def printMLN(path=None):
out_str = ""
for ci in Clust.clusts:
cl = Clust.getClust(ci)
out_str = "{}\t{}\n".format(cl._clustIdx,cl)
for aci in cl._argClusts:
ac = cl._argClusts[aci]
out_str += "\t{}: ".format(aci)
out_str += "\t".join(["{}: {}".format(k, v)
for k, v in ac._argNum_cnt.items()])
out_str += "\n\t"
out_str += "\t".join(["{}: {}: {}".format(k,
ArgType.getArgType(k).toString(),
v)
for k, v in ac._argTypeIdx_cnt.items()])
out_str += "\n\t"
out_str += "\t".join(["{}: {}: {}".format(k,
Clust.getClust(k),
v)
for k, v in ac._chdClustIdx_cnt.items()])
out_str += "\n"
if path is not None:
dst = "{}/{}.mln".format(path,
os.path.basename(os.path.dirname(path)))
with open(dst, 'w') as f:
f.write(out_str)
return None
else:
return out_str
def printParse(path=None):
out_str = ""
for rnid, pt in Part.rootNodeId_part.items():
out_str += "{}\t{}\n".format(rnid, pt._relTreeRoot.getTreeStr())
out_str += "\t{}: {}\n".format(pt._clustIdx,
Clust.getClust(pt._clustIdx).toString())
if pt._parPart is None:
out_str += "\t\n\t\n"
else:
arg = pt._parPart.getArgument(pt._parArgIdx)
out_str += "\t{}\t{}\t{}\n".format(pt._parPart._relTreeRoot.getId(),
pt._parPart._clustIdx,
Clust.getClust(pt._parPart._clustIdx))
out_str += "\t{}: {}: {}\n".format(pt._parPart.getArgClust(pt._parArgIdx),
arg._path.getArgType(),
ArgType.getArgType(arg._path.getArgType()))
if path is not None:
dst = "{}/{}.parse".format(path,
os.path.basename(os.path.dirname(path)))
with open(dst, 'w') as f:
f.write(out_str)
return None
else:
return out_str
def execComposePart(self, pp, cp):
parClustIdx = pp._clustIdx
chdClustIdx = cp._clustIdx
pcl = Clust.getClust(parClustIdx)
ccl = Clust.getClust(chdClustIdx)
dep = pp.getArguments()[cp._parArgIdx]._path.getDep()
pp._relTreeRoot.addChild(dep, cp._relTreeRoot)
nrti = RelType.getRelType(pp._relTreeRoot)
ncl = Clust.getClust(next(iter(Clust.getClustsWithRelType(nrti))))
nci = ncl.getId()
pp.removeArgument(cp._parArgIdx, clust_only=True)
for argIdx, arg in pp.getArguments().items():
pp.unsetArgClust(argIdx)
arg._argPart.unsetParent()
pp.changeClust(nci, nrti, clust_only=True)
for argIdx, arg in pp.getArguments().items():
ati = arg._path.getArgType()
aci = -1
if ati not in ncl._argTypeIdx_argClustIdxs:
aci = ncl.createArgClust(ati)
elif len(ncl._argTypeIdx_argClustIdxs[ati]) == 0:
aci = ncl.createArgClust(ati)
else:
aci = next(iter(ncl._argTypeIdx_argClustIdxs[ati]))
arg._argPart.setParent(pp, argIdx)
pp.setArgClustOnly(argIdx, aci)
pp.setRelTypeIdx(nrti)
for argIdx, arg in cp.getArguments():
ati = arg._path.getArgType()
aci = -1
if ati not in ncl._argTypeIdx_argClustIdxs:
aci = ncl.createArgClust(ati)
elif len(ncl._argTypeIdx_argClustIdxs[ati]) == 0:
aci = ncl.createArgClust(ati)
else:
aci = next(iter(ncl._argTypeIdx_argClustIdxs[ati]))
cp.unsetArgClustOnly(argIdx)
pp.setArgClust(pp.addArgument(arg), aci)
arg._argPart.setParent(pp, pp.addArgument(arg))
cp.destroy()
return None
def genString(self):
self._str = "OP_{}:".format(self._op)
if self._op == SearchOp.OP_MERGE_CLUST:
c1 = Clust.getClust(self._clustIdx1)
c2 = Clust.getClust(self._clustIdx2)
self._str += "{} == {}".format(c1.toString(), c2.toString())
elif self._op == SearchOp.OP_MERGE_ROLE:
self._str += "{}:{}:{}".format(self._clustIdx, self._argIdx1,
self._argIdx2)
elif self._op == SearchOp.OP_COMPOSE:
rc = Clust.getClust(self._parClustIdx)
ac = Clust.getClust(self._chdClustIdx)
self._str += "{} ++ {}".format(rc.toString(), ac.toString())
def setArgs(self, art_id, sent_id, sent, idx):
this_part = Part.getPartByRootNodeId(
genTreeNodeID(art_id, sent_id, idx))
this_node = this_part.getRelTreeRoot()
node_clust = Clust.getClust(this_part.getClustIdx())
children = sent.get_children(idx)
if children is None:
return None
else:
for dependency, child_index in children:
child_node_id = genTreeNodeID(art_id, sent_id, child_index)
path = Path(dependency)
argTypeIdx = path.getArgType()
child_part = Part.getPartByRootNodeId(child_node_id)
if child_part.getParPart() is not None:
continue
arg = Argument(this_node, path, child_part)
argIdx = this_part.addArgument(arg)
child_part.setParent(this_part, argIdx)
