def getEdgeFeats(self, k_edge, v_edge, tag, curr_filename, my_nodes, my_edges):
"""
extract features for one edge.
cache features before conjoining with edge tag.
k_edge: (1,2), v_edge: AmrEdge
my_nodes: (1,) -> AmrNode, my_edges: (2,1) -> AmrEdge
"""
feat_vec = FeatureVector()
# edge features have been extracted
if self.curr_filename == curr_filename and k_edge in self.curr_feats:
feat_vec = self.curr_feats[k_edge]
else: # new edge
# if this is a new file, clear cache
if self.curr_filename != curr_filename:
self.curr_filename = curr_filename
self.curr_feats = {}
# extract features and add to cache
for feat_func in self.edge_feat_funcs:
feat_vec += feat_func(v_edge)
self.curr_feats[k_edge] = feat_vec
# conjoin features with tag
new_feat_vec = FeatureVector()
for k, v in feat_vec.iteritems():
new_feat_vec[(str(tag),) + k] = v
# return edge features conjoined with tag
return new_feat_vec
def getEdgeFeats(self, k_edge, v_edge, tag, curr_filename, my_nodes,
my_edges):
"""
extract features for one edge.
cache features before conjoining with edge tag.
k_edge: (1,2), v_edge: AmrEdge
my_nodes: (1,) -> AmrNode, my_edges: (2,1) -> AmrEdge
"""
feat_vec = FeatureVector()
# edge features have been extracted
if self.curr_filename == curr_filename and k_edge in self.curr_feats:
feat_vec = self.curr_feats[k_edge]
else: # new edge
# if this is a new file, clear cache
if self.curr_filename != curr_filename:
self.curr_filename = curr_filename
self.curr_feats = {}
# extract features and add to cache
for feat_func in self.edge_feat_funcs:
feat_vec += feat_func(v_edge)
self.curr_feats[k_edge] = feat_vec
# conjoin features with tag
new_feat_vec = FeatureVector()
for k, v in feat_vec.iteritems():
new_feat_vec[(str(tag), ) + k] = v
# return edge features conjoined with tag
return new_feat_vec
def ffEdgeNodeDepth(self, edge):
"""
extract node depth features for edge
"""
feat_vec = FeatureVector()
node1 = edge.node1
node2 = edge.node2
for k, v in self.ffNodeDepth(node1).iteritems():
feat_vec[('e', 'n1') + k] = v
for k, v in self.ffNodeDepth(node2).iteritems():
feat_vec[('e', 'n2') + k] = v
node1_foremost = 1e10
node2_foremost = 1e10
for source in node1.sources:
node_depth = len((source.graph_idx).split('.'))
if node1_foremost > node_depth:
node1_foremost = node_depth
for source in node2.sources:
node_depth = len((source.graph_idx).split('.'))
if node2_foremost > node_depth:
node2_foremost = node_depth
# concatenate foremost occurrence of node1 and node2
feat_vec[('e', 'n1', 'dep', 'fmst', str(node1_foremost), 'n2', 'dep',
'fmst', str(node2_foremost))] = 1.0
return feat_vec
def oracle(self, instance):
"""
an instance includes:
my_nodes: (1,) -> AmrNode1, (2,) -> AmrNode2, ...
my_edges: (1,2) -> AmrEdge1, (2,1) -> AmrEdge2,...
root_nodes: (1,), (3,),... nodes that are root of sentence
selected_nodes: (1,), (3,),... nodes contained in summary graph
selected_edges: (1,2), (3,1),... edges contained in summary graph
"""
logger.debug('start oracle decoding...')
curr_filename = instance.filename
my_nodes, oracle_nodes, _ = instance.nodes # nodes and selected nodes
my_edges, oracle_edges = instance.edges # edges and selected edges
# features that are associated with oracle graph
feat_vec = FeatureVector()
for k_edge, v_edge in my_edges.iteritems():
tag = 1 if k_edge in oracle_edges else 0 # use oracle tag
feat_vec += self.feat_extr.getEdgeFeats(k_edge, v_edge, tag,
curr_filename, my_nodes,
my_edges)
for k_node, v_node in my_nodes.iteritems():
tag = 1 if k_node in oracle_nodes else 0 # use oracle tag
feat_vec += self.feat_extr.getNodeFeats(k_node, v_node, tag,
curr_filename, my_nodes,
my_edges)
score_true = self.weights.dot(feat_vec)
# return features associated with oracle graph
return feat_vec, oracle_nodes, oracle_edges, score_true
def ffNodeBias(self, node):
"""
add a bias term to node
"""
feat_vec = FeatureVector()
feat_vec[('n', 'bias')] = 1.0
return feat_vec
def ffEdgeNonNullFreq(self, edge):
"""
extract a binary feature for edge frequency (Non-NULL edges)
freq == 0, freq >= 1, freq >= 2, freq >= 5, freq >= 10
"""
feat_vec = FeatureVector()
edge_freq = 0
for source in edge.sources:
if source.relation != 'NULL':
edge_freq += 1
# binary feature for edge frequency
feat_vec[('e', 'freq', 'non_null',
