""" 第八章「评论文本补充分析」独立脚本(**不修改**主报告核心逻辑)。 流程(按细类分组):清洗(中文分词 + 停用词)→ **词云图(可选)** → 词频 / 关键词突出度 → 词对共现 → 主题归纳 → 规则化叙事小结 → 文末可选 **专用 LLM**(结构化 JSON + ``PROBE_TEXT_MINING_SYSTEM`` + ``_call_llm``;**非** ``COMMENT_GROUPS_SYSTEM``、**非**已废弃的星级子集预设口语短语情感条形图)。 依赖(请自行安装):: pip install jieba scikit-learn numpy wordcloud matplotlib 用法(在 ``backend`` 目录下):: python -m pipeline.demos.chapter8_text_mining_probe --run-dir \"../data/JD/pipeline_runs/20260413_104252_低GI\" python -m pipeline.demos.chapter8_text_mining_probe --run-dir \"...\" --out chapter8_probe.md python -m pipeline.demos.chapter8_text_mining_probe --run-dir \"...\" --live-llm python -m pipeline.demos.chapter8_text_mining_probe --run-dir \"...\" --live-llm --llm-chunked 输出:默认写入 ``/chapter8_text_mining_probe.md``。 环境变量 ``MA_PROBE_LLM_DIAG=1``:向 stderr 打印每个 LLM 分块的 JSON 大小与耗时(定位「哪一类超时」)。 嵌入竞品报告:流水线默认开启(``get_default_report_config`` 中 ``chapter8_text_mining_probe``: true);若任务显式关闭则为 false。开启时会生成本稿并调用 ``markdown_embed_body_for_competitor_report`` 写入 ``competitor_analysis.md`` 的 **第八章第二节(评论文本补充分析)**,替代原「关注词 + 场景」条图及对应两段大模型;**不再**嵌入原「评价正负面粗判」预设口语短语扇形图/条形图及同口径大模型块。 """ from __future__ import annotations import argparse import json import os import re import sys from collections import Counter from pathlib import Path from typing import Any BACKEND_ROOT = Path(__file__).resolve().parents[2] if str(BACKEND_ROOT) not in sys.path: sys.path.insert(0, str(BACKEND_ROOT)) os.environ.setdefault("DJANGO_SETTINGS_MODULE", "config.settings") import django # noqa: E402 django.setup() JCR_ROOT = BACKEND_ROOT / "crawler_copy" / "jd_pc_search" if str(JCR_ROOT) not in sys.path: sys.path.insert(0, str(JCR_ROOT)) from pipeline.competitor_report import jd_report as jcr # noqa: E402 import jd_keyword_pipeline as kpl # noqa: E402 from pipeline.csv.schema import MERGED_FIELD_TO_CSV_HEADER # noqa: E402 from pipeline.llm.generate import _call_llm # noqa: E402 探针专用,不新增 generate 导出 # 导入失败时**不得** ``sys.exit``:本模块会被 ``runner`` 在 Web 请求中 import,退出会整进程 500。 _PROBE_TEXT_MINING_DEPS_OK = False _PROBE_TEXT_MINING_IMPORT_ERROR = "" try: import numpy as np # noqa: WPS433 import jieba # noqa: WPS433 from sklearn.decomposition import LatentDirichletAllocation # noqa: WPS433 from sklearn.feature_extraction.text import ( # noqa: WPS433 CountVectorizer, TfidfVectorizer, ) _PROBE_TEXT_MINING_DEPS_OK = True except ImportError as e: np = None # type: ignore[assignment] jieba = None # type: ignore[assignment] LatentDirichletAllocation = None # type: ignore[assignment] CountVectorizer = None # type: ignore[assignment] TfidfVectorizer = None # type: ignore[assignment] _PROBE_TEXT_MINING_IMPORT_ERROR = str(e) try: import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt # noqa: WPS433 from wordcloud import