argClustIdxs = node_clust.getArgClustIdxs(argTypeIdx)
argClustIdx = -1
if argClustIdxs is None:
argClustIdx = node_clust.createArgClust(argTypeIdx)
else:
argClustIdx = next(iter(argClustIdxs))
this_part.setArgClust(argIdx, argClustIdx, clust_only=True)
setArgs(art_id, sent_id, sent, child_index)
return None
def createAgenda(self, verbose=False):
if verbose:
clust_cnt = len(Part.getClustPartRootNodeIds())
milestones = set([x for x in range(1, 10, 1)])
i = 0
for clust_id in Part.getClustPartRootNodeIds():
clust = Clust.getClust(clust_id)
if clust.getType() != 'C':
continue
elif clust.isStop():
continue
# # self.logc.write("Adding to agenda for cluster {}\n".format(clust_id))
self.addAgendaForNewClust(clust_id, verbose)
if verbose:
i += 1
done = math.floor(i * 10 / clust_cnt)
if done in milestones:
milestones.remove(done)
print("{}% complete.".format(done * 10))
# self.logc.close()
return None
def changeClustRemap(self,
newClustIdx,
argClustIdx_newArgClustIdx,
clust_only=False):
if not clust_only:
oldClustIdx = self.getClustIdx()
ocl = Clust.getClust(oldClustIdx)
self.changeClust(newClustIdx,
self.getRelTypeIdx(),
clust_only=clust_only)
argIdx_newArgClustIdx = {}
for ai, arg in self._args.items():
oaci = self._argIdx_argClustIdx.pop(ai)
self._argClustIdx_argIdxs[oaci].remove(ai)
if len(self._argClustIdx_argIdxs[oaci]) == 0:
del self._argClustIdx_argIdxs[oaci]
argIdx_newArgClustIdx[ai] = argClustIdx_newArgClustIdx[oaci]
if not clust_only:
ocl.onPartUnsetArg(self, arg, oaci)
for ai in self._args:
aci = argIdx_newArgClustIdx[ai]
self.setArgClust(ai, aci, clust_only=clust_only)
return None
def setArgClust(self, argIdx, argClustIdx, clust_only=False):
oldArgClustIdx = -1
if argIdx in self._argIdx_argClustIdx:
oldArgClustIdx = self.getArgClust(argIdx)
if oldArgClustIdx != argClustIdx:
self._argIdx_argClustIdx[argIdx] = argClustIdx
if argClustIdx not in self._argClustIdx_argIdxs:
self._argClustIdx_argIdxs[argClustIdx] = set()
self._argClustIdx_argIdxs[argClustIdx].add(argIdx)
arg = self.getArgument(argIdx)
if not clust_only:
cl = Clust.getClust(self._clustIdx)
if oldArgClustIdx < 0:
if not clust_only:
cl.onPartSetArg(self, arg, argClustIdx)
else:
self._argClustIdx_argIdxs[oldArgClustIdx].remove(argIdx)
if len(self._argClustIdx_argIdxs[oldArgClustIdx]) == 0:
del self._argClustIdx_argIdxs[oldArgClustIdx]
if not clust_only:
cl.onPartSetArg(self, arg, argClustIdx, oldArgClustIdx)
return None
def unsetArgClust(self, argIdx, clust_only=False):
oldArgClustIdx = self._argIdx_argClustIdx.pop(argIdx)
arg = self.getArgument(argIdx)
self._argClustIdx_argIdxs[oldArgClustIdx].remove(argIdx)
if len(self._argClustIdx_argIdxs[oldArgClustIdx]) == 0:
del self._argClustIdx_argIdxs[oldArgClustIdx]
if not clust_only:
cl = Clust.getClust(self.getClustIdx())
cl.onPartUnsetArg(self, arg, oldArgClustIdx)
return None
def setClust(self, clustIdx, clust_only=False):
self._clustIdx = clustIdx
rootID = self.getRelTreeRoot().getId()
if clustIdx not in Part.clustIdx_partRootNodeIds:
Part.clustIdx_partRootNodeIds[clustIdx] = SortedSet()
Part.clustIdx_partRootNodeIds[clustIdx].add(rootID)
if not clust_only:
cl = Clust.getClust(clustIdx)
cl.onPartSetClust(self)
return None
def addAgendaMC(self, clustIdx1, clustIdx2, neighType):
if not (self._skipMC or clustIdx1 == clustIdx2):
type1 = Clust.getClust(clustIdx1).getType()
type2 = Clust.getClust(clustIdx2).getType()
if type2 == 'C' and type1 == 'C':
op = SearchOp()
op._op = SearchOp.OP_MERGE_CLUST
op._clustIdx1 = min((clustIdx1, clustIdx2))
op._clustIdx2 = max((clustIdx1, clustIdx2))
if not self.moveAgendaToScore(op):
if op not in self._mc_neighs:
self._mc_neighs[op] = set()
if len(self._mc_neighs[op]) + 1 >= ParseParams.minMCCnt:
self._agendaToScore.add(op)
del self._mc_neighs[op]
else:
self._mc_neighs[op].add(neighType)
## self.logc.write("\t\tMerge Op: {}; mc_neighs: {}, agendaToScore: {}\n".format(op, len(self._mc_neighs), len(self._agendaToScore)))
return None
def initializeSent(self, ai, sj, sent, verbose=False):
'''
Create TreeNode, Part, and Clust for each token in a sentence,
also adding/assigning RelTypes.
