diff --git "a/app_agent.py" "b/app_agent.py"
new file mode 100644--- /dev/null
+++ "b/app_agent.py"
@@ -0,0 +1,3040 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+app_agent.py
+重譯文本分析系統 AI Agent（雙 Agent 架構）
+────────────────────────────────────────────
+主 Agent (LLM1): google/gemma-4-31b-it
+ → 研究情境詢問 & 最終研究報告撰寫
+工作程序Agent(LLM2): google/gemma-4-26b-a4b-it
+ → 文本上傳引導 / 格式驗證 / 指標計算 / 統計分析
+"""
+
+# ============================================================================
+# SECTION 1: Imports & Configuration
+# ============================================================================
+import sys, io, os, re, json, uuid, shutil, warnings, functools, collections
+import threading, zipfile, traceback
+
+# 強制 UTF-8 輸出（避免 Windows cp950 編碼錯誤）
+if sys.platform.startswith("win"):
+ try:
+ sys.stdout.reconfigure(encoding="utf-8", errors="replace", line_buffering=True)
+ sys.stderr.reconfigure(encoding="utf-8", errors="replace", line_buffering=True)
+ except AttributeError:
+ sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf-8", errors="replace")
+ sys.stderr = io.TextIOWrapper(sys.stderr.buffer, encoding="utf-8", errors="replace")
+from pathlib import Path
+from datetime import datetime
+from collections import Counter
+
+import numpy as np
+import pandas as pd
+import scipy.stats as stats
+import matplotlib
+matplotlib.use('Agg')
+import matplotlib.pyplot as plt
+import matplotlib.font_manager as fm
+from sklearn.ensemble import RandomForestClassifier
+from sklearn.model_selection import train_test_split
+from sklearn.metrics import accuracy_score, classification_report
+from sklearn.preprocessing import LabelEncoder, StandardScaler
+from sklearn.cluster import KMeans, AgglomerativeClustering
+from sklearn.decomposition import PCA
+from scipy.cluster.hierarchy import dendrogram, linkage
+from docx import Document
+from docx.shared import Pt, Cm, RGBColor
+from docx.enum.table import WD_TABLE_ALIGNMENT
+from docx.enum.text import WD_ALIGN_PARAGRAPH
+import urllib.request, urllib.error
+import gradio as gr
+
+warnings.filterwarnings('ignore')
+print = functools.partial(print, flush=True)
+
+# ── 載入 .env 環境變數 ───────────────────────────────────────────────────────
+try:
+ from dotenv import load_dotenv
+ load_dotenv(Path(__file__).parent / ".env")
+except ImportError:
+ print("[警告] 未安裝 python-dotenv，略過 .env 載入（請執行 pip install python-dotenv）")
+
+# ── 路徑常數 ────────────────────────────────────────────────────────────────
+BASE_DIR = Path(__file__).parent
+CORPUS_INFO = BASE_DIR / "01-Corpus" / "03-Info"
+SESSIONS_DIR = BASE_DIR / "sessions"
+SESSIONS_DIR.mkdir(exist_ok=True)
+
+# ── LLM 設定 ────────────────────────────────────────────────────────────────
+LLM1_API_KEY = os.environ.get("LLM1_API_KEY", "")
+LLM1_MODEL = os.environ.get("LLM1_MODEL", "google/gemma-4-31b-it")
+
+LLM2_API_KEY = os.environ.get("LLM2_API_KEY", "")
+LLM2_MODEL = os.environ.get("LLM2_MODEL", "google/gemma-4-26b-a4b-it")
+
+ALPHA = 0.05
+
+# ── POS 集合（與 Prog02 相同）───────────────────────────────────────────────
+Vset = ['VC','VH','VE','VJ','VA','V2','VK','VL','VD','VCL','VHC','VB','VF','VI','VAC','VG']
+Nset = ['Na','Nb','Nc','Ncd','Nd','Nf','Nh','Nv']
+Nhset = ['Nh']
+ADVset = ['D','Da','Dfa','Dfb','Dk']
+Cset = ['Caa','Cbb']
+Caset = ['Caa']
+Cbset = ['Cbb']
+DETset = ['Neu','Nes','Nep','Neqa']
+POSTset= ['Cab','Cba','Neqb','Ng']
+Tset = ['I','T']
+Aset = ['A']
+
+# ── 研究情境定義 ─────────────────────────────────────────────────────────────
+SCENARIOS = {
+ "1": {
+ "title": "同一原文的人工翻譯 vs 機器翻譯",
+ "short": "人機對比",
+ "upload_prompt": (
+ "請上傳欲分析的文本，文本檔名設計用 **A_xx.txt、B_xx.txt** "
+ "作為區分人工或機器翻譯。\n"
+ "上傳的文本必須使用 [ckip_segmentation_sumin](https://github.com/suhsiung/ckip_segmentation_sumin) "
+ "工具完成斷詞後的 `.txt` 檔案，"
+ "內容格式為：`詞_詞性`、`_詞性` 或空白行。"
+ ),
+ "desc": "焦點在人機差異的語言特徵。",
+ "scenario_module_key": "scenario_2",
+ },
+ "2": {
+ "title": "全譯 vs 編譯（��譯、改寫）",
+ "short": "全譯vs編譯",
+ "upload_prompt": (
+ "請上傳欲分析的文本，文本檔名設計用 **A_xx.txt、B_xx.txt** "
+ "作為區分全譯或編譯。\n"
+ "上傳的文本必須使用 [ckip_segmentation_sumin](https://github.com/suhsiung/ckip_segmentation_sumin) "
+ "工具完成斷詞後的 `.txt` 檔案，"
+ "內容格式為：`詞_詞性`、`_詞性` 或空白行。"
+ ),
+ "desc": "焦點在文本完整性和改寫策略。",
+ "scenario_module_key": "scenario_3",
+ },
+ "3": {
+ "title": "不同譯者的同期譯本",
+ "short": "譯者風格",
+ "upload_prompt": (
+ "請上傳欲分析的文本，文本檔名設計用 **A_xx.txt、B_xx.txt、C_xx.txt…** "
+ "作為區分不同的譯者。\n"
+ "上傳的文本必須使用 [ckip_segmentation_sumin](https://github.com/suhsiung/ckip_segmentation_sumin) "
+ "工具完成斷詞後的 `.txt` 檔案，"
+ "內容格式為：`詞_詞性`、`_詞性` 或空白行。"
+ ),
+ "desc": "焦點在個人風格差異。排除時代因素，討論譯者個人的語言偏好。",
+ "scenario_module_key": "scenario_4",
+ },
+ "4": {
+ "title": "同一原文的不同時期人工譯本",
+ "short": "歷時重譯",
+ "upload_prompt": (
+ "請上傳欲分析的文本，文本檔名設計用 **A_xx.txt、B_xx.txt、C_xx.txt…** "
+ "作為區分時期。\n"
+ "上傳的文本必須使用 [ckip_segmentation_sumin](https://github.com/suhsiung/ckip_segmentation_sumin) "
+ "工具完成斷詞後的 `.txt` 檔案，"
+ "內容格式為：`詞_詞性`、`_詞性` 或空白行。"
+ ),
+ "desc": "分析焦點在歷時變化、語言現代化、翻譯規範的演變。時間維度是核心變數。",
+ "scenario_module_key": "scenario_1",
+ },
+}
+
+# ── 指標分類（同 prog03）────────────────────────────────────────────────────
+INDEX_CATEGORIES = {
+ "基礎統計": {
+ "AA_01": "總字數", "AA_02": "相異字數", "AA_03": "總詞數",
+ "AA_04": "相異詞數", "AA_05": "段落數", "AA_06": "總句數",
+ },
+ "詞彙難易度": {"AC_01": "非常用詞比率", "AC_02": "難詞比率"},
+ "詞彙多樣性": {"AD_02": "MSTTR", "AD_03": "RTTR"},
+ "語意複雜度": {
+ "AG_01": "詞彙密度1", "AG_02": "詞彙密度2",
+ "AG_05": "每千詞成語數1", "AG_06": "每千詞成語數2",
+ },
+ "語法複雜度": {
+ "AH_01": "平均句長", "AH_03": "句長標準差",
+ "AH_06": "平均小句長度", "AH_08": "小句標準差",
+ },
+ "篇章連貫": {
+ "AJ_01": "對等關聯連接詞密度",
+ "AJ_02": "連接詞總密度",
+ "AJ_03": "從屬連接詞密度",
+ "AJ_04": "指示代詞密度",
+ "AJ_05": "一般代名詞密度",
+ "AJ_06": "指示照應代詞密度",
+ "AJ_07": "第一人稱代詞密度",
+ "AJ_08": "第二人稱代詞密度",
+ "AJ_09": "第三人稱代詞密度",
+ "AJ_10": "泛指與反身代詞密度",
+ "AJ_11": "泛指與反身代詞密度2",
+ },
+ "詞性標記": {
+ "AF_01": "A非謂形容詞", "AF_02": "Caa對等連接詞", "AF_03": "Cab連接詞",
+ "AF_04": "Cba連接詞", "AF_05": "Cbb關聯連接詞", "AF_06": "D副詞",
+ "AF_07": "Da數量副詞", "AF_08": "DE的之得地",
+ "AF_09": "Dfa動詞前程度副詞", "AF_10": "Dfb動詞後程度副詞",
+ "AF_11": "Di時態標記", "AF_12": "Dk句副詞", "AF_14": "I感嘆詞",
+ "AF_15": "Na普通名詞", "AF_16": "Nb專有名詞", "AF_17": "Nc地方詞",
+ "AF_18": "Ncd位置詞", "AF_19": "Nd時間詞", "AF_20": "Nep指代定詞",
+ "AF_21": "Neqa數量定詞", "AF_22": "Neqb後置數量定詞",
+ "AF_23": "Nes特指定詞", "AF_24": "Neu數詞定詞", "AF_25": "Nf量詞",
+ "AF_26": "Ng後置詞", "AF_27": "Nh代名詞", "AF_28": "Nv名物化動詞",
+ "AF_29": "P介詞", "AF_30": "SHI是", "AF_31": "T語助詞",
+ "AF_32": "V2有", "AF_33": "VA動作不及物動詞", "AF_34": "VAC動作使動動詞",
+ "AF_35": "VB動作類及物動詞", "AF_36": "VC動作及物動詞",
+ "AF_37": "VCL動作接地方賓語動詞", "AF_38": "VD雙賓動詞",
+ "AF_39": "VE動作句賓動詞", "AF_40": "VF動作謂賓動詞",
+ "AF_41": "VG分類動詞", "AF_42": "VH狀態不及物動詞",
+ "AF_43": "VHC狀態使動動詞", "AF_44": "VI狀態類及物動詞",
+ "AF_45": "VJ狀態及物動詞", "AF_46": "VK狀態句賓動詞",
+ "AF_47": "VL狀態謂賓動詞",
+ },
+}
+INDEX_NAME_MAP = {}
+INDEX_CATEGORY_MAP = {}
+for _cat, _items in INDEX_CATEGORIES.items():
+ for _code, _name in _items.items():
+ INDEX_NAME_MAP[_code] = _name
+ INDEX_CATEGORY_MAP[_code] = _cat
+
+# ── LLM 報告用情境模組（同 prog04）─────────────────────────────────────────
+SCENARIO_MODULE_1 = """## 二、情境特定分析：歷時變化分析
+
+在陳述差異時，必須將「時間順序」作為組織軸線。
+請依照以下子項目分析：
+
+### 2.1 語言現代化的跡象
+- 詞彙難易度（AC_01, AC_02）是否隨時間呈現規律變化？
+- 成語比率（AG_06）在不同時期譯本中的變化方向？
+- 若出現非單調變化，明確標示此為反常模式，需進一步解釋。
+
+### 2.2 句式與篇章結構的演變
+- 段落數、句數（AA_05, AA_06）的變化是否反映閱讀習慣的改變？
+- 詞性分布中，連接詞、代詞比例的變化是否呈現顯著化或隱化趨勢？
+
+### 2.3 譯本間的親疏關係
+- 根據分群結果，時序相鄰的譯本是否聚為一群？
+- 若出現時序相隔較遠的譯本反而更相似的情況，標記為需進一步探討的現象。
+
+### 禁令補充
+- 不得將「較新」直接等同於「較現代」或「較好」。"""
+
+SCENARIO_MODULE_2 = """## 二、情境特定分析：人機差異分析
+
+### 2.1 機器翻譯典型特徵的檢視
+根據量化數據檢視機器翻譯是否呈現以下已知特徵：
+- 詞彙多樣性是否偏低？（MSTTR, RTTR）
+- 句長是否較為整齊（變異數較小）？
+
+### 2.2 人工翻譯的差異化特徵
+- 人工譯本在哪些指標上顯著偏離機器譯本？
+
+### 禁令補充
+- 不得對機器翻譯系統的品質做整體評價。
+- 不得預設「人工一定優於機器」或反之。"""
+
+SCENARIO_MODULE_3 = """## 二、情境特定分析：改寫策略分析
+
+【重要限制】：本系統目前未進行原文對齊，所有關於刪減/增補的陳述必須明確標示為「推測」。
+
+### 3.1 規模差異的量化
+- 總字數、總詞數（AA_01, AA_03）的差距幅度
+
+### 3.2 改寫策略的語言訊號
+- 編譯本的詞彙密度是否反而較高？（可能反映資訊壓縮策略）
+- 編譯本的平均句長是否較短？（可能反映簡化策略）
+
+### 禁令補充
+- 嚴禁使用「忠實度」、「完整性」等規範性詞彙。"""
+
+SCENARIO_MODULE_4 = """## 二、情境特定分析：個人風格差異分析
+
+【前提】：所有譯本產出於相近時期，時代因素已被控制。
+
+### 2.1 核心風格指標的差異
+- 在哪些指標上，譯者間差異達到統計顯著？
+
+### 2.2 風格穩定性的檢視
+- 譯者內變異 vs 譯者間變異的比較
+
+### 禁令補充
+- 不得將個人風格與「翻譯能力」掛鉤。"""
+
+SCENARIO_MODULES = {
+ "scenario_1": SCENARIO_MODULE_1,
+ "scenario_2": SCENARIO_MODULE_2,
+ "scenario_3": SCENARIO_MODULE_3,
+ "scenario_4": SCENARIO_MODULE_4,
+}
+
+SYSTEM_PROMPT_BASE = """# 角色定義
+你是一位翻譯研究的量化分析助理。你的工作是將統計數據轉化為結構化的觀察與假設，供研究者進一步驗證。
+
+# 嚴格禁令
+1. 僅根據本次提供的量化數據進行分析，不得引入外部知識。
+2. 不得使用「優秀」、「忠實」、「準確」、「流暢」等評價性詞彙。
+3. 不得推測譯者的主觀意圖，除非數據直接支持且標明為推論。
+4. 遇到數據中未涵蓋的問題，明確回答「本次數據無法回答此問題」。
+5. 所有統計顯著性的陳述必須基於提供的 p 值或效果量。
+
+# 輸入資料結構
+【研究情境】：{scenario_type}
+【譯本 metadata】：{metadata}
+【量化分析結果】：（見 user message 中的 JSON 資料）
+【使用者補充關注點】：{user_focus}
+
+# 通用輸出結構
+輸出必須包含以下四個區塊，缺一不可：
+
+## 一、數據事實描述
+僅陳述統計結果，每項陳述後標註對應的指標代碼與數值。
+區分：達統計顯著的差異（標明 p 值與效果量）、未達顯著但數值差距較大的趨勢、無明顯差異的指標。
+
+{scenario_module}
+
+## 三、待驗證事項
+- 需研究者回到原文比對的具體項目
+- 數據中的異常值或可疑模式
+- 本分析無法處理的面向
+
+## 四、本次分析的已知限制
+明確列出：分詞系統潛在誤差、樣本量對統計檢定的影響、指標本身的理論侷限。"""
+
+# ── matplotlib 中文字型 ──────────────────────────────────────────────────────
+def _find_cjk_font():
+ candidates = ["Microsoft JhengHei","Microsoft YaHei","SimHei",
+ "Noto Sans CJK TC","Noto Sans CJK SC"]
+ available = {f.name for f in fm.fontManager.ttflist}
+ for c in candidates:
+ if c in available:
+ return c
+ return None
+
+_cjk = _find_cjk_font()
+if _cjk:
+ plt.rcParams['font.family'] = _cjk
+plt.rcParams['axes.unicode_minus'] = False
+
+# ============================================================================
+# SECTION 2: Resource Loading (load once at startup)
+# ============================================================================
+
+_RESOURCES = {}
+
+def _load_resources():
+ global _RESOURCES
+ print("【系統啟動】載入參考語料資源...")
