def build_branch(self, rule):
if isinstance(rule, UnaryRule):
return Tree(rule.parent, [Tree(rule.child)])
elif isinstance(rule, BinaryRule):
return Tree(rule.parent,
[Tree(rule.left_child),
Tree(rule.right_child)])
def back_trace(self, back, label, i, j):
#print label, i, j
if not label in back[i][j]: # leaf
if label[0] == "#":
return Tree(label[1:])
return Tree(label)
pointer = back[i][j][label]
children = []
if type(pointer) == tuple and len(pointer) == 3:
split, left, right = pointer
left_child = self.back_trace(back, left, i, split)
right_child = self.back_trace(back, right, split, j)
children = [left_child, right_child]
"""
if right[0] == "@":
children = [left_child] + right_child.children
else:
children = [left_child, right_child]
"""
elif type(pointer) == str:
children = [self.back_trace(back, pointer, i, j)]
else:
raise ValueError, "Invalid Pointer"
return Tree(label, children)
def annotate_tree(cls, unannotated_tree):
"""
Currently, the only annotation done is a lossless binarization
"""
# TODO: change the annotation from a lossless binarization to a
# finite-order markov process (try at least 1st and 2nd order)
# mark nodes with the label of their parent nodes, giving a second
# order vertical markov process
label = unannotated_tree.label
children = unannotated_tree.children
new_children = []
if len(children) == 1:
new_children.append(children[0])
if len(children) == 2:
for n in [0, 1]:
other = (n + 1) % 2
my_labels = re.split("\^", children[n].label)
other_labels = re.split("\^", children[other].label)
new_label = my_labels[0] + '^' + other_labels[0]
new_children.append(
TreeAnnotations.annotate_tree(
Tree(new_label, child.children)))
return Tree(label, new_children)
def buildTree(self,left_pos,right_pos,label):
result = self.back[left_pos][right_pos][label]
#print "Label:%s\tResult:%s" % (label,result)
if isinstance(result,float):
return Tree(label,[Tree(self.sentence[left_pos],[])])
elif isinstance(result,tuple):
return Tree(label,[self.buildTree(left_pos,result[0],result[1]),self.buildTree(result[0],right_pos,result[2])])
else:
return Tree(label,[self.buildTree(left_pos,right_pos,result)])
def strip_leaves(self, tree):
if tree.is_leaf():
return None
if tree.is_preterminal():
return Tree(tree.label)
children = []
for child in tree.children:
children.append(self.strip_leaves(child))
return Tree(tree.label, children)
def binarize_tree(cls, tree):
label = tree.label
if tree.is_leaf():
return Tree(label)
if len(tree.children) == 1:
return Tree(label, [TreeAnnotations.binarize_tree(tree.children[0])])
intermediate_label = "@%s->" % label
intermediate_tree = TreeAnnotations.binarize_tree_helper(
tree, 0, intermediate_label)
return Tree(label, intermediate_tree.children)
def get_best_parse(self, sentence):
"""
Should return a Tree.
'sentence' is a list of strings (words) that form a sentence.
"""
# TODO: implement this method
nonterms = self.lexicon.get_all_tags()
score = collections.defaultdict(lambda:collections.defaultdict(lambda:0))
back = collections.defaultdict(lambda:collections.defaultdict(lambda:[]))
for i,w in enumerate(sentence):
for A in nonterms:
prob = self.lexicon.score_tagging(w,A)
if prob > 0:
score[(i,i+1)][A] = prob
added = True
while added:
added = False
for j in range(len(nonterms)-1):
A = nonterms[j]
B = nonterms[j+1]
prob = self.lexicon.score_tagging(B,A) * score[(i,i+1)][B]
if prob > score[(i,i+1)][A]:
score[(i,i+1)][A] = prob
back[(i,i+1)][A] = B
added = True
for span in range(2,len(sentence)):
for begin in range(len(sentence)-span):
end = begin + span
for split in range(begin+1,end-1):
for i in range(len(nonterms)-2):
A = nonterms[i]
B = nonterms[i+1]
C = nonterms[i+2]
prob = score[(begin,end)][A] * score[(split,end)][C] * self.lexicon.score_tagging(A,B)
if prob > score[(begin,end)][A]:
score[(begin,end)][A] = prob
back[(begin,end)][A] = (split,B,C)
added = True
while added:
added = False
for j in range(len(nonterms)-1):
A = nonterms[j]
B = nonterms[j+1]
prob = self.lexicon.score_tagging(A,B) * score[(begin,end)][B]
if prob > score[(begin,end)][A]:
score[(begin,end)][A] = prob
back[(begin,end)][A] = B
added = True
return Tree("ROOT",[Tree("Test",[])])
def binarize_tree(cls, tree):
label = tree.label