Increment the root count for the cluster assigned to the root
token (tokens with a parent of ROOT).
Finally, run CreateArgs() to define the parent-child relation-
ships. This call is recursive, traversing the whole dependency
tree for each sentence.
'''
self.numTkns += len(sent.get_tokens()) - 1
roots = sent.get_children(0)
if roots is None:
return None
elif len(roots) == 0:
return None
for k in range(1, len(sent.get_tokens())):
Parse.part_from_node(ai, sj, sent, k, sent.get_token(k))
# if len(roots) == 1:
for _, idx in roots:
sub_node_id = genTreeNodeID(ai, sj, idx)
# Is this global set really necessary? I don't think it is...
self.rootTreeNodeIds.add(sub_node_id)
node_part = Part.getPartByRootNodeId(sub_node_id)
if node_part is None:
continue
ncl = Clust.getClust(node_part.getClustIdx())
ncl.incRootCnt()
self.createArgs(ai, sj, sent, idx, verbose)
return None
def scoreMCForParent(self, clustIdx1, clustIdx2):
scr = 0
if clustIdx1 in Clust.clustIdx_parArgs and clustIdx2 in Clust.clustIdx_parArgs:
parents1 = Clust.clustIdx_parArgs[clustIdx1]
parents2 = Clust.clustIdx_parArgs[clustIdx2]
for par_arg in parents1:
if par_arg in parents2:
par_clust_id, arg_clust_id = par_arg
pcl = Clust.getClust(par_clust_id)
if pcl is None:
print("ERR: ScoreMC parent cluster is null: {}, {}".
format(clustIdx1, clustIdx2))
continue
ac = pcl._argClusts[arg_clust_id]
c1 = ac._chdClustIdx_cnt[clustIdx1]
c2 = ac._chdClustIdx_cnt[clustIdx2]
scr += ParseParams.priorNumParam
scr += Scorer.updateScore(c1, c2)
return scr
def changeClust(self, newClustIdx, newRelTypeIdx, clust_only=False):
oldClustIdx = self.getClustIdx()
rootID = self.getRelTreeRoot().getId()
Part.clustIdx_partRootNodeIds[oldClustIdx].discard(rootID)
if clust_only:
self._relTypeIdx = newRelTypeIdx
else:
ocl = Clust.getClust(oldClustIdx)
ocl.onPartUnsetClust(self)
self.setRelTypeIdx(newRelTypeIdx)
self.setClust(newClustIdx, clust_only=clust_only)
parent = self.getParPart()
if parent is None:
if newClustIdx in Clust.clustIdx_rootCnt:
Clust.clustIdx_rootCnt[newClustIdx] += 1
else:
Clust.clustIdx_rootCnt[newClustIdx] = 1
Clust.clustIdx_rootCnt[newClustIdx] -= 1
else:
parent_clust_id = parent.getClustIdx()
paci = parent.getArgClust(self.getParArgIdx())
pcl = Clust.getClust(parent_clust_id)
pac = pcl._argClusts[paci]
pac._chdClustIdx_cnt[oldClustIdx] -= 1
if newClustIdx in pac._chdClustIdx_cnt:
pac._chdClustIdx_cnt[newClustIdx] += 1
else:
pac._chdClustIdx_cnt[newClustIdx] = 1
pa = (parent_clust_id, paci)
Clust.clustIdx_parArgs[oldClustIdx][pa] -= 1
if newClustIdx not in Clust.clustIdx_parArgs:
Clust.clustIdx_parArgs[newClustIdx] = {}
if pa in Clust.clustIdx_parArgs[newClustIdx]:
Clust.clustIdx_parArgs[newClustIdx][pa] += 1
else:
Clust.clustIdx_parArgs[newClustIdx][pa] = 1
opci = (parent_clust_id, oldClustIdx)
npci = (parent_clust_id, newClustIdx)
ptnid = (parent.getRelTreeRoot().getId(), rootID)
Part.pairClustIdxs_pairPartRootNodeIds[opci].discard(ptnid)
if len(Part.pairClustIdxs_pairPartRootNodeIds[opci]) == 0:
del Part.pairClustIdxs_pairPartRootNodeIds[opci]
# Part.clustIdx_pairClustIdxs[oldClustIdx].discard(opci)
# Part.clustIdx_pairClustIdxs[parent_clust_id].discard(opci)
if npci not in Part.pairClustIdxs_pairPartRootNodeIds:
Part.pairClustIdxs_pairPartRootNodeIds[npci] = set()
Part.pairClustIdxs_pairPartRootNodeIds[npci].add(ptnid)
# Part.clustIdx_pairClustIdxs[parent_clust_id].add(npci)
# if newClustIdx not in Part.clustIdx_pairClustIdxs:
# Part.clustIdx_pairClustIdxs[newClustIdx] = set()
# Part.clustIdx_pairClustIdxs[newClustIdx].add(npci)
return None
def unsetRelTypeIdx(self):
old_type = self._relTypeIdx
cl = Clust.getClust(self._clustIdx)
cl.onPartUnsetRelTypeIdx(old_type)
return None
def scoreOpCompose(self, rcidx, acidx):
def update_score_from_dict(scr, d, orig_d):
for key, cnt in d.items():
origcnt = orig_d[key]
# assert origcnt >= cnt
scr -= xlogx(origcnt)
if cnt > 0:
scr += xlogx(cnt)
else:
scr += ParseParams.priorNumParam
return scr
# get parent and child root-node id numbers
parChdNids = Part.getPairPartRootNodeIds(rcidx, acidx)
if parChdNids is None:
return -10000
score = 0
rcl = Clust.getClust(rcidx)
acl = Clust.getClust(acidx)