+ def _read(fname):
+ p = CORPUS_INFO / fname
+ if not p.exists():
+ print(f" 警告：找不到 {fname}")
+ return []
+ with open(p, 'r', encoding='utf-8') as f:
+ return [l.replace('\n','') for l in f.readlines()]
+
+ _RESOURCES['T01'] = _read('T01_成語5335.txt')
+ _RESOURCES['T01_B'] = _read('T01_成語20389.txt')
+ _RESOURCES['T02'] = _read('T02_國小常用詞.txt')
+ _RESOURCES['T03'] = _read('T03_國小常用詞3000.txt')
+ _RESOURCES['T04'] = _read('T04_現代漢語3000.txt')
+ _RESOURCES['T05'] = _read('T05_現代漢語8000.txt')
+ _RESOURCES['T06_V500'] = _read('T06_Sinica高頻動詞前500.txt')
+ _RESOURCES['T06_N3000'] = _read('T06_Sinica高頻詞前3000詞_常用名詞.txt')
+ _RESOURCES['T06_A3000'] = _read('T06_Sinica高頻詞前3000詞_常用形容詞.txt')
+ _RESOURCES['T06_D3000'] = _read('T06_Sinica高頻詞前3000詞_常用副詞.txt')
+ _RESOURCES['T06_N8000'] = _read('T06_Sinica高頻詞前8000詞_常用名詞.txt')
+ _RESOURCES['T06_A8000'] = _read('T06_Sinica高頻詞前8000詞_常用形容詞.txt')
+ _RESOURCES['T06_D8000'] = _read('T06_Sinica高頻詞前8000詞_常用副詞.txt')
+
+ strokes_path = CORPUS_INFO / 'strokesDictionary.json'
+ if strokes_path.exists():
+ with open(strokes_path, 'r', encoding='utf-8') as f:
+ _RESOURCES['strokes'] = json.load(f)
+ else:
+ _RESOURCES['strokes'] = {}
+ print(" 警告：找不到 strokesDictionary.json")
+
+ print("【系統啟動】資源載入完成。")
+
+_load_resources()
+
+# ============================================================================
+# SECTION 3: LLM API Wrappers
+# ============================================================================
+
+def _call_openrouter(api_key, model, system_prompt, user_prompt, max_tokens=16000, temperature=0.5):
+ """通用 OpenRouter API 呼叫（同步）。"""
+ url = "https://openrouter.ai/api/v1/chat/completions"
+ headers = {
+ "Authorization": f"Bearer {api_key}",
+ "Content-Type": "application/json",
+ "HTTP-Referer": "https://retranslation-agent.local",
+ "X-Title": "Retranslation Analysis Agent",
+ }
+ payload = {
+ "model": model,
+ "messages": [
+ {"role": "system", "content": system_prompt},
+ {"role": "user", "content": user_prompt},
+ ],
+ "max_tokens": max_tokens,
+ "temperature": temperature,
+ }
+ data = json.dumps(payload).encode("utf-8")
+ req = urllib.request.Request(url, data=data, headers=headers, method="POST")
+ try:
+ with urllib.request.urlopen(req, timeout=180) as resp:
+ result = json.loads(resp.read().decode("utf-8"))
+ return result["choices"][0]["message"]["content"]
+ except urllib.error.HTTPError as e:
+ body = e.read().decode("utf-8") if e.fp else ""
+ raise RuntimeError(f"OpenRouter API HTTP 錯誤 ({e.code}): {body}")
+ except Exception as e:
+ raise RuntimeError(f"OpenRouter API 呼叫失敗: {e}")
+
+
+def call_main_agent(system_prompt, user_prompt):
+ """主 Agent (LLM1) 呼叫。"""
+ return _call_openrouter(LLM1_API_KEY, LLM1_MODEL, system_prompt, user_prompt)
+
+
+def call_worker_agent(system_prompt, user_prompt):
+ """工作程序 Agent (LLM2) 呼叫。"""
+ return _call_openrouter(LLM2_API_KEY, LLM2_MODEL, system_prompt, user_prompt)
+
+
+# ============================================================================
+# SECTION 4: Text Format Validation
+# ============================================================================
+
+def validate_ckip_format(filepath: Path):
+ """
+ 檢查檔案是否為 ckip_segmentation_sumin 工具產生的斷詞格式。
+ 判斷依據：
+ 1. 非空行中，含底線（_）的 token 佔比 > 70%
+ 2. 至少出現一個 CKIP 特徵標記（CATEGORY 系列或標準詞性碼）
+ 返回 (is_valid: bool, reason: str)
+ """
+ CKIP_SIGNATURES = re.compile(
+ r'(PERIODCATEGORY|COMMACATEGORY|PAUSECATEGORY|SEMICOLONCATEGORY|'
+ r'QUESTIONCATEGORY|EXCLAMATIONCATEGORY|PARENTHESISCATEGORY|'
+ r'DASHCATEGORY|ETCCATEGORY|COLONCATEGORY|SPCHANGECATEGORY|'
+ r'\b(?:Na|Nb|Nc|Ncd|Nd|Nf|Nh|Nv|VA|VB|VC|VD|VE|VF|VG|VH|VHC|'
+ r'VI|VJ|VK|VL|VAC|VCL|DE|Di|SHI|FW|Caa|Cab|Cba|Cbb|'
+ r'Neu|Nes|Nep|Neqa|Neqb|Ng|P|D|Da|Dfa|Dfb|Dk|I|T|A)\b)'
+ )
+ try:
+ with open(filepath, 'r', encoding='utf-8-sig') as f:
+ lines = f.readlines()
+ except Exception as e:
+ return False, f"無法讀取檔案：{e}"
+
+ total_tokens = 0
+ underscore_tokens = 0
+ has_ckip_signature = False
+ sample_lines = [l.rstrip() for l in lines if l.strip()][:100]
+
+ if not sample_lines:
+ return False, "檔案為空。"
+
+ for line in sample_lines:
+ tokens = line.split()
+ for tok in tokens:
+ total_tokens += 1
+ if '_' in tok:
+ underscore_tokens += 1
+ if not has_ckip_signature and CKIP_SIGNATURES.search(tok):
+ has_ckip_signature = True
+
+ if total_tokens == 0:
+ return False, "檔案無有效內容。"
+
+ ratio = underscore_tokens / total_tokens
+ if ratio < 0.70:
+ return False, (
+ f"token 中含底線（詞_詞性）的比例僅 {ratio:.0%}，"
+ "不符合 ckip_segmentation_sumin 的輸出格式。"
+ )
+ if not has_ckip_signature:
+ return False, "未偵測到 CKIP 詞性標記，可能非 ckip_segmentation_sumin 產生的檔案。"
+
+ return True, "格式符合 ckip_segmentation_sumin 輸出。"
+
+
+# ============================================================================
+# SECTION 5: Indicator Calculation (adapted from Prog02)
+# ============================================================================
+
+def _text2strokes(tokens, strokes_db):
+ words, strokes, missing = [], [], []
+ for tok in tokens:
+ for char in tok:
+ if char in strokes_db:
+ words.append(char)
+ strokes.append(strokes_db[char]['strokes_total'])
+ else:
+ missing.append(char)
+ return words, strokes, missing
+
+
+def calculate_text_index(corpus_file_path: Path, resources: dict) -> dict:
+ """
+ 計算單一文本的所有語言學指標。
+ 改寫自 Prog02_indicators_multi-v3_1150509.py 的 calculateTextIndex()。
+ """
+ fname = corpus_file_path.name
+
+ with open(corpus_file_path, 'r', encoding='utf-8-sig') as xf:
+ text0 = xf.read()
+
+ text0 = text0.replace('', '')
+ text1 = text0.split('\n')
+ text2 = [xx.split() for xx in text1]
+ text3_1, text3_2 = [], []
+ for i, row in enumerate(text2):
+ t1, t2 = [], []
+ for xxi in row:
+ parts = xxi.split('_')
+ t1.append(parts[0])
+ t2.append(parts[1] if len(parts) > 1 else '')
+ text3_1.append(t1)
+ text3_2.append(t2)
+
+ # ── 標點 / 句子統計 ───────────────────────────────────────────────────────
+ com1 = sum(row.count("COMMACATEGORY") for row in text3_2)
+ ss1 = sum(row.count("PERIODCATEGORY") for row in text3_2)
+ ss2 = sum(row.count("SEMICOLONCATEGORY") for row in text3_2)
+ ss3 = sum(row.count("QUESTIONCATEGORY") for row in text3_2)
+ ss4 = sum(row.count("EXCLAMATIONCATEGORY")for row in text3_2)
+ sentenceTotal = ss1 + ss2 + ss3 + ss4
+ if sentenceTotal == 0:
+ sentenceTotal = 1 # 防止除以零
+
+ A_index1 = com1 / sentenceTotal
+
+ # ── 展開成一維 ─────────────────────────────────────────────────────────────
+ text3_1_long = [t for row in text3_1 for t in row]
+ text3_2_long = [t for row in text3_2 for t in row]
+
+ sen_pos_set = ["PERIODCATEGORY","SEMICOLONCATEGORY","QUESTIONCATEGORY","EXCLAMATIONCATEGORY"]
+ sec_pos_set = sen_pos_set + ["COMMACATEGORY","COLONCATEGORY","DASHCATEGORY"]
+
+ def _calc_lengths(boundary_set):
+ lengths, xlen = [], 0
+ for sss in text3_2_long:
+ if sss in boundary_set:
+ lengths.append(xlen); xlen = 0
+ elif 'CATEGORY' not in sss:
+ xlen += 1
+ return lengths
+
+ sen_length = _calc_lengths(sen_pos_set)
+ sec_length = _calc_lengths(sec_pos_set)
+ if not sen_length: sen_length = [0]
+ if not sec_length: sec_length = [0]
+
+ mean_sentenceLength = np.mean(sen_length)
+ var_sentenceLength = np.var(sen_length)
+ mean_sectionLength = np.mean(sec_length)
+ var_sectionLength = np.var(sec_length)
+ sd_sectionLength = np.sqrt(var_sectionLength)
+
+ # ── 標點符號總數 ───────────────────────────────────────────────────────────
+ categoryTotal = sum(1 for t in text3_2_long if 'CATEGORY' in t)
+
+ # ── 詞 / 詞性統計 ───────────────────────────────────────────────────────────
+ tokens, tokens_pos, TokenAddPos, uniTokens, uni_TokenAddPos = [], [], [], [], []
+ for ii, row in enumerate(text3_1):
+ for jj in range(len(row)):
+ if 'CATEGORY' not in text3_2[ii][jj]:
+ w = text3_1[ii][jj]
+ p = text3_2[ii][jj]
+ tokens.append(w)
+ tokens_pos.append(p)
+ TokenAddPos.append(f"{w}-{p}")
+ if w not in uniTokens: uniTokens.append(w)
+ if f"{w}-{p}" not in uni_TokenAddPos: uni_TokenAddPos.append(f"{w}-{p}")
+
+ tokensTotal = len(tokens)
+ uniTokensTotal = len(uni_TokenAddPos)
+ if tokensTotal == 0: tokensTotal = 1
+
+ wordTotal = sum(len(w) for w in tokens)
+ words_0 = ''.join(tokens)
+ uni_words = set(words_0)
+ uni_words_len = len(uni_words)
+
+ # 詞長
+ tokens_len = [len(w) for w in tokens]
+ mean_tokens_len = np.mean(tokens_len) if tokens_len else 0
+
+ # 段落數
+ lineTotal = sum(1 for row in text3_1 if row)
+
+ # 基礎比值
+ A_index2 = wordTotal / sentenceTotal
+ A_index3 = tokensTotal / sentenceTotal
+ A_index4 = categoryTotal / sentenceTotal
+ A_index5 = wordTotal / max(categoryTotal, 1)
+ A_index6 = tokensTotal / max(categoryTotal, 1)
+ A_index7 = wordTotal / max(lineTotal, 1)
+ A_index8 = tokensTotal / max(lineTotal, 1)
+
+ # ── 詞頻比對 ──────────────────────────────────────────────────────────────
+ R = resources
+ elementaryTotal = sum(1 for w in tokens if w in R.get('T02', []))
+ elementaryTotal3 = sum(1 for w in tokens if w in R.get('T03', []))
+ elementaryTotal4 = sum(1 for w in tokens if w in R.get('T04', []))
+ elementaryTotal5 = sum(1 for w in tokens if w in R.get('T05', []))
+
+ xx = Counter(tokens)
+ idiomTotal = sum(v for k,v in xx.items() if k in R.get('T01', []))
+ idiomTotal2 = sum(v for k,v in xx.items() if k in R.get('T01_B',[]))
+
+ tokens_V = [tokens[i] for i in range(len(tokens)) if tokens_pos[i] in Vset]
