Re: Review Request 114632: Improve pdf title extraction
> On Dec. 26, 2013, 1:57 a.m., Christoph Feck wrote: > > Hm, you broke the comment :) What do you mean? It all works fine here. - Luis --- This is an automatically generated e-mail. To reply, visit: https://git.reviewboard.kde.org/r/114632/#review46156 --- On Dec. 23, 2013, 4:14 p.m., Luis Silva wrote: > > --- > This is an automatically generated e-mail. To reply, visit: > https://git.reviewboard.kde.org/r/114632/ > --- > > (Updated Dec. 23, 2013, 4:14 p.m.) > > > Review request for Baloo and Vishesh Handa. > > > Repository: kfilemetadata > > > Description > --- > > A good portion of scientific papers in my collection had a doi or an index > number in the title. These are in general short string chains, shorter than > the real title. > I improve extraction of titles from pdf's by setting a minimum size below > which parsing of the first page is forced. > The cut-off size is arbitrarily set to 25 characters (three "big words"). > > > Diffs > - > > src/extractors/popplerextractor.cpp > b056581f51d10b632799586eed3cc15ac539fe80 > > Diff: https://git.reviewboard.kde.org/r/114632/diff/ > > > Testing > --- > > This improved the title extraction on my pdf collection of scientific papers > by quite a lot. > > > Thanks, > > Luis Silva > > >> Visit http://mail.kde.org/mailman/listinfo/kde-devel#unsub to unsubscribe <<
Re: Review Request 114632: Improve pdf title extraction
> On Dec. 26, 2013, 1:57 a.m., Christoph Feck wrote: > > Hm, you broke the comment :) > > Luis Silva wrote: > What do you mean? It all works fine here. Yes, because the compiler does not read comments. - Christoph --- This is an automatically generated e-mail. To reply, visit: https://git.reviewboard.kde.org/r/114632/#review46156 --- On Dec. 23, 2013, 4:14 p.m., Luis Silva wrote: > > --- > This is an automatically generated e-mail. To reply, visit: > https://git.reviewboard.kde.org/r/114632/ > --- > > (Updated Dec. 23, 2013, 4:14 p.m.) > > > Review request for Baloo and Vishesh Handa. > > > Repository: kfilemetadata > > > Description > --- > > A good portion of scientific papers in my collection had a doi or an index > number in the title. These are in general short string chains, shorter than > the real title. > I improve extraction of titles from pdf's by setting a minimum size below > which parsing of the first page is forced. > The cut-off size is arbitrarily set to 25 characters (three "big words"). > > > Diffs > - > > src/extractors/popplerextractor.cpp > b056581f51d10b632799586eed3cc15ac539fe80 > > Diff: https://git.reviewboard.kde.org/r/114632/diff/ > > > Testing > --- > > This improved the title extraction on my pdf collection of scientific papers > by quite a lot. > > > Thanks, > > Luis Silva > > >> Visit http://mail.kde.org/mailman/listinfo/kde-devel#unsub to unsubscribe <<
Re: Review Request 114632: Improve pdf title extraction
> On Dec. 26, 2013, 1:57 a.m., Christoph Feck wrote: > > Hm, you broke the comment :) > > Luis Silva wrote: > What do you mean? It all works fine here. > > Christoph Feck wrote: > Yes, because the compiler does not read comments. Aside this, the approach seems too naive? DOIs have a defined structure, leading "doi: 10" (ignoring the case and making colon and whitespace optional) and in general the "problematic" tokens will have a massive digit overhead - so this could be used as additional test ( < 25 && looksLikeIndex()) - Thomas --- This is an automatically generated e-mail. To reply, visit: https://git.reviewboard.kde.org/r/114632/#review46156 --- On Dec. 23, 2013, 4:14 p.m., Luis Silva wrote: > > --- > This is an automatically generated e-mail. To reply, visit: > https://git.reviewboard.kde.org/r/114632/ > --- > > (Updated Dec. 23, 2013, 4:14 p.m.) > > > Review request for Baloo and Vishesh Handa. > > > Repository: kfilemetadata > > > Description > --- > > A good portion of scientific papers in my collection had a doi or an index > number in the title. These are in general short string chains, shorter than > the real title. > I improve extraction of titles from pdf's by setting a minimum size below > which parsing of the first page is forced. > The cut-off size is arbitrarily set to 25 characters (three "big words"). > > > Diffs > - > > src/extractors/popplerextractor.cpp > b056581f51d10b632799586eed3cc15ac539fe80 > > Diff: https://git.reviewboard.kde.org/r/114632/diff/ > > > Testing > --- > > This improved the title extraction on my pdf collection of scientific papers > by quite a lot. > > > Thanks, > > Luis Silva > > >> Visit http://mail.kde.org/mailman/listinfo/kde-devel#unsub to unsubscribe <<
