def __init__(self, training_ds, cf):
self.training_ds = training_ds
self.cf = cf
t = ""
self.esdc_structures = [
ExtendedSdc(esdcType="PATH",
entireText=t,
r=[TextStandoff(t, (0, 0))],
l=[
ExtendedSdc(esdcType="OBJECT",
f=[TextStandoff(t, (0, 0))])
]),
#ExtendedSdc(esdcType="EVENT", entireText=t,
# r=[TextStandoff(t, (0, 0))],
# l=[TextStandoff(t, (0, 0))],
# l2=[TextStandoff(t, (0, 0))])
]
self.esdc_field_to_texts = {}
groundings = set([])
for ex in self.training_ds.observations:
for sdc in ex.sdcs:
for field in ExtendedSdc.fieldNames:
self.esdc_field_to_texts.setdefault(field, [])
text = sdc.childText(field)
if text != "":
self.esdc_field_to_texts[field].append(text)
if len(groundings) < 10:
for glist in ex.annotation.groundings:
groundings.update(glist)
self.groundings = list(groundings)
def testBreadthFirstTraverseLoop(self):
cmd = "Move in front of the pallet of boxes in the center and wait."
esdc = ExtendedSdc(
'EVENT',
r=[TextStandoff(cmd, (0, 4)),
TextStandoff(cmd, (5, 7))],
l2=[],
f=[TextStandoff(cmd, (8, 13))])
esdc.l = [esdc]
def callback(esdcFromParent):
print "callback"
dataStructures.breadthFirstTraverse(esdc, callback)
def word_tokenize(self, string):
"""
Tokenize a string, returning a tuple. The first element is a
list of starting locations for the tokens and the second
element is a list of tokens.
"""
tokens = self.nltkTokenizer.tokenize(string)
indexes = []
startIdx = 0
for token in tokens:
idx = string.index(token, startIdx)
indexes.append(idx)
startIdx = idx + len(token)
if len(tokens) > 0:
lastToken = tokens[-1]
if len(lastToken) > 1 and lastToken[-1] in ('?', '.', '!'):
lastCharacter = lastToken[-1]
tokens[-1] = lastToken[0:-1]
tokens.append(lastCharacter)
indexes.append(indexes[-1] + len(lastToken) - 1)
return [
TextStandoff(string, (i, i + len(token)))
for i, token in zip(indexes, tokens)
]
def recipesToDataset(recipes):
observations = []
task_planner = nodeSearch.BeamSearch(CostFnRandom())
fe = kitchen_features.GGGFeatures()
for recipe in recipes:
print "training", recipe.name
for idx in range(recipe.num_instructions):
instruction = recipe.idx_to_instruction(idx)
esdc = ExtendedSdc("EVENT", instruction,
r=TextStandoff(instruction, (0, len(instruction))))
ggg = ggg_from_esdc(esdc)
states = recipe.idx_to_states(idx)
example_id = "%s_%d" % (recipe.name, idx)
features = fe.features(ggg, ggg.factors[0], states)
obs = dataset.ContinuousObservation(example_id, True, True, features,
sdcs=[esdc])
observations.append(obs)
negative_obs = negativeExamples(task_planner,
fe,
recipe.idx_to_start_state(idx),
esdc,
ggg,
example_id, states)
observations.extend(negative_obs)
return dataset.ContinuousDataset(observations)
def standoffForString(relation, sentence):
try:
if relation.startswith("prep"):
preposition = relation[relation.index("_") + 1:].replace("_", " ")
index = sentence.lower().find(preposition + " ")
if index == -1:
index = sentence.lower().find(preposition)
string = sentence[index:index + len(preposition)]
standoffs = []
offset = index
for token in string.split(" "):
standoffs.append(
TextStandoff(sentence, (offset, offset + len(token))))
offset += len(token) + 1
return (standoffs, preposition)
else:
raise ValueError("Invalid relation: " + ` relation ` + " in " +
` sentence `)
except:
print "relation", relation
print "sentence", sentence
raise
def updateRep(self):
if not hasattr(self, "frozen"):
self.frozen = False
for key, values in self:
for value in values:
