def _toforest(self, lmforest, sent, cache, level=0):
if self in cache:
return cache[self]
this_node = Node("", "X [0-0]", 1, Vector(), sent) # no node fv, temporary iden=""
is_root = level == 0 # don't include final </s> </s>
for prev_states, extra_fv, extra_words in self.backptrs:
prev_nodes = [p._toforest(lmforest, sent, cache, level+1) for p in prev_states if p is not None]
if is_root:
extra_words = extra_words[:-LMState.lm.order+1]
edge = Hyperedge(this_node, prev_nodes, extra_fv,
prev_nodes + LMState.lm.ppqstr(extra_words).split())
edge.rule = Rule("a(\"a\")", "b", "")
edge.rule.ruleid = 1
this_node.add_edge(edge)
cache[self] = this_node
this_node.iden = str(len(cache)) # post-order id
lmforest.add_node(this_node)
if is_root:
lmforest.root = this_node
return this_node
def recover_oracle(self):
'''oracle is already stored implicitly in the forest
returns best_score, best_parseval, best_tree, edgelist
'''
edgelist = self.root.get_oracle_edgelist()
fv = Hyperedge.deriv2fvector(edgelist)
tr = Hyperedge.deriv2tree(edgelist)
bleu_p1 = self.bleu.rescore(tr)
return bleu_p1, tr, fv, edgelist
def extract_oracle(forest):
"""oracle is already stored implicitly in the forest
returns best_score, best_parseval, best_tree, edgelist
"""
global implicit_oracle
implicit_oracle = True
edgelist = get_edgelist(forest.root)
fv = Hyperedge.deriv2fvector(edgelist)
tr = Hyperedge.deriv2tree(edgelist)
return fv[0], Parseval(), tr, edgelist
def start_state(root):
''' None -> <s>^{g-1} . TOP </s>^{g-1} '''
## LMState.cache = {}
lmstr = LMState.lm.raw_startsyms()
lhsstr = lmstr + [root] + LMState.lm.raw_stopsyms()
edge = Hyperedge(None, [root], Vector(), lhsstr)
edge.lmlhsstr = LMState.lm.startsyms() + [root] + LMState.lm.stopsyms()
edge.rule = Rule.parse("ROOT(TOP) -> x0 ### ")
sc = root.bestres[0] if FLAGS.futurecost else 0
return LMState(None, [DottedRule(edge, dot=len(lmstr))], LMState.lm.startsyms(),
step=0, score=sc) # future cost
def _oracle(self, forest):
sc, parseval, tr, edgelist = forest_oracle(forest, forest.goldtree)
forest.oracle_tree = tr
forest.oracle_fvector = Hyperedge.deriv2fvector(edgelist)
if Decoder.MAX_NUM_BRACKETS < 0:
forest.oracle_size_ratio = 1
else:
forest.oracle_size_ratio = len(tr.all_label_spans()) / Decoder.MAX_NUM_BRACKETS
def add_lex_th(self, lhs, node, isBP):
''' add lexical translation rules '''
ruleset = self.ruleset
if lhs in ruleset:
rules = ruleset[lhs]
# add all translation hyperedges
for rule in rules:
newrhs = [x[1:-1] for x in rule.rhs]
tfedge = Hyperedge(node, [], Vector(rule.fields), newrhs)
tfedge.rule = rule
node.edges.append(tfedge)
elif not isBP: # add a default translation hyperedge (monotonic)
defword = '"%s"' % node.word
rule = Rule(lhs, [defword], self.deffields)
tfedge = Hyperedge(node, [], Vector(self.deffields), [defword[1:-1]])
tfedge.rule = rule
ruleset.add_rule(rule)
node.edges.append(tfedge)
def load(filename, lower=True, sentid=0):
'''now return a generator! use load().next() for singleton.
and read the last line as the gold tree -- TODO: optional!
