File size: 10,840 Bytes
a8e2bd1
e4cf54e
a8e2bd1
e4cf54e
 
a8e2bd1
e4cf54e
 
 
 
 
 
 
 
 
 
a8e2bd1
e4cf54e
a8e2bd1
e4cf54e
a8e2bd1
 
e4cf54e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a8e2bd1
 
e4cf54e
a8e2bd1
 
 
 
e4cf54e
 
a8e2bd1
 
 
 
e4cf54e
a8e2bd1
 
 
 
 
 
e4cf54e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a8e2bd1
 
e4cf54e
a8e2bd1
 
e4cf54e
 
 
 
a8e2bd1
 
 
e4cf54e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a8e2bd1
e4cf54e
a8e2bd1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e4cf54e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a8e2bd1
 
 
 
 
 
 
 
 
 
 
 
e4cf54e
 
 
 
 
 
 
a8e2bd1
 
 
 
 
e4cf54e
a8e2bd1
 
 
 
 
 
 
e4cf54e
 
 
 
 
 
a8e2bd1
 
e4cf54e
 
 
 
 
 
 
 
a8e2bd1
e4cf54e
 
 
 
 
 
a8e2bd1
 
 
 
 
 
 
 
 
e4cf54e
 
 
 
 
 
a8e2bd1
e4cf54e
 
 
 
 
 
 
 
 
a8e2bd1
 
 
 
 
 
e4cf54e
 
 
a8e2bd1
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
"""
Chinese slang / evasion lexicon layer.

General-purpose models do not recognise livestream-style slang like 收米 / 上车 /
日结, so this layer scores it with a plain word table. Three steps:

1. normalize() folds the text — fullwidth to halfwidth, zero-width characters
   stripped, homoglyphs restored (Cyrillic / Greek / lookalike Han characters),
   traditional to simplified, lowercased.
2. strip_separators() additionally removes in-word separators, which breaks the
   "split the word up" trick (叚*币 / 假-币 / 假 币).
3. score() matches the table on all three tracks and adds two bonuses:
   - combo bonus: a money/scam term AND a contact-evasion term in one message
   - evasion bonus: the term only matches AFTER normalizing / de-separating

Usage (called before the judge in bot.py):
    s, terms = lexicon.score(text)
    if s >= config.LEXICON_HARD_THRESHOLD:   # hard hit — spam, no AI call needed
        ...
    else:                                     # soft hit — feed terms to the AI
        ...

The table is extensible through config.LEXICON_EXTRA, so an operator can add
terms without touching code.

Why the evasion bonus exists
----------------------------
**Deliberate obfuscation is itself evidence of intent.** Somebody discussing or
complaining about counterfeit money has no reason to write 假 as 叚. Making that
substitution means the sender knows the word gets blocked — which is an
admission that what they are posting is the thing that gets blocked.

So the same word carries a completely different risk depending on how it was
written:
    "假币"    → could be news, a complaint, a question → weight 4, under the
                hard threshold, goes to the AI for context
    "叚*币"   → 4 + evasion bonus 3 = 7 → over the threshold, deleted at once

This does not depend on enumerating every lookalike character: swap in any rare
character or insert any separator and, as long as the folded text spells the
term, the bonus applies automatically.
"""

import re
import unicodedata

import config

# Zero-width / directional control characters (spam inserts these mid-word to
# break keyword matching).
_ZERO_WIDTH = dict.fromkeys(
    map(ord, "​‌‍‎‏‪‫‬⁠"), None
)

# Common homoglyphs: Cyrillic / Greek → Latin (letters disguised as English).
_HOMOGLYPH = {
    "а": "a", "е": "e", "о": "o", "р": "p", "с": "c", "х": "x", "у": "y",
    "ѕ": "s", "і": "i", "ј": "j", "к": "k", "н": "h", "в": "b", "м": "m", "т": "t",
    "ο": "o", "ρ": "p", "α": "a", "ν": "v", "τ": "t", "ϲ": "c",
}

