{ "id": 287940, "title": "28th Place Solution(Contains additional reading material)", "content": "

Congratulations to the winners in the medal place! Also, thanks for everyone contributing to this competition!

\n

Then, I'm happy that I get my second silver medal in this cometition which is my first time to participate a multilingual languge competition (maybe I have some chance to get a gold one 😆). My effective solution almost come from new translated data, ensembling method and diversity model with different trained data.

\n\n
    \n

  1. logits average (like almost everyone do)

    2. \n

  2. logits sum vote:
    \nThis method is designed for different pretrained model with different tokenizer. I choose the top 3 different predicted answers for a single sample, keeping span score, then sum all span score from different model, finally get the top 1 ranking answer as the final result.

    \n

    The two methods have the similar socre in private /publie LB.

    3. \n
\n\n

Besides, I also tried some other methods, having not get a better result. beacure I have not my own training GPU and not do enough experiment, my teammates does😨. So I list it for your reference hoping helpful:

\n\n

Finally, thank @thedrcat @abhishek @nbroad for your notebooks( Extremely agonizing text tokenizing processing🙄 ! ).

\n

Update:

\n

Score vote maybe a better ensembling method in this competition for private LB! My best submissions also come from it but I havn't choosed them😂
\n\"Wechat-IMG1495.png\"

", "comments": [ { "id": 1584935, "author_name": "Mintwater", "author_url": "", "post_date": "2021-11-16T21:56:26.207000", "content": "

Amazing job! Big congrats!

", "votes": 1, "replies": [ { "id": 1585010, "author_name": "van73", "author_url": "", "post_date": "2021-11-17T01:47:04.510000", "content": "

Thanks!
\nCongratulations too！

", "votes": 0, "replies": [] } ] }, { "id": 1584476, "author_name": "Ayush Singh", "author_url": "", "post_date": "2021-11-16T15:01:25.040000", "content": "

Well written and nice solution .. congrats

", "votes": 1, "replies": [ { "id": 1584611, "author_name": "van73", "author_url": "", "post_date": "2021-11-16T16:19:01.213000", "content": "

Thanks for your comment!

", "votes": 0, "replies": [] } ] } ], "comment_count": 4, "messages": [ { "id": 1583719, "postDate": "2021-11-16T02:54:51.777Z", "content": "

Congratulations to the winners in the medal place! Also, thanks for everyone contributing to this competition!

\n

Then, I'm happy that I get my second silver medal in this cometition which is my first time to participate a multilingual languge competition (maybe I have some chance to get a gold one 😆). My effective solution almost come from new translated data, ensembling method and diversity model with different trained data.

\n\n
    \n

  1. logits average (like almost everyone do)

    2. \n

  2. logits sum vote:
    \nThis method is designed for different pretrained model with different tokenizer. I choose the top 3 different predicted answers for a single sample, keeping span score, then sum all span score from different model, finally get the top 1 ranking answer as the final result.

    \n

    The two methods have the similar socre in private /publie LB.

    3. \n
\n\n

Besides, I also tried some other methods, having not get a better result. beacure I have not my own training GPU and not do enough experiment, my teammates does😨. So I list it for your reference hoping helpful:

\n\n

Finally, thank @thedrcat @abhishek @nbroad for your notebooks( Extremely agonizing text tokenizing processing🙄 ! ).

\n

Update:

\n

Score vote maybe a better ensembling method in this competition for private LB! My best submissions also come from it but I havn't choosed them😂
\n\"Wechat-IMG1495.png\"