if self.entireText == None:
self.entireText = value.entireText
else:
assert self.entireText == value.entireText, (
self.entireText, value.entireText, key, str(value),
[str(x) for x in values])
self.fields[key] = sorted(values, key=lambda x: x.range[0])
minRange = len(self.entireText)
maxRange = 0
for key, values in self:
for value in values:
minValueRange, maxValueRange = value.range
if minValueRange < minRange:
minRange = minValueRange
if maxValueRange > maxRange:
maxRange = maxValueRange
self.range = minRange, maxRange
self.startIdx = minRange
self.endIdx = maxRange
self.text = self.entireText[self.startIdx:self.endIdx]
self.flattenedEsdcs = flattenedEsdcs([self])
self.checkRep()
self._repr = self.recomputeRepr()
self._hash = hash(self._repr)
self.hash_string = fasthash(self._repr)
self.standoff = TextStandoff(self.entireText, self.range)
def loadSdc(xmlElement, text):
argMap = {}
for field in xmlElement.childNodes:
if isinstance(field, xml.dom.minidom.Element):
start = int(field.getAttribute("start"))
end = int(field.getAttribute("end"))
argMap[str(field.nodeName)] = TextStandoff(text, (start, end))
return Annotation(**argMap)
def __init__(self, **args):
self.annotationMap = args
self.entireText = None
self.instructionIdx = None
for key, value in self.annotationMap.iteritems():
if not (value is None):
if self.entireText is None:
self.entireText = value.entireText
else:
assert self.entireText == value.entireText, (
self.entireText, value.entireText)
for key in Annotation.keys:
if not (key
in self.annotationMap) or self.annotationMap[key] is None:
self.annotationMap[key] = TextStandoff(self.entireText, (0, 0))
self.range = TextStandoff(self.entireText,
enclosingRange(self.annotationMap.values()))
def testCorrectStandoffs(self):
annotations = yamlReader.load("data/forklift_open_ended.yaml")
esdc1 = annotations[0].esdcs[0]
esdc2 = annotations[1].esdcs[0]
new_entire_text = esdc1.entireText + " " + esdc2.entireText
sentence_standoff = TextStandoff(new_entire_text, esdc1.range)
correctStandoffs(sentence_standoff, esdc1)
self.assertEqual(esdc1.entireText, new_entire_text)
def setUp(self):
self.sentence = "Pick up the tire pallet."
self.esdc1 = ExtendedSdc("EVENT",
r=TextStandoff(self.sentence, (0, 7)),
l=TextStandoff(self.sentence, (8, 23)))
self.esdc2 = ExtendedSdc("EVENT",
r=TextStandoff(self.sentence, (0, 7)),
l=TextStandoff(self.sentence, (8, 23)))
self.esdc3 = ExtendedSdc("EVENT",
r=TextStandoff(self.sentence, (0, 7)),
l=TextStandoff(self.sentence, (8, 22)))
self.sentence2 = "Pick up the tire pallet near the truck."
self.childEsdc = ExtendedSdc("OBJECT",
f=TextStandoff(self.sentence2, (8, 23)),
r=TextStandoff(self.sentence2, (24, 28)),
l=TextStandoff(self.sentence2, (29, 38)))
self.parentEsdc = ExtendedSdc("EVENT",
r=TextStandoff(self.sentence2, (0, 7)),
l=self.childEsdc)
def tokenize(self, string):
instructions = string
sentences = self.tokenizer.tokenize(instructions)
standoffs = []
lastStart = 0
for sentence in sentences:
startIdx = instructions.index(sentence, lastStart)
endIdx = startIdx + len(sentence)
standoffs.append(TextStandoff(string, (startIdx, endIdx)))
lastStart = endIdx
for s1 in standoffs:
for s2 in standoffs:
assert s1 == s2 or not s1.overlaps(s2)
return standoffs
def testNestedRepeatedStrings(self):
from esdcs.dataStructures import ExtendedSdc, ExtendedSdcGroup
from standoff import TextStandoff
txt = "Move to the right side of the trailer of the trailer on the right and wait."
