def test_get_valid_actions_in_world_without_date_columns(self):
question_tokens = [Token(x) for x in ['what', 'was', 'the', 'first', 'title', '?']]
table_file = self.FIXTURES_ROOT / 'data' / 'corenlp_processed_tables' / 'TEST-4.table'
table_context = TableQuestionContext.read_from_file(table_file, question_tokens)
# The table does not have a date column.
assert "date" not in table_context.column_types.values()
world = WikiTablesVariableFreeWorld(table_context)
actions = world.get_valid_actions()
assert set(actions.keys()) == {
"<r,<g,s>>",
"<r,<f,<n,r>>>",
"<r,<c,r>>",
"<r,<g,r>>",
"<r,<r,<f,n>>>",
"<r,<t,<s,r>>>",
"<n,<n,<n,d>>>",
"<r,<f,n>>",
"<r,r>",
"<r,n>",
"d",
"n",
"s",
"t",
"f",
"r",
"@start@",
}
assert set([str(type_) for type_ in world.get_basic_types()]) == {'n', 'd', 's', 'r', 't',
'f', 'g', 'c'}
check_productions_match(actions['s'],
['[<r,<g,s>>, r, f]',
'[<r,<g,s>>, r, t]'])
def forward(self, token_reps: List[torch.Tensor],
world: WikiTablesVariableFreeWorld) -> Dict:
actions = world.get_valid_actions()
actions = self._filter_abstract(actions)
prod_score_dic = self._score_prod(token_reps, actions, world)
# sketch_lfs = self.get_all_skethch_lf(actions, prod_score_dic, world)
sketches = self.get_all_sketches(actions, prod_score_dic, world)
logger.info("%s skethces generated", len(sketches))
score_list = []
for actions, score in sketches:
score_list.append(score)
score_vec = torch.stack(score_list, 0)
lf_prob = F.softmax(score_vec, dim=0)
m = Categorical(lf_prob)
lf_sample_t = m.sample()
lf_sample_idx = lf_sample_t.item()
sampled_lf_actions, sampled_score = sketches[lf_sample_idx]
sampled_log_probs = m.log_prob(lf_sample_t)
slot_rep = self._gen_slot_rep(prod_score_dic, sampled_lf_actions)
return (sampled_lf_actions, sampled_log_probs, slot_rep)
def evaluate(self, ques_rep: torch.Tensor, sampled_actions: List[str],
slot_dic: Dict, target_list: List,
world: WikiTablesVariableFreeWorld) -> Dict:
_, _token_rnn_feat_size = ques_rep.size()
assert self.token_rnn_feat_size == _token_rnn_feat_size
id2column, column2id, column_type_dic, column_reps = self.collect_column_reps(
world.table_context)
actions = world.get_valid_actions()
filtered_actions = self.filter_functions(actions)
possible_paths = self.get_all_sequences(ques_rep, column2id, column_reps, \
sampled_actions, filtered_actions, \
slot_dic, world)
max_path, max_score = possible_paths[0]
for candidate_path, candidate_score in possible_paths[1:]:
if candidate_score > max_score:
max_path = candidate_path
lf = world.get_logical_form(max_path)
if world._executor.evaluate_logical_form(lf, target_list):
return True
else:
return False
def forward_enumerate(self, token_reps: List[torch.Tensor],
world: WikiTablesVariableFreeWorld) -> Dict:
actions = world.get_valid_actions()
actions = self._filter_abstract(actions)
prod_score_dic = self._score_prod(token_reps, actions, world)
# sketch_lfs = self.get_all_skethch_lf(actions, prod_score_dic, world)
sketches = self.get_all_sketches(actions, prod_score_dic, world)
logger.info("%s skethces generated", len(sketches))
score_list = []
for actions, score in sketches:
score_list.append(score)
score_vec = torch.stack(score_list, 0)
lf_prob = F.softmax(score_vec, dim=0)
for i, (lf_actions, lf_score) in enumerate(sketches):
slot_rep = self._gen_slot_rep(prod_score_dic, lf_actions)
yield (lf_actions, torch.log(lf_prob[i]), slot_rep)
def forward(self, ques_rep: torch.Tensor, sampled_actions: List[str],
slot_dic: Dict, target_list: List,
world: WikiTablesVariableFreeWorld) -> Dict:
"""
It takes in a sampled path and finish the selection part
based on alignments to the question, table and fileter/same_as function.
Operations for selecting one row: filter_eq, filter_in
Operations for selecting multiple rows: all filters and all_rows
"""
_, _token_rnn_feat_size = ques_rep.size()
assert self.token_rnn_feat_size == _token_rnn_feat_size
id2column, column2id, column_type_dic, column_reps = self.collect_column_reps(
world.table_context)
actions = world.get_valid_actions()
filtered_actions = self.filter_functions(actions)
possible_paths = self.get_all_sequences(ques_rep, column2id, column_reps, \
sampled_actions, filtered_actions, \
slot_dic, world)
correct_lf = []
candidate_scores = []
gold_ids = []
for candidate_path, candidate_score in possible_paths:
lf = world.get_logical_form(candidate_path)
candidate_scores.append(candidate_score)
if world._executor.evaluate_logical_form(lf, target_list):
correct_lf.append(lf)
gold_ids.append(1)
else:
gold_ids.append(0)
gold_id_v = torch.FloatTensor(gold_ids)
if torch.sum(gold_id_v) == 0:
return 0
else:
score_v = torch.stack(candidate_scores, 0)
score_prob = F.softmax(score_v, 0)
reward_v = gold_id_v * score_prob
return torch.sum(reward_v, 0)
def predict(self, token_reps: List[torch.Tensor],
world: WikiTablesVariableFreeWorld) -> Dict:
actions = world.get_valid_actions()
actions = self._filter_abstract(actions)
prod_score_dic = self._score_prod(token_reps, actions, world)
# sketch_lfs = self.get_all_skethch_lf(actions, prod_score_dic, world)
sketches = self.get_all_sketches(actions, prod_score_dic, world)
logger.info("%s skethces generated", len(sketches))
score_list = []
for actions, score in sketches:
score_list.append(score)
score_vec = torch.stack(score_list, 0)
lf_prob = F.softmax(score_vec, dim=0)
max_v, max_id = torch.max(lf_prob, dim=0)
max_lf_actions, max_score = sketches[max_id]
max_log_probs = torch.log(lf_prob[max_id])
slot_rep = self._gen_slot_rep(prod_score_dic, max_lf_actions)
return (max_lf_actions, max_log_probs, slot_rep)
class TestWikiTablesVariableFreeWorld(AllenNlpTestCase):
def setUp(self):
super().setUp()
question_tokens = [
Token(x)
for x in ['what', 'was', 'the', 'last', 'year', '2013', '?']
