Misleading HTML captions for some datapoints

[JaMe76]

Hi,

Thank you for providing that amazing dataset, which is undoubtably one of the best, if not the best for this subject.

When converting the HTML into a kind of tile structure, I sometimes ran into issues that might be problematic:

I assume that for each row, the sum of all column spans, where cells without colspan attribute are considered having column span equal to one, is the same. Same holds true for each column, so no cells can overlap. However, when tiling the table with its cells taking into account their positions and spans, I realized that there are some HTML strings that either allow cell overlapping or some rows having more columns than others.

I checked the tiling for the first 2000 tables of the training set and discovered 14 instances with that peculiarity. Below are four examples where I highlighted, according to my view, the problematic cell annotation within the HTML string.

Thanks

NB. rs, cs corresponds to rowspan, colspan, resp.
PMC3900820_003_00

PMC3900788_002_00

PMC4723163_004_00

PMC3230392_003_00

[ajjimeno]

Thank you for the feedback. The data set has been generated automatically using the HTML provided by the journals and the images of the tables that we recovered from the articles. It could be that in some cases the HTML might not fully represent the table image. The number of such cases should be small.

…

On Tue, Sep 8, 2020 at 4:02 AM JaMe76 ***@***.***> wrote: Hi, Thank you for providing that amazing dataset, which is undoubtably one of the best, if not the best for this subject. When converting the HTML into a kind of tile structure, I sometimes ran into issues that might be problematic: I assume that for each row, the sum of all column spans, where cells without colspan attribute are considered having column span equal to one, is the same. Same holds true for each column, so no cells can overlap. However, when tiling the table with its cells taking into account their positions and spans, I realized that there are some HTML strings that either allow cell overlapping or some rows having more columns than others. I checked the tiling for the first 2000 tables of the training set and discovered 14 instances with that peculiarity. Below are four examples where I highlighted, according to my view, the problematic cell annotation within the HTML string. Thanks NB. rs, cs corresponds to rowspan, colspan, resp. PMC3900820_003_00 [image: PMC3900820_003_00] <https://user-images.githubusercontent.com/50324223/92409705-cea04680-f141-11ea-959d-a72ff839a2bd.jpg> PMC3900788_002_00 [image: PMC3900788_002_00] <https://user-images.githubusercontent.com/50324223/92409712-d3fd9100-f141-11ea-844c-f325aee67cc1.jpg> PMC4723163_004_00 [image: PMC4723163_004_00] <https://user-images.githubusercontent.com/50324223/92409715-d829ae80-f141-11ea-86b7-692d123a3e92.jpg> PMC3230392_003_00 [image: PMC3230392_003_00] <https://user-images.githubusercontent.com/50324223/92409717-dbbd3580-f141-11ea-91fa-a793169daa92.jpg> — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#13>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA6BZDIUIEX2EFUJS67CCVDSEUN2DANCNFSM4Q6WCGGQ> .

[ZaishengLi]

Thank you for providing that amazing dataset.
I also found some structural annotation errors in the html. I think it is normal for the data set to have slight noise. But I am not sure whether the final data set used for evaluation will be artificially corrected for noise. If it will be artificially corrected, then I will tend to write scripts to filter out noisy samples during the training process.
Thanks
@ajjimeno

[ajjimeno]

Thanks for the feedback. We estimated that the number of errors was small. We did not correct the tables in the test set. This is not the same for PubLayNet, where both validation and test sets have gone through a cleaning phase.

…

On Thu, Dec 17, 2020 at 2:15 PM Rerise-Lee ***@***.***> wrote: Thank you for providing that amazing dataset. I also found some structural annotation errors in the html. I think it is normal for the data set to have slight noise. But I am not sure whether the final data set used for evaluation will be artificially corrected for noise. If it will be artificially corrected, then I will tend to write scripts to filter out noisy samples during the training process. Thanks @ajjimeno <https://github.com/ajjimeno> — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#13 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA6BZDLPMQDGG6PQDW36SATSVFZTLANCNFSM4Q6WCGGQ> .