{
  "id": 131064,
  "title": "Is Negative Sampling the huge trick in this competation?",
  "url": "/competitions/tensorflow2-question-answering/discussion/131064",
  "author_name": "",
  "post_date": "2020-02-18T03:07:38.689255Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I have seen Negative Sampling in all top solutions 💯 </p>",
  "messages": [
    {
      "id": "748824",
      "postDate": "02/18/2020 03:07:38",
      "content": "<p>I have seen Negative Sampling in all top solutions 💯 </p>",
      "rawMarkdown": "I have seen Negative Sampling in all top solutions 💯",
      "votes": null
    },
    {
      "id": "748875",
      "postDate": "02/18/2020 04:35:22",
      "content": "<p>It's the quantity and quality of negative sampling that made the difference. To clarify, the bert joint paper had already addressed the problem of imbalance during training by negative sampling with a rate of 2% but it wasn't the best solution to the problem as you remain unsure of the quality of that 2% and its effect on training. </p>\n\n<p>What top scorers did was to either increase the empty answers to long answers ratio like what <a href=\"/seesee\">@seesee</a> did or increase the quality of the samples like what <a href=\"/wowfattie\">@wowfattie</a> did. Kudos to both of them.</p>",
      "rawMarkdown": "It's the quantity and quality of negative sampling that made the difference. To clarify, the bert joint paper had already addressed the problem of imbalance during training by negative sampling with a rate of 2% but it wasn't the best solution to the problem as you remain unsure of the quality of that 2% and its effect on training. \n\nWhat top scorers did was to either increase the empty answers to long answers ratio like what @seesee did or increase the quality of the samples like what @wowfattie did. Kudos to both of them.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 748875,
      "author_name": "msheriey",
      "author_url": "",
      "post_date": "02/18/2020 04:35:22",
      "content": "<p>It's the quantity and quality of negative sampling that made the difference. To clarify, the bert joint paper had already addressed the problem of imbalance during training by negative sampling with a rate of 2% but it wasn't the best solution to the problem as you remain unsure of the quality of that 2% and its effect on training. </p>\n\n<p>What top scorers did was to either increase the empty answers to long answers ratio like what <a href=\"/seesee\">@seesee</a> did or increase the quality of the samples like what <a href=\"/wowfattie\">@wowfattie</a> did. Kudos to both of them.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "748824": "I have seen Negative Sampling in all top solutions 💯",
    "748875": "It's the quantity and quality of negative sampling that made the difference. To clarify, the bert joint paper had already addressed the problem of imbalance during training by negative sampling with a rate of 2% but it wasn't the best solution to the problem as you remain unsure of the quality of that 2% and its effect on training. \n\nWhat top scorers did was to either increase the empty answers to long answers ratio like what @seesee did or increase the quality of the samples like what @wowfattie did. Kudos to both of them."
  },
  "source": "meta"
}