+ tokens_N = [tokens[i] for i in range(len(tokens)) if tokens_pos[i] in Nset]
+ tokens_A = [tokens[i] for i in range(len(tokens)) if tokens_pos[i] in Aset] or ['XX']
+ tokens_ADV = [tokens[i] for i in range(len(tokens)) if tokens_pos[i] in ADVset]
+
+ VelementaryTotal500 = sum(1 for w in tokens_V if w in R.get('T06_V500', []))
+ NelementaryTotal3000 = sum(1 for w in tokens_N if w in R.get('T06_N3000',[]))
+ AelementaryTotal3000 = sum(1 for w in tokens_A if w in R.get('T06_A3000',[]))
+ ADVelementaryTotal3000= sum(1 for w in tokens_ADV if w in R.get('T06_D3000',[]))
+ NelementaryTotal8000 = sum(1 for w in tokens_N if w in R.get('T06_N8000',[]))
+ AelementaryTotal8000 = sum(1 for w in tokens_A if w in R.get('T06_A8000',[]))
+ ADVelementaryTotal8000= sum(1 for w in tokens_ADV if w in R.get('T06_D8000',[]))
+
+ # ── 詞性計數 ─────────────────────────────────────────────────────────────
+ pos_A_fre = tokens_pos.count("A")
+ pos_V_fre = sum(1 for p in tokens_pos if p in Vset)
+ pos_N_fre = sum(1 for p in tokens_pos if p in Nset)
+ pos_Nh_fre2 = sum(1 for p in tokens_pos if p in Nhset)
+ pos_ADV_fre = sum(1 for p in tokens_pos if p in ADVset)
+ pos_ASP_fre = tokens_pos.count("Di")
+ pos_C_fre = sum(1 for p in tokens_pos if p in Cset)
+ pos_Ca_fre = sum(1 for p in tokens_pos if p in Caset)
+ pos_Cb_fre = sum(1 for p in tokens_pos if p in Cbset)
+ pos_DET_fre = sum(1 for p in tokens_pos if p in DETset)
+ pos_P_fre = tokens_pos.count("P")
+ pos_POST_fre = sum(1 for p in tokens_pos if p in POSTset)
+ pos_T_fre = sum(1 for p in tokens_pos if p in Tset)
+ pos_DE_fre = tokens_pos.count("DE")
+ pos_SHI_fre = tokens_pos.count("SHI")
+ pos_V2_fre = tokens_pos.count("V2")
+
+ # 個別詞性
+ def _pc(p): return tokens_pos.count(p)
+ pos_Caa_fre = _pc("Caa"); pos_Cab_fre = _pc("Cab")
+ pos_Cba_fre = _pc("Cba"); pos_Cbb_fre = _pc("Cbb")
+ pos_D_fre = _pc("D"); pos_Da_fre = _pc("Da")
+ pos_Dfa_fre = _pc("Dfa"); pos_Dfb_fre = _pc("Dfb")
+ pos_Di_fre = _pc("Di"); pos_Dk_fre = _pc("Dk")
+ pos_I_fre = _pc("I"); pos_Na_fre = _pc("Na")
+ pos_Nb_fre = _pc("Nb"); pos_Nc_fre = _pc("Nc")
+ pos_Ncd_fre = _pc("Ncd"); pos_Nd_fre = _pc("Nd")
+ pos_Nep_fre = _pc("Nep"); pos_Neqa_fre= _pc("Neqa")
+ pos_Neqb_fre= _pc("Neqb");pos_Nes_fre = _pc("Nes")
+ pos_Neu_fre = _pc("Neu"); pos_Nf_fre = _pc("Nf")
+ pos_Ng_fre = _pc("Ng"); pos_Nh_fre = _pc("Nh")
+ pos_Nv_fre = _pc("Nv")
+ pos_VA_fre = _pc("VA"); pos_VAC_fre = _pc("VAC")
+ pos_VB_fre = _pc("VB"); pos_VC_fre = _pc("VC")
+ pos_VCL_fre = _pc("VCL"); pos_VD_fre = _pc("VD")
+ pos_VE_fre = _pc("VE"); pos_VF_fre = _pc("VF")
+ pos_VG_fre = _pc("VG"); pos_VH_fre = _pc("VH")
+ pos_VHC_fre = _pc("VHC"); pos_VI_fre = _pc("VI")
+ pos_VJ_fre = _pc("VJ"); pos_VK_fre = _pc("VK")
+ pos_VL_fre = _pc("VL")
+
+ # 內容詞
+ cont1 = pos_N_fre + pos_A_fre + pos_V_fre
+ cont2 = cont1 + pos_ADV_fre
+
+ # AJ 指標
+ AJ04_set = ['什麼','哪','什','甚麼','啥']
+ AJ05_set = ['它','牠','它們','牠們','其','之','前者','後者','兩者','二者',
+ '以上','以下','誰','誰人','何者','什麼','甚麼','啥']
+ AJ06_set = ['它','牠','它們','牠們','其','之','前者','後者','兩者','二者','以上','以下']
+ AJ07_set = ['我','我們','我方','我倆','咱們','咱','俺','儂','咱家','咱們','予','余',
+ '老子','我等','吾等']
+ AJ08_set = ['你','妳','你們','你方','妳們','妳方','你倆','妳倆']
+ AJ09_set = ['他','她','他們','她們','他倆']
+ AJ10_set = ['對方','個人','人家','各位','各自','自己','自我','本人','自身',
+ '別人','彼此','雙方','大家','大夥兒','大眾','他人','私人','彼方']
+ AJ02_set = ['Caa','Cbb','Cab','Cba']
+ AJ03_set = ['Cbb','Cba']
+
+ pos_Nep_AJ04_fre = sum(1 for i in range(len(tokens))
+ if tokens_pos[i]=="Nep" and tokens[i] not in AJ04_set)
+ pos_Nh_AJ05_fre = sum(1 for i in range(len(tokens))
+ if tokens_pos[i]=="Nh" and tokens[i] not in AJ05_set)
+ pos_Nh_AJ06_fre = sum(1 for i in range(len(tokens))
+ if tokens_pos[i]=="Nh" and tokens[i] in AJ06_set)
+ pos_Nh_AJ07_fre = sum(1 for i in range(len(tokens))
+ if tokens_pos[i]=="Nh" and tokens[i] in AJ07_set)
+ pos_Nh_AJ08_fre = sum(1 for i in range(len(tokens))
+ if tokens_pos[i]=="Nh" and tokens[i] in AJ08_set)
+ pos_Nh_AJ09_fre = sum(1 for i in range(len(tokens))
+ if tokens_pos[i]=="Nh" and tokens[i] in AJ09_set)
+ pos_Nh_AJ10_fre = sum(1 for i in range(len(tokens))
+ if tokens_pos[i]=="Nh" and tokens[i] in AJ10_set)
+ AJ02_fre = sum(1 for p in tokens_pos if p in AJ02_set)
+ AJ03_fre = sum(1 for p in tokens_pos if p in AJ03_set)
+
+ # ── 豐富度指標 ─────────────────────────────────────────────────────────────
+ TTR = uniTokensTotal / tokensTotal
+ perK = 1000
+ kk = len(TokenAddPos) // perK
+ kk_TTR = []
+ for ii in range(1, kk+1):
+ seg = TokenAddPos[(ii-1)*1000 : ii*1000]
+ kk_TTR.append(len(set(seg)) / perK)
+ STD_TTR = sum(kk_TTR) / len(kk_TTR) if kk_TTR else 0
+ Guiraud_R = uniTokensTotal / pow(tokensTotal, 0.5)
+ LexicalDensity1 = cont1 / tokensTotal
+ LexicalDensity2 = cont2 / tokensTotal
+
+ return {
+ "Text": [fname],
+ "AA_01": [wordTotal],
+ "AA_02": [uni_words_len],
+ "AA_03": [tokensTotal],
+ "AA_04": [uniTokensTotal],
+ "AA_05": [lineTotal],
+ "AA_06": [sentenceTotal],
+ "AC_01": [round((tokensTotal - elementaryTotal4) / tokensTotal * 1000, 4)],
+ "AC_02": [round((tokensTotal - elementaryTotal5) / tokensTotal * 1000, 4)],
+ "AD_02": [round(STD_TTR, 4)],
+ "AD_03": [round(Guiraud_R, 4)],
+ "AG_01": [round(LexicalDensity1, 4)],
+ "AG_02": [round(LexicalDensity2, 4)],
+ "AG_05": [round(idiomTotal / tokensTotal * 1000, 4)],
+ "AG_06": [round(idiomTotal2 / tokensTotal * 1000, 4)],
+ "AH_01": [round(mean_sentenceLength, 4)],
+ "AH_03": [round(np.sqrt(var_sentenceLength), 4)],
+ "AH_06": [round(mean_sectionLength, 4)],
+ "AH_08": [round(sd_sectionLength, 4)],
+ "AJ_01": [round(pos_C_fre / tokensTotal * 1000, 4)],
+ "AJ_02": [round(AJ02_fre / tokensTotal * 1000, 4)],
+ "AJ_03": [round(AJ03_fre / tokensTotal * 1000, 4)],
+ "AJ_04": [round(pos_Nep_AJ04_fre / tokensTotal * 1000, 4)],
+ "AJ_05": [round(pos_Nh_AJ05_fre / tokensTotal * 1000, 4)],
+ "AJ_06": [round(pos_Nh_AJ06_fre / tokensTotal * 1000, 4)],
+ "AJ_07": [round(pos_Nh_AJ07_fre / tokensTotal * 1000, 4)],
+ "AJ_08": [round(pos_Nh_AJ08_fre / tokensTotal * 1000, 4)],
+ "AJ_09": [round(pos_Nh_AJ09_fre / tokensTotal * 1000, 4)],
+ "AJ_10": [round(pos_Nh_AJ10_fre / tokensTotal * 1000, 4)],
+ "AJ_11": [round(pos_Nh_AJ10_fre / tokensTotal * 1000, 4)],
+ "AF_01": [round(pos_A_fre / tokensTotal * 1000, 4)],
+ "AF_02": [round(pos_Caa_fre / tokensTotal * 1000, 4)],
+ "AF_03": [round(pos_Cab_fre / tokensTotal * 1000, 4)],
+ "AF_04": [round(pos_Cba_fre / tokensTotal * 1000, 4)],
+ "AF_05": [round(pos_Cbb_fre / tokensTotal * 1000, 4)],
+ "AF_06": [round(pos_D_fre / tokensTotal * 1000, 4)],
+ "AF_07": [round(pos_Da_fre / tokensTotal * 1000, 4)],
+ "AF_08": [round(pos_DE_fre / tokensTotal * 1000, 4)],
+ "AF_09": [round(pos_Dfa_fre / tokensTotal * 1000, 4)],
+ "AF_10": [round(pos_Dfb_fre / tokensTotal * 1000, 4)],
+ "AF_11": [round(pos_Di_fre / tokensTotal * 1000, 4)],
+ "AF_12": [round(pos_Dk_fre / tokensTotal * 1000, 4)],
+ "AF_14": [round(pos_I_fre / tokensTotal * 1000, 4)],
+ "AF_15": [round(pos_Na_fre / tokensTotal * 1000, 4)],
+ "AF_16": [round(pos_Nb_fre / tokensTotal * 1000, 4)],
+ "AF_17": [round(pos_Nc_fre / tokensTotal * 1000, 4)],
+ "AF_18": [round(pos_Ncd_fre / tokensTotal * 1000, 4)],
+ "AF_19": [round(pos_Nd_fre / tokensTotal * 1000, 4)],
+ "AF_20": [round(pos_Nep_fre / tokensTotal * 1000, 4)],
+ "AF_21": [round(pos_Neqa_fre/ tokensTotal * 1000, 4)],
+ "AF_22": [round(pos_Neqb_fre/ tokensTotal * 1000, 4)],
+ "AF_23": [round(pos_Nes_fre / tokensTotal * 1000, 4)],
+ "AF_24": [round(pos_Neu_fre / tokensTotal * 1000, 4)],
+ "AF_25": [round(pos_Nf_fre / tokensTotal * 1000, 4)],
+ "AF_26": [round(pos_Ng_fre / tokensTotal * 1000, 4)],
+ "AF_27": [round(pos_Nh_fre / tokensTotal * 1000, 4)],
+ "AF_28": [round(pos_Nv_fre / tokensTotal * 1000, 4)],
+ "AF_29": [round(pos_P_fre / tokensTotal * 1000, 4)],
+ "AF_30": [round(pos_SHI_fre / tokensTotal * 1000, 4)],
+ "AF_31": [round(pos_T_fre / tokensTotal * 1000, 4)],
+ "AF_32": [round(pos_V2_fre / tokensTotal * 1000, 4)],
+ "AF_33": [round(pos_VA_fre / tokensTotal * 1000, 4)],
+ "AF_34": [round(pos_VAC_fre / tokensTotal * 1000, 4)],
+ "AF_35": [round(pos_VB_fre / tokensTotal * 1000, 4)],
+ "AF_36": [round(pos_VC_fre / tokensTotal * 1000, 4)],
+ "AF_37": [round(pos_VCL_fre / tokensTotal * 1000, 4)],
+ "AF_38": [round(pos_VD_fre / tokensTotal * 1000, 4)],
+ "AF_39": [round(pos_VE_fre / tokensTotal * 1000, 4)],
+ "AF_40": [round(pos_VF_fre / tokensTotal * 1000, 4)],
+ "AF_41": [round(pos_VG_fre / tokensTotal * 1000, 4)],
+ "AF_42": [round(pos_VH_fre / tokensTotal * 1000, 4)],
+ "AF_43": [round(pos_VHC_fre / tokensTotal * 1000, 4)],
+ "AF_44": [round(pos_VI_fre / tokensTotal * 1000, 4)],
+ "AF_45": [round(pos_VJ_fre / tokensTotal * 1000, 4)],
+ "AF_46": [round(pos_VK_fre / tokensTotal * 1000, 4)],
+ "AF_47": [round(pos_VL_fre / tokensTotal * 1000, 4)],
+ }
+
+
+def run_indicator_calculation(file_paths: list, output_dir: Path, resources: dict):
+ """
+ 對所有文本計算指標，儲存 xlsx + csv，回傳 (DataFrame, xlsx_path, csv_path)。
+ """
+ output_dir.mkdir(parents=True, exist_ok=True)
+ result_df = pd.DataFrame()
+ for fp in file_paths:
+ try:
+ row_dict = calculate_text_index(Path(fp), resources)
+ result_df = pd.concat([result_df, pd.DataFrame(row_dict)],
+ axis=0, ignore_index=True)
