      \u002Frss-news  NewsPanel: 情绪筛选\u002F利好利空强度排序\u002F统计栏情绪分布条形图\n      个股终端「个股资讯」Tab: 每篇利好\u002F利空标签 (stock_news 关联 enrichment)\n```\n\n**情绪词库**：finance_lexicon 已从最初的 230 词扩充到 **51,887 词**（sentiment_pos 27,227 \u002F sentiment_neg 24,089 \u002F event 571），来源见第 6 节。\n\n## 6. 情绪词库增强（从 230 词 → 5 万词）\n\n情绪词分两级来源，统一存 finance_lexicon：\n\n### 6.1 现成情感词典（通用，可复现导入）\n| 来源 | 词数 | 说明 |\n|---|---|---|\n| **DLUT 情感本体库** (github.com\u002Fyizhanmiao\u002FDLUT-Emotionontology) | 22,001 | 中文情感词汇本体，强度 1-9，极性 1正\u002F2负 |\n| **pysenti** (github.com\u002Fshibing624\u002Fpysenti) | 9,870 | 内置情感词典，连续分数 -7~+7 |\n| **NTUSD 台大词典** (github.com\u002Fntunlplab\u002FNTUSD) | 20,485 | 繁体，Big5 编码，需转简体 |\n\n### 6.2 RSS 标题提炼（金融语境，最精准）\n用 41.5 万条 Huntly 标题 + enrichment 已有情绪标签，做**共现统计**（对数似然比 LLR 筛选）：\n- bearish 标题高频词 → 负向金融情绪词（下跌\u002F暴跌\u002F跌破\u002F违规\u002F立案\u002F警示）\n- bullish 标题高频词 → 正向金融情绪词（涨超\u002F涨停\u002F新高\u002F买入\u002F上调）\n- 产出 5,705 词，贴合 A股\u002F监管\u002F业绩语境，通用词典覆盖不到的（如\"暂停开户\"\"被重锤\"）都在这\n\n### 6.3 复现导入\n```bash\n# 词表已固化: backend\u002Fscripts\u002Fdata\u002Ffinance_sentiment_lexicon.tsv (51,213 词)\n# 导入脚本: backend\u002Fscripts\u002Fimport_sentiment_lexicon.py\n# 本地\npython3 backend\u002Fscripts\u002Fimport_sentiment_lexicon.py\n# 容器\ndocker exec quantmind python3 \u002Fapp\u002Fbackend\u002Fscripts\u002Fimport_sentiment_lexicon.py\n```\n\n导入后重载 matcher：\n```bash\ncurl -s -X POST -H \"$AUTH\" \"$BASE\u002Fapi\u002Fv1\u002Fnews\u002Fenrichment\u002Frun\"  # 触发重算\n```\n\n### 6.4 新增情绪词的完整流程\n1. 从新数据源提取词（如再跑 RSS 统计）\n2. 写入 `backend\u002Fscripts\u002Fdata\u002Ffinance_sentiment_lexicon.tsv`（`term\\tpos|neg\\tweight`）\n3. 跑 `import_sentiment_lexicon.py` 导入\n4. 触发 `\u002Fenrichment\u002Frun` 或 rebuild 让新词生效\n\n## 2. 情绪不生效 \u002F 全是中性 的排查（最常见）\n\n**症状**：RSS 面板里利好\u002F利空占比极低（\u003C5%），几乎全 neutral，置信度恒 0.3。\n\n**根因**：FinBERT 模型没装上（容器缺 `transformers`），回退纯字典法，而字典法词太少打不出分。\n\n**判断方法**：查 `model_version` 有没有 `+finbert` 后缀。\n```bash\ndocker exec quantmind python3 -c \"\nimport asyncio\nfrom backend.shared.database_manager_v2 import get_session\nfrom sqlalchemy import text\nasync def main():\n    async with get_session() as s:\n        r = await s.execute(text(\\\"SELECT sentiment_label, COUNT(*) FROM news_article_enrichment WHERE model_version LIKE '%finbert%' GROUP BY sentiment_label ORDER BY 2 DESC\\\"))\n        print('+finbert 分布:', r.fetchall())\n        r = await s.execute(text(\\\"SELECT COUNT(*) FROM news_article_enrichment WHERE model_version NOT LIKE '%finbert%'\\\"))\n        print('无 finbert 条数:', r.fetchone()[0])\nasyncio.run(main())\n\"\n```\n- 有 `+finbert` 且有 bullish\u002Fbearish 分布 → FinBERT 正常\n- 全 neutral \u002F 无 `+finbert` → FinBERT 没生效，按第 3 节处理\n\n## 3. 安装 \u002F 修复 FinBERT\n\n### 3.1 检查 transformers 是否在容器里\n```bash\ndocker exec quantmind python3 -c \"import transformers; print(transformers.