ltg
/

ssa-perin / mtool /graph.py
larkkin's picture
Add code and readme
c45d283
# -*- coding: utf-8; -*-
# GraphaLogue Analyzer
# Marco Kuhlmann <marco.kuhlmann@liu.se>
# Stephan Oepen <oe@ifi.uio.no>
from datetime import datetime;
import html;
import operator;
from pathlib import Path;
import sys;
import score.core;
#
# default values on edge attributes, which will be removed in normalization.
# because all constants are normalized to lowercase strings prior to testing
# for default values, we need to deal in the normalized values here.
#
ATTRIBUTE_DEFAULTS = {"remote": "false",
"effective": "false", "member": "false"};
FLAVORS = {"dm": 0, "psd": 0, "ptg": 0,
"eds": 1, "ptg": 1, "ucca": 1,
"amr": 2, "drg": 2};
class Node(object):
def __init__(self, id, label = None, properties = None, values = None,
anchors = None, top = False, type = 1, anchorings = None):
self.id = id
self.type = type;
self.label = label;
self.properties = properties;
self.values = values;
self.anchorings = anchorings;
self.incoming_edges = set()
self.outgoing_edges = set()
self.anchors = anchors;
self.is_top = top
def set_property(self, name, value):
if self.properties and self.values:
try:
i = self.properties.index(name);
self.values[i] = value;
except ValueError:
self.properties.append(name);
self.values.append(value);
else:
self.properties = [name];
self.values = [value];
def set_anchoring(self, name, value):
#
# _fix_me_
# this (currently only used in the AMR overlay) will not work in the
# general case, where all three arrays should correspond in order.
# (22-jun-20; oe)
if self.properties and self.anchorings:
try:
i = self.properties.index(name);
self.anchorings[i] = value;
except ValueError:
self.properties.append(name);
self.anchorings.append(value);
else:
self.properties = [name];
self.anchorings = [value];
def add_anchor(self, anchor):
if anchor is not None:
if self.anchors is None: self.anchors = [anchor];
elif anchor not in self.anchors: self.anchors.append(anchor);
def is_root(self):
return len(self.incoming_edges) == 0
def is_leaf(self):
return len(self.outgoing_edges) == 0
def is_singleton(self):
return self.is_root() and self.is_leaf() and not self.is_top
def normalize(self, actions, input = None, trace = 0):
def union(anchors):
characters = set();
for anchor in anchors:
if "from" in anchor and "to" in anchor:
for i in range(anchor["from"], anchor["to"]):
characters.add(i);
result = [];
last = start = None;
for i in sorted(characters):
if start is None: start = i;
if last is None:
last = i;
continue;
elif i == last + 1 \
or all(c in score.core.SPACE for c in input[last:i]):
last = i;
continue;
else:
result.append({"from": start, "to": last + 1});
last = start = i;
if len(characters) > 0:
result.append({"from": start, "to": i + 1});
if anchors != result:
old = [anchor for anchor in anchors if anchor not in result];
new = [anchor for anchor in result if anchor not in anchors];
print("{} ==> {} [{}]".format(old, new, input),
file = sys.stderr);
return result;
def trim(anchor, input):
if "from" in anchor and "to" in anchor:
i = max(anchor["from"], 0);
j = min(anchor["to"], len(input));
while i < j and input[i] in score.core.PUNCTUATION: i += 1;
while j > i and input[j - 1] in score.core.PUNCTUATION: j -= 1;
if trace and (i != anchor["from"] or j != anchor["to"]):
print("{} ({}) --> <{}:{}> ({})"
"".format(anchor,
input[anchor["from"]:anchor["to"]],
i, j, input[i:j]),
file = sys.stderr);
anchor["from"] = i;
anchor["to"] = j;
if self.anchors is not None and "anchors" in actions:
self.anchors = union(self.anchors);
if self.anchors is not None and len(self.anchors) > 0 and input:
for anchor in self.anchors: trim(anchor, input);
elif isinstance(self.anchors, list) and len(self.anchors) == 0:
self.anchors = None;
if "case" in actions:
if self.label is not None:
self.label = str(self.label).lower();
if self.properties and self.values:
for i in range(len(self.properties)):
self.properties[i] = str(self.properties[i]).lower();
self.values[i] = str(self.values[i]).lower();
def compare(self, node):
