def connect_graph_if_not_connected(graph):
try:
encoded = penman.encode(graph, model=amr_model)
return graph, ParsedStatus.OK
except:
pass
nxgraph = nx.MultiGraph()
variables = graph.variables()
for v1, _, v2 in graph.triples:
if v1 in variables and v2 in variables:
nxgraph.add_edge(v1, v2)
elif v1 in variables:
nxgraph.add_edge(v1, v1)
triples = graph.triples.copy()
new_triples = []
addition = f'a{len(variables) + 1}'
triples.append(penman.Triple(addition, ':instance', 'and'))
for i, conn_set in enumerate(nx.connected_components(nxgraph), start=1):
edge = f':op{i}'
conn_set = sorted(conn_set, key=lambda x: int(x[1:]))
conn_set = [c for c in conn_set if c in variables]
node = conn_set[0]
new_triples.append(penman.Triple(addition, edge, node))
triples = new_triples + triples
metadata = graph.metadata
graph = penman.Graph(triples)
graph.metadata.update(metadata)
penman.encode(graph, model=amr_model)
return graph, ParsedStatus.FIXED
def __str__(self):
if self.penman:
return penman.encode(self.penman)
else:
return legacy_graph_printer(self.get_metadata(), self.nodes,
self.root, self.edges)
def pgraph_to_gstring(pgraph):
pgraph = deepcopy(pgraph)
pgraph.metadata = {}
gstring = penman.encode(pgraph, indent=0)
gstring = gstring.replace('\n', ' ')
gstring = re.sub(' +', ' ', gstring)
return gstring
def test_for_decode_encode_issue(gold):
graph = penman.decode(gold, model=NoOpModel())
test = penman.encode(graph, indent=6, compact=True, model=NoOpModel())
gold = to_graph_line(gold)
test = to_graph_line(test)
is_good = test == gold
return graph, is_good
def _get_nodes_and_backreferences(self, graph):
graph_ = copy.deepcopy(graph)
graph_.metadata = {}
linearized = penman.encode(graph_)
linearized_nodes = self._tokenize_encoded_graph(linearized)
if self.use_pointer_tokens:
remap = {}
for i in range(1, len(linearized_nodes)):
nxt = linearized_nodes[i]
lst = linearized_nodes[i - 1]
if nxt == '/':
remap[lst] = f'<pointer:{len(remap)}>'
i = 1
linearized_nodes_ = [linearized_nodes[0]]
while i < (len(linearized_nodes)):
nxt = linearized_nodes[i]
lst = linearized_nodes_[-1]
if nxt in remap:
if lst == '(' and linearized_nodes[i + 1] == '/':
nxt = remap[nxt]
i += 1
elif lst.startswith(':'):
nxt = remap[nxt]
linearized_nodes_.append(nxt)
i += 1
linearized_nodes = linearized_nodes_
if self.remove_pars:
linearized_nodes = [n for n in linearized_nodes if n != '(']
backreferences = list(range(len(linearized_nodes)))
return linearized_nodes, backreferences
def to_format(self,
passage,
metadata=True,
wikification=True,
verbose=False,
use_original=True,
default_label=None,
**kwargs):
self.wikification = wikification
if use_original:
original = passage.extra.get("original")
if original:
return original
textutil.annotate(passage, as_array=True)
if self.wikification:
if verbose:
print("Wikifying passage...")
