First of all I want to thank my teammates here.I will briefly introduce our solution here.
\n\n1.Model structure。we design different models structure. We mainly refer to the solution of ccf internet sentiment analysis，concat different cls embedding . here it is the link BDCI2019-SENTIMENT-CLASSIFICATION
\n 2.We found 30 labels through analysis, one is the question-related evaluation, and the other is the answer-related evaluation. In order to make the model learn better, we have designed the q model to remove the problem-related label and the a model to process the answer Related labels。 it is better than qa models.
\n3.different model test. roberta base >roberta large >xlnet base >bert base > t5 base.
1.we want that our model learns features that are not only considered in text, so we add host and \ncategory embeeding features annd other Statistical Features。 it improve both cv and lb about 0.005.
\n\n1.We also did text cleaning to remove stop words and some symbols， it improve about 0.002
\n\n1.Our Best Private model scored 0.42787 ,but we dont't select it. it is stacking by roberta large and roberta base and xlnet base.
\nblend.loc[:,targets] = roberta_large_oof_test.loc[:,targets].values*0.4+0.3*roberta_base_oof_test.loc[:,targets].values+\\\n xlnet_base_oof_test.loc[:,targets].values*0.3
stacking improve both cv and lb about 0.02 . it help much.
", "comments": [ { "id": 749635, "author_name": "", "author_url": "", "post_date": "2020-02-18T20:27:04.020000", "content": "tq u for sharing ur hardwork
", "votes": 15, "replies": [] }, { "id": 742934, "author_name": "CPMP", "author_url": "", "post_date": "2020-02-11T16:42:33.643000", "content": "Great job, congrats. Which statistical features did you use? We didn't found useful ones besides category embeddings.
", "votes": 1, "replies": [ { "id": 743129, "author_name": "BestPredict", "author_url": "", "post_date": "2020-02-11T20:13:57.773000", "content": "some statistical features like word num, answer word nums vs question word nums, but it help. little.
", "votes": 0, "replies": [] }, { "id": 743139, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-02-11T20:29:01.587000", "content": "I posted some EDA of meta features here. Word count in the answer was a helpful one (especially for question_type_spelling and answer_level_of_information).
nice work and you have shared a lot of valuable things in a lot of games and learned a lot from you @cdeotte
", "votes": 1, "replies": [] }, { "id": 881623, "author_name": "ich bin MICHA", "author_url": "", "post_date": "2020-06-11T08:03:40.437000", "content": "Iuiiii
", "votes": 0, "replies": [] } ] }, { "id": 862005, "author_name": "哈尔的移动城堡", "author_url": "", "post_date": "2020-05-26T11:11:23.933000", "content": "今年考研究生，哈工深真的那么好吗？
", "votes": -2, "replies": [] }, { "id": 1060437, "author_name": "Vijaybhandare123", "author_url": "", "post_date": "2020-10-26T07:35:01.830000", "content": "Congratulations …..
\nWhat was the loss function used ?
Congrats @bestpredict and team. Thanks for sharing your solution.
", "votes": 0, "replies": [] }, { "id": 742770, "author_name": "VIKRAM KUSHWAH", "author_url": "", "post_date": "2020-02-11T14:33:02.530000", "content": "hey congratulations!!!
", "votes": 0, "replies": [] }, { "id": 742435, "author_name": "RB", "author_url": "", "post_date": "2020-02-11T09:45:03.957000", "content": "Congratulations and thank you for write up. Seems like your stacking doesn't include Bert. \nHow did you come up with stacking weights ? 0.4 for Roberta large, 0.3 for Roberta Base and 0.3 for Xlnet.
", "votes": 0, "replies": [ { "id": 743127, "author_name": "BestPredict", "author_url": "", "post_date": "2020-02-11T20:12:01.403000", "content": "based on oof train
", "votes": 0, "replies": [] } ] }, { "id": 742416, "author_name": "liuze", "author_url": "", "post_date": "2020-02-11T09:35:18.653000", "content": "Congratulations, what we use 0,1 reseting way means? thank you
Calculate the threshold from the training set， if value< thresold 1 , then reset value to 0, and if value >thresold2 ,then reset value to 1
", "votes": 1, "replies": [] }, { "id": 742888, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-02-11T16:18:20.200000", "content": "Awesome. Congrats on a great job. When you say \"from training set\", do you mean analyze the OOF from 5-Fold?