#
# keep track of node-local pieces of information that either occur in
# both nodes (i.e. match), or only in the first or second of them. in
# guiding the MCES search, we (apparently) use the net gain of matching
# pieces /minus/ those not matching on either side. that does not lead
# to monotonicity, in the sense of cumulative scores moving either up
# or down as more node correspondences are fixed, but for guiding the
# MCES search monotonicity fortunately is not a requirement either.
#
count1 = both = count2 = 0;
if node is None:
if self.is_top:
count1 += 1;
if self.label is not None:
count1 += 1;
if self.properties is not None:
count1 += len(self.properties);
return both - count1 - count2, count1, both, count2;
if self.is_top:
if node.is_top: both += 1;
else: count1 += 1;
else:
if node.is_top: count2 += 1;
else: both += 1;
if self.label is not None:
if self.label == node.label:
both += 1;
else:
count1 += 1;
if node.label is not None: count2 += 1;
if self.properties is not None:
if node.properties is None:
count1 += len(self.properties);
else:
properties1 = {(property, self.values[i])
for i, property in enumerate(self.properties)};
properties2 = {(property, node.values[i])
for i, property in enumerate(node.properties)};
n = len(properties1 & properties2);
count1 += len(properties1) - n;
both += n;
count2 += len(properties2) - n;
elif node.properties is not None:
count2 += len(node.properties);
return both - count1 - count2, count1, both, count2;
def encode(self):
json = {"id": self.id};
if self.label:
json["label"] = self.label;
if self.properties and self.values or self.anchorings:
json["properties"] = self.properties;
if self.values:
json["values"] = self.values;
if self.anchorings:
json["anchorings"] = self.anchorings;
if self.anchors:
json["anchors"] = self.anchors;
return json;
@staticmethod
def decode(json):
id = json["id"]
label = json.get("label", None)
properties = json.get("properties", None)
values = json.get("values", None)
anchorings = json.get("anchorings", None)
anchors = json.get("anchors", None)
return Node(id=id, label=label, properties=properties, values=values, anchors=anchors, anchorings=anchorings)
def dot(self, stream, input = None, ids = False, strings = False,
errors = None, overlay = False):
shapes = ["square", "oval", "diamond", "triangle"];
if errors is not None and "correspondences" in errors:
correspondences = {g: s for g, s in errors["correspondences"]};
else:
correspondences = None;
missing = [None, [], [], None];
surplus = [None, [], [], None];
if errors is not None:
if "labels" in errors and "missing" in errors["labels"]:
for id, label in errors["labels"]["missing"]:
if id == self.id: missing[0] = label;
if "properties" in errors and "missing" in errors["properties"]:
for id, property, value in errors["properties"]["missing"]:
if id == self.id:
missing[1].append(property); missing[2].append(value);
if "anchors" in errors and "missing" in errors["anchors"]:
for id, anchor in errors["anchors"]["missing"]:
if id == self.id: missing[3] = anchor;
if correspondences is not None and self.id in correspondences:
key = correspondences[self.id];
if "labels" in errors and "surplus" in errors["labels"]:
for id, label in errors["labels"]["surplus"]:
if id == key: surplus[0] = label;
if "properties" in errors and "surplus" in errors["properties"]:
for id, property, value in errors["properties"]["surplus"]:
if id == key:
surplus[1].append(property); surplus[2].append(value);
if "anchors" in errors and "surplus" in errors["anchors"]:
for id, anchor in errors["anchors"]["surplus"]:
if id == key: surplus[3] = anchor;
if self.label \
or ids and not overlay \
or self.properties and self.values \
or self.anchors \
or missing[0] is not None or len(missing[1]) > 0 \
or missing[3] is not None \
or surplus[0] is not None or len(surplus[1]) > 0 \
or surplus[3] is not None:
if self.type in {0, 1, 2, 3}:
shape = "shape={}, ".format(shapes[self.type]);
else:
shape = "";
color = "color=blue, " if overlay else "";
print(" {} [ {}{}label=<<table align=\"center\" border=\"0\" cellspacing=\"0\">"
"".format(self.id, shape, color), end = "", file = stream);
if ids and not overlay:
print("<tr><td colspan=\"2\">#{}</td></tr>"
"".format(self.id), end = "", file = stream);
if self.label:
if missing[0]: font = "<font color=\"red\">";
elif overlay: font = "<font color=\"blue\">";
else: font = "<font>";
print("<tr><td colspan=\"2\">{}{}</font></td></tr>"
"".format(font, html.escape(self.label, False)),
end = "", file = stream);
if surplus[0]:
font = "<font color=\"blue\">";
