dwarf-fortress-annals/dfannals/deslop.py

"""AI 味机械诊断：把「文笔是否自然」变成可度量的指标。

规则来自 oh-story-claudecode 的 story-deslop 与 banned-words（MIT），做了两处适配：

1. **分层**。原文把「仿佛、一丝、坚定、冰冷」放同一张一级表，但「坚定」「冰冷」在
   史书叙事里是正常词（矮人要塞的"冰冷"可能是字面温度），一律判罪会刷满假阳性。
   因此分成硬伤词（几乎只在 AI 文本里成串出现）与密度词（真人也用，只看密度）。
2. **按密度而非布尔判定**。AI 味是密度问题：出现一次是风格，出现二十次是模板。
   输出「每千字命中数」，供闸门与扩写模型参考。

语义类判断（比喻是否套话、是否把一件事掰成三段写）机械检查假阳性太高，
交给扩写模型和自评闸门，这里不猜。
"""
from __future__ import annotations

import re
from dataclasses import dataclass, field

from dfannals.chronicle import cjk_length

# 硬伤词：出现在正常中文里的概率很低，命中一处就该改
HARD_WORDS: dict[str, tuple[str, ...]] = {
    "情态": ("仿佛", "犹如", "宛若", "如同", "一丝", "一抹", "些许", "几分", "隐约",
             "毫无征兆", "几不可闻", "微不可察"),
    "动作": ("深吸一口气", "不禁"),
    "表情": ("眼中闪过", "嘴角勾起", "眉头微皱", "眉眼低垂", "瞳孔微缩", "瞳孔收缩",
             "瞳孔一缩", "指节泛白", "眼神锐利", "目光锐利"),
    "心理": ("心中一动", "心头一震", "心下了然", "心中暗道", "心底泛起", "不由得", "心中一凛"),
    "判断": ("不容置疑", "不容置喙", "不易察觉", "显而易见", "毫无疑问", "不可否认", "前所未有"),
    "过渡": ("不由自主", "情不自禁", "自然而然", "话锋一转", "闪烁着光芒"),
}

# 密度词：真人写作也用，只有成串出现才算模板腔
SOFT_WORDS: dict[str, tuple[str, ...]] = {
    "形容": ("坚定", "狡黠", "深邃", "凛冽", "冰冷"),
    "语境敏感": ("突然", "陡然", "骤然", "猛然", "好像", "似乎", "瞬间", "猛地", "死死地"),
    "弱化副词": ("缓缓", "微微", "轻轻", "淡淡"),
}

# 句式与标点：每条是（规则名，说明，正则）
PATTERNS: tuple[tuple[str, str, str], ...] = (
    ("不是A而是B", "最毒的 AI 句式，直接写 B", r"不是[^，。；！？\n]{1,24}[，,]\s*(?:而)?是"),
    ("万能动宾", "「，带着……」万能状语", r"[，,]\s*带着[^，。；！？\n]{0,20}"),
    ("声音描写", "「声音不大，却带着……」", r"声音(?:不大|很轻|很平|平静)[^。\n]{0,14}(?:却|但)?带着"),
    ("认知直述", "告诉而非展示，改为动作或台词", r"(?:他|她|他们|她们)(?:这才|终于)?(?:明白|意识到|感到|知道)"),
    ("章末预告", "「他不知道的是……」空泛预告", r"(?:他|她)不知道的是"),
    ("收束腔", "「这一刻／才刚刚开始」式总结", r"这一刻[，,]|从这一刻开始|才刚刚开始|原来[，,]"),
    ("抽象命运", "命运＋齿轮／棋局／獠牙", r"命运[^。\n]{0,8}(?:齿轮|棋局|獠牙|改写|安排)"),
    ("心理模板", "「心中一震／心头一凛」式心理惊吓", r"心(?:中|头|底)(?:一|猛|顿)(?:震|凛|颤)"),
    ("过渡模板", "「取而代之的是」", r"取而代之"),
    ("告诉而非展示", "「显得有些X」「散发着一股X气息」", r"显得(?:有些|十分|非常)?[\u4e00-\u9fff]{1,4}|散发(?:着|出)?[^。\n]{0,8}(?:气息|气场)"),
    ("公式化对话标签", "「好的，他说道」", r"[，,](?:他|她)(?:说道|开口道|沉声道|低声道)"),
    ("英文引号", "模型常直接吐 ASCII 直角引号，中文排版用“”", r'"'),
    ("破折号", "正文禁用，改用句号、逗号或动作断句", r"——|──|(?<!-)--(?!-)"),
    ("省略号", "正文禁用，用动作或短句表达停顿", r"……|\.\.\."),
)