WIKIFIER.wikify_passage(passage)
if verbose:
print("Expanding names...")
self._expand_names(passage.layer(layer1.LAYER_ID))
triples = list(self._to_triples(
passage, default_label=default_label)) or [("y", INSTANCE, "yes")]
return (self.header(passage, **kwargs) if metadata else
[]) + (penman.encode(penman.Graph(triples)).split("\n"))
def __str__(self):
if self.penman:
return ' '.join(self.tokens) + '\n\n' + penman.encode(self.penman)
else:
return legacy_graph_printer(self.get_metadata(), self.nodes,
self.root, self.edges)
def toJAMRString(self):
"""
FIXME: Just modifies ::node line with respect to the original
"""
output = penman.encode(self.penman)
# Try first to just modify existing JAMR annotation
new_lines = []
modified = False
for line in output.split('\n'):
if line.startswith('# ::node'):
modified = True
items = line.split('\t')
node_id = items[1]
if node_id in self.alignments:
start = min(self.alignments[node_id])
dend = max(self.alignments[node_id]) + 1
if len(items) == 4:
items[-1] = f'{start}-{dend}'
elif len(items) == 3:
items.append(f'{start}-{dend}')
else:
raise Exception()
line = '\t'.join(items)
new_lines.append(line)
# if not we write it ourselves
if not modified:
from ipdb import set_trace
set_trace(context=30)
print()
return ('\n'.join(new_lines)) + '\n'
def to_format(self, passage, metadata=True, wikification=True):
textutil.annotate(passage)
lines = ["# ::id " + passage.ID,
"# ::tok " + " ".join(t.text for t in passage.layer(layer0.LAYER_ID).all)] if metadata else []
if wikification:
WIKIFIER.wikify_passage(passage)
self._expand_names(passage.layer(layer1.LAYER_ID))
return lines + [penman.encode(penman.Graph(list(self._to_triples(passage)))) or "(y / yes)"]
def build(self, concepts, relations):
self.concepts = concepts
self.relations = relations
triples = self.build_instance_triples() # add self.names
triples += self.build_edge_attrib_triples()
graph = penman.graph.Graph(triples)
string = penman.encode(graph, indent=6)
# Strip the uniqueness post tag (ie.. 2007@attr1@ -> 2007)
string = re.sub(r'@attr\d+@', '', string)
return string
def get_penman(self, return_type='object', indent=None):
if return_type == 'object':
return pm.decode(self.penman)
elif return_type == 'str':
if indent is None:
return self.penman
else:
return pm.encode(pm.decode(self.penman),
top=self.top,
indent=indent)
def main(args):
pattern = re.compile(r'''[\s()":/,\\'#]+''')
with open(args.input,
encoding='utf-8') as f, open(args.output,
mode='w',
encoding='utf-8') as out:
for amr_data in f.readlines():
if amr_data == '' or amr_data is None:
break
amr_data = json.loads(amr_data)
amr_nodes = amr_data.pop('nodes')
amr_edges = amr_data.pop('edges', [])
triples = []
concepts = []
for node in amr_nodes:
short_name = f'c{node["id"]}'
concept = node["label"]
if pattern.search(concept):
concept = f"\"{concept}\""
concepts.append(concept)
triples.append((short_name, 'instance', concept))
for attr, value in zip(node.get('properties', []),
node.get('values', [])):
if pattern.search(value):
value = f"\"{value}\""
triples.append((short_name, attr, value))
for edge in amr_edges:
src = f'c{edge["source"]}'
target = f'c{edge["target"]}'
label = edge["label"]
target = f"\"{target}\"" if pattern.search(target) else target
triples.append((src, label, target))
top = amr_data.pop('tops')[0]
# id = amr_data['id']
# snt = json.dumps(amr_data['input'])
# token = json.dumps(amr_data['token'])
# lemma = json.dumps(amr_data['lemma'])
# upos = json.dumps(amr_data['upos'])
# xpos = json.dumps(amr_data['xpos'])
# ner = json.dumps(amr_data['ner'])
for key, value in amr_data.items():
amr_data[key] = json.dumps(value, ensure_ascii=False)
graph = Graph(triples, top=f"c{top}", metadata=amr_data)
graph_en = pp.encode(graph)
graph_de = pp.decode(graph_en)
out.write(graph_en + '\n\n')
def link_to_graph(number, snt, string):
amr = get_amr_from_snt(snt=snt, amr_list=amr_list)