", "rawMarkdown": "Congratulations to the winners in the medal place! Also, thanks for everyone contributing to this competition!\n\nThen, I'm happy that I get my second silver medal in this cometition which is my first time to participate a multilingual languge competition (maybe I have some chance to get a gold one 😆). My effective solution almost come from new translated data, ensembling method and diversity model with different trained data.\n\n- translated data :\n[https://www.kaggle.com/vanle73/nq-translated-data](https://www.kaggle.com/vanle73/nq-translated-data)\nIt contains two version data with different number samples. I translated with Google API before I truncated the span keeping the maxium length less than 5000.\n- ensembling method\n1. logits average (like almost everyone do)\n1. logits sum vote:\n This method is designed for different pretrained model with different tokenizer. I choose the top 3 different predicted answers for a single sample, keeping span score, then sum all span score from different model, finally get the top 1 ranking answer as the final result. \n \n The two methods have the similar socre in private /publie LB.\n\n- diversity model with different trained data\nAt the beginning, I regard this competition as a weakly supervised problem, So I try to denoise the data using HardEm algorithm iteratively, it actually have a obvious improvement relative to just using origin data, but still below the baseline. Then I try to do neg-samlping to filter some samples, it still have some improvement, but a low score. Then I found this competition is a \"data competition\" after I add xquad/mlqa data. \nSo I just translated the similar data ---NQ which I joined last two years. It actually bring more benefit. My finally ensembling method also contain all model mentioned above because of diversity.\n\nBesides, I also tried some other methods, having not get a better result. beacure I have not my own training GPU and not do enough experiment, my teammates does😨. So I list it for your reference hoping helpful:\n\n- Pretained model :\n[indic-bert](https://indicnlp.ai4bharat.org/indic-bert/)\n[hindiBert](https://huggingface.co/mrm8488/HindiBERTa)\n[roberta-hindi](https://huggingface.co/flax-community/roberta-hindi)\n[info-xlm](https://github.com/microsoft/unilm/tree/master/infoxlm)\n[xlm-align](https://github.com/CZWin32768/XLM-Align)\n\n- Different model structure:\n[FILTER](https://github.com/yuwfan/FILTER)\nVECO: Variable and Flexible Cross-lingual Pre-training for Language Understanding and Generation\n\n- Ensembleing:\n[BERT for Question Answering on SQuAD 2.0](https://web.stanford.edu/class/archive/cs/cs224n/cs224n.1194/reports/default/15848021.pdf)\n[Gestalt: a Stacking Ensemble for SQuAD2.0](https://www.linkedin.com/pulse/stacking-ensemble-squad20-mohamed-el-geish/)\n[Applying Ensembling Methods to BERT to Boost Model Performance](https://web.stanford.edu/class/archive/cs/cs224n/cs224n.1194/reports/default/15775971.pdf)\n[Ensemble ALBERT and RoBERTa for Span Prediction in Question Answering](https://aclanthology.org/2021.dialdoc-1.9.pdf)\n\n- offline translated model:\nHi2en:\nhttps://huggingface.co/Helsinki-NLP/opus-mt-hi-en\nEn2hi:\nhttps://huggingface.co/Helsinki-NLP/opus-mt-en-hi\nTmail2en:\nhttps://huggingface.co/Helsinki-NLP/opus-mt-dra-en\nEn2tmail:\nhttps://huggingface.co/Helsinki-NLP/opus-mt-en-dra\n mBART-50:\nhttps://huggingface.co/facebook/mbart-large-50\n\n- weakly supervised method:\nMML& HardEM:\n A Discrete Hard EM Approach for Weakly Supervised Question Answering\n https://doi.org/10.18653/v1/D19-1284. \nMutual Information Maximization Approach:\n A Mutual Information Maximization Approach for the Spurious Solution Problem in Weakly Supervised Question Answering\n https://arxiv.org/abs/2106.07174\n\nFinally, thank @thedrcat @abhishek @nbroad for your notebooks( Extremely agonizing text tokenizing processing🙄 ! ).\n\nUpdate:\n\nScore vote maybe a better ensembling method in this competition for private LB! My best submissions also come from it but I havn't choosed them😂\n[![Wechat-IMG1495.png](https://i.postimg.cc/xdL305yM/Wechat-IMG1495.png)](https://postimg.cc/Y4qWR1PC)", "votes": 10 }, { "id": 1584935, "postDate": "2021-11-16T21:56:26.207Z", "content": "

Amazing job! Big congrats!

", "rawMarkdown": "Amazing job! Big congrats!", "votes": 1, "replies": [ { "id": 1585010, "postDate": "2021-11-17T01:47:04.510Z", "content": "

Thanks!
\nCongratulations too！

", "rawMarkdown": "Thanks! \nCongratulations too！" } ] }, { "id": 1584476, "postDate": "2021-11-16T15:01:25.040Z", "content": "

Well written and nice solution .. congrats

", "rawMarkdown": "Well written and nice solution .. congrats ", "votes": 1, "replies": [ { "id": 1584611, "postDate": "2021-11-16T16:19:01.213Z", "content": "

Thanks for your comment!

", "rawMarkdown": "Thanks for your comment!" } ] } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T16:21:22Z", "competition_id": 30060, "competition_slug": "chaii-hindi-and-tamil-question-answering", "team_id": 7344551, "solution_url": "/competitions/chaii-hindi-and-tamil-question-answering/writeups/nlp-f3-28th-place-solution-contains-additional-rea" }