Re: Review Request 114632: Improve pdf title extraction
--- This is an automatically generated e-mail. To reply, visit: https://git.reviewboard.kde.org/r/114632/ --- (Updated Jan. 6, 2014, 5:47 p.m.) Review request for Baloo and Vishesh Handa. Repository: kfilemetadata Description --- A good portion of scientific papers in my collection had a doi or an index number in the title. These are in general short string chains, shorter than the real title. I improve extraction of titles from pdf's by setting a minimum size below which parsing of the first page is forced. The cut-off size is arbitrarily set to 25 characters (three "big words"). Diffs (updated) - src/extractors/popplerextractor.cpp b056581f51d10b632799586eed3cc15ac539fe80 Diff: https://git.reviewboard.kde.org/r/114632/diff/ Testing --- This improved the title extraction on my pdf collection of scientific papers by quite a lot. Thanks, Luis Silva >> Visit http://mail.kde.org/mailman/listinfo/kde-devel#unsub to unsubscribe <<
Re: Review Request 114632: Improve pdf title extraction
> On Dec. 26, 2013, 1:57 a.m., Christoph Feck wrote: > > Hm, you broke the comment :) > > Luis Silva wrote: > What do you mean? It all works fine here. > > Christoph Feck wrote: > Yes, because the compiler does not read comments. > > Thomas Lübking wrote: > Aside this, the approach seems too naive? > DOIs have a defined structure, leading "doi: 10" (ignoring the case and > making colon and whitespace optional) and in general the "problematic" tokens > will have a massive digit overhead - so this could be used as additional test > ( < 25 && looksLikeIndex()) @Christoph: Just (finally) understood what you meant with "breaking the comment". I uploaded a new patch that (hopefully) fixes the issue in the correct way. @Thomas: The approach was meant to be naive. In this simple form, this patch takes care of all index-like cases as well as most other short garbage titles without further parsing. What would be the point of actually knowing if a very short title was actually a doi or an index? - Luis --- This is an automatically generated e-mail. To reply, visit: https://git.reviewboard.kde.org/r/114632/#review46156 --- On Dec. 23, 2013, 4:14 p.m., Luis Silva wrote: > > --- > This is an automatically generated e-mail. To reply, visit: > https://git.reviewboard.kde.org/r/114632/ > --- > > (Updated Dec. 23, 2013, 4:14 p.m.) > > > Review request for Baloo and Vishesh Handa. > > > Repository: kfilemetadata > > > Description > --- > > A good portion of scientific papers in my collection had a doi or an index > number in the title. These are in general short string chains, shorter than > the real title. > I improve extraction of titles from pdf's by setting a minimum size below > which parsing of the first page is forced. > The cut-off size is arbitrarily set to 25 characters (three "big words"). > > > Diffs > - > > src/extractors/popplerextractor.cpp > b056581f51d10b632799586eed3cc15ac539fe80 > > Diff: https://git.reviewboard.kde.org/r/114632/diff/ > > > Testing > --- > > This improved the title extraction on my pdf collection of scientific papers > by quite a lot. > > > Thanks, > > Luis Silva > > >> Visit http://mail.kde.org/mailman/listinfo/kde-devel#unsub to unsubscribe <<
Re: Review Request 114632: Improve pdf title extraction
> On Dec. 26, 2013, 1:57 a.m., Christoph Feck wrote:
> > Hm, you broke the comment :)
>
> Luis Silva wrote:
> What do you mean? It all works fine here.