# Parent count, child count, and count of times they occur
# together.
rtc_new = rcl._ttlCnt
atc_new = acl._ttlCnt
ratc_new = 0
raRootCnt = 0
parArg_cnt = dict()
rRelTypeIdx_newcnt = dict()
aRelTypeIdx_newcnt = dict()
raRelTypeIdx_newcnt = dict()
rArgClustIdx_argNum_cnt = dict()
aArgClustIdx_argNum_cnt = dict()
rNewArgClustIdx_argNum_cnt = dict()
aNewArgClustIdx_argNum_cnt = dict()
rArgClustIdx_argTypeIdx_cnt = dict()
aArgClustIdx_argTypeIdx_cnt = dict()
rNewArgClustIdx_argTypeIdx_cnt = dict()
aNewArgClustIdx_argTypeIdx_cnt = dict()
rArgClustIdx_chdClustIdx_cnt = dict()
aArgClustIdx_chdClustIdx_cnt = dict()
rNewArgClustIdx_chdClustIdx_cnt = dict()
aNewArgClustIdx_chdClustIdx_cnt = dict()
rArgClustIdx_partCnt = dict()
aArgClustIdx_partCnt = dict()
rNewArgClustIdx_partCnt = dict()
aNewArgClustIdx_partCnt = dict()
rArgClustIdx_argCnt = dict()
aArgClustIdx_argCnt = dict()
rNewArgClustIdx_argCnt = dict()
aNewArgClustIdx_argCnt = dict()
# For each parent-child pair:
for pcnid in parChdNids:
pp, cp = Part.getPartByRootNodeId(
pcnid[0]), Part.getPartByRootNodeId(pcnid[1])
rtc_new -= 1
atc_new -= 1
ratc_new += 1
rrt = pp.getRelTypeIdx()
art = cp.getRelTypeIdx()
raArgClustidx = pp.getArgClust(cp._parArgIdx)
# Decrement individual relType counts and increment the combined
# relType count for this pair
rRelTypeIdx_newcnt = dec_key(rRelTypeIdx_newcnt,
rrt,
base=rcl._relTypeIdx_cnt[rrt])
aRelTypeIdx_newcnt = dec_key(aRelTypeIdx_newcnt,
art,
base=acl._relTypeIdx_cnt[art])
raRelTypeIdx_newcnt = inc_key(raRelTypeIdx_newcnt, (rrt, art))
pp_par = pp.getParPart()
# If the parent has a parent, increment the parArg count, otherwise
# increment the root count.
if pp_par is not None:
ai = pp.getParArgIdx()
ppi = pp_par.getClustIdx()
aci = pp_par.getArgClust(ai)
parArg_cnt = inc_key(parArg_cnt, (ppi, aci))
else:
raRootCnt += 1
# For each argClust on the parent part, decrement the old parent
# part count and argClust count, and increment the new ones. The
# trick is don't copy/increment the counts for argument shared by
# this pair.
for arg_ci in pp._argClustIdx_argIdxs:
an = len(pp._argClustIdx_argIdxs[arg_ci])
ac = rcl._argClusts[arg_ci]
rArgClustIdx_partCnt = dec_key(rArgClustIdx_partCnt,
arg_ci,
base=len(
ac._partRootTreeNodeIds))
if arg_ci not in rArgClustIdx_argNum_cnt:
rArgClustIdx_argNum_cnt[arg_ci] = {}
rArgClustIdx_argNum_cnt[arg_ci] = \
dec_key(rArgClustIdx_argNum_cnt[arg_ci],
an,
base=ac._argNum_cnt[an])
newArgNum = an
if arg_ci == raArgClustidx:
newArgNum -= 1
if newArgNum == 0:
continue
if arg_ci not in rNewArgClustIdx_argNum_cnt:
rNewArgClustIdx_argNum_cnt[arg_ci] = {}
rNewArgClustIdx_argNum_cnt[arg_ci] = \
inc_key(rNewArgClustIdx_argNum_cnt[arg_ci], newArgNum)
rNewArgClustIdx_partCnt = inc_key(rNewArgClustIdx_partCnt,
arg_ci)