and there is an empty line at the end
'''
file = getfile(filename)
line = None
total_time = 0
num_sents = 0
while True:
start_time = time.time()
##'\tThe complicated language in ...\n"
## tag is often missing
try:
if line is None or line == "\n":
line = "\n"
while line == "\n":
line = file.readline() # emulate seek
tag, sent = line.split("\t")
except:
## no more forests
break
num_sents += 1
sent = sent.split()
cased_sent = sent [:]
if lower:
sent = [w.lower() for w in sent] # mark johnson: lowercase all words
num = int(file.readline())
forest = Forest(num, sent, cased_sent, tag)
forest.labelspans = {}
forest.short_edges = {}
delta = num_spu = 0
for i in xrange(1, num+1):
## '2\tDT* [0-1]\t1 ||| 1232=2 ...\n'
## node-based features here: wordedges, greedyheavy, word(1), [word(2)], ...
line = file.readline()
try:
keys, fields = line.split(" ||| ")
except:
keys = line
fields = ""
iden, labelspan, size = keys.split("\t") ## iden can be non-ints
size = int(size)
fvector = FVector.parse(fields)
node = Node(iden, labelspan, size, fvector, sent)
forest.add_node(node)
if cache_same:
if labelspan in forest.labelspans:
node.same = forest.labelspans[labelspan]
node.fvector = node.same.fvector
else:
forest.labelspans[labelspan] = node
for j in xrange(size):
is_oracle = False
## '\t1 ||| 0=8.86276 1=2 3\n'
tails, fields = file.readline().strip().split(" ||| ")
if tails[0] == "*": #oracle edge
is_oracle = True
tails = tails[1:]
tails = tails.split() ## could be non-integers
tailnodes = []
for x in tails:
assert x in forest.nodes, "BAD TOPOL ORDER: node #%s is referred to " % x + \
"(in a hyperedge of node #%s) before being defined" % iden
## topological ordering
tail = forest.nodes[x]
tailnodes.append(tail)
use_same = False
if fields[-1] == "~":
use_same = True
fields = fields[:-1]
fvector = FVector.parse(fields)
edge = Hyperedge(node, tailnodes, fvector)
if cache_same:
short_edge = edge.shorter()
if short_edge in forest.short_edges:
edge.same = forest.short_edges[short_edge]
if use_same:
edge.fvector += edge.same.fvector
else:
forest.short_edges[short_edge] = edge
node.add_edge(edge)
if is_oracle:
node.oracle_edge = edge
if node.sp_terminal():
node.word = node.edges[0].subs[0].word
## splitted nodes 12-3-4 => (12, 3, 4)
tmp = sorted([(map(int, x.iden.split("-")), x) for x in forest.nodeorder])
forest.nodeorder = [x for (_, x) in tmp]
forest.rehash()
sentid += 1
## print >> logs, "sent #%d %s, %d words, %d nodes, %d edges, loaded in %.2lf secs" \
## % (sentid, forest.tag, forest.len, num, forest.num_edges, time.time() - basetime)
forest.root = node
node.set_root(True)
line = file.readline()
if line is not None and line.strip() != "":
if line[0] == "(":
forest.goldtree = Tree.parse(line.strip(), trunc=True, lower=True)
line = file.readline()
else:
line = None
total_time += time.time() - start_time
if num_sents % 100 == 0:
print >> logs, "... %d sents loaded (%.2lf secs per sent) ..." \
% (num_sents, total_time/num_sents)
yield forest
Forest.load_time = total_time
print >> logs, "%d forests loaded in %.2lf secs (avg %.2lf per sent)" \
% (num_sents, total_time, total_time/num_sents)
def load(filename, lower=False, sentid=0):
'''now return a generator! use load().next() for singleton.
and read the last line as the gold tree -- TODO: optional!
and there is an empty line at the end
'''
file = getfile(filename)
line = None
total_time = 0
num_sents = 0
while True:
start_time = time.time()
##'\tThe complicated language in ...\n"
## tag is often missing
try:
if line is None or line == "\n":
line = "\n"
while line == "\n":
line = file.readline() # emulate seek
tag, sent = line.split("\t")
except:
## no more forests
break
num_sents += 1
sent = sent.split()
cased_sent = sent[:]
if lower:
sent = [w.lower()
for w in sent] # mark johnson: lowercase all words
num = int(file.readline())
forest = Forest(num, sent, cased_sent, tag)
forest.labelspans = {}
forest.short_edges = {}
delta = num_spu = 0
for i in xrange(1, num + 1):