if tree.is_leaf():
return Tree(label)
if len(tree.children) <= 2:
children = [
TreeAnnotations.binarize_tree(child) for child in tree.children
]
return Tree(label, children)
intermediate_label = '@%s->' % label
intermediate_tree = TreeAnnotations.binarize_tree_helper(
tree, 0, intermediate_label)
return Tree(label, intermediate_tree.children)
def merge(self, left_tree, right_tree):
span = len(left_tree.get_yield()) + len(right_tree.get_yield())
maxval = max(self.span_to_categories[span].values())
for key in self.span_to_categories[span]:
if self.span_to_categories[span][key] == maxval:
most_freq_label = key
break
return Tree(most_freq_label, [left_tree, right_tree])
def make_tree(begin, end, A, depth=0):
s = ' ' * (depth)
#print s, 'make_tree(%d, %d, "%s")' % (begin, end, str(A)),
backptrs = back[begin][end][A]
#print D(backptrs)
tag = A.parent
if not backptrs:
#print s, '**', str(A), tag, A.child
return Tree(tag, [Tree(A.child)])
if len(backptrs) == 1:
[B] = backptrs
child = make_tree(begin, end, B, depth + 1)
return Tree(tag, [child])
elif len(backptrs) == 3:
[split, B, C] = backptrs
childB = make_tree(begin, split, B, depth + 1)
childC = make_tree(split, end, C, depth + 1)
return Tree(tag, [childB, childC])
def binarize_tree(cls, tree, suffix):
label = tree.label + suffix
nextsuffix = "^%s" % tree.label
if tree.is_leaf():
return Tree(tree.label)
if len(tree.children) == 1:
return Tree(
label,
[TreeAnnotations.binarize_tree(tree.children[0], nextsuffix)])
if len(tree.children) == 2:
return Tree(label, [
TreeAnnotations.binarize_tree(tree.children[0], nextsuffix),
TreeAnnotations.binarize_tree(tree.children[1], nextsuffix)
])
intermediate_label = "@%s->" % label
intermediate_tree = TreeAnnotations.binarize_tree_helper(
tree, 0, intermediate_label)
return Tree(label, intermediate_tree.children)
def binarize_tree_helper(cls, tree, num_children_generated,
intermediate_label):
left_tree = tree.children[num_children_generated]
children = []
children.append(TreeBinarization.binarize_tree(left_tree))
if num_children_generated < len(tree.children) - 1:
right_tree = TreeBinarization.binarize_tree_helper(
tree, num_children_generated + 1,
intermediate_label + "_" + left_tree.label)
children.append(right_tree)
return Tree(intermediate_label, children)
def build(self, rule):
origin = rule.origin
if origin == (None, None):
print str(rule)
return self.build_branch(rule)
else:
if isinstance(rule, UnaryRule):
# print str(rule)
tag = rule.child
next_rule = self.search_rule(origin, tag)
return Tree(rule.parent, [self.build(next_rule)])
elif isinstance(rule, BinaryRule):
# print str(rule)
left_origin, right_origin = origin[0], origin[1]
left_tag, right_tag = rule.left_child, rule.right_child
next_left_rule = self.search_rule(left_origin, left_tag)
next_right_rule = self.search_rule(right_origin, right_tag)
return Tree(rule.parent, [self.build(next_left_rule), self.build(next_right_rule)])
def binarize(self, tree):
if tree.is_leaf():
return
if len(tree.children) > 2:
new_label = "@%s_%s" % (tree.label, tree. children[0].label)
new_children = tree.children[1:]
new_subtree = Tree(new_label, new_children)
tree.children = [tree.children[0], new_subtree]
for child in tree.children:
self.binarize(child)
def binarize_tree_helper(cls, tree, num_children_generated,
intermediate_label):
left_tree = tree.children[num_children_generated]
children = []
nextsuffix = "^%s" % tree.label
children.append(TreeAnnotations.binarize_tree(left_tree, nextsuffix))
if num_children_generated < len(tree.children) - 2:
right_tree = TreeAnnotations.binarize_tree_helper(
tree, num_children_generated + 1,
intermediate_label + "_" + left_tree.label)
children.append(right_tree)
else:
right_tree = left_tree = tree.children[num_children_generated + 1]
children.append(
TreeAnnotations.binarize_tree(right_tree, nextsuffix))
return Tree(intermediate_label, children)
def build_tag_tree(self, words, tags, cur_position):
leaf_tree = Tree(words[cur_position])
tag_tree = Tree(tags[cur_position], [leaf_tree])
return tag_tree
def add_root(self, tree):
return Tree("ROOT", [tree])
def get_best_parse(self, sentence):
"""
Should return a Tree.
'sentence' is a list of strings (words) that form a sentence.
