"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# Collect words from Google Analogy dataset\n",
"SUB_CATEGORY = \": family\"\n",
"\n",
"# TODO3: Plot t-SNE for the words in the SUB_CATEGORY `: family`\n",
"family_words = sorted({w for q in data[data[\"SubCategory\"] == SUB_CATEGORY][\"Question\"]\n",
" for w in q.lower().split()})\n",
"family_words = [w for w in family_words if w in model.key_to_index]\n",
"vectors = np.array([model[w] for w in family_words])\n",
"# t-SNE：把 100 維壓成 2 維。字只有幾十個，所以 perplexity 要調小（必須小於樣本數）。\n",
"points = TSNE(n_components=2, perplexity=15, random_state=42, init=\"pca\").fit_transform(vectors)\n",
"plt.figure(figsize=(12, 9))\n",
"plt.scatter(points[:, 0], points[:, 1], s=15)\n",
"for (x, y), w in zip(points, family_words):\n",
" plt.annotate(w, (x, y), fontsize=9)\n",
"\n",
"\n",
"plt.title(\"Word Relationships from Google Analogy Task\")\n",
"plt.show()\n",
"plt.savefig(\"word_relationships.png\", bbox_inches=\"tight\")\n"
]
},
{
"cell_type": "markdown",
"id": "c7945139",
"metadata": {
"id": "DKRPJxgKXH4j"
},
"source": [
"### Part III: Train your own word embeddings"
]
},
{
"cell_type": "markdown",
"id": "f18b21fd",
"metadata": {
"id": "VC_0fE1UzL8T"
},
"source": [
"### Get the latest English Wikipedia articles and do sampling.\n",
"- Usually, we start from Wikipedia dump (https://dumps.wikimedia.org/enwiki/latest/enwiki-latest-pages-articles.xml.bz2). However, the downloading step will take very long. Also, the cleaning step for the Wikipedia corpus ([`gensim.corpora.wikicorpus.WikiCorpus`](https://radimrehurek.com/gensim/corpora/wikicorpus.html#gensim.corpora.wikicorpus.WikiCorpus)) will take much time. Therefore, we provide cleaned files for you."
]
},
{
"cell_type": "code",
"execution_count": 15,
"id": "210cd691",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T11:42:10.209112Z",
"iopub.status.busy": "2026-10-05T11:42:10.208107Z",
"iopub.status.idle": "2026-10-05T11:42:10.303671Z",
"shell.execute_reply": "2026-10-05T11:42:10.303169Z"
},
"id": "FkubArwCCYxR"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"data\\wiki_texts_part_0.txt.gz 1.51 GB\n",
"data\\wiki_texts_part_1.txt.gz 0.84 GB\n",
"data\\wiki_texts_part_2.txt.gz 0.67 GB\n",
"data\\wiki_texts_part_3.txt.gz 0.60 GB\n",
"data\\wiki_texts_part_4.txt.gz 0.57 GB\n",
"data\\wiki_texts_part_5.txt.gz 0.58 GB\n",
"data\\wiki_texts_part_6.txt.gz 0.56 GB\n",
"data\\wiki_texts_part_7.txt.gz 0.53 GB\n",
"data\\wiki_texts_part_8.txt.gz 0.55 GB\n",
"data\\wiki_texts_part_9.txt.gz 0.55 GB\n",
"data\\wiki_texts_part_10.txt.gz 0.00 GB\n"
]
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# [Data loading changed] Only the data-loading part is modified (allowed by the assignment rules).\n",
"# Original template comment(s):\n",
"# Download the split Wikipedia files\n",
"# Each file contain 562365 lines (articles).\n",
"# 原本是 11 行 `!gdown --id ...`。改成 Python 迴圈：檔案已經存在就跳過，斷線後可以重跑。\n",
"# Each file contain 562365 lines (articles), except the last file.\n",
"import os, gdown\n",
"DATA_DIR = \"data\"\n",
"os.makedirs(DATA_DIR, exist_ok=True)\n",
"WIKI_IDS = [\n",
" \"1jiu9E1NalT2Y8EIuWNa1xf2Tw1f1XuGd\", \"1ABblLRd9HXdXvaNv8H9fFq984bhnowoG\",\n",
" \"1z2VFNhpPvCejTP5zyejzKj5YjI_Bn42M\", \"1VKjded9BxADRhIoCzXy_W8uzVOTWIf0g\",\n",
" \"16mBeG26m9LzHXdPe8UrijUIc6sHxhknz\", \"17JFvxOH-kc-VmvGkhG7p3iSZSpsWdgJI\",\n",
" \"19IvB2vOJRGlrYulnTXlZECR8zT5v550P\", \"1sjwO8A2SDOKruv6-8NEq7pEIuQ50ygVV\",\n",
" \"1s7xKWJmyk98Jbq6Fi1scrHy7fr_ellUX\", \"17eQXcrvY1cfpKelLbP2BhQKrljnFNykr\",\n",
" \"1J5TAN6bNBiSgTIYiPwzmABvGhAF58h62\",\n",
"]\n",
"for i, fid in enumerate(WIKI_IDS):\n",
" path = os.path.join(DATA_DIR, f\"wiki_texts_part_{i}.txt.gz\")\n",
" if not os.path.exists(path):\n",
" gdown.download(id=fid, output=path, quiet=True)\n",
" print(path, f\"{os.path.getsize(path)/1e9:.2f} GB\")\n"
]
},
{
"cell_type": "code",
"execution_count": 16,
"id": "c5010838",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T11:42:10.338774Z",
"iopub.status.busy": "2026-10-05T11:42:10.338774Z",
"iopub.status.idle": "2026-10-05T11:42:10.342061Z",
"shell.execute_reply": "2026-10-05T11:42:10.342061Z"
},
"id": "8S3ibNT3C8Xk"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"all parts downloaded\n"
]
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# [Data loading changed] Only the data-loading part is modified (allowed by the assignment rules).\n",
"# Original template comment(s):\n",
"# Download the split Wikipedia files\n",
"# Each file contain 562365 lines (articles), except the last file.\n",
"# 上一格已經把 11 個檔都下載完（模板原本把下載拆成兩格）。\n",
"print(\"all parts downloaded\")\n"
]
},
{
"cell_type": "code",
"execution_count": 17,
"id": "7abcd7cf",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T11:42:10.378091Z",
"iopub.status.busy": "2026-10-05T11:42:10.377087Z",
"iopub.status.idle": "2026-10-05T11:42:10.381439Z",
"shell.execute_reply": "2026-10-05T11:42:10.381439Z"
},
"id": "DUg_c79BC7OL"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"11 parts: ['wiki_texts_part_0.txt.gz', 'wiki_texts_part_1.txt.gz', 'wiki_texts_part_2.txt.gz', 'wiki_texts_part_3.txt.gz', 'wiki_texts_part_4.txt.gz', 'wiki_texts_part_5.txt.gz', 'wiki_texts_part_6.txt.gz', 'wiki_texts_part_7.txt.gz', 'wiki_texts_part_8.txt.gz', 'wiki_texts_part_9.txt.gz', 'wiki_texts_part_10.txt.gz']\n"
]
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# [Data loading changed] Only the data-loading part is modified (allowed by the assignment rules).\n",
"# Original template comment(s):\n",
"# Extract the downloaded wiki_texts_parts files.\n",
"# 原本是 `!gunzip -k`（解壓縮並保留壓縮檔）＋ 下一格 `!cat`（合併成一個大檔）。\n",
"# 這樣同一份資料會在硬碟上存三份（壓縮檔＋解壓檔＋合併檔），本機硬碟吃不消。\n",
"# 改成：不解壓、不合併，下一步直接用 gzip 一行一行讀壓縮檔。結果完全一樣。\n",
"import glob, gzip, re\n",
"wiki_parts = sorted(glob.glob(os.path.join(DATA_DIR, \"wiki_texts_part_*.txt.gz\")),\n",
" key=lambda p: int(re.search(r\"part_(\\d+)\", p).group(1)))\n",
"print(len(wiki_parts), \"parts:\", [os.path.basename(p) for p in wiki_parts])\n"
]
},
{
"cell_type": "code",
"execution_count": 18,
"id": "06330269",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T11:42:10.415988Z",
"iopub.status.busy": "2026-10-05T11:42:10.415988Z",
"iopub.status.idle": "2026-10-05T11:42:10.420034Z",
"shell.execute_reply": "2026-10-05T11:42:10.420034Z"
},
"id": "7duk2RbYDB02"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"total compressed size: 6.93 GB\n"
]
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# [Data loading changed] Only the data-loading part is modified (allowed by the assignment rules).\n",
"# Original template comment(s):\n",
"# Combine the extracted wiki_texts_parts files.\n",
"# 不需要合併：wiki_parts 這個清單按 0 到 10 的順序排好，依序讀就等於讀合併檔。\n",
"print(\"total compressed size:\", round(sum(os.path.getsize(p) for p in wiki_parts) / 1e9, 2), \"GB\")\n"
]
},
{
"cell_type": "code",
"execution_count": 19,
"id": "6f2b85c2",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T11:42:10.461341Z",
"iopub.status.busy": "2026-10-05T11:42:10.461341Z",
"iopub.status.idle": "2026-10-05T11:42:10.476186Z",
"shell.execute_reply": "2026-10-05T11:42:10.476186Z"
},
"id": "givLH7NrDs6X"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"0 anarchism is political philosophy and movement that is against all forms of authority and seeks to abolish the institutions it claims maintain unnecessary coercion and hierarchy typically including th ...\n",
"1 albedo change in greenland the map shows the difference between the amount of sunlight greenland reflected in the summer of versus the average percent it reflected between and some areas reflect close ...\n",
"2 or is the first letter and the first vowel letter of the latin alphabet used in the modern english alphabet and others worldwide its name in english is pronounced plural aes it is similar in shape to ...\n",
"3 alabama is state in the southeastern region of the united states it borders tennessee to the north georgia to the east florida and the gulf of mexico to the south and mississippi to the west alabama i ...\n",
"4 in greek mythology achilles or achilleus was hero of the trojan war who was known as being the greatest of all the greek warriors the central character in homer iliad he was the son of the nereid thet ...\n",
"5 abraham lincoln february april was an american lawyer politician and statesman who served as the th president of the united states from until his assassination in he led the united states through the ...\n",
"6 aristotle aristotélēs bc was an ancient greek philosopher and polymath his writings cover broad range of subjects spanning the natural sciences philosophy linguistics economics politics psychology and ...\n",
"7 an american in paris is jazz influenced symphonic poem or tone poem for orchestra by american composer george gershwin first performed in it was inspired by the time that gershwin had spent in paris a ...\n",
"8 the academy award for best production design recognizes achievement for art direction in film the category original name was best art direction but was changed to its current name in for the th academ ...\n",
"9 the academy awards of merit commonly known as the oscars or academy awards are awards for artistic and technical merit given for excellence within the american and international film industry they are ...\n"
]
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# [Data loading changed] Only the data-loading part is modified (allowed by the assignment rules).\n",
"# Original template comment(s):\n",
"# Check the first ten lines of the combined file\n",
"# 原本是 `!head -n 10`。每一行是一整篇文章，很長，所以每行只印前 200 個字元。\n",
"with gzip.open(wiki_parts[0], \"rt\", encoding=\"utf-8\") as f:\n",
" for i, line in enumerate(f):\n",
" print(i, line[:200].strip(), \"...\")\n",
" if i == 9:\n",
" break\n"
]
},
{
"cell_type": "markdown",
"id": "fb905716",
"metadata": {
"id": "Hfwx92QCEhrq"
},
"source": [
"Please note that we used the default parameters of [`gensim.corpora.wikicorpus.WikiCorpus`](https://radimrehurek.com/gensim/corpora/wikicorpus.html#gensim.corpora.wikicorpus.WikiCorpus) for cleaning the Wiki raw file. Thus, words with one character were discarded."