'0')] = 1.0 if edge_freq == 0 else 0.0
feat_vec[('e', 'freq', 'non_null',
'1')] = 1.0 if edge_freq >= 1 else 0.0
feat_vec[('e', 'freq', 'non_null',
'2')] = 1.0 if edge_freq >= 2 else 0.0
feat_vec[('e', 'freq', 'non_null',
'5')] = 1.0 if edge_freq >= 5 else 0.0
feat_vec[('e', 'freq', 'non_null',
'10')] = 1.0 if edge_freq >= 10 else 0.0
return feat_vec
def ffEdgeBias(self, edge):
"""
add a bias term to edge
"""
feat_vec = FeatureVector()
feat_vec[('e', 'bias')] = 1.0
return feat_vec
def ffEdgeNodeSpan(self, edge):
"""
extract features from node spans of edge
"""
feat_vec = FeatureVector()
node1 = edge.node1
node2 = edge.node2
for k, v in self.ffNodeSpan(node1).iteritems():
feat_vec[('e', 'n1') + k] = v
for k, v in self.ffNodeSpan(node2).iteritems():
feat_vec[('e', 'n2') + k] = v
node1_longest = -1
node2_longest = -1
for source in node1.sources:
node_span = source.end_idx - source.start_idx
if node1_longest < node_span:
node1_longest = node_span
for source in node2.sources:
node_span = source.end_idx - source.start_idx
if node2_longest < node_span:
node2_longest = node_span
# concatenate foremost occurrence of node1 and node2
feat_vec[('e', 'n1', 'span', 'lgst', str(node1_longest), 'n2', 'span',
'lgst', str(node2_longest))] = 1.0
return feat_vec
def ffEdgeNodePos(self, edge):
"""
extract node position features for edge
"""
feat_vec = FeatureVector()
node1 = edge.node1
node2 = edge.node2
for k, v in self.ffNodePos(node1).iteritems():
feat_vec[('e', 'n1') + k] = v
for k, v in self.ffNodePos(node2).iteritems():
feat_vec[('e', 'n2') + k] = v
node1_foremost = 1e10
node2_foremost = 1e10
for source in node1.sources:
if node1_foremost > source.line_num:
node1_foremost = source.line_num
for source in node2.sources:
if node2_foremost > source.line_num:
node2_foremost = source.line_num
# concatenate foremost occurrence of node1 and node2
feat_vec[('e', 'n1', 'posit', 'fmst', str(node1_foremost), 'n2',
'posit', 'fmst', str(node2_foremost))] = 1.0
return feat_vec
def ffNodeConcept(self, node):
"""
extract node concept feature
"""
feat_vec = FeatureVector()
# node concept
feat_vec[('n', 'cpt', node.concept)] = 1.0
return feat_vec
def ffNodeCollapsedEntity(self, node):
"""
extract features from collapsed concept node
"""
feat_vec = FeatureVector()
# named entity or not
feat_vec[('n', 'nam-ent')] = 1.0 if '_' in node.concept else 0.0
feat_vec[('n', 'date-ent')] = 1.0 if (
node.concept).startswith('date-entity') else 0.0
return feat_vec
def ffEdgeIsNull(self, edge):
"""
extract a binary feature indicating a NULL edge or not
"""
feat_vec = FeatureVector()
is_null = 1.0
for source in edge.sources:
if source.relation != 'NULL':
is_null = 0.0
break
feat_vec[('e', 'is_null')] = is_null
return feat_vec
def ffEdgeNodeConcept(self, edge):
"""
extract node concept features for edge
"""
feat_vec = FeatureVector()
node1 = edge.node1
node2 = edge.node2
for k, v in self.ffNodeConcept(node1).iteritems():
feat_vec[('e', 'n1') + k] = v
for k, v in self.ffNodeConcept(node2).iteritems():
feat_vec[('e', 'n2') + k] = v
return feat_vec
def ffEdgeNodeFreq(self, edge):
"""
extract node frequency features for an edge
"""
feat_vec = FeatureVector()
node1 = edge.node1
node2 = edge.node2
for k, v in self.ffNodeFreq(node1).iteritems():
feat_vec[('e', 'n1') + k] = v
for k, v in self.ffNodeFreq(node2).iteritems():
feat_vec[('e', 'n2') + k] = v
return feat_vec
def ffEdgeNodeCollapsedEntity(self, edge):
"""
extract features from collapsed concept node
"""
feat_vec = FeatureVector()
node1 = edge.node1
node2 = edge.node2
for k, v in self.ffNodeSpan(node1).iteritems():
feat_vec[('e', 'n1') + k] = v
for k, v in self.ffNodeSpan(node2).iteritems():
feat_vec[('e', 'n2') + k] = v
return feat_vec
def ffNodeFreq(self, node):
"""
extract node frequency features
"""
feat_vec = FeatureVector()
# node frequency
node_freq = len(node.sources)
feat_vec[('n', 'freq', '0')] = 1.0 if node_freq == 0 else 0.0
feat_vec[('n', 'freq', '1')] = 1.0 if node_freq >= 1 else 0.0
feat_vec[('n', 'freq', '2')] = 1.0 if node_freq >= 2 else 0.0
feat_vec[('n', 'freq', '5')] = 1.0 if node_freq >= 5 else 0.0
feat_vec[('n', 'freq', '10')] = 1.0 if node_freq >= 10 else 0.0
return feat_vec
def ffEdgeRel(self, edge):
"""
extract a binary feature for edge relation.