"""
# TODO: implement this method
score = collections.defaultdict(lambda: \
collections.defaultdict(lambda: \
collections.defaultdict(lambda: 0.0)))
back = collections.defaultdict(lambda: \
collections.defaultdict(lambda: \
collections.defaultdict(lambda: None)))
tags = self.lexicon.get_all_tags()
wordN = len(sentence)
i = 0
for word in sentence:
iplus = i + 1
for tag in tags:
prob = self.lexicon.score_tagging(word, tag)
if prob > score[i][iplus][tag]:
score[i][iplus][tag] = prob
word_tree = Tree(word, [])
tag_tree = Tree(tag, [word_tree])
back[i][iplus][tag] = tag_tree
added = True
while added:
added = False
for child in self.grammar.unary_rules_by_child:
for rule in self.grammar.get_unary_rules_by_child(child):
prob = score[i][iplus][child] * rule.score
parent = rule.parent
if prob > score[i][iplus][parent]:
score[i][iplus][parent] = prob
child_tree = back[i][iplus][child]
back[i][iplus][parent] = Tree(parent, [child_tree])
added = True
i = i + 1
for span in range(2, wordN + 1):
for begin in range(0, wordN + 1 - span):
end = begin + span
for split in range(begin + 1, end):
for left_child in self.grammar.binary_rules_by_left_child:
for rule in self.grammar.get_binary_rules_by_left_child(
left_child):
right_child = rule.right_child
parent = rule.parent
prob = score[begin][split][left_child] *\
score[split][end][right_child] *\
rule.score
if prob > score[begin][end][parent]:
score[begin][end][parent] = prob
left_tree = back[begin][split][left_child]
right_tree = back[split][end][right_child]
back[begin][end][parent] = \
Tree(parent, [left_tree,right_tree])
added = True
while added:
added = False
if self.grammar.get_unary_rules_by_child(
parent):
rules = self.grammar.get_unary_rules_by_child(
parent)
for rule in rules:
prob = score[begin][end][
parent] * rule.score
newparent = rule.parent
if prob > score[begin][end][newparent]:
score[begin][end][newparent] = prob
tree = back[begin][end][parent]
back[begin][end][newparent] = Tree(
newparent, [tree])
added = True
return TreeAnnotations.unannotate_tree(back[0][wordN]['ROOT'])
def tree(self):
return Tree(self.label, [Tree(self.word)])
def tree(self):
return Tree(self.label, [self.left.tree(), self.right.tree()])
def tree(self):
return Tree(self.label, [self.parent.tree()])
def add_root(self, tree):
return Tree('ROOT', [tree])
def get_best_parse(self, sentence):
"""
Should return a Tree.
'sentence' is a list of strings (words) that form a sentence.
"""
# TODO: implement this method
# initialize
length = len(sentence)
score = []
for i in xrange(length+1):
score.append([])
for j in xrange(length+1):
score[i].append(collections.defaultdict(float))
back = []
for i in xrange(length+1):
back.append([])
for j in xrange(length+1):
back[i].append(collections.defaultdict())
# process terminal symbols
for i in xrange(length):
word = sentence[i]
for tag in self.lexicon.get_all_tags():
score[i][i+1][tag] = self.lexicon.score_tagging(word, tag)
back[i][i+1][tag] = "#" + word
# handle unaries
added = True
while added:
added = False
for tag in score[i][i+1].keys():
for unary_rule in self.grammar.get_unary_rules_by_child(tag):
prob = unary_rule.score * score[i][i+1][tag]
if prob != 0 and prob > score[i][i+1][unary_rule.parent]:
score[i][i+1][unary_rule.parent] = prob
back[i][i+1][unary_rule.parent] = tag
added = True
# process nonterminal symbols
for span in xrange(2, length+1):
for begin in xrange(length+1-span):
end = begin + span
for split in xrange(begin + 1, end):
left = score[begin][split]
right = score[split][end]
for left_tag in left:
for binary_rule in self.grammar.get_binary_rules_by_left_child(left_tag):
if binary_rule.right_child in right:
prob = left[left_tag] * right[binary_rule.right_child] * binary_rule.score
if prob > score[begin][end][binary_rule.parent]:
score[begin][end][binary_rule.parent] = prob
back[begin][end][binary_rule.parent] = (split, left_tag, binary_rule.right_child)
# handle unaries
added = True
while added:
added = False
for tag in score[begin][end].keys():
for unary_rule in self.grammar.get_unary_rules_by_child(tag):
prob = unary_rule.score * score[begin][end][tag]
if prob != 0 and prob > score[begin][end][unary_rule.parent]:
score[begin][end][unary_rule.parent] = prob
back[begin][end][unary_rule.parent] = tag
added = True
"""
# debug
print sentence
print "score:"
for i in xrange(len(score)):
for j in xrange(len(score[i])):
print "(%d %d):" % (i, j), ["%s : %.02f" % (key, value) for key,value in score[i][j].items()],
print
print "back:"
for i in xrange(len(back)):
print i,
for j in xrange(len(back[i])):
print dict(back[i][j]),
print
"""
if not "ROOT" in back[0][length]:
return Tree("", [Tree("ROOT")])
result = self.back_trace(back, "ROOT", 0, length)
return TreeAnnotations.unannotate_tree(result)
