{
  "id": 120025,
  "title": "What is the significance of multiple long_answer_candidates ?",
  "url": "/competitions/tensorflow2-question-answering/discussion/120025",
  "author_name": "",
  "post_date": "2019-12-03T07:13:00.627817200Z",
  "votes": null,
  "comment_count": 5,
  "views": 0,
  "content": "<p>I have some doubt, what is the purpose of multiple long_answer_candidates? Is it useful somehow? If not any other kind of significance in terms of understanding?</p>",
  "messages": [
    {
      "id": "686457",
      "postDate": "12/03/2019 07:13:00",
      "content": "<p>I have some doubt, what is the purpose of multiple long_answer_candidates? Is it useful somehow? If not any other kind of significance in terms of understanding?</p>",
      "rawMarkdown": "I have some doubt, what is the purpose of multiple long_answer_candidates? Is it useful somehow? If not any other kind of significance in terms of understanding?",
      "votes": null
    },
    {
      "id": "686563",
      "postDate": "12/03/2019 09:47:05",
      "content": "<p>We can choose long answer from long answer candidates.\nIt is easier than predicting all kinds of possibilities for <code>start_token</code> and <code>end_token</code> in document tokens, and this is the official setting of <a href=\"https://ai.google.com/research/NaturalQuestions\">Natural Questions</a>.</p>\n\n<p>For example, the official baseline model <a href=\"https://github.com/google-research/language/tree/master/language/question_answering/bert_joint\">bert-joint</a> predict only short answer if sample has short answer, then choose long answer from candidates as predicted short answer should be within the long answer span.</p>",
      "rawMarkdown": "We can choose long answer from long answer candidates.\nIt is easier than predicting all kinds of possibilities for `start_token` and `end_token` in document tokens, and this is the official setting of [Natural Questions](https://ai.google.com/research/NaturalQuestions).\n\nFor example, the official baseline model [bert-joint](https://github.com/google-research/language/tree/master/language/question_answering/bert_joint) predict only short answer if sample has short answer, then choose long answer from candidates as predicted short answer should be within the long answer span.",
      "votes": null
    },
    {
      "id": "686938",
      "postDate": "12/03/2019 17:57:45",
      "content": "<p>Okay thanks, but the what  candidate id inside annotations field used for?</p>",
      "rawMarkdown": "Okay thanks, but the what  candidate id inside annotations field used for?",
      "votes": null
    },
    {
      "id": "687072",
      "postDate": "12/03/2019 22:04:49",
      "content": "<p>I think that the candidate_index field tells you which of the long answer candidates (if any) is the correct one.</p>",
      "rawMarkdown": "I think that the candidate_index field tells you which of the long answer candidates (if any) is the correct one.",
      "votes": null
    },
    {
      "id": "687233",
      "postDate": "12/04/2019 05:24:08",
      "content": "<p>M kind of not getting it here. If candidate_index is the correct answer index, what is the purpose of other start_index and end_index, provided in long_answer_candidate list . Is my question valid?</p>",
      "rawMarkdown": "M kind of not getting it here. If candidate_index is the correct answer index, what is the purpose of other start_index and end_index, provided in long_answer_candidate list . Is my question valid?",
      "votes": null
    },
    {
      "id": "687339",
      "postDate": "12/04/2019 09:12:42",
      "content": "<p>There is a more detailed explanation of the long answer candidates here:</p>\n\n<p><a href=\"https://github.com/google-research-datasets/natural-questions\">https://github.com/google-research-datasets/natural-questions</a></p>\n\n<p>The long answer candidates are provided for convenience and are just the locations of certain html tags.  The start token and end token entries in the long answer dict (in the \"annotations\" list) should match the start and end tokens of the long answer candidate indicated by the candidate index. </p>",
      "rawMarkdown": "There is a more detailed explanation of the long answer candidates here:\n\nhttps://github.com/google-research-datasets/natural-questions\n\nThe long answer candidates are provided for convenience and are just the locations of certain html tags.  The start token and end token entries in the long answer dict (in the \"annotations\" list) should match the start and end tokens of the long answer candidate indicated by the candidate index.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 686563,
      "author_name": "kentaronakanishi",
      "author_url": "",
      "post_date": "12/03/2019 09:47:05",
      "content": "<p>We can choose long answer from long answer candidates.\nIt is easier than predicting all kinds of possibilities for <code>start_token</code> and <code>end_token</code> in document tokens, and this is the official setting of <a href=\"https://ai.google.com/research/NaturalQuestions\">Natural Questions</a>.</p>\n\n<p>For example, the official baseline model <a href=\"https://github.com/google-research/language/tree/master/language/question_answering/bert_joint\">bert-joint</a> predict only short answer if sample has short answer, then choose long answer from candidates as predicted short answer should be within the long answer span.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 686938,
      "author_name": "s4sarath",
      "author_url": "",
      "post_date": "12/03/2019 17:57:45",
      "content": "<p>Okay thanks, but the what  candidate id inside annotations field used for?</p>",
      "votes": null,
      "replies": [
        {
          "id": 687072,
          "author_name": "particlebbq",
          "author_url": "",
          "post_date": "12/03/2019 22:04:49",
          "content": "<p>I think that the candidate_index field tells you which of the long answer candidates (if any) is the correct one.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 687233,
          "author_name": "s4sarath",
          "author_url": "",
          "post_date": "12/04/2019 05:24:08",
          "content": "<p>M kind of not getting it here. If candidate_index is the correct answer index, what is the purpose of other start_index and end_index, provided in long_answer_candidate list . Is my question valid?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 687339,
          "author_name": "particlebbq",
          "author_url": "",
          "post_date": "12/04/2019 09:12:42",
          "content": "<p>There is a more detailed explanation of the long answer candidates here:</p>\n\n<p><a href=\"https://github.com/google-research-datasets/natural-questions\">https://github.com/google-research-datasets/natural-questions</a></p>\n\n<p>The long answer candidates are provided for convenience and are just the locations of certain html tags.  The start token and end token entries in the long answer dict (in the \"annotations\" list) should match the start and end tokens of the long answer candidate indicated by the candidate index. </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "686457": "I have some doubt, what is the purpose of multiple long_answer_candidates? Is it useful somehow? If not any other kind of significance in terms of understanding?",
    "686563": "We can choose long answer from long answer candidates.\nIt is easier than predicting all kinds of possibilities for `start_token` and `end_token` in document tokens, and this is the official setting of [Natural Questions](https://ai.google.com/research/NaturalQuestions).\n\nFor example, the official baseline model [bert-joint](https://github.com/google-research/language/tree/master/language/question_answering/bert_joint) predict only short answer if sample has short answer, then choose long answer from candidates as predicted short answer should be within the long answer span.",
    "686938": "Okay thanks, but the what  candidate id inside annotations field used for?",
    "687072": "I think that the candidate_index field tells you which of the long answer candidates (if any) is the correct one.",
    "687233": "M kind of not getting it here. If candidate_index is the correct answer index, what is the purpose of other start_index and end_index, provided in long_answer_candidate list . Is my question valid?",
    "687339": "There is a more detailed explanation of the long answer candidates here:\n\nhttps://github.com/google-research-datasets/natural-questions\n\nThe long answer candidates are provided for convenience and are just the locations of certain html tags.  The start token and end token entries in the long answer dict (in the \"annotations\" list) should match the start and end tokens of the long answer candidate indicated by the candidate index."
  },
  "source": "meta"
}