>
> Christoph Feck wrote:
> Yes, because the compiler does not read comments.
>
> Thomas Lübking wrote:
> Aside this, the approach seems too naive?
> DOIs have a defined structure, leading "doi: 10" (ignoring the case and
> making colon and whitespace optional) and in general the "problematic" tokens
> will have a massive digit overhead - so this could be used as additional test
> ( < 25 && looksLikeIndex())
>
> Luis Silva wrote:
> @Christoph: Just (finally) understood what you meant with "breaking the
> comment". I uploaded a new patch that (hopefully) fixes the issue in the
> correct way.
> @Thomas: The approach was meant to be naive. In this simple form, this
> patch takes care of all index-like cases as well as most other short garbage
> titles without further parsing. What would be the point of actually knowing
> if a very short title was actually a doi or an index?
echo "The Lord of the Rings" | wc -m
22
And that's not a short title - not to mention the typical Stephen King ("It")
or other languages that use hanzi, kanji or hanja and will never met your
arbitrary 25 glyph requirement.
Though many academic papers (in western cultures at least) in fact have clumsy
long titles, that doesn't hold for other document types.
OTOH, if the "title" (=index) is some (md5, sha*) hash of the text, that will
easily outnumber 25 glyphs.
So the more honest solution seems to just omit the title field altogether.
The alternative (don't know how expensive the document scan is) would be to
check whether the title field seems like reasonable text, what could invoke the
digit ratio, the longest non-digit sequence ("0x12a21f56ea5") and maybe whether
there's any digitless word at all.
- Thomas
---
This is an automatically generated e-mail. To reply, visit:
https://git.reviewboard.kde.org/r/114632/#review46156
---
On Jan. 6, 2014, 5:47 p.m., Luis Silva wrote:
>
> ---
> This is an automatically generated e-mail. To reply, visit:
> https://git.reviewboard.kde.org/r/114632/
> ---
>
> (Updated Jan. 6, 2014, 5:47 p.m.)
>
>
> Review request for Baloo and Vishesh Handa.
>
>
> Repository: kfilemetadata
>
>
> Description
> ---
>
> A good portion of scientific papers in my collection had a doi or an index
> number in the title. These are in general short string chains, shorter than
> the real title.
> I improve extraction of titles from pdf's by setting a minimum size below
> which parsing of the first page is forced.
> The cut-off size is arbitrarily set to 25 characters (three "big words").
>
>
> Diffs
> -
>
> src/extractors/popplerextractor.cpp
> b056581f51d10b632799586eed3cc15ac539fe80
>
> Diff: https://git.reviewboard.kde.org/r/114632/diff/
>
>
> Testing
> ---
>
> This improved the title extraction on my pdf collection of scientific papers
> by quite a lot.
>
>
> Thanks,
>
> Luis Silva
>
>
>> Visit http://mail.kde.org/mailman/listinfo/kde-devel#unsub to unsubscribe <<
Re: Review Request 114632: Improve pdf title extraction
> On Dec. 26, 2013, 1:57 a.m., Christoph Feck wrote:
> > Hm, you broke the comment :)
>
> Luis Silva wrote:
> What do you mean? It all works fine here.
>
> Christoph Feck wrote:
> Yes, because the compiler does not read comments.
>
> Thomas Lübking wrote:
> Aside this, the approach seems too naive?
