{ "id": 238428, "title": "What Approach Are you using??", "url": "/competitions/coleridgeinitiative-show-us-the-data/discussion/238428", "author_name": "", "post_date": "2021-05-12T06:59:03.687647400Z", "votes": 30, "comment_count": 20, "views": 0, "content": "

Hi all ,
\nI was just thinking to start this competition but I am new to entity extraction and I just see one kernel to learn and to start this competition. I was wondering if the community can share bits and pieces on what they are using at this time , it will serve as a motivator and a set piece for anyone to start with this competition

\n

Is NER the right approach for this competition ?

", "messages": [ { "id": "1303637", "postDate": "05/12/2021 06:59:03", "content": "

Hi all ,
\nI was just thinking to start this competition but I am new to entity extraction and I just see one kernel to learn and to start this competition. I was wondering if the community can share bits and pieces on what they are using at this time , it will serve as a motivator and a set piece for anyone to start with this competition

\n

Is NER the right approach for this competition ?

", "rawMarkdown": "Hi all ,\nI was just thinking to start this competition but I am new to entity extraction and I just see one kernel to learn and to start this competition. I was wondering if the community can share bits and pieces on what they are using at this time , it will serve as a motivator and a set piece for anyone to start with this competition\n\nIs NER the right approach for this competition ?", "votes": null }, { "id": "1303721", "postDate": "05/12/2021 07:58:27", "content": "

May be in private LB.

", "rawMarkdown": "May be in private LB.", "votes": null }, { "id": "1303898", "postDate": "05/12/2021 10:23:38", "content": "

NER: https://www.kaggle.com/tungmphung/pytorch-bert-for-named-entity-recognition
\nQA: https://www.kaggle.com/theoviel/bert-for-question-answering-baseline-training
\nMLM: https://www.kaggle.com/tungmphung/coleridge-bert-masked-dataset-modeling

\n

Hope this helps!

", "rawMarkdown": "NER: https://www.kaggle.com/tungmphung/pytorch-bert-for-named-entity-recognition\nQA: https://www.kaggle.com/theoviel/bert-for-question-answering-baseline-training\nMLM: https://www.kaggle.com/tungmphung/coleridge-bert-masked-dataset-modeling\n\nHope this helps!", "votes": null }, { "id": "1304202", "postDate": "05/12/2021 13:35:50", "content": "

you just bagged one silver medal and now on to this :)

", "rawMarkdown": "you just bagged one silver medal and now on to this :)", "votes": null }, { "id": "1304249", "postDate": "05/12/2021 14:08:26", "content": "

Thanks man

", "rawMarkdown": "Thanks man", "votes": null }, { "id": "1304251", "postDate": "05/12/2021 14:09:07", "content": "

I am just addicted

", "rawMarkdown": "I am just addicted", "votes": null }, { "id": "1304760", "postDate": "05/12/2021 21:03:30", "content": "

He is everywhere ;)

", "rawMarkdown": "He is everywhere ;)", "votes": null }, { "id": "1304983", "postDate": "05/13/2021 03:42:01", "content": "

Hahaha, I thought I misclicked in the Shopee discussion.

", "rawMarkdown": "Hahaha, I thought I misclicked in the `Shopee` discussion.", "votes": null }, { "id": "1305414", "postDate": "05/13/2021 08:55:43", "content": "

i have tried ner with crf , but i do not get a better score than string model . So , I am confused that should i try some other ner model or Data preprocessing ? or feature extracted ? By the way , I am new to kaggle competition and nlp . Is there some body giving some advice about it ? i will be very appreciate for your help

", "rawMarkdown": "i have tried ner with crf , but i do not get a better score than string model . So , I am confused that should i try some other ner model or Data preprocessing ? or feature extracted ? By the way , I am new to kaggle competition and nlp . Is there some body giving some advice about it ? i will be very appreciate for your help", "votes": null }, { "id": "1305553", "postDate": "05/13/2021 11:09:52", "content": "

Thank You, I was struggling in having a good starting point. Thank You again

", "rawMarkdown": "Thank You, I was struggling in having a good starting point. Thank You again", "votes": null }, { "id": "1306001", "postDate": "05/13/2021 15:11:02", "content": "

I found a very nice paper where the authors are doing something very similar to what this competition is asking for. It seemed a very good starting point to me.