__version__)\"\n# celery worker 也要有（enrich\u002Frebuild 在 celery 里跑）\ndocker exec quantmind-celery python3 -c \"import transformers; print(transformers.__version__)\"\n```\n\n### 3.2 缺的话手动装（临时修复，镜像重建后固化）\n```bash\ndocker exec quantmind pip install --no-cache-dir transformers\ndocker exec quantmind-celery pip install --no-cache-dir transformers\n```\n\n### 3.3 下载\u002F同步 FinBERT 权重（~100MB，离线可用）\n```bash\n# 触发下载（首次）\ndocker exec quantmind python3 -c \"\nfrom backend.services.api.news import sentiment as s\ns._model_ready=False; s._model_failed=False\ns._try_load(); print('ready=', s._model_ready)\n\"\n# 权重缓存位置: \u002Froot\u002F.cache\u002Fhuggingface\u002Fhub\u002Fmodels--bardsai--finance-sentiment-zh-base\n# celery worker 复用（主容器导出 → tar 管道 → celery 导入）\ndocker exec quantmind tar -C \u002Froot\u002F.cache\u002Fhuggingface\u002Fhub -cf - models--bardsai--finance-sentiment-zh-base \\\n  | docker exec -i quantmind-celery sh -c 'mkdir -p \u002Froot\u002F.cache\u002Fhuggingface\u002Fhub && tar -C \u002Froot\u002F.cache\u002Fhuggingface\u002Fhub -xf -'\n```\n\n### 3.4 验证 FinBERT 可用\n```bash\ndocker exec quantmind-celery python3 -c \"\nimport os\nos.environ['HF_HUB_OFFLINE']='1'\nfrom transformers import pipeline\np = pipeline('sentiment-analysis', model='bardsai\u002Ffinance-sentiment-zh-base', device=-1)\nprint(p('公司业绩暴雷，股价暴跌')[0])   # 期望 label=negative, score>0.9\n\"\n```\n\n### 3.5 重启 celery worker 让新代码\u002F新依赖生效\n```bash\ndocker restart quantmind-celery\nsleep 8\n# 确认后台加载成功\ndocker exec quantmind-celery python3 -c \"\nfrom backend.services.api.news import sentiment as s\ns._ensure_loading()\nimport time\nfor _ in range(40):\n    if s.is_available(): break\n    time.sleep(1)\nprint('FinBERT ready:', s.is_available())\n\"\n```\n\n## 4. 触发全量重算（历史情绪重打）\n\n`model_version` 变了（比如补上 `+finbert` 后）需要重算历史才能让旧文章带上情绪。force=true 全量重算 41 万条 Huntly 文章，约 3-4 小时后台跑。\n\n```bash\nTOKEN=$(curl -s -X POST $BASE\u002Fapi\u002Fv1\u002Fauth\u002Flogin -H \"Content-Type: application\u002Fjson\" \\\n  -d '{\"username\":\"admin\",\"password\":\"admin123\",\"tenant_id\":\"default\"}' \\\n  | python3 -c \"import sys,json; print(json.load(sys.stdin).get('access_token',''))\")\n\n# 全量重算（force=true 覆盖已 enrich 的）\ncurl -s -X POST -H \"Authorization: Bearer $TOKEN\" \\\n  \"$BASE\u002Fapi\u002Fv1\u002Fnews\u002Fenrichment\u002Frebuild-all?force=true\"\n# 或增量（跳过 model_version 已是最新的）\ncurl -s -X POST -H \"Authorization: Bearer $TOKEN\" \\\n  \"$BASE\u002Fapi\u002Fv1\u002Fnews\u002Fenrichment\u002Frebuild-all\"\n\n# 查进度\ncurl -s -H \"Authorization: Bearer $TOKEN\" \"$BASE\u002Fapi\u002Fv1\u002Fnews\u002Fenrichment\u002Frebuild-progress\"\n```\n\n> 注意：rebuild 在 celery worker 的后台线程跑，**重启 celery 会中断 rebuild**。重算期间别重启 celery。\n\n## 5. 字典法增强（FinBERT 不可用时的兜底）\n\n情绪是两级融合，FinBERT 置信度 \u003C0.55 或不可用时用字典法。