"""
feat_vec = FeatureVector()
edge_freq = len(edge.sources)
# primary and secondary relation (edge relation entropy)?
rel_freq = Counter()
for source in edge.sources:
rel_freq[source.relation] += 1
rels = rel_freq.most_common()
if rels: # primary relation
(rel, count) = rels[0]
feat_vec[('e', 'rel', 'fst', rel)] = 1.0
# relative frequency
per_fst_rel = count / edge_freq
feat_vec[('e', 'rel', 'fst', rel,
'p1')] = 1.0 if per_fst_rel >= 0.5 else 0.0
feat_vec[('e', 'rel', 'fst', rel,
'p2')] = 1.0 if per_fst_rel >= 0.66 else 0.0
feat_vec[('e', 'rel', 'fst', rel,
'p3')] = 1.0 if per_fst_rel >= 0.75 else 0.0
# secondary relation and relative frequency
if len(rels) > 1:
(sec_rel, sec_count) = rels[1]
feat_vec[('e', 'rel', 'sec', sec_rel)] = 1.0
# relative frequency
per_sec_rel = sec_count / edge_freq
feat_vec[('e', 'rel', 'sec', sec_rel,
'p1')] = 1.0 if per_sec_rel >= 0.25 else 0.0
feat_vec[('e', 'rel', 'sec', sec_rel,
'p2')] = 1.0 if per_sec_rel >= 0.33 else 0.0
feat_vec[('e', 'rel', 'sec', sec_rel,
'p3')] = 1.0 if per_sec_rel >= 0.5 else 0.0
# combine first and secondary relation
feat_vec[('e', 'rel', 'fst', rel, 'sec', sec_rel)] = 1.0
return feat_vec
def ffNodeSpan(self, node):
"""
extract features from node spans
"""
feat_vec = FeatureVector()
# longest span of node
# average span of node
node_longest = -1
node_average = 0.0
node_freq = len(node.sources)
for source in node.sources:
node_span = source.end_idx - source.start_idx
node_average += node_span
if node_longest < node_span:
node_longest = node_span
if node_freq > 0:
node_average /= node_freq
if node_freq > 0:
feat_vec[('n', 'span', 'lgst',
'0')] = 1.0 if node_longest >= 0 else 0.0
feat_vec[('n', 'span', 'lgst',
'1')] = 1.0 if node_longest >= 1 else 0.0
feat_vec[('n', 'span', 'lgst',
'2')] = 1.0 if node_longest >= 2 else 0.0
feat_vec[('n', 'span', 'lgst',
'5')] = 1.0 if node_longest >= 5 else 0.0
feat_vec[('n', 'span', 'lgst',
'10')] = 1.0 if node_longest >= 10 else 0.0
feat_vec[('n', 'span', 'avg',
'0')] = 1.0 if node_longest >= 0 else 0.0
feat_vec[('n', 'span', 'avg',
'1')] = 1.0 if node_longest >= 1 else 0.0
feat_vec[('n', 'span', 'avg',
'2')] = 1.0 if node_longest >= 2 else 0.0
feat_vec[('n', 'span', 'avg',
'5')] = 1.0 if node_longest >= 5 else 0.0
feat_vec[('n', 'span', 'avg',
'10')] = 1.0 if node_longest >= 10 else 0.0
return feat_vec
def ffNodePos(self, node):
"""
extract node position features
"""
feat_vec = FeatureVector()
# foremost occurrence of node
# average occurrence position of node
node_foremost = 1e10
node_average = 0.0
node_freq = len(node.sources)
for source in node.sources:
node_average += source.line_num
if node_foremost > source.line_num:
node_foremost = source.line_num
if node_freq > 0:
node_average /= node_freq
if node_freq > 0:
# foremost occurrence
feat_vec[('n', 'posit', 'fmst',
'5')] = 1.0 if node_foremost >= 5 else 0.0
feat_vec[('n', 'posit', 'fmst',
'6')] = 1.0 if node_foremost >= 6 else 0.0
feat_vec[('n', 'posit', 'fmst',
'7')] = 1.0 if node_foremost >= 7 else 0.0
feat_vec[('n', 'posit', 'fmst',
'10')] = 1.0 if node_foremost >= 10 else 0.0
feat_vec[('n', 'posit', 'fmst',
'15')] = 1.0 if node_foremost >= 15 else 0.0
# average occurrence
feat_vec[('n', 'posit', 'avg',
'5')] = 1.0 if node_average >= 5 else 0.0
feat_vec[('n', 'posit', 'avg',
'6')] = 1.0 if node_average >= 6 else 0.0
feat_vec[('n', 'posit', 'avg',
'7')] = 1.0 if node_average >= 7 else 0.0
feat_vec[('n', 'posit', 'avg',
'10')] = 1.0 if node_average >= 10 else 0.0
feat_vec[('n', 'posit', 'avg',
'15')] = 1.0 if node_average >= 15 else 0.0
return feat_vec
def ffNodeDepth(self, node):
"""
depth of node in graph topology
"""
feat_vec = FeatureVector()
# foremost occurrence of node1, node2
# average occurrence position of node1, node2
node_foremost = 1e10
node_average = 0.0
node_freq = len(node.sources)
for source in node.sources:
node_depth = len((source.graph_idx).split('.'))