print("<tr><td colspan=\"2\">{}{}</font></td></tr>"
"".format(font, html.escape(surplus[0], False)),
end = "", file = stream);
def __anchors__(anchors, color):
print("<tr><td colspan=\"2\"><font color=\"{}\">{{"
"".format(color), end = "", file = stream);
for index in anchors:
print("{}{}".format("&thinsp;" if index != anchors[0] else "", index),
end = "", file = stream);
print("}</font></td></tr>", end = "", file = stream);
if self.anchors is not None:
if overlay:
__anchors__(self.anchors, "blue");
else:
print("<tr><td colspan=\"2\">", end = "", file = stream);
for anchor in self.anchors:
if strings and input:
print("{}<font face=\"Courier\">{}</font>"
"".format(",&nbsp;" if anchor != self.anchors[0] else "",
html.escape(input[anchor["from"]:anchor["to"]])),
end = "", file = stream);
else:
print("{}〈{}:{}〉"
"".format("&thinsp;" if anchor != self.anchors[0] else "",
anchor["from"], anchor["to"]),
end = "", file = stream);
print("</td></tr>", end = "", file = stream);
if missing[3]: __anchors__(missing[3], "red");
if surplus[3]: __anchors__(surplus[3], "blue");
def __properties__(names, values, color):
font = "<font color=\"{}\">".format(color);
for name, value in zip(names, values):
print("<tr><td sides=\"l\" border=\"1\" align=\"left\">{}{}</font>"
"</td><td sides=\"r\" border=\"1\" align=\"left\">{}{}</font></td></tr>"
"".format(font, html.escape(name, False),
font, html.escape(value), False),
end = "", file = stream);
if self.properties and self.values:
if not overlay:
for name, value in zip(self.properties, self.values):
i = None;
try:
i = missing[1].index(name);
except:
pass;
if i is None or missing[2][i] != value:
__properties__([name], [value], "black");
else:
__properties__(self.properties, self.values, "blue");
if len(missing[1]) > 0: __properties__(missing[1], missing[2], "red");
if len(surplus[1]) > 0: __properties__(surplus[1], surplus[2], "blue");
print("</table>> ];", file = stream);
elif overlay is None or self.id < 0:
shape = "{}, label=\" \"".format(shapes[0]) if self.type == 0 else "point";
print(" {} [ shape={}, width=0.2 ];"
"".format(self.id, shape), file = stream);
def __key(self):
return self.id
def __eq__(self, other):
return self.__key() == other.__key()
def __lt__(self, other):
return self.__key() < other.__key()
def __hash__(self):
return hash(self.__key())
class Edge(object):
def __init__(self, id, src, tgt, lab, normal = None,
attributes = None, values = None, anchors = None):
self.id = id;
self.src = src;
self.tgt = tgt;
self.lab = lab;
self.normal = normal;
self.attributes = attributes;
self.values = values;
self.anchors = anchors;
def is_loop(self):
return self.src == self.tgt
def min(self):
return min(self.src, self.tgt)
def max(self):
return max(self.src, self.tgt)
def endpoints(self):
return self.min(), self.max()
def length(self):
return self.max() - self.min()
def normalize(self, actions, trace = 0):
if "edges" in actions:
if self.normal is None \
and self.lab is not None:
label = self.lab;
if label == "mod":
self.normal = "domain";
elif label.endswith("-of-of") \
or label.endswith("-of") \
and label not in {"consist-of" "subset-of"} \
and not label.startswith("prep-"):
self.normal = label[:-3];
if self.normal:
target = self.src;
self.src = self.tgt;
self.tgt = target;
self.lab = self.normal;
self.normal = None;
if "case" in actions:
if self.lab is not None:
self.lab = str(self.lab).lower();
if self.normal is not None:
self.normal = str(self.normal).lower();
if self.attributes and self.values:
for i in range(len(self.attributes)):
self.attributes[i] = str(self.attributes[i]).lower();
self.values[i] = str(self.values[i]).lower();
if "attributes" in actions and self.attributes and self.values:
#
# drop (attribute, value) pairs whose value is the default value
#
attribute_value_pairs = [
(attribute, value) for attribute, value
in zip(self.attributes, self.values)
if attribute not in ATTRIBUTE_DEFAULTS
or ATTRIBUTE_DEFAULTS[attribute] != value]
self.attributes, self.values \
= tuple(map(list, zip(*attribute_value_pairs))) or ([], [])
def encode(self):
json = {"id": self.id};
if self.src is not None: json["source"] = self.src;
if self.tgt is not None: json["target"] = self.tgt;
if self.lab: json["label"] = self.lab;
if self.normal: json["normal"] = self.normal;
if self.attributes and self.values:
json["attributes"] = self.attributes;
json["values"] = self.values;
if self.anchors: json["anchors"] = self.anchors;
return json;
@staticmethod
def decode(json):
id = json.get("id", None);
src = json.get("source", None);
