def coordPath(this,target):
if depG.has_node(this) and depG.has_node(target):
paths1 = [NX.shortest_path(depG,x,this)
for x in parentTokens
if depG.has_node(x)]
paths2 = [NX.shortest_path(depG,x,target)
for x in parentTokens
if depG.has_node(x)]
# same coordination group if there is a pair of
# shortest paths from event that start with the same edge
# and this edge is not between the tokens of the event
start1 = set( [depG[p[0]][p[1]]['type'] for p in paths1
if (p and
len(p)>=2 and
not p[1] in parentTokens)] )
start2 = set( [depG[p[0]][p[1]]['type'] for p in paths2
if (p and
len(p)>=2 and
not p[1] in parentTokens)] )
return(start1.intersection(start2))
return(False)
def coordPath(this, target):
if depG.has_node(this) and depG.has_node(target):
paths1 = [
NX.shortest_path(depG, x, this) for x in parentTokens
if depG.has_node(x)
]
paths2 = [
NX.shortest_path(depG, x, target) for x in parentTokens
if depG.has_node(x)
]
# same coordination group if there is a pair of
# shortest paths from event that start with the same edge
# and this edge is not between the tokens of the event
start1 = set([
depG[p[0]][p[1]]['type'] for p in paths1
if (p and len(p) >= 2 and not p[1] in parentTokens)
])
start2 = set([
depG[p[0]][p[1]]['type'] for p in paths2
if (p and len(p) >= 2 and not p[1] in parentTokens)
])
return (start1.intersection(start2))
return (False)
def analyse(self):
"""
Analyses the parse, determines the unflattening strategy
for each event, and modifies the cache accordingly.
"""
def getCoordGrouping(edges):
"""
Workhorse function for determining the grouping of arguments.
@type edges: list of 3-tuples
@param edges: out-going edges of an event to be grouped
@rtype: list of lists of 3-tuples
@return: grouped edges
"""
def coordPath(this, target):
if depG.has_node(this) and depG.has_node(target):
paths1 = [
NX.shortest_path(depG, x, this) for x in parentTokens
if depG.has_node(x)
]
paths2 = [
NX.shortest_path(depG, x, target) for x in parentTokens
if depG.has_node(x)
]
# same coordination group if there is a pair of
# shortest paths from event that start with the same edge
# and this edge is not between the tokens of the event
start1 = set([
depG[p[0]][p[1]]['type'] for p in paths1
if (p and len(p) >= 2 and not p[1] in parentTokens)
])
start2 = set([
depG[p[0]][p[1]]['type'] for p in paths2
if (p and len(p) >= 2 and not p[1] in parentTokens)
])
return (start1.intersection(start2))
return (False)
def connected(e1, e2):
# do not continue if not within sentence
if self.mapping.has_key(e1):
for t1 in self.mapping[e1]:
# do not continue if not within sentence
if self.mapping.has_key(e2):
for t2 in self.mapping[e2]:
if coordPath(t1, t2):
return (True)
return (False)
if not edges:
return ([])
# where does the parent node belong to?
sentence = self.sentences[Analyser.findSentenceId(edges[0][0])]
depG = self.depGs[sentence]
edgemap = dict([(x[2].attrib['e2'], x) for x in edges])
connG = NX.Graph()
for x in edgemap.values():
connG.add_node(x)
pid = edges[0][2].attrib['e1']
parentTokens = []
if self.mapping.has_key(pid):
parentTokens = self.mapping[pid]
for e1 in edgemap.keys():
for e2 in edgemap.keys():
if not e1 == e2:
if connected(e1, e2):
connG.add_edge(edgemap[e1], edgemap[e2])
return (NX.connected_components(connG))
def getGrouping(node):
"""
Generates the edge groups describing the unflattened event.