node_average += node_depth
if node_foremost > node_depth:
node_foremost = node_depth
if node_freq > 0:
node_average /= node_freq
if node_freq > 0:
feat_vec[('n', 'dep', 'fmst',
'1')] = 1.0 if node_foremost >= 1 else 0.0
feat_vec[('n', 'dep', 'fmst',
'2')] = 1.0 if node_foremost >= 2 else 0.0
feat_vec[('n', 'dep', 'fmst',
'3')] = 1.0 if node_foremost >= 3 else 0.0
feat_vec[('n', 'dep', 'fmst',
'4')] = 1.0 if node_foremost >= 4 else 0.0
feat_vec[('n', 'dep', 'fmst',
'5')] = 1.0 if node_foremost >= 5 else 0.0
feat_vec[('n', 'dep', 'avg',
'1')] = 1.0 if node_average >= 1 else 0.0
feat_vec[('n', 'dep', 'avg',
'2')] = 1.0 if node_average >= 2 else 0.0
feat_vec[('n', 'dep', 'avg',
'3')] = 1.0 if node_average >= 3 else 0.0
feat_vec[('n', 'dep', 'avg',
'4')] = 1.0 if node_average >= 4 else 0.0
feat_vec[('n', 'dep', 'avg',
'5')] = 1.0 if node_average >= 5 else 0.0
return feat_vec
def ffEdgeFreq(self, edge):
"""
extract a binary feature for edge frequency.
freq == 0, freq >= 1, freq >= 2, freq >= 5, freq >= 10
"""
feat_vec = FeatureVector()
# edge frequency
edge_freq = len(edge.sources)
# binary feature for edge frequency
feat_vec[('e', 'freq', '0')] = 1.0 if edge_freq == 0 else 0.0
feat_vec[('e', 'freq', '1')] = 1.0 if edge_freq >= 1 else 0.0
feat_vec[('e', 'freq', '2')] = 1.0 if edge_freq >= 2 else 0.0
feat_vec[('e', 'freq', '5')] = 1.0 if edge_freq >= 5 else 0.0
feat_vec[('e', 'freq', '10')] = 1.0 if edge_freq >= 10 else 0.0
return feat_vec
def ffEdgePos(self, edge):
"""
extract features from edge occurrences.
"""
feat_vec = FeatureVector()
# foremost position in all edge occurrences
# average position across all edge occurrences
foremost = 1e10
average = 0.0
edge_freq = len(edge.sources)
for source in edge.sources:
average += source.line_num
if foremost > source.line_num:
foremost = source.line_num
if edge_freq > 0:
average /= edge_freq
if edge_freq > 0: # edge exists
# foremost occurrence
feat_vec[('e', 'posit', 'fmst',
'5')] = 1.0 if foremost >= 5 else 0.0
feat_vec[('e', 'posit', 'fmst',
'6')] = 1.0 if foremost >= 6 else 0.0
feat_vec[('e', 'posit', 'fmst',
'7')] = 1.0 if foremost >= 7 else 0.0
feat_vec[('e', 'posit', 'fmst',
'10')] = 1.0 if foremost >= 10 else 0.0
feat_vec[('e', 'posit', 'fmst',
'15')] = 1.0 if foremost >= 15 else 0.0
# average occurrence
feat_vec[('e', 'posit', 'avg', '5')] = 1.0 if average >= 5 else 0.0
feat_vec[('e', 'posit', 'avg', '6')] = 1.0 if average >= 6 else 0.0
feat_vec[('e', 'posit', 'avg', '7')] = 1.0 if average >= 7 else 0.0
feat_vec[('e', 'posit', 'avg',
'10')] = 1.0 if average >= 10 else 0.0
feat_vec[('e', 'posit', 'avg',
'15')] = 1.0 if average >= 15 else 0.0
return feat_vec
feat_vec[('n', 'nam-ent')] = 1.0 if '_' in node.concept else 0.0
feat_vec[('n', 'date-ent')] = 1.0 if (node.concept).startswith('date-entity') else 0.0
return feat_vec
if __name__ == '__main__':
input_dir = '/Users/user/Data/SemanticSumm/Proxy/gold/split/dev/'
body_file = 'aligned-amr-release-1.0-dev-proxy-body.txt'
summ_file = 'aligned-amr-release-1.0-dev-proxy-summary.txt'
corpus = buildCorpus(os.path.join(input_dir, body_file),
os.path.join(input_dir, summ_file))
feat_extr = FeatureExtractor()
feat_vec = FeatureVector()
for inst in corpus:
curr_filename = inst.filename
my_nodes, s_nodes = inst.nodes
my_edges, s_edges = inst.edges
# logger.debug('extracting features for file: %s' % curr_filename)
# for k_edge, v_edge in my_edges.iteritems():
# for tag in [0,1]:
# feat_vec += feat_extr.getEdgeFeats(k_edge, v_edge, tag, curr_filename, my_nodes, my_edges)
logger.debug('extracting features for file: %s' % curr_filename)
for k_node, v_node in my_nodes.iteritems():
for tag in [0,1]:
feat_vec += feat_extr.getNodeFeats(k_node, v_node, tag, curr_filename, my_nodes, my_edges)
def learnParamsAdaGrad(self,
decoder,
corpus,
param_file,
loss_func,
num_passes=10,
oracle_len='nolen'):
"""
learn parameters using Structured Perceptron, Ramp Loss, (Hinge??)