+ except Exception as e:
+ print(f" 計算 {fp} 時發生錯誤：{e}")
+ traceback.print_exc()
+
+ timestamp = datetime.now().strftime("%Y%m%d%H%M%S")
+ xlsx_path = output_dir / f"textIndex_calculated_1_{timestamp}.xlsx"
+ csv_path = output_dir / "textIndex_calculated.csv"
+ result_df.to_excel(str(xlsx_path), index=False)
+ result_df.to_csv(str(csv_path), index=False, encoding='utf-8-sig')
+ return result_df, xlsx_path, csv_path
+
+
+# ============================================================================
+# SECTION 6: Statistical Analysis (adapted from prog03, no tkinter)
+# ============================================================================
+
+def cohens_d(g1, g2):
+ n1, n2 = len(g1), len(g2)
+ if n1 < 2 or n2 < 2:
+ return 0.0
+ var1, var2 = np.var(g1, ddof=1), np.var(g2, ddof=1)
+ pooled = np.sqrt(((n1-1)*var1 + (n2-1)*var2) / (n1+n2-2))
+ return (np.mean(g1) - np.mean(g2)) / pooled if pooled != 0 else 0.0
+
+
+def eta_squared(group_data):
+ all_data = np.concatenate(group_data)
+ grand_mean = all_data.mean()
+ ss_between = sum(len(g) * (np.mean(g) - grand_mean)**2 for g in group_data)
+ ss_total = np.sum((all_data - grand_mean)**2)
+ return ss_between / ss_total if ss_total != 0 else 0.0
+
+
+def effect_label(val, method):
+ if method == "t-test":
+ v = abs(val)
+ return "大" if v>=0.8 else "中" if v>=0.5 else "小" if v>=0.2 else "極小"
+ else:
+ return "大" if val>=0.14 else "中" if val>=0.06 else "小" if val>=0.01 else "極小"
+
+
+def run_statistical_analysis(csv_path: Path, output_dir: Path):
+ """
+ 執行完整統計分析（敘述統計、差異分析、隨機森林、分群）。
+ 回傳 dict 包含所有輸出檔路徑與摘要資訊。
+ """
+ output_dir.mkdir(parents=True, exist_ok=True)
+
+ # ── 1. 讀入資料 ────────────────────────────────────────────────────────────
+ df = pd.read_csv(str(csv_path))
+
+ df['Type1'] = df['Text'].str[0].str.upper()
+ df['Type2'] = df['Type1']
+ GROUP_COL = 'Type1'
+ N_CLUSTERS = max(2, len(df['Type1'].unique()))
+
+ # 重新命名欄位（加中文名稱）
+ rename_map = {}
+ for col in df.columns:
+ if col in INDEX_NAME_MAP:
+ rename_map[col] = f"{col}_{INDEX_NAME_MAP[col]}"
+ df.rename(columns=rename_map, inplace=True)
+ numeric_cols = df.select_dtypes(include=[np.number]).columns.tolist()
+
+ # ── 2. 敘述統計（涵蓋所有計算指標，與語言指標分析一致）────────────────────
+ def _desc_row(cat, code, cname, rcol):
+ s = df[rcol].dropna()
+ if len(s) == 0:
+ return None
+ return {
+ "類別": cat, "指標代碼": code, "指標名稱": cname,
+ "文本個數": int(s.count()),
+ "平均數": round(float(s.mean()), 4),
+ "中位數": round(float(s.median()), 4),
+ "標準差": round(float(s.std()), 4),
+ "最小值": round(float(s.min()), 4),
+ "最大值": round(float(s.max()), 4),
+ "變異係數(%)": round(float(s.std()/s.mean()*100), 2) if s.mean()!=0 else None,
+ }
+
+ all_desc_rows = []
+ covered = set()
+ # (a) 已分類指標（依分類順序呈現）
+ for cat, items in INDEX_CATEGORIES.items():
+ for code, cname in items.items():
+ rcol = f"{code}_{cname}"
+ if rcol not in df.columns:
+ continue
+ r = _desc_row(cat, code, cname, rcol)
+ if r:
+ all_desc_rows.append(r)
+ covered.add(rcol)
+ # (b) 未分類指標（如 AJ 系列），補入確保與指標計算數量一致
+ for rcol in numeric_cols:
+ if rcol in covered:
+ continue
+ parts = rcol.split("_")
+ code = (parts[0] + "_" + parts[1]) if len(parts) >= 2 else rcol
+ cname = INDEX_NAME_MAP.get(code, code)
+ r = _desc_row("其他指標", code, cname, rcol)
+ if r:
+ all_desc_rows.append(r)
+ desc_df = pd.DataFrame(all_desc_rows)
+ desc_path = output_dir / "01_descriptive_stats.csv"
+ desc_df.to_csv(str(desc_path), index=False, encoding='utf-8-sig')
+
+ # ── 3. 差異性分析 ──────────────────────────────────────────────────────────
+ groups = df[GROUP_COL].unique()
+ n_groups = len(groups)
+ method_name = "Independent t-test (Welch)" if n_groups == 2 else "One-way ANOVA"
+
+ diff_results = []
+ for col in numeric_cols:
+ gdata = [df.loc[df[GROUP_COL]==g, col].dropna().values for g in groups]
+ if any(len(g)==0 for g in gdata): continue
+ row = {"指標": col}
+ for gi, g in enumerate(groups):
+ row[f"{g}_N"] = len(gdata[gi])
+ row[f"{g}_M"] = round(float(np.mean(gdata[gi])), 4)
+ row[f"{g}_SD"] = round(float(np.std(gdata[gi], ddof=1)), 4) if len(gdata[gi])>1 else 0.0
+ if n_groups == 2:
+ stat, p = stats.ttest_ind(gdata[0], gdata[1], equal_var=False)
+ es = cohens_d(gdata[0], gdata[1])
+ es_name = "Cohen's d"; method = "t-test"
+ else:
+ stat, p = stats.f_oneway(*gdata)
+ es = eta_squared(gdata)
+ es_name = "η²"; method = "ANOVA"
+ row["方法"] = method
+ row["統計量"] = round(float(stat), 4) if not np.isnan(stat) else None
+ row["p值"] = round(float(p), 4) if not np.isnan(p) else None
+ row["顯著"] = ("***" if p is not None and p<0.001 else
+ "**" if p is not None and p<0.01 else
+ "*" if p is not None and p<0.05 else "ns")
+ row["效果量名稱"] = es_name
+ row["效果量"] = round(float(es), 4)
+ row["效果大小"] = effect_label(es, method)
+ diff_results.append(row)
+
+ diff_df = pd.DataFrame(diff_results)
+ sig_df = diff_df[diff_df['顯著'] != "ns"]
+ diff_path = output_dir / "02_difference_analysis.csv"
+ diff_df.to_csv(str(diff_path), index=False, encoding='utf-8-sig')
+
+ # boxplot（最多 12 個顯著指標）
+ box_path = None
+ if len(sig_df) > 0:
+ sig_cols = sig_df['指標'].tolist()[:12]
+ n_row = max(1, (len(sig_cols) + 2) // 3)
+ fig, axes = plt.subplots(n_row, 3, figsize=(14, 4*n_row))
+ axes = np.array(axes).flatten()
+ for i, col in enumerate(sig_cols):
+ ax = axes[i]
+ pd_data = [df.loc[df[GROUP_COL]==g, col].dropna().values for g in groups]
+ ax.boxplot(pd_data, labels=groups)
+ pv = diff_df.loc[diff_df['指標']==col, 'p值'].values
+ pv_str = f"p={pv[0]:.4f}" if len(pv)>0 else ""
+ ax.set_title(f"{col}\n({pv_str})", fontsize=8)
+ for j in range(len(sig_cols), len(axes)):
+ axes[j].set_visible(False)
+ plt.suptitle(f"Significant Variables by {GROUP_COL}", fontsize=11, y=1.01)
+ plt.tight_layout()
+ box_path = output_dir / "02_significant_boxplots.png"
+ plt.savefig(str(box_path), dpi=150, bbox_inches='tight')
+ plt.close()
+
+ # ── 4. 隨機森林 ────────────────────────��───────────────────────────────────
+ le = LabelEncoder()
+ y = le.fit_transform(df[GROUP_COL])
+ X = df[numeric_cols].fillna(df[numeric_cols].median())
+
+ min_class = pd.Series(y).value_counts().min()
+ test_sz = max(len(np.unique(y)), int(len(df) * 0.2))
+
+ if len(df) > 6 and min_class >= 2 and test_sz < len(df):
+ X_train, X_test, y_train, y_test = train_test_split(
+ X, y, test_size=test_sz, random_state=42, stratify=y)
+ do_eval = True
+ else:
+ X_train, y_train = X, y
+ X_test, y_test = X, y
+ do_eval = False
+
+ rf = RandomForestClassifier(n_estimators=500, random_state=42)
+ rf.fit(X_train, y_train)
+ y_pred = rf.predict(X_test)
+ accuracy = accuracy_score(y_test, y_pred)
+
+ imp_df = pd.DataFrame({"指標": numeric_cols, "重要性": rf.feature_importances_})\
+ .sort_values("重要性", ascending=False).reset_index(drop=True)
+ imp_path = output_dir / "03_feature_importance.csv"
+ imp_df.to_csv(str(imp_path), index=False, encoding='utf-8-sig')
+
+ top_n = min(20, len(imp_df))
+ fig, ax = plt.subplots(figsize=(10, 6))
+ ax.barh(imp_df['指標'][:top_n][::-1], imp_df['重要性'][:top_n][::-1], color='steelblue')
+ ax.set_xlabel("Feature Importance")
+ ax.set_title(f"Random Forest — Top {top_n} Feature Importance")
+ plt.tight_layout()
+ imp_fig_path = output_dir / "03_feature_importance.png"
+ plt.savefig(str(imp_fig_path), dpi=150, bbox_inches='tight')
+ plt.close()
+
+ # ── 5. 非監督式分群 ─────────────────────────────────────────────────────────
+ scaler = StandardScaler()
+ X_scaled = scaler.fit_transform(X)
+ pca = PCA(n_components=min(2, X_scaled.shape[1]), random_state=42)
+ X_pca = pca.fit_transform(X_scaled)
+
+ # 階層式分群
+ linkage_mat = linkage(X_scaled, method='ward')
+ fig, ax = plt.subplots(figsize=(10, 5))
+ dendrogram(linkage_mat, labels=df['Text'].values, leaf_rotation=45, leaf_font_size=8, ax=ax)
+ ax.set_title("Hierarchical Clustering Dendrogram (Ward)")
+ ax.set_xlabel("Sample"); ax.set_ylabel("Distance")
+ plt.tight_layout()
+ dendro_path = output_dir / "04a_dendrogram.png"
+ plt.savefig(str(dendro_path), dpi=150, bbox_inches='tight')
+ plt.close()
+
+ hc = AgglomerativeClustering(n_clusters=N_CLUSTERS, linkage='ward')
+ df['HC_Cluster'] = hc.fit_predict(X_scaled)
+
+ # K-means + Elbow
+ max_k = min(8, max(2, len(df)-1))
+ inertia = []
+ for k in range(1, max_k+1):
+ km = KMeans(n_clusters=k, random_state=42, n_init=10)
+ km.fit(X_scaled)
+ inertia.append(km.inertia_)
+ fig, ax = plt.subplots(figsize=(7, 4))
+ ax.plot(list(range(1, max_k+1)), inertia, marker='o', color='coral')
+ ax.set_xlabel("Number of Clusters (K)"); ax.set_ylabel("Inertia (SSE)")
+ ax.set_title("K-means Elbow Curve"); ax.set_xticks(list(range(1, max_k+1)))
+ plt.tight_layout()
+ elbow_path = output_dir / "04b_kmeans_elbow.png"
+ plt.savefig(str(elbow_path), dpi=150, bbox_inches='tight')
+ plt.close()
+
+ km_final = KMeans(n_clusters=N_CLUSTERS, random_state=42, n_init=10)
+ df['KM_Cluster'] = km_final.fit_predict(X_scaled)
+
+ # PCA 2D 視覺化
+ fig, axes = plt.subplots(1, 2, figsize=(14, 5))
+ for ax, lcol, title in zip(axes,
+ ['HC_Cluster','KM_Cluster'],
+ ['Hierarchical Clustering', f'K-means (K={N_CLUSTERS})']):
+ ulabels = sorted(df[lcol].unique())
+ colors = plt.cm.tab10(np.linspace(0, 1, len(ulabels)))
+ for idx, cl in enumerate(ulabels):
+ mask = df[lcol] == cl
+ ax.scatter(X_pca[mask,0], X_pca[mask,1],
