def create_features(env, program_tks, max_n_exp):
"""Create features to be used in the score function."""
# Unique decoder variables used in the program
variables = set(
[tk for tk in program_tks if tk[0] == 'v' and tk[1:].isdigit()])
ns = env.interpreter.namespace
# Variables already present in the namespace i.e. entities and column names
ns_vars = [tk for tk in variables if tk in ns]
str_vals = [unicode(ns[tk]['value']) for tk in ns_vars]
vars_to_val = dict(zip(ns_vars, str_vals))
# Tokens which represent entities
val_ent = env.ent_props['val_ent']
ent_vars = [tk for tk, val in vars_to_val.iteritems() if val in val_ent]
ents_index = [val_ent[vars_to_val[tk]][0] for tk in ent_vars]
entities = [env.entities[i] for i in ents_index]
sum_entity_length = {
'datetime_list': 0.0,
'string_list': 0.0,
'num_list': 0.0
}
ent_lengths = [e['length'] for e in entities]
for ent, w in zip(entities, ent_lengths):
sum_entity_length[ent['type']] += w
# Feature vector
feature_size = 12 if env.trigger_words_dict else 11
f = np.zeros(feature_size)
# Entity features
ent_sum = sum(sum_entity_length.values())
ques_ent_sum = sum(env.ent_props['sum_entity_length'].values())
# No of fractional entities where each entity is weighted by its length.
f[0] = ent_sum / (ques_ent_sum + EPS)
# No of fractional entities for each entity type weighted by its length
f[1:4] = [
v / (env.ent_props['sum_entity_length'][k] + EPS)
for k, v in sum_entity_length.iteritems()
]
# No of fractional entities
f[4] = len(ent_vars) / (len(env.entities) + EPS)
# Feature that the longest entity is present in the program or not, if there
# is than one entity that has the maximum length, this feature
# represents the fraction of entities present in the program that has the
# max length.
max_len = max(ent_lengths) if ent_lengths else 0.0
if max_len == env.ent_props['max_length']:
max_ent_sum = data_utils.max_sum(ent_lengths)
f[5] = max_ent_sum / (env.ent_props['max_sum'] + EPS)
# Column features
# Only consider columns represnting differing things, for example,
# year-date and year-string are considered the same column
column_vars, column_vals = [], set()
for v in ns_vars:
if v not in ent_vars:
str_val = vars_to_val[v].split('-')[0]
if str_val not in column_vals:
column_vars.append(v)
column_vals.add(str_val)
cols = env.de_vocab.lookup(column_vars)
col_features = [env.id_feature_dict[col][0] for col in cols]
col_sum, ques_col_sum = sum(col_features), env.col_props['sum']
f[6] = col_sum / (ques_col_sum + EPS)
f[7] = sum([(i > 0) for i in col_features]) / (env.col_props['num'] + EPS)
max_w = max(col_features) if col_features else 0.0
if max_w == env.col_props['max']:
f[8] = data_utils.max_sum(col_features) / (env.col_props['max_sum'] +
EPS)
# (1 - n) where n = fractional number of expressions in the program
num_exp = program_tks.count('(')
f[9] = 1.0 - (num_exp / max_n_exp)
# Function tokens
if env.trigger_words_dict:
fn_score = 0.0
fn_tks = [tk for tk in program_tks if tk in env.trigger_words_dict]
if fn_tks:
for tk in fn_tks:
if env.ques_tokens & env.trigger_words_dict[tk]:
fn_score += 1.0
fn_score /= len(fn_tks)
f[10] = fn_score
# Approximate set interection similarity feature
denominator = ((2 * ques_ent_sum + ques_col_sum) * num_exp)
f[-1] = (2 * ent_sum + col_sum) / (denominator + EPS)
return f
def __init__(self,
en_vocab,
de_vocab,
question_annotation,
answer,
constant_value_embedding_fn,
score_fn,
interpreter,
constants=None,
punish_extra_work=True,
init_interp=True,
trigger_words_dict=None,
max_cache_size=1e4,
name='qa_programming'):
self.name = name
self.en_vocab = en_vocab