tgt = json.get("target", None);
lab = json.get("label", None);
if lab == "": lab = None;
normal = json.get("normal", None)
attributes = json.get("attributes", None)
if attributes is None:
attributes = json.get("properties", None)
if attributes is not None:
print("Edge.decode(): "
"interpreting deprecated ‘properties’ on edge object.",
file = sys.stderr);
values = json.get("values", None)
anchors = json.get("anchors", None)
return Edge(id, src, tgt, lab, normal, attributes, values, anchors)
def dot(self, stream, input = None, strings = False,
errors = None, overlay = False):
def __missing__():
if errors is not None and "edges" in errors \
and "missing" in errors["edges"]:
for source, target, label in errors["edges"]["missing"]:
if source == self.src and target == self.tgt and label == self.lab:
return True;
return False;
if self.attributes and self.values:
style = ", style=dashed";
label = "<<table align=\"center\" border=\"0\" cellspacing=\"0\">";
if self.lab: label += "<tr><td colspan=\"1\">{}</td></tr>".format(self.lab);
#
# _fix_me_
# currently assuming that all values are boolean where presence of
# the attribute means True. (oe; 21-apr-20)
#
if self.attributes and self.values:
for attribute, _ in zip(self.attributes, self.values):
label += "<tr><td>{}</td></tr>".format(attribute);
label += "</table>>";
else:
label = self.lab;
if label and self.normal:
if label[:-3] == self.normal:
label = "(" + self.normal + ")-of";
else:
label = label + " (" + self.normal + ")";
if label: label = "\"{}\"".format(label);
style = "";
if overlay:
color = ", color=blue, fontcolor=blue";
elif __missing__():
color = ", color=red, fontcolor=red";
else:
color = "";
print(" {} -> {} [ label={}{}{} ];"
"".format(self.src, self.tgt, label if label else "\"\"",
style, color),
file = stream);
def __key(self):
return self.tgt, self.src, self.lab
def __eq__(self, other):
return self.__key() == other.__key()
def __lt__(self, other):
return self.__key() < other.__key()
def __hash__(self):
return hash(self.__key())
class Graph(object):
def __init__(self, id, flavor = None, framework = None):
self.id = id;
self.time = datetime.utcnow();
self._language = None;
self._provenance = None;
self._source = None;
self._targets = None;
self.input = None;
self.nodes = [];
self.edges = set();
self.flavor = FLAVORS.get(framework) if flavor is None else flavor;
self.framework = framework;
def language(self, value = None):
if value is not None: self._language = value;
return self._language;
def provenance(self, value = None):
if value is not None: self._provenance = value;
return self._provenance;
def source(self, value = None):
if value is not None: self._source = value;
return self._source;
def targets(self, value = None):
if value is not None: self._targets = value;
return self._targets;
def size(self):
return len(self.nodes);
def inject(self, information):
if isinstance(information, str): information = eval(information);
for key, value in information.items():
if key == "id": self.id = value;
elif key == "time": self.item = value;
elif key == "language": self._language = value;
elif key == "provenance": self._provenance = value;
elif key == "source": self._source = value;
elif key == "targets": self._targets = value;
elif key == "input": self.input = value;
elif key == "flavor": self.flavor = value;
elif key == "framework": self.framework = value;
else:
print("Graph.inject(): ignoring invalid key ‘{}’"
"".format(key), file = sys.stderr);
def add_node(self, id = None, label = None,
properties = None, values = None,
anchors = None, top = False, type = 1, anchorings = None):
node = Node(id if id is not None else len(self.nodes),
label = label, properties = properties, values = values,
anchors = anchors, top = top, type = type,
anchorings = anchorings);
self.nodes.append(node)
return node
def find_node(self, id):
for node in self.nodes:
if node.id == id: return node;
def add_edge(self, src, tgt, lab, normal = None,
attributes = None, values = None, anchors = None):
self.store_edge(Edge(id=len(self.edges), src=src, tgt=tgt, lab=lab, normal=normal,
attributes=attributes, values=values, anchors=anchors));
def store_edge(self, edge, robust = False):
self.edges.add(edge)
source = self.find_node(edge.src);
if source is None and not robust:
raise ValueError("Graph.add_edge(): graph #{}: "
"invalid source node {}."
"".format(self.id, self.src))
if source: source.outgoing_edges.add(edge)
target = self.find_node(edge.tgt);
if target is None and not robust:
raise ValueError("Graph.add_edge(): graph #{}: "
"invalid target node {}."
