AnAutomaticPencil's picture
data preprocessing update
a6326c7
from const import (
SUMMARY,
EMOTIONS,
EMOTION,
UTTERANCE,
ASPECTS,
TARGET,
VALUE,
OPINION,
SENTIMENT,
CATEGORY,
CHARACTERS,
DIALOG,
START,
END,
BELIEF_STATE,
DOMAIN,
INFORMED_SLOT_VALUE_TABLE,
SLOT,
VALUES,
RELATION,
SQL,
SLOT_VALUE_TABLE,
SLOTS_TO_FILL,
ROLE_RELATIONS,
REWRITTEN,
ROLES_TO_SELECT,
ACTIVE_INTENTS,
TRAIN_SPLIT,
OPTION_LABEL,
CANDIDATES,
)
from typing import Dict
import re
import random
import copy
import json
def extract_summary(dial: Dict, **kwargs):
"""
`dial` is the full dialog.
"""
return dial[SUMMARY]
def extract_turn_emotion(turn: Dict, sep: str, **kwargs):
if EMOTIONS not in turn:
return None
return sep.join(map(lambda x: x[EMOTION], turn[EMOTIONS]))
def extract_turn_emotion_wrapper(sep: str):
def extract_turn_emotion_func(turn: Dict, **kwargs):
return extract_turn_emotion(turn, sep)
return extract_turn_emotion_func
def extract_turn_utterance(turn: Dict, **kwargs):
return turn[UTTERANCE]
def extract_aspects(turn: Dict, ext_aspect_sep: str, int_aspect_sep: str):
if not turn[ASPECTS]:
return "None"
aspects = turn[ASPECTS]
tgt_seq = []
for aspect in aspects:
aspect_seq = []
if TARGET in aspect:
aspect_seq.append(aspect[TARGET][VALUE])
if CATEGORY in aspect:
aspect_seq.append(aspect[CATEGORY])
if OPINION in aspect:
aspect_seq.append(aspect[OPINION][VALUE])
if SENTIMENT in aspect:
aspect_seq.append(aspect[SENTIMENT])
tgt_seq.append(int_aspect_sep.join(aspect_seq))
return ext_aspect_sep.join(tgt_seq)
def extract_aspects_wrapper(ext_aspect_sep: str, int_aspect_sep: str):
def extract_aspects_func(turn: Dict, **kwargs):
return extract_aspects(turn, ext_aspect_sep, int_aspect_sep)
return extract_aspects_func
def rebuild_utterance_with_characters(turn: Dict, split):
if split == "train":
utterance = turn[UTTERANCE]
parts = []
pre = 0
for character in turn[CHARACTERS]:
parts.append(utterance[pre : character[START]])
parts.append(
f"[{utterance[character[START]: character[END]]} | {character[VALUE]}]"
)
pre = character[END]
parts.append(utterance[pre:])
return "".join(parts)
else:
tuples = []
for character in turn[CHARACTERS]:
tuples.append(f"{character[VALUE]}, {character[START]}, {character[END]}")
if not tuples:
return "None"
return " | ".join(tuples)
def extract_characters(example):
for turn_id, turn in enumerate(example[DIALOG]):
if CHARACTERS not in turn:
continue
for character in turn[CHARACTERS]:
yield turn_id, character[VALUE], (character[END],)
def extract_belief_state(
turn,
value_sep,
domain_sep,
slot_sep,
domain_prompt_op,
ontology=None,
do_train=True,
):
domain_bs = dict()
bs = turn[BELIEF_STATE]
# spare_bs = {domain: {slot for slot in ontology[domain]} for domain in ontology}
for state in bs:
domain = state[DOMAIN]
if domain not in domain_bs:
domain_bs[domain] = dict()
if INFORMED_SLOT_VALUE_TABLE not in state:
continue
for svp in state[INFORMED_SLOT_VALUE_TABLE]:
slot = svp[SLOT]
values = svp[VALUES]
relation = svp[RELATION]
if slot not in domain_bs[domain]:
domain_bs[domain][slot] = {"relation": relation, "values": []}
domain_bs[domain][slot]["values"] += list(map(lambda x: x[VALUE], values))
# spare_bs[domain].remove(slot)
domain_bs_list = []
for domain in domain_bs:
svp_list = []
for slot in domain_bs[domain]:
val_str = value_sep.join(domain_bs[domain][slot]["values"])
svp_list.append(f"{slot} {domain_bs[domain][slot]['relation']} {val_str}")
# control whether to add spare slots
# for slot in sorted(spare_bs[domain]):
# svp_list.append(f"{slot} = None")
if not svp_list:
continue
if do_train:
# shuffle for training
random.shuffle(svp_list)
# append a slot separator at the end to alleviate the problem of end point prediction of T5
svt_str = slot_sep.join(svp_list) + slot_sep
domain_bs_list.append(f"{domain}{domain_prompt_op}{svt_str.strip()}")
if not domain_bs_list:
return "None"
return domain_sep.join(domain_bs_list)
def extract_belief_state_wrapper(value_sep, domain_sep, slot_sep, domain_prompt_op):
def extract_belief_state_func(turn, ontology, do_train=True, **kwargs):
return extract_belief_state(
turn,
value_sep,
domain_sep,
slot_sep,
domain_prompt_op,
ontology,
do_train=do_train,
)
return extract_belief_state_func
def normalize(query: str) -> str:
def comma_fix(s):
# Remove spaces in front of commas
return s.replace(" , ", ", ")
def white_space_fix(s):
# Remove double and triple spaces
return " ".join(s.split())
def lower(s):
# Convert everything except text between (single or double) quotation marks to lower case
return re.sub(
r"\b(?<!['\"])(\w+)(?!['\"])\b", lambda match: match.group(1).lower(), s
)
def space_fix(sql: str):
def agg_fix(sql: str):
return re.sub(
r"(count|max|min|sum|avg)\s\(",
lambda match: match.group(0).replace(" ", ""),
sql,
)
def brackets_fix(sql: str):
sql = re.sub(r"\(\s", lambda match: match.group(0)[:-1], sql)
sql = re.sub(r"\s\)", lambda match: match.group(0)[1:], sql)
return sql
def double_chars_op_fix(sql: str):
return re.sub(
r"((>|<|!)\s=)",
lambda match: match.group(0).replace(" ", ""),
sql,
)
return double_chars_op_fix(brackets_fix(agg_fix(sql)))
return space_fix(comma_fix(white_space_fix(lower(query))))
def extract_sql(turn, split):
if SQL not in turn:
return None
_normalize = normalize if split == "train" else (lambda x: x)
return _normalize(turn[SQL])
def extract_slots_without_intents(turn, value_sep, slot_sep):
if SLOTS_TO_FILL not in turn or not turn[SLOTS_TO_FILL][SLOT_VALUE_TABLE]:
return "None"
slots = []
for svp in turn[SLOTS_TO_FILL][SLOT_VALUE_TABLE]:
slots.append(
svp[SLOT]
+ " "
+ svp[RELATION]
+ " "
+ value_sep.join(map(lambda x: x[VALUE], svp[VALUES]))
)
return (slot_sep.join(slots) + slot_sep).strip()
def extract_slots_without_intents_wrapper(value_sep, slot_sep):
def extract_slots_without_intents_func(turn, **kwargs):
return extract_slots_without_intents(turn, value_sep, slot_sep)
return extract_slots_without_intents_func
def extract_role_relation_without_turn(dialog, relation_sep):
return relation_sep.join(map(lambda x: x[RELATION], dialog[ROLE_RELATIONS]))
def extract_role_relation_without_turn_wrapper(relation_sep):
def extract_role_relation_without_turn_func(dialog, **kwargs):
return extract_role_relation_without_turn(dialog, relation_sep)
return extract_role_relation_without_turn_func
def extrac_rewritten(turn, **kwargs):
if REWRITTEN not in turn:
return None
return turn[REWRITTEN]
def extract_options(turn, knowledge, split=None):
if ROLES_TO_SELECT not in turn:
return None
if split == TRAIN_SPLIT:
return knowledge[turn[ROLES_TO_SELECT][0]]
else:
return json.dumps(
{OPTION_LABEL: turn[ROLES_TO_SELECT][0], CANDIDATES: knowledge}
)
# def extract_roles_wrapper(role_sep):
# def extract_roles_func(turn, knowledge, split=None):
# return extract_options(turn, know)
# return extract_roles_func
def extract_intents(turn, intent_sep):
if not turn[ACTIVE_INTENTS]:
return "None"
return intent_sep.join(
map(lambda intent: intent.replace("_", " "), turn[ACTIVE_INTENTS])
)
def extract_intents_wrapper(intent_sep):
def extract_intents_func(turn, **kwargs):
return extract_intents(turn, intent_sep)
return extract_intents_func