## '2\tDT* [0-1]\t1 ||| 1232=2 ...\n'
## node-based features here: wordedges, greedyheavy, word(1), [word(2)], ...
line = file.readline()
try:
keys, fields = line.split(" ||| ")
except:
keys = line
fields = ""
iden, labelspan, size = keys.split(
"\t") ## iden can be non-ints
size = int(size)
fvector = FVector(fields) # TODO: myvector
node = Node(iden, labelspan, size, fvector, sent)
forest.add_node(node)
if cache_same:
if labelspan in forest.labelspans:
node.same = forest.labelspans[labelspan]
node.fvector = node.same.fvector
else:
forest.labelspans[labelspan] = node
for j in xrange(size):
is_oracle = False
## '\t1 ||| 0=8.86276 1=2 3\n'
tails, fields = file.readline().strip().split(" ||| ")
if tails[0] == "*": #oracle edge
is_oracle = True
tails = tails[1:]
tails = tails.split() ## could be non-integers
tailnodes = []
for x in tails:
assert x in forest.nodes, "BAD TOPOL ORDER: node #%s is referred to " % x + \
"(in a hyperedge of node #%s) before being defined" % iden
## topological ordering
tail = forest.nodes[x]
tailnodes.append(tail)
use_same = False
if fields[-1] == "~":
use_same = True
fields = fields[:-1]
fvector = FVector(fields)
edge = Hyperedge(node, tailnodes, fvector)
if cache_same:
short_edge = edge.shorter()
if short_edge in forest.short_edges:
edge.same = forest.short_edges[short_edge]
if use_same:
edge.fvector += edge.same.fvector
else:
forest.short_edges[short_edge] = edge
node.add_edge(edge)
if is_oracle:
node.oracle_edge = edge
if node.sp_terminal():
node.word = node.edges[0].subs[0].word
## splitted nodes 12-3-4 => (12, 3, 4)
tmp = sorted([(map(int, x.iden.split("-")), x)
for x in forest.nodeorder])
forest.nodeorder = [x for (_, x) in tmp]
forest.rehash()
sentid += 1
## print >> logs, "sent #%d %s, %d words, %d nodes, %d edges, loaded in %.2lf secs" \
## % (sentid, forest.tag, forest.len, num, forest.num_edges, time.time() - basetime)
forest.root = node
node.set_root(True)
line = file.readline()
if line is not None and line.strip() != "":
if line[0] == "(":
## forest.goldtree = Tree.parse(line.strip(), trunc=True, lower=False)
line = file.readline()
else:
line = None
total_time += time.time() - start_time
if num_sents % 100 == 0:
print >> logs, "... %d sents loaded (%.2lf secs per sent) ..." \
% (num_sents, total_time/num_sents)
yield forest
Forest.load_time = total_time
if num_sents > 0:
print >> logs, "%d forests loaded in %.2lf secs (avg %.2lf per sent)" \
% (num_sents, total_time, total_time/num_sents)
def forest_oracle(forest, goldtree, del_puncs=False, prune_results=False):
""" returns best_score, best_parseval, best_tree, edgelist
now non-recursive topol-sort-style
"""
if hasattr(forest.root, "oracle_edge"):
return extract_oracle(forest)
## modifies forest also!!
if del_puncs:
idx_mapping, newforest = check_puncs(forest, goldtree.tag_seq)
else:
idx_mapping, newforest = lambda x: x, forest
goldspans = merge_labels(goldtree.all_label_spans(), idx_mapping)
goldbrs = set(goldspans) ## including TOP
for node in newforest:
if node.is_terminal():
results = Oracles.unit("(%s %s)" % (node.label, node.word)) ## multiplication unit
else:
a, b = (
(0, 0)
if node.is_spurious()
else ((1, 1) if (merge_label((node.label, node.span), idx_mapping) in goldbrs) else (1, 0))
)
label = "" if node.is_spurious() else node.label
results = Oracles() ## addition unit
for edge in node.edges:
edgeres = Oracles.unit() ## multiplication unit
for sub in edge.subs:
assert hasattr(sub, "oracles"), "%s ; %s ; %s" % (node, sub, edge)
edgeres = edgeres * sub.oracles
## nodehead = (a, RES((b, -edge.fvector[0], label, [edge]))) ## originally there is label
assert 0 in edge.fvector, edge
nodehead = (a, RES((b, -edge.fvector[0], [edge])))
results += nodehead * edgeres ## mul
if prune_results:
prune(results)
node.oracles = results
if debug:
print >> logs, node.labelspan(), "\n", results, "----------"
res = (-1, RES((-1, 0, []))) * newforest.root.oracles ## scale, remove TOP match
num_gold = len(goldspans) - 1 ## omit TOP. N.B. goldspans, not brackets! (NP (NP ...))