> DOIs have a defined structure, leading "doi: 10" (ignoring the case and
> making colon and whitespace optional) and in general the "problematic" tokens
> will have a massive digit overhead - so this could be used as additional test
> ( < 25 && looksLikeIndex())
>
> Luis Silva wrote:
> @Christoph: Just (finally) understood what you meant with "breaking the
> comment". I uploaded a new patch that (hopefully) fixes the issue in the
> correct way.
> @Thomas: The approach was meant to be naive. In this simple form, this
> patch takes care of all index-like cases as well as most other short garbage
> titles without further parsing. What would be the point of actually knowing
> if a very short title was actually a doi or an index?
>
> Thomas Lübking wrote:
> echo "The Lord of the Rings" | wc -m
> 22
>
> And that's not a short title - not to mention the typical Stephen King
> ("It") or other languages that use hanzi, kanji or hanja and will never met
> your arbitrary 25 glyph requirement.
> Though many academic papers (in western cultures at least) in fact have
> clumsy long titles, that doesn't hold for other document types.
>
> OTOH, if the "title" (=index) is some (md5, sha*) hash of the text, that
> will easily outnumber 25 glyphs.
>
> So the more honest solution seems to just omit the title field altogether.
>
> The alternative (don't know how expensive the document scan is) would be
> to check whether the title field seems like reasonable text, what could
> invoke the digit ratio, the longest non-digit sequence ("0x12a21f56ea5") and
> maybe whether there's any digitless word at all.
Honestly I don't even know why there is the rule for needing a space, looking
at my shelf of books i can see "Cryptonomicon", "Azogue", "Portico",
"Hyperion", "Endymion", "1984", and then I have stopped. Please, don't try to
be that much clever, i can understand if you want to rule out stuff like
"Microsoft Word - something.doc", but imho you're being already too broad with
the rule of "it includes microsoft". What about if i have a manual about
"Microsoft Visual Basic"?
Honestly omiting or mangling the title is a very bad thing to do. If you have a
sensible thing to run over the 1500 test pdf files i have here i'm happy to
help.
- Albert
---
This is an automatically generated e-mail. To reply, visit:
https://git.reviewboard.kde.org/r/114632/#review46156
---
On Jan. 6, 2014, 5:47 p.m., Luis Silva wrote:
>
> ---
> This is an automatically generated e-mail. To reply, visit:
> https://git.reviewboard.kde.org/r/114632/
> ---
>
> (Updated Jan. 6, 2014, 5:47 p.m.)
>
>
> Review request for Baloo and Vishesh Handa.
>
>
> Repository: kfilemetadata
>
>
> Description
> ---
>
> A good portion of scientific papers in my collection had a doi or an index
> number in the title. These are in general short string chains, shorter than
> the real title.
> I improve extraction of titles from pdf's by setting a minimum size below
> which parsing of the first page is forced.
> The cut-off size is arbitrarily set to 25 characters (three "big words").
>
>
> Diffs
> -
>
> src/extractors/popplerextractor.cpp
> b056581f51d10b632799586eed3cc15ac539fe80
>
> Diff: https://git.reviewboard.kde.org/r/114632/diff/
>
>
> Testing
> ---
>
> This improved the title extraction on my pdf collection of scientific papers
> by quite a lot.
>
>
> Thanks,
>
> Luis Silva
>
>
>> Visit http://mail.kde.org/mailman/listinfo/kde-devel#unsub to unsubscribe <<
Re: Review Request 114632: Improve pdf title extraction
> On Dec. 26, 2013, 1:57 a.m., Christoph Feck wrote:
> > Hm, you broke the comment :)
>
> Luis Silva wrote:
> What do you mean? It all works fine here.
>
> Christoph Feck wrote:
> Yes, because the compiler does not read comments.
>
> Thomas Lübking wrote:
> Aside this, the approach seems too naive?
