Repository navigation
Generate training data: label the people a review panel names - #790
Merged
Merged
Conversation
`ore` sets a panel beside its front matter naming who reviewed the article and where they work, and files each review as a `<sub-article>` of its own, so both are declared and neither was read. Nine of the ten documents the header corpus holds now label them, and background falls from 20.0% to 17.4% of what the `ore` header region sets. The person takes a label of their own, written as a note because GROBID has no element for a reviewer and writes every kind it has no element for that way. Where they work takes the ordinary affiliation mark-up, which is what the guidelines say of an editor's and what GROBID's corpus does beside a reviewer: only the person's label is in dispute, and that one is provisional on the issue raised upstream. Two things the panel needs that a byline does not. The values are looked for in the column the page sets beside the article and nowhere else, because the reports name the same people again at the end of the document and a value that reached them took the review region into the header. And inside that column they are matched at the bar a reference's parent takes, because the panel drops a part the deposit carries -- a region between the city and the country -- which costs two fifths of the match. A reply to a review is filed as a `<sub-article>` inside the report it answers, and its contributor is an author of the article, so only the report's own stub is read. The segmentation, reference-segmenter and citation corpora are byte-identical.
The panel sets a reviewer's name and where they work with a comma between them, and the comma belongs to neither, so a span that matched starting on one gives it back. Where the affiliation is looked for is now the end of the name it came with, which also keeps a name whole: `Noorlizawati Abd Rahim` had half of itself taken into the affiliation beside it.
A deposit declares the departments in front of the institution and the region between the city and the country; the panel sets neither, so under half of what is declared can be all the page carries and no bar low enough to take that is safe generally. The forms are tried instead -- each run of the declared parts with the leading ones dropped, with and without the part before the last, longest first -- and the one the page sets is the one that matches, at the ordinary threshold. A form the page does not set simply finds nothing. Every panel entry now names where its reviewer works, where six did not.
The forms were tried at the panel's lowered bar rather than the ordinary one, so a longer form that merely overlaps the page cleared it first and labelled only as far as the page carried it -- which left the country outside the affiliation on one panel entry. A form the page sets matches outright, so the ordinary bar is the one to try them at. The panel's lower bar stays for the declared form, where it is still what reaches three documents.
Test fixtures and the comments beside them read as placeholders rather than as something that could be taken for a real person, institution or publication. Only the generated data carries what a deposit declares.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
part of https://github.com/eLifePathways/ScienceBeam2.0/issues/205
An open review platform sets a panel beside the front matter naming who
reviewed the article and where they work, and files each review as a
<sub-article>of its own, so both are declared and neither was read. Thegenerated header data taught that a person's name followed by their institution
is background, which is the opposite of what GROBID's half of the mixed corpus
teaches wherever a header carries the same shape. With no class or sample
weighting in either trainer, that is a full training example teaching against
it rather than a weak one.
The person takes a label of their own, written as a note because GROBID has no
element for a reviewer and writes every kind it has no element for that way.
Where they work takes the ordinary affiliation mark-up, which is what the
guidelines say of an editor's and what GROBID's corpus does beside a reviewer.
Only the person's label is in dispute, and that one is provisional on an issue
raised upstream: GROBID annotates a reviewer as the editor, which its own
guidelines do not describe.
Three things the panel needs that a byline does not. The values are looked for
in the column beside the article and nowhere else, because the reports name the
same people again after the body and a value that reached them took the review
region into the header. A reply to a review is filed inside the report it
answers and its contributor is an author of the article, so only the report's
own stub is read. And the panel sets a shortened form of each affiliation,
dropping the departments the deposit declares in front of the institution and
the region it declares between the city and the country, so the forms are tried
and the one the page sets is the one that matches.
The segmentation, reference-segmenter and citation corpora are unchanged.