]
},
{
"cell_type": "code",
"execution_count": 20,
"id": "621fdd1f",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T11:42:10.513442Z",
"iopub.status.busy": "2026-10-05T11:42:10.513442Z",
"iopub.status.idle": "2026-10-05T11:45:16.915879Z",
"shell.execute_reply": "2026-10-05T11:45:16.915879Z"
},
"id": "vUAzButoP03w"
},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"\r",
" 0%| | 0/11 [00:00 big 會讓比較級、最高級、複數、過去式的答案消失\n",
"TOKEN_RE = re.compile(r\"^[a-z]+$\")\n",
"\n",
"def preprocess_file(src, dst):\n",
" n_in = n_tokens = 0\n",
" with open(src, encoding=\"utf-8\") as fin, open(dst, \"w\", encoding=\"utf-8\", newline=\"\\n\") as fout:\n",
" for line in fin:\n",
" raw = line.split()\n",
" tokens = [w for w in raw if TOKEN_RE.match(w)]\n",
" n_in += len(raw)\n",
" n_tokens += len(tokens)\n",
" fout.write(\" \".join(tokens) + \"\\n\")\n",
" print(f\"{src}: removed non-[a-z] tokens {n_in - n_tokens:,} of {n_in:,} ({(n_in - n_tokens) / n_in:.2%})\")\n",
" return n_tokens\n",
"\n",
"# 超參數：向量長度跟 GloVe 一樣是 100，比較才公平；skip-gram 是老師投影片的重點。\n",
"# max_final_vocab：只留最常見的 30 萬個字（老師建議的「只保留高頻字」）。\n",
"# 注意：設了 max_final_vocab 之後，gensim 會自動把 min_count 往上調到剛好留下 30 萬字（實際值印在下面）。\n",
"# seed：固定初始值；但 workers>1 時多執行緒的計算順序不固定，訓練結果仍會有小幅差異（報告第 5 題 (a) 有量）。\n",
"W2V_PARAMS = dict(vector_size=100, window=5, sg=1, negative=5, sample=1e-3,\n",
" min_count=5, max_final_vocab=300_000, epochs=5, workers=15, seed=42)\n",
"\n",
"def train_w2v(clean_path):\n",
" t = time.time()\n",
" m = Word2Vec(corpus_file=clean_path, **W2V_PARAMS)\n",
" print(f\"trained on {clean_path}: {(time.time() - t) / 60:.1f} min, vocab {len(m.wv.key_to_index):,}, \"\n",
" f\"effective min_count {m.effective_min_count}\")\n",
" return m\n",
"\n",
"clean_path = \"wiki_sampled_20_clean.txt\"\n",
"n_tokens = preprocess_file(output_path, clean_path)\n",
"print(f\"20% wiki tokens after preprocessing: {n_tokens:,}\")\n",
"my_model = train_w2v(clean_path)\n",
"my_model.save(\"w2v_wiki20.model\")\n"
]
},
{
"cell_type": "code",
"execution_count": 22,
"id": "7a162bbc",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T12:27:09.315308Z",
"iopub.status.busy": "2026-10-05T12:27:09.315308Z",
"iopub.status.idle": "2026-10-05T12:27:09.343647Z",
"shell.execute_reply": "2026-10-05T12:27:09.343647Z"
},
"id": "qWiQF70izxP7"
},
"outputs": [],
"source": [
"data = pd.read_csv(\"questions-words.csv\")"
]
},
{
"cell_type": "code",
"execution_count": 23,
"id": "9124ab13",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T12:27:09.379734Z",
"iopub.status.busy": "2026-10-05T12:27:09.379734Z",
"iopub.status.idle": "2026-10-05T12:28:55.099934Z",
"shell.execute_reply": "2026-10-05T12:28:55.099934Z"
},
"id": "q6xpqgdIy5x1"
},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"\r",
" 0%| | 0/19544 [00:00"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"text/plain": [
"
"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# Collect words from Google Analogy dataset\n",
"SUB_CATEGORY = \": family\"\n",
"\n",
"# TODO7: Plot t-SNE for the words in the SUB_CATEGORY `: family`\n",
"family_words = sorted({w for q in data[data[\"SubCategory\"] == SUB_CATEGORY][\"Question\"]\n",
" for w in q.lower().split()})\n",
"family_words = [w for w in family_words if w in my_wv.key_to_index]\n",
"vectors = np.array([my_wv[w] for w in family_words])\n",
"# t-SNE：把 100 維壓成 2 維。字只有幾十個，所以 perplexity 要調小（必須小於樣本數）。\n",
"points = TSNE(n_components=2, perplexity=15, random_state=42, init=\"pca\").fit_transform(vectors)\n",
"plt.figure(figsize=(12, 9))\n",
"plt.scatter(points[:, 0], points[:, 1], s=15)\n",
"for (x, y), w in zip(points, family_words):\n",
" plt.annotate(w, (x, y), fontsize=9)\n",
"\n",
"\n",
"plt.title(\"Word Relationships from Google Analogy Task\")\n",
"plt.show()\n",
"plt.savefig(\"word_relationships.png\", bbox_inches=\"tight\")\n"
]
},
{
"cell_type": "markdown",
"id": "231bfb89",
"metadata": {},
"source": [
"Running environment: Local — Windows 11 (10.0.26200), CPU: Intel Core i7-11800H (8 cores / 16 threads), RAM 16 GB\n",
"\n",
"Python version: 3.12.3\n",
"\n",
"> 說明：除了附出處連結的外部資料（GloVe 規模、論文）與執行環境，以下數字都來自本筆記本的執行輸出（差值與平均字數由輸出數字相減或相除），括號內標出是哪一格的輸出，用 Jupyter 畫面左邊的執行序號表示（例如〔In [23]〕）。比較實驗一律先對照第 5 題 (a) 量到的雜訊範圍；標「（推論）」的是根據原理的解釋，沒有另外用實驗驗證。\n",
"\n",
"##### 1. Which embedding model do you use? What are the pre-processing steps? What are the hyperparameter settings? (5%)\n",
"\n",
"Answer:\n",
"\n",
"**模型**\n",
"- Part II：`glove-wiki-gigaword-100`（Stanford GloVe，Wikipedia＋Gigaword 約 60 億字訓練，40 萬字、100 維、全部小寫；https://nlp.stanford.edu/projects/glove/ ）。\n",
"- Part III：Gensim `Word2Vec`（Skip-gram, `sg=1`），用 20% Wikipedia 自己訓練。\n",
"\n",
"**抽樣**〔In [20]〕：以「篇」為單位，`random.seed(42)`，每篇擲一個亂數 `r`，`r < 0.2` 就留下，共 1,124,733 / 5,623,655 篇（20.00%）。同一個 `r` 同時產生 5%、10% 樣本，所以 5% ⊂ 10% ⊂ 20%，三份只差在資料量。直接串流讀取 11 個 `.gz` 檔，不解壓、不合併，記憶體裡只放一篇文章。\n",
"\n",
"**前處理**〔In [21]〕\n",
"- 只保留 `[a-z]+` 的字。助教的資料已經是小寫、去標點、斷好字，所以這一步只刪掉 0.72% 的字（4,828,696 / 670,881,822）。\n",
"- 字典最多保留 30 萬個最常見的字（`max_final_vocab=300000`）。\n",
"- **刻意不移除停用詞**：考卷有 206 題含 NLTK 停用詞（family 86 題，如 he/she/his/her；currency 58 題，如 Korea won 的 won）〔In [35]〕，移除後這些題目必錯；老師上課講搜尋引擎建索引時也提到，移除停用詞「以前很重要，但現在一點都不重要也不會去做」，而且可能刪掉重要的東西（W2 01:39:50；投影片 W1 p47 \"Avoid to broke the semantics\"）；本作業的實驗結果也支持不刪。第 5 題 (b) 有實驗。\n",
"- **刻意不做詞形還原**：語法類考的就是字形變化（bigger、biggest、cars、danced），還原成原形會讓答案消失。\n",
"\n",
"**超參數**：`vector_size=100`（與 GloVe-100 相同，方便比較）、`window=5`、`sg=1`、`negative=5`、`sample=1e-3`、`min_count=5`、`max_final_vocab=300000`、`epochs=5`、`workers=15`、`seed=42`。\n",
"- 注意：因為設了 `max_final_vocab`，gensim 會自動把門檻往上調。20% 維基的實際門檻是出現 **30** 次以上才收進字典（`effective min_count 30`）；5%、10% 分別是 9、16〔In [21], In [31]〕。\n",
"- `seed` 固定初始值，但多執行緒訓練的計算順序不固定，結果仍有小幅變動（第 5 題 (a)）。\n",
"- 20% 維基處理後 666,053,126 字，訓練 37.9 分鐘，字典 299,405 字〔In [21]〕。\n",
"\n",
"**答題方式**：3CosAdd，`most_similar(positive=[b, c], negative=[a])`（gensim 會排除題目的 a、b、c 三個字）；題目先轉小寫；題目的字不在字典裡就算答錯（本次 0 題）。\n",
"\n",
"##### 2. What will the performance be like if you sample 5%, 10% and 20% of wiki text in TODO4? (10%, 3% for each)\n",
"\n",
"Answer:\n",
"\n",
"〔In [21], In [27], In [31]〕\n",
"\n",
"| | 5% | 10% | 20% | 5%→20% | 雜訊範圍* |\n",
"|---|---|---|---|---|---|\n",
"| 處理後字數 | 166,194,401 | 333,052,711 | 666,053,126 | | |\n",
"| 實際字典門檻 | 9 | 16 | 30 | | |\n",
"| 訓練時間 | 9.1 min | 19.7 min | 37.9 min | | |\n",
"| **Overall** | 46.12 | 47.15 | **48.39** | +2.27 | 1.05 |\n",
"| Semantic | 55.35 | 55.60 | 56.70 | +1.35 | 1.80 |\n",
"| Syntactic | 38.45 | 40.13 | 41.48 | +3.03 | 1.26 |\n",
"| family | 68.77 | 73.32 | 74.31 | +5.54 | 4.74 |\n",
"| gram4-superlative | 17.02 | 22.37 | 24.33 | +7.31 | 2.76 |\n",
"| gram2-opposite | 11.82 | 14.29 | 16.01 | +4.19 | 2.22 |\n",
"| capital-common-countries | 81.82 | 76.48 | 77.87 | −3.95 | 2.77 |\n",
"| OOV 題數 | 107 | 0 | 0 | | |\n",
"\n",
"\\* 雜訊範圍＝同樣用 5% 維基、只換亂數種子訓練 4 次，最高與最低的差（第 5 題 (a)〔In [34]〕）。差距的絕對值大於它，才當作真的有差。\n",
"\n",
"觀察：\n",
"- **整體隨資料量上升，但幅度小**：資料每多一倍（5→10→20%），整體約多 1 分（+1.03、+1.24），訓練時間卻跟著加倍。整體 +2.27 大於雜訊 1.05，是可信的上升。\n",
"- **語法類的上升明確，語意類加總不明確**：Syntactic +3.03（雜訊 1.26），最高級 +7.31、反義字首 +4.19 都超過雜訊；Semantic 加總只 +1.35，**小於雜訊 1.80，不能下結論**。語意類的五個小類都超過各自的雜訊，但方向不一：family +5.54、capital-world +2.98、currency +2.66 上升，capital-common-countries −3.95、city-in-state −1.86 下降，加總後互相抵消。（推論）biggest、worst、unaware 這類字形比較少見，需要更多出現次數才學得穩。\n",
"- **5% 有 107 題 OOV，10% 以後沒有**：5% 的資料裡有些考題的字出現次數低於實際門檻（9 次），沒有收進字典。\n",
"- capital-common-countries（−3.95）、city-in-state（−1.86）的下降超過雜訊，但雜訊只用 4 個種子、在 5% 量的，所以只說「語意類沒有隨資料穩定上升」，不推論成「資料多反而變差」。\n",
"- 限制：雜訊是在 5% 量的，20% 的雜訊可能不同。\n",
"\n",
"##### 3. What is the performance for different categories or sub-categories when trained on different corpora? (15%)\n",
"\n",
"3.1. Present your results. (5%)\n",
"\n",
"Answer:\n",
"\n",
"語料：SemEval 2016/2017 Task 3 的論壇問答（Qatar Living 生活論壇）。為了把「資料量」與「內容種類」分開，另外從 5% 維基依序切出**同樣字數**（72,920,923 字，論壇是 72,919,300 字）的對照組；兩者用相同前處理與相同參數訓練〔In [27], In [28]〕。\n",
"\n",
"| | Wiki control | Forum | Forum − Wiki |\n",
"|---|---|---|---|\n",
"| **Overall** | **41.46** | 25.78 | −15.68 |\n",
"| Semantic | 45.77 | 12.84 | −32.93 |\n",
"| Syntactic | 37.87 | 36.53 | −1.34 |\n",
"| capital-common-countries | 76.48 | 33.60 | −42.88 |\n",
"| capital-world | 54.00 | 12.51 | −41.49 |\n",
"| currency | 6.93 | 1.27 | −5.66 |\n",
"| city-in-state | 33.81 | 2.92 | −30.89 |\n",
"| family | 66.21 | 63.24 | −2.97 |\n",
"| gram1-adjective-to-adverb | 10.28 | 13.10 | +2.82 |\n",
"| gram2-opposite | 8.87 | 14.04 | +5.17 |\n",
"| gram3-comparative | 53.08 | 64.94 | +11.86 |\n",
"| gram4-superlative | 17.20 | 40.55 | +23.35 |\n",
"| gram5-present-participle | 31.82 | 51.33 | +19.51 |\n",
"| gram6-nationality-adjective | 76.55 | 21.70 | −54.85 |\n",
"| gram7-past-tense | 40.96 | 33.78 | −7.18 |\n",
"| gram8-plural | 36.26 | 41.97 | +5.71 |\n",
"| gram9-plural-verbs | 32.99 | 41.49 | +8.50 |\n",
"| OOV 題數 | 169 | 3,510 | |\n",
"\n",
"答案覆蓋率（四個字都在字典裡的比例）：capital-world 98.3% vs **48.6%**、currency 86.8% vs **39.0%**、city-in-state 100% vs 68.3%、family 100% vs 91.3%。\n",
"\n",
"為了排除「論壇字典比較小、很多字查不到」的影響，另外只算**四個字在兩本字典都查得到**的 15,684 題〔In [28]〕：Overall 42.80 vs 32.12、Semantic 52.80 vs 22.10、Syntactic 37.91 vs 37.01；論壇仍在 superlative（42.90 vs 18.28）、present-participle（51.33 vs 31.82）勝出，在 nationality（22.81 vs 77.12）、past-tense（33.78 vs 40.96）落後。\n",
"\n",
"判斷規則：論壇和對照組沒有重跑，所以借用第 5 題 (a) 在 5% 維基量到的各小類雜訊範圍，差距超過該小類的雜訊才判勝負（例如 family −2.97 小於雜訊 4.74，不判勝負；Syntactic −1.34 剛好超過雜訊 1.26：論壇略低，但差距很小）。\n",
"\n",
"3.2. Introduce the corpus you selected and explain what are the differences between the Wikipedia corpus and your corpus. (including data size, topic difference, structural difference … ) (5%)\n",
"\n",
"Answer:\n",
"\n",
"- **來源**：gensim-data 的 `semeval-2016-2017-task3-subtaskA-unannotated`（https://github.com/RaRe-Technologies/gensim-data ；SemEval-2016 Task 3, https://alt.qcri.org/semeval2016/task3/ ）。內容是卡達 Qatar Living 論壇的發問與留言，共 189,941 個討論串〔In [27]〕。只取標題、內文、每則留言三種文字，刪除 HTML 標籤與網址後，用與維基相同的規則斷詞（小寫、2～15 個字母），並刪掉少於 3 個字的段落，最後留下 2,118,254 段文字。\n",