]
self.table_file = self.FIXTURES_ROOT / 'data' / 'wikitables' / 'sample_table.tsv'
self.table_kg = TableQuestionKnowledgeGraph.read_from_file(
self.table_file, question_tokens)
self.world = WikiTablesVariableFreeWorld(self.table_kg)
def test_get_valid_actions_returns_correct_set(self):
# This test is long, but worth it. These are all of the valid actions in the grammar, and
# we want to be sure they are what we expect.
valid_actions = self.world.get_valid_actions()
assert set(valid_actions.keys()) == {
"<r,<l,s>>",
"<r,<n,<l,r>>>",
"<r,<l,r>>",
"<r,<r,<l,n>>>",
"<r,<s,<l,r>>>",
"<n,<n,<n,d>>>",
"<r,<d,<l,r>>>",
"<r,<l,n>>",
"<r,r>",
"<r,n>",
"d",
"n",
"s",
"l",
"r",
"@start@",
}
check_productions_match(valid_actions['<r,<l,s>>'], ['mode', 'select'])
check_productions_match(valid_actions['<r,<n,<l,r>>>'], [
'filter_number_equals', 'filter_number_greater',
'filter_number_greater_equals', 'filter_number_lesser',
'filter_number_lesser_equals', 'filter_number_not_equals'
])
check_productions_match(valid_actions['<r,<l,r>>'],
['argmax', 'argmin', 'same_as'])
check_productions_match(valid_actions['<r,<r,<l,n>>>'], ['diff'])
check_productions_match(valid_actions['<r,<s,<l,r>>>'],
['filter_in', 'filter_not_in'])
check_productions_match(valid_actions['<n,<n,<n,d>>>'], ['date'])
check_productions_match(valid_actions['<r,<d,<l,r>>>'], [
'filter_date_equals', 'filter_date_greater',
'filter_date_greater_equals', 'filter_date_lesser',
'filter_date_lesser_equals', 'filter_date_not_equals'
])
check_productions_match(valid_actions['<r,<l,n>>'],
['average', 'max', 'min', 'sum'])
check_productions_match(valid_actions['<r,r>'],
['first', 'last', 'next', 'previous'])
check_productions_match(valid_actions['<r,n>'], ['count'])
# These are the columns in table, and are instance specific.
check_productions_match(valid_actions['l'], [
'fb:row.row.year', 'fb:row.row.league',
'fb:row.row.avg_attendance', 'fb:row.row.division',
'fb:row.row.regular_season', 'fb:row.row.playoffs',
'fb:row.row.open_cup'
])
check_productions_match(valid_actions['@start@'], ['d', 'n', 's'])
# We merged cells and parts in SEMPRE to strings in this grammar.
check_productions_match(valid_actions['s'], [
'fb:cell.2', 'fb:cell.2001', 'fb:cell.2005', 'fb:cell.4th_round',
'fb:cell.4th_western', 'fb:cell.5th', 'fb:cell.6_028',
'fb:cell.7_169', 'fb:cell.did_not_qualify',
'fb:cell.quarterfinals', 'fb:cell.usl_a_league',
'fb:cell.usl_first_division', 'fb:part.4th', 'fb:part.western',
'fb:part.5th', '[<r,<l,s>>, r, l]'
])
check_productions_match(valid_actions['d'],
['[<n,<n,<n,d>>>, n, n, n]'])
check_productions_match(valid_actions['n'], [
'-1', '0', '1', '2013', '[<r,<l,n>>, r, l]',
'[<r,<r,<l,n>>>, r, r, l]', '[<r,n>, r]'
])
check_productions_match(valid_actions['r'], [
'all_rows', '[<r,<d,<l,r>>>, r, d, l]', '[<r,<l,r>>, r, l]',
'[<r,<n,<l,r>>>, r, n, l]', '[<r,<s,<l,r>>>, r, s, l]',
'[<r,r>, r]'
])
def test_world_processes_logical_forms_correctly(self):
logical_form = "(select (filter_in all_rows fb:cell.usl_a_league fb:row.row.league) fb:row.row.year)"
expression = self.world.parse_logical_form(logical_form)
