To save time in the model runs, it may be a good idea to pre-select pieces of the articles text that contain the dataset titles. It could be sections, regular length slices or sentences. This is especially important since the documents that we get are sometimes very long. For example c12a3ad2-17bc-4b8b-8779-77ba2680ae96 is over 11 million characters.
A very simple way to pre-select title candidates may be to search for keywords that occur in the titles, like \"survey\". This notebook illustrates that 11 keywords are enough to cover 88.5% of the titles resent in the train set.
\nAnother notebook by @bowdenjr also looks at words in the titles, but without elimination of the stopwords. However it adds analysis of bi- tri- and quad-grams.
\nNext step: research words occurring in the vicinity of the titles.
", "messages": [ { "id": "1264362", "postDate": "04/06/2021 05:24:37", "content": "To save time in the model runs, it may be a good idea to pre-select pieces of the articles text that contain the dataset titles. It could be sections, regular length slices or sentences. This is especially important since the documents that we get are sometimes very long. For example c12a3ad2-17bc-4b8b-8779-77ba2680ae96 is over 11 million characters.
A very simple way to pre-select title candidates may be to search for keywords that occur in the titles, like \"survey\". This notebook illustrates that 11 keywords are enough to cover 88.5% of the titles resent in the train set.
\nAnother notebook by @bowdenjr also looks at words in the titles, but without elimination of the stopwords. However it adds analysis of bi- tri- and quad-grams.
\nNext step: research words occurring in the vicinity of the titles.
", "rawMarkdown": "To save time in the model runs, it may be a good idea to pre-select pieces of the articles text that contain the dataset titles. It could be sections, regular length slices or sentences. This is especially important since the documents that we get are sometimes very long. For example `c12a3ad2-17bc-4b8b-8779-77ba2680ae96` is over 11 million characters.\n\nA very simple way to pre-select title candidates may be to search for keywords that occur in the titles, like \"survey\". [This notebook](https://www.kaggle.com/poedator/coleridge-words-in-titles) illustrates that 11 keywords are enough to cover 88.5% of the titles resent in the train set. \n\n[Another notebook by @bowdenjr](https://www.kaggle.com/bowdenjr/coleridge-publication-titles-eda) also looks at words in the titles, but without elimination of the stopwords. However it adds analysis of bi- tri- and quad-grams.\n\nNext step: research words occurring in the vicinity of the titles.", "votes": null }, { "id": "1264383", "postDate": "04/06/2021 05:51:12", "content": "I like your idea about using a subset of sentences (those which reference data).
\nHowever, in addition to keywords 'Survey', 'Dataset', 'Data', there are several other words such as
\n'Study', 'Census', 'Genome Sequences', etc which are not easy to identify.
11 keywords are enough to cover 88.5% of the titles
\nIn my experience in Kaggle, discarding 11.5% of \"important\" data may not help in getting medals!
I recommend directly starting with NLP deep learning models, instead of attempting solutions based on keywords.
", "rawMarkdown": "I like your idea about using a subset of sentences (those which reference data).\n\nHowever, in addition to keywords 'Survey', 'Dataset', 'Data', there are several other words such as \n'Study', 'Census', 'Genome Sequences', etc which are not easy to identify.\n\n`11 keywords are enough to cover 88.5% of the titles`\nIn my experience in Kaggle, discarding 11.5% of \"important\" data may not help in getting medals!\n\nI recommend directly starting with NLP deep learning models, instead of attempting solutions based on keywords.", "votes": null }, { "id": "1264440", "postDate": "04/06/2021 06:51:45", "content": "@sirishks You made a few good points. To be clear, my post was just to open discussion of ways to pre-filter text before running heavier models.
\n88% may be high or low depending on context.
\nSome keywords may look relevant but overlap with others. Keeping them is OK if it is not slowing the processing.
There's also the issue of identifying phrases such as \"csv data\" or \"txt data\". We get heavily penalized for stuff like that.
", "rawMarkdown": "There's also the issue of identifying phrases such as \"csv data\" or \"txt data\". We get heavily penalized for stuff like that.", "votes": null }, { "id": "1266875", "postDate": "04/08/2021 06:21:40", "content": "Do you mean that phrases such as \"csv data\" cause False Positives? I tend to agree, but how to test for that precesely? Do you think that a custom list of stopwords (like csv and txt) may help here ?
