{
  "id": 74150,
  "title": "Conditional Random Fields",
  "url": "/competitions/quora-insincere-questions-classification/discussion/74150",
  "author_name": "",
  "post_date": "2018-12-09T08:32:19.094968200Z",
  "votes": 13,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I also experimented with using Condional Random fields on top of LSTM or GRU layer. You can find a keras implementation in one of my recent kernels:</p>\n\n<p><a href=\"https://www.kaggle.com/christofhenkel/gru-crf\">GRU + CRF</a></p>\n\n<p>Might be interesting to discuss some opinions on using CRF for Quora here.</p>",
  "messages": [
    {
      "id": "435992",
      "postDate": "12/09/2018 08:32:19",
      "content": "<p>I also experimented with using Condional Random fields on top of LSTM or GRU layer. You can find a keras implementation in one of my recent kernels:</p>\n\n<p><a href=\"https://www.kaggle.com/christofhenkel/gru-crf\">GRU + CRF</a></p>\n\n<p>Might be interesting to discuss some opinions on using CRF for Quora here.</p>",
      "rawMarkdown": "I also experimented with using Condional Random fields on top of LSTM or GRU layer. You can find a keras implementation in one of my recent kernels:\n\n[GRU + CRF][1]\n\nMight be interesting to discuss some opinions on using CRF for Quora here.\n\n  [1]: https://www.kaggle.com/christofhenkel/gru-crf",
      "votes": null
    },
    {
      "id": "436197",
      "postDate": "12/09/2018 21:24:28",
      "content": "<p>I ran an experiment by following this paper, BiLSTM-CNNs-CRF [1], without any fine tuning, the f1 score is around 0.684 </p>\n\n<p>[1] <a href=\"https://arxiv.org/pdf/1603.01354.pdf\">https://arxiv.org/pdf/1603.01354.pdf</a></p>",
      "rawMarkdown": "I ran an experiment by following this paper, BiLSTM-CNNs-CRF [1], without any fine tuning, the f1 score is around 0.684 \n\n[1] https://arxiv.org/pdf/1603.01354.pdf",
      "votes": null
    },
    {
      "id": "436575",
      "postDate": "12/10/2018 15:00:51",
      "content": "<p>This doesn't make sense to me.  CRFs impose a conditional probabilistic model on the labelling of adjacent tokens (at least for linear chain CRFs), in a sequence labelling task.  However, this problem is not a sequence labelling task - it's a sequence classification task, and there are no adjacent labels being predicted over which a CRF would appear to be useful.\nCan you clarify how you are mapping to a sequence labelling task?</p>",
      "rawMarkdown": "This doesn't make sense to me.  CRFs impose a conditional probabilistic model on the labelling of adjacent tokens (at least for linear chain CRFs), in a sequence labelling task.  However, this problem is not a sequence labelling task - it's a sequence classification task, and there are no adjacent labels being predicted over which a CRF would appear to be useful.\nCan you clarify how you are mapping to a sequence labelling task?",
      "votes": null
    },
    {
      "id": "1234932",
      "postDate": "03/11/2021 17:32:21",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/konohayui\" target=\"_blank\">@konohayui</a> , can you share your implementatios about that paper?</p>",
      "rawMarkdown": "Hi @konohayui , can you share your implementatios about that paper?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1234932,
      "author_name": "rodyoukai",
      "author_url": "",
      "post_date": "03/11/2021 17:32:21",
      "content": "<p>Hi <a href=\"https://www.kaggle.com/konohayui\" target=\"_blank\">@konohayui</a> , can you share your implementatios about that paper?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 436197,
      "author_name": "konohayui",
      "author_url": "",
      "post_date": "12/09/2018 21:24:28",
      "content": "<p>I ran an experiment by following this paper, BiLSTM-CNNs-CRF [1], without any fine tuning, the f1 score is around 0.684 </p>\n\n<p>[1] <a href=\"https://arxiv.org/pdf/1603.01354.pdf\">https://arxiv.org/pdf/1603.01354.pdf</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 436575,
      "author_name": "stevedraper",
      "author_url": "",
      "post_date": "12/10/2018 15:00:51",
      "content": "<p>This doesn't make sense to me.  CRFs impose a conditional probabilistic model on the labelling of adjacent tokens (at least for linear chain CRFs), in a sequence labelling task.  However, this problem is not a sequence labelling task - it's a sequence classification task, and there are no adjacent labels being predicted over which a CRF would appear to be useful.\nCan you clarify how you are mapping to a sequence labelling task?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "435992": "I also experimented with using Condional Random fields on top of LSTM or GRU layer. You can find a keras implementation in one of my recent kernels:\n\n[GRU + CRF][1]\n\nMight be interesting to discuss some opinions on using CRF for Quora here.\n\n  [1]: https://www.kaggle.com/christofhenkel/gru-crf",
    "436197": "I ran an experiment by following this paper, BiLSTM-CNNs-CRF [1], without any fine tuning, the f1 score is around 0.684 \n\n[1] https://arxiv.org/pdf/1603.01354.pdf",
    "436575": "This doesn't make sense to me.  CRFs impose a conditional probabilistic model on the labelling of adjacent tokens (at least for linear chain CRFs), in a sequence labelling task.  However, this problem is not a sequence labelling task - it's a sequence classification task, and there are no adjacent labels being predicted over which a CRF would appear to be useful.\nCan you clarify how you are mapping to a sequence labelling task?",
    "1234932": "Hi @konohayui , can you share your implementatios about that paper?"
  },
  "source": "meta"
}