WordCloud # noqa: WPS433 _WORDCLOUD_AVAILABLE = True except ImportError: plt = None # type: ignore[assignment] WordCloud = None # type: ignore[assignment] _WORDCLOUD_AVAILABLE = False # 精简中文停用词(可换外部文件);与业务无关,仅用于探针 _STOP_BASIC: frozenset[str] = frozenset( """ 的 了 和 是 在 也 有 就 都 很 啊 还 吗 吧 呢 呀 哦 噢 哈 呵 与 及 或 等 为 被 让 从 到 把 而 又 对 中 这 那 其 一个 一些 没有 不是 可以 这样 我们 你们 他们 它们 它 会 要 能 去 来 做 用 给 自己 这个 那个 什么 怎么 如果 因为 所以 但是 而且 然后 还是 或者 还有 就是 只是 只是 已经 还是 非常 真的 比较 特别 感觉 觉得 认为 看到 收到 东西 商品 产品 卖家 买家 店铺 京东 物流 快递 包装 评价 评论 购买 买 卖 收到 天 次 个 款 种 条 块 """.split() ) # --- 评论文本补充分析 · 专用 LLM(与正式报告 ``COMMENT_GROUPS_SYSTEM`` / 已废弃的预设口语短语情感口径均不同)--- PROBE_TEXT_MINING_SYSTEM = """你是用户研究与文本挖掘方向的助手。 输入 JSON 为「第八章第二节评论文本补充分析」的**专用**结果(``schema_version``=1),**不是**正式竞品报告里的关注词规则统计、也不是已废弃的星级子集预设口语短语 Lexicon。其中的数字与词表来自**中文分词 + 统计工具**(词频、关键词突出度、共现、主题归纳),与业务侧子串计数**口径不同**。 每个 ``groups`` 元素含: - ``probe_status``:``ok`` 表示该细类已完成分词与统计;``skipped`` 表示样本过少等未下钻。 - ``word_freq_top`` / ``tfidf_top`` / ``cooccurrence_top``:统计型特征(开放词表)。 - ``lda``:无监督自动归纳的主题词;**仅为探索**,同一词可出现在多主题,**禁止**当作严格品类或固定标签。 - ``sample_text_snippets``:与正式报告管线同源的评价短摘录(若有),仅用于**对照语境**;若与关键词突出度焦点冲突,以**整句原文**为准,并在段末或「使用注意」中可点明「统计与语义可能不一致」。 **任务**:对 ``groups`` 中**每一项**输出对应 Markdown(**顺序与输入一致**): - 对 ``probe_status == "ok"``:以 ``#### `` + 与该条 ``group`` 字段**完全一致**的细类名作为小节标题(勿用 ``##`` 一级标题);每段约 **100~260 字**。 - 内容须包含:①用 **1~2 句**概括该细类评论**主要讨论焦点**(综合词频与关键词突出度,**不要罗列具体数字**);② **1~2 句**说明共现词对**暗示**哪些维度常一起出现(**非因果**);③若 ``lda.topics`` 非空,**1~2 句**说明主题粗分侧重点,并**明确**算法无监督、**不**与矩阵细类一一对应;④用 **1~2 句**体现 ``sample_text_snippets`` 中的**用户语气与关切**(以**转述**为主);若必须引用原文,**全小节合计**仅 **一处**极短引号内容(**≤40 字**,**不要**输出 ``【细类…SKU…店铺…】`` 等长前缀);若无可用摘录则写明;⑤ **使用场景(仅从评论推断)**:用 **0~2 句**概括**何时、何地、何人、如何搭配**等(如早餐、加餐、控糖人群、配牛奶等)——**只能**依据本细类 ``word_freq``/``tfidf``/``cooccurrence``/``lda`` 与摘录中**已出现或可合理概括**的信息;**禁止**套用正式报告「场景分组」或其它外部场景分类;若统计与摘录中**均无**场景线索,**一句**写明「评论中未体现清晰使用场景」即可。 - 对 ``probe_status == "skipped"``:该小节仅 **一句**说明原因。 **禁止**:编造数据中未出现的品牌、价格、医学功效或疗效承诺;不要把 ``keyword`` 监测词写进「用户原话」;不要输出 Markdown 表格;不要声称本段与「正式报告第八章末」完全同源——本任务为**补充分析解读**。**禁止**把输入里的 ``sample_text_snippets`` **逐条罗列**、**多条整段复制**到输出(那不是归纳,是重复贴评论)。 全文末可另起一段 **「使用注意」**(简短):点明开放词表统计与人工阅读差异、主题归纳局限、小样本细类不可靠。 总字数约 **800~4500 字**(细类多则偏长)。仅输出正文 Markdown,不要用代码围栏包裹全文。""" PROBE_TEXT_MINING_USER_PREFIX = ( "请根据以下 JSON 撰写「评论文本补充分析」解读正文(Markdown)。\n\n" ) def _is_noise_token(w: str) -> bool: if len(w) < 2: return True if w in ("ldquo", "rdquo", "nbsp", "mdash"): return True if re.match(r"^[a-z]{1,8}$", w) and w not in ("gi",): return True return False def _word_freq_from_cut_docs(cut_docs: list[str]) -> dict[str, int]: c: Counter[str] = Counter() for line in cut_docs: for w in line.split(): if _is_noise_token(w): continue c[w] += 1 return dict(c) def _font_path_chinese() -> str | None: windir = os.environ.get("WINDIR", r"C:\Windows") candidates = [ Path(windir) / "Fonts" / "msyh.ttc", Path(windir) / "Fonts" / "msyhbd.ttc", Path(windir) / "Fonts" / "simhei.ttf", Path(windir) / "Fonts" / "simsun.ttc", Path("/usr/share/fonts/truetype/wqy/wqy-microhei.ttc"), Path("/usr/share/fonts/truetype/noto/NotoSansCJK-Regular.ttc"), ] for p in candidates: try: if p.is_file(): return str(p) except OSError: continue return None def _slug_safe(gname: str) -> str: s = re.sub(r'[<>:"/\\|?*]', "_", (gname or "").strip()) return s[:80] if len(s) > 80 else s def _save_wordcloud_png( freq: dict[str, int], out_path: Path, *, font_path: str | None, ) -> str: """写入 PNG;成功返回空串,失败返回错误说明。""" if not _WORDCLOUD_AVAILABLE or WordCloud is None or plt is None: return "wordcloud/matplotlib 未安装" if not freq: return "词频为空" try: out_path.parent.mkdir(parents=True, exist_ok=True) wc = WordCloud( font_path=font_path, width=960, height=540, background_color="white", max_words=220, relative_scaling=0.35, colormap="viridis", prefer_horizontal=0.88, min_font_size=10, ).generate_from_frequencies(freq) fig, ax = plt.subplots(figsize=(10.5, 6), dpi=120) ax.imshow(wc, interpolation="bilinear") ax.axis("off") fig.tight_layout(pad=0) fig.savefig(out_path, bbox_inches="tight", facecolor="white") plt.close(fig) except Exception as e: return str(e) return "" def _load_run( run_dir: Path, ) -> tuple[str, list[dict[str, str]], list[dict[str, str]]]: run_dir = run_dir.resolve() merged_path = run_dir / kpl.FILE_MERGED_CSV if not merged_path.is_file(): raise FileNotFoundError(f"缺少合并表: {merged_path}") _, merged_rows = jcr._read_csv_rows(merged_path) _, comment_rows = jcr._read_csv_rows(run_dir / kpl.FILE_COMMENTS_FLAT_CSV) meta_path = run_dir / kpl.FILE_RUN_META_JSON meta: dict[str, Any] | None = None if meta_path.is_file(): try: meta = json.loads(meta_path.read_text(encoding="utf-8")) except json.JSONDecodeError: meta = None kw = "" if meta and str(meta.get("keyword") or "").strip(): kw = str(meta.get("keyword")).strip() return kw, merged_rows, comment_rows def _cut_one(text: str) -> list[str]: s = (text or "").strip() if not s: return [] raw = jieba.lcut(s) out: list[str] = [] for w in raw: w = w.strip() if len(w) < 2: continue if w in _STOP_BASIC: continue if re.match(r"^[0-9\s\W_]+$", w): continue out.append(w) return out def _docs_cut(texts: list[str]) -> list[str]: """每条评论 → 空格连接的分词串,供 sklearn。""" rows: list[str] = [] for t in texts: toks = _cut_one(t) if toks: rows.append(" ".join(toks)) return rows def _term_freq_top(cut_docs: list[str], top_k: int) -> list[tuple[str, int]]: c: Counter[str] = Counter() for line in cut_docs: for w in line.split(): c[w] += 1 return c.most_common(top_k) def _tfidf_top(cut_docs: list[str], top_k: int) -> list[tuple[str, float]]: if not cut_docs: return [] vec = TfidfVectorizer(max_features=min(2000, max(50, len(cut_docs) * 3))) try: X = vec.fit_transform(cut_docs) except ValueError: return [] feats = np.array(vec.get_feature_names_out()) scores = np.asarray(X.mean(axis=0)).ravel() idx = np.argsort(-scores)[:top_k] return [(str(feats[i]), float(scores[i])) for i in idx] def _cooc_top_pairs( cut_docs: list[str], vocab_cap: int, pair_top: int, ) -> list[tuple[str, str, int]]: """同一条评论内无序词对共现(过滤低频词)。""" term_freq = Counter() for line in cut_docs: for w in set(line.split()): term_freq[w] += 1 top_terms = {w for w, _ in term_freq.most_common(vocab_cap)} pair_c: Counter[tuple[str, str]] = Counter() for line in cut_docs: toks = sorted({w for w in line.split() if w in top_terms}) for i in range(len(toks)): for j in range(i + 1, len(toks)): a, b = toks[i], toks[j] if a > b: a, b = b, a pair_c[(a, b)] += 1 return [(a, b, n) for (a, b), n in pair_c.most_common(pair_top)] def _lda_topics( cut_docs: list[str], n_topics: int, n_top_words: int, ) -> tuple[list[list[str]], str]: if len(cut_docs) < 4: return [], "文本条数过少,跳过 LDA。" n_topics = max(2, min(n_topics, len(cut_docs) // 2)) try: vec = CountVectorizer(max_df=0.95, min_df=2, max_features=800) X = vec.fit_transform(cut_docs) except ValueError as e: return [], f"LDA 向量化失败:{e}" if X.shape[0] < 3 or X.shape[1] < 3: return [], "矩阵过稀疏,跳过 LDA。" lda = LatentDirichletAllocation( n_components=n_topics, max_iter=30, learning_method="batch", random_state=42, n_jobs=1, ) try: lda.fit(X) except Exception as e: return [], f"LDA 拟合失败:{e}" names = vec.get_feature_names_out() topics: list[list[str]] = [] for topic_idx, topic in enumerate(lda.components_): top_ix = np.argsort(-topic)[:n_top_words] topics.append([str(names[i]) for i in top_ix]) return topics, "" def _md_escape(s: str) -> str: return (s or "").replace("|", "\\|").replace("\n", " ") def _narrative_stub( n_raw: int, n_cut: int, tf_top: list[tuple[str, int]], tfidf_top: list[tuple[str, float]], cooc: list[tuple[str, str, int]], lda_topics: list[list[str]], lda_note: str, ) -> str: lines: list[str] = [ f"本细类有效评论约 **{n_cut}** 条(原始非空 **{n_raw}** 条,经分词去停用后用于建模)。", "", "**词频 Top**:" + ( "、".join(f"「{w}」({n})" for w, n in tf_top[:12]) if tf_top else "(无)" ) + "。", "", "**关键词突出度 Top**(相对区分度):" + ( "、".join(f"「{w}」({s:.3f})" for w, s in tfidf_top[:12]) if tfidf_top else "(无)" ) + "。", "", "**共现较强的词对**(同条评论内,供联想维度):" + ( ";".join(f"「{a}」-「{b}」({c})" for a, b, c in cooc[:10]) if cooc else "(无)" ) + "。", "", ] if lda_note: lines.append(f"*主题归纳:{lda_note}*") lines.append("") elif lda_topics: lines.append("**自动归纳的主题(无监督,仅作探索)**:") for i, words in enumerate(lda_topics): lines.append(f"- 主题 {i + 1}:{'、'.join(words)}") lines.append("") lines.append( "> 