+ c=[colors[idx]], label=f"Cluster {cl}", s=100, edgecolors='k')
+ for i, txt in enumerate(df['Text']):
+ ax.annotate(txt[:8], (X_pca[i,0], X_pca[i,1]),
+ fontsize=7, textcoords='offset points', xytext=(5,3))
+ ax.set_xlabel(f"PC1 ({pca.explained_variance_ratio_[0]:.1%})" if X_scaled.shape[1]>=2 else "PC1")
+ ax.set_ylabel(f"PC2 ({pca.explained_variance_ratio_[1]:.1%})" if X_scaled.shape[1]>=2 else "PC2")
+ ax.set_title(title); ax.legend()
+ plt.suptitle("Clustering Results (PCA 2D Projection)", fontsize=12)
+ plt.tight_layout()
+ cluster_fig_path = output_dir / "04c_clustering_pca2d.png"
+ plt.savefig(str(cluster_fig_path), dpi=150, bbox_inches='tight')
+ plt.close()
+
+ # 分群結果儲存
+ result_cols = ['Text','Type1','Type2',GROUP_COL,'HC_Cluster','KM_Cluster']
+ result_cols = [c for c in result_cols if c in df.columns]
+ cluster_csv_path = output_dir / "05_cluster_results.csv"
+ df[result_cols].to_csv(str(cluster_csv_path), index=False, encoding='utf-8-sig')
+
+ # ── 6. 產出 docx 技術報告 ───────────────────���──────────────────────────────
+ timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+ report_path = output_dir / f"Final_Report-{timestamp}.docx"
+ doc = Document()
+ style = doc.styles['Normal']
+ style.font.name = '微軟正黑體'
+ style.font.size = Pt(10)
+
+ title = doc.add_heading('重譯分析技術報告', level=0)
+ title.alignment = WD_ALIGN_PARAGRAPH.CENTER
+ doc.add_paragraph(f'報告產出時間：{datetime.now().strftime("%Y-%m-%d %H:%M:%S")}',
+ style='Normal').alignment = WD_ALIGN_PARAGRAPH.CENTER
+ doc.add_paragraph(f'分析文本數：{df.shape[0]} 筆',
+ style='Normal').alignment = WD_ALIGN_PARAGRAPH.CENTER
+ doc.add_paragraph(f'分組方式：{GROUP_COL}',
+ style='Normal').alignment = WD_ALIGN_PARAGRAPH.CENTER
+ doc.add_paragraph('')
+
+ # 敘述統計
+ doc.add_heading('一、敘述性統計', level=1)
+ TABLE_HEADERS = ["指標代碼","指標名稱","文本個數","平均數","中位數","標準差","最小值","最大值","變異係數(%)"]
+ for cat, items in INDEX_CATEGORIES.items():
+ cat_data = desc_df[desc_df['類別'] == cat]
+ if cat_data.empty: continue
+ doc.add_heading(f'（{cat}）', level=2)
+ tbl = doc.add_table(rows=1+len(cat_data), cols=len(TABLE_HEADERS))
+ tbl.style = 'Light Shading Accent 1'
+ for j, h in enumerate(TABLE_HEADERS):
+ cell = tbl.rows[0].cells[j]
+ cell.text = h
+ for p in cell.paragraphs:
+ p.alignment = WD_ALIGN_PARAGRAPH.CENTER
+ for r in p.runs: r.bold = True; r.font.size = Pt(8)
+ for i, (_, row) in enumerate(cat_data.iterrows(), start=1):
+ vals = [str(row['指標代碼']), str(row['指標名稱']),
+ str(int(row['文本個數'])),
+ f"{row['平均數']:.4f}", f"{row['中位數']:.4f}",
+ f"{row['標準差']:.4f}", f"{row['最小值']:.4f}",
+ f"{row['最大值']:.4f}",
+ f"{row['變異係數(%)']:.2f}" if row['變異係數(%)'] is not None else "-"]
+ for j, val in enumerate(vals):
+ cell = tbl.rows[i].cells[j]
+ cell.text = val
+ for p in cell.paragraphs:
+ p.alignment = WD_ALIGN_PARAGRAPH.CENTER
+ for r in p.runs: r.font.size = Pt(8)
+ doc.add_paragraph('')
+
+ # 差異性分析
+ doc.add_page_break()
+ doc.add_heading('二、差異性分析', level=1)
+ doc.add_paragraph(f'分組依據：{GROUP_COL}（共 {n_groups} 組）。採用 {method_name}，顯著水準 α = {ALPHA}。')
+ doc.add_paragraph(f'共 {len(diff_df)} 個指標，其中 {len(sig_df)} 個達顯著差異（p < {ALPHA}）。')
+ group_sub_headers = []
+ for g in groups:
+ group_sub_headers.extend([f"{g}\nN", f"{g}\nM", f"{g}\nSD"])
+ diff_headers = ["指標"] + group_sub_headers + ["方法","統計量","p值","顯著","效果量","效果大小"]
+ diff_tbl = doc.add_table(rows=1+len(diff_df), cols=len(diff_headers))
+ diff_tbl.style = 'Light Shading Accent 1'
+ for j, h in enumerate(diff_headers):
+ cell = diff_tbl.rows[0].cells[j]
+ cell.text = h
+ for p in cell.paragraphs:
+ p.alignment = WD_ALIGN_PARAGRAPH.CENTER
+ for r in p.runs: r.bold = True; r.font.size = Pt(7)
+ for i, (_, row) in enumerate(diff_df.iterrows(), start=1):
+ vals = [str(row['指標'])]
+ for g in groups:
+ vals.extend([str(int(row[f"{g}_N"])),
+ f"{row[f'{g}_M']:.4f}",
+ f"{row[f'{g}_SD']:.4f}"])
+ vals.extend([str(row['方法']),
+ f"{row['統計量']:.4f}" if row['統計量'] is not None else "-",
+ f"{row['p值']:.4f}" if row['p值'] is not None else "-",
+ str(row['顯著']), f"{row['效果量']:.4f}", str(row['效果大小'])])
+ for j, val in enumerate(vals):
+ cell = diff_tbl.rows[i].cells[j]
+ cell.text = val
+ for p in cell.paragraphs:
+ p.alignment = WD_ALIGN_PARAGRAPH.CENTER
+ for r in p.runs:
+ r.font.size = Pt(7)
+ if row['顯著'] != "ns" and j == len(vals)-4:
+ r.font.color.rgb = RGBColor(0xCC, 0x00, 0x00)
+ doc.add_paragraph('')
+ if box_path and box_path.exists():
+ doc.add_heading('顯著指標箱形圖', level=2)
+ doc.add_picture(str(box_path), width=Cm(16))
+ doc.paragraphs[-1].alignment = WD_ALIGN_PARAGRAPH.CENTER
+
+ # 隨機森林
+ doc.add_page_break()
+ doc.add_heading('三、監督式學習 — 隨機森林特徵重要性', level=1)
+ doc.add_paragraph(f'模型準確率（Accuracy）：{accuracy:.4f}')
+ if imp_fig_path.exists():
+ doc.add_picture(str(imp_fig_path), width=Cm(15))
+ doc.paragraphs[-1].alignment = WD_ALIGN_PARAGRAPH.CENTER
+
+ # 分群
+ doc.add_page_break()
+ doc.add_heading('四、非監督式學習 — 集群分析', level=1)
+ if dendro_path.exists():
+ doc.add_heading('樹狀圖（階層式分群）', level=2)
+ doc.add_picture(str(dendro_path), width=Cm(15))
+ doc.paragraphs[-1].alignment = WD_ALIGN_PARAGRAPH.CENTER
+ if elbow_path.exists():
+ doc.add_heading('Elbow Curve（K-means）', level=2)
+ doc.add_picture(str(elbow_path), width=Cm(12))
+ doc.paragraphs[-1].alignment = WD_ALIGN_PARAGRAPH.CENTER
+ if cluster_fig_path.exists():
+ doc.add_heading('分群視覺化（PCA 2D）', level=2)
+ doc.add_picture(str(cluster_fig_path), width=Cm(16))
+ doc.paragraphs[-1].alignment = WD_ALIGN_PARAGRAPH.CENTER
+
+ doc.save(str(report_path))
+
+ return {
+ "df": df,
+ "desc_path": desc_path,
+ "diff_path": diff_path,
+ "imp_path": imp_path,
+ "imp_fig_path": imp_fig_path,
+ "box_path": box_path,
+ "dendro_path": dendro_path,
+ "elbow_path": elbow_path,
+ "cluster_fig_path": cluster_fig_path,
+ "cluster_csv_path": cluster_csv_path,
+ "report_path": report_path,
+ "sig_count": len(sig_df),
+ "total_indicators": len(diff_df),
+ "accuracy": accuracy,
+ "groups": list(groups),
+ "n_groups": n_groups,
+ "method_name": method_name,
+ "GROUP_COL": GROUP_COL,
+ "pca_ev": pca.explained_variance_ratio_.tolist(),
+ }
+
+
+# ============================================================================
+# SECTION 7: LLM Report Writing (adapted from prog04, uses LLM1)
+# ============================================================================
+
+def load_analysis_data_for_llm(csv_path: Path, output_dir: Path) -> dict:
+ """讀取分析資料組成 JSON 結構。"""
+ data = {}
+ if csv_path.exists():
+ raw = pd.read_csv(str(csv_path))
+ aj_cols = [c for c in raw.columns if c.startswith('AJ_')]
+ raw.drop(columns=aj_cols, inplace=True, errors='ignore')
+ data["texts"] = raw['Text'].tolist()
+ data["text_count"] = len(raw)
+ else:
+ data["texts"] = []; data["text_count"] = 0
+
+ for fname, key in [
+ ("01_descriptive_stats.csv", "descriptive_stats"),
+ ("02_difference_analysis.csv", "difference_analysis"),
+ ("03_feature_importance.csv", "feature_importance_top20"),
+ ("05_cluster_results.csv", "cluster_results"),
+ ]:
+ p = output_dir / fname
+ if p.exists():
+ d = pd.read_csv(str(p))
+ if key == "feature_importance_top20":
+ data[key] = d.head(20).to_dict(orient='records')
+ else:
+ data[key] = d.to_dict(orient='records')
+ if key == "difference_analysis":
+ sig = d[d['顯著'] != 'ns'] if '顯著' in d.columns else pd.DataFrame()
+ data["significant_indicators"] = sig.to_dict(orient='records')
+ data["significant_count"] = len(sig)
+ data["total_indicators"] = len(d)
+ else:
+ data[key] = []
+ if "significant_count" not in data:
+ data["significant_count"] = 0
+ if "total_indicators" not in data:
+ data["total_indicators"] = 0
+ return data
+
+
+def run_report_writing(scenario_key: str, user_metadata: str, user_focus: str,
+ csv_path: Path, output_dir: Path) -> dict:
+ """
+ 呼叫主 Agent (LLM1) 撰寫研究報告，儲存 docx + txt，回傳路徑 dict。
+ """
+ output_dir.mkdir(parents=True, exist_ok=True)
+ analysis_data = load_analysis_data_for_llm(csv_path, output_dir)
+
+ scenario_info = SCENARIOS.get(scenario_key, SCENARIOS["1"])
+ scenario_mod_key = scenario_info.get("scenario_module_key", "scenario_1")
+ scenario_module = SCENARIO_MODULES.get(scenario_mod_key, "")
+
+ metadata_str = f"共 {analysis_data['text_count']} 個譯本\n"
+ for i, t in enumerate(analysis_data["texts"], 1):
+ metadata_str += f" 譯本{i}：{t}\n"
+ if user_metadata:
+ metadata_str += f"\n【使用者補充】：\n{user_metadata}"
+
+ system_prompt = SYSTEM_PROMPT_BASE.format(
+ scenario_type = scenario_info["title"],
+ metadata = metadata_str,
+ user_focus = user_focus if user_focus else "無",
+ scenario_module = scenario_module,
+ )
+ user_prompt = (
+ f"請根據以下量化分析結果，按照 system prompt 中的輸出結構進行深度分析。\n\n"
+ f"【量化分析資料（JSON）】：\n```json\n"
+ f"{json.dumps(analysis_data, ensure_ascii=False, indent=2)}\n```\n\n"
+ "請嚴格按照四個區塊（一、數據事實描述 / 二、情境特定分析 / "
+ "三、待驗證事項 / 四、已知限制）輸出分析報告。"
+ )
+
+ result_text = call_main_agent(system_prompt, user_prompt)
+
+ timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+ docx_path = output_dir / f"LLM_Analysis-{timestamp}.docx"
+ txt_path = output_dir / f"LLM_Analysis-{timestamp}.txt"
+
+ # 存 docx
+ doc = Document()
+ style = doc.styles['Normal']
+ style.font.name = '微軟正黑體'
+ style.font.size = Pt(11)
+ heading = doc.add_heading('LLM 深度重譯研究報告', level=0)
+ heading.alignment = WD_ALIGN_PARAGRAPH.CENTER
+ doc.add_paragraph(f'研究情境：{scenario_info["title"]}',