"- **資料量**：72,919,300 字，約是 20% 維基（666,053,126 字）的 1/9，所以另做同字數的維基對照組。字典大小：論壇 100,310 字、對照組 224,050 字〔In [28]〕。\n",
"- **主題**：維基涵蓋歷史、地理、人物、科學；論壇集中在簽證、工作、薪水、租屋、購物等在卡達生活的問題。\n",
"- **文體與結構**：維基是編輯校稿過的長篇說明文，20% 樣本平均每篇約 592 字（666,053,126 / 1,124,733）；論壇是短貼文，平均每段約 34 字（72,919,300 / 2,118,254），口語、問句多、錯字與縮寫多。例如論壇裡 salary 最像的字是 slary、salry、salery、sallary、salaray；visa 最像的是 rp、viza、iqama（卡達居留證）〔In [29]〕。\n",
"\n",
"3.3. Explain why the performance increases or decreases. (5%)\n",
"\n",
"Answer:\n",
"\n",
"在資料量相同的前提下，差異主要來自內容。下表是同一批字在兩份語料中每百萬字的出現次數〔In [30]〕：\n",
"\n",
"| 字 | Wiki control | Forum | Forum / Wiki |\n",
"|---|---|---|---|\n",
"| cheapest | 0.96 | 27.50 | 28.64 |\n",
"| biggest | 37.26 | 59.97 | 1.61 |\n",
"| best | 553.01 | 988.65 | 1.79 |\n",
"| looking | 56.31 | 680.74 | 12.09 |\n",
"| cheaper | 7.78 | 80.27 | 10.32 |\n",
"| going | 104.22 | 914.05 | 8.77 |\n",
"| better | 116.98 | 990.40 | 8.47 |\n",
"| goes | 66.47 | 235.33 | 3.54 |\n",
"| capital | 160.38 | 43.23 | 0.27 |\n",
"| brazilian | 42.43 | 33.46 | 0.79 |\n",
"| albanian | 13.37 | 0.75 | 0.06 |\n",
"| illinois | 90.02 | 1.10 | 0.01 |\n",
"\n",
"- **語意類大幅落後**：論壇很少談世界首都、美國州名（capital 只有維基的 0.27 倍、illinois 0.01 倍）。很多答案根本不在論壇字典裡（capital-world 覆蓋率 48.6%）；而且就算只看兩邊都查得到的題目，論壇的 Semantic 仍只有 22.10 vs 52.80，代表這些字即使出現，也很少出現在「首都／國家」的前後文中，關係學不起來。\n",
"- **比較級、最高級、-ing、第三人稱動詞反而勝出**：網友常寫 cheapest（28.64 倍）、looking（12.09 倍）、better（8.47 倍）、goes（3.54 倍）。（推論）這些字形在論壇中出現得更多、前後文更一致，所以字形關係學得比較好。但頻率不能解釋全部：biggest 只多 1.61 倍、best 1.79 倍，最高級卻大贏 +23.35；brazilian 在論壇也有維基的 0.79 倍，國籍類卻大輸 −54.85。（推論）關鍵可能不只是單字出現次數，而是「國家–國籍」「原級–最高級」這種成對關係有沒有在同樣的前後文裡反覆出現。另一個可能的因素是論壇字典只有對照組的一半大，干擾的候選字比較少；排除查不到的題目後論壇仍在 superlative、present-participle 勝出，但字典大小的影響沒有被完全排除。\n",
"- **錯字變成「最像的字」**：錯字與正確字出現的前後文高度相似，而詞向量是用前後文學出來的，所以 slary 成了 salary 的近鄰。這也呼應投影片 W1 p94 \"Vector search is not robust to typos\"。\n",
"- 結論：**語料的主題與文體決定詞向量學到哪些關係**——想讓某類關係學好，語料裡就要有大量該類的用法（投影片 W1 p93 \"What vectors should I use? It depends.\"）。\n",
"- 限制：第 5 題 (a) 的雜訊是在 5% 維基量的，論壇與對照組沒有重跑；差距小於該小類雜訊的（例如 family −2.97）不下結論。\n",
"\n",
"##### 4. Select a few words and use their embeddings to retrieve the five most similar words. What do you observe? (10%)\n",
"\n",
"Answer:\n",
"\n",
"挑了 7 個字，各有目的：king（考卷中的字）、cat（老師上課的例子）、apple 與 bank（一字多義）、good（看反義詞）、taiwan（專有名詞）、computer（一般名詞）〔In [32], In [33]〕。\n",
"\n",
"| 字 | GloVe | Wiki 20% |\n",
"|---|---|---|\n",
"| king | prince, queen, son, brother, monarch | nangklao, chlothar, queen, suryavarman, bodawpaya |\n",
"| cat | dog, rabbit, cats, monkey, pet | rabbit, dog, sourpuss, mouse, pet |\n",
"| apple | microsoft, ibm, intel, software, dell | blackberry, iphone, goldieblox, tvos, raspberry |\n",
"| bank | banks, banking, credit, investment, financial | guaranty, savings, ameriprise, indymac, onewest |\n",
"| good | better, sure, really, kind, very | sure, decent, lovely, bad, thankful |\n",
"| taiwan | mainland, china, taiwanese, taipei, hong | taipei, china, guangdong, hainan, fujian |\n",
"| computer | computers, software, technology, pc, hardware | computing, computers, software, mainframe, hardware |\n",
"\n",
"觀察：\n",
"1. **冷門字效應**：Wiki 20% 的鄰居常是冷門字。king 的鄰居 nangklao（泰國國王）只出現 34 次、suryavarman 53 次、chlothar 76 次；bank 的 onewest 38 次、ameriprise 49 次〔In [33]〕——都只出現幾十次，在 30 萬字的字典裡屬於最冷門的一群（字典門檻是 30 次）。（推論）一個字只出現幾十次、而且幾乎都在國王或銀行的上下文裡，它的向量就被拉到 king、bank 旁邊。改成只在最常見的 5 萬字裡找，king 的鄰居變成 queen, throne, prince, reigned, ruler；bank 變成 savings, bancorp, banking, lenders, jpmorgan，合理得多。GloVe 的鄰居也比較「一般」。\n",
"2. **反義詞很近**：good 的鄰居有 bad，similarity(good, bad) = 0.70。兩者出現在相似的前後文（\"a ___ movie\"），詞向量分不出正反。\n",
"3. **一字多義被常見意思主導**：apple 前 15 名全是科技相關（blackberry, iphone, tvos, ipad, android…；raspberry 在這裡是 Raspberry Pi），similarity(apple, banana) = 0.41、(apple, fruit) = 0.37，遠低於 (apple, microsoft) = 0.65。bank 前 5 名都是金融義，但 similarity(bank, river) = 0.49 反而高於 (bank, money) = 0.40。（推論）河岸義可能沒有消失，而是和金融義混在同一個向量裡；不過 money 是很泛用的字，這組比較只是旁證。這是「一個字只有一個向量」的靜態詞向量的限制（投影片 W1 p60 正是用 bank 舉例；W2 p39 上下文詞向量才能解決）。\n",
"4. **鄰居反映語料內容**：cat 的鄰居 sourpuss 只出現 64 次。回查訓練語料印出的前 5 處上下文〔In [33]〕，有 2 處是卡通角色名單（\"toon characters including … gandy goose sourpuss dinky duck …\"），其餘是字源說明、電視集名、樂團名。（推論）它只出現 64 次，向量容易被少數特定上下文（例如動物卡通角色名單）決定，因此靠近 cat。\n",
"5. **taiwan 與 computer**：taiwan 在兩個模型的鄰居都是中國的地名與省份（taipei、china、guangdong、fujian），反映語料中的地理與歷史脈絡；computer 的鄰居兩邊都是同類詞（computers、software、hardware），是 7 個字裡最「乖」的一組——一般名詞、用法單一時，詞向量的表現最符合直覺。\n",
"6. 兩個模型的 cosine 分數尺度不同，只比較各自的排名，不拿分數互相比大小。\n",
"\n",
"##### 5. … Anything that can strengthen your report. (5%)\n",
"\n",
"Answer:\n",
"\n",
"**(a) 雜訊有多大**〔In [34]〕：同樣用 5% 維基、只換亂數種子訓練 4 次（seed 42, 1, 2, 3），Overall 在 46.05～47.10 之間（範圍 1.05），Semantic 範圍 1.80、Syntactic 1.26，小類範圍 1.50～4.74（family、comparative、present-participle 最大，約 4.7）。之後所有比較都用這個範圍判斷差距是否可信。\n",
"\n",
"**(b) 移除停用詞**〔In [35]〕：同一份 5% 維基，只差在有沒有移除 NLTK 停用詞。\n",
"\n",
"| | 保留 | 移除 | 差 | 超過雜訊？ |\n",
"|---|---|---|---|---|\n",
"| Overall | 46.12 | 45.25 | −0.87 | 否 |\n",
"| Semantic | 55.35 | 55.41 | +0.06 | 否 |\n",
"| Syntactic | 38.45 | 36.81 | −1.65 | 是 |\n",
"| family | 68.77 | 53.95 | −14.82 | 是 |\n",
"| gram9-plural-verbs | 37.82 | 26.55 | −11.26 | 是 |\n",
"| gram3-comparative | 44.14 | 37.01 | −7.13 | 是 |\n",
"| gram8-plural | 38.59 | 41.89 | +3.30 | 是 |\n",
"| OOV 題數 | 107 | 284 | | |\n",
"\n",
"整體分數的變化在雜訊範圍內，但小類有明顯的此消彼長：family 大跌（he/she/his/her 被刪掉，題目必錯），plural-verbs 與 comparative 也下跌。（推論）停用詞（is、has、more、than…）是判斷字形的線索。只看整體分數會以為「刪不刪都沒差」。另外，`sample=1e-3` 本來就會隨機略過高頻字，效果和部分移除停用詞相似，可能也減小了整體差異。這支持不移除停用詞的決定。\n",
"\n",
"**(c) 超參數（5% 維基，一次只改一個）**〔In [36]〕\n",
"\n",
"| 設定 | Overall | Semantic | Syntactic | Overall 差 |\n",
"|---|---|---|---|---|\n",
"| 基準：Skip-gram, window 5, 100 維 | 46.12 | 55.35 | 38.45 | — |\n",
"| CBOW (`sg=0`) | 52.58 | 58.60 | 47.58 | +6.46 |\n",
"| `window=10` | 41.75 | 51.22 | 33.87 | −4.37 |\n",
"| `vector_size=300` | 57.44 | 68.54 | 48.22 | +11.32 |\n",
"\n",
"三個 Overall 差距都遠大於雜訊 1.05。\n",
"- **300 維**：只用 5% 資料就超過 100 維的 20%（48.39），city-in-state +25.78、capital-common-countries +13.64。在本設定下，向量維度的影響比資料量（5%→20%：+2.27）大。但只做了一次，而且老師提醒維度不是越大越好（「設一萬……效果不會比較好」，W3 01:22:17），也不和 GloVe-300 比較，所以不推論到其他設定。\n",
"- **CBOW**：Overall、Semantic、Syntactic 都較好，語法類大幅領先（comparative +25.00、superlative +20.59）；但 14 個小類裡有 3 類較差，其中 nationality-adjective −1.81 超過雜訊。老師上課以 Skip-gram 為主（W3 01:12:12），原論文（Mikolov et al., 2013, https://arxiv.org/abs/1301.3781 ）也是 Skip-gram 語意類較好；這是「這份資料、這組參數」下的結果。\n",
"- **window=10**：family −18.18、capital-common-countries −9.29。（推論）一般認為較大的視窗偏向「主題相關」而非「用法相似」；但語意類整體也下降，與這個說法不完全吻合，所以只當作可能的解釋。\n",
"\n",
"**(d) Top-k 命中率**〔In [37]〕：Wiki 20% 的 top-1 / 3 / 5 / 10 為 48.39 / 61.53 / 66.32 / 71.74（GloVe：63.11 / 73.68 / 77.73 / 82.01）。正確答案常排在第 2～3 名，呼應老師說的 queen「可能在前三名」。\n",
"\n",
"**(e) 錯題分析**〔In [38]〕：Wiki 20% 答錯 10,087 題，其中 0 題是 OOV，4,563 題（45.2%）的正確答案在第 2～10 名。錯法可以分成幾類：\n",
"- 拼法變體：Argentina → argentinian（考卷答案 argentinean）、Belarus → belarusian（belorussian）、Slovakia → slovak（slovakian）——這些其實是考卷的拼法限制。\n",
"- 方向相反：large → smaller（larger）、long → shorter（longer）——反義詞在向量空間中太近。\n",
"- 同類選錯：Baghdad → syria（iraq）、Armenia → hryvnia（dram）、Cambodia → kyat（riel）。\n",
"- 近親字：father → stepmother（mother）、groom → bridesmaid（bride）。\n",
"- 被其他詞義帶走：great → britain（greater）、child → childbearing（children）。\n",
"\n",
"**(f) 換算答案的公式**〔In [39]〕：同一份 Wiki 20% 向量改用 3CosMul（Levy & Goldberg, 2014, https://aclanthology.org/W14-1618/ ），Overall 48.39 → 44.66，14 個小類全部下降。原因未驗證；（推論）可能與字典中大量冷門字有關（見第 4 題觀察 1）。\n",
"\n",
"**(g) PCA 與 t-SNE**〔In [40]〕：PCA 的縱軸只有微弱的性別傾向：he、his、king 在最上，mom、grandma、aunt 在下，但 she、her、queen 也在上半，dad、grandpa、husband 在下半，分不乾淨；t-SNE 則把每一對（boy/girl, groom/bride, dad/mom, king/queen）貼得很近、分群清楚。PCA 是線性投影，較能保留整體方向；t-SNE 是非線性，主要保留局部鄰居。\n",
"\n",
"**(h) 覆蓋率**〔In [41]〕：GloVe 與 Wiki 20% 在 14 個小類都是 100%，所以 Wiki 20% 的錯誤都不是「查不到字」造成的。\n",
"\n",
"##### 6. Generative AI Usage\n",
"\n",
"Answer:\n",
"\n",
"I used Claude (Anthropic) through Claude Code on my own computer for this assignment.\n",
"- **Code**: All code added to the template was generated by Claude — TODO1–TODO7, the data-loading changes, the code for Question 3, all extra experiments below the report, and the section headings below the report. Every AI-generated code cell has a `# [Generative AI]` comment at the top.\n",
"- **Experiment design**: The choice of the forum corpus, the same-size Wikipedia control, the noise measurement and the other extra experiments were proposed by Claude.\n",
"- **Execution**: Claude ran the whole notebook on my computer (environment above). All experimental numbers and figures in this report come from the outputs of this notebook.\n",
"- **Report**: The report text, including the analysis and tables, was written by Claude based on these outputs.\n",
"- **What I did**: ⟦WHAT_I_DID⟧\n"
]
},
{
"cell_type": "markdown",
"id": "01a559c1",
"metadata": {},
"source": [
"## Code For 3. What is the performance for different categories or sub-categories when trained on different corpora? (15%)"
]
},
{
"cell_type": "code",
"execution_count": 25,
"id": "8691fa4a",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T12:28:55.763691Z",
"iopub.status.busy": "2026-10-05T12:28:55.763691Z",
"iopub.status.idle": "2026-10-05T12:32:57.783812Z",
"shell.execute_reply": "2026-10-05T12:32:57.783812Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"GloVe (pre-trained): overall 63.11% (OOV questions 0)\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 20%: overall 48.39% (OOV questions 0)\n"
]
}
],
"source": [
"# Write your code here\n",
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 共用工具：用任何一份詞向量考一次試。\n",
"# 每題一次拿前 10 名候選字，記下正確答案排第幾名（之後算 top-k、做錯題分析都用得到）。\n",
"results, details, models = {}, {}, {}\n",
"\n",
"def evaluate(wv, name, method=\"add\", max_k=10):\n",
" sim = wv.most_similar if method == \"add\" else wv.most_similar_cosmul\n",
" ranks, preds, oov = [], [], 0\n",
" for q in data[\"Question\"]:\n",
" a, b, c, d = q.lower().split()\n",
" if all(w in wv.key_to_index for w in (a, b, c)):\n",
" cands = [w for w, _ in sim(positive=[b, c], negative=[a], topn=max_k)]\n",