以上为主题探索与统计摘要,**不等同**于业务结论;若与星级、规则词表冲突,以人工抽样为准。" ) return "\n".join(lines) def _truncate_probe_payload(payload: dict[str, Any]) -> dict[str, Any]: """压缩摘录长度,避免单次 JSON 顶满上下文。""" out: dict[str, Any] = dict(payload) groups: list[Any] = [] for g in (out.get("groups") or []): if not isinstance(g, dict): groups.append(g) continue g2 = dict(g) g2.pop("focus_hit_lines", None) sn = g2.get("sample_text_snippets") if isinstance(sn, list): # 条数略减,降低模型「照抄罗列」倾向;仍以转述为主见系统提示 g2["sample_text_snippets"] = [str(x)[:200] for x in sn[:6]] groups.append(g2) out["groups"] = groups return out def _merge_snippets_from_comment_groups( probe_rows: list[dict[str, Any]], *, merged_rows: list[dict[str, str]], comment_rows: list[dict[str, str]], ) -> None: """把正式 ``build_comment_groups_llm_payload`` 中的摘录并入补充分析行(原地修改)。""" sku_h = MERGED_FIELD_TO_CSV_HEADER["sku_id"] title_h = MERGED_FIELD_TO_CSV_HEADER["title"] fb = jcr._consumer_feedback_by_matrix_group( merged_rows=merged_rows, comment_rows=comment_rows, sku_header=sku_h, ) pl = jcr.build_comment_groups_llm_payload( feedback_groups=fb, merged_rows=merged_rows, sku_header=sku_h, title_h=title_h, ) by_g = {str(x.get("group")): x for x in pl if isinstance(x, dict)} for row in probe_rows: if row.get("probe_status") != "ok": continue gname = str(row.get("group") or "") src = by_g.get(gname) if not src: continue row["comment_flat_rows"] = src.get("comment_flat_rows") sn = src.get("sample_text_snippets") if isinstance(sn, list): row["sample_text_snippets"] = [str(x)[:220] for x in sn[:8]] def _probe_llm_diag_enabled() -> bool: return os.environ.get("MA_PROBE_LLM_DIAG", "").strip().lower() in ( "1", "true", "yes", ) def _run_probe_text_mining_llm( payload: dict[str, Any], *, chunked: bool, ) -> str: """补充分析专用:``PROBE_TEXT_MINING_SYSTEM`` + 结构化 JSON;可选按细类拆分调用。 - **分块模式**:每个 ``probe_status == ok`` 的细类**单独**请求;某一类失败时**保留**已成功类的正文,该类下追加失败说明(不再整段被外层 ``except`` 吃掉)。 - 设置环境变量 ``MA_PROBE_LLM_DIAG=1`` 时向 **stderr** 打印每类 JSON 字符数与耗时,便于定位超时发生在哪一类。 """ if not payload.get("groups"): return "> **补充分析 LLM 解读**:无分组数据,跳过。" import time as _time diag = _probe_llm_diag_enabled() try: if not chunked: p = _truncate_probe_payload(payload) raw = json.dumps(p, ensure_ascii=False) if len(raw) > 88_000: raw = ( raw[:82_000] + "\n\n…(JSON 过长已截断,仅依据可见字段撰写。)\n" ) if diag: print( f"[probe-llm] mode=single json_chars={len(raw)}", file=sys.stderr, flush=True, ) t0 = _time.perf_counter() try: out = _call_llm( PROBE_TEXT_MINING_SYSTEM, PROBE_TEXT_MINING_USER_PREFIX + raw, ).strip() except Exception as e: if diag: print( f"[probe-llm] mode=single FAIL after " f"{_time.perf_counter() - t0:.1f}s: {e}", file=sys.stderr, flush=True, ) return f"> **补充分析 LLM 解读**调用失败:{e}" if diag: print( f"[probe-llm] mode=single OK " f"{_time.perf_counter() - t0:.1f}s out_chars={len(out)}", file=sys.stderr, flush=True, ) return out kw = str(payload.get("keyword") or "") note = str(payload.get("probe_note") or "") parts: list[str] = [] for g in payload.get("groups") or []: if not isinstance(g, dict): continue gname = str(g.get("group") or "?") if g.get("probe_status") != "ok": parts.append( f"#### {gname}\n\n" f"*(评论量不足,未做词云与主题归纳:{g.get('reason', '')})*" ) continue mini = { "schema_version": payload.get("schema_version", 1), "keyword": kw, "probe_note": note, "groups": [g], } raw = json.dumps(mini, ensure_ascii=False) if len(raw) > 48_000: raw = raw[:44_000] + "\n…\n" if diag: print( f"[probe-llm] mode=chunked group={gname!r} json_chars={len(raw)}", file=sys.stderr, flush=True, ) t0 = _time.perf_counter() try: chunk_out = _call_llm( PROBE_TEXT_MINING_SYSTEM, PROBE_TEXT_MINING_USER_PREFIX + raw, ).strip() parts.append(chunk_out) if diag: print( f"[probe-llm] group={gname!r} OK " f"{_time.perf_counter() - t0:.1f}s out_chars={len(chunk_out)}", file=sys.stderr, flush=True, ) except Exception as e: if diag: print( f"[probe-llm] group={gname!r} FAIL " f"{_time.perf_counter() - t0:.1f}s: {e}", file=sys.stderr, flush=True, ) parts.append( f"#### {gname}\n\n" f"> **本细类补充分析 LLM 调用失败**:{e}" ) return "\n\n---\n\n".join(parts) except Exception as e: return f"> **补充分析 LLM 解读**调用失败:{e}" def _ensure_probe_dependencies() -> None: if not _PROBE_TEXT_MINING_DEPS_OK: raise ImportError( "第八章评论文本补充分析依赖未安装,请在 backend 环境下执行:" "pip install jieba scikit-learn numpy wordcloud\n" f"原始错误: {_PROBE_TEXT_MINING_IMPORT_ERROR}" ) def build_markdown( run_dir: Path, *, min_texts: int, lda_topics_n: int, top_k_words: int, cooc_vocab: int, cooc_pairs: int, live_llm: bool, llm_chunked: bool, wordcloud_enabled: bool, wordcloud_max: int, ) -> str: _ensure_probe_dependencies() kw, merged, comments = _load_run(run_dir) sku_h = MERGED_FIELD_TO_CSV_HEADER["sku_id"] groups = jcr._consumer_feedback_by_matrix_group( merged_rows=merged, comment_rows=comments, sku_header=sku_h, ) lines: list[str] = [ "# 八、消费者反馈与用户画像(评论文本补充分析 · 实验稿)", "", f"- **运行目录**:`{run_dir}`", f"- **监测词(run_meta)**:{kw or '—'}", f"- **生成脚本**:`pipeline.demos.chapter8_text_mining_probe`", "", "## 8.0 说明", "", "本稿为**独立补充分析**,流程为:清洗评论 → 词云(可选)→ 词频与关键词 → 词对共现 → 主题归纳 → 文字小结;" "文末可选用**专用提示词**由大模型解读(与已废弃的预设口语短语情感口径不同)。" "**细类划分与 SKU 归因**与主报告一致;其余为中文分词与统计工具做的开放词表分析,**不替代**正式报告中的规则统计。", "", "---", "", ] if wordcloud_enabled and not _WORDCLOUD_AVAILABLE: lines.extend( [ "> **词云**:当前环境未安装 ``wordcloud`` / ``matplotlib``,已跳过出图。" "请执行:``pip install wordcloud matplotlib``。", "", ] ) elif wordcloud_enabled and not _font_path_chinese(): lines.extend( [ "> **词云字体**:未检测到常见中文字体路径,词云可能出现方框;" "Windows 可确认 ``C:\\Windows\\Fonts\\msyh.ttc`` 是否存在。", "", ] ) wc_n = 0 probe_rows: list[dict[str, Any]] = [] for gname, _cr_rows, texts in groups: n_raw = len([t for t in texts if (t or "").strip()]) if n_raw < min_texts: probe_rows.append( { "group": gname, "probe_status": "skipped", "reason": f"有效文本 {n_raw} 条,低于 min_texts={min_texts}", } ) lines.extend( [ f"## {gname}", "", f"*本细类有效评论仅 {n_raw} 条,低于分析所需最少条数({min_texts} 条),故未生成词云与主题图。