", "votes": 0, "replies": [] }, { "id": 743125, "author_name": "BestPredict", "author_url": "", "post_date": "2020-02-11T20:10:40.557000", "content": "", "votes": 0, "replies": [] }, { "id": 743131, "author_name": "BestPredict", "author_url": "", "post_date": "2020-02-11T20:15:39.303000", "content": "yes ,based oof train
", "votes": 1, "replies": [] } ] }, { "id": 742308, "author_name": "Ailurophile", "author_url": "", "post_date": "2020-02-11T08:18:02.770000", "content": "Congratulations!\nThanks for Sharing your Approach & Insights!! @bestpredict
", "votes": 0, "replies": [] }, { "id": 742193, "author_name": "xia", "author_url": "", "post_date": "2020-02-11T06:21:11.377000", "content": "Nice
", "votes": 0, "replies": [ { "id": 742233, "author_name": "BestPredict", "author_url": "", "post_date": "2020-02-11T07:07:08.120000", "content": "thank u ,brother, thank teammates
", "votes": 0, "replies": [] } ] }, { "id": 742129, "author_name": "Rainbow Cat", "author_url": "", "post_date": "2020-02-11T04:56:42.857000", "content": "Excellect jobs. Thank you so much for learning so much.
", "votes": 0, "replies": [ { "id": 742130, "author_name": "BestPredict", "author_url": "", "post_date": "2020-02-11T04:57:10.987000", "content": "thank u, my brother
", "votes": 0, "replies": [] } ] }, { "id": 744570, "author_name": "corochann", "author_url": "", "post_date": "2020-02-12T23:52:28.980000", "content": "Congrats, thank you for sharing.
", "votes": 0, "replies": [] } ], "comment_count": 23, "messages": [ { "id": 742115, "postDate": "2020-02-11T04:46:39.240Z", "content": "First of all I want to thank my teammates here.I will briefly introduce our solution here.
\n\n1.Model structure。we design different models structure. We mainly refer to the solution of ccf internet sentiment analysis，concat different cls embedding . here it is the link BDCI2019-SENTIMENT-CLASSIFICATION
\n 2.We found 30 labels through analysis, one is the question-related evaluation, and the other is the answer-related evaluation. In order to make the model learn better, we have designed the q model to remove the problem-related label and the a model to process the answer Related labels。 it is better than qa models.
\n3.different model test. roberta base >roberta large >xlnet base >bert base > t5 base.
1.we want that our model learns features that are not only considered in text, so we add host and \ncategory embeeding features annd other Statistical Features。 it improve both cv and lb about 0.005.
\n\n1.We also did text cleaning to remove stop words and some symbols， it improve about 0.002
\n\n1.Our Best Private model scored 0.42787 ,but we dont't select it. it is stacking by roberta large and roberta base and xlnet base.
\nblend.loc[:,targets] = roberta_large_oof_test.loc[:,targets].values*0.4+0.3*roberta_base_oof_test.loc[:,targets].values+\\\n xlnet_base_oof_test.loc[:,targets].values*0.3
stacking improve both cv and lb about 0.02 . it help much.
", "rawMarkdown": "\nFirst of all I want to thank my teammates here.I will briefly introduce our solution here.\n\n## models\n1.Model structure。we design different models structure. We mainly refer to the solution of ccf internet sentiment analysis，concat different cls embedding . here it is the link [BDCI2019-SENTIMENT-CLASSIFICATION](https://github.com/cxy229/BDCI2019-SENTIMENT-CLASSIFICATION) \n 2.We found 30 labels through analysis, one is the question-related evaluation, and the other is the answer-related evaluation. In order to make the model learn better, we have designed the q model to remove the problem-related label and the a model to process the answer Related labels。 it is better than qa models. \n3.different model test. roberta base >roberta large >xlnet base >bert base > t5 base. \n \n## Post-processing \n1. Analysis and evaluation methods and competition data，we use 0,1 reseting way. it improve lb 0.05 or more. \n\n\n## Features\n1.we want that our model learns features that are not only considered in text, so we add host and \ncategory embeeding features annd other Statistical Features。 it improve both cv and lb about 0.005. \n\n##Text clean \n1.We also did text cleaning to remove stop words and some symbols， it improve about 0.002 \n\n\n\n## Stacking \n\n1.Our Best Private model scored 0.42787 ,but we dont't select it. it is stacking by roberta large and roberta base and xlnet base. \n`blend.loc[:,targets] = roberta_large_oof_test.loc[:,targets].values*0.4+0.3*roberta_base_oof_test.loc[:,targets].values+\\\n xlnet_base_oof_test.loc[:,targets].values*0.3` \n\nstacking improve both cv and lb about 0.02 . it help much.\n", "votes": 45 }, { "id": 749635, "postDate": "2020-02-18T20:27:04.020Z", "content": "tq u for sharing ur hardwork
", "rawMarkdown": "tq u for sharing ur hardwork", "votes": 15 }, { "id": 742934, "postDate": "2020-02-11T16:42:33.643Z", "content": "Great job, congrats. Which statistical features did you use? We didn't found useful ones besides category embeddings.
", "rawMarkdown": "Great job, congrats. Which statistical features did you use? We didn't found useful ones besides category embeddings.", "votes": 1, "replies": [ { "id": 743129, "postDate": "2020-02-11T20:13:57.773Z", "content": "some statistical features like word num, answer word nums vs question word nums, but it help. little.
