First of all, I would like to thank the excellent notebooks and discussions from https://www.kaggle.com/abhishek/two-longformers-are-better-than-1 @abhishek
\nhttps://www.kaggle.com/c/feedback-prize-2021/discussion/308992 @hengck23
\nhttps://www.kaggle.com/librauee/infer-fast-ensemble-models @librauee
\nI learned a lot from their work. This is the second kaggle competition we have participated in, and although we are one short of gold, we are already very satisfied.
Merge consecutive predictions with same label, for example we merge [B-Lead, I-Lead, I-Lead], [B-Lead, I-Lead] into one single prediction. We only do this operation when the label is in ['Lead', 'Position', 'Concluding', 'Rebuttal'], since there are not consecutive predictions for these labels in the training data.
Filter \"Lead\" and \"Concluding\". There are only one Lead label and Concluding Label in almost all the trainging data, so we only keep the predictions that has higher score than threshold. Besides, we found that merge two Lead can increase cv further.
concluding_df = sorted(concluding_df, key=lambda x: np.mean(x[4]), reverse=True)\nnew_begin = min(concluding_df[0][3][0], concluding_df[1][3][0])\nnew_end = max(concluding_df[0][3][-1], concluding_df[1][3][-1])\nSince the score is based on the overlap between prediction and ground truth, so we extend the predictions from word_list[begin:end] to word_list[begin - 1: end + 1]. Hoping the extended predictions can better hit ground truth and accross the 50% threshold.
Scaling. The probabilities of each token are multiplied by a factor. The factors are obtained through genetic algorithm search.
There are some other attempts but didn't work well. These attempts are included in the inference notebook.
Next target: gold medal!!!
", "votes": 2, "replies": [] }, { "id": 1724559, "author_name": "GUNER", "author_url": "", "post_date": "2022-03-16T10:23:35.270000", "content": "Congrats on your big win!
\nHow did assign the the tokenwise weights in scaling?
we use optuna to find best weights
", "votes": 1, "replies": [] } ] }, { "id": 1724534, "author_name": "Wonho Song", "author_url": "", "post_date": "2022-03-16T09:57:27.903000", "content": "Thanks for sharing, I enjoy to read ur solution. I think ur cv score from deberta-xlarge is high enough. What is your cv score when u do ensemble?
", "votes": 0, "replies": [ { "id": 1724541, "author_name": "TelmaZzzz", "author_url": "", "post_date": "2022-03-16T10:02:57.957000", "content": "Our ensemble model can get 0.715 at fold0 (10 folds) with simple post-processing, and 0.722 can be achieved using the above post-processing
", "votes": 1, "replies": [] } ] }, { "id": 1724460, "author_name": "newbieclass", "author_url": "", "post_date": "2022-03-16T09:00:48.143000", "content": "good job!
", "votes": 0, "replies": [] } ], "comment_count": 6, "messages": [ { "id": 1724408, "postDate": "2022-03-16T08:28:40.780Z", "content": "First of all, I would like to thank the excellent notebooks and discussions from https://www.kaggle.com/abhishek/two-longformers-are-better-than-1 @abhishek
\nhttps://www.kaggle.com/c/feedback-prize-2021/discussion/308992 @hengck23
\nhttps://www.kaggle.com/librauee/infer-fast-ensemble-models @librauee
\nI learned a lot from their work. This is the second kaggle competition we have participated in, and although we are one short of gold, we are already very satisfied.
Merge consecutive predictions with same label, for example we merge [B-Lead, I-Lead, I-Lead], [B-Lead, I-Lead] into one single prediction. We only do this operation when the label is in ['Lead', 'Position', 'Concluding', 'Rebuttal'], since there are not consecutive predictions for these labels in the training data.
Filter \"Lead\" and \"Concluding\". There are only one Lead label and Concluding Label in almost all the trainging data, so we only keep the predictions that has higher score than threshold. Besides, we found that merge two Lead can increase cv further.
concluding_df = sorted(concluding_df, key=lambda x: np.mean(x[4]), reverse=True)\nnew_begin = min(concluding_df[0][3][0], concluding_df[1][3][0])\nnew_end = max(concluding_df[0][3][-1], concluding_df[1][3][-1])\nSince the score is based on the overlap between prediction and ground truth, so we extend the predictions from word_list[begin:end] to word_list[begin - 1: end + 1]. Hoping the extended predictions can better hit ground truth and accross the 50% threshold.
Scaling. The probabilities of each token are multiplied by a factor. The factors are obtained through genetic algorithm search.
There are some other attempts but didn't work well. These attempts are included in the inference notebook.
Next target: gold medal!!!
", "rawMarkdown": "Next target: gold medal!!!", "votes": 2 }, { "id": 1724559, "postDate": "2022-03-16T10:23:35.270Z", "content": "Congrats on your big win!
\nHow did assign the the tokenwise weights in scaling?
we use optuna to find best weights
", "rawMarkdown": "we use optuna to find best weights", "votes": 1 } ] }, { "id": 1724534, "postDate": "2022-03-16T09:57:27.903Z", "content": "Thanks for sharing, I enjoy to read ur solution. I think ur cv score from deberta-xlarge is high enough. What is your cv score when u do ensemble?
", "rawMarkdown": "Thanks for sharing, I enjoy to read ur solution. I think ur cv score from deberta-xlarge is high enough. What is your cv score when u do ensemble?", "replies": [ { "id": 1724541, "postDate": "2022-03-16T10:02:57.957Z", "content": "Our ensemble model can get 0.715 at fold0 (10 folds) with simple post-processing, and 0.722 can be achieved using the above post-processing
", "rawMarkdown": "Our ensemble model can get 0.715 at fold0 (10 folds) with simple post-processing, and 0.722 can be achieved using the above post-processing", "votes": 1 } ] }, { "id": 1724460, "postDate": "2022-03-16T09:00:48.143Z", "content": "good job!
", "rawMarkdown": "good job! " } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T12:35:51Z", "competition_id": 31779, "competition_slug": "feedback-prize-2021", "team_id": 8025185, "solution_url": "/competitions/feedback-prize-2021/writeups/15th-solution" }