def search(
tables_directory: str,
data: JsonDict,
output_path: str,
max_path_length: int,
max_num_logical_forms: int,
use_agenda: bool,
output_separate_files: bool,
conservative_agenda: bool,
) -> None:
print(f"Starting search with {len(data)} instances", file=sys.stderr)
language_logger = logging.getLogger("allennlp.semparse.domain_languages.wikitables_language")
language_logger.setLevel(logging.ERROR)
tokenizer = SpacyTokenizer()
if output_separate_files and not os.path.exists(output_path):
os.makedirs(output_path)
if not output_separate_files:
output_file_pointer = open(output_path, "w")
for instance_data in data:
utterance = instance_data["question"]
question_id = instance_data["id"]
if utterance.startswith('"') and utterance.endswith('"'):
utterance = utterance[1:-1]
# For example: csv/200-csv/47.csv -> tagged/200-tagged/47.tagged
table_file = instance_data["table_filename"].replace("csv", "tagged")
target_list = instance_data["target_values"]
tokenized_question = tokenizer.tokenize(utterance)
table_file = f"{tables_directory}/{table_file}"
context = TableQuestionContext.read_from_file(table_file, tokenized_question)
world = WikiTablesLanguage(context)
walker = ActionSpaceWalker(world, max_path_length=max_path_length)
correct_logical_forms = []
if use_agenda:
agenda = world.get_agenda(conservative=conservative_agenda)
allow_partial_match = not conservative_agenda
all_logical_forms = walker.get_logical_forms_with_agenda(
agenda=agenda, max_num_logical_forms=10000, allow_partial_match=allow_partial_match
)
else:
all_logical_forms = walker.get_all_logical_forms(max_num_logical_forms=10000)
for logical_form in all_logical_forms:
if world.evaluate_logical_form(logical_form, target_list):
correct_logical_forms.append(logical_form)
if output_separate_files and correct_logical_forms:
with gzip.open(f"{output_path}/{question_id}.gz", "wt") as output_file_pointer:
for logical_form in correct_logical_forms:
print(logical_form, file=output_file_pointer)
elif not output_separate_files:
print(f"{question_id} {utterance}", file=output_file_pointer)
if use_agenda:
print(f"Agenda: {agenda}", file=output_file_pointer)
if not correct_logical_forms:
print("NO LOGICAL FORMS FOUND!", file=output_file_pointer)
for logical_form in correct_logical_forms[:max_num_logical_forms]:
print(logical_form, file=output_file_pointer)
print(file=output_file_pointer)
if not output_separate_files:
output_file_pointer.close()
def text_to_instance(self, # type: ignore
logical_forms: List[str],
table_lines: List[List[str]],
question: str) -> Instance:
# pylint: disable=arguments-differ
tokenized_question = self._tokenizer.tokenize(question.lower())
tokenized_question.insert(0, Token(START_SYMBOL))
tokenized_question.append(Token(END_SYMBOL))
question_field = TextField(tokenized_question, self._question_token_indexers)
table_context = TableQuestionContext.read_from_lines(table_lines, tokenized_question)
world = WikiTablesLanguage(table_context)
action_sequences_list: List[List[str]] = []
action_sequence_fields_list: List[TextField] = []
for logical_form in logical_forms:
try:
action_sequence = world.logical_form_to_action_sequence(logical_form)
action_sequence = reader_utils.make_bottom_up_action_sequence(action_sequence,
world.is_nonterminal)
action_sequence_field = TextField([Token(rule) for rule in action_sequence],
self._rule_indexers)
action_sequences_list.append(action_sequence)
action_sequence_fields_list.append(action_sequence_field)
except ParsingError as error:
logger.debug(f'Parsing error: {error.message}, skipping logical form')
logger.debug(f'Question was: {question}')
logger.debug(f'Logical form was: {logical_form}')
logger.debug(f'Table info was: {table_lines}')
except:
logger.error(logical_form)
raise
if not action_sequences_list:
return None
all_production_rule_fields: List[List[Field]] = []
for action_sequence in action_sequences_list:
all_production_rule_fields.append([])
for production_rule in action_sequence:
_, rule_right_side = production_rule.split(' -> ')
is_global_rule = not world.is_instance_specific_entity(rule_right_side)
field = ProductionRuleField(production_rule, is_global_rule=is_global_rule)
all_production_rule_fields[-1].append(field)
action_field = ListField([ListField(production_rule_fields) for production_rule_fields in
all_production_rule_fields])
fields = {'action_sequences': ListField(action_sequence_fields_list),
'target_tokens': question_field,
'world': MetadataField(world),
'actions': action_field}
return Instance(fields)
def text_to_instance(
self, # type: ignore
question: str,
table_lines: List[List[str]],
target_values: List[str] = None,
offline_search_output: List[str] = None) -> Instance:
"""
Reads text inputs and makes an instance. We pass the ``table_lines`` to ``TableQuestionContext``, and that
method accepts this field either as lines from CoreNLP processed tagged files that come with the dataset,
or simply in a tsv format where each line corresponds to a row and the cells are tab-separated.
