def _add_target_prob_to_res(self, res, row_data):
res[Target.TARGET_PROB_FIELD] = parse_json_array(
row_data[DFColumn.TARGET_PROBS])
res[Target.TARGET_LABEL_FIELD] = parse_json_array(
row_data[DFColumn.TARGET_LABELS])
res[Target.TARGET_LOGITS_FIELD] = parse_json_array(
row_data[DFColumn.TARGET_LOGITS])
def preprocess_row(self, row_data: Dict[str, Any]) -> Dict[str, Any]:
sequence = data.parse_json_array(row_data[DFColumn.UTTERANCE])
features_list = [
self.featurizer.featurize(InputRecord(raw_text=utterance))
for utterance in sequence
]
return {
# features
DatasetFieldName.TEXT_FIELD:
[utterance.tokens for utterance in features_list],
# labels
DatasetFieldName.DOC_LABEL_FIELD:
row_data[DFColumn.DOC_LABEL],
DatasetFieldName.UTTERANCE_FIELD:
row_data[DFColumn.UTTERANCE],
}
def preprocess_row(self, row_data: Dict[str, Any]) -> Dict[str, Any]:
sequence = data.parse_json_array(row_data[DFColumn.UTTERANCE])
features_list = [
self.featurizer.featurize(InputRecord(raw_text=utterance))
for utterance in sequence
]
res = {
# features
ModelInput.WORD_FEAT:
[utterance.tokens for utterance in features_list],
# labels
DatasetFieldName.DOC_LABEL_FIELD:
row_data[DFColumn.DOC_LABEL],
DatasetFieldName.UTTERANCE_FIELD:
row_data[DFColumn.UTTERANCE],
}
if DFColumn.DENSE_FEAT in row_data:
res[ModelInput.DENSE_FEAT] = row_data.get(DFColumn.DENSE_FEAT)
return res
def preprocess_row(self, row_data: Dict[str, Any]) -> Dict[str, Any]:
"""Preprocess steps for a single input row: 1. apply tokenization to a
sequence of utterances; 2. process dictionary features to align with
the last utterance. 3. align word labels with the last utterance.
Args:
row_data (Dict[str, Any]): Dict of one row data with column names as keys.
Keys includes "doc_label", "word_label", "text", "dict_feat",
"word weight" and "doc weight".
Returns:
Dict[str, Any]: Preprocessed dict of one row data includes:
"seq_word_feat" (list of list of string)
tokenized words of sequence of utterances
"word_feat" (list of string)
tokenized words of last utterance
"raw_word_label" (string)
raw word label
"token_range" (list of tuple)
token ranges of word labels, each tuple contains the start
position index and the end position index
"utterance" (list of string)
raw utterances
"word_label" (list of string)
list of labels of words in last utterance
"doc_label" (string)
doc label for intent classification
"word_weight" (float)
weight of word label
"doc_weight" (float)
weight of document label
"dict_feat" (tuple, optional)
tuple of three lists, the first is the label of each words,
the second is the weight of the feature, the third is the
length of the feature.
"""
sequence = data.parse_json_array(row_data[RawData.TEXT])
# ignore dictionary feature for context sentences other than the last one
features_list = [
self.featurizer.featurize(InputRecord(raw_text=utterance))
for utterance in sequence[:-1]
]
# adding dictionary feature for the last (current) message
features_list.append(
self.featurizer.featurize(
InputRecord(
raw_text=sequence[-1],
raw_gazetteer_feats=row_data.get(ModelInput.DICT, ""),
)))
res = {
# features
ModelInput.SEQ: [utterance.tokens for utterance in features_list],
ModelInput.TEXT:
features_list[-1].tokens,
ModelInput.DICT: (
features_list[-1].gazetteer_feats,
features_list[-1].gazetteer_feat_weights,
features_list[-1].gazetteer_feat_lengths,
),
ModelInput.CHAR:
features_list[-1].characters,
ModelInput.PRETRAINED:
features_list[-1].pretrained_token_embedding,
# labels
DocLabelConfig._name:
row_data[RawData.DOC_LABEL],
# extra data
# TODO move the logic to FloatField
ExtraField.DOC_WEIGHT:
row_data.get(RawData.DOC_WEIGHT) or 1.0,
ExtraField.WORD_WEIGHT:
row_data.get(RawData.WORD_WEIGHT) or 1.0,
ExtraField.RAW_WORD_LABEL:
row_data[RawData.WORD_LABEL],
ExtraField.UTTERANCE:
row_data[RawData.TEXT],
ExtraField.TOKEN_RANGE:
features_list[-1].token_ranges,
}
if WordLabelConfig._name in self.labels:
# TODO move it into word label field
res[WordLabelConfig._name] = data.align_slot_labels(
features_list[-1].token_ranges,
row_data[RawData.WORD_LABEL],
self.labels[WordLabelConfig._name].use_bio_labels,
)
return res