\n

To summarize their approach :

\n
    \n
  1. Create BIO labelling for your dataset, for eg- label this sentence [\"We\", \"are\", \"using\", \"COVID\", \"19\", \"Dataset\", \"in\", \"our\", \"paper\"] as [\"O\", \"O\", \"O\", \"B\", \"I\", \"I\", \"O\", \"O\", \"O\"]. Here, B is first word of a dataset, I is word that is a part of a dataset and O is any other word in our dataset.
    2. \n
  2. Experiment with various text to vector conversion algorithms (word embedding algorithms), for eg - Word2Vec, Bag of Words etc.
    3. \n
  3. Train Bi-directional LSTM to predict BIO labels for a given sentence.
    4. \n
  4. It may output sequences (e.g., “OIO”) that are invalid under the BIO labeling scheme. To penalize such invalid label sequences, add a Conditional Random Field (CRF) layer on top of the Bi-LSTM network.
    5. \n
\n

The paper also talks about extracting lexical features out of text, but it may not help much in this competition. And one can also try BERT instead of Bi-directional LSTM.

\n

EDIT : I have created a notebook where I try to label Coleridge dataset using BIO scheme link

", "rawMarkdown": "I found a very nice [paper](https://link.springer.com/content/pdf/10.1007%2F978-3-030-50417-5_23.pdf) where the authors are doing something very similar to what this competition is asking for. It seemed a very good starting point to me.\n\nTo summarize their approach : \n\n1. **Create BIO labelling for your dataset**, for eg- label this sentence [\"We\", \"are\", \"using\", \"COVID\", \"19\", \"Dataset\", \"in\", \"our\", \"paper\"] as [\"O\", \"O\", \"O\", \"B\", \"I\", \"I\", \"O\", \"O\", \"O\"]. Here, B is first word of a dataset, I is word that is a part of a dataset and O is any other word in our dataset.\n2. **Experiment with various text to vector conversion algorithms** (word embedding algorithms), for eg - Word2Vec, Bag of Words etc.\n3. **Train Bi-directional LSTM** to predict BIO labels for a given sentence.\n4. It may output sequences (e.g., “OIO”) that are invalid under the BIO labeling scheme. To penalize such invalid label sequences, **add a Conditional Random Field (CRF) layer** on top of the Bi-LSTM network.\n\nThe paper also talks about extracting lexical features out of text, but it may not help much in this competition. And one can also try BERT instead of Bi-directional LSTM.\n\n**EDIT : I have created a notebook where I try to label Coleridge dataset using BIO scheme [link](https://www.kaggle.com/devashishprasad/bio-labeled-dataset)**", "votes": null }, { "id": "1306478", "postDate": "05/13/2021 19:33:02", "content": "

Hi Thanks for sharing the results I am almost finished with the baseline , will start sharing results very soon

", "rawMarkdown": "Hi Thanks for sharing the results I am almost finished with the baseline , will start sharing results very soon", "votes": null }, { "id": "1307167", "postDate": "05/14/2021 09:42:50", "content": "

I knew it was going to be an interesting competition when I saw him show up.

", "rawMarkdown": "I knew it was going to be an interesting competition when I saw him show up.", "votes": null }, { "id": "1307241", "postDate": "05/14/2021 10:34:33", "content": "

@tanulsingh077 You're already 25th so I guess you've found a working approach. I'm guessing NER but I can't tell for sure.

\n

Also if someone managed to get any good results with starting from my QA approach (here) let me know :)

", "rawMarkdown": "tanulsingh077 You're already 25th so I guess you've found a working approach. I'm guessing NER but I can't tell for sure. \n\nAlso if someone managed to get any good results with starting from my QA approach ([here](https://www.kaggle.com/theoviel/bert-for-question-answering-baseline-training)) let me know :)", "votes": null }, { "id": "1307318", "postDate": "05/14/2021 11:25:48", "content": "

HI @theoviel glad that you commented , we are still not using NER we were just verifying few things , the public lb in this competition is purely misleading and you can get to 0.55 with just string and pattern matching but it won't mean anything on private , you can read more about it here
\nhttps://www.kaggle.com/c/coleridgeinitiative-show-us-the-data/discussion/238005

\n

But we have two pipelines ready will be sharing all my results very soon

\n

you are one of my idols , I might be shameless here but I would be on the cloud nine if you would join our team ;-)

", "rawMarkdown": "HI @theoviel glad that you commented , we are still not using NER we were just verifying few things , the public lb in this competition is purely misleading and you can get to 0.55 with just string and pattern matching but it won't mean anything on private , you can read more about it here \nhttps://www.kaggle.com/c/coleridgeinitiative-show-us-the-data/discussion/238005\n\nBut we have two pipelines ready will be sharing all my results very soon\n\nyou are one of my idols , I might be shameless here but I would be on the cloud nine if you would join our team ;-)", "votes": null }, { "id": "1307319", "postDate": "05/14/2021 11:25:59", "content": "

Thanks for sharing

", "rawMarkdown": "Thanks for sharing", "votes": null }, { "id": "1307333", "postDate": "05/14/2021 11:34:46", "content": "

Thanks for the kind words @tanulsingh077 !