esdcs = [ExtendedSdc('EVENT', r=[TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (0, 4))],l2=[],l=[ExtendedSdc('PATH', r=[TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (5, 7))],l2=[],l=[ExtendedSdc('OBJECT', r=[TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (23, 25))],l2=[ExtendedSdc('OBJECT', r=[],l2=[],l=[],f=[TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (41, 44)), TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (45, 52))])],l=[ExtendedSdc('OBJECT', r=[],l2=[],l=[],f=[TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (26, 29)), TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (30, 37))])],f=[ExtendedSdc('OBJECT', r=[],l2=[],l=[],f=[TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (8, 11)), TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (12, 17)), TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (18, 22))])])],f=[])],f=[ExtendedSdc('OBJECT', r=[],l2=[],l=[],f=[], entireText='Move to the right side of the trailer of the trailer on the right and wait.')]), ExtendedSdc('OBJECT', r=[TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (53, 55))],l2=[],l=[ExtendedSdc('OBJECT', r=[],l2=[],l=[],f=[TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (56, 59)), TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (60, 65))])],f=[ExtendedSdc('OBJECT', r=[],l2=[],l=[],f=[TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (41, 44)), TextStandoff("Move to the right side of the trailer of the trailer on the right and wait.", (45, 52))])])]
entireText, yamlData = esdcIo.toYaml(ExtendedSdcGroup(esdcs))
rereadAnnotations = esdcIo.fromYaml(entireText, yamlData)
try:
self.assertEqual(list(rereadAnnotations), esdcs)
except:
print "start with", [e.asPrettyMap() for e in esdcs]
print "ended with", [e.asPrettyMap() for e in rereadAnnotations]
raise
def testCorrectStandoffsImmutable(self):
annotations = yamlReader.load("data/forklift_open_ended.yaml")
esdc1 = annotations[0].esdcs[0]
esdc2 = annotations[1].esdcs[0]
old_entire_text = esdc1.entireText
new_entire_text = esdc1.entireText + " " + esdc2.entireText
sentence_standoff = TextStandoff(new_entire_text, esdc1.range)
correctedEsdc1 = correctStandoffsImmutable(sentence_standoff,
ExtendedSdcGroup([esdc1]))
self.assertEqual(esdc1.entireText, old_entire_text)
self.assertEqual(correctedEsdc1.entireText, new_entire_text)
print str(correctedEsdc1[0])
self.assertEqual(" ".join(x.text for x in correctedEsdc1[0]["f"]),
"Forklift")
def standoffFromToken(txt, token):
return TextStandoff(txt, (token.beginPosition(), token.endPosition()))
def handleEsdcContents(argMap, entireText):
outputDict = {}
for argName, argValue in argMap.iteritems():
assert argName in ExtendedSdc.fieldNames, ("Arg " + ` argName ` +
" not in names." +
" Value: " + ` argValue `)
if argName == "id":
outputDict["esdc_id"] = argValue
elif isinstance(argValue, str):
matches = list(re.finditer(re.escape(argValue), entireText))
if len(matches) == 1:
match = matches[0]
tokens = []
matchText = match.group()
currentIndex = 0
for token in matchText.split():
# we've ensured the index both exists and is unique.
tokenIdx = matchText[currentIndex:].index(
token) + currentIndex
standoff = TextStandoff(
entireText, (match.start() + tokenIdx,
match.start() + tokenIdx + len(token)))
currentIndex = tokenIdx + len(token)
tokens.append(standoff)
outputDict[argName] = tokens
else:
candidates = [[match.start(), match.end()]
for match in matches]
token = argValue
tokenizer = IndexedTokenizer()
for candidate in candidates:
print "candidate", candidate
for standoff in tokenizer.tokenize(argValue):
print "- -", standoff.text
start_idx = standoff.start
print " - [%d, %d]" % (candidate[0] + start_idx,
candidate[0] + start_idx +
len(token))
raise ValueError("Must provide indices for token: '" +
argValue + "' in text '" + entireText + "'."
" matches: " + ` candidates `)
elif isEsdc(argValue):
outputDict[argName] = list(fromYaml(entireText, argValue))
elif isWordList(argValue):
tokens = []
try:
for token, (start, end) in argValue:
substring = entireText[start:end]
if substring != token:
print "candidates"
for match in re.finditer(token, entireText):
print[match.start(), match.end()]
raise ValueError("Token '" + token +
"' must correspond" + " to index " +
` (start, end) ` + "and not '" +
substring + "'.")
