It turns out you can do surprisingly well (I mean not great, I got ~0.63 but better than you might initially think) using a character only model: https://www.kaggle.com/hamishdickson/char-level-only-lb0-630
\n\nMy intuition around this is it's learning things like capitalisation and punctuation - which we know is important in how quora decide if something is sincere, but I've noticed a lot of people are inadvertently lowering/stripping in public kernels
", "messages": [ { "id": 462080, "postDate": "2019-01-27T16:14:56.080Z", "content": "It turns out you can do surprisingly well (I mean not great, I got ~0.63 but better than you might initially think) using a character only model: https://www.kaggle.com/hamishdickson/char-level-only-lb0-630
\n\nMy intuition around this is it's learning things like capitalisation and punctuation - which we know is important in how quora decide if something is sincere, but I've noticed a lot of people are inadvertently lowering/stripping in public kernels
", "rawMarkdown": "It turns out you can do surprisingly well (I mean not great, I got ~0.63 but better than you might initially think) using a character only model: https://www.kaggle.com/hamishdickson/char-level-only-lb0-630\n\nMy intuition around this is it's learning things like capitalisation and punctuation - which we know is important in how quora decide if something is sincere, but I've noticed a lot of people are inadvertently lowering/stripping in public kernels", "votes": 3 }, { "id": 462099, "postDate": "2019-01-27T17:05:10.007Z", "content": "A great point! I've not entered this competition - so I say this fully intending that it's taken with more than a grain or two of salt - but I have noticed in reading quite a few kernels and blog posts, that sometimes people head straight to the cleaning of data without considering whether or not they're throwing away useful information.
", "rawMarkdown": "A great point! I've not entered this competition - so I say this fully intending that it's taken with more than a grain or two of salt - but I have noticed in reading quite a few kernels and blog posts, that sometimes people head straight to the cleaning of data without considering whether or not they're throwing away useful information.", "votes": 1 }, { "id": 462380, "postDate": "2019-01-28T08:10:54.160Z", "content": "I made several attempts with char level model and got similar score (F1 around 0.62-0.63). But I thought it's not good results :) \nHowever I haven't tried to use it with other models yet, may be it can add some value in ensemble.
", "rawMarkdown": "I made several attempts with char level model and got similar score (F1 around 0.62-0.63). But I thought it's not good results :) \nHowever I haven't tried to use it with other models yet, may be it can add some value in ensemble." }, { "id": 462674, "postDate": "2019-01-28T17:44:59.330Z", "content": "I tried too, characters (a-z 0-9 white-space) -> Embedding -> 1D CNN, got f1 score 0.63 at most :-(\nNeeded 20 - 30 epochs and > 8000 min to get to this score (local hold-out set), so I did not proceed, but it was fun :-)
", "rawMarkdown": "I tried too, characters (a-z 0-9 white-space) -> Embedding -> 1D CNN, got f1 score 0.63 at most :-(\nNeeded 20 - 30 epochs and > 8000 min to get to this score (local hold-out set), so I did not proceed, but it was fun :-)", "votes": 2, "isDeleted": true, "replies": [ { "id": 462686, "postDate": "2019-01-28T18:15:18.307Z", "content": "Oh interesting you got basically the same results with a CNN as I got with an RNN. The time was a killer for me too :(
", "rawMarkdown": "Oh interesting you got basically the same results with a CNN as I got with an RNN. The time was a killer for me too :(" }, { "id": 463584, "postDate": "2019-01-30T08:37:01.973Z", "content": "If you're interested I managed to get this down to 10 epochs and ~2800s .. which is probably still way too slow to use, but at least I got to the bad result quicker 🙃
", "rawMarkdown": "If you're interested I managed to get this down to 10 epochs and ~2800s .. which is probably still way too slow to use, but at least I got to the bad result quicker 🙃" }, { "id": 464148, "postDate": "2019-01-31T08:58:40.300Z", "content": "Interesting, so you get f1 = 0.63 in 2800 sec, this is an improvement, it shows character level has potential.
", "rawMarkdown": "Interesting, so you get f1 = 0.63 in 2800 sec, this is an improvement, it shows character level has potential.", "isDeleted": true } ] } ], "comments": [ { "id": 462099, "author_name": "ChrisBow", "author_url": "", "post_date": "2019-01-27T17:05:10.007000", "content": "A great point! I've not entered this competition - so I say this fully intending that it's taken with more than a grain or two of salt - but I have noticed in reading quite a few kernels and blog posts, that sometimes people head straight to the cleaning of data without considering whether or not they're throwing away useful information.
", "votes": 1, "replies": [] }, { "id": 462380, "author_name": "antklen", "author_url": "", "post_date": "2019-01-28T08:10:54.160000", "content": "I made several attempts with char level model and got similar score (F1 around 0.62-0.63). But I thought it's not good results :) \nHowever I haven't tried to use it with other models yet, may be it can add some value in ensemble.
", "votes": 0, "replies": [] }, { "id": 462674, "author_name": "", "author_url": "", "post_date": "2019-01-28T17:44:59.330000", "content": "I tried too, characters (a-z 0-9 white-space) -> Embedding -> 1D CNN, got f1 score 0.63 at most :-(\nNeeded 20 - 30 epochs and > 8000 min to get to this score (local hold-out set), so I did not proceed, but it was fun :-)
", "votes": 2, "replies": [ { "id": 462686, "author_name": "Hamish", "author_url": "", "post_date": "2019-01-28T18:15:18.307000", "content": "Oh interesting you got basically the same results with a CNN as I got with an RNN. The time was a killer for me too :(
", "votes": 0, "replies": [] }, { "id": 463584, "author_name": "Hamish", "author_url": "", "post_date": "2019-01-30T08:37:01.973000", "content": "If you're interested I managed to get this down to 10 epochs and ~2800s .. which is probably still way too slow to use, but at least I got to the bad result quicker 🙃
", "votes": 0, "replies": [] }, { "id": 464148, "author_name": "", "author_url": "", "post_date": "2019-01-31T08:58:40.300000", "content": "Interesting, so you get f1 = 0.63 in 2800 sec, this is an improvement, it shows character level has potential.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "462080": "It turns out you can do surprisingly well (I mean not great, I got ~0.63 but better than you might initially think) using a character only model: https://www.kaggle.com/hamishdickson/char-level-only-lb0-630\n\nMy intuition around this is it's learning things like capitalisation and punctuation - which we know is important in how quora decide if something is sincere, but I've noticed a lot of people are inadvertently lowering/stripping in public kernels", "462099": "A great point! I've not entered this competition - so I say this fully intending that it's taken with more than a grain or two of salt - but I have noticed in reading quite a few kernels and blog posts, that sometimes people head straight to the cleaning of data without considering whether or not they're throwing away useful information.", "462380": "I made several attempts with char level model and got similar score (F1 around 0.62-0.63). But I thought it's not good results :) \nHowever I haven't tried to use it with other models yet, may be it can add some value in ensemble.", "462674": "I tried too, characters (a-z 0-9 white-space) -> Embedding -> 1D CNN, got f1 score 0.63 at most :-(\nNeeded 20 - 30 epochs and > 8000 min to get to this score (local hold-out set), so I did not proceed, but it was fun :-)" } }