"".format(self.id, self.tgt))
if target: target.incoming_edges.add(edge)
return edge
def add_input(self, text, id = None, quiet = False):
if not id: id = self.id;
if isinstance(text, str):
self.input = text;
elif isinstance(text, Path):
file = text / (str(id) + ".txt");
if not file.exists() and not quiet:
print("add_input(): no text for {}.".format(file),
file = sys.stderr);
else:
with file.open() as stream:
input = stream.readline();
if input.endswith("\n"): input = input[:len(input) - 1];
self.input = input;
else:
input = text.get(id);
if input:
self.input = input;
elif not quiet:
print("add_input(): no text for key {}.".format(id),
file = sys.stderr);
def anchor(self):
n = len(self.input);
i = 0;
def skip():
nonlocal i;
while i < n and self.input[i] in {" ", "\t"}:
i += 1;
def scan(candidates):
for candidate in candidates:
if self.input.startswith(candidate, i):
return len(candidate);
skip();
for node in self.nodes:
for j in range(len(node.anchors) if node.anchors else 0):
if isinstance(node.anchors[j], str):
form = node.anchors[j];
m = None;
if self.input.startswith(form, i):
m = len(form);
else:
for old, new in {("‘", "`"), ("’", "'"), ("`", "'"),
("“", "\""), ("”", "\"")}:
form = form.replace(old, new);
if self.input.startswith(form, i):
m = len(form);
break;
#
# _fix_me_
# the block below looks weird: it would seem to accept any
# of the punctuation marks given to scan(), irrespective
# of the current .form. value? (oe; 27-apr-20)
#
if not m:
m = scan({"“", "\"", "``"}) or scan({"‘", "`"}) \
or scan({"”", "\"", "''"}) or scan({"’", "'"}) \
or scan({"—", "—", "---", "--"}) \
or scan({"…", "...", ". . ."});
if m:
node.anchors[j] = {"from": i, "to": i + m};
i += m;
skip();
else:
raise Exception("failed to anchor |{}| in |{}| ({})"
"".format(form, self.input, i));
def normalize(self, actions, trace = 0):
for node in self.nodes:
node.normalize(actions, self.input, trace);
for edge in self.edges:
edge.normalize(actions, trace);
#
# recompute cached edge relations, to reflect the new state of affairs
#
if "edges" in actions:
for node in self.nodes:
node.outgoing_edges.clear();
node.incoming_edges.clear();
for edge in self.edges:
self.find_node(edge.src).outgoing_edges.add(edge);
self.find_node(edge.tgt).incoming_edges.add(edge);
def prettify(self, trace = 0):
if self.framework == "drg":
boxes = {"IMP", "DIS", "DUP", "NOT", "POS", "NEC",
"ALTERNATION", "ATTRIBUTION", "BACKGROUND",
"COMMENTARY", "CONDITION", "CONTINUATION", "CONTRAST",
"CONSEQUENCE", "ELABORATION", "EXPLANATION", "INSTANCE",
"NARRATION", "NEGATION", "NECESSITY",
"POSSIBILITY", "PARALLEL", "PRECONDITION",
"RESULT", "TOPIC", "PRESUPPOSITION"};
for node in self.nodes:
if node.is_top or node.is_root():
node.type = 0;
#
# _fix_me_
# but what about more deeply nested boxes? (24-aug-20; oe)
#
for edge in node.outgoing_edges:
if edge.lab in boxes:
self.find_node(edge.tgt).type = 0;
elif len(node.incoming_edges) == len(node.outgoing_edges) == 1:
if next(iter(node.incoming_edges)).lab is None \
and next(iter(node.outgoing_edges)).lab is None:
node.type = 2;
def score(self, graph, correspondences, errors = None):
#
# accommodate the various conventions for node correspondence matrices;
# anyway, entries are indices into the .nodes. list, not identifiers.
# _fix_me_
# double-check for correspondences from SMATCH. (oe; 19-apr-20)
#
if isinstance(correspondences, list) and len(correspondences) > 0:
if isinstance(correspondences[0], tuple):
correspondences = {i: j if j is not None else -1
for i, j in correspondences};
elif isinstance(correspondences[0], int):
correspondences = {i: j if j is not None else -1
for i, j in enumerate(correspondences)};
#
# all tuples use node identifiers from the gold graph, where there is
# a correspondence; otherwise we (appear to) synthesize new unique
# identifiers for remaining nodes from both graphs.
#
identities1 = dict();
identities2 = dict();
for i, pair in enumerate(correspondences.items()):
identities1[self.nodes[pair[0]].id] = i;
if pair[1] >= 0:
identities2[graph.nodes[pair[1]].id] = i;
i = len(correspondences);
for node in self.nodes:
if node.id not in identities1:
identities1[node.id] = i;
i += 1;
for node in graph.nodes:
if node.id not in identities2:
identities2[node.id] = i;
i += 1;
#
# map 'corresponding' identifiers back to the original graphs
#
def native(id, identities):
for key, value in identities.items():
if id == value: return key;
def tuples(graph, identities):