> DOIs have a defined structure, leading "doi: 10" (ignoring the case and
> making colon and whitespace optional) and in general the "problematic" tokens
> will have a massive digit overhead - so this could be used as additional test
> ( < 25 && looksLikeIndex())
>
> Luis Silva wrote:
> @Christoph: Just (finally) understood what you meant with "breaking the
> comment". I uploaded a new patch that (hopefully) fixes the issue in the
> correct way.
> @Thomas: The approach was meant to be naive. In this simple form, this
> patch takes care of all index-like cases as well as most other short garbage
> titles without further parsing. What would be the point of actually knowing
> if a very short title was actually a doi or an index?
>
> Thomas Lübking wrote:
> echo "The Lord of the Rings" | wc -m
> 22
>
> And that's not a short title - not to mention the typical Stephen King
> ("It") or other languages that use hanzi, kanji or hanja and will never met
> your arbitrary 25 glyph requirement.
> Though many academic papers (in western cultures at least) in fact have
> clumsy long titles, that doesn't hold for other document types.
>
> OTOH, if the "title" (=index) is some (md5, sha*) hash of the text, that
> will easily outnumber 25 glyphs.
>
> So the more honest solution seems to just omit the title field altogether.
>
> The alternative (don't know how expensive the document scan is) would be
> to check whether the title field seems like reasonable text, what could
> invoke the digit ratio, the longest non-digit sequence ("0x12a21f56ea5") and
> maybe whether there's any digitless word at all.
>
> Albert Astals Cid wrote:
> Honestly I don't even know why there is the rule for needing a space,
> looking at my shelf of books i can see "Cryptonomicon", "Azogue", "Portico",
> "Hyperion", "Endymion", "1984", and then I have stopped. Please, don't try to
> be that much clever, i can understand if you want to rule out stuff like
> "Microsoft Word - something.doc", but imho you're being already too broad
> with the rule of "it includes microsoft". What about if i have a manual about
> "Microsoft Visual Basic"?
>
> Honestly omiting or mangling the title is a very bad thing to do. If you
> have a sensible thing to run over the 1500 test pdf files i have here i'm
> happy to help.
Would it make sense to refactor the code to use the (PDF supplied) document
title, and, if for whatever reason it is believed to be wrong, append the
extracted text that is believed to be a better title?
- Christoph
---
This is an automatically generated e-mail. To reply, visit:
https://git.reviewboard.kde.org/r/114632/#review46156
---
On Jan. 6, 2014, 5:47 p.m., Luis Silva wrote:
>
> ---
> This is an automatically generated e-mail. To reply, visit:
> https://git.reviewboard.kde.org/r/114632/
> ---
>
> (Updated Jan. 6, 2014, 5:47 p.m.)
>
>
> Review request for Baloo and Vishesh Handa.
>
>
> Repository: kfilemetadata
>
>
> Description
> ---
>
> A good portion of scientific papers in my collection had a doi or an index
> number in the title. These are in general short string chains, shorter than
> the real title.
> I improve extraction of titles from pdf's by setting a minimum size below
> which parsing of the first page is forced.
> The cut-off size is arbitrarily set to 25 characters (three "big words").
>
>
> Diffs
> -
>
> src/extractors/popplerextractor.cpp
> b056581f51d10b632799586eed3cc15ac539fe80
>
> Diff: https://git.reviewboard.kde.org/r/114632/diff/
>
>
> Testing
> ---
>
> This improved the title extraction on my pdf collection of scientific papers
> by quite a lot.
>
>
> Thanks,
>
> Luis Silva
>
>
>> Visit http://mail.kde.org/mailman/listinfo/kde-devel#unsub to unsubscribe <<
Re: Review Request 114632: Improve pdf title extraction
> On Dec. 26, 2013, 1:57 a.m., Christoph Feck wrote:
> > Hm, you broke the comment :)
>
> Luis Silva wrote:
> What do you mean? It all works fine here.
>
> Christoph Feck wrote:
> Yes, because the compiler does not read comments.
>
> Thomas Lübking wrote:
> Aside this, the approach seems too naive?