"""
logger.debug('start learning parameters...')
shuffle(corpus) # shuffle corpus
avg_weights = FeatureVector()
curr_instances = 0
node_perf = PerfScore(0.0, 0.0, 0.0) # node performance
edge_perf = PerfScore(0.0, 0.0, 0.0) # edge performance
eta = 1.0 # stepsize
l2reg = 0.0 #
node_cost_scaling = 1.0 # cost scaling factor
edge_cost_scaling = 1.0 # cost scaling factor
sumSq = FeatureVector()
for curr_num_passes in xrange(1, num_passes + 1):
logger.debug('#curr_num_passes#: %d' % curr_num_passes)
for instance in corpus:
curr_instances += 1
logger.debug('processing instance %d...' % curr_instances)
# perceptron loss
if loss_func.startswith('perceptron'):
gradient, selected_nodes, selected_edges, score_pred = decoder.decode(
instance, oracle_len)
plus_feats, oracle_nodes, oracle_edges, score_true = decoder.oracle(
instance)
curr_loss = score_pred - score_true # @UnusedVariable
gradient -= plus_feats
decoder.weights -= eta * gradient
# ramp loss + cost-augmented decoding
if loss_func.startswith('ramp'):
node_cost, edge_cost = node_cost_scaling, edge_cost_scaling
gradient, _, _, score_plus_cost = decoder.decode(
instance, oracle_len, node_cost, edge_cost)
plus_feats, selected_nodes, selected_edges, score_minus_cost = decoder.decode(
instance, oracle_len, -1.0 * node_cost,
-1.0 * edge_cost)
_, oracle_nodes, oracle_edges, score_true = decoder.oracle(
instance)
curr_loss = score_plus_cost - score_minus_cost # @UnusedVariable
gradient -= plus_feats
for k, v in gradient.iteritems():
if v == 0.0: continue
sumSq[k] = sumSq.get(k, 0.0) + v * v
decoder.weights[k] = decoder.weights.get(
k, 0.0) - eta * v / sqrt(sumSq[k])
if l2reg != 0.0:
for k, v in decoder.weights.iteritems():
if v == 0.0: continue
value = l2reg * v
sumSq[k] = sumSq.get(k, 0.0) + value * value
decoder.weights[k] = v - eta * value / sqrt(
sumSq[k])
# hinge-loss + cost-augmented decoding
if loss_func.startswith('hinge'):
node_cost, edge_cost = node_cost_scaling, edge_cost_scaling
gradient, selected_nodes, selected_edges, score_plus_cost = decoder.decode(
instance, oracle_len, node_cost, edge_cost)
plus_feats, oracle_nodes, oracle_edges, score_true = decoder.oracle(
instance)
curr_loss = score_plus_cost - score_true # @UnusedVariable
gradient -= plus_feats
for k, v in gradient.iteritems():
if v == 0.0: continue
sumSq[k] = sumSq.get(k, 0.0) + v * v
decoder.weights[k] = decoder.weights.get(
k, 0.0) - eta * v / sqrt(sumSq[k])
if l2reg != 0.0:
for k, v in decoder.weights.iteritems():
if v == 0.0: continue
value = l2reg * v
sumSq[k] = sumSq.get(k, 0.0) + value * value
decoder.weights[k] = v - eta * value / sqrt(
sumSq[k])
# use gold nodes and edges to calculate P/R/F
num_gold_nodes, num_gold_edges = instance.gold
# P/R/F scores of nodes and edges, for current instance
# Edge recall can not reach %100 since decoding produces only tree structure
intersect_nodes = set(selected_nodes) & set(oracle_nodes)
curr_node_perf = getPRFScores(len(intersect_nodes),
len(selected_nodes),
num_gold_nodes)
logPRFScores('train_node', curr_node_perf)
intersect_edges = set(selected_edges) & set(oracle_edges)
curr_edge_perf = getPRFScores(len(intersect_edges),
len(selected_edges),
num_gold_edges)
logPRFScores('train_edge', curr_edge_perf)
# P/R/F scores of nodes and edges, averaged across all curr_instances
node_perf = PerfScore(
*[sum(x) for x in zip(node_perf, curr_node_perf)])
edge_perf = PerfScore(
*[sum(x) for x in zip(edge_perf, curr_edge_perf)])
logPRFScores(
'train_node_avg',
PerfScore(node_perf.prec / curr_instances,
node_perf.rec / curr_instances,
node_perf.fscore / curr_instances))
logPRFScores(
'train_edge_avg',
PerfScore(edge_perf.prec / curr_instances,
edge_perf.rec / curr_instances,
edge_perf.fscore / curr_instances))
# averaging weight vectors
avg_weights += decoder.weights
# output averaged weight vectors to file
curr_weights = FeatureVector()
curr_weights += avg_weights * (1 / curr_num_passes)
if param_file:
with codecs.open(param_file, 'w', 'utf-8') as outfile:
outfile.write('#curr_num_passes#: %d\n' % curr_num_passes)
outfile.write('%s\n' % curr_weights.toString())
final_weights = FeatureVector()
final_weights += avg_weights * (1 / num_passes)
return final_weights
def __init__(self):
self.gilp = GurobiILP()
self.weights = FeatureVector()
self.feat_extr = FeatureExtractor()
return
def __init__(self):
self.gilp = GurobiILP()
self.weights = FeatureVector()
self.feat_extr = FeatureExtractor()
return
feat_vec[('n', 'date-ent')] = 1.0 if (
node.concept).startswith('date-entity') else 0.0
return feat_vec
if __name__ == '__main__':
input_dir = '/Users/user/Data/SemanticSumm/Proxy/gold/split/dev/'
body_file = 'aligned-amr-release-1.0-dev-proxy-body.txt'
summ_file = 'aligned-amr-release-1.0-dev-proxy-summary.txt'
corpus = buildCorpus(os.path.join(input_dir, body_file),
os.path.join(input_dir, summ_file))
feat_extr = FeatureExtractor()
feat_vec = FeatureVector()
for inst in corpus:
curr_filename = inst.filename
my_nodes, s_nodes = inst.nodes
my_edges, s_edges = inst.edges
# logger.debug('extracting features for file: %s' % curr_filename)
# for k_edge, v_edge in my_edges.iteritems():
# for tag in [0,1]:
# feat_vec += feat_extr.getEdgeFeats(k_edge, v_edge, tag, curr_filename, my_nodes, my_edges)
logger.debug('extracting features for file: %s' % curr_filename)
for k_node, v_node in my_nodes.iteritems():
for tag in [0, 1]:
feat_vec += feat_extr.getNodeFeats(k_node, v_node, tag,
class Decoder(object):
"""
Implement of decoder for structured prediction
"""
def __init__(self):
self.gilp = GurobiILP()
self.weights = FeatureVector()
self.feat_extr = FeatureExtractor()
return
def decode(self, instance, oracle_len='nolen', node_cost=None, edge_cost=None):
"""
an instance includes:
my_nodes: (1,) -> AmrNode1, (2,) -> AmrNode2, ...