\n

I have not decided whether I actually want to spend time on this competition yet, I need a break after the HubMAP one. So at the moment I'm not ready to join a team, sorry !

", "rawMarkdown": "Thanks for the kind words @tanulsingh077 !\n\nI have not decided whether I actually want to spend time on this competition yet, I need a break after the HubMAP one. So at the moment I'm not ready to join a team, sorry !", "votes": null }, { "id": "1307452", "postDate": "05/14/2021 12:55:49", "content": "

thanks for sharing. I was in similar lines but here texts are too long to balance BIO tokens. in fact i used two tokens dataset and non dataset but couldn't beat the baseline

", "rawMarkdown": "thanks for sharing. I was in similar lines but here texts are too long to balance BIO tokens. in fact i used two tokens dataset and non dataset but couldn't beat the baseline", "votes": null }, { "id": "1334874", "postDate": "06/03/2021 20:42:39", "content": "

Thank you @theoviel
\nbut by trying q/a i am not getting good score

", "rawMarkdown": "Thank you @theoviel \nbut by trying q/a i am not getting good score", "votes": null }, { "id": "1339860", "postDate": "06/07/2021 13:47:15", "content": "

Would you like to team up?

", "rawMarkdown": "Would you like to team up?", "votes": null }, { "id": "1340030", "postDate": "06/07/2021 15:57:09", "content": "

you could have tagged. glad to join. send me the request if it is for me.

", "rawMarkdown": "you could have tagged. glad to join. send me the request if it is for me.", "votes": null } ], "comments": [ { "id": 1303721, "author_name": "vanle73", "author_url": "", "post_date": "05/12/2021 07:58:27", "content": "

May be in private LB.

", "votes": null, "replies": [] }, { "id": 1303898, "author_name": "hannes82", "author_url": "", "post_date": "05/12/2021 10:23:38", "content": "

NER: https://www.kaggle.com/tungmphung/pytorch-bert-for-named-entity-recognition
\nQA: https://www.kaggle.com/theoviel/bert-for-question-answering-baseline-training
\nMLM: https://www.kaggle.com/tungmphung/coleridge-bert-masked-dataset-modeling

\n

Hope this helps!

", "votes": null, "replies": [ { "id": 1304249, "author_name": "tanulsingh077", "author_url": "", "post_date": "05/12/2021 14:08:26", "content": "

Thanks man

", "votes": null, "replies": [] }, { "id": 1305553, "author_name": "osamir", "author_url": "", "post_date": "05/13/2021 11:09:52", "content": "

Thank You, I was struggling in having a good starting point. Thank You again

", "votes": null, "replies": [] } ] }, { "id": 1304202, "author_name": "udaygurugubelli", "author_url": "", "post_date": "05/12/2021 13:35:50", "content": "

you just bagged one silver medal and now on to this :)

", "votes": null, "replies": [ { "id": 1304251, "author_name": "tanulsingh077", "author_url": "", "post_date": "05/12/2021 14:09:07", "content": "

I am just addicted

", "votes": null, "replies": [] }, { "id": 1304760, "author_name": "nischaydnk", "author_url": "", "post_date": "05/12/2021 21:03:30", "content": "

He is everywhere ;)

", "votes": null, "replies": [] }, { "id": 1304983, "author_name": "chienhsianghung", "author_url": "", "post_date": "05/13/2021 03:42:01", "content": "

Hahaha, I thought I misclicked in the Shopee discussion.

", "votes": null, "replies": [] }, { "id": 1307167, "author_name": "xj609210970", "author_url": "", "post_date": "05/14/2021 09:42:50", "content": "

I knew it was going to be an interesting competition when I saw him show up.

", "votes": null, "replies": [] }, { "id": 1339860, "author_name": "syavaprd", "author_url": "", "post_date": "06/07/2021 13:47:15", "content": "

Would you like to team up?

