Grew

Graph Rewriting for NLP

grewpy Tutorial

Grewpy tutorial: Modify data

Download the notebook here.

import grewpy
from grewpy import Corpus, CorpusDraft, Request
grewpy.set_config("sud") # ud or basic
corpus = Corpus("SUD_English-PUD")

connected to port: 61637

Access data in a corpus

# Access to the corpus
sentence = corpus[1]
print("A corpus is a set of graphs:", type(sentence))

A corpus is a set of graphs: <class 'grewpy.graph.Graph'>

# Each graph is a sentence and contains all its information
print("Sentence metadata:")
sentence.meta

Sentence metadata:

{'sent_id': 'n01001013',
 'text': 'For those who follow social media transitions on Capitol Hill, this will be a little different.',
 '_filename': 'en_pud-sud-test.conllu'}

# Sentence order, which in this case is the same as the token's id
print(sentence.order)

['0', '1', '2', '3', '4', '5', '6', '7', '8', '9', '10', '11', '12', '13', '14', '15', '16', '17', '18']

# Token features, which make possible to access every token feature
print(sentence.features)

# e.g get all upos of the sentence
print([sentence.features[id]['upos'] for id in sentence.features if id != "0"])

{'0': {'form': '__0__'}, '1': {'form': 'For', 'lemma': 'for', 'textform': 'For', 'upos': 'ADP', 'wordform': 'For', 'xpos': 'IN'}, '2': {'Number': 'Plur', 'PronType': 'Dem', 'form': 'those', 'lemma': 'those', 'textform': 'those', 'upos': 'PRON', 'wordform': 'those', 'xpos': 'DT'}, '3': {'PronType': 'Rel', 'form': 'who', 'lemma': 'who', 'textform': 'who', 'upos': 'PRON', 'wordform': 'who', 'xpos': 'WP'}, '4': {'Mood': 'Ind', 'Tense': 'Pres', 'VerbForm': 'Fin', 'form': 'follow', 'lemma': 'follow', 'textform': 'follow', 'upos': 'VERB', 'wordform': 'follow', 'xpos': 'VBP'}, '5': {'Degree': 'Pos', 'form': 'social', 'lemma': 'social', 'textform': 'social', 'upos': 'ADJ', 'wordform': 'social', 'xpos': 'JJ'}, '6': {'Number': 'Sing', 'form': 'media', 'lemma': 'media', 'textform': 'media', 'upos': 'NOUN', 'wordform': 'media', 'xpos': 'NN'}, '7': {'Number': 'Plur', 'form': 'transitions', 'lemma': 'transition', 'textform': 'transitions', 'upos': 'NOUN', 'wordform': 'transitions', 'xpos': 'NNS'}, '8': {'form': 'on', 'lemma': 'on', 'textform': 'on', 'upos': 'ADP', 'wordform': 'on', 'xpos': 'IN'}, '9': {'Number': 'Sing', 'form': 'Capitol', 'lemma': 'Capitol', 'textform': 'Capitol', 'upos': 'PROPN', 'wordform': 'Capitol', 'xpos': 'NNP'}, '10': {'Number': 'Sing', 'SpaceAfter': 'No', 'form': 'Hill', 'lemma': 'Hill', 'textform': 'Hill', 'upos': 'PROPN', 'wordform': 'Hill', 'xpos': 'NNP'}, '11': {'form': ',', 'lemma': ',', 'textform': ',', 'upos': 'PUNCT', 'wordform': ',', 'xpos': ','}, '12': {'Number': 'Sing', 'PronType': 'Dem', 'form': 'this', 'lemma': 'this', 'textform': 'this', 'upos': 'PRON', 'wordform': 'this', 'xpos': 'DT'}, '13': {'VerbForm': 'Fin', 'form': 'will', 'lemma': 'will', 'textform': 'will', 'upos': 'AUX', 'wordform': 'will', 'xpos': 'MD'}, '14': {'VerbForm': 'Inf', 'form': 'be', 'lemma': 'be', 'textform': 'be', 'upos': 'AUX', 'wordform': 'be', 'xpos': 'VB'}, '15': {'Definite': 'Ind', 'PronType': 'Art', 'form': 'a', 'lemma': 'a', 'textform': 'a', 'upos': 'DET', 'wordform': 'a', 'xpos': 'DT'}, '16': {'Degree': 'Pos', 'form': 'little', 'lemma': 'little', 'textform': 'little', 'upos': 'ADJ', 'wordform': 'little', 'xpos': 'JJ'}, '17': {'Degree': 'Pos', 'SpaceAfter': 'No', 'form': 'different', 'lemma': 'different', 'textform': 'different', 'upos': 'ADJ', 'wordform': 'different', 'xpos': 'JJ'}, '18': {'form': '.', 'lemma': '.', 'textform': '.', 'upos': 'PUNCT', 'wordform': '.', 'xpos': '.'}}
['ADP', 'PRON', 'PRON', 'VERB', 'ADJ', 'NOUN', 'NOUN', 'ADP', 'PROPN', 'PROPN', 'PUNCT', 'PRON', 'AUX', 'AUX', 'DET', 'ADJ', 'ADJ', 'PUNCT']