my_edges: (1,2) -> AmrEdge1, (2,1) -> AmrEdge2,...
selected_nodes: (1,), (3,),... nodes contained in summary graph
selected_edges: (1,2), (3,1),... edges contained in summary graph
"""
logger.debug('start feature extraction...')
curr_filename = instance.filename
my_nodes, oracle_nodes, root_nodes = instance.nodes # nodes and selected nodes
my_edges, oracle_edges = instance.edges # edges and selected edges
num_gold_nodes, num_gold_edges = instance.gold # number of gold nodes and edges
node_weights = {}
edge_weights = {}
# get edge weights
num_nonnegative_edges = 0
for k_edge, v_edge in my_edges.iteritems():
for tag in [0,1]:
edge_feats = self.feat_extr.getEdgeFeats(k_edge, v_edge, tag, curr_filename, my_nodes, my_edges)
edge_weights[k_edge + (tag,)] = self.weights.dot(edge_feats)
# cost-augmented decoding
if edge_cost is not None:
if tag == 0 and k_edge not in oracle_edges: # true negative
curr_edge_cost = 0
if tag == 1 and k_edge in oracle_edges: # true positive
curr_edge_cost = 0
if tag == 1 and k_edge not in oracle_edges: # false positive
curr_edge_cost = edge_cost
if tag == 0 and k_edge in oracle_edges: # false negative
curr_edge_cost = edge_cost
edge_weights[k_edge + (tag,)] += curr_edge_cost
if tag == 1 and edge_weights[k_edge + (tag,)] > 0.0: num_nonnegative_edges += 1
# count number of non-negative edges
logger.debug('[num_nonnegative_edges]: %d' % num_nonnegative_edges)
# get node weights
num_nonnegative_nodes = 0
for k_node, v_node in my_nodes.iteritems():
for tag in [0,1]:
node_feats = self.feat_extr.getNodeFeats(k_node, v_node, tag, curr_filename, my_nodes, my_edges)
node_weights[k_node + (tag,)] = self.weights.dot(node_feats)
# cost-augmented decoding
if node_cost is not None:
if tag == 0 and k_node not in oracle_nodes:
curr_node_cost = 0
if tag == 1 and k_node in oracle_nodes:
curr_node_cost = 0
if tag == 1 and k_node not in oracle_nodes: # false positive
curr_node_cost = node_cost
if tag == 0 and k_node in oracle_nodes: # false negative
curr_node_cost = node_cost
node_weights[k_node + (tag,)] += curr_node_cost
if tag == 1 and node_weights[k_node + (tag,)] > 0.0: num_nonnegative_nodes += 1
# count number of non-negative nodes
logger.debug('[num_nonnegative_nodes]: %d' % num_nonnegative_nodes)
# run Gurobi ILP decoder using node and edge weights
# optionally set the decoded summary length (#nodes and #edges)
logger.debug('start ILP decoding...')
num_selected_nodes = num_gold_nodes if oracle_len == 'nodes' else 0
num_selected_edges = num_gold_edges if oracle_len == 'edges' else 0
selected_nodes, selected_edges, score_pred = self.gilp.decode(node_weights, edge_weights, root_nodes,
num_selected_nodes=num_selected_nodes,
num_selected_edges=num_selected_edges)
logger.debug('[num_gold_nodes]: %d' % num_gold_nodes)
logger.debug('[num_selected_nodes]: %d' % len(selected_nodes))
logger.debug('[num_gold_edges]: %d' % num_gold_edges)
logger.debug('[num_selected_edges]: %d' % len(selected_edges))
# features that are associated with the decoded graph
feat_vec = FeatureVector()
for k_edge, v_edge in my_edges.iteritems():
tag = 1 if k_edge in selected_edges else 0 # use decoded tag
feat_vec += self.feat_extr.getEdgeFeats(k_edge, v_edge, tag, curr_filename, my_nodes, my_edges)
for k_node, v_node in my_nodes.iteritems():
tag = 1 if k_node in selected_nodes else 0 # use decoded tag
feat_vec += self.feat_extr.getNodeFeats(k_node, v_node, tag, curr_filename, my_nodes, my_edges)
# return features associated with decoded graph
return feat_vec, selected_nodes, selected_edges, score_pred
def oracle(self, instance):
"""
an instance includes:
my_nodes: (1,) -> AmrNode1, (2,) -> AmrNode2, ...
my_edges: (1,2) -> AmrEdge1, (2,1) -> AmrEdge2,...