# Same as above, but for child part, and we don't skip anything.
for arg_ci in cp._argClustIdx_argIdxs:
an = len(cp._argClustIdx_argIdxs[arg_ci])
ac = acl._argClusts[arg_ci]
aArgClustIdx_partCnt = dec_key(aArgClustIdx_partCnt,
arg_ci,
base=len(
ac._partRootTreeNodeIds))
if arg_ci not in aArgClustIdx_argNum_cnt:
aArgClustIdx_argNum_cnt[arg_ci] = {}
aArgClustIdx_argNum_cnt[arg_ci] = \
dec_key(aArgClustIdx_argNum_cnt[arg_ci],
an,
base=ac._argNum_cnt[an])
if arg_ci not in aNewArgClustIdx_argNum_cnt:
aNewArgClustIdx_argNum_cnt[arg_ci] = {}
aNewArgClustIdx_argNum_cnt[arg_ci] = \
inc_key(aNewArgClustIdx_argNum_cnt[arg_ci], an)
aNewArgClustIdx_partCnt = inc_key(aNewArgClustIdx_partCnt,
arg_ci)
args = pp.getArguments()
# For all the parent's arguments
for ai, arg in args.items():
arg_part = arg._argPart
child_clust_id = arg_part._clustIdx
aci = pp.getArgClust(ai)
ac = rcl._argClusts[aci]
ati = arg._path.getArgType()
# Drop the old arguments
rArgClustIdx_argCnt = dec_key(rArgClustIdx_argCnt,
aci,
base=ac._ttlArgCnt)
if aci not in rArgClustIdx_argTypeIdx_cnt:
rArgClustIdx_argTypeIdx_cnt[aci] = {}
rArgClustIdx_argTypeIdx_cnt[aci] = \
dec_key(rArgClustIdx_argTypeIdx_cnt[aci],
ati,
base=ac._argTypeIdx_cnt[ati])
if aci not in rArgClustIdx_chdClustIdx_cnt:
rArgClustIdx_chdClustIdx_cnt[aci] = {}
rArgClustIdx_chdClustIdx_cnt[aci] = \
dec_key(rArgClustIdx_chdClustIdx_cnt[aci],
child_clust_id,
base=ac._chdClustIdx_cnt[child_clust_id])
# Add the new arguments, except for the child part we're possibly
# absorbing
if arg_part.getRelTreeRoot().getId() != cp.getRelTreeRoot(
).getId():
rNewArgClustIdx_argCnt = inc_key(rNewArgClustIdx_argCnt,
aci)
if aci not in rNewArgClustIdx_argTypeIdx_cnt:
rNewArgClustIdx_argTypeIdx_cnt[aci] = {}
rNewArgClustIdx_argTypeIdx_cnt[aci] = \
inc_key(rNewArgClustIdx_argTypeIdx_cnt[aci], ati)
if aci not in rNewArgClustIdx_chdClustIdx_cnt:
rNewArgClustIdx_chdClustIdx_cnt[aci] = {}
rNewArgClustIdx_chdClustIdx_cnt[aci] = \
inc_key(rNewArgClustIdx_chdClustIdx_cnt[aci], child_clust_id)
args = cp.getArguments()
for ai, arg in args.items():
ap = arg._argPart
cci = ap._clustIdx
aci = cp.getArgClust(ai)
ac = acl._argClusts[aci]
ati = arg._path.getArgType()
# Drop the old arguments
aArgClustIdx_argCnt = dec_key(aArgClustIdx_argCnt,
aci,
base=ac._ttlArgCnt)
if aci not in aArgClustIdx_argTypeIdx_cnt:
aArgClustIdx_argTypeIdx_cnt[aci] = {}
aArgClustIdx_argTypeIdx_cnt[aci] = \
dec_key(aArgClustIdx_argTypeIdx_cnt[aci],
ati,
base=ac._argTypeIdx_cnt[ati])
if aci not in aArgClustIdx_chdClustIdx_cnt:
aArgClustIdx_chdClustIdx_cnt[aci] = dict()
aArgClustIdx_chdClustIdx_cnt[aci] = \
dec_key(aArgClustIdx_chdClustIdx_cnt[aci],
cci,
base=ac._chdClustIdx_cnt[cci])
# Add the new arguments
aNewArgClustIdx_argCnt = inc_key(aNewArgClustIdx_argCnt, aci)
if aci not in aNewArgClustIdx_argTypeIdx_cnt:
aNewArgClustIdx_argTypeIdx_cnt[aci] = {}
aNewArgClustIdx_argTypeIdx_cnt[aci] = \
inc_key(aNewArgClustIdx_argTypeIdx_cnt[aci], ati)
if aci not in aNewArgClustIdx_chdClustIdx_cnt:
aNewArgClustIdx_chdClustIdx_cnt[aci] = {}
aNewArgClustIdx_chdClustIdx_cnt[aci] = \
inc_key(aNewArgClustIdx_chdClustIdx_cnt[aci], cci)
if raRootCnt > 0:
origRootCnt = Clust.clustIdx_rootCnt[rcidx]
if origRootCnt > raRootCnt:
score += xlogx(raRootCnt) \
+ xlogx(origRootCnt - raRootCnt) \
- xlogx(origRootCnt)
score -= ParseParams.priorNumParam
denomor = xlogx(rcl._ttlCnt)
denomnr = xlogx(rtc_new)
score = update_score_from_dict(score, rRelTypeIdx_newcnt,
rcl._relTypeIdx_cnt)
score += denomor
score -= denomnr
denomoa = xlogx(acl._ttlCnt)
denomna = xlogx(atc_new)
score = update_score_from_dict(score, aRelTypeIdx_newcnt,
