def buildAdjacencyMatrixWithShortestPaths(tokenElements,
dependencyElements,
entityElements,
pairElement,
directed=True,
weight_by_distance=False):
#Punctuation dependencies are mostly junk
dependencyElements = removeDependencies(dependencyElements, ["punct"])
parseGraph = ParseGraph.ParseGraph(tokenElements, dependencyElements)
parseGraph.markNamedEntities(entityElements)
e1Id = pairElement.get("e1")
e2Id = pairElement.get("e2")
entity1TokenIds = parseGraph.getNamedEntityTokenIds([e1Id])
entity2TokenIds = parseGraph.getNamedEntityTokenIds([e2Id])
binaryPaths = parseGraph.buildBinaryPaths(entity1TokenIds, entity2TokenIds)
shortestPaths = ParseGraph.getShortestPaths(binaryPaths)
parseGraph.setAllDependencyWeights(0.3)
parseGraph.setDependencyWeightsByPath(shortestPaths, 0.9)
if weight_by_distance:
parseGraph.reduceWeightByDistance(0.9, 0.5)
parseGraph.setPPIPrefixForDependencies("sp", 0.9) # shortest path prefix
parseGraph.maskNames(e1Id, e2Id)
parseGraph.addPositionTags(entity1TokenIds, entity2TokenIds)
if pairElement.get("interaction") == "True":
output = 1.
else:
output = -1.
adjMatrix, labels = parseGraph.buildAdjacencyMatrix(floattype)
return adjMatrix, labels, output
def buildAdjacencyMatrixWithShortestPaths(tokenElements, dependencyElements, entityElements, pairElement, directed = True, weight_by_distance = False):
#Punctuation dependencies are mostly junk
dependencyElements = removeDependencies(dependencyElements, ["punct"])
parseGraph = ParseGraph.ParseGraph(tokenElements, dependencyElements)
parseGraph.markNamedEntities(entityElements)
e1Id = pairElement.get("e1")
e2Id = pairElement.get("e2")
entity1TokenIds = parseGraph.getNamedEntityTokenIds( [e1Id] )
entity2TokenIds = parseGraph.getNamedEntityTokenIds( [e2Id] )
binaryPaths = parseGraph.buildBinaryPaths(entity1TokenIds, entity2TokenIds)
shortestPaths = ParseGraph.getShortestPaths(binaryPaths)
parseGraph.setAllDependencyWeights(0.3)
parseGraph.setDependencyWeightsByPath(shortestPaths, 0.9)
if weight_by_distance:
parseGraph.reduceWeightByDistance(0.9, 0.5)
parseGraph.setPPIPrefixForDependencies("sp", 0.9) # shortest path prefix
parseGraph.maskNames(e1Id, e2Id)
parseGraph.addPositionTags(entity1TokenIds, entity2TokenIds)
if pairElement.get("interaction") == "True":
output = 1.
else:
output = -1.
adjMatrix, labels = parseGraph.buildAdjacencyMatrix(floattype)
return adjMatrix, labels, output
def buildAdjacencyMatrix(tokenElements, dependencyElements,
entityElements, metamapElements, pairElement, matrixSettings):
"""
Parameters
----------
tokenElements : cElementTree.Element
List of <token> elements, which represent all the tokens of the
sentence. Each <token> contains an ID, the text, POS tag and
character offset of the token.
dependencyElements : cElementTree.Element
List of <dependency> elements, which represent all the
dependencies of the (dependency-parsed) sentence. Each
<dependency> contains an ID, a source, a target and dependency
type.
entityElements : cElementTree.Element
List of <entity> elements, which represent protein mentions of
the sentence. Each <entity> contains an ID, the text and the
character offsets of the mention.
metamapElements: ??? or None
There is code to extract metamappings from an analysis XML file
(in GraphMatrices.build_sentence_dict), but there are no example
corpora which have these attributes!
pairElement : cElementTree.Element
a <pair> element, which contains an ID, the IDs of both
entities and truth value stating whether there's an interaction
between the two.
matrixSettings : MatrixSettings
contains settings for creating an adjacency matrix
Returns
-------
matrix_tuple : tuple of (adjMatrix, labels, output)
TODO ???