+ style='Normal').alignment = WD_ALIGN_PARAGRAPH.CENTER
+ doc.add_paragraph(f'報告時間：{datetime.now().strftime("%Y-%m-%d %H:%M:%S")}',
+ style='Normal').alignment = WD_ALIGN_PARAGRAPH.CENTER
+ doc.add_page_break()
+ for line in result_text.split('\n'):
+ stripped = line.strip()
+ if stripped.startswith('#### '): doc.add_heading(stripped[5:], level=4)
+ elif stripped.startswith('### '): doc.add_heading(stripped[4:], level=3)
+ elif stripped.startswith('## '): doc.add_heading(stripped[3:], level=2)
+ elif stripped.startswith('# '): doc.add_heading(stripped[2:], level=1)
+ elif stripped.startswith('- '): doc.add_paragraph(stripped[2:], style='List Bullet')
+ elif stripped.startswith('* '): doc.add_paragraph(stripped[2:], style='List Bullet')
+ elif stripped: doc.add_paragraph(stripped)
+ doc.save(str(docx_path))
+
+ # 存 txt
+ with open(str(txt_path), 'w', encoding='utf-8') as f:
+ f.write(f"研究情境：{scenario_info['title']}\n")
+ f.write(f"時間：{datetime.now().strftime('%Y-%m-%d %H:%M:%S')}\n")
+ f.write("=" * 60 + "\n\n")
+ f.write(result_text)
+
+ return {
+ "docx_path": docx_path,
+ "txt_path": txt_path,
+ "result_text": result_text,
+ }
+
+
+# ============================================================================
+# SECTION 8: (Reserved)
+# ============================================================================
+
+
+# ============================================================================
+# SECTION 9: Conversation Flow & State Machine
+# ============================================================================
+
+STATES = {
+ "SCENARIO_SELECT": "研究情境選擇",
+ "UPLOAD_GUIDE": "文本上傳說明",
+ "UPLOAD_WAIT": "等待文本上傳",
+ "FORMAT_CHECK": "格式檢查中",
+ "CALCULATING": "指標計算中",
+ "ANALYZING": "統計分析中",
+ "REPORT_WRITING": "研究報告撰寫中",
+ "COMPLETE": "分析完成",
+}
+
+def _assistant_msg(text: str):
+ return {"role": "assistant", "content": text}
+
+def _user_msg(text: str):
+ return {"role": "user", "content": text}
+
+
+def make_initial_state(session_id: str) -> dict:
+ sess_dir = SESSIONS_DIR / session_id
+ sess_dir.mkdir(parents=True, exist_ok=True)
+ (sess_dir / "uploads").mkdir(exist_ok=True)
+ (sess_dir / "output").mkdir(exist_ok=True)
+ return {
+ "session_id": session_id,
+ "session_dir": str(sess_dir),
+ "state": "SCENARIO_SELECT",
+ "scenario": None,
+ "uploaded_files": [],
+ "invalid_info": {},
+ "cleaned_files": [],
+ "output_files": [],
+ "zip_path": None,
+ "user_metadata": "",
+ "user_focus": "",
+ "analysis_result": None,
+ "done_steps": [], # 已完成步驟（1-based）
+ "current_step": None, # 目前執行步驟
+ }
+
+
+def get_intro_message() -> str:
+ sys_p = (
+ "你是「重譯文本分析系統」的主 Agent（智能助理），"
+ "請用繁體中文、專業而友善的語氣向使用者介紹本系統，"
+ "並說明四種研究情境，最後邀請使用者選擇情境。"
+ "回覆請控制在 400 字以內。"
+ )
+ user_p = (
+ "請介紹本系統並列出以下四種研究情境，請使用者選擇：\n"
+ "1. 同一原文的人工翻譯 vs 機器翻譯\n"
+ "2. 全譯 vs 編譯（節譯、改寫）\n"
+ "3. 不同譯者的同期譯本\n"
+ "4. 同一原文的不同時期人工譯本\n"
+ "詢問使用者：「請輸入數字 1、2、3 或 4 選擇您要進行的研究情境。」"
+ )
+ try:
+ return call_main_agent(sys_p, user_p)
+ except Exception as e:
+ return (
+ "歡迎使用 TransPrism II 分析引擎 SpectraEngine。\n\n"
+ "請依照您的研究設計選擇分析情境：\n\n"
+ "1 人工翻譯 vs 機器翻譯\n"
+ "2 全譯 vs 編譯／節譯／改寫\n"
+ "3 同期不同譯者譯本比較\n"
+ "4 不同時期重譯本的歷時比較\n\n"
+ "輸入 1、2、3 或 4 。"
+ )
+
+
+def _build_todo_list(sc_num: str, done_steps: list = None, current_step: int = None) -> str:
+ """
+ 即時 to-do list 執行進度（確定性，不依賴 LLM）。
+ done_steps: 已完成步驟（1-based list）
+ current_step: 目前執行步驟（1-based int）
+ """
+ done_steps = done_steps or []
+ sc_info = SCENARIOS.get(sc_num, {})
+ sc_title = sc_info.get("short", sc_info.get("title", ""))
+
+ steps = [
+ (1, "情境選擇"),
+ (2, "上傳引導"),
+ (3, "文本上傳"),
+ (4, "格式驗證"),
+ (5, "指標計算"),
+ (6, "統計分析"),
+ (7, "報告撰寫"),
+ (8, "成果下載"),
+ ]
+
+ total = len(steps)
+ done_count = len([s for s in done_steps if s <= total])
+ pct = int(done_count / total * 100)
+
+ lines = [
+ f"**執行進度｜{sc_title}**",
+ f"`{'█' * (done_count * 2)}{'░' * ((total - done_count) * 2)}` {pct}%",
+ "",
+ ]
+ for num, name in steps:
+ if num in done_steps:
+ lines.append(f"✅ ~~{num}. {name}~~")
+ elif num == current_step:
+ lines.append(f"⏳ **{num}. {name}**")
+ else:
+ lines.append(f"☐ {num}. {name}")
+
+ return "\n\n".join(lines)
+
+
+def handle_scenario_select(user_input: str, state: dict):
+ """處理研究情境選擇輸入，回傳 (reply_text, new_state)。"""
+ choice = user_input.strip()
+ m = re.search(r'[1-4]', choice)
+ if not m:
+ return ("請輸入數字 **1、2、3 或 4** 選擇研究情境。", state)
+
+ sc_num = m.group()
+ sc_info = SCENARIOS[sc_num]
+ state["scenario"] = sc_num
+ state["done_steps"] = [1]
+ state["current_step"] = 2
+
+ sys_p = (
+ "你是重譯文本分析系統的主 Agent，請用繁體中文確認使用者的研究情境選擇，"
+ "並說明接下來將由工作程序 Agent 協助上傳文本。語氣專業友善，50 字以內。"
+ )
+ user_p = f"使用者選擇了研究情境 {sc_num}：{sc_info['title']}"
+ try:
+ confirm_text = call_main_agent(sys_p, user_p)
+ except Exception:
+ confirm_text = f"已記錄研究情境：**{sc_info['title']}**。"
+
+ state["state"] = "UPLOAD_GUIDE"
+ return (confirm_text, state)
+
+
+def get_upload_guide_message(scenario_num: str) -> str:
+ """工作程序 Agent 根據情境提供上傳說明。"""
+ sc_info = SCENARIOS[scenario_num]
+ sys_p = (
+ "你是重譯文本分析系統的工作程序 Agent，請用繁體中文說明文本上傳格式要求。"
+ "語氣清晰、條列式，120 字以內。"
+ "必須強調：文本必須使用 https://github.com/suhsiung/ckip_segmentation_sumin 工具完成斷詞，"
+ "才能上傳。最後提示使用者上傳檔案。"
+ )
+ user_p = f"研究情境：{sc_info['title']}\n上傳說明：\n{sc_info['upload_prompt']}"
+ try:
+ return call_worker_agent(sys_p, user_p)
+ except:
+ return (
+ f"**【工作程序 Agent】文本上傳說明**\n\n"
+ f"{sc_info['upload_prompt']}\n\n"
+ "請在下方選擇並上傳文本檔案（可多選）。"
+ )
+
+
+def handle_uploaded_files(files, state: dict):
+ """
+ 接收 Gradio 上傳的檔案，複製到 session 目錄，回傳狀態更新與訊息。
+ """
+ if not files:
+ return "尚未偵測到上傳檔案，請重新上傳。", state
+
+ uploads_dir = Path(state["session_dir"]) / "uploads"
+ uploads_dir.mkdir(parents=True, exist_ok=True)
+
+ saved_paths = []
+ for f in files:
+ src = Path(f.name) if hasattr(f, 'name') else Path(str(f))
+ dst = uploads_dir / src.name
+ shutil.copy2(str(src), str(dst))
+ saved_paths.append(str(dst))
+
+ state["uploaded_files"] = saved_paths
+ state["state"] = "FORMAT_CHECK"
+
+ # 工作程序 Agent 說明開始檢查
+ file_list_str = "\n".join([Path(p).name for p in saved_paths])
+ sys_p = (
+ "你是重譯文本分析系統的工作程序 Agent，"
+ "請用繁體中文簡短告知使用者已收到檔案並正在進行格式檢查，50 字以內。"
+ )
+ user_p = f"已接收以下 {len(saved_paths)} 個檔案：\n{file_list_str}"
+ try:
+ msg = call_worker_agent(sys_p, user_p)
+ except:
+ msg = f"已接收 **{len(saved_paths)}** 個檔案，正在進行格式檢查……"
+
+ return msg, state
+
+
+def run_format_check(state: dict):
+ """
+ 檢查所有上傳檔案是否為 ckip_segmentation_sumin 產生的格式。
+ 通過 → CALCULATING；不通過 → UPLOAD_WAIT（要求重新上傳）。
+ 返回 (reply_text, new_state, has_errors)
+ """
+ failed = {}
+ for fp in state["uploaded_files"]:
+ ok, reason = validate_ckip_format(Path(fp))
+ if not ok:
+ failed[Path(fp).name] = reason
+
+ if not failed:
+ state["state"] = "CALCULATING"
+ state["cleaned_files"] = state["uploaded_files"]
+ sys_p = (
+ "你是重譯文本分析系統的工作程序 Agent，"
+ "請用繁體中文告知使用者所有檔案已���認為 ckip_segmentation_sumin 斷詞格式，"
+ "即將開始計算指標，30 字以內。"
+ )
+ try:
+ msg = call_worker_agent(sys_p, "所有檔案格式驗證通過。")
+ except:
+ msg = "[OK] 所有檔案均符合 ckip_segmentation_sumin 斷詞格式，即將開始計算文本指標……"
+ return msg, state, False
+
+ # 有檔案不符
+ fail_list = "\n".join([f"- **{fn}**：{reason}" for fn, reason in failed.items()])
+ sys_p = (
+ "你是重譯文本分析系統的工作程序 Agent，請用繁體中文告知使用者："
+ "部分檔案不符合 ckip_segmentation_sumin 斷詞工具的輸出格式，"
+ "必須重新使用該工具斷詞後再上傳，語氣清晰，80 字以內。"
+ "請附上工具連結：https://github.com/suhsiung/ckip_segmentation_sumin"
+ )
+ try:
+ msg = call_worker_agent(sys_p, f"格式不符的檔案：\n{fail_list}")
+ msg = msg + f"\n\n{fail_list}"
+ except:
+ msg = (
+ f"[WARNING] 以下檔案不符合 ckip_segmentation_sumin 斷詞格式，請重新處理後再上傳：\n\n"
+ f"{fail_list}\n\n"
+ "請使用 [ckip_segmentation_sumin](https://github.com/suhsiung/ckip_segmentation_sumin) "
+ "工具完成斷詞後，重新上傳檔案。"
+ )
+ state["uploaded_files"] = []
+ state["state"] = "UPLOAD_WAIT"
+ return msg, state, True
+
+
+def run_computation_pipeline_streaming(state: dict):
+ """
+ 執行完整計算流程（指標計算 → 統計分析 → LLM 報告），
+ 使用 generator 逐步 yield (progress_msg, is_final, state)。
+ """
+ output_dir = Path(state["session_dir"]) / "output"
+ output_dir.mkdir(parents=True, exist_ok=True)
+
+ results_paths = []
+ anal = None
+
+ # ── 步驟 1：指標計算 ─────────────────────────────────────────────────────
+ yield "【步驟 1/3】正在計算文本指標……", False, state
+ try:
+ idx_df, xlsx_path, csv_path = run_indicator_calculation(
+ state["cleaned_files"], output_dir, _RESOURCES
+ )
+ results_paths.extend([str(xlsx_path), str(csv_path)])
+ yield f" [OK] 指標計算完成（{len(idx_df)} 個文本）", False, state
+ except Exception as e:
+ state["state"] = "COMPLETE"
+ yield f"[ERROR] 指標計算失敗：{e}", True, state
+ return
+
+ # ── 步驟 2：統計分析 ─────────────────────────────────────────────────────
+ yield "【步驟 2/3】正在執行統計分析……", False, state
+ try:
+ anal = run_statistical_analysis(csv_path, output_dir)
+ for key in ["desc_path","diff_path","imp_path","imp_fig_path",
+ "box_path","dendro_path","elbow_path","cluster_fig_path",
+ "cluster_csv_path","report_path"]:
+ p = anal.get(key)
+ if p and Path(str(p)).exists():
+ results_paths.append(str(p))
+ state["analysis_result"] = {
+ k: str(v) if isinstance(v, Path) else v
+ for k, v in anal.items() if k != "df"
+ }
+ yield (
+ f" [OK] 統計分析完成（{anal['sig_count']}/{anal['total_indicators']} 個顯著指標，"
+ f"隨機森林準確率 {anal['accuracy']:.4f}）"
+ ), False, state
+ except Exception as e:
+ state["state"] = "COMPLETE"