# Cells (and parts) get mapped to strings.
assert str(expression) == "S0(F30(R,string:usl_a_league,C2),C6)"
def test_world_gets_correct_actions(self):
logical_form = "(select (filter_in all_rows fb:cell.usl_a_league fb:row.row.league) fb:row.row.year)"
expression = self.world.parse_logical_form(logical_form)
expected_sequence = [
'@start@ -> s', 's -> [<r,<l,s>>, r, l]', '<r,<l,s>> -> select',
'r -> [<r,<s,<l,r>>>, r, s, l]', '<r,<s,<l,r>>> -> filter_in',
'r -> all_rows', 's -> fb:cell.usl_a_league',
'l -> fb:row.row.league', 'l -> fb:row.row.year'
]
assert self.world.get_action_sequence(expression) == expected_sequence
def test_world_gets_logical_form_from_actions(self):
logical_form = "(select (filter_in all_rows fb:cell.usl_a_league fb:row.row.league) fb:row.row.year)"
expression = self.world.parse_logical_form(logical_form)
action_sequence = self.world.get_action_sequence(expression)
reconstructed_logical_form = self.world.get_logical_form(
action_sequence)
assert logical_form == reconstructed_logical_form
def test_world_processes_logical_forms_with_number_correctly(self):
logical_form = "(select (filter_number_greater all_rows 2013 fb:row.row.year) fb:row.row.year)"
expression = self.world.parse_logical_form(logical_form)
# Cells (and parts) get mapped to strings.
assert str(expression) == "S0(F10(R,num:2013,C6),C6)"
def test_world_processes_logical_forms_with_date_correctly(self):
logical_form = "(select (filter_date_greater all_rows (date 2013 -1 -1) fb:row.row.year) fb:row.row.year)"
expression = self.world.parse_logical_form(logical_form)
# Cells (and parts) get mapped to strings.
assert str(expression) == "S0(F20(R,T0(num:2013,num:~1,num:~1),C6),C6)"
def _get_world_with_question_tokens(
self, tokens: List[Token]) -> WikiTablesVariableFreeWorld:
table_kg = TableQuestionKnowledgeGraph.read_from_file(
self.table_file, tokens)
world = WikiTablesVariableFreeWorld(table_kg)
return world
def test_get_agenda(self):
tokens = [
Token(x)
for x in ['what', 'was', 'the', 'last', 'year', '2000', '?']
]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {
'n -> 2000', 'l -> fb:row.row.year', '<r,<l,r>> -> argmax'
}
tokens = [
Token(x) for x in [
'what', 'was', 'the', 'difference', 'in', 'attendance',
'between', 'years', '2001', 'and', '2005', '?'
]
]
world = self._get_world_with_question_tokens(tokens)
# The agenda contains strings here instead of numbers because 2001 and 2005 actually link to
# entities in the table whereas 2000 (in the previous case) does not.
assert set(world.get_agenda()) == {
's -> fb:cell.2001', 's -> fb:cell.2005', 'l -> fb:row.row.year',
'<r,<r,<l,n>>> -> diff'
}
tokens = [
Token(x) for x in [
'what', 'was', 'the', 'total', 'avg.', 'attendance', 'in',
'years', '2001', 'and', '2005', '?'
]
]
world = self._get_world_with_question_tokens(tokens)
# The agenda contains cells here instead of numbers because 2001 and 2005 actually link to
# entities in the table whereas 2000 (in the previous case) does not.
assert set(world.get_agenda()) == {
's -> fb:cell.2001', 's -> fb:cell.2005', 'l -> fb:row.row.year',
'l -> fb:row.row.avg_attendance', '<r,<l,n>> -> sum'
}
tokens = [
Token(x) for x in
['when', 'was', 'the', 'least', 'avg.', 'attendance', '?']
]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {
'l -> fb:row.row.avg_attendance', '<r,<l,r>> -> argmin'
}
tokens = [
Token(x)
for x in ['what', 'is', 'the', 'least', 'avg.', 'attendance', '?']
]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {
'l -> fb:row.row.avg_attendance', '<r,<l,n>> -> min'
}
class TestWikiTablesVariableFreeWorld(AllenNlpTestCase):
def setUp(self):
super().setUp()
question_tokens = [
Token(x)
for x in ['what', 'was', 'the', 'last', 'year', '2013', '?']
]
self.table_file = self.FIXTURES_ROOT / 'data' / 'wikitables' / 'sample_table.tagged'
self.table_context = TableQuestionContext.read_from_file(
self.table_file, question_tokens)
self.world_with_2013 = WikiTablesVariableFreeWorld(self.table_context)
usl_league_tokens = [
Token(x) for x in [
'what', 'was', 'the', 'last', 'year', 'with', 'usl', 'a',
'league', '?'
]
]
self.world_with_usl_a_league = self._get_world_with_question_tokens(
usl_league_tokens)
def _get_world_with_question_tokens(
self, tokens: List[Token]) -> WikiTablesVariableFreeWorld:
table_context = TableQuestionContext.read_from_file(
self.table_file, tokens)
world = WikiTablesVariableFreeWorld(table_context)
return world
def test_get_valid_actions_returns_correct_set(self):