self.de_vocab = de_vocab
self.end_action = self.de_vocab.end_id
self.score_fn = score_fn
self.interpreter = interpreter
self.answer = answer
self.question_annotation = question_annotation
self.constant_value_embedding_fn = constant_value_embedding_fn
self.constants = constants
self.punish_extra_work = punish_extra_work
self.error = False
self.trigger_words_dict = trigger_words_dict
tokens = question_annotation['tokens']
if 'pos_tags' in question_annotation:
self.ques_tokens = set(tokens + question_annotation['pos_tags'])
else:
self.ques_tokens = set(tokens)
en_inputs = en_vocab.lookup(tokens)
self.n_builtin = len(de_vocab.vocab) - interpreter.max_mem
self.n_mem = interpreter.max_mem
self.n_exp = interpreter.max_n_exp
max_n_constants = self.n_mem - self.n_exp
constant_spans = []
constant_values = []
if constants is None:
constants = []
for c in constants:
constant_spans.append([-1, -1])
constant_values.append(c['value'])
if init_interp:
self.interpreter.add_constant(value=c['value'], type=c['type'])
for entity in question_annotation['entities']:
# Use encoder output at start and end (inclusive) step
# to create span embedding.
constant_spans.append(
[entity['token_start'], entity['token_end'] - 1])
constant_values.append(entity['value'])
if init_interp:
self.interpreter.add_constant(value=entity['value'],
type=entity['type'])
constant_value_embeddings = [
constant_value_embedding_fn(value) for value in constant_values
]
if len(constant_values) > (self.n_mem - self.n_exp):
tf.logging.info(
'Not enough memory slots for example {}, which has {} constants.'
.format(self.name, len(constant_values)))
constant_spans = constant_spans[:max_n_constants]
constant_value_embeddings = constant_value_embeddings[:max_n_constants]
self.context = (en_inputs, constant_spans, constant_value_embeddings,
question_annotation['features'],
question_annotation['tokens'])
# Create output features.
prop_features = question_annotation['prop_features']
self.id_feature_dict = {}
for name, id in de_vocab.vocab.iteritems():
self.id_feature_dict[id] = [0]
if name in self.interpreter.namespace:
val = self.interpreter.namespace[name]['value']
if ((isinstance(val, str) or isinstance(val, unicode))
and val in prop_features):
self.id_feature_dict[id] = prop_features[val]
# Create features to make calculation of score function easy
entities = question_annotation['entities']
for e in entities:
if e['type'] != 'datetime_list':
e['length'] = e['token_end'] - e['token_start']
else:
# For datetime entities, either token_end or token_start is incorrect,
# so need to look at the entity itself for calculating the length
# Also, we shouldn't consider 'xxxx' or 'xx' while calculating the
# entity length
e['length'] = len([
x for x in e['value'][0].replace('x', '').split('-') if x
])
entity_lengths = [e['length'] for e in entities]
max_entity_length = max(entity_lengths) if entity_lengths else 0.0
max_entity_sum = data_utils.max_sum(entity_lengths)
sum_entity_length = {
'datetime_list': 0,
'string_list': 0,
'num_list': 0
}
for e, w in zip(entities, entity_lengths):
sum_entity_length[e['type']] += w
self.entities = entities
self.ent_props = dict(max_sum=max_entity_sum,
max_length=max_entity_length,
val_ent=value_to_index(entities),
sum_entity_length=sum_entity_length)
col_features = [v[0] for v in self.id_feature_dict.values()]
self.col_props = dict(sum=sum(col_features),
max=max(col_features) if col_features else 0.0,
max_sum=data_utils.max_sum(col_features),
num=sum([i > 0 for i in col_features]))
self.cache = SearchCache(name=name, max_elements=max_cache_size)
self.use_cache = False
self.reset()