字典词在 `finance_lexicon` 表（kind=sentiment_pos\u002Fneg，weight 为强度）。\n\n扩充方式：\n1. 改仓库 `backend\u002Fscripts\u002Fseed_a_share_stocks.py` 的 `_BUILTIN_SENTIMENT_POS\u002FNEG` 常量（已含 2026-08 扩充的 ~80 个词）\n2. 重跑 seed：\n```bash\ndocker exec quantmind python3 \u002Fapp\u002Fbackend\u002Fscripts\u002Fseed_a_share_stocks.py\n```\n3. 或直接 SQL 插入（幂等，先查后插）：\n```bash\ndocker exec quantmind python3 -c \"\nimport asyncio\nfrom backend.shared.database_manager_v2 import get_session\nfrom sqlalchemy import text\nasync def main():\n    async with get_session() as s:\n        await s.execute(text(\\\"INSERT INTO finance_lexicon (term, kind, weight, enabled) VALUES ('涨停','sentiment_pos',1.0,true)\\\"))\n        await s.commit()\nasyncio.run(main())\n\"\n```\n4. 词表改动立即生效（NewsMatcher 每 600s 重载 \u002F 可触发 `\u002Fenrichment\u002Frun` 重算新文章）\n\n## 6. 情绪数据怎么用（前端已就绪）\n\n| 位置 | 说明 |\n|---|---|\n| `\u002Frss-news` NewsPanel | 情绪筛选（利好\u002F利空\u002F中性）、利好\u002F利空强度排序、统计栏情绪分布条形图（红利好\u002F绿利空\u002F灰中性） |\n| 个股终端 → 个股资讯 Tab | 每篇标题前利好\u002F利空标签（stock_news 按 huntly_page_id join enrichment） |\n| `\u002Fnews\u002Farticles` API | `sentiment=bullish\u002Fbearish\u002Fneutral` 过滤 + `sort=sentiment_bullish\u002Fsentiment_bearish` |\n| `\u002Fnews\u002Fenrichment\u002Fstats` | 当前筛选下的情绪分布计数（前端统计栏数据源） |\n\n## 7. 相关代码文件\n\n- `backend\u002Fservices\u002Fapi\u002Fnews\u002Fsentiment.py` — FinBERT 懒加载\u002F后台加载\u002F失败重试（`_RETRY_AFTER` 冷却 300s）\n- `backend\u002Fservices\u002Fapi\u002Fnews\u002Fenricher.py` — enrich 管线，0.6字典+0.4FinBERT 融合\n- `backend\u002Fservices\u002Fapi\u002Fnews\u002Fmatcher.py` — Aho-Corasick 匹配 + 字典分\n- `backend\u002Fscripts\u002Fseed_a_share_stocks.py` — finance_lexicon 内置词（含情绪词）\n- `backend\u002Fscripts\u002Fimport_sentiment_lexicon.py` — **5 万情绪词批量导入脚本（可复现）**\n- `backend\u002Fscripts\u002Fdata\u002Ffinance_sentiment_lexicon.tsv` — **合并情绪词表 51,213 词（RSS提炼+DLUT+pysenti+NTUSD）**\n- `backend\u002Fservices\u002Fapi\u002Frouters\u002Fnews.py` — \u002Fnews\u002F* 路由（articles\u002Fenrichment\u002Fstats\u002Fsources）\n- `backend\u002Fservices\u002Fapi\u002Frouters\u002Fstock_terminal.py` — \u002Fstock-terminal\u002Fnews 个股资讯（带情绪标签）\n- `electron\u002Fsrc\u002Ffeatures\u002Fnews\u002Fcomponents\u002FNewsPanel.tsx` — RSS 面板（情绪分布条形图）\n- `electron\u002Fsrc\u002Ffeatures\u002Fstock-terminal\u002Fcomponents\u002Ftabs\u002FNewsTab.tsx` — 个股资讯情绪标签\n- `docker\u002FDockerfile.oss` + `requirements\u002Fai.txt` — transformers 依赖 + FinBERT 权重预下载\n\n## 8. 