html_penman = penman.encode(amr.penman,
indent=4).replace("\"", "'")
if amr is not None:
return string.replace(
snt,
f'<a id="link_to_{str(number)}" value="{str(number)}">'
f'<input type="hidden" id="amr_hidden_link_hidden_to_{str(number)}" '
f'value="{html_penman}"/>'
f'{snt}'
f'</a>')
def build(self, concepts, relations):
self.concepts = concepts
self.relations = relations
self.used_arcs = defaultdict(
set) # keep track of edge names aready seen (key is source_id)
triples = self.build_instance_triples() # add self.names
triples += self.build_edge_attrib_triples()
graph = penman.graph.Graph(triples)
string = penman.encode(graph, indent=6)
# Strip the uniqueness post tag (ie.. 2007@attr1@ -> 2007)
string = re.sub(r'@attr\d+@', '', string)
return string
def tokenize_amr(self, graph):
if self.raw_graph:
graph_ = copy.deepcopy(graph)
graph_.metadata = {}
linearized = penman.encode(graph_)
linearized = re.sub(r"\s+", ' ', linearized)
bpe_tokens = [self.bos_token] + self._tokenize(linearized)[:1022]
bpe_token_ids = [self.encoder.get(b, self.unk_token_id) for b in bpe_tokens]
bpe_backreferences = list(range(len(bpe_token_ids)))
return bpe_tokens, bpe_token_ids, bpe_backreferences
else:
return super().tokenize_amr(graph)
def parse_sents(self,
sents,
add_metadata=True,
return_penman=False,
disable_progress=True,
pbar_desc=None):
assert isinstance(sents, list)
# Loop though batches
gen_graphs = []
dataloader = torch.utils.data.DataLoader(sents,
batch_size=self.batch_size,
shuffle=False)
pbar = tqdm(total=len(dataloader.dataset),
disable=disable_progress,
ncols=100,
desc=pbar_desc)
for batch in dataloader:
# I'm Ignoring potential for clipped sentences. If return_overflowing_tokens=True is passed into
# the the lower level call to batch_encode_plus(), I could get these back out if needed.
# Bart supports up to 1024 input tokens so this would have to be paragraphs of text, all
# concatenated into a single "sent", in order to overflow. This isn't valid for AMR anyway.
x, _ = self.tokenizer.batch_encode_sentences(batch,
device=self.device)
# model.config.max_length=20 is the base model. Set this much higher for generating AMR graphs.
with torch.no_grad():
model_out = self.model.generate(**x,
max_length=512,
num_beams=self.num_beams)
# re-encode the model output
assert len(model_out) == len(batch)
for tokk, sent in zip(model_out, batch):
graph, status, _ = self.tokenizer.decode_amr(
tokk.tolist(), restore_name_ops=self.restore_name_ops)
# Handle status errors (also has ParsedStatus.FIXED for fixed disconnected graphs)
if status == ParsedStatus.BACKOFF:
graph = self.invalid_graph
# In Penman 1.2.0, metadata does not impact penam.Graph.__eq__() so code checking for
# Inference.invalid_graph should still work, even if 'snt' metadata is different.
if add_metadata:
graph.metadata['snt'] = sent
gen_graphs.append(graph)
pbar.update(len(batch))
pbar.close()
# Return the penman graphs
if return_penman:
return gen_graphs
# The required behavior across all parse_mdoels, is to return graphs as strings by default
gstrings = [
penman.encode(g, indent=6, model=amr_model) for g in gen_graphs
]
return gstrings
def parse_spans(self, spans, add_metadata=True):
sio_f = io.StringIO()
for i, span in enumerate(spans):
sent = span.text
tokens = list(span)
entry = '# ::snt %s\n' % sent
entry += '(d / dummy)\n' # not-used but required for proper AMR file format
pen_graph = annotate_graph(entry, tokens)
amr_string = penman.encode(pen_graph)
sio_f.write(amr_string + '\n')
if i != len(spans)-1:
sio_f.write('\n')
sio_f.seek(0)
return self.parse_file_handle(sio_f, add_metadata)
def to_format(self, passage, metadata=True, wikification=True, verbose=False, use_original=True):
if use_original:
original = passage.extra.get("original")
if original:
return original
textutil.annotate(passage, as_array=True)
lines = self.header(passage) if metadata else []
if wikification:
if verbose:
print("Wikifying passage...")