"""
m = matrixSettings
#Punctuation dependencies are mostly junk
dependencyElements = removeDependencies(dependencyElements, m.removeDependencies)
parseGraph = ParseGraph.ParseGraph(tokenElements, dependencyElements, m.mergeDependencies)
parseGraph.shortestPathMethod = "dijkstra"
parseGraph.markNamedEntities(entityElements)
e1Id = pairElement.get("e1")
e2Id = pairElement.get("e2")
entity1TokenIds = parseGraph.getNamedEntityTokenIds( [e1Id] )
entity2TokenIds = parseGraph.getNamedEntityTokenIds( [e2Id] )
interactionWordTokenIds = parseGraph.getTokenIdsByText(m.interactionWords, False)
# Give dependencies base weights
parseGraph.setAllDependencyWeights(m.depBaseWeight)
# Set dependencies' weights based on paths
pathStyles = ParseGraph.splitPathStyles(m.paths)
for style in pathStyles:
paths = []
if style["type"] == "binary":
paths = parseGraph.buildBinaryPaths(entity1TokenIds, entity2TokenIds, style["length"], style["direction"]=="directed", m.pathTimeout)
elif style["type"].find("tertiary") != -1:
paths = parseGraph.buildTertiaryPaths(entity1TokenIds, interactionWordTokenIds, entity2TokenIds, style["type"]=="closest_tertiary", style["length"], style["direction"]=="directed", m.pathTimeout)
if m.markInteractionWords == MatrixSettings.markedInteractionWords.fromTertiaryPaths:
parseGraph.setPPIInteractionWords(paths)
for i in range(len(paths)):
paths[i] = paths[i][0]
if paths != None:
parseGraph.setDependencyWeightsByPath(paths, style["weight"])
# Reduce dependencies' weights by distance from threshold
if m.weightByDistance:
parseGraph.reduceWeightByDistance(m.depWeightReductionThreshold, m.depWeightReductionFactor)
# Set dependency prefixes
if m.depPrefixThreshold > 0.0:
parseGraph.setPPIPrefixForDependencies(m.depPrefix, m.depPrefixThreshold) # f.e. shortest path prefix
# Set token texts
if m.tokenPPIText == MatrixSettings.ppiTexts.full:
parseGraph.ppiTextFromOriginalText()
elif m.tokenPPIText == MatrixSettings.ppiTexts.stem:
parseGraph.ppiTextFromStems() # FIXME: not implemented!
else:
print >> sys.stderr, "Illegal ppiText setting", m.tokenPPIText
sys.exit(1)
# Add metamap codes
if metamapElements != None:
metamapDict = {}
for metamapElement in metamapElements:
metamapDict[metamapElement.get("tokenid")] = metamapElement.get("basecodes").split(",")
parseGraph.addMetamapCodes(metamapDict)
if m.maskPPIText:
parseGraph.maskNames(e1Id, e2Id)
if m.tokenPositionTags:
parseGraph.addPositionTags(entity1TokenIds, entity2TokenIds)
output = 1. if pairElement.get("interaction") == "True" else -1.
adjMatrix, labels = parseGraph.buildAdjacencyMatrix(floattype, m.directed, m.linearOrderWeight)
return adjMatrix, labels, output
def buildAdjacencyMatrix(tokenElements, dependencyElements, entityElements,
metamapElements, pairElement, matrixSettings):
"""
Parameters
----------
tokenElements : cElementTree.Element
List of <token> elements, which represent all the tokens of the
sentence. Each <token> contains an ID, the text, POS tag and
character offset of the token.
dependencyElements : cElementTree.Element
List of <dependency> elements, which represent all the
dependencies of the (dependency-parsed) sentence. Each
<dependency> contains an ID, a source, a target and dependency
type.
