Deep syntax
The goal of the deep syntax is to give a linguistic description of sentences which is closer to a semantic representation. More information about deep syntax can be found on the Deep-sequoia project.
For the sentence:
- “La souris a été mangée par le chat.” ["The mouse was eaten by the cat."].
the deep structure (following Deep-sequoia guidelines) is:
With Grew, this representation can be computed from the surface syntax in two steps:
- A general representation (called deep_and_surf) encodes both surface and deep syntax in the same structure.
- A projection from the deep_and_surf to the deep structure
Building the mixed structure
The GRS used to build the mixed deep_and_surf structure can be obtained by:
git clone https://gitlab.inria.fr/grew/SSQtoDSQ.git
The input of the GRS which produced the deep_and_surf structure is the surf structure.
We recall here the surface structure (see Dependency parsing page) for our example sentence and we suppose that the file test.surf.conll
contains the CoNLL-U description below:
# sent_id = 00000
1 La le D DET _ 2 det _ _
2 souris souris N NC s=c 5 suj _ _
3 a avoir V V m=ind 5 aux.tps _ _
4 été être V VPP m=part|t=past 5 aux.pass _ _
5 mangée manger V VPP m=part|t=past _ _ _ _
6 par par P P _ 5 p_obj.agt _ _
7 le le D DET _ 8 det _ _
8 chat chat N NC s=c 6 obj.p _ _
9 . . PONCT PONCT _ 5 ponct _ _
The mixed structure is then computed with the command:
grew transform -grs SSQtoDSQ/grs/main_dsq.grs -i test.surf.conll -o test.deep_and_surf.conll
which produces the file test.deep_and_surf.conll
which contains the code below corresponding the next figure
# global.columns = ID FORM LEMMA UPOS XPOS FEATS HEAD DEPREL DEPS MISC
# sent_id = 00000
1 La le D DET _ 2 det _ _
2 souris souris N NC s=c 5 suj:suj _ _
3 a avoir V V m=ind|void=y 5 S:aux.tps _ _
4 été être V VPP m=part|t=past|void=y 5 S:aux.pass _ _
5 mangée manger V VPP m=part|t=past _ _ _ _
6 par par P P _ 5 p_obj.agt _ _
7 le le D DET _ 8 det _ _
8 chat chat N NC s=c 6 obj.p _ _
9 . . PONCT PONCT _ 5 ponct _ _
Building the deep structure
The deep structure is a projection form the mixed structure.
This projection is realised with a GRS file sequoia_proj.grs
which can be download with the commands:
wget https://gitlab.inria.fr/sequoia/deep-sequoia/raw/master/tools/sequoia_proj.grs
The deep structure is then computed with the command:
grew transform -grs sequoia_proj.grs -strat deep -i test.deep_and_surf.conll -o test.deep.conll
The output test.deep.conll
is given below (code and picture):
# global.columns = ID FORM LEMMA UPOS XPOS FEATS HEAD DEPREL DEPS MISC
# sent_id = 00000
1 La le D DET _ 2 det _ _
2 souris souris N NC s=c 5 suj _ _
3 a avoir V V m=ind|void=y 0|1|2|4|5|6|7|8|9 void|void|void|void|void|void|void|void|void _ _
4 été être V VPP m=part|t=past|void=y 0|1|2|3|5|6|7|8|9 void|void|void|void|void|void|void|void|void _ _
5 mangée manger V VPP m=part|t=past _ _ _ _
6 par par P P _ 5 p_obj.agt _ _
7 le le D DET _ 8 det _ _
8 chat chat N NC s=c 6 obj.p _ _
9 . . PONCT PONCT _ 5 ponct _ _