{
  "id": 161048,
  "title": "Has anyone tried to split 6 test langs into classes?",
  "url": "/competitions/jigsaw-multilingual-toxic-comment-classification/discussion/161048",
  "author_name": "godelscat",
  "post_date": "2020-06-23T15:16:27.834000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi. Congrats to all winners, and thanks for sharing your winning solutions, I really learnt a lot today.</p>\n\n<p>In my case, also mentioned in <a href=\"https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification/discussion/160862\">1st place solution</a>, monolingual model helps a lot while blending. In this paper <a href=\"https://arxiv.org/pdf/1908.09324.pdf\">arxiv:1908.09324</a>, they shown that clustering similar languages together and training one single model could outperform monolingual models. As for this competition, following their results[Fig. 3], we can split test languages into three classes, (it, pt, fr, es ), (tr, ), (ru, ), and train three models. I guess this may boost final model a bit.</p>\n\n<p>Has anyone tried similar ideas? </p>",
  "messages": [
    {
      "id": 898534,
      "postDate": "2020-06-23T15:16:27.833Z",
      "content": "<p>Hi. Congrats to all winners, and thanks for sharing your winning solutions, I really learnt a lot today.</p>\n\n<p>In my case, also mentioned in <a href=\"https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification/discussion/160862\">1st place solution</a>, monolingual model helps a lot while blending. In this paper <a href=\"https://arxiv.org/pdf/1908.09324.pdf\">arxiv:1908.09324</a>, they shown that clustering similar languages together and training one single model could outperform monolingual models. As for this competition, following their results[Fig. 3], we can split test languages into three classes, (it, pt, fr, es ), (tr, ), (ru, ), and train three models. I guess this may boost final model a bit.</p>\n\n<p>Has anyone tried similar ideas? </p>",
      "rawMarkdown": "Hi. Congrats to all winners, and thanks for sharing your winning solutions, I really learnt a lot today.\n\nIn my case, also mentioned in [1st place solution](https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification/discussion/160862), monolingual model helps a lot while blending. In this paper [arxiv:1908.09324](https://arxiv.org/pdf/1908.09324.pdf), they shown that clustering similar languages together and training one single model could outperform monolingual models. As for this competition, following their results[Fig. 3], we can split test languages into three classes, (it, pt, fr, es ), (tr, ), (ru, ), and train three models. I guess this may boost final model a bit.\n\nHas anyone tried similar ideas? ",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "898534": "Hi. Congrats to all winners, and thanks for sharing your winning solutions, I really learnt a lot today.\n\nIn my case, also mentioned in [1st place solution](https://www.kaggle.com/c/jigsaw-multilingual-toxic-comment-classification/discussion/160862), monolingual model helps a lot while blending. In this paper [arxiv:1908.09324](https://arxiv.org/pdf/1908.09324.pdf), they shown that clustering similar languages together and training one single model could outperform monolingual models. As for this competition, following their results[Fig. 3], we can split test languages into three classes, (it, pt, fr, es ), (tr, ), (ru, ), and train three models. I guess this may boost final model a bit.\n\nHas anyone tried similar ideas? "
  }
}