{
  "id": 118294,
  "title": "After reading this nice paper, I have 2 questions",
  "url": "/competitions/tensorflow2-question-answering/discussion/118294",
  "author_name": "",
  "post_date": "2019-11-20T15:06:00.713302200Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Here <a href=\"https://arxiv.org/pdf/1901.08634.pdf\">https://arxiv.org/pdf/1901.08634.pdf</a> seems nicely describe how the public LB 0.70 notebook works. But after I read the paper, I have questions:\n1. when training, how do the examples with short answer span form the output as output form is\n[start_tokens=????, end_tokens=???, answer_type_tokens='short'] ? (I can't understand the expression in thesis)\n2. (also related to last question) in evaluation, when we have short answer indice, how to get the long answer span indice without candidates of long answers provided? what is 'top level' really means for usage of the code:\n<code>\n    for c in example.candidates:\n      start = short_span.start_token_idx\n      end = short_span.end_token_idx\n      ## print(c['top_level'],c['start_token'],start,c['end_token'],end)\n      if c[\"top_level\"] and c[\"start_token\"] &lt;= start and c[\"end_token\"] &gt;= end:\n        long_span = Span(c[\"start_token\"], c[\"end_token\"])\n        break\n</code>\nHope everyone understand the bert_utils.py without hardcore hacking the code only!</p>",
  "messages": [
    {
      "id": "677763",
      "postDate": "11/20/2019 15:06:00",
      "content": "<p>Here <a href=\"https://arxiv.org/pdf/1901.08634.pdf\">https://arxiv.org/pdf/1901.08634.pdf</a> seems nicely describe how the public LB 0.70 notebook works. But after I read the paper, I have questions:\n1. when training, how do the examples with short answer span form the output as output form is\n[start_tokens=????, end_tokens=???, answer_type_tokens='short'] ? (I can't understand the expression in thesis)\n2. (also related to last question) in evaluation, when we have short answer indice, how to get the long answer span indice without candidates of long answers provided? what is 'top level' really means for usage of the code:\n<code>\n    for c in example.candidates:\n      start = short_span.start_token_idx\n      end = short_span.end_token_idx\n      ## print(c['top_level'],c['start_token'],start,c['end_token'],end)\n      if c[\"top_level\"] and c[\"start_token\"] &lt;= start and c[\"end_token\"] &gt;= end:\n        long_span = Span(c[\"start_token\"], c[\"end_token\"])\n        break\n</code>\nHope everyone understand the bert_utils.py without hardcore hacking the code only!</p>",
      "rawMarkdown": "Here https://arxiv.org/pdf/1901.08634.pdf seems nicely describe how the public LB 0.70 notebook works. But after I read the paper, I have questions:\n1. when training, how do the examples with short answer span form the output as output form is\n[start_tokens=????, end_tokens=???, answer_type_tokens='short'] ? (I can't understand the expression in thesis)\n2. (also related to last question) in evaluation, when we have short answer indice, how to get the long answer span indice without candidates of long answers provided? what is 'top level' really means for usage of the code:\n```\n    for c in example.candidates:\n      start = short_span.start_token_idx\n      end = short_span.end_token_idx\n      ## print(c['top_level'],c['start_token'],start,c['end_token'],end)\n      if c[\"top_level\"] and c[\"start_token\"] &lt;= start and c[\"end_token\"] &gt;= end:\n        long_span = Span(c[\"start_token\"], c[\"end_token\"])\n        break\n```\nHope everyone understand the bert_utils.py without hardcore hacking the code only!",
      "votes": null
    },
    {
      "id": "677803",
      "postDate": "11/20/2019 15:49:24",
      "content": "<p><a href=\"/httpwwwfszyc\">@httpwwwfszyc</a> \nfor 2nd, if you got a short answer, its respective paragraph is the long answer for this question\n<img src=\"http://ssl.gstatic.com/mljam/compressed_version/brewing_process.gif\" alt=\"\"></p>\n\n<p>for 1st, starttokens and endtokens have an array of 512 which have indices from 0 to 511 based on the answer</p>",
      "rawMarkdown": "httpwwwfszyc \nfor 2nd, if you got a short answer, its respective paragraph is the long answer for this question\n![](http://ssl.gstatic.com/mljam/compressed_version/brewing_process.gif)\n\nfor 1st, starttokens and endtokens have an array of 512 which have indices from 0 to 511 based on the answer",
