finetune_data.py

# coding=utf-8

import argparse
import json
import random
from ltp import LTP
from tqdm import tqdm

def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument("--input", type=str, required=True)
    parser.add_argument("--output", type=str, required=True)
    parser.add_argument("--ltp_model", type=str, required=True)
    parser.add_argument("--basic_hanzi", type=str, default="confusion/basic_hanzi_2500.txt")
    parser.add_argument("--sound_confusion", type=str, default="confusion/sound_confusion.txt")
    parser.add_argument("--shape_confusion", type=str, default="confusion/shape_confusion.txt")
    parser.add_argument("--same_ratio", type=float, default=0.1)
    parser.add_argument("--repeat_ratio", type=float, default=0.15)
    parser.add_argument("--delete_ratio", type=float, default=0.15)
    parser.add_argument("--sound_ratio", type=float, default=0.5)
    parser.add_argument("--shape_ratio", type=float, default=0.1)
    parser.add_argument("--whitelist", type=str, default="一二三四五六七八九十")
    parser.add_argument("--seed", type=int, default=42)
    args = parser.parse_args()
    return args

def isChinese(word):
    for ch in word:
        cp = ord(ch)
        if cp >= 0x4E00 and cp <= 0x9FA5:
            continue
        return False
    return True

def load_hanzi(path):
    hanzi = set()
    with open(path, mode="r", encoding="utf-8") as handle:
        for line in handle:
            line = line.strip()
            assert len(line) == 1
            hanzi.update(line)
    return hanzi

def load_confusion_set(path, hanzi):
    confusion_set = {}
    with open(path, mode="r", encoding="utf-8") as handle:
        for line in handle:
            line = line.strip().split()
            if len(line) < 2:   continue
            key, val = line[0], []
            for c in line[1]:
                if c in hanzi and c not in val and c != key:
                    val.append(c)
            if val:
                confusion_set[key] = val
    return confusion_set

def do_mask(sent, args):
    # 分词和词性标注
    cws, pos = args.ltp.pipeline(sent, tasks=["cws", "pos"], return_dict=False)
    
    n = len(cws)
    i = random.choice(range(n))
    word = cws[i]
    if not isChinese(word):
        i = random.choice(range(n))
        word = cws[i]
    if not isChinese(word):
        return sent

    p = random.random()
    p1 = args.same_ratio
    p2 = p1 + args.repeat_ratio
    p3 = p2 + args.delete_ratio
    p4 = p3 + args.sound_ratio
    p5 = p4 + args.shape_ratio
    assert abs(p5 - 1) < 0.001
    if p < p1:
        return sent
    if p < p2:
        # 字词冗余
        cws[i] += word
        return ''.join(cws)
    if pos[i] in ['nh', 'ns']:
        # 不修改人名地名
        return sent
    chars = list(word)
    k = random.choice(range(len(word)))
    c = chars[k]
    if c in args.whitelist:
        return sent
    if p < p3:
        if len(word) < 2:
            return sent
        chars[k] = ''
        cws[i] = ''.join(chars)
        return ''.join(cws)
    if p < p4:
        if c in args.sound_set:
            chars[k] = random.choice(args.sound_set[c])
    else:
        if c in args.shape_set:
            chars[k] = random.choice(args.shape_set[c])
    cws[i] = ''.join(chars)
    return ''.join(cws)

if __name__ == "__main__":
    args = parse_args()
    random.seed(args.seed)
    # 常用汉字集合
    hanzi = load_hanzi(args.basic_hanzi)

    # 混淆集
    args.sound_set = load_confusion_set(args.sound_confusion, hanzi)
    args.shape_set = load_confusion_set(args.shape_confusion, hanzi)
    args.hanzi = list(hanzi)

    # ltp中文分词模型
    args.ltp = LTP(args.ltp_model)

    output = open(args.output, mode="w")
    with open(args.input, mode="r", encoding="utf-8") as handle:
        for line in tqdm(handle):
            sent = line.strip()
            if len(sent) < 4:
                continue
            source = do_mask(sent, args)
            label = int(source != sent)
            output.write(
                json.dumps({"source": source, "target": sent, "label": label}, ensure_ascii=False)
            )
            output.write("\n")