# Lookalike Han characters / traditional / Japanese shinjitai → simplified.
# NFKC does **nothing** for these: it handles fullwidth and compatibility forms,
# not distinct characters that merely look alike, and it does no traditional →
# simplified conversion. 叚 (U+53DA) and 假 (U+5047) are two separate characters.
_CJK_HOMOGLYPH = {
    # Lookalike substitutions actually observed in spam
    "叚": "假", "仮": "假", "葭": "假",
    "帀": "币", "巿": "币",
    # Traditional / variant → simplified (outside NFKC's remit)
    "幣": "币", "鈔": "钞", "偽": "伪", "僞": "伪", "貨": "货", "錢": "钱",
    "髙": "高", "證": "证", "護": "护",
    "銀": "银", "帳": "账", "號": "号", "軟": "软", "體": "体",
    "電": "电", "報": "报", "聯": "联", "係": "系", "繫": "系",
    "點": "点", "擊": "击", "賣": "卖", "買": "买", "貸": "贷",
}

# Characters commonly pushed into the middle of a word as separators (the * in
# 叚*币). Stripping zero-width characters is not enough — these are visible
# characters and NFKC leaves them alone.
_SEPARATORS = re.compile(r"[\*\-_\.·・~||/\\\s、,,。::;;'\"“”‘’()()\[\]【】<>《》!!??##]+")


def normalize(text):
    """Fullwidth → halfwidth, drop zero-width, restore homoglyphs, lowercase."""
    if not text:
        return ""
    t = unicodedata.normalize("NFKC", text)              # fullwidth → halfwidth
    t = t.translate(_ZERO_WIDTH)                          # drop zero-width/marks
    t = "".join(_HOMOGLYPH.get(ch, ch) for ch in t)      # Cyrillic/Greek → Latin
    t = "".join(_CJK_HOMOGLYPH.get(ch, ch) for ch in t)  # lookalike Han → simplified
    return t.lower()


def strip_separators(text):
    """Remove in-word separators, defeating 叚*币 / 假-币 / 假 币 style splitting.

    This crosses legitimate punctuation boundaries ("美国。币安" → "美国币安"), so
    it is used **only as an auxiliary match track and every hit on it carries the
    evasion bonus** — it is never a verdict on its own.
    """
    return _SEPARATORS.sub("", text or "")


# Table: term -> (meaning, weight, category)
# Categories: money=earning/recruiting  contact=contact-detail evasion
#             pay=payment/crypto  scam=fraud scheme
#             fake=counterfeit currency/documents (serious, and the ad format is
#                  "the account itself is the contact detail")
# Higher weight = more suspicious. A single term is usually not enough to call
# something spam on its own (see LEXICON_HARD_THRESHOLD) — combinations and the
# threshold keep false positives down (收米, for example, is an ordinary word in a
# livestream-tipping context).
SLANG = {
    # ---- Earning / recruiting ----
    "收米": ("收钱", 3, "money"),
    "上车": ("入局/加入项目", 2, "money"),
    "车头": ("项目发起人", 2, "money"),
    "带单": ("带人下注/投资", 3, "money"),
    "日结": ("日结工资(刷单诈骗常见)", 2, "money"),
    "日入": ("日收入(夸张收益诱导)", 2, "money"),
    "刷单": ("刷单兼职诈骗", 3, "money"),
    "兼职": ("兼职引流", 1, "money"),
    "口子": ("放贷/诈骗渠道", 3, "money"),
    "洗码": ("赌场洗码", 3, "money"),
    "跑分": ("跑分洗钱", 3, "money"),
    "卡商": ("贩卖银行卡/账号", 3, "money"),
    "杀猪盘": ("杀猪盘诈骗", 3, "scam"),
    "反水": ("赌博返利", 3, "money"),
    "包赢": ("赌博诱导", 3, "scam"),
    "稳赚": ("虚假收益", 2, "scam"),
    "躺赚": ("虚假收益", 2, "scam"),
    "内部消息": ("荐股诈骗", 2, "scam"),
    "带你飞": ("带单诱导", 2, "money"),
    # ---- Counterfeit currency ----
    # Serious offence, but the plain spelling gets weight 4 (under the hard
    # threshold) so context still goes to the AI; obfuscated it picks up the
    # evasion bonus and hard-hits on its own.
    # Note 冥币 (joss paper, a funeral good) is deliberately NOT in this table.
    "假币": ("伪造货币", 4, "fake"),
    "假钞": ("伪造钞票", 4, "fake"),
    "伪钞": ("伪造钞票", 4, "fake"),
    "假钱": ("伪造货币", 3, "fake"),
    # These are pure trade jargon — they do not turn up in ordinary conversation,
    # so one occurrence is enough. Weight 6 clears the threshold by itself.
    "高仿钞": ("高仿伪钞", 6, "fake"),
    "仿真钞": ("仿真伪钞", 6, "fake"),
    "1:1真钞": ("伪钞话术", 6, "fake"),
    # 练功券 is a bank note-counting practice pad — a legal product — so it only
    # gets 4 and the AI decides from context.
    "练功券": ("点钞练习券(常被用作伪钞幌子)", 4, "fake"),
    # ---- Contact-detail evasion ----
    "薇": ("微信", 2, "contact"),
    "威": ("微信", 2, "contact"),
    "维": ("微信", 2, "contact"),
    "魏": ("微信", 2, "contact"),
    "vx": ("微信", 2, "contact"),
    "vxin": ("微信", 2, "contact"),
    "威信": ("微信", 2, "contact"),
    "扣扣": ("QQ", 2, "contact"),
    "企鹅": ("QQ", 2, "contact"),
    "扣v": ("加QQ/微信", 2, "contact"),
    "纸飞机": ("Telegram", 2, "contact"),
    "电报": ("Telegram", 1, "contact"),
    # A bare 飞机 cannot go in — "我坐飞机去北京" would be a false positive. Only
    # the multi-character forms that unambiguously mean a contact handle.
    "飞机号": ("Telegram 账号", 2, "contact"),
    "联系飞机": ("Telegram 联系", 2, "contact"),
    "飞机搜": ("Telegram 搜索", 2, "contact"),
    "tg号": ("Telegram 账号", 2, "contact"),
    "电报号": ("Telegram 账号", 2, "contact"),
    "蝙蝠": ("BatChat 加密聊天", 2, "contact"),
    "皮皮虾": ("加密聊天软件", 2, "contact"),
    "私我": ("私聊引流", 1, "contact"),
    "详聊": ("私下详谈引流", 1, "contact"),
    "加我": ("引流加好友", 1, "contact"),
    # ---- Payment / crypto ----
    "usdt": ("USDT 加密货币支付", 2, "pay"),
    "泰达": ("USDT", 2, "pay"),
    "承兑": ("加密货币承兑洗钱", 3, "pay"),
    "代收": ("第三方代收款", 2, "pay"),
    "四方": ("四方支付(灰产收款)", 3, "pay"),
}