相关技能\n\n- **[[quantmind-operations]]** — RSS 新闻对接与分析（文章拉取\u002F过滤\u002F富化统计）\n- **[[quantdb-fields]]** — 数据字段口径\n",{"name":5,"description":7},"main","d69096e239f2a89aabc2cd078c47664a9d3b1e11","05-技能与代理开发\u002Fnews-sentiment-finbert",3,"https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts","https:\u002F\u002Fgithub.com\u002Flza6","个人vx：Tf00798\r\n个人✈：@TFGY999","2026-10-08 14:29:22",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add lza6\u002FArmor-piercing-system-prompts --skill news-sentiment-finbert",{"tool":38,"command":39},"bunx","bunx skills add lza6\u002FArmor-piercing-system-prompts --skill news-sentiment-finbert",{"tool":41,"command":42},"pnpm","pnpm dlx skills add lza6\u002FArmor-piercing-system-prompts --skill news-sentiment-finbert","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F05-技能与代理开发\u002Fnews-sentiment-finbert\u002FSKILL.md",true,[],[47,54,59,66,73,80],{"skillNo":48,"name":49,"nameZh":6,"description":50,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":51,"categories":52,"keywords":53,"publishedAt":19},"01M7FFG485KY83TD6HSC4YPCMC","obsidian-clipper-template-creator","Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator\u002FSKILL.md",[],[],{"skillNo":55,"name":49,"nameZh":6,"description":50,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":56,"categories":57,"keywords":58,"publishedAt":19},"01M7FFG0T6AZ0STEJPSB3CX84K","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobsidian-clipper-template-creator-antigravity-awesome-skills-main\u002FSKILL.md",[],[],{"skillNo":60,"name":61,"nameZh":6,"description":62,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":63,"categories":64,"keywords":65,"publishedAt":19},"01M7FFFY0QEM14SKWPDYGMXM2P","python-api","Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP\u002FmAP semantics, and source-style import diagnostics.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Fpython-api\u002FSKILL.md",[],[],{"skillNo":67,"name":68,"nameZh":6,"description":69,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":70,"categories":71,"keywords":72,"publishedAt":19},"01M7FFFVZQZ2GJ0C0TGP7JHD20","file-evaluation","Folder-based PASCAL VOC AP\u002FmAP evaluation from ground-truth and detection text files.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002Fsub-skills\u002Ffile-evaluation\u002FSKILL.md",[],[],{"skillNo":74,"name":75,"nameZh":6,"description":76,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":77,"categories":78,"keywords":79,"publishedAt":19},"01M7FFFSFNYSE77SJ6PE1CK2DA","object-detection-metrics","Guides legacy Object-Detection-Metrics PASCAL VOC AP\u002FmAP evaluation from detection text folders and direct Python API classes.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fobject-detection-metrics\u002FSKILL.md",[],[],{"skillNo":81,"name":82,"nameZh":6,"description":83,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":84,"categories":85,"keywords":86,"publishedAt":19},"01M7FFFP0GCK1RVXBCPQ8APBMT","nx-workspace-patterns","Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.","https:\u002F\u002Fgithub.com\u002Flza6\u002FArmor-piercing-system-prompts\u002Fblob\u002Fmain\u002F07-后端与架构\u002Fnx-workspace-patterns\u002FSKILL.md",[],[]]