#
# .identities. is a hash table mapping node identifiers into the
# 'corresponding' identifier space, such that paired nodes (and
# only these) share the same identifier.
#
def identify(id):
return identities[id] if identities is not None else id;
tops = set();
labels = set();
properties = set();
anchors = set();
edges = set();
attributes = set();
for node in graph.nodes:
identity = identify(node.id);
if node.is_top: tops.add(identity);
if node.label is not None: labels.add((identity, node.label));
if node.properties is not None:
for property, value in zip(node.properties, node.values):
properties.add((identity, property, value.lower()));
if node.anchors is not None:
anchor = score.core.anchor(node);
if graph.input:
anchor = score.core.explode(graph.input, anchor);
else:
anchor = tuple(anchor);
anchors.add((identity, anchor));
for edge in graph.edges:
identity \
= (identify(edge.src), identify(edge.tgt), edge.lab);
edges.add(identity);
if edge.attributes and edge.values:
for attribute, value in zip(edge.attributes, edge.values):
attributes.add(tuple(list(identity) + [attribute, value]));
return tops, labels, properties, anchors, edges, attributes;
def count(gold, system, key):
if errors is not None:
missing = gold - system;
surplus = system - gold;
if len(missing) > 0 or len(surplus) > 0 and key not in errors:
errors[key] = dict();
if key == "tops":
if missing:
errors[key]["missing"] \
= [native(id, identities1) for id in missing];
if surplus:
errors[key]["surplus"] \
= [native(id, identities2) for id in surplus];
elif key == "labels":
if missing:
errors[key]["missing"] \
= [(native(id, identities1), label)
for id, label in missing];
if surplus:
errors[key]["surplus"] \
= [(native(id, identities2), label)
for id, label in surplus];
elif key == "properties":
if missing:
errors[key]["missing"] \
= [(native(id, identities1), property, value)
for id, property,value in missing];
if surplus:
errors[key]["surplus"] \
= [(native(id, identities2), property, value)
for id, property, value in surplus];
elif key == "anchors":
if missing:
errors[key]["missing"] \
= [(native(id, identities1), list(sorted(anchor)))
for id, anchor in missing];
if surplus:
errors[key]["surplus"] \
= [(native(id, identities2), list(sorted(anchor)))
for id, anchor in surplus];
elif key == "edges":
if missing:
errors[key]["missing"] \
= [(native(source, identities1),
native(target, identities1), label)
for source, target, label in missing];
if surplus:
errors[key]["surplus"] \
= [(native(source, identities2),
native(target, identities2), label)
for source, target, label in surplus];
elif key == "attributes":
if missing:
errors[key]["missing"] \
= [(native(source, identities1),
native(target, identities1), label,
attribute, value)
for source, target, label, attribute, value
in missing];
if surplus:
errors[key]["surplus"] \
= [(native(source, identities2),
native(target, identities2), label,
attribute, value)
for source, target, label, attribute, value
in surplus];
return {"g": len(gold), "s": len(system), "c": len(gold & system)};
if correspondences is None or len(correspondences) == 0:
return count(set(), set()), count(set(), set()), \
count(set(), set()), count(set(), set()), \
count(set(), set()), count(set(), set());
gtops, glabels, gproperties, ganchors, gedges, gattributes \
= tuples(self, identities1);
stops, slabels, sproperties, sanchors, sedges, sattributes \
= tuples(graph, identities2);
if errors is not None:
errors[self.framework][self.id] = errors \
= {"correspondences": [(self.nodes[g].id, graph.nodes[s].id)
for g, s in correspondences.items()
if s >= 0]}
return count(gtops, stops, "tops"), \
count(glabels, slabels, "labels"), \
count(gproperties, sproperties, "properties"), \
count(ganchors, sanchors, "anchors"), \
count(gedges, sedges, "edges"), \
count(gattributes, sattributes, "attributes");
def encode(self, version = 1.1):
json = {"id": self.id};
if self.flavor is not None:
json["flavor"] = self.flavor;
if self.framework:
json["framework"] = self.framework;
json["version"] = version;
if self.time is not None:
json["time"] = self.time.strftime("%Y-%m-%d");
else:
json["time"] = datetime.now().strftime("%Y-%m-%d");
if self._language is not None: json["language"] = self._language;
if self._source is not None: json["source"] = self._source;
if self._provenance is not None: json["provenance"] = self._provenance;
if self._targets is not None: json["targets"] = self._targets;
if self.input:
json["input"] = self.input;
if self.nodes:
tops = [node.id for node in self.nodes if node.is_top];
if len(tops):
json["tops"] = tops;
json["nodes"] = [node.encode() for node in self.nodes];
if self.edges:
json["edges"] = [edge.encode() for edge in
sorted(self.edges, key = operator.attrgetter("id"))];
return json;
@staticmethod
def decode(json, robust = False):
graph = Graph(json["id"], json.get("flavor"), json.get("framework"))
try:
graph.time = datetime.strptime(json["time"], "%Y-%m-%d")
except:
graph.time = datetime.strptime(json["time"], "%Y-%m-%d (%H:%M)")
graph.input = json.get("input")
graph.language(json.get("language"))
graph.source(json.get("source"))
graph.provenance(json.get("provenance"))
graph.targets(json.get("targets"))
nodes = json.get("nodes")
if nodes is not None:
for j in nodes:
node = Node.decode(j)
graph.add_node(node.id, node.label, node.properties,
node.values, node.anchors, top = False, anchorings=node.anchorings)
edges = json.get("edges")
if edges is not None:
for j in edges:
edge = Edge.decode(j);
if edge.id is None: edge.id = len(graph.edges);
graph.store_edge(edge, robust = robust);
tops = json.get("tops")
if tops is not None:
for i in tops:
node = graph.find_node(i)
if node is not None:
node.is_top = True
else:
raise ValueError("Graph.decode(): graph #{}: "
"invalid top node {}."