", "votes": null, "replies": [] }, { "id": 1340030, "author_name": "udaygurugubelli", "author_url": "", "post_date": "06/07/2021 15:57:09", "content": "

you could have tagged. glad to join. send me the request if it is for me.

", "votes": null, "replies": [] } ] }, { "id": 1305414, "author_name": "coffeebear98", "author_url": "", "post_date": "05/13/2021 08:55:43", "content": "

i have tried ner with crf , but i do not get a better score than string model . So , I am confused that should i try some other ner model or Data preprocessing ? or feature extracted ? By the way , I am new to kaggle competition and nlp . Is there some body giving some advice about it ? i will be very appreciate for your help

", "votes": null, "replies": [ { "id": 1306478, "author_name": "tanulsingh077", "author_url": "", "post_date": "05/13/2021 19:33:02", "content": "

Hi Thanks for sharing the results I am almost finished with the baseline , will start sharing results very soon

", "votes": null, "replies": [] } ] }, { "id": 1306001, "author_name": "devashishprasad", "author_url": "", "post_date": "05/13/2021 15:11:02", "content": "

I found a very nice paper where the authors are doing something very similar to what this competition is asking for. It seemed a very good starting point to me.

\n

To summarize their approach :

\n
    \n
  1. Create BIO labelling for your dataset, for eg- label this sentence [\"We\", \"are\", \"using\", \"COVID\", \"19\", \"Dataset\", \"in\", \"our\", \"paper\"] as [\"O\", \"O\", \"O\", \"B\", \"I\", \"I\", \"O\", \"O\", \"O\"]. Here, B is first word of a dataset, I is word that is a part of a dataset and O is any other word in our dataset.
    2. \n
  2. Experiment with various text to vector conversion algorithms (word embedding algorithms), for eg - Word2Vec, Bag of Words etc.
    3. \n
  3. Train Bi-directional LSTM to predict BIO labels for a given sentence.
    4. \n
  4. It may output sequences (e.g., “OIO”) that are invalid under the BIO labeling scheme. To penalize such invalid label sequences, add a Conditional Random Field (CRF) layer on top of the Bi-LSTM network.
    5. \n
\n

The paper also talks about extracting lexical features out of text, but it may not help much in this competition. And one can also try BERT instead of Bi-directional LSTM.

\n

EDIT : I have created a notebook where I try to label Coleridge dataset using BIO scheme link

", "votes": null, "replies": [ { "id": 1307319, "author_name": "tanulsingh077", "author_url": "", "post_date": "05/14/2021 11:25:59", "content": "

Thanks for sharing

", "votes": null, "replies": [] }, { "id": 1307452, "author_name": "udaygurugubelli", "author_url": "", "post_date": "05/14/2021 12:55:49", "content": "

thanks for sharing. I was in similar lines but here texts are too long to balance BIO tokens. in fact i used two tokens dataset and non dataset but couldn't beat the baseline

", "votes": null, "replies": [] } ] }, { "id": 1307241, "author_name": "theoviel", "author_url": "", "post_date": "05/14/2021 10:34:33", "content": "

@tanulsingh077 You're already 25th so I guess you've found a working approach. I'm guessing NER but I can't tell for sure.

\n

Also if someone managed to get any good results with starting from my QA approach (here) let me know :)

", "votes": null, "replies": [ { "id": 1307318, "author_name": "tanulsingh077", "author_url": "", "post_date": "05/14/2021 11:25:48", "content": "

HI @theoviel glad that you commented , we are still not using NER we were just verifying few things , the public lb in this competition is purely misleading and you can get to 0.55 with just string and pattern matching but it won't mean anything on private , you can read more about it here
\nhttps://www.kaggle.com/c/coleridgeinitiative-show-us-the-data/discussion/238005

\n

But we have two pipelines ready will be sharing all my results very soon

\n

you are one of my idols , I might be shameless here but I would be on the cloud nine if you would join our team ;-)

", "votes": null, "replies": [] }, { "id": 1307333, "author_name": "theoviel", "author_url": "", "post_date": "05/14/2021 11:34:46", "content": "

Thanks for the kind words @tanulsingh077 !

\n

I have not decided whether I actually want to spend time on this competition yet, I need a break after the HubMAP one. So at the moment I'm not ready to join a team, sorry !