tokens.append(TextStandoff(entireText, (start, end)))
except:
print "Problem with", argValue
raise
outputDict[argName] = tokens
else:
raise ValueError("Must be strings or ESDCs: " + ` argValue `)
return outputDict
def testDegreeOfOverlap(self):
self.assertEqual(
TextStandoff("Testing 123", (0, 1)).degreeOfOverlap(
TextStandoff("Testing 123", (0, 1))), 1)
self.assertEqual(
TextStandoff("Testing 123", (0, 1)).degreeOfOverlap(
TextStandoff("Testing 123", (1, 2))), 0)
self.assertEqual(
TextStandoff("Testing 123", (1, 2)).degreeOfOverlap(
TextStandoff("Testing 123", (0, 1))), 0)
self.assertEqual(
TextStandoff("Testing 123", (0, 10)).degreeOfOverlap(
TextStandoff("Testing 123", (1, 2))), 1)
self.assertEqual(
TextStandoff("Testing 123", (0, 10)).degreeOfOverlap(
TextStandoff("Testing 123", (9, 10))), 1)
self.assertEqual(
TextStandoff("Testing 123", (0, 10)).degreeOfOverlap(
TextStandoff("Testing 123", (10, 11))), 0)
self.assertEqual(
TextStandoff("Testing 123", (0, 10)).degreeOfOverlap(
TextStandoff("Testing 123", (8, 11))), 2)
self.assertEqual(
TextStandoff("Testing 123", (8, 11)).degreeOfOverlap(
TextStandoff("Testing 123", (0, 10))), 2)
self.assertEqual(
TextStandoff("Testing 123", (0, 5)).degreeOfOverlap(
TextStandoff("Testing 123", (-1, 10))), 5)
self.assertEqual(
TextStandoff("Testing 123", (0, 5)).degreeOfOverlap(
TextStandoff("Testing 123", (6, 125))), 0)
def testContains(self):
self.assertEqual(
TextStandoff("Testing 123",
(0, 1)).contains(TextStandoff("Testing 123", (0, 1))),
True)
self.assertEqual(
TextStandoff("Testing 123",
(0, 1)).contains(TextStandoff("Testing 123", (1, 2))),
False)
self.assertEqual(
TextStandoff("Testing 123",
(1, 2)).contains(TextStandoff("Testing 123", (0, 1))),
False)
self.assertEqual(
TextStandoff("Testing 123",
(0, 10)).contains(TextStandoff("Testing 123",
(1, 2))), True)
self.assertEqual(
TextStandoff("Testing 123", (0, 10)).contains(
TextStandoff("Testing 123", (9, 10))), True)
self.assertEqual(
TextStandoff("Testing 123", (0, 10)).contains(
TextStandoff("Testing 123", (10, 11))), False)
def testOverlaps(self):
self.assertEqual(
TextStandoff("Testing 123",
(0, 1)).overlaps(TextStandoff("Testing 123", (0, 1))),
True)
self.assertEqual(
TextStandoff("Testing 123",
(0, 1)).overlaps(TextStandoff("Testing 123", (1, 2))),
False)
self.assertEqual(
TextStandoff("Testing 123",
(0, 10)).overlaps(TextStandoff("Testing 123",
(1, 2))), True)
self.assertEqual(
TextStandoff("Testing 123", (0, 10)).overlaps(
TextStandoff("Testing 123", (9, 10))), True)
self.assertEqual(
TextStandoff("Testing 123", (0, 10)).overlaps(
TextStandoff("Testing 123", (10, 11))), False)
self.assertEqual(
TextStandoff("Testing 123", (10, 11)).overlaps(
TextStandoff("Testing 123", (0, 10))), False)
def testBefore(self):
self.assertEqual(
TextStandoff("Testing 123",
(0, 1)).before(TextStandoff("Testing 123", (0, 1))),
False)
self.assertEqual(
TextStandoff("Testing 123",
(0, 1)).before(TextStandoff("Testing 123", (1, 2))),
True)
self.assertEqual(
TextStandoff("Testing 123",
(1, 2)).before(TextStandoff("Testing 123", (0, 1))),
False)
self.assertEqual(
TextStandoff("Testing 123",
(0, 10)).before(TextStandoff("Testing 123", (1, 2))),
False)
self.assertEqual(
TextStandoff("Testing 123",
(0, 10)).before(TextStandoff("Testing 123", (9, 10))),
False)
self.assertEqual(
TextStandoff("Testing 123",
(0, 10)).before(TextStandoff("Testing 123",
(10, 11))), True)
def extractEsdcList(self, sentence, dependencies):
"""
Returns a list of ESDCs, without hierarchy.