# This test is long, but worth it. These are all of the valid actions in the grammar, and
# we want to be sure they are what we expect.
valid_actions = self.world_with_2013.get_valid_actions()
assert set(valid_actions.keys()) == {
"<r,<g,s>>",
"<r,<f,<n,r>>>",
"<r,<c,r>>",
"<r,<g,r>>",
"<r,<r,<f,n>>>",
"<r,<t,<s,r>>>",
"<n,<n,<n,d>>>",
"<r,<m,<d,r>>>",
"<r,<f,n>>",
"<r,r>",
"<r,n>",
"d",
"n",
"s",
"m",
"t",
"f",
"r",
"@start@",
}
check_productions_match(valid_actions['<r,<g,s>>'], ['mode', 'select'])
check_productions_match(valid_actions['<r,<f,<n,r>>>'], [
'filter_number_equals', 'filter_number_greater',
'filter_number_greater_equals', 'filter_number_lesser',
'filter_number_lesser_equals', 'filter_number_not_equals'
])
check_productions_match(valid_actions['<r,<c,r>>'],
['argmax', 'argmin'])
check_productions_match(valid_actions['<r,<g,r>>'], ['same_as'])
check_productions_match(valid_actions['<r,<r,<f,n>>>'], ['diff'])
check_productions_match(valid_actions['<r,<t,<s,r>>>'],
['filter_in', 'filter_not_in'])
check_productions_match(valid_actions['<n,<n,<n,d>>>'], ['date'])
check_productions_match(valid_actions['<r,<m,<d,r>>>'], [
'filter_date_equals', 'filter_date_greater',
'filter_date_greater_equals', 'filter_date_lesser',
'filter_date_lesser_equals', 'filter_date_not_equals'
])
check_productions_match(valid_actions['<r,<f,n>>'],
['average', 'max', 'min', 'sum'])
check_productions_match(valid_actions['<r,r>'],
['first', 'last', 'next', 'previous'])
check_productions_match(valid_actions['<r,n>'], ['count'])
# These are the columns in table, and are instance specific.
check_productions_match(valid_actions['m'], ['date_column:year'])
check_productions_match(
valid_actions['f'],
['number_column:avg_attendance', 'number_column:division'])
check_productions_match(valid_actions['t'], [
'string_column:league', 'string_column:playoffs',
'string_column:open_cup', 'string_column:regular_season'
])
check_productions_match(valid_actions['@start@'], ['d', 'n', 's'])
# The question does not produce any strings. It produces just a number.
check_productions_match(valid_actions['s'], ['[<r,<g,s>>, r, g]'])
check_productions_match(valid_actions['d'],
['[<n,<n,<n,d>>>, n, n, n]'])
check_productions_match(valid_actions['n'], [
'2013', '-1', '[<r,<f,n>>, r, f]', '[<r,<r,<f,n>>>, r, r, f]',
'[<r,n>, r]'
])
check_productions_match(valid_actions['r'], [
'all_rows', '[<r,<m,<d,r>>>, r, m, d]', '[<r,<g,r>>, r, g]',
'[<r,<c,r>>, r, c]', '[<r,<f,<n,r>>>, r, f, n]',
'[<r,<t,<s,r>>>, r, t, s]', '[<r,r>, r]'
])
def test_parsing_logical_form_with_string_not_in_question_fails(self):
logical_form_with_usl_a_league = """(select (filter_in all_rows string_column:league usl_a_league)
date_column:year)"""
logical_form_with_2013 = """(select (filter_date_greater all_rows date_column:year (date 2013 -1 -1))
date_column:year)"""
with self.assertRaises(ParsingError):
self.world_with_2013.parse_logical_form(
logical_form_with_usl_a_league)
self.world_with_usl_a_league.parse_logical_form(
logical_form_with_2013)
def test_world_processes_logical_forms_correctly(self):
logical_form = "(select (filter_in all_rows string_column:league string:usl_a_league) date_column:year)"
expression = self.world_with_usl_a_league.parse_logical_form(
logical_form)
f = types.name_mapper.get_alias
# Cells (and parts) get mapped to strings.
# Column names are mapped in local name mapping. For the global names, we can get their
# aliases from the name mapper.
assert str(
expression
) == f"{f('select')}({f('filter_in')}({f('all_rows')},C2,string:usl_a_league),C0)"
def test_world_gets_correct_actions(self):
logical_form = "(select (filter_in all_rows string_column:league string:usl_a_league) date_column:year)"
expression = self.world_with_usl_a_league.parse_logical_form(
logical_form)
expected_sequence = [
'@start@ -> s', 's -> [<r,<g,s>>, r, m]', '<r,<g,s>> -> select',
'r -> [<r,<t,<s,r>>>, r, t, s]', '<r,<t,<s,r>>> -> filter_in',
'r -> all_rows', 't -> string_column:league',
's -> string:usl_a_league', 'm -> date_column:year'
]
assert self.world_with_usl_a_league.get_action_sequence(
expression) == expected_sequence
def test_world_gets_logical_form_from_actions(self):
logical_form = "(select (filter_in all_rows string_column:league string:usl_a_league) date_column:year)"
expression = self.world_with_usl_a_league.parse_logical_form(
logical_form)
action_sequence = self.world_with_usl_a_league.get_action_sequence(
expression)
reconstructed_logical_form = self.world_with_usl_a_league.get_logical_form(
action_sequence)
assert logical_form == reconstructed_logical_form
def test_world_processes_logical_forms_with_number_correctly(self):
tokens = [
Token(x) for x in [
'when', 'was', 'the', 'attendance', 'higher', 'than', '3000',
'?'
]
]
world = self._get_world_with_question_tokens(tokens)
logical_form = """(select (filter_number_greater all_rows number_column:avg_attendance 3000)
date_column:year)"""
expression = world.parse_logical_form(logical_form)
f = types.name_mapper.get_alias
# Cells (and parts) get mapped to strings.
# Column names are mapped in local name mapping. For the global names, we can get their
# aliases from the name mapper.
assert str(
expression
) == f"{f('select')}({f('filter_number_greater')}({f('all_rows')},C6,num:3000),C0)"
def test_world_processes_logical_forms_with_date_correctly(self):
logical_form = """(select (filter_date_greater all_rows date_column:year (date 2013 -1 -1))
date_column:year)"""
expression = self.world_with_2013.parse_logical_form(logical_form)
f = types.name_mapper.get_alias
# Cells (and parts) get mapped to strings.
# Column names are mapped in local name mapping. For the global names, we can get their
# aliases from the name mapper.
assert str(expression) == \
f"{f('select')}({f('filter_date_greater')}({f('all_rows')},C0,{f('date')}(num:2013,num:~1,num:~1)),C0)"
def test_get_agenda(self):
tokens = [
Token(x)
for x in ['what', 'was', 'the', 'last', 'year', '2000', '?']
