def convert_spider(
spider_example: SpiderExample,
schema: Schema,
wordpiece_tokenizer,
generate_sql,
anonymize_values,
abstract_sql=False,
table_schemas: Optional[List[TableSchema]] = None,
allow_value_generation=False,
) -> Optional[NLToSQLExample]:
"""
Converts a Spider example to the standard format.
Args:
spider_example: JSON object for SPIDER example in original format.
schema: JSON object for SPIDER schema in converted format.
wordpiece_tokenizer: language.bert.tokenization.FullTokenizer instance.
generate_sql: If True, will populate SQL.
anonymize_values: If True, anonymizes values in SQL.
abstract_sql: If True, use under-specified FROM clause.
table_schemas: required if abstract_sql, list of TableSchema tuples.
allow_value_generation: Allow value generation.
Returns:
NLToSQLExample instance.
"""
if spider_example["question"] in WRONG_TRAINING_EXAMPLES:
return None
sql_query: sqlparse.sql.Statement = sqlparse.parse(
preprocess_sql(spider_example["query"].rstrip("; ").lower())
)[0]
nl_query = " ".join(spider_example["question_toks"])
example = NLToSQLExample.empty(nl_query)
# Set the input
populate_utterance(example, schema, wordpiece_tokenizer)
# Set the output
successful_copy = True
if generate_sql:
if abstract_sql:
assert table_schemas
successful_copy = abstract_sql_converters.populate_abstract_sql(
example, spider_example["query"], table_schemas, anonymize_values
)
else:
try:
successful_copy = populate_sql(sql_query, example, anonymize_values)
except ParseError:
return None
# If the example contained an unsuccessful copy action, and values should not
# be generated, then return an empty example.
if not successful_copy and not allow_value_generation:
return None
return example
def convert_spider(spider_example,
schema,
wordpiece_tokenizer,
generate_sql,
anonymize_values,
abstract_sql=False,
table_schemas=None):
"""Converts a Spider example to the standard format.
Args:
spider_example: JSON object for SPIDER example in original format.
schema: JSON object for SPIDER schema in converted format.
wordpiece_tokenizer: language.bert.tokenization.FullTokenizer instance.
generate_sql: If True, will populate SQL.
anonymize_values: If True, anonymizes values in SQL.
abstract_sql: If True, use under-specified FROM clause.
table_schemas: required if abstract_sql, list of TableSchema tuples.
Returns:
NLToSQLExample instance.
"""
if spider_example['question'] in WRONG_TRAINING_EXAMPLES:
return None
sql_query = spider_example['query'].rstrip('; ')
sql_query = sqlparse.parse(preprocess_sql(sql_query.lower()))[0]
example = NLToSQLExample()
# Set the input
populate_utterance(example, ' '.join(spider_example['question_toks']),
schema, wordpiece_tokenizer)
# Set the output
if generate_sql:
if abstract_sql:
abstract_sql_converters.populate_abstract_sql(
example, spider_example['query'], table_schemas)
else:
populate_sql(sql_query, example, anonymize_values)
return example
def convert_wikisql(input_example,
schema,
tokenizer,
generate_sql,
anonymize_values,
use_abstract_sql,
tables_schema=None,
allow_value_generation=False):
"""Converts a WikiSQL example into a NLToSQLExample."""
example = NLToSQLExample()
try:
try:
example = populate_utterance(example, input_example[0], schema,
tokenizer)
except ValueError as e:
print(e)
return None
# WikiSQL databases have a single table.
assert len(schema) == 1
# Some preprocessing of the WikiSQL SQL queries.
sql = input_example[1].rstrip('; ')
sql = sql.replace('TABLE', schema.keys()[0])
sql = sql.replace('_FIELD', '')
string_split_sql = sql.split(' ')
if string_split_sql[1].lower() in {
'count', 'min', 'max', 'avg', 'sum'
}:
# Add parentheses around the column that's an argument of any of these
# aggregate functions (because gold annotations don't have it).
sql = ' '.join(string_split_sql[0:2] +
['(', string_split_sql[2], ')'] +
string_split_sql[3:])
sql = normalize_sql(sql, replace_period=False)
try:
sql = preprocess_sql(sql)
except UnicodeDecodeError as e:
return None
sql = sql.lower()
parsed_sql = sqlparse.parse(sql)[0]
successful_copy = True
if generate_sql:
try:
if use_abstract_sql:
successful_copy = abstract_sql_converters.populate_abstract_sql(
example, sql, tables_schema, anonymize_values)
else:
successful_copy = populate_sql(parsed_sql, example,
anonymize_values)
except (ParseError, ValueError, AssertionError, KeyError,
IndexError, abstract_sql.ParseError,
abstract_sql.UnsupportedSqlError) as e:
return None
if not successful_copy and not allow_value_generation:
return None
if not example.gold_sql_query.actions:
return None
elif example.gold_sql_query.actions[-1].symbol == '=':
return None
except UnicodeEncodeError as e:
print(e)
return None
return example
def convert_wikisql(
input_example: Union[Tuple[str, str], Tuple[str, str, Any]],
schema,
tokenizer,
generate_sql: bool,
anonymize_values: bool,
use_abstract_sql: bool,
tables_schema=None,
allow_value_generation=False,
) -> Optional[NLToSQLExample]:
"""
Converts a WikiSQL example into a NLToSQLExample.
"""
if len(input_example) == 2:
# https://github.com/python/mypy/issues/1178
nl_str, sql_str = cast(Tuple[str, str], input_example)
else:
nl_str, sql_str, _ = cast(Tuple[str, str, Any], input_example)
example = NLToSQLExample.empty(nl_str)
try:
try:
populate_utterance(example, schema, tokenizer)
except ValueError as e:
print("Couldn't populate utterance in wikisql example:")
print(e)
return None
# WikiSQL databases have a single table.
assert len(schema) == 1
# Some preprocessing of the WikiSQL SQL queries.
sql = input_example[1].rstrip("; ")
sql = sql.replace("TABLE", list(schema.keys())[0])
sql = sql.replace("_FIELD", "")
string_split_sql = sql.split(" ")
if string_split_sql[1].lower() in {
"count", "min", "max", "avg", "sum"
}:
# Add parentheses around the column that's an argument of any of these
# aggregate functions (because gold annotations don't have it).
sql = " ".join(string_split_sql[0:2] +
["(", string_split_sql[2], ")"] +
string_split_sql[3:])
sql = normalize_sql(sql, replace_period=False)
try:
sql = preprocess_sql(sql)
except UnicodeDecodeError:
print("Couldn't preprocess sql in wikisql example:")
return None
sql = sql.lower()
parsed_sql = sqlparse.parse(sql)[0]
successful_copy = True
if generate_sql:
try:
if use_abstract_sql:
successful_copy = abstract_sql_converters.populate_abstract_sql(
example, sql, tables_schema, anonymize_values)
else:
successful_copy = populate_sql(parsed_sql, example,
anonymize_values)
except (
ParseError,
ValueError,
AssertionError,
KeyError,
IndexError,
abstract_sql.ParseError,
abstract_sql.UnsupportedSqlError,
):
return None
if not successful_copy and not allow_value_generation:
return None
if not example.gold_sql_query.actions:
return None
elif example.gold_sql_query.actions[-1].symbol == "=":
return None
except UnicodeEncodeError as e:
print(e)
return None
return example
def convert_michigan(
nl_query: str,
sql_str: str,
schema: Schema,
tokenizer: FullTokenizer,
generate_sql: bool,
anonymize_values: bool,
abstract_sql: bool,
table_schemas: Optional[List[TableSchema]],
allow_value_generation: bool,
) -> Optional[NLToSQLExample]:
"""
Converts a Michigan example to the standard format.
Args:
nl_query: natural language query
sql: SQL query
schema: JSON object for SPIDER schema in converted format.
wordpiece_tokenizer: language.bert.tokenization.FullTokenizer instance.
generate_sql: If True, will populate SQL.
anonymize_values: If True, anonymizes values in SQL.
abstract_sql: If True, use under-specified FROM clause.
table_schemas: required if abstract_sql, list of TableSchema tuples.
allow_value_generation: Allow value generation.
Returns:
NLToSQLExample instance.
"""
example = NLToSQLExample.empty(nl_query)
populate_utterance(example, schema, tokenizer)
# gold_sql_query =
# Set the output
successful_copy = True
if generate_sql:
if abstract_sql:
assert table_schemas
successful_copy = abstract_sql_converters.populate_abstract_sql(
example, sql_str, table_schemas, anonymize_values)
else:
sql_query: sqlparse.sql.Statement = sqlparse.parse(
preprocess_sql(sql_str.rstrip("; ").lower()))[0]
try:
successful_copy = sql_parsing.populate_sql(
sql_query, example, anonymize_values)
except sql_parsing.ParseError as e:
print(e)
successful_copy = False
# If the example contained an unsuccessful copy action, and values should not
# be generated, then return an empty example.
if not successful_copy and not allow_value_generation:
return None
return example
if generate_sql:
raise ValueError(
"Generating annotated SQL is not yet supported for Michigan datasets. "
"Tried to annotate: " + sql_query)