DialogZoo / src /preprocess /Google_Simulated_Dialogue.py
AnAutomaticPencil's picture
data preprocessing update
a6326c7
from utils import read_json_file, write_jsonl_file, parse, write_json_file
import os
def get_aligned_index(text, non_space_len, include_tail_space=True):
cursor = 0
while non_space_len >= 0 and cursor < len(text):
if not include_tail_space and non_space_len == 0:
break
if include_tail_space and non_space_len == 0 and text[cursor] != " ":
break
if text[cursor] != " ":
non_space_len -= 1
cursor += 1
return cursor
def parse_slot2index(slots, text, tokens):
slot2index = dict()
prefix_length = [0]
for token in tokens:
prefix_length.append(prefix_length[-1] + len(token))
for trip in slots:
start = get_aligned_index(text, prefix_length[trip["start"]])
end = get_aligned_index(text, prefix_length[trip["exclusive_end"]], False)
if trip["slot"] not in slot2index:
slot2index[trip["slot"]] = dict()
value = text[start:end]
# if text not in [
# "there are none available for 8 pm , but there is one at 6.15 pm . would you like 6.15 pm instead ?",
# "8 pm is unavailable at 8 pm has been declined by oren hummus . you can choose another time such as 7.15 pm , 7.30 pm , or 8.30 pm .",
# "8 pm is not available on wednesday . i was unable to book your table on wednesday night at cetrella for 3 people .",
# ]:
# assert value not in slot2index[trip["slot"]], value
if value not in slot2index[trip["slot"]]:
slot2index[trip["slot"]][value] = (start, end)
return slot2index
def get_slot_index(slot2index, value, slot):
if slot not in slot2index or value not in slot2index[slot]:
# print(slot, value)
return -1, -1
return slot2index[slot][value]
def parse_dialogue_acts(acts, utter_dict, domain):
dialog_acts = []
text = utter_dict["text"]
# print(utter_dict)
# print(acts)
# print()
slot2index = parse_slot2index(utter_dict["slots"], text, utter_dict["tokens"])
act_slot_values = dict()
for act in acts:
if act["type"] not in act_slot_values:
act_slot_values[act["type"]] = dict()
if "slot" in act:
if act["slot"] not in act_slot_values[act["type"]]:
act_slot_values[act["type"]][act["slot"]] = []
if "value" in act:
act_slot_values[act["type"]][act["slot"]].append(act["value"])
for act in act_slot_values:
svt = []
for slot, values in act_slot_values[act].items():
svp = {
"slot": slot,
}
if values:
svp["values"] = []
svp["relation"] = "="
for value in values:
start, end = get_slot_index(slot2index, value, slot)
if start != -1:
assert value == text[start:end], f"{value} {text[start: end]}"
svp["values"].append(
{"value": value, "start": start, "end": end}
)
else:
svp["values"].append({"value": value})
svt.append(svp)
dialog_acts.append({"act": act, "slot_value_table": svt, "domain": domain})
return dialog_acts
def parse_dialogue_state(state, intent, domain, schema):
svt = []
for pair in state:
svt.append(
{
"slot": pair["slot"],
"values": [{"value": pair["value"]}],
"relation": "=",
}
)
schema[domain].add(pair["slot"])
dialog_state = [
{"intent": intent, "informed_slot_value_table": svt, "domain": domain}
]
return dialog_state, schema
def reformat(args, file, domain):
path = args.input_dir + "/" + file + ".json"
data = read_json_file(path)
processed_data = []
schema = {domain: set()}
for origin_dial in data:
dialog = {"turn": "multi", "locale": "en", "dialog": []}
origin_turns = origin_dial["turns"]
intent = origin_turns[0]["user_intents"]
assert len(intent) == 1
intent = intent[0]
for origin_turn in origin_turns:
# system
if "system_utterance" in origin_turn:
new_turn = {
"roles": ["SYSTEM"],
"utterance": origin_turn["system_utterance"]["text"],
"dialog_acts": parse_dialogue_acts(
origin_turn["system_acts"],
origin_turn["system_utterance"],
domain,
),
}
dialog["dialog"].append(new_turn)
# user
bs, schema = parse_dialogue_state(
origin_turn["dialogue_state"], intent, domain, schema
)
new_turn = {
"roles": ["USER"],
"utterance": origin_turn["user_utterance"]["text"],
"dialog_acts": parse_dialogue_acts(
origin_turn["user_acts"], origin_turn["user_utterance"], domain
),
"belief_state": bs,
}
dialog["dialog"].append(new_turn)
processed_data.append(dialog)
write_jsonl_file(processed_data, args.output_dir + "/" + file + ".jsonl")
schema[domain] = sorted(list(schema[domain]))
ontology = {domain: {slot: True for slot in schema[domain]}}
write_json_file(ontology, os.path.join(args.output_dir, f"{file}_ontology.json"))
def preprocess(args, domain):
reformat(args, "train", domain)
reformat(args, "dev", domain)
reformat(args, "test", domain)
if __name__ == "__main__":
args = parse()
preprocess(args, args.domain)