]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {
'n -> 2000', '<r,r> -> last', 'm -> date_column:year'
}
tokens = [
Token(x) for x in [
'what', 'was', 'the', 'difference', 'in', 'attendance',
'between', 'years', '2001', 'and', '2005', '?'
]
]
world = self._get_world_with_question_tokens(tokens)
# "year" column does not match because "years" occurs in the question.
assert set(world.get_agenda()) == {
'n -> 2001', 'n -> 2005', '<r,<r,<f,n>>> -> diff'
}
tokens = [
Token(x) for x in [
'what', 'was', 'the', 'total', 'avg.', 'attendance', 'in',
'years', '2001', 'and', '2005', '?'
]
]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {
'n -> 2001', 'n -> 2005', '<r,<f,n>> -> sum',
'f -> number_column:avg_attendance'
}
tokens = [
Token(x) for x in
['when', 'was', 'the', 'least', 'avg.', 'attendance', '?']
]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {
'<r,<c,r>> -> argmin', 'f -> number_column:avg_attendance'
}
tokens = [
Token(x)
for x in ['what', 'is', 'the', 'least', 'avg.', 'attendance', '?']
]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {
'<r,<f,n>> -> min', 'f -> number_column:avg_attendance'
}
tokens = [
Token(x)
for x in ['when', 'did', 'the', 'team', 'not', 'qualify', '?']
]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {'s -> string:qualify'}
tokens = [
Token(x) for x in [
'when', 'was', 'the', 'avg.', 'attendance', 'at', 'least',
'7000', '?'
]
]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {
'<r,<f,<n,r>>> -> filter_number_greater_equals',
'f -> number_column:avg_attendance', 'n -> 7000'
}
tokens = [
Token(x) for x in [
'when', 'was', 'the', 'avg.', 'attendance', 'more', 'than',
'7000', '?'
]
]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {
'<r,<f,<n,r>>> -> filter_number_greater',
'f -> number_column:avg_attendance', 'n -> 7000'
}
tokens = [
Token(x) for x in [
'when', 'was', 'the', 'avg.', 'attendance', 'at', 'most',
'7000', '?'
]
]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {
'<r,<f,<n,r>>> -> filter_number_lesser_equals',
'f -> number_column:avg_attendance', 'n -> 7000'
}
tokens = [Token(x) for x in ['what', 'was', 'the', 'top', 'year', '?']]
world = self._get_world_with_question_tokens(tokens)
assert set(
world.get_agenda()) == {'<r,r> -> first', 'm -> date_column:year'}
tokens = [
Token(x) for x in
['what', 'was', 'the', 'year', 'in', 'the', 'bottom', 'row', '?']
]
world = self._get_world_with_question_tokens(tokens)
assert set(
world.get_agenda()) == {'<r,r> -> last', 'm -> date_column:year'}
class TestWikiTablesVariableFreeWorld(AllenNlpTestCase):
def setUp(self):
super().setUp()
question_tokens = [Token(x) for x in ['what', 'was', 'the', 'last', 'year', '2013', '?']]
self.table_file = self.FIXTURES_ROOT / 'data' / 'wikitables' / 'sample_table.tagged'
self.table_context = TableQuestionContext.read_from_file(self.table_file, question_tokens)
self.world_with_2013 = WikiTablesVariableFreeWorld(self.table_context)
usl_league_tokens = [Token(x) for x in ['what', 'was', 'the', 'last', 'year', 'with', 'usl',
'a', 'league', '?']]
self.world_with_usl_a_league = self._get_world_with_question_tokens(usl_league_tokens)
def _get_world_with_question_tokens(self, tokens: List[Token]) -> WikiTablesVariableFreeWorld:
table_context = TableQuestionContext.read_from_file(self.table_file, tokens)
world = WikiTablesVariableFreeWorld(table_context)
return world
def test_get_valid_actions_returns_correct_set(self):
# This test is long, but worth it. These are all of the valid actions in the grammar, and
# we want to be sure they are what we expect.
valid_actions = self.world_with_2013.get_valid_actions()
assert set(valid_actions.keys()) == {
"<r,<g,s>>",
"<r,<f,<n,r>>>",
"<r,<c,r>>",
"<r,<g,r>>",
"<r,<r,<f,n>>>",
"<r,<t,<s,r>>>",
"<n,<n,<n,d>>>",
"<r,<m,<d,r>>>",
"<r,<f,n>>",
"<r,r>",
"<r,n>",
"d",
"n",
"s",
"m",
"t",
"f",
"r",
"@start@",
}
check_productions_match(valid_actions['<r,<g,s>>'],
['mode', 'select'])
check_productions_match(valid_actions['<r,<f,<n,r>>>'],
['filter_number_equals', 'filter_number_greater',
'filter_number_greater_equals', 'filter_number_lesser',
'filter_number_lesser_equals', 'filter_number_not_equals'])
check_productions_match(valid_actions['<r,<c,r>>'],
['argmax', 'argmin'])
check_productions_match(valid_actions['<r,<g,r>>'],
['same_as'])
check_productions_match(valid_actions['<r,<r,<f,n>>>'],
['diff'])
check_productions_match(valid_actions['<r,<t,<s,r>>>'],
['filter_in', 'filter_not_in'])
check_productions_match(valid_actions['<n,<n,<n,d>>>'],
['date'])
check_productions_match(valid_actions['<r,<m,<d,r>>>'],
['filter_date_equals', 'filter_date_greater',
'filter_date_greater_equals', 'filter_date_lesser',
'filter_date_lesser_equals', 'filter_date_not_equals'])
check_productions_match(valid_actions['<r,<f,n>>'],
['average', 'max', 'min', 'sum'])
check_productions_match(valid_actions['<r,r>'],
['first', 'last', 'next', 'previous'])
check_productions_match(valid_actions['<r,n>'],
['count'])