", "votes": null, "replies": [] } ] }, { "id": 1334874, "author_name": "ajaypawar123", "author_url": "", "post_date": "06/03/2021 20:42:39", "content": "

Thank you @theoviel
\nbut by trying q/a i am not getting good score

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1303637": "Hi all ,\nI was just thinking to start this competition but I am new to entity extraction and I just see one kernel to learn and to start this competition. I was wondering if the community can share bits and pieces on what they are using at this time , it will serve as a motivator and a set piece for anyone to start with this competition\n\nIs NER the right approach for this competition ?", "1303721": "May be in private LB.", "1303898": "NER: https://www.kaggle.com/tungmphung/pytorch-bert-for-named-entity-recognition\nQA: https://www.kaggle.com/theoviel/bert-for-question-answering-baseline-training\nMLM: https://www.kaggle.com/tungmphung/coleridge-bert-masked-dataset-modeling\n\nHope this helps!", "1304202": "you just bagged one silver medal and now on to this :)", "1304249": "Thanks man", "1304251": "I am just addicted", "1304760": "He is everywhere ;)", "1304983": "Hahaha, I thought I misclicked in the `Shopee` discussion.", "1305414": "i have tried ner with crf , but i do not get a better score than string model . So , I am confused that should i try some other ner model or Data preprocessing ? or feature extracted ? By the way , I am new to kaggle competition and nlp . Is there some body giving some advice about it ? i will be very appreciate for your help", "1305553": "Thank You, I was struggling in having a good starting point. Thank You again", "1306001": "I found a very nice [paper](https://link.springer.com/content/pdf/10.1007%2F978-3-030-50417-5_23.pdf) where the authors are doing something very similar to what this competition is asking for. It seemed a very good starting point to me.\n\nTo summarize their approach : \n\n1. **Create BIO labelling for your dataset**, for eg- label this sentence [\"We\", \"are\", \"using\", \"COVID\", \"19\", \"Dataset\", \"in\", \"our\", \"paper\"] as [\"O\", \"O\", \"O\", \"B\", \"I\", \"I\", \"O\", \"O\", \"O\"]. Here, B is first word of a dataset, I is word that is a part of a dataset and O is any other word in our dataset.\n2. **Experiment with various text to vector conversion algorithms** (word embedding algorithms), for eg - Word2Vec, Bag of Words etc.\n3. **Train Bi-directional LSTM** to predict BIO labels for a given sentence.\n4. It may output sequences (e.g., “OIO”) that are invalid under the BIO labeling scheme. To penalize such invalid label sequences, **add a Conditional Random Field (CRF) layer** on top of the Bi-LSTM network.\n\nThe paper also talks about extracting lexical features out of text, but it may not help much in this competition. And one can also try BERT instead of Bi-directional LSTM.\n\n**EDIT : I have created a notebook where I try to label Coleridge dataset using BIO scheme [link](https://www.kaggle.com/devashishprasad/bio-labeled-dataset)**", "1306478": "Hi Thanks for sharing the results I am almost finished with the baseline , will start sharing results very soon", "1307167": "I knew it was going to be an interesting competition when I saw him show up.", "1307241": "tanulsingh077 You're already 25th so I guess you've found a working approach. I'm guessing NER but I can't tell for sure. \n\nAlso if someone managed to get any good results with starting from my QA approach ([here](https://www.kaggle.com/theoviel/bert-for-question-answering-baseline-training)) let me know :)", "1307318": "HI @theoviel glad that you commented , we are still not using NER we were just verifying few things , the public lb in this competition is purely misleading and you can get to 0.55 with just string and pattern matching but it won't mean anything on private , you can read more about it here \nhttps://www.kaggle.com/c/coleridgeinitiative-show-us-the-data/discussion/238005\n\nBut we have two pipelines ready will be sharing all my results very soon\n\nyou are one of my idols , I might be shameless here but I would be on the cloud nine if you would join our team ;-)", "1307319": "Thanks for sharing", "1307333": "Thanks for the kind words @tanulsingh077 !\n\nI have not decided whether I actually want to spend time on this competition yet, I need a break after the HubMAP one. So at the moment I'm not ready to join a team, sorry !", "1307452": "thanks for sharing. I was in similar lines but here texts are too long to balance BIO tokens. in fact i used two tokens dataset and non dataset but couldn't beat the baseline", "1334874": "Thank you @theoviel \nbut by trying q/a i am not getting good score", "1339860": "Would you like to team up?", "1340030": "you could have tagged. glad to join. send me the request if it is for me." }, "source": "meta" }