root_nodes: (1,), (3,),... nodes that are root of sentence
selected_nodes: (1,), (3,),... nodes contained in summary graph
selected_edges: (1,2), (3,1),... edges contained in summary graph
"""
logger.debug('start oracle decoding...')
curr_filename = instance.filename
my_nodes, oracle_nodes, _ = instance.nodes # nodes and selected nodes
my_edges, oracle_edges = instance.edges # edges and selected edges
# features that are associated with oracle graph
feat_vec = FeatureVector()
for k_edge, v_edge in my_edges.iteritems():
tag = 1 if k_edge in oracle_edges else 0 # use oracle tag
feat_vec += self.feat_extr.getEdgeFeats(k_edge, v_edge, tag, curr_filename, my_nodes, my_edges)
for k_node, v_node in my_nodes.iteritems():
tag = 1 if k_node in oracle_nodes else 0 # use oracle tag
feat_vec += self.feat_extr.getNodeFeats(k_node, v_node, tag, curr_filename, my_nodes, my_edges)
score_true = self.weights.dot(feat_vec)
# return features associated with oracle graph
return feat_vec, oracle_nodes, oracle_edges, score_true
def decode(self,
instance,
oracle_len='nolen',
node_cost=None,
edge_cost=None):
"""
an instance includes:
my_nodes: (1,) -> AmrNode1, (2,) -> AmrNode2, ...
my_edges: (1,2) -> AmrEdge1, (2,1) -> AmrEdge2,...
selected_nodes: (1,), (3,),... nodes contained in summary graph
selected_edges: (1,2), (3,1),... edges contained in summary graph
"""
logger.debug('start feature extraction...')
curr_filename = instance.filename
my_nodes, oracle_nodes, root_nodes = instance.nodes # nodes and selected nodes
my_edges, oracle_edges = instance.edges # edges and selected edges
num_gold_nodes, num_gold_edges = instance.gold # number of gold nodes and edges
node_weights = {}
edge_weights = {}
# get edge weights
num_nonnegative_edges = 0
for k_edge, v_edge in my_edges.iteritems():
for tag in [0, 1]:
edge_feats = self.feat_extr.getEdgeFeats(
k_edge, v_edge, tag, curr_filename, my_nodes, my_edges)
edge_weights[k_edge + (tag, )] = self.weights.dot(edge_feats)
# cost-augmented decoding
if edge_cost is not None:
if tag == 0 and k_edge not in oracle_edges: # true negative
curr_edge_cost = 0
if tag == 1 and k_edge in oracle_edges: # true positive
curr_edge_cost = 0
if tag == 1 and k_edge not in oracle_edges: # false positive
curr_edge_cost = edge_cost
if tag == 0 and k_edge in oracle_edges: # false negative
curr_edge_cost = edge_cost
edge_weights[k_edge + (tag, )] += curr_edge_cost
if tag == 1 and edge_weights[k_edge + (tag, )] > 0.0:
num_nonnegative_edges += 1
# count number of non-negative edges
logger.debug('[num_nonnegative_edges]: %d' % num_nonnegative_edges)
# get node weights
num_nonnegative_nodes = 0
for k_node, v_node in my_nodes.iteritems():
for tag in [0, 1]:
node_feats = self.feat_extr.getNodeFeats(
k_node, v_node, tag, curr_filename, my_nodes, my_edges)
node_weights[k_node + (tag, )] = self.weights.dot(node_feats)
# cost-augmented decoding
if node_cost is not None:
if tag == 0 and k_node not in oracle_nodes:
curr_node_cost = 0
if tag == 1 and k_node in oracle_nodes:
curr_node_cost = 0
if tag == 1 and k_node not in oracle_nodes: # false positive
curr_node_cost = node_cost
if tag == 0 and k_node in oracle_nodes: # false negative
curr_node_cost = node_cost
node_weights[k_node + (tag, )] += curr_node_cost
if tag == 1 and node_weights[k_node + (tag, )] > 0.0:
num_nonnegative_nodes += 1
# count number of non-negative nodes
logger.debug('[num_nonnegative_nodes]: %d' % num_nonnegative_nodes)
# run Gurobi ILP decoder using node and edge weights
# optionally set the decoded summary length (#nodes and #edges)
logger.debug('start ILP decoding...')
num_selected_nodes = num_gold_nodes if oracle_len == 'nodes' else 0
num_selected_edges = num_gold_edges if oracle_len == 'edges' else 0
selected_nodes, selected_edges, score_pred = self.gilp.decode(
node_weights,
edge_weights,
root_nodes,
num_selected_nodes=num_selected_nodes,
num_selected_edges=num_selected_edges)
logger.debug('[num_gold_nodes]: %d' % num_gold_nodes)
logger.debug('[num_selected_nodes]: %d' % len(selected_nodes))
logger.debug('[num_gold_edges]: %d' % num_gold_edges)
logger.debug('[num_selected_edges]: %d' % len(selected_edges))
# features that are associated with the decoded graph
feat_vec = FeatureVector()
for k_edge, v_edge in my_edges.iteritems():
tag = 1 if k_edge in selected_edges else 0 # use decoded tag
feat_vec += self.feat_extr.getEdgeFeats(k_edge, v_edge, tag,
curr_filename, my_nodes,
my_edges)
for k_node, v_node in my_nodes.iteritems():
tag = 1 if k_node in selected_nodes else 0 # use decoded tag
feat_vec += self.feat_extr.getNodeFeats(k_node, v_node, tag,
curr_filename, my_nodes,
my_edges)
# return features associated with decoded graph
return feat_vec, selected_nodes, selected_edges, score_pred
def learnParamsAdaGrad(self, decoder, corpus, param_file, loss_func, num_passes=10, oracle_len='nolen'):
"""
learn parameters using Structured Perceptron, Ramp Loss, (Hinge??)