" preds.append(cands[0])\n",
" ranks.append(cands.index(d) + 1 if d in cands else None)\n",
" else:\n",
" preds.append(None)\n",
" ranks.append(None)\n",
" oov += 1\n",
" hits = np.array([r == 1 for r in ranks])\n",
" res = {\"Overall\": hits.mean() * 100}\n",
" for col in [\"Category\", \"SubCategory\"]:\n",
" for v in data[col].unique():\n",
" res[v] = hits[(data[col] == v).values].mean() * 100\n",
" res[\"OOV questions\"] = oov\n",
" results[name], details[name] = res, {\"ranks\": ranks, \"preds\": preds}\n",
" print(f\"{name}: overall {res['Overall']:.2f}% (OOV questions {oov})\")\n",
" return res\n",
"\n",
"def answer_coverage(wv):\n",
" # 每個小類裡，四個字全部都在字典裡的題目佔幾 %（答案不在字典裡，再聰明也答不對）\n",
" ok = data[\"Question\"].map(lambda q: all(w in wv.key_to_index for w in q.lower().split()))\n",
" return ok.groupby(data[\"SubCategory\"], sort=False).mean().mul(100).round(1)\n",
"\n",
"models[\"GloVe (pre-trained)\"], models[\"Wiki 20%\"] = model, my_wv\n",
"_ = evaluate(model, \"GloVe (pre-trained)\")\n",
"_ = evaluate(my_wv, \"Wiki 20%\")\n"
]
},
{
"cell_type": "code",
"execution_count": 26,
"id": "33d80533",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T12:32:57.846506Z",
"iopub.status.busy": "2026-10-05T12:32:57.846506Z",
"iopub.status.idle": "2026-10-05T12:32:58.486234Z",
"shell.execute_reply": "2026-10-05T12:32:58.486234Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"keys: ['RelComments', 'RelQuestion', 'THREAD_SEQUENCE'] | question keys: ['RELQ_DATE', 'RELQ_ID', 'RelQSubject', 'RELQ_USERNAME', 'RELQ_USERID', 'RelQBody', 'RELQ_CATEGORY']\n",
"subject: Thailand:IT Minsitry blocks CNN; Facebook;\n",
"first comment: have they blocked porn??? ![](\"http://www.qatarliving.com/files/images/Da.gif\")
\n"
]
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 3 題：換一份「不是維基百科」的語料。\n",
"# 選的是 SemEval 2016/2017 Task 3 的論壇問答：卡達生活論壇（Qatar Living）網友的發問與留言。\n",
"# 出處：gensim-data https://github.com/RaRe-Technologies/gensim-data （資料集 semeval-2016-2017-task3-subtaskA-unannotated）\n",
"# SemEval-2016 Task 3: Community Question Answering, https://alt.qcri.org/semeval2016/task3/\n",
"# 注意：gensim 幫這份資料附的讀檔程式用了舊版 smart_open 的寫法，新版會 ImportError。\n",
"# 所以只用 gensim 下載（return_path=True 只回傳檔案位置），讀檔自己來：每一行是一個討論串的 JSON。\n",
"import gensim.downloader, itertools, json\n",
"from gensim.utils import simple_preprocess\n",
"\n",
"FORUM_PATH = gensim.downloader.load(\"semeval-2016-2017-task3-subtaskA-unannotated\", return_path=True)\n",
"\n",
"def iter_threads():\n",
" with gzip.open(FORUM_PATH, \"rt\", encoding=\"utf-8\") as f:\n",
" for line in f:\n",
" yield json.loads(line)\n",
"\n",
"def thread_texts(thread):\n",
" # 只拿真正的文字：發問標題、發問內容、每一則留言。使用者名稱、日期、編號都不要。\n",
" q = thread.get(\"RelQuestion\", {})\n",
" yield q.get(\"RelQSubject\") or \"\"\n",
" yield q.get(\"RelQBody\") or \"\"\n",
" for c in thread.get(\"RelComments\", []):\n",
" yield c.get(\"RelCText\") or \"\"\n",
"\n",
"first = next(iter_threads())\n",
"print(\"keys:\", list(first.keys()), \"| question keys:\", list(first[\"RelQuestion\"].keys()))\n",
"print(\"subject:\", first[\"RelQuestion\"].get(\"RelQSubject\"))\n",
"print(\"first comment:\", first[\"RelComments\"][0][\"RelCText\"][:200])\n"
]
},
{
"cell_type": "code",
"execution_count": 27,
"id": "e8b757af",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T12:32:58.546655Z",
"iopub.status.busy": "2026-10-05T12:32:58.546655Z",
"iopub.status.idle": "2026-10-05T12:35:14.673546Z",
"shell.execute_reply": "2026-10-05T12:35:14.673546Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"forum: 189,941 threads, 2,118,254 posts, 72,919,300 tokens\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"wiki_sampled_5.txt: removed non-[a-z] tokens 1,211,749 of 167,406,150 (0.72%)\n",
"5% wiki tokens: 166,194,401\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"wiki control: 72,920,923 tokens\n"
]
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 用跟維基百科一樣的規則前處理：先拿掉網頁語法和網址，再小寫、只留 2～15 個字母的英文字（WikiCorpus 的預設）。\n",
"HTML_URL_RE = re.compile(r\"<[^>]+>|https?://\\S+|www\\.\\S+\")\n",
"forum_tokens = forum_lines = n_threads = 0\n",
"with open(\"forum_clean.txt\", \"w\", encoding=\"utf-8\", newline=\"\\n\") as fout:\n",
" for thread in iter_threads():\n",
" n_threads += 1\n",
" for text in thread_texts(thread):\n",
" text = HTML_URL_RE.sub(\" \", text)\n",
" toks = [w for w in simple_preprocess(text, min_len=2, max_len=15) if TOKEN_RE.match(w)]\n",
" if len(toks) >= 3:\n",
" fout.write(\" \".join(toks) + \"\\n\")\n",
" forum_tokens += len(toks)\n",
" forum_lines += 1\n",
"print(f\"forum: {n_threads:,} threads, {forum_lines:,} posts, {forum_tokens:,} tokens\")\n",
"\n",
"# 對照組：從維基 5% 裡依序拿文章，拿到「字數跟論壇一樣多」就停。\n",
"# 這樣兩份資料一樣大，分數的差別就只剩「內容種類」造成的。\n",
"n5 = preprocess_file(\"wiki_sampled_5.txt\", \"wiki_sampled_5_clean.txt\")\n",
"print(f\"5% wiki tokens: {n5:,}\")\n",
"ctrl_tokens = 0\n",
"with open(\"wiki_sampled_5_clean.txt\", encoding=\"utf-8\") as fin, open(\"wiki_control_clean.txt\", \"w\", encoding=\"utf-8\", newline=\"\\n\") as fout:\n",
" for line in fin:\n",
" if ctrl_tokens >= forum_tokens:\n",
" break\n",
" fout.write(line)\n",
" ctrl_tokens += len(line.split())\n",
"print(f\"wiki control: {ctrl_tokens:,} tokens\")\n"
]
},
{
"cell_type": "code",
"execution_count": 28,
"id": "4c0ba87d",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T12:35:14.732911Z",
"iopub.status.busy": "2026-10-05T12:35:14.732911Z",
"iopub.status.idle": "2026-10-05T12:44:28.567664Z",
"shell.execute_reply": "2026-10-05T12:44:28.567131Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"trained on forum_clean.txt: 3.9 min, vocab 100,310, effective min_count 5\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"trained on wiki_control_clean.txt: 3.9 min, vocab 224,050, effective min_count 5\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Forum (same size): overall 25.78% (OOV questions 3510)\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki control (same size): overall 41.46% (OOV questions 169)\n"
]
},
{
"data": {
"text/html": [
"
\n"
]
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 3 題：換一份「不是維基百科」的語料。\n",
"# 選的是 SemEval 2016/2017 Task 3 的論壇問答：卡達生活論壇（Qatar Living）網友的發問與留言。\n",
"# 出處：gensim-data https://github.com/RaRe-Technologies/gensim-data （資料集 semeval-2016-2017-task3-subtaskA-unannotated）\n",
"# SemEval-2016 Task 3: Community Question Answering, https://alt.qcri.org/semeval2016/task3/\n",
"# 注意：gensim 幫這份資料附的讀檔程式用了舊版 smart_open 的寫法，新版會 ImportError。\n",
"# 所以只用 gensim 下載（return_path=True 只回傳檔案位置），讀檔自己來：每一行是一個討論串的 JSON。\n",
"import gensim.downloader, itertools, json\n",
"from gensim.utils import simple_preprocess\n",
"\n",
"FORUM_PATH = gensim.downloader.load(\"semeval-2016-2017-task3-subtaskA-unannotated\", return_path=True)\n",
"\n",
"def iter_threads():\n",
" with gzip.open(FORUM_PATH, \"rt\", encoding=\"utf-8\") as f:\n",
" for line in f:\n",
" yield json.loads(line)\n",
"\n",
"def thread_texts(thread):\n",
" # 只拿真正的文字：發問標題、發問內容、每一則留言。使用者名稱、日期、編號都不要。\n",
" q = thread.get(\"RelQuestion\", {})\n",
" yield q.get(\"RelQSubject\") or \"\"\n",
" yield q.get(\"RelQBody\") or \"\"\n",
" for c in thread.get(\"RelComments\", []):\n",
" yield c.get(\"RelCText\") or \"\"\n",
"\n",
"first = next(iter_threads())\n",
"print(\"keys:\", list(first.keys()), \"| question keys:\", list(first[\"RelQuestion\"].keys()))\n",
"print(\"subject:\", first[\"RelQuestion\"].get(\"RelQSubject\"))\n",
"print(\"first comment:\", first[\"RelComments\"][0][\"RelCText\"][:200])\n"
]
},
{
"cell_type": "code",
"execution_count": 27,
"id": "e8b757af",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T12:32:58.546655Z",
"iopub.status.busy": "2026-10-05T12:32:58.546655Z",
"iopub.status.idle": "2026-10-05T12:35:14.673546Z",
"shell.execute_reply": "2026-10-05T12:35:14.673546Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"forum: 189,941 threads, 2,118,254 posts, 72,919,300 tokens\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"wiki_sampled_5.txt: removed non-[a-z] tokens 1,211,749 of 167,406,150 (0.72%)\n",
"5% wiki tokens: 166,194,401\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"wiki control: 72,920,923 tokens\n"
]
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 用跟維基百科一樣的規則前處理：先拿掉網頁語法和網址，再小寫、只留 2～15 個字母的英文字（WikiCorpus 的預設）。\n",
"HTML_URL_RE = re.compile(r\"<[^>]+>|https?://\\S+|www\\.\\S+\")\n",
"forum_tokens = forum_lines = n_threads = 0\n",
"with open(\"forum_clean.txt\", \"w\", encoding=\"utf-8\", newline=\"\\n\") as fout:\n",
" for thread in iter_threads():\n",
" n_threads += 1\n",
" for text in thread_texts(thread):\n",
" text = HTML_URL_RE.sub(\" \", text)\n",
" toks = [w for w in simple_preprocess(text, min_len=2, max_len=15) if TOKEN_RE.match(w)]\n",
" if len(toks) >= 3:\n",
" fout.write(\" \".join(toks) + \"\\n\")\n",
" forum_tokens += len(toks)\n",
" forum_lines += 1\n",
"print(f\"forum: {n_threads:,} threads, {forum_lines:,} posts, {forum_tokens:,} tokens\")\n",
"\n",
"# 對照組：從維基 5% 裡依序拿文章，拿到「字數跟論壇一樣多」就停。\n",
"# 這樣兩份資料一樣大，分數的差別就只剩「內容種類」造成的。\n",
"n5 = preprocess_file(\"wiki_sampled_5.txt\", \"wiki_sampled_5_clean.txt\")\n",
"print(f\"5% wiki tokens: {n5:,}\")\n",
"ctrl_tokens = 0\n",
"with open(\"wiki_sampled_5_clean.txt\", encoding=\"utf-8\") as fin, open(\"wiki_control_clean.txt\", \"w\", encoding=\"utf-8\", newline=\"\\n\") as fout:\n",
" for line in fin:\n",
" if ctrl_tokens >= forum_tokens:\n",
" break\n",
" fout.write(line)\n",
" ctrl_tokens += len(line.split())\n",
"print(f\"wiki control: {ctrl_tokens:,} tokens\")\n"
]
},
{
"cell_type": "code",
"execution_count": 28,
"id": "4c0ba87d",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T12:35:14.732911Z",
"iopub.status.busy": "2026-10-05T12:35:14.732911Z",
"iopub.status.idle": "2026-10-05T12:44:28.567664Z",
