AnAutomaticPencil's picture
data preprocessing update
a6326c7
from utils import parse, read_json_file, write_jsonl_file
import os
def get_prefix_lengths(lengths):
prefix_lengths = [0]
for length in lengths:
prefix_lengths.append(prefix_lengths[-1] + length)
return prefix_lengths
def custom_join(tokens, start, end):
joined_str = ""
for i in range(start, end):
joined_str += tokens[i]
if (
i == end - 1
or tokens[i + 1] in ["-", ","]
or tokens[i + 1].startswith("'")
or tokens[i] == "-"
):
continue
joined_str += " "
if joined_str.count('"') > 1:
start = joined_str.index('"')
end = joined_str[start + 1 :].index('"')
joined_str = (
joined_str[: start + 1]
+ joined_str[start + 1 :][:end].strip()
+ joined_str[start + 1 :][end:]
)
return joined_str
def parse_characters(utterance, tokens, character_entities):
utterance_char_idx = 0
characters = []
for sent_idx, sent in enumerate(character_entities):
prefix_lengths = get_prefix_lengths(map(len, tokens[sent_idx]))
for character in sent:
span = ", ".join(character[2:])
scan_length = prefix_lengths[character[0]]
start = utterance_char_idx
while scan_length >= 0 and start < len(utterance):
if scan_length == 0 and utterance[start] != " ":
break
if utterance[start] != " ":
scan_length -= 1
start += 1
scan_length = prefix_lengths[character[1]] - prefix_lengths[character[0]]
end = start
while scan_length > 0 and end < len(utterance):
if utterance[end] != " ":
scan_length -= 1
end += 1
characters.append({"value": span, "start": start, "end": end})
# print(utterance)
# print(tokens)
# print(character_entities)
# print(
# utterance_char_idx,
# start,
# end,
# utterance[start:end],
# tokens[sent_idx][character[0] : character[1]],
# custom_join(tokens[sent_idx], character[0], character[1]),
# )
if utterance[start:end] == "Emily- noooo":
continue
assert utterance[start:end] == custom_join(
tokens[sent_idx], character[0], character[1]
)
# update utterance char idx
scan_length = prefix_lengths[-1]
while scan_length >= 0 and utterance_char_idx < len(utterance):
if scan_length == 0 and utterance[utterance_char_idx] != " ":
break
if utterance[utterance_char_idx] != " ":
scan_length -= 1
utterance_char_idx += 1
return characters
def preprocess(args, split):
input_file = os.path.join(args.input_dir, f"character-identification-{split}.json")
if split == "trn":
split = "train"
elif split == "tst":
split = "test"
output_file = os.path.join(args.output_dir, f"{split}.jsonl")
episodes = read_json_file(input_file)["episodes"]
processed_data = []
for episode in episodes:
scenes = episode["scenes"]
for scene in scenes:
utterances = scene["utterances"]
dialog = {
"turn": "multi",
"locale": "en",
"dialog": [],
}
roles = ["#GENERAL#", "#OTHER#", "#ALL#"]
# labels = set()
for example in utterances:
utterance = example["transcript"]
tokens = example["tokens"]
character_entities = example["character_entities"]
characters = parse_characters(utterance, tokens, character_entities)
dialog["dialog"].append(
{
"roles": example["speakers"],
"utterance": utterance,
"characters": characters,
}
)
roles += example["speakers"]
# for character in characters:
# labels.add(character["value"])
# assert not len(labels - set(roles)), f"\n{sorted(labels)} \n {sorted(set(roles))}"
dialog["knowledge"] = {"type": "lsit", "value": sorted(roles)}
processed_data.append(dialog)
write_jsonl_file(processed_data, output_file)
if __name__ == "__main__":
args = parse()
preprocess(args, "trn")
preprocess(args, "dev")
preprocess(args, "tst")