best_parseval = None
for num_test in res:
## num_matched, score, tree_str, edgelist = res[num_test]
num_matched, score, edgelist = res[num_test]
this = Parseval.get_parseval(num_matched, num_test, num_gold)
if best_parseval is None or this < best_parseval:
best_parseval = this
best_score = score
## best_tree = tree_str
best_edgelist = edgelist
best_tree = Hyperedge.deriv2tree(best_edgelist)
## annotate the forest for oracle so that next-time you can preload oracle
for edge in best_edgelist:
edge.head.oracle_edge = edge
## very careful here: desymbol !
## return -best_score, best_parseval, Tree.parse(desymbol(best_tree)), best_edgelist
return -best_score, best_parseval, best_tree, best_edgelist
def load(filename, is_tforest=False, lower=False, sentid=0, first=None, lm=None):
'''now returns a generator! use load().next() for singleton.
and read the last line as the gold tree -- TODO: optional!
and there is an empty line at the end
'''
if first is None: # N.B.: must be here, not in the param line (after program initializes)
first = FLAGS.first
file = getfile(filename)
line = None
total_time = 0
num_sents = 0
while True:
start_time = time.time()
##'\tThe complicated language in ...\n"
## tag is often missing
line = file.readline() # emulate seek
if len(line) == 0:
break
try:
## strict format, no consecutive breaks
# if line is None or line == "\n":
# line = "\n"
# while line == "\n":
# line = file.readline() # emulate seek
tag, sent = line.split("\t") # foreign sentence
except:
## no more forests
yield None
continue
num_sents += 1
# caching the original, word-based, true-case sentence
sent = sent.split() ## no splitting with " "
cased_sent = sent [:]
if lower:
sent = [w.lower() for w in sent] # mark johnson: lowercase all words
#sent = words_to_chars(sent, encode_back=True) # split to chars
## read in references
refnum = int(file.readline().strip())
refs = []
for i in xrange(refnum):
refs.append(file.readline().strip())
## sizes: number of nodes, number of edges (optional)
num, nedges = map(int, file.readline().split("\t"))
forest = Forest(sent, cased_sent, tag, is_tforest)
forest.tag = tag
forest.refs = refs
forest.bleu = Bleu(refs=refs) ## initial (empty test) bleu; used repeatedly later
forest.labelspans = {}
forest.short_edges = {}
forest.rules = {}
for i in xrange(1, num+1):
## '2\tDT* [0-1]\t1 ||| 1232=2 ...\n'
## node-based features here: wordedges, greedyheavy, word(1), [word(2)], ...
line = file.readline()
try:
keys, fields = line.split(" ||| ")
except:
keys = line
fields = ""
iden, labelspan, size = keys.split("\t") ## iden can be non-ints
size = int(size)
fvector = Vector(fields) #
## remove_blacklist(fvector)
node = Node(iden, labelspan, size, fvector, sent)
forest.add_node(node)
if cache_same:
if labelspan in forest.labelspans:
node.same = forest.labelspans[labelspan]
node.fvector = node.same.fvector
else:
forest.labelspans[labelspan] = node
for j in xrange(size):
is_oracle = False
## '\t1 ||| 0=8.86276 1=2 3\n'
## N.B.: can't just strip! "\t... ||| ... ||| \n" => 2 fields instead of 3
tails, rule, fields = file.readline().strip("\t\n").split(" ||| ")
if tails != "" and tails[0] == "*": #oracle edge
is_oracle = True
tails = tails[1:]
tails = tails.split() ## N.B.: don't split by " "!
tailnodes = []
lhsstr = [] # 123 "thank" 456
lmstr = []
lmscore = 0
lmlhsstr = []
for x in tails:
if x[0]=='"': # word
word = desymbol(x[1:-1])
lhsstr.append(word) ## desymbol here and only here; ump will call quoteattr
if lm is not None:
this = lm.word2index(word)
lmscore += lm.ngram.wordprob(this, lmstr)
lmlhsstr.append(this)
lmstr += [this,]
else: # variable
assert x in forest.nodes, "BAD TOPOL ORDER: node #%s is referred to " % x + \
"(in a hyperedge of node #%s) before being defined" % iden
tail = forest.nodes[x]
tailnodes.append(tail)
lhsstr.append(tail)
if lm is not None:
lmstr = [] # "..." "..." x0 "..."