"shell.execute_reply": "2026-10-05T12:44:28.567131Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"trained on forum_clean.txt: 3.9 min, vocab 100,310, effective min_count 5\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"trained on wiki_control_clean.txt: 3.9 min, vocab 224,050, effective min_count 5\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Forum (same size): overall 25.78% (OOV questions 3510)\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki control (same size): overall 41.46% (OOV questions 169)\n"
]
},
{
"data": {
"text/html": [
"\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Wiki control (same size) Forum (same size) \\\n",
"Overall 41.46 25.78 \n",
"Semantic 45.77 12.84 \n",
"Syntactic 37.87 36.53 \n",
": capital-common-countries 76.48 33.60 \n",
": capital-world 54.00 12.51 \n",
": currency 6.93 1.27 \n",
": city-in-state 33.81 2.92 \n",
": family 66.21 63.24 \n",
": gram1-adjective-to-adverb 10.28 13.10 \n",
": gram2-opposite 8.87 14.04 \n",
": gram3-comparative 53.08 64.94 \n",
": gram4-superlative 17.20 40.55 \n",
": gram5-present-participle 31.82 51.33 \n",
": gram6-nationality-adjective 76.55 21.70 \n",
": gram7-past-tense 40.96 33.78 \n",
": gram8-plural 36.26 41.97 \n",
": gram9-plural-verbs 32.99 41.49 \n",
"OOV questions 169.00 3510.00 \n",
"\n",
" Wiki 20% \n",
"Overall 48.39 \n",
"Semantic 56.70 \n",
"Syntactic 41.48 \n",
": capital-common-countries 77.87 \n",
": capital-world 72.81 \n",
": currency 10.28 \n",
": city-in-state 35.51 \n",
": family 74.31 \n",
": gram1-adjective-to-adverb 15.93 \n",
": gram2-opposite 16.01 \n",
": gram3-comparative 49.10 \n",
": gram4-superlative 24.33 \n",
": gram5-present-participle 31.34 \n",
": gram6-nationality-adjective 80.49 \n",
": gram7-past-tense 43.85 \n",
": gram8-plural 41.67 \n",
": gram9-plural-verbs 40.92 \n",
"OOV questions 0.00 "
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"text/html": [
"| \n", " | Wiki control (same size) | \n", "Forum (same size) | \n", "Wiki 20% | \n", "
|---|---|---|---|
| Overall | \n", "41.46 | \n", "25.78 | \n", "48.39 | \n", "
| Semantic | \n", "45.77 | \n", "12.84 | \n", "56.70 | \n", "
| Syntactic | \n", "37.87 | \n", "36.53 | \n", "41.48 | \n", "
| : capital-common-countries | \n", "76.48 | \n", "33.60 | \n", "77.87 | \n", "
| : capital-world | \n", "54.00 | \n", "12.51 | \n", "72.81 | \n", "
| : currency | \n", "6.93 | \n", "1.27 | \n", "10.28 | \n", "
| : city-in-state | \n", "33.81 | \n", "2.92 | \n", "35.51 | \n", "
| : family | \n", "66.21 | \n", "63.24 | \n", "74.31 | \n", "
| : gram1-adjective-to-adverb | \n", "10.28 | \n", "13.10 | \n", "15.93 | \n", "
| : gram2-opposite | \n", "8.87 | \n", "14.04 | \n", "16.01 | \n", "
| : gram3-comparative | \n", "53.08 | \n", "64.94 | \n", "49.10 | \n", "
| : gram4-superlative | \n", "17.20 | \n", "40.55 | \n", "24.33 | \n", "
| : gram5-present-participle | \n", "31.82 | \n", "51.33 | \n", "31.34 | \n", "
| : gram6-nationality-adjective | \n", "76.55 | \n", "21.70 | \n", "80.49 | \n", "
| : gram7-past-tense | \n", "40.96 | \n", "33.78 | \n", "43.85 | \n", "
| : gram8-plural | \n", "36.26 | \n", "41.97 | \n", "41.67 | \n", "
| : gram9-plural-verbs | \n", "32.99 | \n", "41.49 | \n", "40.92 | \n", "
| OOV questions | \n", "169.00 | \n", "3510.00 | \n", "0.00 | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" coverage: wiki control coverage: forum\n",
"SubCategory \n",
": capital-common-countries 100.0 100.0\n",
": capital-world 98.3 48.6\n",
": currency 86.8 39.0\n",
": city-in-state 100.0 68.3\n",
": family 100.0 91.3\n",
": gram1-adjective-to-adverb 100.0 100.0\n",
": gram2-opposite 100.0 100.0\n",
": gram3-comparative 100.0 100.0\n",
": gram4-superlative 94.1 100.0\n",
": gram5-present-participle 100.0 100.0\n",
": gram6-nationality-adjective 100.0 95.1\n",
": gram7-past-tense 100.0 100.0\n",
": gram8-plural 100.0 100.0\n",
": gram9-plural-verbs 100.0 100.0"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Accuracy on questions whose 4 words are in BOTH vocabularies:\n"
]
},
{
"data": {
"text/html": [
"| \n", " | coverage: wiki control | \n", "coverage: forum | \n", "
|---|---|---|
| SubCategory | \n", "\n", " | \n", " |
| : capital-common-countries | \n", "100.0 | \n", "100.0 | \n", "
| : capital-world | \n", "98.3 | \n", "48.6 | \n", "
| : currency | \n", "86.8 | \n", "39.0 | \n", "
| : city-in-state | \n", "100.0 | \n", "68.3 | \n", "
| : family | \n", "100.0 | \n", "91.3 | \n", "
| : gram1-adjective-to-adverb | \n", "100.0 | \n", "100.0 | \n", "
| : gram2-opposite | \n", "100.0 | \n", "100.0 | \n", "
| : gram3-comparative | \n", "100.0 | \n", "100.0 | \n", "
| : gram4-superlative | \n", "94.1 | \n", "100.0 | \n", "
| : gram5-present-participle | \n", "100.0 | \n", "100.0 | \n", "
| : gram6-nationality-adjective | \n", "100.0 | \n", "95.1 | \n", "
| : gram7-past-tense | \n", "100.0 | \n", "100.0 | \n", "
| : gram8-plural | \n", "100.0 | \n", "100.0 | \n", "
| : gram9-plural-verbs | \n", "100.0 | \n", "100.0 | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Wiki control (same size) Forum (same size)\n",
"questions used 15684.00 15684.00\n",
"Overall 42.80 32.12\n",
"Semantic 52.80 22.10\n",
"Syntactic 37.91 37.01\n",
": capital-common-countries 76.48 33.60\n",
": capital-world 62.03 25.74\n",
": currency 11.26 3.64\n",
": city-in-state 36.34 4.28\n",
": family 70.13 69.26\n",
": gram1-adjective-to-adverb 10.28 13.10\n",
": gram2-opposite 8.87 14.04\n",
": gram3-comparative 53.08 64.94\n",
": gram4-superlative 18.28 42.90\n",
": gram5-present-participle 31.82 51.33\n",
": gram6-nationality-adjective 77.12 22.81\n",
": gram7-past-tense 40.96 33.78\n",
": gram8-plural 36.26 41.97\n",
": gram9-plural-verbs 32.99 41.49"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"forum_model = train_w2v(\"forum_clean.txt\")\n",
"ctrl_model = train_w2v(\"wiki_control_clean.txt\")\n",
"models[\"Forum (same size)\"], models[\"Wiki control (same size)\"] = forum_model.wv, ctrl_model.wv\n",
"_ = evaluate(forum_model.wv, \"Forum (same size)\")\n",
"_ = evaluate(ctrl_model.wv, \"Wiki control (same size)\")\n",
"\n",
"cols = [\"Wiki control (same size)\", \"Forum (same size)\", \"Wiki 20%\"]\n",
"display(pd.DataFrame({k: results[k] for k in cols}).round(2))\n",
"# 答案覆蓋率：題目四個字都在字典裡的比例。覆蓋率低 = 很多題連答的機會都沒有。\n",
"display(pd.DataFrame({\"coverage: wiki control\": answer_coverage(ctrl_model.wv),\n",
" \"coverage: forum\": answer_coverage(forum_model.wv)}))\n",
"\n",
"# 公平一點：只算「四個字在兩本字典裡都查得到」的題目，排除「字典裡沒有這個字」的影響。\n",
"both = data[\"Question\"].map(lambda q: all(w in forum_model.wv.key_to_index and w in ctrl_model.wv.key_to_index\n",
" for w in q.lower().split())).values\n",
"fair = {}\n",
"for name in [\"Wiki control (same size)\", \"Forum (same size)\"]:\n",
" hits = np.array([r == 1 for r in details[name][\"ranks\"]])\n",
" row = {\"questions used\": int(both.sum()), \"Overall\": hits[both].mean() * 100}\n",
" for col in [\"Category\", \"SubCategory\"]:\n",
" for v in data[col].unique():\n",
" m = (data[col] == v).values & both\n",
" row[v] = hits[m].mean() * 100 if m.sum() else float(\"nan\")\n",
" fair[name] = row\n",
"print(\"Accuracy on questions whose 4 words are in BOTH vocabularies:\")\n",
"display(pd.DataFrame(fair).round(2))\n"
]
},
{
"cell_type": "code",
"execution_count": 29,
"id": "58d7012b",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T12:44:28.631411Z",
"iopub.status.busy": "2026-10-05T12:44:28.631411Z",
"iopub.status.idle": "2026-10-05T12:44:28.676426Z",
"shell.execute_reply": "2026-10-05T12:44:28.675914Z"
}
},
"outputs": [
{
"data": {
"text/html": [
"| \n", " | Wiki control (same size) | \n", "Forum (same size) | \n", "
|---|---|---|
| questions used | \n", "15684.00 | \n", "15684.00 | \n", "
| Overall | \n", "42.80 | \n", "32.12 | \n", "
| Semantic | \n", "52.80 | \n", "22.10 | \n", "
| Syntactic | \n", "37.91 | \n", "37.01 | \n", "
| : capital-common-countries | \n", "76.48 | \n", "33.60 | \n", "
| : capital-world | \n", "62.03 | \n", "25.74 | \n", "
| : currency | \n", "11.26 | \n", "3.64 | \n", "
| : city-in-state | \n", "36.34 | \n", "4.28 | \n", "
| : family | \n", "70.13 | \n", "69.26 | \n", "
| : gram1-adjective-to-adverb | \n", "10.28 | \n", "13.10 | \n", "
| : gram2-opposite | \n", "8.87 | \n", "14.04 | \n", "
| : gram3-comparative | \n", "53.08 | \n", "64.94 | \n", "
| : gram4-superlative | \n", "18.28 | \n", "42.90 | \n", "
| : gram5-present-participle | \n", "31.82 | \n", "51.33 | \n", "
| : gram6-nationality-adjective | \n", "77.12 | \n", "22.81 | \n", "
| : gram7-past-tense | \n", "40.96 | \n", "33.78 | \n", "
| : gram8-plural | \n", "36.26 | \n", "41.97 | \n", "
| : gram9-plural-verbs | \n", "32.99 | \n", "41.49 | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" word corpus \\\n",
"0 visa Wiki control (same size) \n",
"1 visa Forum (same size) \n",
"2 salary Wiki control (same size) \n",
"3 salary Forum (same size) \n",
"4 doha Wiki control (same size) \n",
"5 doha Forum (same size) \n",
"6 family Wiki control (same size) \n",
"7 family Forum (same size) \n",
"8 king Wiki control (same size) \n",
"9 king Forum (same size) \n",
"10 paris Wiki control (same size) \n",
"11 paris Forum (same size) \n",
"12 computer Wiki control (same size) \n",
"13 computer Forum (same size) \n",
"\n",
" top 5 \n",
"0 visas, passport, bdtc, passports, zwartendijk \n",
"1 rp, viza, iqama, residance, vissa \n",
"2 salaries, remuneration, wages, repayment, paid \n",
"3 slary, salry, salery, sallary, salaray \n",
"4 qatar, manama, abbasiyyin, dhabi, dubai \n",
"5 qatar, doah, qata, qatat, qutar \n",
"6 father, relatives, phyllanthaceae, grandparents, parents \n",
"7 familly, fmaily, rescidence, pernament, husbant \n",
"8 queen, prince, monarch, harthacnut, eystein \n",
"9 queen, ypsilanti, edsel, bhumibol, edshel \n",
"10 fontainebleau, marseille, bercy, brussels, reims \n",
"11 amsterdam, cancun, france, bangkok, london \n",
"12 computers, software, mainframe, hardware, computing \n",
"13 soundcard, desktop, pc, computers, hardware "
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 同一個字，在論壇和維基（一樣大小）學到的鄰居差在哪\n",
"rows = []\n",
