sapienzanlp/gsrl

GSRL is a seq2seq model for end-to-end dependency- and span-based SRL (IJCAI2021).

gsrl's Issues

Need some help with linearization

Thank you very much for your contribution in gsrl!
Your design for NESTED SPAN-BASED SRL really inspired me , but I have trouble with how to merge some SRL sentences into one NESTED sentence , I would be very grateful if you could make this preprocessing method for this part in public .

Question about the evaluation

Thanks for publishing your code. When I reading the code for the evaluation, I have a question concerning the generation of the prediction file.
In

gsrl/src/modelling/evaluation.py

Line 81 in 64291a7

def write_props_format_predictions(out, predictions, gold_pred):

, if the input sentence is [A, B, C, D, E, F, G], and we assume the golden predicate is [A, C, F] (the role is B, D, G, we assume each predicate has only one role), and the model predicts the predicate is [A, F] (and the corresponding role is B, G, respectively). Based on the code, the results will be something like

A	(V*)	*
-	Arg0*)	*	
C	*	(V*)	*
-	*	*
-	*	*
F	*	(V*)	
-	*	(Arg0*)

namely, that some columns will have two predicates, and some rows will have more columns (the third column has three rows, while others have two). Will this code generate this way? If so, will this negatively influence the performance?

Question About Dataset Preprocessing.

Hi, thanks for your great work. I am trying to train the gsrl model on conll-2012 dataset. However I met some trouble during dataset preprocessing.
I have the official dataset (Ontonotes5), which looks like

$ tar zxvf LDC2013T19.tgz
$ tree -L 3 -d ontonotes-release-5.0/data/files/data
ontonotes-release-5.0/
??? data
    ??? files
        ??? data
            ??? arabic
            ?   ??? annotations
            ?       ??? nw
            ??? chinese
            ?   ??? annotations
            ?       ??? bc
            ?       ??? bn
            ?       ??? mz
            ?       ??? nw
            ?       ??? tc
            ?       ??? wb
            ??? english
                ??? annotations
                    ??? bc
                    ??? bn
                    ??? mz
                    ??? nw
                    ??? pt
                    ??? tc
                    ??? wb

I am wondering what does the final data struture look like (train.txt,dev.txt,test.txt) and how to transform the raw dataset into yours, is there a transform script ?
Thanks~

API / module

Hi, first off thank you very much for your work overall! You guy did so much on SRL :)
I need to run SRL on another dataset and as I needed predicate senses and a dependency based output, thus your model was just perfect. Is there a fairly easy way to run predictions on new unstructured sentences ? Any suggestions? or Are you maybe planning to make a pip module?