#!/usr/bin/env perl | |
# Expands enhanced relation labels in French. For certain relations, the French | |
# treebanks provide a double label, consisting of a canonical and a final | |
# relation. This is to neutralize diathesis: final passive subject is canonical | |
# object etc. (Candito et al. 2017). Unfortunately it is not possible to | |
# represent a double label in the CoNLL-U file while complying with the UD | |
# guidelines, specifically with the limited set of characters that are allowed | |
# in a relation label. Therefore, the labels have been encoded in a form that | |
# makes the file formally valid but the label poorly readable by human users, | |
# e.g., "nsubj:passxoxobjenh". This script expands the labels to improve their | |
# readability while preserving the information they encode. It is thus a kind | |
# of antipole of the script fix_edeps_in_french.pl, although it will not restore | |
# the relation labels exactly in the form in which the other script reads them. | |
# | |
# With --list-edeprels, the script will also list the resulting edeprels to STDERR. | |
# Copyright © 2020 Dan Zeman <zeman@ufal.mff.cuni.cz> | |
# License: GNU GPL | |
use utf8; | |
use open ':utf8'; | |
binmode(STDIN, ':utf8'); | |
binmode(STDOUT, ':utf8'); | |
binmode(STDERR, ':utf8'); | |
use Getopt::Long; | |
my $list_edeprels = 0; | |
GetOptions | |
( | |
'list-edeprels' => \$list_edeprels | |
); | |
my %edeprels; | |
while(<>) | |
{ | |
my $line = $_; | |
if($line =~ m/^\d/) | |
{ | |
my @f = split(/\t/, $line); | |
my $edeps = $f[8]; | |
if($edeps ne '_') | |
{ | |
my @edeps = map {m/^(\d+(?:\.\d+)?):(.+)$/; [$1, $2]} (split(/\|/, $edeps)); | |
foreach my $edep (@edeps) | |
{ | |
my $h = $edep->[0]; | |
my $d = $edep->[1]; | |
# Relations that are specific to the enhanced graph (they don't have a counterpart in the basic tree) | |
# have the suffix "enh". Sometimes it is not preceded by a colon because the maximum allowed number of | |
# colons would be exceeded. | |
#$d =~ s/([^:])enh$/$1:enh/; | |
# Marie: the ":enh" suffix is only here to *easily spot* the enhanced dependencies, | |
# it is used (in a different form) in our internal format. Since the information on | |
# what is the original non-enhanced dependency is given in columns 7 and 8, you should | |
# completely remove all the :enh suffixes in column 9, they are useless and harmful. | |
$d =~ s/:?enh$//; | |
# The final and canonical relations are separated by "xox" or ":xox", a sequence that does not occur | |
# anywhere else in the labels. Again, the colon may not be there and may have to be restored. | |
$d =~ s/([^:])xox/$1:xox/g; | |
# Finally, replace ":xox" by something looking more like a delimiter. | |
$d =~ s/:xox/\@/g; | |
# In rare cases, a colon may also be missing between a main universal relation and its subtype. Restore it. | |
$d =~ s/(nsubj)(caus|xxx)/$1:$2/g; | |
# It is a bug in the original annotation that some relations have the ":xxx" subtype. | |
# Marie: the XXX suffixes in labels are an internal mark. Could you | |
# - remove the :xxx suffix in the Sequoia (i.e. keep "obj@nsubj" as label) | |
# - change all "obj@nsubj:xxx" into "obj" in the FQB files? | |
# This script does not know whether it is processing Sequoia or FQB. | |
# But there are 9 occurrences in FQB and 1 in Sequoia, so we are taking the FQB rule as default. | |
$d =~ s/^obj\@nsubj:xxx$/obj/; | |
$edeprels{$d}++; | |
$edep->[0] = $h; | |
$edep->[1] = $d; | |
} | |
$edeps = join('|', map {"$_->[0]:$_->[1]"} (sort {my $r = $a->[0] <=> $b->[0]; unless($r) {$r = $a->[1] cmp $b->[1]} $r} (@edeps))); | |
} | |
$f[8] = $edeps; | |
$line = join("\t", @f); | |
} | |
print($line); | |
} | |
if($list_edeprels) | |
{ | |
my @edeprels = sort(keys(%edeprels)); | |
foreach my $edeprel (@edeprels) | |
{ | |
print STDERR ("$edeprel\n"); | |
} | |
} | |