> DOIs have a defined structure, leading "doi: 10" (ignoring the case and
> making colon and whitespace optional) and in general the "problematic" tokens
> will have a massive digit overhead - so this could be used as additional test
> ( < 25 && looksLikeIndex())
>
> Luis Silva wrote:
> @Christoph: Just (finally) understood what you meant with "breaking the
> comment". I uploaded a new patch that (hopefully) fixes the issue in the
> correct way.
> @Thomas: The approach was meant to be naive. In this simple form, this
> patch takes care of all index-like cases as well as most other short garbage
> titles without further parsing. What would be the point of actually knowing
> if a very short title was actually a doi or an index?
>
> Thomas Lübking wrote:
> echo "The Lord of the Rings" | wc -m
> 22
>
> And that's not a short title - not to mention the typical Stephen King
> ("It") or other languages that use hanzi, kanji or hanja and will never met
> your arbitrary 25 glyph requirement.
> Though many academic papers (in western cultures at least) in fact have
> clumsy long titles, that doesn't hold for other document types.
>
> OTOH, if the "title" (=index) is some (md5, sha*) hash of the text, that
> will easily outnumber 25 glyphs.
>
> So the more honest solution seems to just omit the title field altogether.
>
> The alternative (don't know how expensive the document scan is) would be
> to check whether the title field seems like reasonable text, what could
> invoke the digit ratio, the longest non-digit sequence ("0x12a21f56ea5") and
> maybe whether there's any digitless word at all.
>
> Albert Astals Cid wrote:
> Honestly I don't even know why there is the rule for needing a space,
> looking at my shelf of books i can see "Cryptonomicon", "Azogue", "Portico",
> "Hyperion", "Endymion", "1984", and then I have stopped. Please, don't try to
> be that much clever, i can understand if you want to rule out stuff like
> "Microsoft Word - something.doc", but imho you're being already too broad
> with the rule of "it includes microsoft". What about if i have a manual about
> "Microsoft Visual Basic"?
>
> Honestly omiting or mangling the title is a very bad thing to do. If you
> have a sensible thing to run over the 1500 test pdf files i have here i'm
> happy to help.
>
> Christoph Feck wrote:
> Would it make sense to refactor the code to use the (PDF supplied)
> document title, and, if for whatever reason it is believed to be wrong,
> append the extracted text that is believed to be a better title?
I can see the point Albert is making that when a pdf has a short (but valid)
pdftitle and an unparseable first page the resulting extracted title will be
gibberish. I also agree that mangling the title just because it seemed to be
small is unacceptable. I must admit that I did not think about the cases of
hanzi, kanji or hanja for which this patch would systematically force the
parsing of the first page of the document.
The issue here is when the pdftitle does not match the real document title. In
my database of academic papers (700+) this happens a lot. Most of my other
documents are either prints to pdf, documents generated from their latex source
or Word documents converted to pdf most (90%) of which lack a pdftitle and so
have to be parsed anyway. From my experience this is a typical situation, at
least amongst academics. Of course, the best operating solution must cater for
the most common personas, not just academics, but in your experience, what
would that be?
- Luis
---
This is an automatically generated e-mail. To reply, visit:
https://git.reviewboard.kde.org/r/114632/#review46156
---
On Jan. 6, 2014, 5:47 p.m., Luis Silva wrote:
>
> ---
> This is an automatically generated e-mail. To reply, visit:
> https://git.reviewboard.kde.org/r/114632/
> ---
>
> (Updated Jan. 6, 2014, 5:47 p.m.)
>
>
> Review request for Baloo and Vishesh Handa.
>
>
> Repository: kfilemetadata
>
>
> Description
> ---
>
> A good portion of scientific papers in my collection had a doi or an index
> number in the title. These are in general short string chains, shorter than
> the real title.
> I improve extraction of titles from pdf's by setting a minimum size below
> which parsing of the first page is forced.
> The cut-off size is arbitrarily set to 25 characters (three "