+ state["output_files"] = results_paths
+ yield f"[ERROR] 統計分析失敗：{e}", True, state
+ return
+
+ # ── 步驟 3：LLM 研究報告 ────────────────────────────────────────────────
+ yield "【步驟 3/3】主 Agent 正在撰寫研究報告（預計需要 30-90 秒）……", False, state
+ try:
+ rpt = run_report_writing(
+ scenario_key = state["scenario"],
+ user_metadata = state.get("user_metadata", ""),
+ user_focus = state.get("user_focus", ""),
+ csv_path = csv_path,
+ output_dir = output_dir,
+ )
+ results_paths.extend([str(rpt["docx_path"]), str(rpt["txt_path"])])
+ yield " [OK] 研究報告撰寫完成", False, state
+ except Exception as e:
+ yield f" [WARNING] LLM 報告撰寫失敗（{e}），其他分析結果仍可下載。", False, state
+
+ # ── 完成 ─────────────────────────────────────────────────────────────────
+ state["output_files"] = results_paths
+ state["state"] = "COMPLETE"
+
+ sig_info = ""
+ if anal:
+ sig_info = (
+ f"顯著指標：**{anal['sig_count']}/{anal['total_indicators']}** 個（p < {ALPHA}）。\n"
+ f"隨機森林分類準確率：**{anal['accuracy']:.4f}**。\n"
+ )
+ summary = (
+ f"**全部分析完成！**\n\n"
+ f"共分析 **{len(state['cleaned_files'])}** 個文本。\n"
+ f"{sig_info}\n"
+ f"各項結果請點選上方各分頁查看。"
+ )
+ yield summary, True, state
+
+
+# ============================================================================
+# SECTION 10: Gradio Interface (Tabbed)
+# ============================================================================
+
+_EMPTY_HINT = (
+ ''
+ '
'
+ '
尚未產生資料
'
+ '
請先於「主頁」完成分析流程
'
+ '
'
+ ' 對話區：統一深色主題，AI 訊息左邊藍線區分'
+ ' 進度面板：情境選擇後才顯示'
+ ' 執行紀錄：最新在上'
+ '
'
+ )
+
+ # ── 分頁 Tab 元件（控制顯示/隱藏）─────────────────────────────────
+ _RESULT_TABS = [tab2, tab3, tab6]
+
+ # ── 所有分頁內容輸出元件 + 3 個 Tab 可見性 ────────────────────────
+ _TAB_OUTPUTS = [
+ tab_idx_md, tab_idx_df, tab_idx_file, # 0-2 語言指標分析
+ tab_desc_md, tab_desc_df, # 3-4 敘述統計
+ tab_diff_head, tab_diff_df, tab_box_img, # 5-7 差異性分析
+ tab_rf_md, tab_imp_df, tab_imp_img, # 8-10 監督式學習
+ tab_clust_md, tab_dendro_img, tab_elbow_img, # 11-13 非監督式學習
+ tab_pca_img, tab_cluster_df, # 14-15
+ tab_stats_file, # 16 其他
+ tab_rpt_md, tab_rpt_file, # 17-18 研究報告
+ tab2, tab3, tab6, # 19-21 (visibility)
+ ]
+ _N_TAB = len(_TAB_OUTPUTS)
+
+ # ── helper ──────────────────────────────────────────────────────────
+ def _log_append(log_history, msg):
+ # 最新訊息放最前面（置頂顯示），以 markdown 字串呈現
+ prev = log_history or ""
+ line = f"`›` {msg}"
+ return f"{line}\n\n{prev}" if prev else line
+
+ def _populate_tabs(state):
+ """streaming 結束後呼叫，讀取 output 檔案並填入各分頁。"""
+ R = [gr.update()] * _N_TAB # 預設全部不動
+ if state is None:
+ return R
+
+ output_dir = Path(state["session_dir"]) / "output"
+ uploaded = state.get("uploaded_files", [])
+
+ # ── 需要 output 目錄存在 ──────────────────────────────────────
+ if not output_dir.exists():
+ return R
+
+ # ── 語言指標分析（index 0-2）──────────────────────────────────
+ csv_path = output_dir / "textIndex_calculated.csv"
+ xlsx_list = sorted(output_dir.glob("textIndex_calculated_1_*.xlsx"))
+ if csv_path.exists():
+ df = pd.read_csv(str(csv_path))
+ # 指標數 = 總欄位數扣除中介欄（Text/Type1/Type2）
+ meta_cols = [c for c in ["Text", "Type1", "Type2"] if c in df.columns]
+ n_indicators = len(df.columns) - len(meta_cols)
+ R[0] = f"### 指標計算結果\n共 **{len(df)}** 個文本，**{n_indicators}** 個語言指標。"
+ R[1] = gr.update(value=df, visible=True)
+ dl = [str(csv_path)]
+ if xlsx_list:
+ dl.append(str(xlsx_list[-1]))
+ R[2] = gr.update(value=dl, visible=True)
+
+ # ── 統計比較分析：敘述統計 + 差異分析（index 3-7）─────────────
+ desc_path = output_dir / "01_descriptive_stats.csv"
+ diff_path = output_dir / "02_difference_analysis.csv"
+ box_path = output_dir / "02_significant_boxplots.png"
+
+ if desc_path.exists():
+ desc = pd.read_csv(str(desc_path))
+ R[3] = f"### 敘述統計\n共 **{len(desc)}** 個指標的敘述統計。"
+ R[4] = gr.update(value=desc, visible=True)
+
+ if diff_path.exists():
+ diff = pd.read_csv(str(diff_path))
+ sig = diff[diff.get('顯著', pd.Series()) != 'ns'] if '顯著' in diff.columns else pd.DataFrame()
+ R[5] = gr.update(
+ value=f"### 差異性分析\n共 **{len(diff)}** 個指標，其中 **{len(sig)}** 個達顯著差異（p < {ALPHA}）。",
+ visible=True,
+ )
+ R[6] = gr.update(value=diff, visible=True)
+
+ if box_path.exists():
+ R[7] = gr.update(value=str(box_path), visible=True)
+
+ # ── 統計比較分析：監督式學習（index 8-10）─────────────────────
+ imp_path = output_dir / "03_feature_importance.csv"
+ imp_fig = output_dir / "03_feature_importance.png"
+ anal = state.get("analysis_result") or {}
+
+ if anal.get("accuracy") is not None:
+ acc = anal["accuracy"]
+ acc_str = f"{acc:.4f}" if isinstance(acc, float) else str(acc)
+ R[8] = f"### 隨機森林分類結果\n模型準確率（Accuracy）：**{acc_str}**"
+
+ if imp_path.exists():
+ imp = pd.read_csv(str(imp_path)).head(20)
+ R[9] = gr.update(value=imp, visible=True)
+ if imp_fig.exists():
+ R[10] = gr.update(value=str(imp_fig), visible=True)
+
+ # ── 統計比較分析：非監督式學習（index 11-15）──────────────────
+ dendro_path = output_dir / "04a_dendrogram.png"
+ elbow_path = output_dir / "04b_kmeans_elbow.png"
+ pca_fig_path = output_dir / "04c_clustering_pca2d.png"
+ cluster_path = output_dir / "05_cluster_results.csv"
+
+ pca_ev = anal.get("pca_ev", [])
+ ev_str = ""
+ if len(pca_ev) >= 2:
+ ev_str = f"\n\nPCA 解釋變異量：PC1 = {pca_ev[0]:.1%}，PC2 = {pca_ev[1]:.1%}"
+ if any(p.exists() for p in [dendro_path, elbow_path, pca_fig_path]):
+ R[11] = f"### 集群分析結果{ev_str}"
+
+ if dendro_path.exists():
+ R[12] = gr.update(value=str(dendro_path), visible=True)
+ if elbow_path.exists():
+ R[13] = gr.update(value=str(elbow_path), visible=True)
+ if pca_fig_path.exists():
+ R[14] = gr.update(value=str(pca_fig_path), visible=True)
+ if cluster_path.exists():
+ cdf = pd.read_csv(str(cluster_path))
+ R[15] = gr.update(value=cdf, visible=True)
+
+ # ── 統計比較分析：其他（下載統計檔案，index 16）───────────────
+ stat_files = [str(p) for p in [desc_path, diff_path] if p.exists()]
+ if stat_files:
+ R[16] = gr.update(value=stat_files, visible=True)
+
+ # ── 研究報告（index 17-18）────────────────────────────────────
+ txt_files = sorted(output_dir.glob("LLM_Analysis-*.txt"), reverse=True)
+ docx_files = sorted(output_dir.glob("LLM_Analysis-*.docx"), reverse=True)
+ final_rpts = sorted(output_dir.glob("Final_Report-*.docx"), reverse=True)
+
+ if txt_files:
+ try:
+ with open(str(txt_files[0]), 'r', encoding='utf-8') as f:
+ rpt_text = f.read()
+ R[17] = f"### LLM 深度研究報告\n\n{rpt_text}"
+ except Exception:
+ R[17] = "### LLM 深度研究報告\n\n（讀取報告時發生錯誤）"
+
+ dl_rpt = []
+ if final_rpts:
+ dl_rpt.append(str(final_rpts[0]))
+ if docx_files:
+ dl_rpt.append(str(docx_files[0]))
+ if txt_files:
+ dl_rpt.append(str(txt_files[0]))
+ if dl_rpt:
+ R[18] = gr.update(value=dl_rpt, visible=True)
+
+ # ── 顯示各分頁 Tab（index 19-21）──────────────────────────────
+ if csv_path.exists():
+ R[19] = gr.update(visible=True) # 語言指標分析
+ if desc_path.exists() or imp_path.exists() \
+ or any(p.exists() for p in [dendro_path, elbow_path, pca_fig_path]):
+ R[20] = gr.update(visible=True) # 統計比較分析
+ if txt_files or docx_files:
+ R[21] = gr.update(visible=True) # 研究報告
+
+ return R
+
+ # ── 啟動：先顯示「啟動中」，再載入 LLM 歡迎訊息 ───────────────────
+ def on_load_init():
+ """第一步：立即顯示啟動中提示。"""
+ sid = str(uuid.uuid4())[:8]
+ state = make_initial_state(sid)
+ return (
+ [_assistant_msg("系統啟動中...")],
+ "`›` 系統啟動中…",
+ state,
+ gr.update(visible=False),
+ gr.update(value=""),
+ )
+
+ def on_load_ready(chat_history, log_history, state):
+ """第二步：顯示固定歡迎訊息。"""
+ intro = (
+ "歡迎使用 TransPrism II 分析引擎 SpectraEngine。\n\n"
+ "請依照您的研究設計選擇分析情境：\n\n"
+ "1 人工翻譯 vs 機器翻譯\n"
+ "2 全譯 vs 編譯／節譯／改寫\n"
+ "3 同期不同譯者譯本比較\n"
+ "4 不同時期重譯本的歷時比較\n\n"
+ "輸入 1、2、3 或 4 。"
+ )
+ return (
+ [_assistant_msg(intro)],
+ "`›` 系統已就緒，等待使用者操作。",
+ )
+
+ demo.load(
+ fn=on_load_init,
+ outputs=[chatbot, exec_log, session_state, upload_row, todo_md],
+ ).then(
+ fn=on_load_ready,
+ inputs=[chatbot, exec_log, session_state],
+ outputs=[chatbot, exec_log],
+ )
+
+ # ── todo 即時更新 helper ─────────────────────────────────────────────
+ def _todo_update(state):
+ """根據 state 的 done_steps / current_step 回傳 todo_md gr.update。"""
+ if not state or not state.get("scenario"):
+ return gr.update(visible=False)
+ return gr.update(
+ visible=True,
+ value=_build_todo_list(
+ state["scenario"],
+ state.get("done_steps", []),
+ state.get("current_step"),
+ )
+ )
+
+ # ── 文字送出（streaming generator）──────────────────────────────────
+ def on_submit(user_text, chat_history, log_history, state):
+ if not user_text or not user_text.strip() or state is None:
+ yield chat_history, log_history, state, "", gr.update(), gr.update()
+ return
+
+ chat_history = chat_history + [_user_msg(user_text)]
+ cur = state["state"]
+
+ if cur == "SCENARIO_SELECT":
+ log_history = _log_append(log_history, "使用者選擇研究情境中……")
+ yield chat_history, log_history, state, "", gr.update(), gr.update()
+ reply, state = handle_scenario_select(user_text, state)
+ if state["state"] == "UPLOAD_GUIDE":
+ log_history = _log_append(log_history, "✅ STEP 1 情境選擇完成")
+ chat_history = chat_history + [_assistant_msg(reply)]
+ yield chat_history, log_history, state, "", gr.update(), _todo_update(state)
+ # STEP 2
+ state["current_step"] = 2
+ log_history = _log_append(log_history, "⏳ STEP 2 上傳引導中……")
+ yield chat_history, log_history, state, "", gr.update(), _todo_update(state)
+ guide = get_upload_guide_message(state["scenario"])
+ chat_history = chat_history + [_assistant_msg(guide)]
+ state["done_steps"] = [1, 2]