SAMPLE_WIDTH = 14
SAMPLES_PER_RULE = 2


@dataclass
class Hit:
    """一条诊断：命中哪个规则、几次、原文长什么样。"""

    rule: str
    category: str
    level: str          # hard = 硬伤词+句式+标点；soft = 只看密度的词
    count: int
    samples: list[str] = field(default_factory=list)

    def line(self) -> str:
        ex = "；".join(self.samples)
        return f"{self.category:<6} {self.rule:<8} ×{self.count:<3} {ex}"


def _context(text: str, start: int, end: int) -> str:
    left = max(0, start - SAMPLE_WIDTH)
    right = min(len(text), end + SAMPLE_WIDTH)
    return text[left:right].replace("\n", " ").strip()


def _scan_word(text: str, word: str) -> tuple[int, list[str]]:
    count, samples, pos = 0, [], text.find(word)
    while pos != -1:
        count += 1
        if len(samples) < SAMPLES_PER_RULE:
            samples.append(_context(text, pos, pos + len(word)))
        pos = text.find(word, pos + len(word))
    return count, samples


def check(text: str) -> list[Hit]:
    """扫一遍文本，返回全部诊断（硬伤在前，同类按命中数降序）。"""
    hits: list[Hit] = []

    for category, words in HARD_WORDS.items():
        for word in words:
            count, samples = _scan_word(text, word)
            if count:
                hits.append(Hit(word, category, "hard", count, samples))

    for rule, _note, pattern in PATTERNS:
        rx = re.compile(pattern)
        found = list(rx.finditer(text))
        if found:
            samples = [_context(text, m.start(), m.end()) for m in found[:SAMPLES_PER_RULE]]
            hits.append(Hit(rule, "句式" if not rule.startswith(("破折号", "省略号", "英文引号")) else "标点",
                            "hard", len(found), samples))

    for category, words in SOFT_WORDS.items():
        for word in words:
            count, samples = _scan_word(text, word)
            if count:
                hits.append(Hit(word, category, "soft", count, samples))

    hits.sort(key=lambda h: (h.level != "hard", -h.count, h.rule))
    return hits


def hard_hits(hits: list[Hit]) -> list[Hit]:
    return [h for h in hits if h.level == "hard"]


def per_k(text: str) -> float:
    """每千字硬伤命中数。空文本返回 0，避免除零。"""
    length = cjk_length(text) or 1
    return sum(h.count for h in hard_hits(check(text))) * 1000 / length


def report(text: str, hits: list[Hit] | None = None, top: int = 14) -> str:
    hits = check(text) if hits is None else hits
    hard = hard_hits(hits)
    soft = [h for h in hits if h.level == "soft"]
    total_hard = sum(h.count for h in hard)
    total_soft = sum(h.count for h in soft)
    length = cjk_length(text)
    density = total_hard * 1000 / (length or 1)

    lines = [
        f"AI 味诊断：{length} 字｜硬伤 {total_hard} 处（{density:.1f}/千字）｜密度词 {total_soft} 处",
    ]
    if hard:
        lines.append("")
        lines.append("硬伤（命中即改）：")
        lines += ["  " + h.line() for h in hard[:top]]
    if soft:
        lines.append("")
        lines.append("密度词（成串才处理）：")
        lines += ["  " + h.line() for h in soft[:top]]
    if not hard and not soft:
        lines.append("  没有命中：用词干净。")
    return "\n".join(lines)


def digest(text: str, top: int = 10) -> str:
    """给扩写模型看的简短诊断：要求它逐条消灭这些。"""
    hits = [h for h in check(text) if h.level == "hard"][:top]
    if not hits:
        return "（机械检查没有发现 AI 味硬伤；重点放在把概述换成场景。）"
    return "\n".join(f"- {h.category}『{h.rule}』×{h.count}｜例：{h.samples[0] if h.samples else ''}"
                     for h in hits)