"for w in [\"visa\", \"salary\", \"doha\", \"family\", \"king\", \"paris\", \"computer\"]:\n",
" for name in [\"Wiki control (same size)\", \"Forum (same size)\"]:\n",
" wv = models[name]\n",
" rows.append({\"word\": w, \"corpus\": name,\n",
" \"top 5\": \", \".join(x for x, _ in wv.most_similar(w, topn=5)) if w in wv.key_to_index else \"(not in vocab)\"})\n",
"pd.set_option(\"display.max_colwidth\", 200)\n",
"display(pd.DataFrame(rows))\n"
]
},
{
"cell_type": "code",
"execution_count": 30,
"id": "b14f1ca1",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T12:44:28.741167Z",
"iopub.status.busy": "2026-10-05T12:44:28.741167Z",
"iopub.status.idle": "2026-10-05T12:44:43.239542Z",
"shell.execute_reply": "2026-10-05T12:44:43.239542Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki control (per 1M): 72,920,923 tokens\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Forum (per 1M): 72,919,300 tokens\n"
]
},
{
"data": {
"text/html": [
"| \n", " | word | \n", "corpus | \n", "top 5 | \n", "
|---|---|---|---|
| 0 | \n", "visa | \n", "Wiki control (same size) | \n", "visas, passport, bdtc, passports, zwartendijk | \n", "
| 1 | \n", "visa | \n", "Forum (same size) | \n", "rp, viza, iqama, residance, vissa | \n", "
| 2 | \n", "salary | \n", "Wiki control (same size) | \n", "salaries, remuneration, wages, repayment, paid | \n", "
| 3 | \n", "salary | \n", "Forum (same size) | \n", "slary, salry, salery, sallary, salaray | \n", "
| 4 | \n", "doha | \n", "Wiki control (same size) | \n", "qatar, manama, abbasiyyin, dhabi, dubai | \n", "
| 5 | \n", "doha | \n", "Forum (same size) | \n", "qatar, doah, qata, qatat, qutar | \n", "
| 6 | \n", "family | \n", "Wiki control (same size) | \n", "father, relatives, phyllanthaceae, grandparents, parents | \n", "
| 7 | \n", "family | \n", "Forum (same size) | \n", "familly, fmaily, rescidence, pernament, husbant | \n", "
| 8 | \n", "king | \n", "Wiki control (same size) | \n", "queen, prince, monarch, harthacnut, eystein | \n", "
| 9 | \n", "king | \n", "Forum (same size) | \n", "queen, ypsilanti, edsel, bhumibol, edshel | \n", "
| 10 | \n", "paris | \n", "Wiki control (same size) | \n", "fontainebleau, marseille, bercy, brussels, reims | \n", "
| 11 | \n", "paris | \n", "Forum (same size) | \n", "amsterdam, cancun, france, bangkok, london | \n", "
| 12 | \n", "computer | \n", "Wiki control (same size) | \n", "computers, software, mainframe, hardware, computing | \n", "
| 13 | \n", "computer | \n", "Forum (same size) | \n", "soundcard, desktop, pc, computers, hardware | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Wiki control (per 1M) Forum (per 1M) forum / wiki\n",
"better 116.98 990.40 8.47\n",
"best 553.01 988.65 1.79\n",
"cheaper 7.78 80.27 10.32\n",
"cheapest 0.96 27.50 28.64\n",
"bigger 13.12 53.58 4.08\n",
"biggest 37.26 59.97 1.61\n",
"looking 56.31 680.74 12.09\n",
"going 104.22 914.05 8.77\n",
"working 187.26 583.48 3.12\n",
"costs 43.99 77.13 1.75\n",
"goes 66.47 235.33 3.54\n",
"wife 212.38 605.50 2.85\n",
"husband 104.85 418.87 3.99\n",
"he 5122.78 3669.63 0.72\n",
"she 1266.80 1885.25 1.49\n",
"capital 160.38 43.23 0.27\n",
"illinois 90.02 1.10 0.01\n",
"albanian 13.37 0.75 0.06\n",
"brazilian 42.43 33.46 0.79\n",
"kwanza 0.19 0.01 0.07"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 3 題補充證據：同樣的字，在兩份一樣大的教材裡各出現幾次（每 100 萬字）。\n",
"from collections import Counter\n",
"\n",
"probe_words = [\"better\", \"best\", \"cheaper\", \"cheapest\", \"bigger\", \"biggest\", \"looking\", \"going\",\n",
" \"working\", \"costs\", \"goes\", \"wife\", \"husband\", \"he\", \"she\",\n",
" \"capital\", \"illinois\", \"albanian\", \"brazilian\", \"kwanza\"]\n",
"wanted = set(probe_words)\n",
"freq = {}\n",
"for name, path in [(\"Wiki control (per 1M)\", \"wiki_control_clean.txt\"), (\"Forum (per 1M)\", \"forum_clean.txt\")]:\n",
" counts, total = Counter(), 0\n",
" with open(path, encoding=\"utf-8\") as f:\n",
" for line in f:\n",
" toks = line.split()\n",
" total += len(toks)\n",
" counts.update(t for t in toks if t in wanted)\n",
" print(f\"{name}: {total:,} tokens\")\n",
" freq[name] = {w: counts[w] / total * 1e6 for w in probe_words}\n",
"table = pd.DataFrame(freq)\n",
"table[\"forum / wiki\"] = table[\"Forum (per 1M)\"] / table[\"Wiki control (per 1M)\"]\n",
"display(table.round(2))\n"
]
},
{
"cell_type": "markdown",
"id": "ead0f6de",
"metadata": {},
"source": [
"## Extra code for report question 2 (5% / 10% / 20%)"
]
},
{
"cell_type": "code",
"execution_count": 31,
"id": "dfee9c91",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T12:44:43.301335Z",
"iopub.status.busy": "2026-10-05T12:44:43.301335Z",
"iopub.status.idle": "2026-10-05T13:18:34.046319Z",
"shell.execute_reply": "2026-10-05T13:18:34.045278Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"wiki_sampled_10.txt: removed non-[a-z] tokens 2,417,608 of 335,470,319 (0.72%)\n",
"10% wiki tokens: 333,052,711\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"trained on wiki_sampled_5_clean.txt: 9.1 min, vocab 278,954, effective min_count 9\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 5%: overall 46.12% (OOV questions 107)\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"trained on wiki_sampled_10_clean.txt: 19.7 min, vocab 294,612, effective min_count 16\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 10%: overall 47.15% (OOV questions 0)\n"
]
},
{
"data": {
"text/html": [
"| \n", " | Wiki control (per 1M) | \n", "Forum (per 1M) | \n", "forum / wiki | \n", "
|---|---|---|---|
| better | \n", "116.98 | \n", "990.40 | \n", "8.47 | \n", "
| best | \n", "553.01 | \n", "988.65 | \n", "1.79 | \n", "
| cheaper | \n", "7.78 | \n", "80.27 | \n", "10.32 | \n", "
| cheapest | \n", "0.96 | \n", "27.50 | \n", "28.64 | \n", "
| bigger | \n", "13.12 | \n", "53.58 | \n", "4.08 | \n", "
| biggest | \n", "37.26 | \n", "59.97 | \n", "1.61 | \n", "
| looking | \n", "56.31 | \n", "680.74 | \n", "12.09 | \n", "
| going | \n", "104.22 | \n", "914.05 | \n", "8.77 | \n", "
| working | \n", "187.26 | \n", "583.48 | \n", "3.12 | \n", "
| costs | \n", "43.99 | \n", "77.13 | \n", "1.75 | \n", "
| goes | \n", "66.47 | \n", "235.33 | \n", "3.54 | \n", "
| wife | \n", "212.38 | \n", "605.50 | \n", "2.85 | \n", "
| husband | \n", "104.85 | \n", "418.87 | \n", "3.99 | \n", "
| he | \n", "5122.78 | \n", "3669.63 | \n", "0.72 | \n", "
| she | \n", "1266.80 | \n", "1885.25 | \n", "1.49 | \n", "
| capital | \n", "160.38 | \n", "43.23 | \n", "0.27 | \n", "
| illinois | \n", "90.02 | \n", "1.10 | \n", "0.01 | \n", "
| albanian | \n", "13.37 | \n", "0.75 | \n", "0.06 | \n", "
| brazilian | \n", "42.43 | \n", "33.46 | \n", "0.79 | \n", "
| kwanza | \n", "0.19 | \n", "0.01 | \n", "0.07 | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Wiki 5% Wiki 10% Wiki 20%\n",
"Overall 46.12 47.15 48.39\n",
"Semantic 55.35 55.60 56.70\n",
"Syntactic 38.45 40.13 41.48\n",
": capital-common-countries 81.82 76.48 77.87\n",
": capital-world 69.83 71.07 72.81\n",
": currency 7.62 9.58 10.28\n",
": city-in-state 37.37 35.47 35.51\n",
": family 68.77 73.32 74.31\n",
": gram1-adjective-to-adverb 12.70 10.99 15.93\n",
": gram2-opposite 11.82 14.29 16.01\n",
": gram3-comparative 44.14 46.92 49.10\n",
": gram4-superlative 17.02 22.37 24.33\n",
": gram5-present-participle 29.92 34.28 31.34\n",
": gram6-nationality-adjective 78.42 76.11 80.49\n",
": gram7-past-tense 44.29 44.87 43.85\n",
": gram8-plural 38.59 40.77 41.67\n",
": gram9-plural-verbs 37.82 41.49 40.92\n",
"OOV questions 107.00 0.00 0.00"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 2 題：5%、10%、20%（助教說這題不用附程式，這裡附上供參考）\n",
"n10 = preprocess_file(\"wiki_sampled_10.txt\", \"wiki_sampled_10_clean.txt\")\n",
"print(f\"10% wiki tokens: {n10:,}\")\n",
"for pct in [5, 10]:\n",
" m = train_w2v(f\"wiki_sampled_{pct}_clean.txt\")\n",
" models[f\"Wiki {pct}%\"] = m.wv\n",
" _ = evaluate(m.wv, f\"Wiki {pct}%\")\n",
"display(pd.DataFrame({k: results[k] for k in [\"Wiki 5%\", \"Wiki 10%\", \"Wiki 20%\"]}).round(2))\n"
]
},
{
"cell_type": "markdown",
"id": "7f4e5782",
"metadata": {},
"source": [
"## Extra code for report question 4 (most similar words)"
]
},
{
"cell_type": "code",
"execution_count": 32,
"id": "820e2535",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T13:18:34.118103Z",
"iopub.status.busy": "2026-10-05T13:18:34.118103Z",
"iopub.status.idle": "2026-10-05T13:18:34.281834Z",
"shell.execute_reply": "2026-10-05T13:18:34.281834Z"
}
},
"outputs": [
{
"data": {
"text/html": [
"| \n", " | Wiki 5% | \n", "Wiki 10% | \n", "Wiki 20% | \n", "
|---|---|---|---|
| Overall | \n", "46.12 | \n", "47.15 | \n", "48.39 | \n", "
| Semantic | \n", "55.35 | \n", "55.60 | \n", "56.70 | \n", "
| Syntactic | \n", "38.45 | \n", "40.13 | \n", "41.48 | \n", "
| : capital-common-countries | \n", "81.82 | \n", "76.48 | \n", "77.87 | \n", "
| : capital-world | \n", "69.83 | \n", "71.07 | \n", "72.81 | \n", "
| : currency | \n", "7.62 | \n", "9.58 | \n", "10.28 | \n", "
| : city-in-state | \n", "37.37 | \n", "35.47 | \n", "35.51 | \n", "
| : family | \n", "68.77 | \n", "73.32 | \n", "74.31 | \n", "
| : gram1-adjective-to-adverb | \n", "12.70 | \n", "10.99 | \n", "15.93 | \n", "
| : gram2-opposite | \n", "11.82 | \n", "14.29 | \n", "16.01 | \n", "
| : gram3-comparative | \n", "44.14 | \n", "46.92 | \n", "49.10 | \n", "
| : gram4-superlative | \n", "17.02 | \n", "22.37 | \n", "24.33 | \n", "
| : gram5-present-participle | \n", "29.92 | \n", "34.28 | \n", "31.34 | \n", "
| : gram6-nationality-adjective | \n", "78.42 | \n", "76.11 | \n", "80.49 | \n", "
| : gram7-past-tense | \n", "44.29 | \n", "44.87 | \n", "43.85 | \n", "
| : gram8-plural | \n", "38.59 | \n", "40.77 | \n", "41.67 | \n", "
| : gram9-plural-verbs | \n", "37.82 | \n", "41.49 | \n", "40.92 | \n", "
| OOV questions | \n", "107.00 | \n", "0.00 | \n", "0.00 | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" word model \\\n",
"0 king GloVe (pre-trained) \n",
"1 king Wiki 20% \n",
"2 cat GloVe (pre-trained) \n",
"3 cat Wiki 20% \n",
"4 apple GloVe (pre-trained) \n",