acl._relTypeIdx_cnt)
score += denomoa
score -= denomna
for cnt in raRelTypeIdx_newcnt.values():
score -= ParseParams.priorNumParam
score += xlogx(cnt)
denomra = xlogx(ratc_new)
score -= denomra
for pi, cnt in parArg_cnt.items():
pc = Clust.getClust(pi[0])
ac = pc._argClusts[pi[1]]
origcnt = ac._chdClustIdx_cnt[rcidx]
if cnt == origcnt:
continue
score -= ParseParams.priorNumParam
score += xlogx(cnt) + xlogx(origcnt - cnt) - xlogx(origcnt)
for aci, ac in rcl._argClusts.items():
origPartCnt = len(ac._partRootTreeNodeIds)
score -= (xlogx(rcl._ttlCnt - origPartCnt) - denomor)
if aci not in rArgClustIdx_partCnt:
score += (xlogx(rtc_new - origPartCnt) - denomnr)
continue
if rArgClustIdx_partCnt[aci] > 0:
score += (xlogx(rtc_new - rArgClustIdx_partCnt[aci]) - denomnr)
score = update_score_from_dict(score, rArgClustIdx_argNum_cnt[aci],
ac._argNum_cnt)
score -= 2 * (xlogx(rArgClustIdx_argCnt[aci]) -
xlogx(ac._ttlArgCnt))
score = update_score_from_dict(score,
rArgClustIdx_argTypeIdx_cnt[aci],
ac._argTypeIdx_cnt)
score = update_score_from_dict(score,
rArgClustIdx_chdClustIdx_cnt[aci],
ac._chdClustIdx_cnt)
# line 570 in Scorer.java
for aci, ac in acl._argClusts.items():
origPartCnt = len(ac._partRootTreeNodeIds)
score -= (xlogx(acl._ttlCnt - origPartCnt) - denomoa)
if aci not in aArgClustIdx_partCnt:
score += (xlogx(atc_new - origPartCnt) - denomna)
continue
if aArgClustIdx_partCnt[aci] > 0:
score += (xlogx(atc_new - aArgClustIdx_partCnt[aci]) - denomna)
score = update_score_from_dict(score, aArgClustIdx_argNum_cnt[aci],
ac._argNum_cnt)
score -= 2 * (xlogx(aArgClustIdx_argCnt[aci]) -
xlogx(ac._ttlArgCnt))
score = update_score_from_dict(score,
aArgClustIdx_argTypeIdx_cnt[aci],
ac._argTypeIdx_cnt)
score = update_score_from_dict(score,
aArgClustIdx_chdClustIdx_cnt[aci],
ac._chdClustIdx_cnt)
for ds in [(rNewArgClustIdx_partCnt, rNewArgClustIdx_argNum_cnt),
(aNewArgClustIdx_partCnt, aNewArgClustIdx_argNum_cnt)]:
for aci, partCnt in ds[0].items():
score += xlogx(ratc_new - partCnt) - denomra
for idx, cnt in ds[1][aci].items():
score += xlogx(cnt)
score -= ParseParams.priorNumParam
for ds in [(rNewArgClustIdx_argCnt, rNewArgClustIdx_argTypeIdx_cnt,
rNewArgClustIdx_chdClustIdx_cnt),
(aNewArgClustIdx_argCnt, aNewArgClustIdx_argTypeIdx_cnt,
aNewArgClustIdx_chdClustIdx_cnt)]:
for aci, argCnt in ds[0].items():
score -= 2 * xlogx(argCnt)
for idx, cnt in ds[1][aci].items():
score += xlogx(cnt)
score -= ParseParams.priorNumParam
for idx, cnt in ds[2][aci].items():
score += xlogx(cnt)
score -= ParseParams.priorNumParam
return score
def execMC(self, op):
#
# Get clusters associated with our op
#
cluster1 = Clust.getClust(op._clustIdx1)
cluster2 = Clust.getClust(op._clustIdx2)
if cluster1 is None or cluster2 is None:
return -1
#
# If cluster 1 has fewer argument clusters than cluster 2, swap them.
# We merge the "smaller" cluster into the larger one.
#
if len(cluster1._argClusts) < len(cluster2._argClusts):
clust_swap = cluster2
cluster2 = cluster1
cluster1 = clust_swap
#
# Align the argument clusters based on scores, and then map over
# any remaining argument clusters from cluster 2 to cluster 1.
#
aci2_aci1 = dict()
scorer = self._parse.scorer
_, aci2_aci1 = scorer.scoreMCForAlign(cluster1, cluster2, aci2_aci1)
for arg_clust_id2 in cluster2._argClusts:
if arg_clust_id2 not in aci2_aci1:
arg_clust = cluster2._argClusts[arg_clust_id2]
for arg_type in arg_clust._argTypeIdx_cnt:
arg_clust_ids = cluster1.getArgClustIdxs(arg_type)
if arg_clust_ids is None:
arg_clust_id1 = cluster1.createArgClust(arg_type)
else:
arg_clust_id1 = next(iter(arg_clust_ids))