*", "", "---", "", ] ) continue cut_docs = _docs_cut(texts) if len(cut_docs) < 2: probe_rows.append( { "group": gname, "probe_status": "skipped", "reason": "分词后不足 2 条", } ) lines.extend( [ f"## {gname}", "", "*分词后不足 2 条,跳过。*", "", "---", "", ] ) continue tf_top = _term_freq_top(cut_docs, top_k_words) tfidf_top = _tfidf_top(cut_docs, top_k_words) cooc = _cooc_top_pairs(cut_docs, cooc_vocab, cooc_pairs) lda_t, lda_err = _lda_topics(cut_docs, lda_topics_n, 12) if (lda_err or "").strip(): lda_obj: dict[str, Any] = { "status": "skipped", "reason": lda_err.strip(), } else: lda_obj = {"status": "ok", "topics": lda_t} probe_rows.append( { "group": gname, "probe_status": "ok", "comment_text_units": n_raw, "jieba_cut_document_count": len(cut_docs), "word_freq_top": [ {"term": w, "count": int(n)} for w, n in tf_top[:20] ], "tfidf_top": [ {"term": w, "score": round(float(s), 4)} for w, s in tfidf_top[:20] ], "cooccurrence_top": [ {"term_a": a, "term_b": b, "joint_count": int(c)} for a, b, c in cooc[:15] ], "lda": lda_obj, } ) lines.extend([f"## {gname}", ""]) if ( wordcloud_enabled and _WORDCLOUD_AVAILABLE and wc_n < wordcloud_max ): freq_wc = _word_freq_from_cut_docs(cut_docs) fn = f"wordcloud_probe__{wc_n:02d}_{_slug_safe(gname)}.png" img_path = run_dir.resolve() / "report_assets" / fn err_wc = _save_wordcloud_png( freq_wc, img_path, font_path=_font_path_chinese(), ) if not err_wc: lines.append( f"![词云(按词频权重)](report_assets/{fn})" ) lines.append("") wc_n += 1 else: lines.append(f"> 词云未生成:{err_wc}") lines.append("") lines.append(_narrative_stub( n_raw, len(cut_docs), tf_top, tfidf_top, cooc, lda_t, lda_err, )) lines.extend(["", "---", ""]) _merge_snippets_from_comment_groups( probe_rows, merged_rows=merged, comment_rows=comments, ) llm_payload: dict[str, Any] = { "schema_version": 1, "keyword": kw, "probe_note": ( "中文分词 + 停用词;关键词突出度/共现/主题归纳为统计库;" "与正式报告第八章第二节图表中的关注词计数等规则统计、已废弃的预设口语短语情感口径均不同;" "评价短摘录合并自 build_comment_groups_llm_payload(与正式管线同源,不含关注词子串计数摘要),供语境对照;" "「使用场景」若出现,须仅能从本 JSON 内统计与摘录推断,不接入正式场景分组。" ), "groups": probe_rows, } lines.extend( [ "", "---", "", "## 评论文本归纳(大模型 · 专用口径)", "", "> 输入为按细类整理后的词频、关键词、共现与主题等统计结果,以及少量原文摘录(供理解语境,**不是**要求逐条复述);" "归纳中的**使用场景**仅能从上述统计推断,**不等同**于正式的「场景分组」章节。", "", ] ) if live_llm: lines.append(_run_probe_text_mining_llm(llm_payload, chunked=llm_chunked)) else: lines.append( "> **补充分析 LLM 解读**:未启用。