入声字情感分析：入声字的情感与舒声字无明显差别

.gitignore

/zh_en.linux
/char.csv
/词表.txt
/结果.txt

README.md

与舒声字相比，入声字表示的情感更负面吗？

wget https://github.com/Heptagon196/Dict/raw/master/dic/zh_en.linux
wget https://github.com/CanCLID/rime-cantonese-upstream/raw/d82d3e3e5fc3d39cc3ec67116385e6be5ec37b17/char.csv
python preprocess.py
python 入声字情感分析.py

结果：

舒声字平均分: 60.90
入声字平均分: 59.42

结果表明入声字的情感与舒声字无明显差别。

preprocess.py

from collections import defaultdict
import re

d = defaultdict(list)

# https://ayaka.shn.hk/hanregex/zh-CN/
def has_han(s):
    return bool(re.search(r'[\u3006\u3007\u4e00-\u9fff\u3400-\u4dbf\U00020000-\U0002a6df\U0002a700-\U0002ebef\U00030000-\U0003134f]', s))

# workaround for cases like 'AAAA (BBBB; CCCC); DDDD'
def convert_all_nested_semicolon_to_正(s):
    last_s = s
    while True:
        s = re.sub(r'(\([^()]*);([^()]*\))', r'\1正\2', s)
        if s == last_s:
            break
        last_s = s
    return s

def convert_正_to_semicolon(s):
    return s.replace('正', ';')

with open('zh_en.linux', encoding='utf-8') as f:
    try:
        while True:
            word = next(f).rstrip('\n')
            explanation = next(f).rstrip('\n')

            if len(word) != 1:
                continue

            assert explanation[:2] == r'\n'
            explanation = explanation[2:]

            explanation = convert_all_nested_semicolon_to_正(explanation)
            items = explanation.split('; ')
            items = [convert_正_to_semicolon(item) for item in items]

            for item in items:
                if not has_han(item):
                    d[word].append(item)
    except StopIteration:
        pass

with open('词表.txt', 'w', encoding='utf-8') as f:
    for k, vs in d.items():
        vfs = [v for v in vs if v[0] != '(']
        v = vfs[0] if vfs else vs[0]
        assert '\t' not in v
        print(k, v, sep='\t', file=f)

requirements.txt

tqdm
ToJyutping==0.2.1
numpy==1.22.3
tensorflow==2.10.0
transformers==4.18.0

入声字情感分析.py

import ToJyutping
from tqdm import tqdm
from transformers import pipeline
import numpy as np
import re

classifier = pipeline('sentiment-analysis')

def handle_one_result(result):
    if result['label'] == 'POSITIVE':
        score = result['score'] * 100
        assert score > 50.
    else:
        score = 100. - result['score'] * 100
        assert score < 50.
    return score

def 分块(lst, n):
    return [lst[i:i+n] for i in range(0, len(lst), n)]

def 根据粤拼判断入声(粵拼: str) -> bool:
    return bool(re.search('[ptk]\d$', 粵拼))

汉字_英文列表 = []
英文列表 = []

with open('词表.txt', encoding='utf-8') as f:
    for line in f:
        汉字, 英文 = line.rstrip('\n').split('\t')
        汉字_英文列表.append((汉字, 英文))
        英文列表.append(英文)

分块英文列表 = 分块(英文列表, 64)

情感分析结果列表 = []

for 英文块 in tqdm(分块英文列表):
    情感分析结果块 = classifier(英文块)
    情感分析结果块 = list(map(handle_one_result, 情感分析结果块))
    情感分析结果列表.extend(情感分析结果块)

舒声字得分列表 = []
入声字得分列表 = []

with open('结果.txt', 'w', encoding='utf-8') as f:
    for (汉字, 英文), 情感分析结果 in zip(汉字_英文列表, 情感分析结果列表):
        粵拼 = ToJyutping.get_jyutping_text(汉字)
        if not 粵拼:
            continue  # 不处理没有读音的情况
        是入声字 = 根据粤拼判断入声(粵拼)
        if not 是入声字:
            舒声字得分列表.append(情感分析结果)
        else:
            入声字得分列表.append(情感分析结果)
        print(汉字, 粵拼, '舒' if not 是入声字 else '入', 英文, 情感分析结果, sep='\t', file=f)

舒声字平均分 = np.array(舒声字得分列表).mean()
入声字平均分 = np.array(入声字得分列表).mean()

print(f'舒声字平均分：{舒声字平均分:.2f}')
print(f'入声字平均分：{入声字平均分:.2f}')