| """题型推断：由模型自己判断题目落在哪个 6×3 格子（不读任何 benchmark 标签）。 |
| |
| 用户定稿 2026-08-21：题型识别本身是模型的活，不能外部告知。 |
| 输出全部落盘（cell_log.jsonl），将来可用于训练一个小分类器替换 LLM 调用。 |
| 消融档见 ablation.py：llm（默认）/ lexical（纯关键词规则，免费）/ none（不分类）。 |
| """ |
|
|
| import json |
| import logging |
| import os |
| import re |
|
|
| from .qtype import SPACE, TIME, spec_for_prompt, is_valid |
|
|
| logger = logging.getLogger(__name__) |
|
|
| PROMPT = """给你一道视频问答题，判断它属于哪一类。 |
| |
| {spec} |
| |
| 题目：{question} |
| 选项：{options} |
| |
| 只输出 JSON：{{"space": "<空间维度>", "time": "<时序维度>", "confidence": 0.0-1.0}}""" |
|
|
| |
| _LEX_SPACE = [ |
| ("COUNT", r"\bhow many\b|\bhow much\b|\bnumber of\b|\bcount\b|几个|多少"), |
| ("TEXT", r"\bwritten\b|\btext\b|\blabel\b|\bsign\b|\bbrand\b|\bword[s]?\b|\bread[s]?\b|文字|写着"), |
| ("ORDER", r"\border\b|\bsequence\b|\bfirst\b|\bthen\b|\bbefore\b|\bafter\b|\bnext\b|顺序|先|然后"), |
| ("SPATIAL", r"\bwhere\b|\bleft\b|\bright\b|\babove\b|\bbelow\b|\bnext to\b|\bposition\b|位置|在哪"), |
| ("STATE", r"\bcolor\b|\bcolour\b|\bopen\b|\bclosed\b|\bstate\b|\bstatus\b|\bhow many times\b|状态|颜色"), |
| ("ACTION", r"\bdoing\b|\baction\b|\bactivity\b|\bperform\b|\btask\b|\bpurpose\b|\bgoal\b|做什么|动作"), |
| ] |
| _LEX_TIME = [ |
| ("GLOBAL", r"\bthroughout\b|\boverall\b|\bprimary\b|\bentire\b|\bwhole video\b|\bin general\b|\bsummar\w*\b|\boverarching\b|\btaking into account\b|全局|整段|总体"), |
| |
| |
| |
| ("PAST", r"\bdid\b|\bwas\b|\bwere\b|\bearlier\b|\bbefore\b|\bpreviously\b|\blast time\b|\bhad\b|之前|过去|上一次"), |
| ("NOW", r"\bnow\b|\bcurrent\w*\b|\bthis moment\b|\bam i\b|\bis (?:he|she|this|the) \w+ing\b|当前|正在"), |
| ] |
|
|
|
|
| def classify_lexical(question: str, options=None) -> dict: |
| """免费的词法分类（消融档 + LLM 失败兜底）。""" |
| q = (question or "").lower() |
| space = next((s for s, pat in _LEX_SPACE if re.search(pat, q)), "ACTION") |
| time = next((t for t, pat in _LEX_TIME if re.search(pat, q)), "PAST") |
| return {"space": space, "time": time, "confidence": 0.3, "method": "lexical"} |
|
|
|
|
| def classify_llm(question: str, options, llm) -> dict: |
| """LLM 判定；解析失败或越界则回落词法。""" |
| opts = " | ".join(f"{chr(65+i)}. {o}" for i, o in enumerate(options or [])) |
| try: |
| r = llm.chat_json(PROMPT.format(spec=spec_for_prompt(), question=question, |
| options=opts or "(无选项)"), max_tokens=200) or {} |
| except Exception as e: |
| logger.warning("题型分类调用失败: %s，回落词法", e) |
| r = {} |
| s, t = str(r.get("space", "")).upper(), str(r.get("time", "")).upper() |
| if not is_valid(s, t): |
| fb = classify_lexical(question, options) |
| fb["llm_raw"] = {k: r.get(k) for k in ("space", "time")} |
| return fb |
| return {"space": s, "time": t, |
| "confidence": float(r.get("confidence", 0.5) or 0.5), "method": "llm"} |
|
|
|
|
| class CellClassifier: |
| """带落盘的题型推断器。mode: llm | lexical | none""" |
|
|
| def __init__(self, llm=None, mode: str = "llm", log_path: str = ""): |
| self.llm, self.mode, self.log_path = llm, mode, log_path |
| if log_path: |
| os.makedirs(os.path.dirname(log_path), exist_ok=True) |
|
|
| def classify(self, question: str, options=None, meta: dict = None) -> dict: |
| if self.mode == "none": |
| out = {"space": None, "time": None, "confidence": 0.0, "method": "none"} |
| elif self.mode == "lexical" or self.llm is None: |
| out = classify_lexical(question, options) |
| else: |
| out = classify_llm(question, options, self.llm) |
| if self.log_path: |
| with open(self.log_path, "a") as f: |
| f.write(json.dumps({**(meta or {}), "question": question, |
| **out}, ensure_ascii=False) + "\n") |
| return out |
|
|