      "votes": null
    },
    {
      "id": "677857",
      "postDate": "11/20/2019 17:11:10",
      "content": "<p>Thank you for 2nd question.\nBut for 1st question, my problem is: In thesis there's a saying\n<code>\nIf all annotated short spans are contained\nin the instance, we set the start and end target indices to point to the smallest span containing all\nthe annotated short answer spans.\n</code>\nThen my problem to the short answer cases is: what exactly is the \"indice to point to the smallest span containing...\" ? Is this meaning that the short answer indice provided by the training dataset exactly?</p>",
      "rawMarkdown": "Thank you for 2nd question.\nBut for 1st question, my problem is: In thesis there's a saying\n```\nIf all annotated short spans are contained\nin the instance, we set the start and end target indices to point to the smallest span containing all\nthe annotated short answer spans.\n```\nThen my problem to the short answer cases is: what exactly is the \"indice to point to the smallest span containing...\" ? Is this meaning that the short answer indice provided by the training dataset exactly?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 677803,
      "author_name": "seshurajup",
      "author_url": "",
      "post_date": "11/20/2019 15:49:24",
      "content": "<p><a href=\"/httpwwwfszyc\">@httpwwwfszyc</a> \nfor 2nd, if you got a short answer, its respective paragraph is the long answer for this question\n<img src=\"http://ssl.gstatic.com/mljam/compressed_version/brewing_process.gif\" alt=\"\"></p>\n\n<p>for 1st, starttokens and endtokens have an array of 512 which have indices from 0 to 511 based on the answer</p>",
      "votes": null,
      "replies": [
        {
          "id": 677857,
          "author_name": "httpwwwfszyc",
          "author_url": "",
          "post_date": "11/20/2019 17:11:10",
          "content": "<p>Thank you for 2nd question.\nBut for 1st question, my problem is: In thesis there's a saying\n<code>\nIf all annotated short spans are contained\nin the instance, we set the start and end target indices to point to the smallest span containing all\nthe annotated short answer spans.\n</code>\nThen my problem to the short answer cases is: what exactly is the \"indice to point to the smallest span containing...\" ? Is this meaning that the short answer indice provided by the training dataset exactly?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "677763": "Here https://arxiv.org/pdf/1901.08634.pdf seems nicely describe how the public LB 0.70 notebook works. But after I read the paper, I have questions:\n1. when training, how do the examples with short answer span form the output as output form is\n[start_tokens=????, end_tokens=???, answer_type_tokens='short'] ? (I can't understand the expression in thesis)\n2. (also related to last question) in evaluation, when we have short answer indice, how to get the long answer span indice without candidates of long answers provided? what is 'top level' really means for usage of the code:\n```\n    for c in example.candidates:\n      start = short_span.start_token_idx\n      end = short_span.end_token_idx\n      ## print(c['top_level'],c['start_token'],start,c['end_token'],end)\n      if c[\"top_level\"] and c[\"start_token\"] &lt;= start and c[\"end_token\"] &gt;= end:\n        long_span = Span(c[\"start_token\"], c[\"end_token\"])\n        break\n```\nHope everyone understand the bert_utils.py without hardcore hacking the code only!",
    "677803": "httpwwwfszyc \nfor 2nd, if you got a short answer, its respective paragraph is the long answer for this question\n![](http://ssl.gstatic.com/mljam/compressed_version/brewing_process.gif)\n\nfor 1st, starttokens and endtokens have an array of 512 which have indices from 0 to 511 based on the answer",
    "677857": "Thank you for 2nd question.\nBut for 1st question, my problem is: In thesis there's a saying\n```\nIf all annotated short spans are contained\nin the instance, we set the start and end target indices to point to the smallest span containing all\nthe annotated short answer spans.\n```\nThen my problem to the short answer cases is: what exactly is the \"indice to point to the smallest span containing...\" ? Is this meaning that the short answer indice provided by the training dataset exactly?"
  },
  "source": "meta"
}