"""
esdcs = []
child_esdcs = []
leftover_deps = []
for relation, gov, dep in dependencies.dependencies:
#print "Relation:", relation
if relation == "root":
continue
if relation == "prep":
esdc = ExtendedSdc("EVENT",
r=gov,
l=ExtendedSdc("EVENT", r=dep))
elif relation.startswith("prep"):
prepStandoff, prep = standoffForString(relation, sentence)
govTag = dependencies.tagForTokenStandoff(gov)
if prep in [
"on", "in", "at", "near", "next to", "in front of",
"behind", "away from", "close to", "closer to"
]:
esdcType = "PLACE"
else:
esdcType = "PATH"
#if relation.startswith("prepc"):
if not govTag in ["NN", "NNS", "NNP"]:
esdc = ExtendedSdc("EVENT",
r=gov,
l=ExtendedSdc(esdcType,
r=prepStandoff,
l=dep))
child_esdcs.extend(esdc.l)
esdcs.append(esdc)
else:
if govTag in ["NN", "NNS", "NNP"]:
esdc = ExtendedSdc("OBJECT",
f=gov,
r=prepStandoff,
l=dep)
else:
esdc = ExtendedSdc(
"EVENT",
r=gov,
l=[ExtendedSdc(esdcType, r=prepStandoff, l=dep)])
child_esdcs.extend(esdc.l)
esdcs.append(esdc)
elif relation == "conj_and":
esdcs.append(ExtendedSdc("EVENT", r=[gov]))
esdcs.append(ExtendedSdc("EVENT", r=[dep]))
elif self.sdh.isa(relation, "arg"):
esdc = ExtendedSdc("EVENT", r=gov, l=dep)
esdcs.append(esdc)
elif self.sdh.isa(relation, "subj"):
esdc = ExtendedSdc("EVENT", f=dep, r=gov)
esdcs.append(esdc)
elif relation == "conj_and":
pass
elif relation == "dep":
esdc = ExtendedSdc("EVENT", r=[dep, gov])
esdcs.append(esdc)
else:
leftover_deps.append((relation, gov, dep))
for relation, gov, dep in leftover_deps:
for esdc in flattenedEsdcs(esdcs): #chain(esdcs, child_esdcs):
for key, valueList in esdc:
if gov in valueList:
valueList.append(dep)
freeze(esdcs)
esdcs = list(sorted(set(esdcs)))
unfreeze(esdcs)
esdcs = mergeAll(esdcs)
esdcs = mergeAll(esdcs)
#esdcs = mergeAll(esdcs)
#for esdc in flattenedEsdcs(esdcs):
# esdc.updateRep()
for esdc in flattenedEsdcs(esdcs):
esdc.updateRep()
if len(esdcs) == 0:
print "sentence", sentence, sentence.__class__
sentence_standoff = TextStandoff(sentence, (0, len(sentence)))
return [ExtendedSdc("EVENT", r=sentence_standoff)]
return esdcs
def testEqual1(self):
e1 = ExtendedSdc(
'EVENT',
r=[TextStandoff("Load the forklift onto the trailer.", (0, 4))],
l2=[
ExtendedSdc(
'PATH',
r=[
TextStandoff("Load the forklift onto the trailer.",
(18, 22))
],
l2=[],
l=[
ExtendedSdc(
'OBJECT',
r=[],
l2=[],
l=[],
f=[
TextStandoff(
"Load the forklift onto the trailer.",
(23, 26)),
TextStandoff(
"Load the forklift onto the trailer.",
(27, 34))
])
],
f=[])
],
l=[
ExtendedSdc(
'OBJECT',
r=[],
l2=[],
l=[],
f=[
TextStandoff("Load the forklift onto the trailer.",
(5, 8)),
TextStandoff("Load the forklift onto the trailer.",
(9, 17))
])
],
f=[
ExtendedSdc('OBJECT',
r=[],
l2=[],
l=[],
f=[],
entireText='Load the forklift onto the trailer.')
])
e2 = ExtendedSdc(
'EVENT',
r=[TextStandoff("Load the forklift onto the trailer.", (0, 4))],
l2=[
ExtendedSdc(
'PATH',
r=[
TextStandoff("Load the forklift onto the trailer.",
(18, 22))
],
l2=[],
l=[
ExtendedSdc(
'OBJECT',
r=[],
l2=[],
l=[],
f=[
TextStandoff(
"Load the forklift onto the trailer.",
(23, 26)),
TextStandoff(
"Load the forklift onto the trailer.",
(27, 34))
])
],
f=[])
],
l=[
ExtendedSdc(
'OBJECT',
r=[],
l2=[],
l=[],
f=[
TextStandoff("Load the forklift onto the trailer.",
(5, 8)),
TextStandoff("Load the forklift onto the trailer.",
(9, 17))
])
],
f=[
ExtendedSdc('OBJECT',
r=[],
l2=[],
l=[],
f=[],
entireText='Load the forklift onto the trailer.')
])
self.assertEqual(e1, e2)
self.assertFalse(e1 != e2)
self.assertTrue(e1 in [e2])
self.assertTrue(e2 in [e1])
def testSetter(self):
newL = [TextStandoff(self.sentence, (0, 4))]
self.esdc1.l = newL
self.assertEqual(self.esdc1.fields['l'], newL)
self.assertNotEqual(self.esdc1, self.esdc2)
def makeNullTextStandoff(text):
return TextStandoff(text, (0, 0))
def make_ggg_for_instruction(self, text):
esdc = ExtendedSdc("EVENT", text, r=TextStandoff(text, (0, len(text))))
ggg = ggg_from_esdc(esdc)
return esdc, ggg