And does the full text help???

[jambo6]

Asking because I'm doing a project where I was looking at full abstracts, and have recently switched to single sentences for relation extraction because labelling was faster + abstracts seemed like it might be overkill. Very interested on thoughts of the best way to do this.

[danich1]

Asking because I'm doing a project where I was looking at full abstracts, and have recently switched to single sentences for relation extraction because labelling was faster + abstracts seemed like it might be overkill. Very interested on thoughts of the best way to do this.

So long story short full text doesn't improve performance as performance is mainly reliant on the label functions generated. If I'm following correctly you are looking at the whole abstracts for relation extraction rather than individual sentences? If so, then you will still end up with examining individual sentences for extraction. Now if you are trying to circumvent the manual effort of hand labeling sentences, then using the my constructed label functions could help.

One strong warning is that performance for this approach relies on the label functions themselves to be effective. Depending on the task, you could end up in a trap where you have to constantly go back and read a whole lot of sentences to generate more label functions just to improve model performance. If this happens manually labeling sentences themselves and not using label functions could be a lot faster.