请使用 ``--live-llm``(需 ``AI_crawler`` 等可用);" "细类很多、单次 JSON 易超长时可加 ``--llm-chunked``(按细类多次调用后拼接)。" ) lines.append("") lines.append("*(完)*") return "\n".join(lines) def markdown_embed_body_for_competitor_report(full_probe_md: str) -> str: """ 将独立补充分析稿转为可嵌入 ``build_competitor_markdown`` 的 **第八章第二节正文**(不含 ``### 8.2`` 标题行;由 ``jd_report`` 统一加标题): 从 ``## 8.0 说明`` 起至文末,并把 ``## …`` 降为 ``#### …``,避免与宿主 ``## 八、`` 冲突。 """ lines = (full_probe_md or "").splitlines() try: start = next( i for i, ln in enumerate(lines) if ln.strip() == "## 8.0 说明" or ln.strip().startswith("## 8.0 说明") ) except StopIteration: return (full_probe_md or "").strip() chunk = lines[start:] out: list[str] = [] for ln in chunk: if ln.startswith("## ") and not ln.startswith("###"): out.append("#### " + ln[3:]) else: out.append(ln) while out and out[-1].strip() in ("*(完)*", ""): out.pop() while out and not out[-1].strip(): out.pop() return "\n".join(out).strip() def main() -> None: if not _PROBE_TEXT_MINING_DEPS_OK: print( "缺少依赖,请先安装:pip install jieba scikit-learn numpy wordcloud\n" f"原始错误: {_PROBE_TEXT_MINING_IMPORT_ERROR}", file=sys.stderr, ) sys.exit(1) ap = argparse.ArgumentParser(description="第八章评论文本补充分析(独立脚本)") ap.add_argument( "--run-dir", type=Path, required=True, help="pipeline_runs 下某批次目录(含 keyword_pipeline_merged.csv、comments_flat.csv)", ) ap.add_argument( "--out", type=Path, default=None, help="输出 Markdown 路径(默认 /chapter8_text_mining_probe.md)", ) ap.add_argument("--min-texts", type=int, default=8, help="细类最少评论条数才分析") ap.add_argument("--lda-topics", type=int, default=4, help="自动主题归纳条数上限(会按样本量裁剪)") ap.add_argument("--top-k-words", type=int, default=30, help="词频/关键词突出度展示长度") ap.add_argument("--cooc-vocab", type=int, default=80, help="共现矩阵保留的高频词数") ap.add_argument("--cooc-pairs", type=int, default=25, help="输出词对数量") ap.add_argument( "--live-llm", action="store_true", help="文末调用补充分析专用 LLM(PROBE_TEXT_MINING_SYSTEM + 结构化 JSON;需 AI_crawler 等)", ) ap.add_argument( "--llm-chunked", action="store_true", help="按细类拆分多次调用同一补充分析提示词,防单次 JSON 过长", ) ap.add_argument( "--no-wordcloud", action="store_true", help="不生成词云 PNG(默认生成,需 wordcloud+matplotlib)", ) ap.add_argument( "--wordcloud-max", type=int, default=40, help="最多为多少个细类各出一张词云(防文件过多)", ) args = ap.parse_args() md = build_markdown( args.run_dir, min_texts=args.min_texts, lda_topics_n=args.lda_topics, top_k_words=args.top_k_words, cooc_vocab=args.cooc_vocab, cooc_pairs=args.cooc_pairs, live_llm=args.live_llm, llm_chunked=args.llm_chunked, wordcloud_enabled=not args.no_wordcloud, wordcloud_max=max(0, args.wordcloud_max), ) out = args.out or (args.run_dir.resolve() / "chapter8_text_mining_probe.md") out.write_text(md, encoding="utf-8") print(f"已写入: {out}", flush=True) if __name__ == "__main__": main()