"5 apple Wiki 20% \n",
"6 bank GloVe (pre-trained) \n",
"7 bank Wiki 20% \n",
"8 good GloVe (pre-trained) \n",
"9 good Wiki 20% \n",
"10 taiwan GloVe (pre-trained) \n",
"11 taiwan Wiki 20% \n",
"12 computer GloVe (pre-trained) \n",
"13 computer Wiki 20% \n",
"\n",
" top 5 (similarity) \n",
"0 prince (0.77), queen (0.75), son (0.70), brother (0.70), monarch (0.70) \n",
"1 nangklao (0.74), chlothar (0.72), queen (0.70), suryavarman (0.70), bodawpaya (0.70) \n",
"2 dog (0.88), rabbit (0.74), cats (0.73), monkey (0.73), pet (0.72) \n",
"3 rabbit (0.73), dog (0.72), sourpuss (0.72), mouse (0.71), pet (0.70) \n",
"4 microsoft (0.74), ibm (0.68), intel (0.68), software (0.68), dell (0.67) \n",
"5 blackberry (0.79), iphone (0.70), goldieblox (0.69), tvos (0.68), raspberry (0.68) \n",
"6 banks (0.81), banking (0.75), credit (0.70), investment (0.69), financial (0.68) \n",
"7 guaranty (0.77), savings (0.77), ameriprise (0.77), indymac (0.76), onewest (0.76) \n",
"8 better (0.89), sure (0.83), really (0.83), kind (0.83), very (0.83) \n",
"9 sure (0.71), decent (0.71), lovely (0.70), bad (0.70), thankful (0.70) \n",
"10 mainland (0.86), china (0.83), taiwanese (0.79), taipei (0.79), hong (0.77) \n",
"11 taipei (0.83), china (0.82), guangdong (0.82), hainan (0.79), fujian (0.79) \n",
"12 computers (0.88), software (0.84), technology (0.76), pc (0.74), hardware (0.73) \n",
"13 computing (0.83), computers (0.79), software (0.79), mainframe (0.77), hardware (0.74) "
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 4 題：挑字，各找 5 個最像的字。同時看現成的 GloVe 和自己訓練的 Wiki 20%。\n",
"# 挑字的原則：一字多義（apple、bank）、反義詞（good）、專有名詞（taiwan）、一般名詞（cat、computer）、考卷裡的字（king）\n",
"# 注意：兩個模型的 cosine 分數尺度不同，只能比「同一個模型內的排名」，不能拿分數互相比大小。\n",
"probe = [\"king\", \"cat\", \"apple\", \"bank\", \"good\", \"taiwan\", \"computer\"]\n",
"rows = []\n",
"for w in probe:\n",
" for name in [\"GloVe (pre-trained)\", \"Wiki 20%\"]:\n",
" wv = models[name]\n",
" if w in wv.key_to_index:\n",
" rows.append({\"word\": w, \"model\": name,\n",
" \"top 5 (similarity)\": \", \".join(f\"{x} ({s:.2f})\" for x, s in wv.most_similar(w, topn=5))})\n",
"pd.set_option(\"display.max_colwidth\", 200)\n",
"display(pd.DataFrame(rows))\n"
]
},
{
"cell_type": "code",
"execution_count": 33,
"id": "4cf48994",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T13:18:34.376973Z",
"iopub.status.busy": "2026-10-05T13:18:34.375975Z",
"iopub.status.idle": "2026-10-05T13:18:35.264063Z",
"shell.execute_reply": "2026-10-05T13:18:35.264063Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"king -> [('nangklao', np.int64(34)), ('chlothar', np.int64(76)), ('queen', np.int64(92609)), ('suryavarman', np.int64(53)), ('bodawpaya', np.int64(77))]\n",
"cat -> [('rabbit', np.int64(7844)), ('dog', np.int64(36766)), ('sourpuss', np.int64(64)), ('mouse', np.int64(13524)), ('pet', np.int64(9983))]\n",
"apple -> [('blackberry', np.int64(1285)), ('iphone', np.int64(2800)), ('goldieblox', np.int64(33)), ('tvos', np.int64(102)), ('raspberry', np.int64(1710))]\n",
"bank -> [('guaranty', np.int64(330)), ('savings', np.int64(8467)), ('ameriprise', np.int64(49)), ('indymac', np.int64(86)), ('onewest', np.int64(38))]\n"
]
},
{
"data": {
"text/html": [
"| \n", " | word | \n", "model | \n", "top 5 (similarity) | \n", "
|---|---|---|---|
| 0 | \n", "king | \n", "GloVe (pre-trained) | \n", "prince (0.77), queen (0.75), son (0.70), brother (0.70), monarch (0.70) | \n", "
| 1 | \n", "king | \n", "Wiki 20% | \n", "nangklao (0.74), chlothar (0.72), queen (0.70), suryavarman (0.70), bodawpaya (0.70) | \n", "
| 2 | \n", "cat | \n", "GloVe (pre-trained) | \n", "dog (0.88), rabbit (0.74), cats (0.73), monkey (0.73), pet (0.72) | \n", "
| 3 | \n", "cat | \n", "Wiki 20% | \n", "rabbit (0.73), dog (0.72), sourpuss (0.72), mouse (0.71), pet (0.70) | \n", "
| 4 | \n", "apple | \n", "GloVe (pre-trained) | \n", "microsoft (0.74), ibm (0.68), intel (0.68), software (0.68), dell (0.67) | \n", "
| 5 | \n", "apple | \n", "Wiki 20% | \n", "blackberry (0.79), iphone (0.70), goldieblox (0.69), tvos (0.68), raspberry (0.68) | \n", "
| 6 | \n", "bank | \n", "GloVe (pre-trained) | \n", "banks (0.81), banking (0.75), credit (0.70), investment (0.69), financial (0.68) | \n", "
| 7 | \n", "bank | \n", "Wiki 20% | \n", "guaranty (0.77), savings (0.77), ameriprise (0.77), indymac (0.76), onewest (0.76) | \n", "
| 8 | \n", "good | \n", "GloVe (pre-trained) | \n", "better (0.89), sure (0.83), really (0.83), kind (0.83), very (0.83) | \n", "
| 9 | \n", "good | \n", "Wiki 20% | \n", "sure (0.71), decent (0.71), lovely (0.70), bad (0.70), thankful (0.70) | \n", "
| 10 | \n", "taiwan | \n", "GloVe (pre-trained) | \n", "mainland (0.86), china (0.83), taiwanese (0.79), taipei (0.79), hong (0.77) | \n", "
| 11 | \n", "taiwan | \n", "Wiki 20% | \n", "taipei (0.83), china (0.82), guangdong (0.82), hainan (0.79), fujian (0.79) | \n", "
| 12 | \n", "computer | \n", "GloVe (pre-trained) | \n", "computers (0.88), software (0.84), technology (0.76), pc (0.74), hardware (0.73) | \n", "
| 13 | \n", "computer | \n", "Wiki 20% | \n", "computing (0.83), computers (0.79), software (0.79), mainframe (0.77), hardware (0.74) | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" word top 5 among the 50k most frequent words\n",
"0 king queen, throne, prince, reigned, ruler\n",
"1 cat rabbit, dog, mouse, pet, mug\n",
"2 apple blackberry, iphone, raspberry, ipad, android\n",
"3 bank savings, bancorp, banking, lenders, jpmorgan\n",
"4 good sure, decent, lovely, bad, thankful\n",
"5 taiwan taipei, china, guangdong, hainan, fujian\n",
"6 computer computing, computers, software, mainframe, hardware"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"apple top 15: ['blackberry', 'iphone', 'goldieblox', 'tvos', 'raspberry', 'xelibri', 'iigs', 'visicalc', 'magix', 'airis', 'cloudflare', 'roxio', 'ipad', 'livescribe', 'android']\n",
"similarity(apple, banana) = 0.41\n",
"similarity(apple, fruit) = 0.37\n",
"similarity(apple, microsoft) = 0.65\n",
"similarity(bank, river) = 0.49\n",
"similarity(bank, money) = 0.40\n",
"similarity(good, bad) = 0.70\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"... of guard dog bushy doozy best friend and hero to tom tom ms lulu the pet psychic sourpuss owner sourpuss owner and friend of millie animals bip and bop two nutty squirrel ...\n",
"... toon characters including fanny zilch mighty mouse heckle and jeckle gandy goose sourpuss dinky duck little roquefort the terry bears dimwit and luno terry pre existing c ...\n",
"... k puck in hockey almost certainly from irish poc according to the oed puss as in sourpuss comes from irish pus pouting mouth rapparee an irish highwayman from ropaire sta ...\n",
"... n year title role notes hangin in episode the princess and the pea today special sourpuss sal episode smiles murielle episode three monkeys of bah roghar part episode thr ...\n",
"... ack flag discharge and flipper in few days before her th birthday her first band sourpuss played set at australia summersault festival where she met tim armstrong frontma ...\n"
]
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 4 題補充證據\n",
"# (1) Wiki 20% 的鄰居裡常有冷門字：印出每個鄰居在 20% 維基裡出現幾次（字典門檻見 TODO5 的 effective min_count）\n",
"for w in [\"king\", \"cat\", \"apple\", \"bank\"]:\n",
" print(w, \"->\", [(x, my_wv.get_vecattr(x, \"count\")) for x, _ in my_wv.most_similar(w, topn=5)])\n",
"\n",
"# (2) 只在最常見的 5 萬個字裡找鄰居（restrict_vocab），冷門字就被排除\n",
"rows = [{\"word\": w, \"top 5 among the 50k most frequent words\":\n",
" \", \".join(x for x, _ in my_wv.most_similar(w, topn=5, restrict_vocab=50000))} for w in probe]\n",
"display(pd.DataFrame(rows))\n",
"\n",
"# (3) 一字多義：apple 的前 15 名鄰居，以及幾組相似度\n",
"print(\"apple top 15:\", [x for x, _ in my_wv.most_similar(\"apple\", topn=15)])\n",
"for a, b in [(\"apple\", \"banana\"), (\"apple\", \"fruit\"), (\"apple\", \"microsoft\"),\n",
" (\"bank\", \"river\"), (\"bank\", \"money\"), (\"good\", \"bad\")]:\n",
" print(f\"similarity({a}, {b}) = {my_wv.similarity(a, b):.2f}\")\n",
"\n",
"# (4) 回到教材查 cat 的鄰居 sourpuss 是什麼：印出前 5 處上下文\n",
"found = []\n",
"with open(\"wiki_sampled_20_clean.txt\", encoding=\"utf-8\") as f:\n",
" for line in f:\n",
" i = line.find(\" sourpuss \")\n",
" if i >= 0:\n",
" found.append(line[max(0, i - 80): i + 90].strip())\n",
" if len(found) == 5:\n",
" break\n",
"for s in found:\n",
" print(\"...\", s, \"...\")\n"
]
},
{
"cell_type": "markdown",
"id": "41a53ade",
"metadata": {},
"source": [
"## Extra code for report question 5 (anything that strengthens the report)"
]
},
{
"cell_type": "code",
"execution_count": 34,
"id": "593f1378",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T13:18:35.375035Z",
"iopub.status.busy": "2026-10-05T13:18:35.375035Z",
"iopub.status.idle": "2026-10-05T13:51:26.573483Z",
"shell.execute_reply": "2026-10-05T13:51:26.571479Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 5% seed=1: overall 46.05% (OOV questions 107)\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 5% seed=2: overall 47.10% (OOV questions 107)\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 5% seed=3: overall 46.34% (OOV questions 107)\n"
]
},
{
"data": {
"text/html": [
"| \n", " | word | \n", "top 5 among the 50k most frequent words | \n", "
|---|---|---|
| 0 | \n", "king | \n", "queen, throne, prince, reigned, ruler | \n", "
| 1 | \n", "cat | \n", "rabbit, dog, mouse, pet, mug | \n", "
| 2 | \n", "apple | \n", "blackberry, iphone, raspberry, ipad, android | \n", "
| 3 | \n", "bank | \n", "savings, bancorp, banking, lenders, jpmorgan | \n", "
| 4 | \n", "good | \n", "sure, decent, lovely, bad, thankful | \n", "
| 5 | \n", "taiwan | \n", "taipei, china, guangdong, hainan, fujian | \n", "
| 6 | \n", "computer | \n", "computing, computers, software, mainframe, hardware | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" seed 42 seed 1 seed 2 seed 3 max - min\n",
"Overall 46.12 46.05 47.10 46.34 1.05\n",
"Semantic 55.35 54.18 55.98 54.72 1.80\n",