# These are the columns in table, and are instance specific.
check_productions_match(valid_actions['m'],
['date_column:year'])
check_productions_match(valid_actions['f'],
['number_column:avg_attendance',
'number_column:division'])
check_productions_match(valid_actions['t'],
['string_column:league',
'string_column:playoffs',
'string_column:open_cup',
'string_column:regular_season'])
check_productions_match(valid_actions['@start@'],
['d', 'n', 's'])
# The question does not produce any strings. It produces just a number.
check_productions_match(valid_actions['s'],
['[<r,<g,s>>, r, m]',
'[<r,<g,s>>, r, f]',
'[<r,<g,s>>, r, t]'])
check_productions_match(valid_actions['d'],
['[<n,<n,<n,d>>>, n, n, n]'])
check_productions_match(valid_actions['n'],
['2013',
'-1',
'[<r,<f,n>>, r, f]',
'[<r,<r,<f,n>>>, r, r, f]',
'[<r,n>, r]'])
check_productions_match(valid_actions['r'],
['all_rows',
'[<r,<m,<d,r>>>, r, m, d]',
'[<r,<g,r>>, r, m]',
'[<r,<g,r>>, r, f]',
'[<r,<g,r>>, r, t]',
'[<r,<c,r>>, r, m]',
'[<r,<c,r>>, r, f]',
'[<r,<f,<n,r>>>, r, f, n]',
'[<r,<t,<s,r>>>, r, t, s]',
'[<r,r>, r]'])
def test_get_valid_actions_in_world_without_number_columns(self):
question_tokens = [Token(x) for x in ['what', 'was', 'the', 'first', 'title', '?']]
table_file = self.FIXTURES_ROOT / 'data' / 'corenlp_processed_tables' / 'TEST-6.table'
table_context = TableQuestionContext.read_from_file(table_file, question_tokens)
# The table does not have a number column.
assert "number" not in table_context.column_types.values()
world = WikiTablesVariableFreeWorld(table_context)
actions = world.get_valid_actions()
assert set(actions.keys()) == {
"<r,<g,s>>",
"<r,<c,r>>",
"<r,<g,r>>",
"<r,<t,<s,r>>>",
"<n,<n,<n,d>>>",
"<r,<m,<d,r>>>",
"<r,r>",
"<r,n>",
"d",
"n",
"s",
"m",
"t",
"r",
"@start@",
}
assert set([str(type_) for type_ in world.get_basic_types()]) == {'n', 'd', 's', 'r', 't',
'm', 'g', 'c'}
check_productions_match(actions['s'],
['[<r,<g,s>>, r, m]',
'[<r,<g,s>>, r, t]'])
def test_get_valid_actions_in_world_without_date_columns(self):
question_tokens = [Token(x) for x in ['what', 'was', 'the', 'first', 'title', '?']]
table_file = self.FIXTURES_ROOT / 'data' / 'corenlp_processed_tables' / 'TEST-4.table'
table_context = TableQuestionContext.read_from_file(table_file, question_tokens)
# The table does not have a date column.
assert "date" not in table_context.column_types.values()
world = WikiTablesVariableFreeWorld(table_context)
actions = world.get_valid_actions()
assert set(actions.keys()) == {
"<r,<g,s>>",
"<r,<f,<n,r>>>",
"<r,<c,r>>",
"<r,<g,r>>",
"<r,<r,<f,n>>>",
"<r,<t,<s,r>>>",
"<n,<n,<n,d>>>",
"<r,<f,n>>",
"<r,r>",
"<r,n>",
"d",
"n",
"s",
"t",
"f",
"r",
"@start@",
}
assert set([str(type_) for type_ in world.get_basic_types()]) == {'n', 'd', 's', 'r', 't',
'f', 'g', 'c'}
check_productions_match(actions['s'],
['[<r,<g,s>>, r, f]',
'[<r,<g,s>>, r, t]'])
def test_get_valid_actions_in_world_without_comparable_columns(self):
question_tokens = [Token(x) for x in ['what', 'was', 'the', 'first', 'title', '?']]