Following this strategy, I think the only thing you can do is to visually inspect the values. I just think the only way to do this correctly is going to be using the typical suspects: Bert, transformers, spacy, et.al.
", "rawMarkdown": "Following this strategy, I think the only thing you can do is to visually inspect the values. I just think the only way to do this correctly is going to be using the typical suspects: Bert, transformers, spacy, et.al.", "votes": null } ], "comments": [ { "id": 1264383, "author_name": "sirishks", "author_url": "", "post_date": "04/06/2021 05:51:12", "content": "I like your idea about using a subset of sentences (those which reference data).
\nHowever, in addition to keywords 'Survey', 'Dataset', 'Data', there are several other words such as
\n'Study', 'Census', 'Genome Sequences', etc which are not easy to identify.
11 keywords are enough to cover 88.5% of the titles
\nIn my experience in Kaggle, discarding 11.5% of \"important\" data may not help in getting medals!
I recommend directly starting with NLP deep learning models, instead of attempting solutions based on keywords.
", "votes": null, "replies": [ { "id": 1264440, "author_name": "poedator", "author_url": "", "post_date": "04/06/2021 06:51:45", "content": "@sirishks You made a few good points. To be clear, my post was just to open discussion of ways to pre-filter text before running heavier models.
\n88% may be high or low depending on context.
\nSome keywords may look relevant but overlap with others. Keeping them is OK if it is not slowing the processing.
There's also the issue of identifying phrases such as \"csv data\" or \"txt data\". We get heavily penalized for stuff like that.
", "votes": null, "replies": [] }, { "id": 1266875, "author_name": "poedator", "author_url": "", "post_date": "04/08/2021 06:21:40", "content": "Do you mean that phrases such as \"csv data\" cause False Positives? I tend to agree, but how to test for that precesely? Do you think that a custom list of stopwords (like csv and txt) may help here ?
Following this strategy, I think the only thing you can do is to visually inspect the values. I just think the only way to do this correctly is going to be using the typical suspects: Bert, transformers, spacy, et.al.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1264362": "To save time in the model runs, it may be a good idea to pre-select pieces of the articles text that contain the dataset titles. It could be sections, regular length slices or sentences. This is especially important since the documents that we get are sometimes very long. For example `c12a3ad2-17bc-4b8b-8779-77ba2680ae96` is over 11 million characters.\n\nA very simple way to pre-select title candidates may be to search for keywords that occur in the titles, like \"survey\". [This notebook](https://www.kaggle.com/poedator/coleridge-words-in-titles) illustrates that 11 keywords are enough to cover 88.5% of the titles resent in the train set. \n\n[Another notebook by @bowdenjr](https://www.kaggle.com/bowdenjr/coleridge-publication-titles-eda) also looks at words in the titles, but without elimination of the stopwords. However it adds analysis of bi- tri- and quad-grams.\n\nNext step: research words occurring in the vicinity of the titles.", "1264383": "I like your idea about using a subset of sentences (those which reference data).\n\nHowever, in addition to keywords 'Survey', 'Dataset', 'Data', there are several other words such as \n'Study', 'Census', 'Genome Sequences', etc which are not easy to identify.\n\n`11 keywords are enough to cover 88.5% of the titles`\nIn my experience in Kaggle, discarding 11.5% of \"important\" data may not help in getting medals!\n\nI recommend directly starting with NLP deep learning models, instead of attempting solutions based on keywords.", "1264440": "sirishks You made a few good points. To be clear, my post was just to open discussion of ways to pre-filter text before running heavier models.\n88% may be high or low depending on context. \nSome keywords may look relevant but overlap with others. Keeping them is OK if it is not slowing the processing.", "1266415": "There's also the issue of identifying phrases such as \"csv data\" or \"txt data\". We get heavily penalized for stuff like that.", "1266875": "Do you mean that phrases such as \"csv data\" cause False Positives? I tend to agree, but how to test for that precesely? Do you think that a custom list of stopwords (like `csv` and `txt`) may help here ?", "1267563": "Following this strategy, I think the only thing you can do is to visually inspect the values. I just think the only way to do this correctly is going to be using the typical suspects: Bert, transformers, spacy, et.al." }, "source": "meta" }