# Evasion bonus: added when a term only matches after normalizing / removing
# separators. 3 is chosen so a single obfuscated fake-currency term (weight 4)
# clears the default hard threshold of 6, while the same term spelled plainly
# stays at 4 and still goes to the AI.
_EVASION_BONUS = 3


def score(text):
    """Return (total, ['term=meaning', ...]). Each term scores at most once.

    Three match tracks:
      raw    the original text, lowercased only  — direct hit, no bonus
      norm   after normalize()                   — needed folding = evasion
      strip  after normalize() + strip_separators() — same
    """
    if not text:
        return 0, []
    raw = text.lower()
    norm = normalize(text)
    strip = strip_separators(norm)
    if not norm:
        return 0, []

    table = dict(SLANG)
    extra = getattr(config, "LEXICON_EXTRA", None) or {}
    table.update(extra)

    total = 0
    matched = []
    cats = set()
    for term, meta in table.items():
        meaning, weight, cat = meta
        hit_raw = term in raw
        if not (hit_raw or term in norm or term in strip):
            continue
        total += weight
        cats.add(cat)
        if hit_raw:
            matched.append(term + "=" + meaning)
        else:
            # Only matched after folding → the sender obfuscated it on purpose,
            # which is evidence of intent in itself.
            total += _EVASION_BONUS
            matched.append(term + "=" + meaning + " (obfuscated)")

    # Combo bonus: an earning/scam/counterfeit term together with a
    # contact-evasion term is strongly suspicious.
    if cats & {"money", "scam", "pay", "fake"} and "contact" in cats:
        total += 3

    return total, matched


def is_hard_spam(text):
    """Normalized score >= the hard threshold (config.LEXICON_HARD_THRESHOLD,
    default 6) → spam outright. Used by the prefilter: a hard slang hit saves an
    AI call and defeats fullwidth / lookalike / zero-width / split-word evasion."""
    s, _ = score(text)
    return s >= getattr(config, "LEXICON_HARD_THRESHOLD", 6)