", "rawMarkdown": "some statistical features like word num, answer word nums vs question word nums, but it help. little." }, { "id": 743139, "postDate": "2020-02-11T20:29:01.587Z", "content": "I posted some EDA of meta features here. Word count in the answer was a helpful one (especially for question_type_spelling and answer_level_of_information).
nice work and you have shared a lot of valuable things in a lot of games and learned a lot from you @cdeotte
", "rawMarkdown": "nice work and you have shared a lot of valuable things in a lot of games and learned a lot from you @cdeotte ", "votes": 1 }, { "id": 881623, "postDate": "2020-06-11T08:03:40.437Z", "content": "Iuiiii
", "rawMarkdown": "Iuiiii" } ] }, { "id": 862005, "postDate": "2020-05-26T11:11:23.933Z", "content": "今年考研究生，哈工深真的那么好吗？
", "rawMarkdown": "今年考研究生，哈工深真的那么好吗？", "votes": -2 }, { "id": 1060437, "postDate": "2020-10-26T07:35:01.830Z", "content": "Congratulations …..
\nWhat was the loss function used ?
Congrats @bestpredict and team. Thanks for sharing your solution.
", "rawMarkdown": "Congrats @bestpredict and team. Thanks for sharing your solution." }, { "id": 742770, "postDate": "2020-02-11T14:33:02.530Z", "content": "hey congratulations!!!
", "rawMarkdown": "hey congratulations!!!" }, { "id": 742435, "postDate": "2020-02-11T09:45:03.957Z", "content": "Congratulations and thank you for write up. Seems like your stacking doesn't include Bert. \nHow did you come up with stacking weights ? 0.4 for Roberta large, 0.3 for Roberta Base and 0.3 for Xlnet.
", "rawMarkdown": "Congratulations and thank you for write up. Seems like your stacking doesn't include Bert. \nHow did you come up with stacking weights ? 0.4 for Roberta large, 0.3 for Roberta Base and 0.3 for Xlnet. ", "replies": [ { "id": 743127, "postDate": "2020-02-11T20:12:01.403Z", "content": "based on oof train
", "rawMarkdown": "based on oof train" } ] }, { "id": 742416, "postDate": "2020-02-11T09:35:18.653Z", "content": "Congratulations, what we use 0,1 reseting way means? thank you
Calculate the threshold from the training set， if value< thresold 1 , then reset value to 0, and if value >thresold2 ,then reset value to 1
", "rawMarkdown": "\nCalculate the threshold from the training set， if value< thresold 1 , then reset value to 0, and if value >thresold2 ,then reset value to 1", "votes": 1 }, { "id": 742888, "postDate": "2020-02-11T16:18:20.200Z", "content": "Awesome. Congrats on a great job. When you say \"from training set\", do you mean analyze the OOF from 5-Fold?
", "rawMarkdown": "Awesome. Congrats on a great job. When you say \"from training set\", do you mean analyze the OOF from 5-Fold?" }, { "id": 743125, "postDate": "2020-02-11T20:10:40.557Z", "rawMarkdown": "" }, { "id": 743131, "postDate": "2020-02-11T20:15:39.303Z", "content": "yes ,based oof train
", "rawMarkdown": "yes ,based oof train", "votes": 1 } ] }, { "id": 742308, "postDate": "2020-02-11T08:18:02.770Z", "content": "Congratulations!\nThanks for Sharing your Approach & Insights!! @bestpredict
", "rawMarkdown": "Congratulations!\nThanks for Sharing your Approach & Insights!! @bestpredict " }, { "id": 742193, "postDate": "2020-02-11T06:21:11.377Z", "content": "Nice
", "rawMarkdown": "Nice", "replies": [ { "id": 742233, "postDate": "2020-02-11T07:07:08.120Z", "content": "thank u ,brother, thank teammates
", "rawMarkdown": "thank u ,brother, thank teammates" } ] }, { "id": 742129, "postDate": "2020-02-11T04:56:42.857Z", "content": "Excellect jobs. Thank you so much for learning so much.
", "rawMarkdown": "Excellect jobs. Thank you so much for learning so much.", "replies": [ { "id": 742130, "postDate": "2020-02-11T04:57:10.987Z", "content": "thank u, my brother
", "rawMarkdown": "thank u, my brother" } ] }, { "id": 744570, "postDate": "2020-02-12T23:52:28.980Z", "content": "Congrats, thank you for sharing.
", "rawMarkdown": "Congrats, thank you for sharing." } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T19:07:40Z", "competition_id": 7968, "competition_slug": "google-quest-challenge", "team_id": 4026431, "solution_url": "/competitions/google-quest-challenge/writeups/cgii-5th-solution" }