# It's possible to access to edges between nodes as successors
print(sentence.sucs)

{'17': [('16', Fs_edge({'1': 'udep', 'deep': 'npmod'}))], '16': [('15', Fs_edge({'1': 'det'}))], '14': [('17', Fs_edge({'1': 'comp', '2': 'pred'}))], '13': [('18', Fs_edge({'1': 'punct'})), ('14', Fs_edge({'1': 'comp', '2': 'aux'})), ('12', Fs_edge({'1': 'subj'})), ('11', Fs_edge({'1': 'punct'})), ('1', Fs_edge({'1': 'udep'}))], '10': [('9', Fs_edge({'1': 'compound'}))], '8': [('10', Fs_edge({'1': 'comp', '2': 'obj'}))], '7': [('8', Fs_edge({'1': 'udep'})), ('6', Fs_edge({'1': 'compound'}))], '6': [('5', Fs_edge({'1': 'mod'}))], '4': [('7', Fs_edge({'1': 'comp', '2': 'obj'})), ('3', Fs_edge({'1': 'subj'}))], '2': [('4', Fs_edge({'1': 'mod', 'deep': 'relcl'}))], '1': [('2', Fs_edge({'1': 'comp', '2': 'obj'}))], '0': [('13', Fs_edge({'1': 'root'}))]}

Modifying a corpus

Corpus is an abstract object which cannot be modified directly:

try:
	corpus[0] = corpus[1]
except TypeError as error_message:
	print (f"{error_message}")

'Corpus' object does not support item assignment

CorpusDraft is an object similar to Corpus but which is mutable. Below, we add the feature Transitive=Yes to all occurrences of verbs with a direct object.

  1. We make the search on corpus (an instance of Corpus).
  2. The modification is done on a CorpusDraft counterpart named draft.
  3. The draft should be transformed again into a Corpus (names corpus2 below) in order to use the count method.
# step 1
req7 = Request().pattern("X[upos=VERB]; Y[upos=NOUN|PROPN|PRON]; X-[comp:obj]->Y")
occurrences = corpus.search(req7)

# step 2
draft = CorpusDraft(corpus)
for occ in occurrences:
    sent_id = occ['sent_id']
    verb_node_id = occ['matching']['nodes']['X']
    draft[sent_id][verb_node_id].update({"Transitive": "Yes"})

# step 3
corpus2 = Corpus(draft)
corpus2.count(Request("pattern { X[Transitive=Yes] }"))

853

It’s possible to modify a whole CorpusDraft with a function getting a graph as input.

def relabel_noun(graph):
    for node in graph:
        if 'upos' in graph[node] and graph[node]['upos'] == 'NOUN':
            graph[node]['upos'] = 'N'
    return graph

draft3 = draft.map(relabel_noun)
# Note that the map function has replaced the apply function which is deprecated in 0.6

# Again, we need to turn the result into a `Corpus` before using the `count` method.
corpus3 = Corpus(draft3)
corpus3.count(Request("pattern { X[upos=N] }"))

4036

Modifying a corpus using a GRS (Graph Rewriting System)

In many cases, it is not required to uses a CorpusDraft and the modification of a corpus can be encoded with graph rewriting rules.

The example above (identifying transitive verbs) can be rephrased as below. See TODO link for an explanation of the without clause in this example.

from grewpy import GRS

s = """
strat main { Onf(tv) }

rule tv {
  pattern { X[upos=VERB]; Y[upos=NOUN|PROPN|PRON]; X-[comp:obj]->Y }
  without { X[Transitive = Yes] }
  commands { X.Transitive = Yes }
}
"""
grs = GRS(s)
corpus2bis = grs.apply(corpus)
corpus2bis.count(Request("pattern { X[Transitive=Yes] }"))

853

For the example, where the upos tag NOUN is changed to N, this can be done with a GRS:

from grewpy import GRS

grs3 = GRS("""
strat main { Onf(noun2n) }

rule noun2n {
  pattern { X[upos=NOUN] }
  commands { X.upos = N }
}
""")
corpus3bis = grs3.apply(corpus)
corpus3bis.count(Request("pattern { X[upos=N] }"))

4036

Similarily to the CorpusDraft above, there is a module GRSDraft which can be inspected and which is mutable.

from grewpy import GRSDraft

s = """
strat main {Onf(cxns)}
package cxns {
    rule existential {
        pattern {X-[comp@expl]->Y; X[lemma=be]}
        without {X[Cxn=Existential]}
        commands {X.Cxn=Existential}
    }
}
"""

grs_draft = GRSDraft(s)

for rule in grs_draft['cxns'].rules():
    print(f"{rule=}")

rule='existential'

A GRSDraft cannot be applied to a corpus, it should be turned into a GRS:

grs = GRS(grs_draft)
corpus.apply(grs)
n_existentials = corpus.count(Request("pattern { X[Cxn=Existential] }"))
print(f"{n_existentials=}")

n_existentials=26