If the task of this competition is regarded as a NER or QA problem, we need to extract the required answers from texts with an average length of 50,000, which seems to be difficult to achieve with BERT or LSTM. (dozens of sliding windows?）
", "messages": [ { "id": "1254213", "postDate": "03/27/2021 12:42:57", "content": "If the task of this competition is regarded as a NER or QA problem, we need to extract the required answers from texts with an average length of 50,000, which seems to be difficult to achieve with BERT or LSTM. (dozens of sliding windows?）
", "rawMarkdown": "If the task of this competition is regarded as a NER or QA problem, we need to extract the required answers from texts with an average length of 50,000, which seems to be difficult to achieve with BERT or LSTM. (dozens of sliding windows?）", "votes": null }, { "id": "1255955", "postDate": "03/29/2021 11:58:57", "content": "Considering similar competitions and their winning solutions, I think a good way of tackling this is the following. Training two models which one of those predicts if any dataset has been mentioned in the text and feeds it to the second model which extracts that dataset from the text.
\nIt needs a lot of pre-processing to make a lot of chunks of texts and label them for the presence of the dataset but I think it will a good way to tackle this problem.
@moeinshariatnia Your first model (presence of dataset) may not be very useful because
\n(i) we already know that at least one dataset is present in each article 😏
\n(ii) same dataset can be repeated multiple number of times (in various chunks)
A better model is to probably predict the number of datasets (single vs multiple) in each article.
", "rawMarkdown": "moeinshariatnia Your first model (presence of dataset) may not be very useful because\n(i) we already know that at least one dataset is present in each article 😏\n(ii) same dataset can be repeated multiple number of times (in various chunks)\n\nA better model is to probably predict the number of datasets (single vs multiple) in each article.", "votes": null }, { "id": "1255992", "postDate": "03/29/2021 12:58:50", "content": "@sirishks it is required. because the texts are too long and you can't feed a complete text into the model. You need to take many chunks out of it and then feed them. These chunks may or may not contain the data. (You can just feed the chunks that have the dataset name because when doing inference on the test set, you do not if the chunk has the dataset or not.)
", "rawMarkdown": "sirishks it is required. because the texts are too long and you can't feed a complete text into the model. You need to take many chunks out of it and then feed them. These chunks may or may not contain the data. (You can just feed the chunks that have the dataset name because when doing inference on the test set, you do not if the chunk has the dataset or not.)", "votes": null } ], "comments": [ { "id": 1255955, "author_name": "moeinshariatnia", "author_url": "", "post_date": "03/29/2021 11:58:57", "content": "Considering similar competitions and their winning solutions, I think a good way of tackling this is the following. Training two models which one of those predicts if any dataset has been mentioned in the text and feeds it to the second model which extracts that dataset from the text.
\nIt needs a lot of pre-processing to make a lot of chunks of texts and label them for the presence of the dataset but I think it will a good way to tackle this problem.
@moeinshariatnia Your first model (presence of dataset) may not be very useful because
\n(i) we already know that at least one dataset is present in each article 😏
\n(ii) same dataset can be repeated multiple number of times (in various chunks)
A better model is to probably predict the number of datasets (single vs multiple) in each article.
", "votes": null, "replies": [] }, { "id": 1255992, "author_name": "moeinshariatnia", "author_url": "", "post_date": "03/29/2021 12:58:50", "content": "@sirishks it is required. because the texts are too long and you can't feed a complete text into the model. You need to take many chunks out of it and then feed them. These chunks may or may not contain the data. (You can just feed the chunks that have the dataset name because when doing inference on the test set, you do not if the chunk has the dataset or not.)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1254213": "If the task of this competition is regarded as a NER or QA problem, we need to extract the required answers from texts with an average length of 50,000, which seems to be difficult to achieve with BERT or LSTM. (dozens of sliding windows?）", "1255955": "Considering similar competitions and their winning solutions, I think a good way of tackling this is the following. Training two models which one of those predicts if any dataset has been mentioned in the text and feeds it to the second model which extracts that dataset from the text. \nIt needs a lot of pre-processing to make a lot of chunks of texts and label them for the presence of the dataset but I think it will a good way to tackle this problem.", "1255983": "moeinshariatnia Your first model (presence of dataset) may not be very useful because\n(i) we already know that at least one dataset is present in each article 😏\n(ii) same dataset can be repeated multiple number of times (in various chunks)\n\nA better model is to probably predict the number of datasets (single vs multiple) in each article.", "1255992": "sirishks it is required. because the texts are too long and you can't feed a complete text into the model. You need to take many chunks out of it and then feed them. These chunks may or may not contain the data. (You can just feed the chunks that have the dataset name because when doing inference on the test set, you do not if the chunk has the dataset or not.)" }, "source": "meta" }