WIKIFIER.wikify_passage(passage)
if verbose:
print("Expanding names...")
self._expand_names(passage.layer(layer1.LAYER_ID))
return lines + (penman.encode(penman.Graph(list(self._to_triples(passage)))).split("\n") or ["(y / yes)"])
def get_amr_line(input_f):
"""
Read the file containing AMRs. AMRs are separated by a blank line.
Each call of get_amr_line() returns the next available AMR (in one-line form).
Note: this function does not verify if the AMR is valid
"""
key, value = [], []
# read_amr = ''
regex1 = r'# ::tok (.+)'
regex = r'# ::snt (.+)'
cur_amr = []
has_content = False
sentence = ''
for line in input_f:
line = line.strip()
if line == "":
if not has_content:
# empty lines before current AMR
continue
else:
# end of current AMR
break
if line.strip().startswith("# ::snt "):
sentence = re.match(regex, line.strip()).group(1)
# updated
# ignore the comment line (starting with "#") in the AMR file
# continue
elif line.strip().startswith('# ::tok '):
tokens = re.match(regex1,
line.strip()).group(1).lower().split()
elif line.strip().startswith('# ::node'):
token = line.split('\t')
key.append(token[2].rstrip())
value.append(token[1])
# level[token[2].rstrip()] = token[1]
elif line.strip().startswith('#'):
continue
else:
has_content = True
# read_amr += line
cur_amr.append(line.strip())
if cur_amr:
g = penman.decode(' '.join(cur_amr))
amr_penam = penman.encode(g)
return "".join(cur_amr), sentence, key, value, amr_penam, tokens
else:
return '', '', '', '', '', ''
def split_multi_sentence(pgraph):
# Get the graph string and variable to concept dictionary
pgraph = deepcopy(pgraph)
gid = pgraph.metadata.get('id', 'none') # for logging
pgraph.metadata = {}
var2concept = {t.source: t.target for t in pgraph.instances()}
gstring = penman.encode(pgraph, indent=0)
# delete the multi-sentence line and any modifiers like (:li, :mode)
glines = gstring.split('\n')
assert glines[0].startswith('(m / multi-sentence')
glines = glines[1:]
while glines:
if glines[0].startswith(':') and not glines[0].startswith(':snt'):
glines = glines[1:]
else:
break
# rejoin the lines remove extra spaces and remove ending paren
gstring = ' '.join(glines)
gstring = re.sub(r' +', ' ', gstring).strip()
assert gstring.endswith(')')
gstring = gstring[:-1]
# Split on the :snt lines and separate each sentence to its own graph
gs_list = [gs.strip() for gs in re.split(':snt\d+', gstring)]
gs_list = [gs for gs in gs_list if gs]
# Convert the separated graphs to penman objects
pgraphs = []
for gidx, gstring in enumerate(gs_list):
try:
pgraph = penman.decode(gstring)
except penman.DecodeError:
logger.error('Error decoding %s %d\n%s' % (gid, gidx, gstring))