"".format(graph.id, i))
return graph
def copy(self):
return Graph.decode(self.encode())
def dot(self, stream, ids = False, strings = False,
errors = None, overlay = False):
if not overlay:
print("digraph \"{}\" {{\n top [ style=invis ];"
"".format(self.id),
file = stream);
for node in self.nodes:
if node.is_top:
if overlay:
color = " [ color=blue ]";
elif errors is not None and "tops" in errors \
and "missing" in errors["tops"] and node.id in errors["tops"]["missing"]:
color = " [ color=red ]";
else:
color = "";
print(" top -> {}{};".format(node.id, color), file = stream);
n = -1;
for node in self.nodes:
node.dot(stream, self.input, ids, strings, errors, overlay);
for edge in self.edges:
if node.id == edge.src:
edge.dot(stream, self.input, strings, errors, overlay);
if errors is not None:
surplus = Graph(self.id, flavor = self.flavor, framework = self.framework);
surplus.add_input(self.input);
mapping = dict();
correspondences = {s: g for g, s in errors["correspondences"]};
if "labels" in errors and "surplus" in errors["labels"]:
for id, label in errors["labels"]["surplus"]:
if id not in correspondences:
mapping[id] = surplus.add_node(id = n, label = label);
n -= 1;
if "properties" in errors and "surplus" in errors["properties"]:
for id, property, value in errors["properties"]["surplus"]:
if id not in correspondences:
if id in mapping:
mapping[id].set_property(property, value);
else:
mapping[id] = surplus.add_node(id = n,
properties = [property],
values = [value]);
n -= 1;
if "anchors" in errors and "surplus" in errors["anchors"]:
for id, anchor in errors["anchors"]["surplus"]:
if id not in correspondences:
if id in mapping:
mapping[id].anchors = anchor;
else:
mapping[id] = surplus.add_node(id = n, anchors = anchor);
n -= 1;
if "tops" in errors and "surplus" in errors["tops"]:
for id in errors["tops"]["surplus"]:
if id in correspondences:
print(" top -> {} [ color=blue ];"
"".format(correspondences[id]), file = stream);
elif id not in mapping:
mapping[id] = surplus.add_node(id = n, top = True);
n -= 1;
else:
mapping[id].is_root = True;
if "edges" in errors and "surplus" in errors["edges"]:
for source, target, label in errors["edges"]["surplus"]:
if source not in mapping:
try:
mapping[source] = surplus.add_node(correspondences[source]);
except KeyError:
mapping[source] = surplus.add_node(n);
n -= 1;
if target not in mapping:
try:
mapping[target] = surplus.add_node(correspondences[target]);
except KeyError:
mapping[target] = surplus.add_node(n);
n -= 1;
surplus.add_edge(mapping[source].id, mapping[target].id, label);
surplus.dot(stream, ids = ids, strings = strings, errors = None, overlay = True);
if not overlay: print("}", file = stream);
def tikz(self, stream):
if self.flavor != 0: # bi-lexical: use tikz-dependency
raise ValueError("TikZ visualization is currently only for flavor-0 graphs.")
graph = self._full_sentence_recovery() # a copy of self with nodes covering all tokens
print(r"\documentclass{article}", file=stream)
print(r"\usepackage[T1]{fontenc}", file=stream)
print(r"\usepackage[utf8]{inputenc}", file=stream)
print(r"\usepackage{tikz-dependency}", file=stream)
print(r"\begin{document}", file=stream)
print(r"\begin{dependency}", file=stream)
print(r"\begin{deptext}", file=stream)
print(r"% id = " + str(graph.id), file=stream)
if graph.input is not None:
print(r"% input = " + str(graph.input), file=stream)
sorted_nodes = sorted((node.id, node) for node in graph.nodes)
id2i = {id: i for i, (id, _) in enumerate(sorted_nodes, start=1)}
print(r" \& ".join(" ".join(graph.input[anchor["from"]:anchor["to"]] for anchor in node.anchors or ())
or node.label for _, node in sorted_nodes) + r" \\", file=stream)
print(r"\end{deptext}", file=stream)
for id, node in sorted_nodes:
if node.is_top:
print(r"\deproot{" + str(id2i[id]) + r"}{TOP}", file=stream)
for edge in graph.edges:
if node.id == edge.tgt:
print(r"\depedge{" + str(id2i[edge.src]) + r"}{" + str(id2i[id]) + r"}{" + str(edge.lab) + r"}", file=stream)
print(r"\end{dependency}", file=stream)
print(r"\end{document}", file=stream)
def displacy(self, stream=None, format="svg", **kwargs):
"""
Use displacy to present dependency graph over sentence.
:param format: can be either "svg" or "html".
kwargs are passed to displacy.render method, see https://spacy.io/usage/visualizers
for possible options.
One can omit the stream argument if specifying `jupyter=True` - this will render the visualization directly
to the jupyter notebook.
