import os from utils import parse, read_json_file, write_jsonl_file, choices import re from tqdm import tqdm import random options = list(range(4)) def refine_roles_of_dialog(example): final = example["options"][0][0] assert final.lower() in ["m", "f"] lowercase = True if final in ["M", "F"]: lowercase = False if lowercase: turns = example["article"].split("m ; f :") roles = ["m :", "f :"] else: turns = example["article"].split("M;F:") roles = ["M:", "F:"] role_idx = 0 if final.lower() == "m" else 1 role_idx = (role_idx + (len(turns) % 2) + 1) % 2 new_turns = [] for turn in turns: if not turn: continue new_turns.append(roles[role_idx]) new_turns.append(turn) role_idx = 1 - role_idx return new_turns def get_an_order_of_choices(label, sep): wrong_choices = options[:label] + options[label + 1 :] random.shuffle(wrong_choices) choices_order = [label] + wrong_choices return sep.join([choices[idx] for idx in choices_order]) def preprocess(args, split, part): indir = os.path.join(os.path.join(args.input_dir, part), split) outfile = os.path.join(os.path.join(args.output_dir, part), f"{split}.jsonl") processed_data = [] for filename in tqdm(os.listdir(indir)): filepath = os.path.join(indir, filename) example = read_json_file(filepath) dial = {"turn": "multi", "locale": "en", "dialog": []} if "plus" not in part: turns = re.split("([mf] :)", example["article"]) else: turns = re.split("([MF]:)", example["article"]) if not turns[0]: turns = turns[1:] assert len(turns) % 2 == 0, example else: turns = refine_roles_of_dialog(example) # print(turns) # print(example) for i in range(0, len(turns), 2): role = turns[i] utterance = turns[i + 1] if "plus" not in part: assert ( len(role) == 3 and role[0] in ["m", "f"] and role[1] == " " and role[2] == ":" ) else: assert len(role) == 2 and role[0] in ["M", "F"] and role[1] == ":" if role[0].lower() == "m": role = "male" else: role = "female" dial["dialog"].append({"roles": [role], "utterance": utterance.strip()}) dial["knowledge"] = {"type": "dict", "value": {}} for idx, option in enumerate(example["options"]): role, utterance = option.split(":", 1) role = role.strip().lower() assert role in ["m", "f"] if role == "m": role = "male" else: role = "female" # utterance = f"{role}: {utterance.strip()}" # dial["dialog"].append( # {"roles": [f"{choices[idx]} choice"], "utterance": utterance} # ) dial["knowledge"]["value"][chr(ord("A") + idx)] = utterance.strip() # label = ord(example["answers"]) - ord("A") # assert 0 <= label < 4, example["answers"] # This task requires to predicts the order of choices. # NOTE: we put the correct answer at the beginning and other options are shuffled. # dial["dialog"][-1]["roles_to_select"] = [get_an_order_of_choices(label, ", ")] dial["dialog"][-1]["roles_to_select"] = [example["answers"]] processed_data.append(dial) write_jsonl_file(processed_data, outfile) if __name__ == "__main__": args = parse() random.seed(args.seed) preprocess(args, "train", "mutual") preprocess(args, "dev", "mutual") preprocess(args, "train", "mutual_plus") preprocess(args, "dev", "mutual_plus")