{
  "id": 123392,
  "title": "How does the joint model fit this competition?",
  "url": "/competitions/tensorflow2-question-answering/discussion/123392",
  "author_name": "",
  "post_date": "2019-12-27T09:50:19.560918900Z",
  "votes": null,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I saw a lot of pub kernels using the bert joint model. But according to <a href=\"https://arxiv.org/abs/1901.08634\">the paper</a>  and <a href=\"https://www.kaggle.com/mmmarchetti/tensorflow-2-0-bert-yes-no-answers\">this post</a>, the model should not be able to predict spans. Instead it just use a given span then classify the type. But the test jsonl does not give short answer spans. In fact it doesn't have short answer at all. So unless I'm wrong, why are there so many people using the joint model?</p>",
  "messages": [
    {
      "id": "704305",
      "postDate": "12/27/2019 09:50:19",
      "content": "<p>I saw a lot of pub kernels using the bert joint model. But according to <a href=\"https://arxiv.org/abs/1901.08634\">the paper</a>  and <a href=\"https://www.kaggle.com/mmmarchetti/tensorflow-2-0-bert-yes-no-answers\">this post</a>, the model should not be able to predict spans. Instead it just use a given span then classify the type. But the test jsonl does not give short answer spans. In fact it doesn't have short answer at all. So unless I'm wrong, why are there so many people using the joint model?</p>",
      "rawMarkdown": "I saw a lot of pub kernels using the bert joint model. But according to [the paper](https://arxiv.org/abs/1901.08634)  and [this post](https://www.kaggle.com/mmmarchetti/tensorflow-2-0-bert-yes-no-answers), the model should not be able to predict spans. Instead it just use a given span then classify the type. But the test jsonl does not give short answer spans. In fact it doesn't have short answer at all. So unless I'm wrong, why are there so many people using the joint model?",
      "votes": null
    },
    {
      "id": "704472",
      "postDate": "12/27/2019 14:23:20",
      "content": "<p>Bert-joint model predict both span and type. You can get this by <code>The key insights in our approach ~</code> in Introduction of the paper.</p>",
      "rawMarkdown": "Bert-joint model predict both span and type. You can get this by `The key insights in our approach ~` in Introduction of the paper.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 704472,
      "author_name": "kentaronakanishi",
      "author_url": "",
      "post_date": "12/27/2019 14:23:20",
      "content": "<p>Bert-joint model predict both span and type. You can get this by <code>The key insights in our approach ~</code> in Introduction of the paper.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "704305": "I saw a lot of pub kernels using the bert joint model. But according to [the paper](https://arxiv.org/abs/1901.08634)  and [this post](https://www.kaggle.com/mmmarchetti/tensorflow-2-0-bert-yes-no-answers), the model should not be able to predict spans. Instead it just use a given span then classify the type. But the test jsonl does not give short answer spans. In fact it doesn't have short answer at all. So unless I'm wrong, why are there so many people using the joint model?",
    "704472": "Bert-joint model predict both span and type. You can get this by `The key insights in our approach ~` in Introduction of the paper."
  },
  "source": "meta"
}