"""
assert stream or kwargs.get("jupyter"), "Either `stream` is given or `jupyter=True` must hold."
assert format in ("svg", "html"), 'format can be either "svg" or "html"'
try:
from spacy import displacy
except ModuleNotFoundError as e:
print("You must install SpaCy in order to use the displacy visualization. \nTry running `pip install spacy`.")
raise e
if self.flavor != 0: # currently supporting only bi-lexical graphs
raise ValueError("displacy visualization is currently only for flavor-0 graphs.")
graph = self._full_sentence_recovery() # a copy of self with nodes covering all tokens
# prepare displacy_dep_input, composed of `words` list and `arcs` list
words = [{"text": n.label, "tag": ""} for n in graph.nodes]
def get_arc(edge: Edge):
src, tgt = edge.src, edge.tgt
direction = u'right' if src < tgt else u'left'
return {'dir': direction,
'start': min(src, tgt),
'end': max(src, tgt),
'label': edge.lab}
arcs = [get_arc(edge) for edge in graph.edges]
displacy_dep_input = {'words': words, 'arcs': arcs}
# render to stream as svg or html
kwargs["page"] = format=="html"
markdown = displacy.render(displacy_dep_input, style='dep', manual=True, **kwargs)
# write svg text to a file
if stream:
stream.write(markdown)
def _full_sentence_recovery(self):
"""
graph nodes may sometimes only include non-singleton nodes, for example when taking the graph from
a model prediction. For this reason, we need to use anchors and the input sentence in order to recover
the original tokenization (thus node-ids and their corresponding text spans).
Here, when necessary, we assume the original tokenization is encoded with spaces in self.input.
But we mainly look for missing character segments (i.e. spans that are not included in anchors)
and produce singleton nodes for them.
The function returns a new Graph, in which recovered nodes are included and thus nodes correspond to
input tokens.
"""
graph = self.copy() # don't change
length = len(graph.input)
def rm_all(lst, items_to_remove):
for item in items_to_remove:
if item in lst:
lst.remove(item)
return lst
def group_consecutive(lst):
# get list of integers, return list of lists, each the maximal consecutive (increasing) set from lst
if not lst:
return []
groups = []
cur_group=[lst[0]]
for i,item in enumerate(lst[1:]):
if item-1 == cur_group[-1]:
cur_group.append(item)
else:
groups.append(cur_group)
cur_group = [item]
groups.append(cur_group)
return groups
# iterate missing ids
node_ids = [n.id for n in graph.nodes]
id2node = {n.id : n for n in graph.nodes}
max_id = max(node_ids)
missing_ids = rm_all(list(range(max_id)), node_ids)
missing_id_groups = group_consecutive(missing_ids)
for id_group in missing_id_groups:
# id_group is a list of consecutive missing ids
if id_group[0]==0:
begin_char = 0
else:
prev_id = id_group[0]-1 # the id of the existing node preceding the missing-id group
prev_node = id2node[prev_id]
begin_char = prev_node.anchors[0]['to']
next_id = id_group[-1]+1
if next_id in id2node:
next_node = id2node[next_id]
end_char = next_node.anchors[0]['from']
else:
end_char = length
omitted_span = graph.input[begin_char:end_char]
# we need to create len(id_group) new nodes for the omitted span.
# Try to align singleton node (i.e. one id) to a token; if num of tokens in omitted_span
# don't match num of missing ids, generate all these nodes with the same anchors to the whole span
tokens = omitted_span.strip().split()
if len(tokens) == len(id_group):
for token, new_id in zip(tokens, id_group):
tok_begin_char = begin_char + omitted_span.find(token)
tok_end_char = tok_begin_char + len(token)
# add new node corresponding to omitted token
graph.add_node(new_id, label=token, anchors=[{"from":tok_begin_char, "to":tok_end_char}])
else:
# add new nodes, all corresponding to omitted span
for new_id in id_group:
graph.add_node(new_id, label=omitted_span, anchors=[{"from": begin_char, "to": end_char}])
# special treatment is required for missing tokens after the last existing node
# (if there are tokens left in self.input not covered by node anchors)
last_end_char_of_nodes = max([n.anchors[0]['to'] for n in graph.nodes])
if last_end_char_of_nodes < length:
# the meaning is that there is some span of the sentence not covered;
# we will add nodes according to num of tokens in this last span
omitted_span = graph.input[last_end_char_of_nodes:]
for i,token in enumerate(omitted_span.strip().split()):
new_id = max_id+1+i
tok_begin_char = last_end_char_of_nodes + omitted_span.find(token)
tok_end_char = tok_begin_char + len(token)
graph.add_node(new_id, label=token, anchors=[{"from":tok_begin_char, "to":tok_end_char}])
# as a finish, sort nodes in graph so that they will again be ordered by id (& realization location)
graph.nodes = list(sorted(graph.nodes))
return graph