Issue with producing <list> and <item> in grobid version 0.7.1

[Tanmay98]

Hi,

• I tried training FULLTEXT model for list and item tags in Grobid 0.7.1-SNAPSHOT version which worked perfectly and produced the corresponding tags in the output of processFullText request.

• Now, I tried implementing the same again for the stable realese 0.7.1 but was unable to get the tags. Same issue occured for the 0.7.2-SNAPSHOT

-For predicting the item and list tags, here is a sample annotation in the training.tei.xml file:

-Any clues on why this might be happening? Since the java code also seems to be exactly same for both versions 0.7.1 and 0.7.1-snapshot

[Tanmay98]

Hi @kermitt2 , I fixed the issue by uncommenting couple of lines in the file TEIFullltextSaxParser.java in grobid-trainer directory.

[kermitt2]

Hi @Tanmay98 !

Sorry for the late answer and thanks for the issue. I think I commented that part on purposes because I am not happy with the current accuracy with lists. I would need to review the training data (which has some annotation issues for this if I remember well) and maybe add some more. Usually when a structure is not reliable enough, I prefer to "mute" it from the current output.

[Tanmay98]

No worries, although I am getting a great accuracy for <item> since I am using more data for training the fulltext model.

[kermitt2]

Ah I see ! Good to know thanks.

Let's keep the issue open until someone can find the time to come back to the list in the public training data :)