You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Dataset Card for chinese-english-technical-patent-glossary-chat

chinese-english-technical-patent-glossary-chat 是一個大規模中英技術名詞翻譯之對話集,總計約 4,537,304 筆,基於中華民國智慧財產局(TIPO)公開之專利技術名詞中英對照詞庫自動組裝而成。每筆以 ShareGPT(messages)與 Alpaca(instruction / input / output)雙格式提供,並附帶 IPC 分類與最近修正日期,適合作為繁中 → 英文專利術語翻譯之 SFT 主力語料。

Dataset Details

Dataset Description

中華民國智慧財產局(TIPO)長期維護「專利技術名詞中英對照詞庫」,收錄專利實務中常見之技術名詞及其英文對照,涵蓋機械、電子、化工、生物、資訊等眾多 IPC 分類。本資料集以該詞庫為基礎,透過程式自動組裝為 SFT 對話格式:

  • system:固定提示語 請將中文翻譯成英文
  • user:中文技術名詞;
  • assistant:對應英文。

由於原始詞庫之條目數量龐大,經組裝後總計約 4,537,304 筆。每筆亦保留 IPC 分類(term_ipc)與 TIPO 最近修正日期(date_modify),方便下游依分類篩選或做術語鑑別。

Dataset Sources

Uses

Direct Use

本資料集主要設計用於:

  • 繁中 → 英文專利術語翻譯之 SFT 訓練;
  • 建立繁中技術名詞之中英對照字典;
  • 作為專利文件翻譯模型之專業術語補強;
  • term_ipc 分類建立領域特化之術語翻譯模型。

Out-of-Scope Use

本資料集不適用於下列用途:

  • 作為一般英文 → 中文翻譯之訓練資料(本資料集為中 → 英單向);
  • 作為專利完整內文之翻譯訓練(本資料集僅為術語級別);
  • 作為非專利領域之一般技術名詞對照。

Dataset Structure

{
  "messages": [
    {"role": "system", "content": "請將中文翻譯成英文"},
    {"role": "user", "content": "空氣預熱"},
    {"role": "assistant", "content": "air preheat"}
  ],
  "instruction": "請將中文翻譯成英文",
  "input": "空氣預熱",
  "output": "air preheat",
  "term_ipc": "-F/22/D/1/38 -F/23/H/1/2 -F/22/D/1/36",
  "date_modify": "20110223"
}
欄位 型別 說明
messages list[{role, content}] ShareGPT 格式對話(system + user + assistant)
instruction string Alpaca instruction(= system prompt)
input string Alpaca input(中文技術名詞)
output string Alpaca output(英文對照)
term_ipc string IPC 分類(可能含多個分類,以空格分隔)
date_modify string TIPO 最近修正日期
統計項目 數值
總筆數 ~4,537,304
檔案 datasets.jsonl(~2.2 GB)

Dataset Creation

Curation Rationale

繁中專利翻譯之訓練資料長期以來都集中於完整專利文件,缺乏「技術名詞級別」之精準對照。本資料集以 TIPO 之官方詞庫為基礎,補足術語級訓練資料之空白。

Source Data

Data Collection and Processing

  • 來源:TIPO 之「專利技術名詞中英對照詞庫」;
  • 處理:以 Python 程式讀取原始詞庫,對每筆中英對照生成固定 prompt 之 ShareGPT / Alpaca 雙格式;
  • 格式:單檔 JSONL。

Who are the source data producers?

原始詞庫由中華民國智慧財產局(TIPO)維護,條目來自專利實務中之審查與翻譯工作。

Annotations

Annotation process

本資料集不包含額外人工標註,所有對話均為自動組裝。

Who are the annotators?

不適用。

Personal and Sensitive Information

本資料集為技術名詞之中英對照,不涉及個人資訊。

Bias, Risks, and Limitations

  • 系統提示固定:system prompt 為單一之「請將中文翻譯成英文」,模型易對提示過擬合;
  • 單向翻譯:僅中 → 英,無反向翻譯;
  • 領域集中:以專利術語為主,不適用於日常用語;
  • 原始詞庫之錯誤:TIPO 詞庫為長期累積之條目,部分對照可能已過時或存在錯誤;
  • 大量短條目:多數條目極短(幾個字),對於 context length 之訓練不具參考價值。

Recommendations

建議使用者:

  • 訓練時混入多種 system prompt 之變體以降低過擬合風險;
  • 搭配完整專利文件之翻譯語料(如 TIPO 專利全文翻譯)一併訓練;
  • term_ipc 進行領域特化訓練;
  • 商業使用前查閱 TIPO 詞庫之原始授權條款。

Citation

@misc{chinese-english-technical-patent-glossary-chat,
  title        = {chinese-english-technical-patent-glossary-chat: Large-scale Chinese-English Technical Patent Term Translation SFT Corpus},
  author       = {Liang Hsun Huang},
  year         = {2024},
  howpublished = {\url{https://huggingface.co/datasets/lianghsun/chinese-english-technical-patent-glossary-chat}},
  note         = {~4.5M Chinese-English patent term translation pairs in ShareGPT and Alpaca formats, built from TIPO glossary.}
}

Dataset Card Authors

Liang Hsun Huang

Dataset Card Contact

Liang Hsun Huang

Downloads last month
8