continue
# If a variable is not in this portion of the graph then penman will treat it like an
# attribute. In this case we need to add an instance for it. The way penman 1.1.0
# works, this will fix the graph.
missing_set = set(t.target for t in pgraph.attributes()
if re_attrib.match(t.target))
if missing_set:
logger.info('%s %d missing variables: %s' %
(gid, gidx, str(missing_set)))
# Add the variables and re-decode the graph
for var in missing_set:
concept = var2concept.get(var, None)
if concept is not None:
pgraph.triples.append((var, ':instance', concept))
pgraphs.append(pgraph)
return pgraphs
def parse_sents(self, sents, add_metadata=True):
assert isinstance(sents, list)
# Annotate the entry then compile it in a StringIO file-type object
# For Simplicity, convert the sentences into an in-memory AMR text file
# This could be simplified but the DataLoader is setup to load from an AMR file
# and this method will create an in-memory file-type object in the AMR format.
sio_f = io.StringIO()
for i, sent in enumerate(sents):
entry = '# ::snt %s\n' % sent
entry += '(d / dummy)\n' # not-used but required for proper AMR file format
pen_graph = annotate_graph(entry)
amr_string = penman.encode(pen_graph)
sio_f.write(amr_string + '\n')
if i != len(sents)-1:
sio_f.write('\n')
sio_f.seek(0)
return self.parse_file_handle(sio_f, add_metadata)
def __init__(self, **kwargs):
"""
:param kwargs:
- string
- triples
- top
- object
-copy
"""
if kwargs:
if len(kwargs) == 3 and ('string' and 'triples'
and 'top') in kwargs:
self.penman = kwargs['string']
self.list_triples = kwargs['graph']
self.top = kwargs['top']
elif len(kwargs) == 2 and ('triples' and 'top') in kwargs:
self.top = kwargs['top']
self.list_triples = kwargs['triples']
self.penman = triple_model_list_to_penman(self.list_triples,
top_id=self.top)
self.list_triples = self.delete_wiki()
elif len(kwargs) == 1 and 'object' in kwargs:
self.penman = pm.encode(kwargs['object'],
top=kwargs['object'].top)
self.list_triples, self.top = penman_to_model(kwargs['object'])
self.list_triples = self.delete_wiki()
elif len(kwargs) == 1 and 'copy' in kwargs:
_source_amr: AMRModel = kwargs['copy']
self.penman = _source_amr.get_penman(return_type='str')
self.list_triples = InstrumentedList(
[Triple(copy=t) for t in _source_amr.get_triples()])
self.list_triples = self.delete_wiki()
self.top = _source_amr.get_top()
self.penman = triple_model_list_to_penman(self.list_triples,
top_id=self.top)
def get_amr_line(input_f):
"""
Read the file containing AMRs. AMRs are separated by a blank line.
Each call of get_amr_line() returns the next available AMR (in one-line form).
Note: this function does not verify if the AMR is valid
"""
regex = r'# ::snt (.+)'
sentence = ''
cur_amr = []
has_content = False
for line in input_f:
line = line.strip()
if line == "":
if not has_content:
# empty lines before current AMR
continue
else:
# end of current AMR
break
if line.strip().startswith('# ::snt'):
sentence = re.match(regex, line.strip()).group(1)
if line.strip().startswith("#"):
# ignore the comment line (starting with "#") in the AMR file
continue
else:
has_content = True
cur_amr.append(line.strip())
if cur_amr:
g = penman.decode(' '.join(cur_amr))
amr_penman = penman.encode(g)
c = PENMANCodec()
t = c.parse(amr_penman)
l = layout.interpret(t)
value, key = t.positions(l, 0)
return "".join(cur_amr), sentence, key, value, amr_penman
else:
return '', '', '', '', ''
def encode(d, properties=True, lnk=True, indent=False):
"""
Serialize a DMRS object to a PENMAN string.