@type node: cElementTree.Element
@param node: event to be unflattened
@rtype: list of lists of 3-tuples
@return: grouped edges
"""
uid = node.attrib['id']
t = node.attrib['type']
# 'neg' edges are not considered and will be
# removed from the final xml
edges = [(x[0], x[1], x[2]['xmlnode'])
for x in self.semG.out_edges(node, True)
if not x[2]['xmlnode'].attrib['type'] == 'neg']
if t in [
'Gene_expression', 'Transcription', 'Translation',
'Protein_catabolism'
]:
result = [[e] for e in edges]
# #sys.stderr.write("Splitting %s (%s) into %s - %s\n"%(uid,t,
# len(result),
# [[y[2].attrib['id']
# for y in x]
# for x in result]))
return (result)
elif t == 'Localization':
result = [[e] for e in edges]
# #sys.stderr.write("Splitting %s (%s) into %s - %s\n"%(uid,t,
# len(result),
# [[y[2].attrib['id']
# for y in x]
# for x in result]))
return (result)
elif t == 'Binding':
# Binding is not perfectly solvable
if self.perfect:
#sys.stderr.write("Skipping %s (%s)\n"%(uid,t))
return ([edges])
groups = getCoordGrouping(edges)
if len(groups) == 1:
# data suggests that regardless of number of members
# in the group, the binding should be split
# (cases of >2 are very rare)
# (the decision to split events with 2 members is about
# 1:1 but splitting is still slightly favoured)
result = [[e] for e in edges]
# #sys.stderr.write("Splitting %s (%s) into %s - %s\n"%(uid,t,
# len(result),
# [[y[2].attrib['id']
# for y in x]
# for x in result]))
return (result)
else:
# two groups should be split to pairwise combinations
# (can 'respectively' be ignored?)
# events with more than two proteins are rare
# so three or more groups should be treated in a
# pairwise manner
result = []
while groups:
g1 = groups.pop()
result.extend([(e1, e2) for g2 in groups for e1 in g1
for e2 in g2])
#sys.stderr.write("Generating inter-group pairs for %s (%s) - %s\n"%(uid,t,[[y[2].attrib['id'] for y in x] for x in result]))
return (result)
elif t == 'Phosphorylation':
result = [[e] for e in edges]
# #sys.stderr.write("Splitting %s (%s) into %s - %s\n"%(uid,t,
# len(result),
# [[y[2].attrib['id']
# for y in x]
# for x in result]))
return (result)
elif t in [
'Regulation', 'Positive_regulation', 'Negative_regulation'
]:
# Regulation is not perfectly solvable
# but for now there no better way than the baseline
# (can 'respectively' be ignored?)
cause = [
x for x in edges if x[2].attrib['type'].startswith('Cause')
]
theme = [
x for x in edges if x[2].attrib['type'].startswith('Theme')
]
if cause and theme:
result = [(ca, th) for ca in cause for th in theme]
#sys.stderr.write("Generating Cause-Theme combinations for %s (%s) - %s\n"%(uid,t,[[y[2].attrib['id'] for y in x] for x in result]))
return (result)
else:
result = [[e] for e in edges]
# #sys.stderr.write("Splitting %s (%s) into %s - %s\n"%(uid,t,
# len(result),
# [[y[2].attrib['id']
# for y in x]
# for x in result]))
return (result)
elif t in ['Protein']:
# Proteins have sites (etc.) as successors
# these edges will be processed later --> do nothing at this point
result = [edges]
else:
if t != "Entity":
sys.stderr.write("Invalid event type: %s\n" % t)
return ([edges])
# unflatten the graph in the cache
counter = Increment()
unprocessed_nodes = set(
[x for x in self.semG.nodes() if not self.semG.out_edges(x)])
while unprocessed_nodes:
next_nodes = set()
for current in unprocessed_nodes:
next_nodes.update(set(self.semG.predecessors(current)))
if self.semG.out_edges(current):
groups = getGrouping(current)
for edges in groups:
evid = counter.get()
newN = ET.Element('entity', current.attrib)
newId = newN.attrib['id'] + '.E' + evid
newN.attrib['id'] = newId
self.semG.add_node(newN)
for e in edges:
newE = ET.Element('interaction', e[2].attrib)
newEid = newE.attrib['id'] + '.E' + evid
newE.attrib['id'] = newEid
newE.attrib['e1'] = newId
self.semG.add_edge(newN, e[1], xmlnode=newE)
for e in self.semG.in_edges(current, True):
newE = ET.Element('interaction',
e[2]['xmlnode'].attrib)
newEid = newE.attrib['id'] + '.E' + evid
newE.attrib['id'] = newEid
newE.attrib['e2'] = newId
self.semG.add_edge(e[0], newN, xmlnode=newE)
self.semG.remove_node(current)
# ensure that nodes-to-be-processed have only out-neighbors
# that have already been processed
removable = set()
for x in next_nodes:
for y in next_nodes:
if NX.shortest_path(self.semG, x, y) and not x == y:
removable.add(x)
unprocessed_nodes = next_nodes - removable
def analyse(self):
"""
Analyses the parse, determines the unflattening strategy
for each event, and modifies the cache accordingly.