table_file = self.FIXTURES_ROOT / 'data' / 'corenlp_processed_tables' / 'TEST-1.table'
table_context = TableQuestionContext.read_from_file(table_file, question_tokens)
# The table does not have date or number columns.
assert "date" not in table_context.column_types.values()
assert "number" not in table_context.column_types.values()
world = WikiTablesVariableFreeWorld(table_context)
actions = world.get_valid_actions()
assert set(actions.keys()) == {
"<r,<g,s>>",
"<r,<g,r>>",
"<r,<t,<s,r>>>",
"<n,<n,<n,d>>>",
"<r,r>",
"<r,n>",
"d",
"n",
"s",
"t",
"r",
"@start@",
}
assert set([str(type_) for type_ in world.get_basic_types()]) == {'n', 'd', 's', 'r', 't', 'g'}
def test_parsing_logical_form_with_string_not_in_question_fails(self):
logical_form_with_usl_a_league = """(select (filter_in all_rows string_column:league usl_a_league)
date_column:year)"""
logical_form_with_2013 = """(select (filter_date_greater all_rows date_column:year (date 2013 -1 -1))
date_column:year)"""
with self.assertRaises(ParsingError):
self.world_with_2013.parse_logical_form(logical_form_with_usl_a_league)
self.world_with_usl_a_league.parse_logical_form(logical_form_with_2013)
@staticmethod
def _get_alias(types_, name) -> str:
if name in types_.generic_name_mapper.common_name_mapping:
return types_.generic_name_mapper.get_alias(name)
elif name in types_.string_column_name_mapper.common_name_mapping:
return types_.string_column_name_mapper.get_alias(name)
elif name in types_.number_column_name_mapper.common_name_mapping:
return types_.number_column_name_mapper.get_alias(name)
elif name in types_.date_column_name_mapper.common_name_mapping:
return types_.date_column_name_mapper.get_alias(name)
else:
return types_.comparable_column_name_mapper.get_alias(name)
def test_world_processes_logical_forms_correctly(self):
logical_form = "(select (filter_in all_rows string_column:league string:usl_a_league) date_column:year)"
expression = self.world_with_usl_a_league.parse_logical_form(logical_form)
f = partial(self._get_alias, types)
# Cells (and parts) get mapped to strings.
# Column names are mapped in local name mapping. For the global names, we can get their
# aliases from the name mapper.
assert str(expression) == f"{f('select')}({f('filter_in')}({f('all_rows')},C2,string:usl_a_league),C0)"
def test_world_gets_correct_actions(self):
logical_form = "(select (filter_in all_rows string_column:league string:usl_a_league) date_column:year)"
expression = self.world_with_usl_a_league.parse_logical_form(logical_form)
expected_sequence = ['@start@ -> s', 's -> [<r,<g,s>>, r, m]', '<r,<g,s>> -> select',
'r -> [<r,<t,<s,r>>>, r, t, s]', '<r,<t,<s,r>>> -> filter_in',
'r -> all_rows', 't -> string_column:league', 's -> string:usl_a_league',
'm -> date_column:year']
assert self.world_with_usl_a_league.get_action_sequence(expression) == expected_sequence
def test_world_gets_logical_form_from_actions(self):
logical_form = "(select (filter_in all_rows string_column:league string:usl_a_league) date_column:year)"
expression = self.world_with_usl_a_league.parse_logical_form(logical_form)
action_sequence = self.world_with_usl_a_league.get_action_sequence(expression)
reconstructed_logical_form = self.world_with_usl_a_league.get_logical_form(action_sequence)
assert logical_form == reconstructed_logical_form
def test_world_processes_logical_forms_with_number_correctly(self):
tokens = [Token(x) for x in ['when', 'was', 'the', 'attendance', 'higher', 'than', '3000',
'?']]
world = self._get_world_with_question_tokens(tokens)
logical_form = """(select (filter_number_greater all_rows number_column:avg_attendance 3000)
date_column:year)"""
expression = world.parse_logical_form(logical_form)
f = partial(self._get_alias, types)
# Cells (and parts) get mapped to strings.
# Column names are mapped in local name mapping. For the global names, we can get their
# aliases from the name mapper.
assert str(expression) == f"{f('select')}({f('filter_number_greater')}({f('all_rows')},C6,num:3000),C0)"
def test_world_processes_logical_forms_with_date_correctly(self):
logical_form = """(select (filter_date_greater all_rows date_column:year (date 2013 -1 -1))
date_column:year)"""
expression = self.world_with_2013.parse_logical_form(logical_form)
f = partial(self._get_alias, types)
# Cells (and parts) get mapped to strings.
# Column names are mapped in local name mapping. For the global names, we can get their
# aliases from the name mapper.
assert str(expression) == \
f"{f('select')}({f('filter_date_greater')}({f('all_rows')},C0,{f('date')}(num:2013,num:~1,num:~1)),C0)"
def test_get_agenda(self):
tokens = [Token(x) for x in ['what', 'was', 'the', 'last', 'year', '2000', '?']]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {'n -> 2000',
'<r,r> -> last',
'm -> date_column:year'}
tokens = [Token(x) for x in ['what', 'was', 'the', 'difference', 'in', 'attendance',
'between', 'years', '2001', 'and', '2005', '?']]
world = self._get_world_with_question_tokens(tokens)