Args:
d: a DMRS object
properties (bool): if `False`, suppress variable properties
lnk: if `False`, suppress surface alignments and strings
indent (bool, int): if `True` or an integer value, add
newlines and indentation
Returns:
a PENMAN-serialization of the DMRS object
"""
if indent is True:
indent = -1
elif indent is False:
indent = None
triples = to_triples(d, properties=properties, lnk=lnk)
g = penman.Graph(triples)
try:
return penman.encode(g, indent=indent)
except penman.PenmanError as exc:
raise PyDelphinException('could not decode with Penman') from exc
def __init__(self, graph, force_annotate=False):
# Convert or copy the input graph to penman format
if isinstance(graph, str):
pgraph = penman.decode(graph, model=NoOpModel())
elif isinstance(graph, penman.graph.Graph):
pgraph = deepcopy(pgraph)
else:
raise ValueError('Code requires either a string a penman graph')
# Annotate if needed (aligner/tagging require annotation)
is_annotated = all([
key in pgraph.metadata for key in ('tokens', 'lemmas', 'pos_tags')
])
if not is_annotated or force_annotate:
sentence = pgraph.metadata[
'snt'] # Sanity check required tag. Throws KeyError if missing
pgraph = annotate_penman(pgraph)
self.annotation_performed = True # for unit-testing and debug
else:
self.annotation_performed = False
# Align the graph. For simplicity, always do this.
# If there are existing alignments they need to be removed.
# See https://penman.readthedocs.io/en/latest/api/penman.surface.html
if penman.surface.alignments(pgraph) or penman.surface.role_alignments(
pgraph):
for key, items in pgraph.epidata.items():
pgraph.epidata[key] = [
x for x in items
if not isinstance(x, penman.surface.AlignmentMarker)
]
pgraph = RBWAligner.from_penman_w_json(pgraph).get_penman_graph()
# get the graph string and pos tags for the tagger
self.metadata = pgraph.metadata.copy()
pos_tags = json.loads(self.metadata['pos_tags'])
pgraph.metadata = {}
gstring = penman.encode(pgraph, model=NoOpModel(), indent=6)
# Tag the graph string
self.gstring_tagged = self.tag(gstring, pos_tags)
def toJAMRString(self):
"""
FIXME: Just modifies ::node line with respect to the original
"""
output = penman.encode(self.penman)
new_lines = []
modified = False
for line in output.split('\n'):
if line.startswith('# ::node'):
modified = True
items = line.split('\t')
node_id = items[1]
start = min(self.alignments[node_id])
dend = max(self.alignments[node_id]) + 1
if len(items) == 4:
items[-1] = f'{start}-{dend}'
elif len(items) == 3:
items.append(f'{start}-{dend}')
else:
raise Exception()
line = '\t'.join(items)
new_lines.append(line)
assert modified
return ('\n'.join(new_lines)) + '\n'
exit()
# Print and generate
gnum = int(gnum)
print('Original:', sents[gnum])
print()
# Get the original graph as a penman object and add back in the sentence to
# metadata (stripped during loading)
pgraph = penman.decode(graphs[gnum])
pgraph.metadata['snt'] = sents[gnum]
# Loop through all variables and select appropriate candidates for the new top variable
candidate_tops = pgraph.variables()
candidate_tops.remove(pgraph.top)
# (optional) Remove nodes with incoming edges - significantly reduces the number of candidates
candidate_tops = [
v for v in candidate_tops if incoming_edge_count(v, pgraph) == 0
]
# Create the list to try, keeping the original top first
new_tops = [pgraph.top] + candidate_tops
new_graphs = [penman.encode(pgraph, top=t) for t in new_tops]
# Get the mapping from top variables to the concept for debug
var2concept = {t.source: t.target for t in pgraph.instances()}
# Generate
print('Generated (first is original top variable):')
gen_sents, _ = gtos.generate(new_graphs, disable_progress=True)
for sent, top in zip(gen_sents, new_tops):
print('top: (%s / %s)' % (top, var2concept[top]))
print(' ', sent)
print()
print('-' * 40)
print()
def get_graph_string(self):
return penman.encode(self.graph, model=NoOpModel(), indent=6)
def __str__(self):
return penman.encode(self)
def triple_model_list_to_penman(triple_model_list, top_id):
return pm.encode(pm.Graph(data=[(triple.source, triple.relation,
triple.target)
for triple in triple_model_list]),
top=top_id)