"""
def getCoordGrouping(edges):
"""
Workhorse function for determining the grouping of arguments.
@type edges: list of 3-tuples
@param edges: out-going edges of an event to be grouped
@rtype: list of lists of 3-tuples
@return: grouped edges
"""
def coordPath(this,target):
if depG.has_node(this) and depG.has_node(target):
paths1 = [NX.shortest_path(depG,x,this)
for x in parentTokens
if depG.has_node(x)]
paths2 = [NX.shortest_path(depG,x,target)
for x in parentTokens
if depG.has_node(x)]
# same coordination group if there is a pair of
# shortest paths from event that start with the same edge
# and this edge is not between the tokens of the event
start1 = set( [depG[p[0]][p[1]]['type'] for p in paths1
if (p and
len(p)>=2 and
not p[1] in parentTokens)] )
start2 = set( [depG[p[0]][p[1]]['type'] for p in paths2
if (p and
len(p)>=2 and
not p[1] in parentTokens)] )
return(start1.intersection(start2))
return(False)
def connected(e1,e2):
# do not continue if not within sentence
if self.mapping.has_key(e1):
for t1 in self.mapping[e1]:
# do not continue if not within sentence
if self.mapping.has_key(e2):
for t2 in self.mapping[e2]:
if coordPath(t1,t2):
return(True)
return(False)
if not edges:
return([])
# where does the parent node belong to?
sentence = self.sentences[Analyser.findSentenceId(edges[0][0])]
depG = self.depGs[sentence]
edgemap = dict( [(x[2].attrib['e2'],x) for x in edges] )
connG = NX.Graph()
for x in edgemap.values():
connG.add_node(x)
pid = edges[0][2].attrib['e1']
parentTokens = []
if self.mapping.has_key(pid):
parentTokens = self.mapping[pid]
for e1 in edgemap.keys():
for e2 in edgemap.keys():
if not e1==e2:
if connected(e1,e2):
connG.add_edge(edgemap[e1],edgemap[e2])
return(NX.connected_components(connG))
def getGrouping(node):
"""
Generates the edge groups describing the unflattened event.
@type node: cElementTree.Element
@param node: event to be unflattened
@rtype: list of lists of 3-tuples
@return: grouped edges
"""
uid = node.attrib['id']
t = node.attrib['type']
# 'neg' edges are not considered and will be
# removed from the final xml
edges = [(x[0],x[1],x[2]['xmlnode'])
for x in self.semG.out_edges(node,True)
if not x[2]['xmlnode'].attrib['type']=='neg']
if t in ['Gene_expression','Transcription',
'Translation','Protein_catabolism']:
result = [[e] for e in edges]
# #sys.stderr.write("Splitting %s (%s) into %s - %s\n"%(uid,t,
# len(result),
# [[y[2].attrib['id']
# for y in x]
# for x in result]))
return(result)
elif t=='Localization':
result = [[e] for e in edges]
# #sys.stderr.write("Splitting %s (%s) into %s - %s\n"%(uid,t,
# len(result),
# [[y[2].attrib['id']
# for y in x]
# for x in result]))
return(result)
elif t=='Binding':
# Binding is not perfectly solvable
if self.perfect:
#sys.stderr.write("Skipping %s (%s)\n"%(uid,t))
return([edges])
groups = getCoordGrouping(edges)
if len(groups)==1:
# data suggests that regardless of number of members
# in the group, the binding should be split
# (cases of >2 are very rare)
# (the decision to split events with 2 members is about
# 1:1 but splitting is still slightly favoured)
result = [[e] for e in edges]
# #sys.stderr.write("Splitting %s (%s) into %s - %s\n"%(uid,t,
# len(result),
# [[y[2].attrib['id']
# for y in x]
# for x in result]))
return(result)
else:
# two groups should be split to pairwise combinations
# (can 'respectively' be ignored?)
# events with more than two proteins are rare
# so three or more groups should be treated in a
# pairwise manner
result = []
while groups:
g1 = groups.pop()
result.extend( [(e1,e2)
for g2 in groups
for e1 in g1
for e2 in g2] )
#sys.stderr.write("Generating inter-group pairs for %s (%s) - %s\n"%(uid,t,[[y[2].attrib['id'] for y in x] for x in result]))
return(result)
elif t=='Phosphorylation':
result = [[e] for e in edges]
# #sys.stderr.write("Splitting %s (%s) into %s - %s\n"%(uid,t,
# len(result),
# [[y[2].attrib['id']
# for y in x]
# for x in result]))
return(result)
elif t in ['Regulation','Positive_regulation',
'Negative_regulation']:
# Regulation is not perfectly solvable
# but for now there no better way than the baseline
# (can 'respectively' be ignored?)
cause = [x for x in edges if
x[2].attrib['type'].startswith('Cause')]
theme = [x for x in edges if
x[2].attrib['type'].startswith('Theme')]
if cause and theme:
result = [(ca,th) for ca in cause for th in theme]
#sys.stderr.write("Generating Cause-Theme combinations for %s (%s) - %s\n"%(uid,t,[[y[2].attrib['id'] for y in x] for x in result]))
return(result)
else:
result = [[e] for e in edges]
# #sys.stderr.write("Splitting %s (%s) into %s - %s\n"%(uid,t,
# len(result),
# [[y[2].attrib['id']
# for y in x]
# for x in result]))
return(result)
elif t in ['Protein']:
# Proteins have sites (etc.) as successors
# these edges will be processed later --> do nothing at this point
result = [edges]
else:
if t != "Entity":
sys.stderr.write("Invalid event type: %s\n"%t)
return([edges])
# unflatten the graph in the cache
counter = Increment()
unprocessed_nodes = set([x for x in self.semG.nodes()
if not self.semG.out_edges(x)])
while unprocessed_nodes:
next_nodes = set()
for current in unprocessed_nodes:
next_nodes.update(set(self.semG.predecessors(current)))
if self.semG.out_edges(current):
groups = getGrouping(current)
for edges in groups:
evid = counter.get()
newN = ET.Element('entity',current.attrib)
newId = newN.attrib['id']+'.E'+evid
newN.attrib['id'] = newId
self.semG.add_node(newN)
for e in edges:
newE = ET.Element('interaction',e[2].attrib)
newEid = newE.attrib['id']+'.E'+evid
newE.attrib['id'] = newEid
newE.attrib['e1'] = newId
self.semG.add_edge(newN,e[1],xmlnode=newE)
for e in self.semG.in_edges(current,True):
newE = ET.Element('interaction',e[2]['xmlnode'].attrib)
newEid = newE.attrib['id']+'.E'+evid
newE.attrib['id'] = newEid
newE.attrib['e2'] = newId
self.semG.add_edge(e[0],newN,xmlnode=newE)
self.semG.remove_node(current)
# ensure that nodes-to-be-processed have only out-neighbors
# that have already been processed
removable = set()
for x in next_nodes:
for y in next_nodes:
if NX.shortest_path(self.semG,x,y) and not x==y:
removable.add(x)
unprocessed_nodes = next_nodes - removable