aci2_aci1[arg_clust_id2] = arg_clust_id1
break
#
# Finally, remap the Parts in cluster 2 to cluster 1 as well.
#
part_ids = set()
part_ids.update(Part.getPartRootNodeIds(cluster2.getId()))
for part_id in part_ids:
pt = Part.getPartByRootNodeId(part_id)
for arg in pt.getArguments().values():
arg._argPart.unsetParent()
pt.changeClustRemap(cluster1.getId(), aci2_aci1)
for argIdx, arg in pt.getArguments().items():
arg._argPart.setParent(pt, argIdx)
Clust.removeClust(cluster2)
return cluster1.getId()
def scoreOpComposePart(self, pp, cp):
score = 0
rcl, acl = [Clust.getClust(x._clustIdx) for x in (pp, cp)]
ptn, ctn = [x._relTreeRoot for x in (pp, cp)]
ptn.addChild(dep, ctn)
nrti = RelType.getRelType(ptn)
if Clust.getClustsWithRelType(nrti) is None:
return score
pai = cp._parArgIdx
pcarg = pp.getArguments()[pai]
dep = pcarg._path.getDep()
orti = pp._relTypeIdx
ncl = Clust.getClust(Clust.getClustsWithRelType(nrti).next())
nci = ncl._clustIdx
if pp.getParPart() is not None:
ppp = pp.getParPart()
ppcl = Clust.getClust(ppp.getClustIdx())
ac = ppcl._argClusts[ppp.getArgClust(pp.getParArgIdx())]
oc = ac._chdClustIdx_cnt[rcl._clustIdx]
nc = ac._chdClustIdx_cnt[nci]
else:
oc = Clust.clustIdx_rootCnt[rcl]
nc = Clust.clustIdx_rootCnt[ncl]
score += log(nc) - log(oc)
for aci, ais in pp._argClustIdx_argIdxs.items():
ac = rcl._argClusts[aci]
score -= (log(ac._argNum_cnt[len(ais)]) - log(ac._ttlArgCnt))
for ai in ais:
arg = pp.getArgument(ai)
score -= (log(ac._chdClustIdx_cnt[arg._argPart._clustIdx]) \
- log(ac._ttlArgCnt))
score -= (log(ac._argTypeIdx_cnt[arg._path.getArgType()]) \
- log(ac._ttlArgCnt))
ai_newaci = dict()
for ai, arg in pp._args.items():
if ai == pai:
pass
else:
ati = arg._path.getArgType()
aci = ncl._argTypeIdx_argClustIdxs[ati].next()
ai_newaci[ai] = aci
newArgClustIdx_ais = dict()
for ai, aci in ai_newaci.items():
if aci not in newArgClustIdx_ais:
newArgClustIdx_ais[aci] = set()
newArgClustIdx_ais[aci].add(ais)
for aci, ais in newArgClustIdx_ais.items():
ac = ncl._argClusts[aci]
score += (log(ac._argNum_cnt[len(ais)]) - log(ac._ttlArgCnt))
for ai in ais:
arg = pp.getArgument(ai)
score -= (log(ac._chdClustIdx_cnt[arg._argPart._clustIdx]) \
- log(ac._ttlArgCnt))
score -= (log(ac._argTypeIdx_cnt[arg._path.getArgType()]) \
- log(ac._ttlArgCnt))
return score
def scoreOpMC(self, op):
# Get our two cluster ids, and make sure cluster 1 was defined earlier
# than cluster 2.
clust_id1, clust_id2 = op._clustIdx1, op._clustIdx2
assert clust_id1 < clust_id2
# Subtract 0 from score? Weird.
score = 0 - ParseParams.priorMerge
# If these clusters appear in conjunction with each other (indicating a
# dissimilarity, otherwise it's redundant) penalize the score.
#
# Clust.pairClustIdx_conjCnt is dictionary of type {(int, int): int}
#
if (clust_id1, clust_id2) in Clust.pairClustIdx_conjCnt:
score -= (ParseParams.priorNumConj \
* Clust.pairClustIdx_conjCnt[(clust_id1, clust_id2)])
# Now get the actual Clust objects
clust1 = Clust.getClust(clust_id1)
clust2 = Clust.getClust(clust_id2)
#
# We update the score by taking xlogx(x+y) - xlogx(x) - xlogx(y)
# where xlogx() == x*log(x)
# Here we calculate on total counts - basically how common these clusters
# are in our corpus. Here we penalize if they are too common.
score -= Scorer.updateScore(clust1._ttlCnt, clust2._ttlCnt)
#
# Then we check for shared relTypes, and score up those with lots of
# shared relTypes.
for reltype, count1 in clust1._relTypeIdx_cnt.items():
if reltype in clust2._relTypeIdx_cnt:
count2 = clust2._relTypeIdx_cnt[reltype]
score += Scorer.updateScore(count1, count2)
score += ParseParams.priorNumParam
#
# Bonus as well if we have lots of sentence roots in these clusters,
# indicating they are semantically important.
if clust_id1 in Clust.clustIdx_rootCnt and clust_id2 in Clust.clustIdx_rootCnt:
root_count1 = Clust.clustIdx_rootCnt[clust_id1]
root_count2 = Clust.clustIdx_rootCnt[clust_id2]
score += Scorer.updateScore(root_count1, root_count2)
score += ParseParams.priorNumParam
#
# Let's compare the parent components as well
#
score += self.scoreMCForParent(clust_id1, clust_id2)
#
# Finally, if cluster 2 has more arguments than cluster 1, reverse them
# before scoring on alignment of their arguments
if len(clust2._argClusts) > len(clust1._argClusts):
clx1 = clust2
clx2 = clust1
else:
clx1 = clust1
clx2 = clust2
score_add, _ = self.scoreMCForAlign(clx1, clx2, dict())
score += score_add
return score
def createArgs(self,
art_id,
sent_id,
sent,
parent_id,
done=set(),
verbose=False):
'''
For each token, get the TreeNode, Part, Cluster and (based on
sentence dependencies) the children tokens.
For each child token, use the dependency relationship to define
a Path and then argument type and Argument defining the parent-
child relationship. Then add/create an ArgClust before recursing
on any grand-child tokens.