"""
logger.debug('start learning parameters...')
shuffle(corpus) # shuffle corpus
avg_weights = FeatureVector()
curr_instances = 0
node_perf = PerfScore(0.0, 0.0, 0.0) # node performance
edge_perf = PerfScore(0.0, 0.0, 0.0) # edge performance
eta = 1.0 # stepsize
l2reg = 0.0 #
node_cost_scaling = 1.0 # cost scaling factor
edge_cost_scaling = 1.0 # cost scaling factor
sumSq = FeatureVector()
for curr_num_passes in xrange(1, num_passes+1):
logger.debug('#curr_num_passes#: %d' % curr_num_passes)
for instance in corpus:
curr_instances += 1
logger.debug('processing instance %d...' % curr_instances)
# perceptron loss
if loss_func.startswith('perceptron'):
gradient, selected_nodes, selected_edges, score_pred = decoder.decode(instance, oracle_len)
plus_feats, oracle_nodes, oracle_edges, score_true = decoder.oracle(instance)
curr_loss = score_pred - score_true # @UnusedVariable
gradient -= plus_feats
decoder.weights -= eta * gradient
# ramp loss + cost-augmented decoding
if loss_func.startswith('ramp'):
node_cost, edge_cost = node_cost_scaling, edge_cost_scaling
gradient, _, _, score_plus_cost = decoder.decode(instance, oracle_len, node_cost, edge_cost)
plus_feats, selected_nodes, selected_edges, score_minus_cost = decoder.decode(instance, oracle_len, -1.0 * node_cost, -1.0 * edge_cost)
_, oracle_nodes, oracle_edges, score_true = decoder.oracle(instance)
curr_loss = score_plus_cost - score_minus_cost # @UnusedVariable
gradient -= plus_feats
for k, v in gradient.iteritems():
if v == 0.0: continue
sumSq[k] = sumSq.get(k, 0.0) + v * v
decoder.weights[k] = decoder.weights.get(k, 0.0) - eta * v / sqrt(sumSq[k])
if l2reg != 0.0:
for k, v in decoder.weights.iteritems():
if v == 0.0: continue
value = l2reg * v
sumSq[k] = sumSq.get(k, 0.0) + value * value
decoder.weights[k] = v - eta * value / sqrt(sumSq[k])
# hinge-loss + cost-augmented decoding
if loss_func.startswith('hinge'):
node_cost, edge_cost = node_cost_scaling, edge_cost_scaling
gradient, selected_nodes, selected_edges, score_plus_cost = decoder.decode(instance, oracle_len, node_cost, edge_cost)
plus_feats, oracle_nodes, oracle_edges, score_true = decoder.oracle(instance)
curr_loss = score_plus_cost - score_true # @UnusedVariable
gradient -= plus_feats
for k, v in gradient.iteritems():
if v == 0.0: continue
sumSq[k] = sumSq.get(k, 0.0) + v * v
decoder.weights[k] = decoder.weights.get(k, 0.0) - eta * v / sqrt(sumSq[k])
if l2reg != 0.0:
for k, v in decoder.weights.iteritems():
if v == 0.0: continue
value = l2reg * v
sumSq[k] = sumSq.get(k, 0.0) + value * value
decoder.weights[k] = v - eta * value / sqrt(sumSq[k])
# use gold nodes and edges to calculate P/R/F
num_gold_nodes, num_gold_edges = instance.gold
# P/R/F scores of nodes and edges, for current instance
# Edge recall can not reach %100 since decoding produces only tree structure
intersect_nodes = set(selected_nodes) & set(oracle_nodes)
curr_node_perf = getPRFScores(len(intersect_nodes), len(selected_nodes), num_gold_nodes)
logPRFScores('train_node', curr_node_perf)
intersect_edges = set(selected_edges) & set(oracle_edges)
curr_edge_perf = getPRFScores(len(intersect_edges), len(selected_edges), num_gold_edges)
logPRFScores('train_edge', curr_edge_perf)
# P/R/F scores of nodes and edges, averaged across all curr_instances
node_perf = PerfScore(*[sum(x) for x in zip(node_perf, curr_node_perf)])
edge_perf = PerfScore(*[sum(x) for x in zip(edge_perf, curr_edge_perf)])
logPRFScores('train_node_avg',
PerfScore(node_perf.prec/curr_instances, node_perf.rec/curr_instances,
node_perf.fscore/curr_instances))
logPRFScores('train_edge_avg',
PerfScore(edge_perf.prec/curr_instances, edge_perf.rec/curr_instances,
edge_perf.fscore/curr_instances))
# averaging weight vectors
avg_weights += decoder.weights
# output averaged weight vectors to file
curr_weights = FeatureVector()
curr_weights += avg_weights * (1/curr_num_passes)
if param_file:
with codecs.open(param_file, 'w', 'utf-8') as outfile:
outfile.write('#curr_num_passes#: %d\n' % curr_num_passes)
outfile.write('%s\n' % curr_weights.toString())
final_weights = FeatureVector()
final_weights += avg_weights * (1/num_passes)
return final_weights