Parameters
----------
question : ``str``
Input question
table_lines : ``List[List[str]]``
The table content optionally preprocessed by CoreNLP. See ``TableQuestionContext.read_from_lines``
for the expected format.
target_values : ``List[str]``, optional
Target values for the denotations the logical forms should execute to. Not required for testing.
offline_search_output : ``List[str]``, optional
List of logical forms, produced by offline search. Not required during test.
"""
# pylint: disable=arguments-differ
tokenized_question = self._tokenizer.tokenize(question.lower())
question_field = TextField(tokenized_question,
self._question_token_indexers)
metadata: Dict[str, Any] = {
"question_tokens": [x.text for x in tokenized_question]
}
table_context = TableQuestionContext.read_from_lines(
table_lines, tokenized_question)
target_values_field = MetadataField(target_values)
world = WikiTablesLanguage(table_context)
world_field = MetadataField(world)
# Note: Not passing any featre extractors when instantiating the field below. This will make
# it use all the available extractors.
table_field = KnowledgeGraphField(
table_context.get_table_knowledge_graph(),
tokenized_question,
self._table_token_indexers,
tokenizer=self._tokenizer,
include_in_vocab=self._use_table_for_vocab,
max_table_tokens=self._max_table_tokens)
production_rule_fields: List[Field] = []
for production_rule in world.all_possible_productions():
_, rule_right_side = production_rule.split(' -> ')
is_global_rule = not world.is_instance_specific_entity(
rule_right_side)
field = ProductionRuleField(production_rule,
is_global_rule=is_global_rule)
production_rule_fields.append(field)
action_field = ListField(production_rule_fields)
fields = {
'question': question_field,
'metadata': MetadataField(metadata),
'table': table_field,
'world': world_field,
'actions': action_field,
'target_values': target_values_field
}
# We'll make each target action sequence a List[IndexField], where the index is into
# the action list we made above. We need to ignore the type here because mypy doesn't
# like `action.rule` - it's hard to tell mypy that the ListField is made up of
# ProductionRuleFields.
action_map = {
action.rule: i
for i, action in enumerate(action_field.field_list)
} # type: ignore
if offline_search_output:
action_sequence_fields: List[Field] = []
for logical_form in offline_search_output:
try:
action_sequence = world.logical_form_to_action_sequence(
logical_form)
index_fields: List[Field] = []
for production_rule in action_sequence:
index_fields.append(
IndexField(action_map[production_rule],
action_field))
action_sequence_fields.append(ListField(index_fields))
except ParsingError as error:
logger.debug(
f'Parsing error: {error.message}, skipping logical form'
)
logger.debug(f'Question was: {question}')
logger.debug(f'Logical form was: {logical_form}')
logger.debug(f'Table info was: {table_lines}')
continue
except KeyError as error:
logger.debug(
f'Missing production rule: {error.args}, skipping logical form'
)
logger.debug(f'Question was: {question}')
logger.debug(f'Table info was: {table_lines}')
logger.debug(f'Logical form was: {logical_form}')
continue
except:
logger.error(logical_form)
raise
if len(action_sequence_fields
) >= self._max_offline_logical_forms:
break
if not action_sequence_fields:
# This is not great, but we're only doing it when we're passed logical form
# supervision, so we're expecting labeled logical forms, but we can't actually
# produce the logical forms. We should skip this instance. Note that this affects
# _dev_ and _test_ instances, too, so your metrics could be over-estimates on the
# full test data.
return None
fields['target_action_sequences'] = ListField(
action_sequence_fields)
if self._output_agendas:
agenda_index_fields: List[Field] = []
for agenda_string in world.get_agenda(conservative=True):
agenda_index_fields.append(
IndexField(action_map[agenda_string], action_field))
if not agenda_index_fields:
agenda_index_fields = [IndexField(-1, action_field)]
fields['agenda'] = ListField(agenda_index_fields)
return Instance(fields)
def _create_grammar_state(self, world: WikiTablesLanguage,
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 : ``WikiTablesLanguage``
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_nonterminal_productions()
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,
world.is_nonterminal)