+ state["current_step"] = 3
+ state["state"] = "UPLOAD_WAIT"
+ log_history = _log_append(log_history, "✅ STEP 2 完成，等待上傳文本")
+ else:
+ chat_history = chat_history + [_assistant_msg(reply)]
+ yield chat_history, log_history, state, "", gr.update(visible=(state["state"] == "UPLOAD_WAIT")), _todo_update(state)
+ return
+
+ if cur == "UPLOAD_WAIT":
+ chat_history = chat_history + [_assistant_msg("請點擊下方「上傳已斷詞文本」選擇檔案並點擊「確認上傳」。")]
+ yield chat_history, log_history, state, "", gr.update(visible=True), gr.update()
+ return
+
+ if cur == "COMPLETE":
+ chat_history = chat_history + [_assistant_msg("分析已完成，請切換上方分頁查看各項結果。\n如需重新分析，請重新整理頁面。")]
+ yield chat_history, log_history, state, "", gr.update(), gr.update()
+ return
+
+ chat_history = chat_history + [_assistant_msg(f"目前狀態：{STATES.get(cur, '未知')}，請依系統提示操作。")]
+ yield chat_history, log_history, state, "", gr.update(), gr.update()
+
+ send_btn.click(
+ fn=on_submit,
+ inputs=[msg_input, chatbot, exec_log, session_state],
+ outputs=[chatbot, exec_log, session_state, msg_input, upload_row, todo_md],
+ ).then(
+ fn=_populate_tabs,
+ inputs=[session_state],
+ outputs=_TAB_OUTPUTS,
+ )
+ msg_input.submit(
+ fn=on_submit,
+ inputs=[msg_input, chatbot, exec_log, session_state],
+ outputs=[chatbot, exec_log, session_state, msg_input, upload_row, todo_md],
+ ).then(
+ fn=_populate_tabs,
+ inputs=[session_state],
+ outputs=_TAB_OUTPUTS,
+ )
+
+ # ── 檔案上傳確認（streaming generator）──────────────────────────────
+ def on_upload_confirm(files, chat_history, log_history, state):
+ if not files or state is None:
+ yield chat_history, log_history, state, gr.update(visible=True), gr.update()
+ return
+
+ # STEP 3（開始處理即關閉上傳彈窗，避免遮罩擋住操作）
+ state["current_step"] = 3
+ log_history = _log_append(log_history, "⏳ STEP 3 接收��傳檔案中……")
+ yield chat_history, log_history, state, gr.update(visible=False), _todo_update(state)
+
+ recv_msg, state = handle_uploaded_files(files, state)
+ chat_history = chat_history + [_assistant_msg(recv_msg)]
+ state["done_steps"] = list(set(state.get("done_steps", []) + [3]))
+ state["current_step"] = 4
+ log_history = _log_append(log_history, f"✅ STEP 3 已接收 {len(files)} 個檔案")
+ yield chat_history, log_history, state, gr.update(visible=False), _todo_update(state)
+
+ # STEP 4
+ log_history = _log_append(log_history, "⏳ STEP 4 格式驗證中……")
+ yield chat_history, log_history, state, gr.update(visible=False), _todo_update(state)
+
+ check_msg, state, has_errors = run_format_check(state)
+ chat_history = chat_history + [_assistant_msg(check_msg)]
+
+ if has_errors:
+ log_history = _log_append(log_history, "⚠️ STEP 4 格式不符，請重新上傳")
+ state["current_step"] = 3
+ yield chat_history, log_history, state, gr.update(visible=False), _todo_update(state)
+ return
+
+ state["done_steps"] = list(set(state.get("done_steps", []) + [4]))
+ state["current_step"] = 5
+ log_history = _log_append(log_history, "✅ STEP 4 格式驗證通過")
+ # 格式通過後：顯示 metadata 說明表單，等使用者填寫後再開始分析
+ state["state"] = "METADATA_INPUT"
+ meta_hint = "請填寫語料 Metadata 說明（原文資訊、群組層與個體層），完成後點「確認並開始分析」。"
+ chat_history = chat_history + [_assistant_msg(meta_hint)]
+ log_history = _log_append(log_history, "📑 請填寫語料 Metadata 說明")
+ yield chat_history, log_history, state, gr.update(visible=False), _todo_update(state)
+
+ # ── metadata 表單 helpers ───────────────────────────────────────────
+ def _build_meta_tables(state):
+ import os
+ files = [os.path.basename(p) for p in state.get("cleaned_files", [])]
+ groups = sorted(set(f[0].upper() for f in files if f))
+ return [[g, "", "", "", ""] for g in groups], [[f, "", "", "", "", ""] for f in files]
+
+ def _df_rows(dfval):
+ try:
+ if isinstance(dfval, pd.DataFrame):
+ return dfval.fillna("").values.tolist()
+ if isinstance(dfval, dict) and "data" in dfval:
+ return dfval["data"]
+ if isinstance(dfval, list):
+ return dfval
+ except Exception:
+ pass
+ return []
+
+ def _compile_metadata(state, title, author, lang, pubyear, version,
+ ngroups, method, group_df, indiv_df):
+ sc = state.get("scenario", "")
+ L = ["## 語料 Metadata", "",
+ "### 一、原文資訊",
+ f"- 原文標題：{title or '（未填）'}",
+ f"- 作者：{author or '（未填）'}",
+ f"- 語言：{lang or '（未填）'}",
+ f"- 出版年：{pubyear or '（未填）'}",
+ f"- 版本：{version or '（未填）'}", ""]
+ if sc in ("3", "4"):
+ L += ["### 二、比較結構",
+ f"- 群組數量：{ngroups}",
+ f"- 比較方式：{method}", ""]
+ gh = ["群組標籤", "翻譯類型", "翻譯方向", "譯者背景", "目標讀者"]
+ L += ["### 三、群組層資訊"]
+ for r in _df_rows(group_df):
+ if not any(str(c).strip() for c in r):
+ continue
+ L.append("- " + "；".join(f"{h}：{str(v).strip()}" for h, v in zip(gh, r) if str(v).strip()))
+ ih = ["檔案", "譯者／系統名稱", "出版年", "譯者母語", "後編輯程度", "重譯序位"]
+ L += ["", "### 四、個體層資訊"]
+ for r in _df_rows(indiv_df):
+ if not any(str(c).strip() for c in r):
+ continue
+ L.append("- " + "；".join(f"{h}：{str(v).strip()}" for h, v in zip(ih, r) if str(v).strip()))
+ return "\n".join(L)
+
+ def _run_computation(chat_history, log_history, state):
+ log_history = _log_append(log_history, "⏳ STEP 5 計算語言指標中……")
+ yield chat_history, log_history, state, _todo_update(state)
+ for prog_msg, is_final, state in run_computation_pipeline_streaming(state):
+ if "指標計算完成" in prog_msg:
+ state["done_steps"] = list(set(state.get("done_steps", []) + [5])); state["current_step"] = 6
+ elif "步驟 2/3" in prog_msg or "統計分析" in prog_msg and "完成" not in prog_msg:
+ state["current_step"] = 6
+ elif "統計分析完成" in prog_msg or ("統計" in prog_msg and "完成" in prog_msg):
+ state["done_steps"] = list(set(state.get("done_steps", []) + [6])); state["current_step"] = 7
+ elif "步驟 3/3" in prog_msg or "報告" in prog_msg and "完成" not in prog_msg:
+ state["current_step"] = 7
+ elif is_final:
+ state["done_steps"] = list(set(state.get("done_steps", []) + [7, 8])); state["current_step"] = None
+ log_history = _log_append(log_history, prog_msg)
+ if is_final:
+ chat_history = chat_history + [_assistant_msg(prog_msg)]
+ yield chat_history, log_history, state, _todo_update(state)
+
+ def _show_metadata_form(state):
+ _NO = gr.update()
+ if not state or state.get("state") != "METADATA_INPUT":
+ return _NO, _NO, _NO, _NO
+ group_rows, indiv_rows = _build_meta_tables(state)
+ show_cmp = state.get("scenario", "") in ("3", "4")
+ return (gr.update(visible=True), gr.update(value=group_rows),
+ gr.update(value=indiv_rows), gr.update(visible=show_cmp))
+
+ upload_btn.click(
+ fn=on_upload_confirm,
+ inputs=[file_input, chatbot, exec_log, session_state],
+ outputs=[chatbot, exec_log, session_state, upload_row, todo_md],
+ ).then(
+ fn=_show_metadata_form,
+ inputs=[session_state],
+ outputs=[metadata_row, meta_group_df, meta_indiv_df, meta_compare_group],
+ )
+
+ cancel_btn.click(
+ fn=lambda: (gr.update(visible=False), None),
+ inputs=None,
+ outputs=[upload_row, file_input],
+ )
+
+ def on_metadata_confirm(title, author, lang, pubyear, version,
+ ngroups, method, group_df, indiv_df,
+ chat_history, log_history, state):
+ if state is None:
+ yield chat_history, log_history, state, gr.update(), _todo_update(state)
+ return
+ state["user_metadata"] = _compile_metadata(state, title, author, lang, pubyear,
+ version, ngroups, method, group_df, indiv_df)
+ log_history = _log_append(log_history, "✅ 已記錄語料 Metadata，開始分析")
+ yield chat_history, log_history, state, gr.update(visible=False), _todo_update(state)
+ for ch, lh, st, todo in _run_computation(chat_history, log_history, state):
+ chat_history, log_history, state = ch, lh, st
+ yield chat_history, log_history, state, gr.update(visible=False), todo
+
+ meta_confirm_btn.click(
+ fn=on_metadata_confirm,
+ inputs=[meta_title, meta_author, meta_lang, meta_pubyear, meta_version,
+ meta_ngroups, meta_method, meta_group_df, meta_indiv_df,
+ chatbot, exec_log, session_state],
+ outputs=[chatbot, exec_log, session_state, metadata_row, todo_md],
+ ).then(fn=_populate_tabs, inputs=[session_state], outputs=_TAB_OUTPUTS)
+
+ def on_metadata_skip(chat_history, log_history, state):
+ if state is None:
+ yield chat_history, log_history, state, gr.update(), _todo_update(state)
+ return
+ state["user_metadata"] = state.get("user_metadata", "") or "（使用者未提供語料 Metadata 說明）"
+ log_history = _log_append(log_history, "⏭️ 已略過 Metadata，直接開始分析")
+ yield chat_history, log_history, state, gr.update(visible=False), _todo_update(state)
+ for ch, lh, st, todo in _run_computation(chat_history, log_history, state):
+ chat_history, log_history, state = ch, lh, st
+ yield chat_history, log_history, state, gr.update(visible=False), todo
+
+ meta_cancel_btn.click(
+ fn=on_metadata_skip,
+ inputs=[chatbot, exec_log, session_state],
+ outputs=[chatbot, exec_log, session_state, metadata_row, todo_md],
+ ).then(fn=_populate_tabs, inputs=[session_state], outputs=_TAB_OUTPUTS)
+
+ return demo
+
+
+# ============================================================================
+# SECTION 11: Entry Point
+# ============================================================================
+
+if __name__ == "__main__":
+ print("=" * 60)
+ print("重譯文本分析系統 AI Agent 啟動中……")
+ print(f"主 Agent : {LLM1_MODEL}")
+ print(f"工作程序 : {LLM2_MODEL}")
+ print(f"BASE_DIR : {BASE_DIR}")
+ print("=" * 60)
+ app = create_app()
+ app.launch(
+ server_name="0.0.0.0",
+ server_port=7860,
+ share=False,
+ inbrowser=False,
+ allowed_paths=[str(SESSIONS_DIR)],
+ )