#
## CHECK TOKENS SO WE DON'T GET STUCK IN A RECURSIVE LOOP IF
## DEPENDENCIES ARE MALFORMED
#
'''
parent_node_id = genTreeNodeID(art_id, sent_id, parent_id)
parent = TreeNode.getTreeNode(parent_node_id)
parent_part = Part.getPartByRootNodeId(parent_node_id)
parent_clust = Clust.getClust(parent_part.getClustIdx())
children = sent.get_children(parent_id)
if children is not None:
for relation, child_id in children:
child_node_id = genTreeNodeID(art_id, sent_id, child_id)
path = Path(relation)
arg_type_id = path.getArgType()
# if child_node_id in done:
# continue
child_part = Part.getPartByRootNodeId(child_node_id)
if child_part is None:
if verbose:
print("Child node id {} has no part".format(
child_node_id))
if child_part.getParPart() is not None:
if verbose:
print("Child node id {} already has "
"parent {}".format(
child_node_id,
child_part.getParPart().getRelTreeRoot().
getId()))
continue
arg = Argument(parent, path, child_part)
arg_id = parent_part.addArgument(arg)
child_part.setParent(parent_part, arg_id)
arg_clust_ids = parent_clust.getArgClustIdxs(arg_type_id)
if arg_clust_ids is None:
arg_clust_id = parent_clust.createArgClust(arg_type_id)
else:
arg_clust_id = next(iter(arg_clust_ids))
parent_part.setArgClust(arg_id, arg_clust_id)
#done.add(child_node_id)
self.createArgs(art_id, sent_id, sent, child_id)
#done.add(parent_node_id)
return None
def setRelTypeIdx(self, newRelTypeIdx):
self._relTypeIdx = newRelTypeIdx
cl = Clust.getClust(self._clustIdx)
cl.onPartSetRelTypeIdx(newRelTypeIdx)
return None
def execCompose(self, op):
parClustIdx = op._parClustIdx
chdClustIdx = op._chdClustIdx
new_clust_id = -1
#
# If either cluster are None, return -1
#
if Clust.getClust(parClustIdx) is None or Clust.getClust(
chdClustIdx) is None:
return -1
new_clust = None
parent_child_pair = (parClustIdx, chdClustIdx)
part_ids = set()
part_ids.update(
Part.pairClustIdxs_pairPartRootNodeIds[parent_child_pair])
deleted_parts = []
for parent_id, child_id in part_ids:
if parent_id in deleted_parts or child_id in deleted_parts:
continue
parent_part = Part.getPartByRootNodeId(parent_id)
child_part = Part.getPartByRootNodeId(child_id)
dep = parent_part.getArguments()[
child_part._parArgIdx]._path.getDep()
parent_part._relTreeRoot.addChild(dep, child_part._relTreeRoot)
nrti = RelType.getRelType(parent_part._relTreeRoot)
if new_clust is None:
# on first loop
rel_clusts = Clust.getClustsWithRelType(nrti)
if rel_clusts is None:
new_clust = Clust.getClust(Clust.createClust(nrti))
elif len(rel_clusts) > 1:
raise Exception
else:
new_clust = Clust.getClust(next(iter(rel_clusts)))
new_clust_id = new_clust.getId()
parent_part.removeArgument(child_part._parArgIdx)
if parent_part.getClustIdx() != new_clust_id:
for argIdx in parent_part.getArguments():
parent_part.unsetArgClust(argIdx)
arg = parent_part.getArgument(argIdx)
arg._argPart.unsetParent()
parent_part.changeClust(new_clust_id, nrti)
for argIdx, arg in parent_part.getArguments().items():
arg_type = arg._path.getArgType()
arg_clust_id = -1
if arg_type not in new_clust._argTypeIdx_argClustIdxs:
arg_clust_id = new_clust.createArgClust(arg_type)
elif len(
new_clust._argTypeIdx_argClustIdxs[arg_type]) == 0:
arg_clust_id = new_clust.createArgClust(arg_type)
else:
arg_clust_id = next(
iter(new_clust._argTypeIdx_argClustIdxs[arg_type]))
arg._argPart.setParent(parent_part, argIdx)
parent_part.setArgClust(argIdx, arg_clust_id)
parent_part.setRelTypeIdx(nrti)
else:
parent_part.unsetRelTypeIdx()
parent_part.setRelTypeIdx(nrti)
#
# Connect the child part's arguments directly to the parent part now
#
for argIdx, arg in child_part.getArguments().items():
child_part.unsetArgClust(argIdx)
arg_type = arg._path.getArgType()
arg_clust_id = -1
if arg_type not in new_clust._argTypeIdx_argClustIdxs:
arg_clust_id = new_clust.createArgClust(arg_type)
elif len(new_clust._argTypeIdx_argClustIdxs[arg_type]) == 0:
arg_clust_id = new_clust.createArgClust(arg_type)
else:
arg_clust_id = next(
iter(new_clust._argTypeIdx_argClustIdxs[arg_type]))
newArgIdx = parent_part.addArgument(arg)
arg._argPart.setParent(parent_part, newArgIdx)
parent_part.setArgClust(newArgIdx, arg_clust_id)
#
# Remove the old child part
#
deleted_parts.append(child_part.getRelTreeRoot().getId())
child_part.destroy()
# Part.clustIdx_pairClustIdxs[parClustIdx].remove(pci)
# Part.clustIdx_pairClustIdxs[chdClustIdx].remove(pci)
del Part.pairClustIdxs_pairPartRootNodeIds[parent_child_pair]
return new_clust_id