lmlhsstr.append(tail) # sync with lhsstr
fvector = Vector(fields)
if lm is not None:
fvector["lm1"] = lmscore # hack
edge = Hyperedge(node, tailnodes, fvector, lhsstr)
edge.lmlhsstr = lmlhsstr
## new
x = rule.split()
edge.ruleid = int(x[0])
if len(x) > 1:
edge.rule = Rule.parse(" ".join(x[1:]) + " ### " + fields)
forest.rules[edge.ruleid] = edge.rule #" ".join(x[1:]) #, None)
else:
edge.rule = forest.rules[edge.ruleid] # cahced rule
node.add_edge(edge)
if is_oracle:
node.oracle_edge = edge
if node.sp_terminal():
node.word = node.edges[0].subs[0].word
## splitted nodes 12-3-4 => (12, 3, 4)
tmp = sorted([(map(int, x.iden.split("-")), x) for x in forest.nodeorder])
forest.nodeorder = [x for (_, x) in tmp]
forest.rehash()
sentid += 1
## print >> logs, "sent #%d %s, %d words, %d nodes, %d edges, loaded in %.2lf secs" \
## % (sentid, forest.tag, forest.len, num, forest.num_edges, time.time() - basetime)
forest.root = node
node.set_root(True)
line = file.readline()
if line is not None and line.strip() != "":
if line[0] == "(":
forest.goldtree = Tree.parse(line.strip(), trunc=True, lower=False)
line = file.readline()
else:
line = None
forest.number_nodes()
#print forest.root.position_id
total_time += time.time() - start_time
if num_sents % 100 == 0:
print >> logs, "... %d sents loaded (%.2lf secs per sent) ..." \
% (num_sents, total_time/num_sents)
forest.subtree() #compute the subtree string for each node
yield forest
if first is not None and num_sents >= first:
break
# better check here instead of zero-division exception
if num_sents == 0:
print >> logs, "NO FORESTS FOUND!!! (empty input file?)"
sys.exit(1)
# yield None # new: don't halt -- WHY?
Forest.load_time = total_time
print >> logs, "%d forests loaded in %.2lf secs (avg %.2lf per sent)" \
% (num_sents, total_time, total_time/(num_sents+0.001))
def add_nonter_th(self, node):
''' add translation hyperedges to non-terminal node '''
ruleset = self.ruleset
tfedges = []
for edge in node.edges:
# enumerate all the possible frags
basefrags = [("%s(" % node.label, [], 1)]
lastchild = len(edge.subs) - 1
if len(edge.subs) >= 5: # this guy has too many children! it cannot be matched!
deflhs = "%s(%s)" % (node.label, " ".join(sub.label for sub in edge.subs))
defrhs = edge.subs
defheight = 1
basefrags = [(deflhs, defrhs, defheight)]
else:
for (id, sub) in enumerate(edge.subs):
oldfrags = basefrags
# cross-product
basefrags = [PatternMatching.combinetwofrags(oldfrag, frag, id, lastchild, self.max_height) \
for oldfrag in oldfrags for frag in sub.frags]
# for each frag add translation hyperedges
for extfrag in basefrags:
(extlhs, extrhs, extheight) = extfrag
# add frags
if extheight <= self.max_height - 1:
node.frags.append(extfrag)
if self.filter:
self.all_lhss.add(extlhs)
else:
# add translation hyperedges
if extlhs in ruleset:
for des in extrhs:
self.descendants[node.iden].add(des.iden) #unit(set(extrhs))
#print self.descendants[node.iden]
rules = ruleset[extlhs]
# add all translation hyperedges
for rule in rules:
rhsstr = [x[1:-1] if x[0]=='"' \
else extrhs[int(x.split('x')[1])] \
for x in rule.rhs]
tfedge = Hyperedge(node, extrhs,\
Vector(rule.fields), rhsstr)
tfedge.rule = rule
tfedges.append(tfedge)
if (not self.filter) and (len(tfedges) == 0): # no translation hyperedge
for des in edge.subs:
self.descendants[node.iden].add(des.iden) #unit(set(edge.subs))
# add a default translation hyperedge
deflhs = "%s(%s)" % (node.label, " ".join(sub.label for sub in edge.subs))
defrhs = ["x%d" % i for i, _ in enumerate(edge.subs)] # N.B.: do not supply str
defrule = Rule(deflhs, defrhs, self.deffields)
tfedge = Hyperedge(node, edge.subs,\
Vector(self.deffields), edge.subs)
tfedge.rule = defrule
ruleset.add_rule(defrule)
tfedges.append(tfedge)
if not self.filter:
# inside replace
node.edges = tfedges