entityElements : cElementTree.Element
List of <entity> elements, which represent protein mentions of
the sentence. Each <entity> contains an ID, the text and the
character offsets of the mention.
metamapElements: ??? or None
There is code to extract metamappings from an analysis XML file
(in GraphMatrices.build_sentence_dict), but there are no example
corpora which have these attributes!
pairElement : cElementTree.Element
a <pair> element, which contains an ID, the IDs of both
entities and truth value stating whether there's an interaction
between the two.
matrixSettings : MatrixSettings
contains settings for creating an adjacency matrix
Returns
-------
matrix_tuple : tuple of (adjMatrix, labels, output)
TODO ???
"""
m = matrixSettings
#Punctuation dependencies are mostly junk
dependencyElements = removeDependencies(dependencyElements,
m.removeDependencies)
parseGraph = ParseGraph.ParseGraph(tokenElements, dependencyElements,
m.mergeDependencies)
parseGraph.shortestPathMethod = "dijkstra"
parseGraph.markNamedEntities(entityElements)
e1Id = pairElement.get("e1")
e2Id = pairElement.get("e2")
entity1TokenIds = parseGraph.getNamedEntityTokenIds([e1Id])
entity2TokenIds = parseGraph.getNamedEntityTokenIds([e2Id])
interactionWordTokenIds = parseGraph.getTokenIdsByText(
m.interactionWords, False)
# Give dependencies base weights
parseGraph.setAllDependencyWeights(m.depBaseWeight)
# Set dependencies' weights based on paths
pathStyles = ParseGraph.splitPathStyles(m.paths)
for style in pathStyles:
paths = []
if style["type"] == "binary":
paths = parseGraph.buildBinaryPaths(
entity1TokenIds, entity2TokenIds, style["length"],
style["direction"] == "directed", m.pathTimeout)
elif style["type"].find("tertiary") != -1:
paths = parseGraph.buildTertiaryPaths(
entity1TokenIds, interactionWordTokenIds, entity2TokenIds,
style["type"] == "closest_tertiary", style["length"],
style["direction"] == "directed", m.pathTimeout)
if m.markInteractionWords == MatrixSettings.markedInteractionWords.fromTertiaryPaths:
parseGraph.setPPIInteractionWords(paths)
for i in range(len(paths)):
paths[i] = paths[i][0]
if paths != None:
parseGraph.setDependencyWeightsByPath(paths, style["weight"])
# Reduce dependencies' weights by distance from threshold
if m.weightByDistance:
parseGraph.reduceWeightByDistance(m.depWeightReductionThreshold,
m.depWeightReductionFactor)
# Set dependency prefixes
if m.depPrefixThreshold > 0.0:
parseGraph.setPPIPrefixForDependencies(
m.depPrefix, m.depPrefixThreshold) # f.e. shortest path prefix
# Set token texts
if m.tokenPPIText == MatrixSettings.ppiTexts.full:
parseGraph.ppiTextFromOriginalText()
elif m.tokenPPIText == MatrixSettings.ppiTexts.stem:
parseGraph.ppiTextFromStems() # FIXME: not implemented!
else:
print >> sys.stderr, "Illegal ppiText setting", m.tokenPPIText
sys.exit(1)
# Add metamap codes
if metamapElements != None:
metamapDict = {}
for metamapElement in metamapElements:
metamapDict[metamapElement.get("tokenid")] = metamapElement.get(
"basecodes").split(",")
parseGraph.addMetamapCodes(metamapDict)
if m.maskPPIText:
parseGraph.maskNames(e1Id, e2Id)
if m.tokenPositionTags:
parseGraph.addPositionTags(entity1TokenIds, entity2TokenIds)
output = 1. if pairElement.get("interaction") == "True" else -1.
adjMatrix, labels = parseGraph.buildAdjacencyMatrix(
floattype, m.directed, m.linearOrderWeight)
return adjMatrix, labels, output