mlqa_en_zh_tw / README.md
zetavg's picture
Update README.md
18d753a
metadata
license: cc-by-3.0
task_categories:
  - question-answering
  - translation
language:
  - zh
  - en
size_categories:
  - 1K<n<10K
pretty_name: MLQA en-zh_tw

MLQA (MultiLingual Question Answering) 中英雙語問答資料集,為原始 MLQA 資料集轉換為台灣正體中文的版本,並將中文與英語版本的相同項目合併,方便供雙語語言模型使用。(致謝:BYVoid/OpenCCvinta/pangu.js

分為 dev 以及 test 兩個 split,各有 302 及 2986 組資料。

範本:

[
  {
    "title": {
      "en": "Curling at the 2014 Winter Olympics",
      "zh_tw": "2014 年冬季奧林匹克運動會冰壺比賽"
    },
    "paragraphs": [
      {
        "context": {
          "en": "Qualification to the curling tournaments at the Winter Olympics was determined through two methods. Nations could qualify teams by earning qualification points from performances at the 2012 and 2013 World Curling Championships. Teams could also qualify through an Olympic qualification event which was held in the autumn of 2013. Seven nations qualified teams via World Championship qualification points, while two nations qualified through the qualification event. As host nation, Russia qualified teams automatically, thus making a total of ten teams per gender in the curling tournaments.",
          "zh_tw": "本屆冬奧會冰壺比賽參加資格有兩種辦法可以取得。各國家或地區可以透過 2012 年和 2013 年的世界冰壺錦標賽,也可以透過 2013 年 12 月舉辦的一次冬奧會資格賽來取得資格。七個國家透過兩屆世錦賽積分之和來獲得資格,兩個國家則透過冬奧會資格賽。作為主辦國,俄羅斯自動獲得參賽資格,這樣就確定了冬奧會冰壺比賽的男女各十支參賽隊伍。"
        },
        "qas": [
          {
            "id": "b08184972e38a79c47d01614aa08505bb3c9b680",
            "question": {
              "zh_tw": "俄羅斯有多少隊獲得參賽資格?",
              "en": "How many teams did Russia qualify for?"
            },
            "answers": {
              "en": [
                {
                  "text": "ten teams",
                  "answer_start": 543
                }
              ],
              "zh_tw": [
                {
                  "text": "十支",
                  "answer_start": 161
                }
              ]
            }
          }
        ]
      }
    ]
  }
]

其餘資訊,詳見:https://github.com/facebookresearch/mlqa

原始資料集

https://github.com/facebookresearch/mlqa ,分別取其中 devtest split 的 context-zh-question-zhcontext-zh-question-encontext-en-question-zh,總共六個檔案。

轉換程序

  1. OpenCC 使用 s2twp.json 配置,將簡體中文轉換為台灣正體中文與臺灣常用詞彙。
  2. 使用 Python 版本的 pangu.js 在中英文(全形與半形文字)之間加上空格。
  3. 將中英文資料集中的相同項目進行合併。

關於轉換的詳細過程,請見:https://github.com/zetavg/LLM-Research/blob/bba5ff7/MLQA_Dataset_Converter_(en_zh_tw).ipynb

已知問題

  • 有些項目的 titleparagraphcontext、問題或是答案可能會缺少其中一種語言的版本。
  • 部分問題與答案可能存在理解偏誤或歧異,例如上方所列範本「2014 年冬季奧林匹克運動會冰壺比賽」的問題「俄羅斯有多少隊獲得參賽資格?」與答案。
  • paragraphcontext 在不同語言的版本下可能長度與涵蓋的內容範圍有很大的落差。例如在 development split 中,title 為 “Adobe Photoshop” 的項目:
    • zh_tw 只有兩句話:「Adobe Photoshop,簡稱 “PS”,是一個由 Adobe 開發和發行的影象處理軟體。該軟體釋出在 Windows 和 Mac OS 上。」
    • en 則是一個段落:“Adobe Photoshop is a raster graphics editor developed and published by Adobe Inc. for Windows and macOS. It was originally created in 1988 by Thomas and John Knoll. Since then, this software has become the industry standard not only in raster graphics editing, but in digital art as a whole. … (下略 127 字)”