# "year" column does not match because "years" occurs in the question.
assert set(world.get_agenda()) == {'n -> 2001',
'n -> 2005',
'<r,<r,<f,n>>> -> diff'}
tokens = [Token(x) for x in ['what', 'was', 'the', 'total', 'avg.', 'attendance', 'in',
'years', '2001', 'and', '2005', '?']]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {'n -> 2001',
'n -> 2005',
'<r,<f,n>> -> sum',
'f -> number_column:avg_attendance'}
tokens = [Token(x) for x in ['when', 'was', 'the', 'least', 'avg.', 'attendance', '?']]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {'<r,<c,r>> -> argmin', 'f -> number_column:avg_attendance'}
tokens = [Token(x) for x in ['what', 'is', 'the', 'least', 'avg.', 'attendance', '?']]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {'<r,<f,n>> -> min', 'f -> number_column:avg_attendance'}
tokens = [Token(x) for x in ['when', 'did', 'the', 'team', 'not', 'qualify', '?']]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {'s -> string:qualify'}
tokens = [Token(x) for x in ['when', 'was', 'the', 'avg.', 'attendance', 'at', 'least',
'7000', '?']]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {'<r,<f,<n,r>>> -> filter_number_greater_equals',
'f -> number_column:avg_attendance', 'n -> 7000'}
tokens = [Token(x) for x in ['when', 'was', 'the', 'avg.', 'attendance', 'more', 'than',
'7000', '?']]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {'<r,<f,<n,r>>> -> filter_number_greater',
'f -> number_column:avg_attendance', 'n -> 7000'}
tokens = [Token(x) for x in ['when', 'was', 'the', 'avg.', 'attendance', 'at', 'most',
'7000', '?']]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {'<r,<f,<n,r>>> -> filter_number_lesser_equals',
'f -> number_column:avg_attendance', 'n -> 7000'}
tokens = [Token(x) for x in ['what', 'was', 'the', 'top', 'year', '?']]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {'<r,r> -> first', 'm -> date_column:year'}
tokens = [Token(x) for x in ['what', 'was', 'the', 'year', 'in', 'the', 'bottom', 'row',
'?']]
world = self._get_world_with_question_tokens(tokens)
assert set(world.get_agenda()) == {'<r,r> -> last', 'm -> date_column:year'}
def _create_grammar_state(self, world: WikiTablesVariableFreeWorld,
possible_actions: List[ProductionRuleArray],
linking_scores: torch.Tensor,
entity_types: torch.Tensor) -> GrammarStatelet:
"""
This method creates the GrammarStatelet object that's used for decoding. Part of
creating that is creating the `valid_actions` dictionary, which contains embedded
representations of all of the valid actions. So, we create that here as well.
The way we represent the valid expansions is a little complicated: we use a
dictionary of `action types`, where the key is the action type (like "global", "linked", or
whatever your model is expecting), and the value is a tuple representing all actions of
that type. The tuple is (input tensor, output tensor, action id). The input tensor has
the representation that is used when `selecting` actions, for all actions of this type.
The output tensor has the representation that is used when feeding the action to the next
step of the decoder (this could just be the same as the input tensor). The action ids are
a list of indices into the main action list for each batch instance.
The inputs to this method are for a `single instance in the batch`; none of the tensors we
create here are batched. We grab the global action ids from the input
``ProductionRuleArrays``, and we use those to embed the valid actions for every
non-terminal type. We use the input ``linking_scores`` for non-global actions.
Parameters
----------
world : ``WikiTablesVariableFreeWorld``
From the input to ``forward`` for a single batch instance.
possible_actions : ``List[ProductionRuleArray]``
From the input to ``forward`` for a single batch instance.
linking_scores : ``torch.Tensor``
Assumed to have shape ``(num_entities, num_question_tokens)`` (i.e., there is no batch
dimension).
entity_types : ``torch.Tensor``
Assumed to have shape ``(num_entities,)`` (i.e., there is no batch dimension).
"""
# TODO(mattg): Move the "valid_actions" construction to another method.
action_map = {}
for action_index, action in enumerate(possible_actions):
action_string = action[0]
action_map[action_string] = action_index
entity_map = {}
for entity_index, entity in enumerate(world.table_graph.entities):
entity_map[entity] = entity_index
valid_actions = world.get_valid_actions()
translated_valid_actions: Dict[str, Dict[str, Tuple[torch.Tensor,
torch.Tensor,
List[int]]]] = {}
for key, action_strings in valid_actions.items():
translated_valid_actions[key] = {}
# `key` here is a non-terminal from the grammar, and `action_strings` are all the valid
# productions of that non-terminal. We'll first split those productions by global vs.
# linked action.
action_indices = [
action_map[action_string] for action_string in action_strings
]
production_rule_arrays = [(possible_actions[index], index)
for index in action_indices]
global_actions = []
linked_actions = []
for production_rule_array, action_index in production_rule_arrays:
if production_rule_array[1]:
global_actions.append(
(production_rule_array[2], action_index))
else:
linked_actions.append(
(production_rule_array[0], action_index))
# Then we get the embedded representations of the global actions if any.
if global_actions:
global_action_tensors, global_action_ids = zip(*global_actions)
global_action_tensor = torch.cat(global_action_tensors, dim=0)
global_input_embeddings = self._action_embedder(
global_action_tensor)
if self._add_action_bias:
global_action_biases = self._action_biases(
global_action_tensor)
global_input_embeddings = torch.cat(
[global_input_embeddings, global_action_biases],
dim=-1)
global_output_embeddings = self._output_action_embedder(
global_action_tensor)
translated_valid_actions[key]['global'] = (
global_input_embeddings, global_output_embeddings,
list(global_action_ids))
# Then the representations of the linked actions.
if linked_actions:
linked_rules, linked_action_ids = zip(*linked_actions)
entities = [rule.split(' -> ')[1] for rule in linked_rules]
entity_ids = [entity_map[entity] for entity in entities]
# (num_linked_actions, num_question_tokens)
entity_linking_scores = linking_scores[entity_ids]
# (num_linked_actions,)
entity_type_tensor = entity_types[entity_ids]
# (num_linked_actions, entity_type_embedding_dim)
entity_type_embeddings = self._entity_type_decoder_embedding(
entity_type_tensor)
translated_valid_actions[key]['linked'] = (
entity_linking_scores, entity_type_embeddings,
list(linked_action_ids))
return GrammarStatelet([START_SYMBOL], translated_valid_actions,
type_declaration.is_nonterminal)