"Syntactic 38.45 39.30 39.72 39.37 1.26\n",
": capital-common-countries 81.82 80.04 82.81 81.82 2.77\n",
": capital-world 69.83 67.44 69.43 68.21 2.39\n",
": currency 7.62 7.39 8.89 8.31 1.50\n",
": city-in-state 37.37 38.27 39.04 37.74 1.66\n",
": family 68.77 67.39 72.13 69.17 4.74\n",
": gram1-adjective-to-adverb 12.70 11.29 14.21 11.19 3.02\n",
": gram2-opposite 11.82 10.84 11.08 13.05 2.22\n",
": gram3-comparative 44.14 46.10 47.82 48.87 4.73\n",
": gram4-superlative 17.02 17.83 18.98 19.79 2.76\n",
": gram5-present-participle 29.92 29.73 31.53 34.47 4.73\n",
": gram6-nationality-adjective 78.42 80.11 78.86 78.42 1.69\n",
": gram7-past-tense 44.29 44.81 44.55 42.37 2.44\n",
": gram8-plural 38.59 40.62 39.86 37.99 2.63\n",
": gram9-plural-verbs 37.82 39.77 38.97 37.70 2.07"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 5 題 (a)：同樣設定、只換亂數種子，結果會晃多少？——先量雜訊，後面的比較才知道哪些差距算數。\n",
"# 用 5% 維基（後面的比較實驗都用 5%）：原本的 seed=42 再加 seed=1、2、3，共 4 次。\n",
"noise_runs = {\"seed 42\": results[\"Wiki 5%\"]}\n",
"for s in [1, 2, 3]:\n",
" m = Word2Vec(corpus_file=\"wiki_sampled_5_clean.txt\", **{**W2V_PARAMS, \"seed\": s})\n",
" _ = evaluate(m.wv, f\"Wiki 5% seed={s}\")\n",
" noise_runs[f\"seed {s}\"] = results[f\"Wiki 5% seed={s}\"]\n",
"noise = pd.DataFrame(noise_runs).drop(index=\"OOV questions\")\n",
"noise[\"max - min\"] = noise.max(axis=1) - noise.min(axis=1)\n",
"display(noise.round(2))\n",
"NOISE = noise[\"max - min\"]\n",
"\n",
"def compare(base, other):\n",
" # 跟 base 比較；差距的絕對值大於「4 次重跑的最大落差」才標 True\n",
" t = pd.DataFrame({base: results[base], other: results[other]}).drop(index=\"OOV questions\")\n",
" t[\"diff\"] = t[other] - t[base]\n",
" t[\"noise (max-min of 4 seeds)\"] = NOISE\n",
" t[\"beyond noise?\"] = t[\"diff\"].abs() > t[\"noise (max-min of 4 seeds)\"]\n",
" return t.round(2)\n"
]
},
{
"cell_type": "code",
"execution_count": 35,
"id": "3c9f2b91",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T13:51:26.683600Z",
"iopub.status.busy": "2026-10-05T13:51:26.683218Z",
"iopub.status.idle": "2026-10-05T14:00:33.289856Z",
"shell.execute_reply": "2026-10-05T14:00:33.289856Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"questions containing a stop word: 206\n",
"SubCategory\n",
": family 86\n",
": gram1-adjective-to-adverb 62\n",
": currency 58\n",
"Name: count, dtype: int64\n",
"['Algeria dinar Korea won', 'Angola kwanza Korea won', 'Argentina peso Korea won', 'Armenia dram Korea won', 'Brazil real Korea won']\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"trained on wiki_sampled_5_nostop.txt: 7.0 min, vocab 278,810, effective min_count 9\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 5% (stop words removed): overall 45.25% (OOV questions 284)\n"
]
},
{
"data": {
"text/html": [
"| \n", " | seed 42 | \n", "seed 1 | \n", "seed 2 | \n", "seed 3 | \n", "max - min | \n", "
|---|---|---|---|---|---|
| Overall | \n", "46.12 | \n", "46.05 | \n", "47.10 | \n", "46.34 | \n", "1.05 | \n", "
| Semantic | \n", "55.35 | \n", "54.18 | \n", "55.98 | \n", "54.72 | \n", "1.80 | \n", "
| Syntactic | \n", "38.45 | \n", "39.30 | \n", "39.72 | \n", "39.37 | \n", "1.26 | \n", "
| : capital-common-countries | \n", "81.82 | \n", "80.04 | \n", "82.81 | \n", "81.82 | \n", "2.77 | \n", "
| : capital-world | \n", "69.83 | \n", "67.44 | \n", "69.43 | \n", "68.21 | \n", "2.39 | \n", "
| : currency | \n", "7.62 | \n", "7.39 | \n", "8.89 | \n", "8.31 | \n", "1.50 | \n", "
| : city-in-state | \n", "37.37 | \n", "38.27 | \n", "39.04 | \n", "37.74 | \n", "1.66 | \n", "
| : family | \n", "68.77 | \n", "67.39 | \n", "72.13 | \n", "69.17 | \n", "4.74 | \n", "
| : gram1-adjective-to-adverb | \n", "12.70 | \n", "11.29 | \n", "14.21 | \n", "11.19 | \n", "3.02 | \n", "
| : gram2-opposite | \n", "11.82 | \n", "10.84 | \n", "11.08 | \n", "13.05 | \n", "2.22 | \n", "
| : gram3-comparative | \n", "44.14 | \n", "46.10 | \n", "47.82 | \n", "48.87 | \n", "4.73 | \n", "
| : gram4-superlative | \n", "17.02 | \n", "17.83 | \n", "18.98 | \n", "19.79 | \n", "2.76 | \n", "
| : gram5-present-participle | \n", "29.92 | \n", "29.73 | \n", "31.53 | \n", "34.47 | \n", "4.73 | \n", "
| : gram6-nationality-adjective | \n", "78.42 | \n", "80.11 | \n", "78.86 | \n", "78.42 | \n", "1.69 | \n", "
| : gram7-past-tense | \n", "44.29 | \n", "44.81 | \n", "44.55 | \n", "42.37 | \n", "2.44 | \n", "
| : gram8-plural | \n", "38.59 | \n", "40.62 | \n", "39.86 | \n", "37.99 | \n", "2.63 | \n", "
| : gram9-plural-verbs | \n", "37.82 | \n", "39.77 | \n", "38.97 | \n", "37.70 | \n", "2.07 | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Wiki 5% Wiki 5% (stop words removed) diff \\\n",
"Overall 46.12 45.25 -0.87 \n",
"Semantic 55.35 55.41 0.06 \n",
"Syntactic 38.45 36.81 -1.65 \n",
": capital-common-countries 81.82 83.99 2.17 \n",
": capital-world 69.83 70.87 1.04 \n",
": currency 7.62 8.89 1.27 \n",
": city-in-state 37.37 37.82 0.45 \n",
": family 68.77 53.95 -14.82 \n",
": gram1-adjective-to-adverb 12.70 12.80 0.10 \n",
": gram2-opposite 11.82 9.98 -1.85 \n",
": gram3-comparative 44.14 37.01 -7.13 \n",
": gram4-superlative 17.02 16.31 -0.71 \n",
": gram5-present-participle 29.92 30.87 0.95 \n",
": gram6-nationality-adjective 78.42 80.61 2.19 \n",
": gram7-past-tense 44.29 41.09 -3.21 \n",
": gram8-plural 38.59 41.89 3.30 \n",
": gram9-plural-verbs 37.82 26.55 -11.26 \n",
"\n",
" noise (max-min of 4 seeds) beyond noise? \n",
"Overall 1.05 False \n",
"Semantic 1.80 False \n",
"Syntactic 1.26 True \n",
": capital-common-countries 2.77 False \n",
": capital-world 2.39 False \n",
": currency 1.50 False \n",
": city-in-state 1.66 False \n",
": family 4.74 True \n",
": gram1-adjective-to-adverb 3.02 False \n",
": gram2-opposite 2.22 False \n",
": gram3-comparative 4.73 True \n",
": gram4-superlative 2.76 False \n",
": gram5-present-participle 4.73 False \n",
": gram6-nationality-adjective 1.69 True \n",
": gram7-past-tense 2.44 True \n",
": gram8-plural 2.63 True \n",
": gram9-plural-verbs 2.07 True "
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 5 題 (b)：為什麼不移除停用詞？先看考卷，再真的做實驗比較。\n",
"# 停用詞表：NLTK English stopwords（https://www.nltk.org/）\n",
"import nltk\n",
"nltk.download(\"stopwords\", quiet=True)\n",
"from nltk.corpus import stopwords\n",
"sw = set(stopwords.words(\"english\"))\n",
"has_sw = data[\"Question\"].map(lambda q: any(w in sw for w in q.lower().split()))\n",
"print(\"questions containing a stop word:\", int(has_sw.sum()))\n",
"print(data[has_sw][\"SubCategory\"].value_counts())\n",
"print(data[has_sw][\"Question\"].head(5).tolist())\n",
"\n",
"# 實驗：同樣是 5% 維基、同樣的參數，只差在「有沒有移除停用詞」\n",
"with open(\"wiki_sampled_5_clean.txt\", encoding=\"utf-8\") as fin, open(\"wiki_sampled_5_nostop.txt\", \"w\", encoding=\"utf-8\", newline=\"\\n\") as fout:\n",
" for line in fin:\n",
" fout.write(\" \".join(w for w in line.split() if w not in sw) + \"\\n\")\n",
"m = train_w2v(\"wiki_sampled_5_nostop.txt\")\n",
"models[\"Wiki 5% (stop words removed)\"] = m.wv\n",
"_ = evaluate(m.wv, \"Wiki 5% (stop words removed)\")\n",
"display(compare(\"Wiki 5%\", \"Wiki 5% (stop words removed)\"))\n"
]
},
{
"cell_type": "code",
"execution_count": 36,
"id": "697699ff",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T14:00:33.355166Z",
"iopub.status.busy": "2026-10-05T14:00:33.355166Z",
"iopub.status.idle": "2026-10-05T14:49:46.977824Z",
"shell.execute_reply": "2026-10-05T14:49:46.976818Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 5% CBOW (sg=0): trained in 9.3 min\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 5% CBOW (sg=0): overall 52.58% (OOV questions 107)\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 5% window=10: trained in 15.2 min\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 5% window=10: overall 41.75% (OOV questions 107)\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 5% vector_size=300: trained in 17.9 min\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 5% vector_size=300: overall 57.44% (OOV questions 107)\n"
]
},
{
"data": {
"text/html": [
"| \n", " | Wiki 5% | \n", "Wiki 5% (stop words removed) | \n", "diff | \n", "noise (max-min of 4 seeds) | \n", "beyond noise? | \n", "
|---|---|---|---|---|---|
| Overall | \n", "46.12 | \n", "45.25 | \n", "-0.87 | \n", "1.05 | \n", "False | \n", "
| Semantic | \n", "55.35 | \n", "55.41 | \n", "0.06 | \n", "1.80 | \n", "False | \n", "
| Syntactic | \n", "38.45 | \n", "36.81 | \n", "-1.65 | \n", "1.26 | \n", "True | \n", "
| : capital-common-countries | \n", "81.82 | \n", "83.99 | \n", "2.17 | \n", "2.77 | \n", "False | \n", "
| : capital-world | \n", "69.83 | \n", "70.87 | \n", "1.04 | \n", "2.39 | \n", "False | \n", "
| : currency | \n", "7.62 | \n", "8.89 | \n", "1.27 | \n", "1.50 | \n", "False | \n", "
| : city-in-state | \n", "37.37 | \n", "37.82 | \n", "0.45 | \n", "1.66 | \n", "False | \n", "
| : family | \n", "68.77 | \n", "53.95 | \n", "-14.82 | \n", "4.74 | \n", "True | \n", "
| : gram1-adjective-to-adverb | \n", "12.70 | \n", "12.80 | \n", "0.10 | \n", "3.02 | \n", "False | \n", "
| : gram2-opposite | \n", "11.82 | \n", "9.98 | \n", "-1.85 | \n", "2.22 | \n", "False | \n", "
| : gram3-comparative | \n", "44.14 | \n", "37.01 | \n", "-7.13 | \n", "4.73 | \n", "True | \n", "
| : gram4-superlative | \n", "17.02 | \n", "16.31 | \n", "-0.71 | \n", "2.76 | \n", "False | \n", "
| : gram5-present-participle | \n", "29.92 | \n", "30.87 | \n", "0.95 | \n", "4.73 | \n", "False | \n", "
| : gram6-nationality-adjective | \n", "78.42 | \n", "80.61 | \n", "2.19 | \n", "1.69 | \n", "True | \n", "
| : gram7-past-tense | \n", "44.29 | \n", "41.09 | \n", "-3.21 | \n", "2.44 | \n", "True | \n", "
| : gram8-plural | \n", "38.59 | \n", "41.89 | \n", "3.30 | \n", "2.63 | \n", "True | \n", "
| : gram9-plural-verbs | \n", "37.82 | \n", "26.55 | \n", "-11.26 | \n", "2.07 | \n", "True | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Wiki 5% Wiki 5% CBOW (sg=0) diff \\\n",
"Overall 46.12 52.58 6.46 \n",
"Semantic 55.35 58.60 3.25 \n",
"Syntactic 38.45 47.58 9.12 \n",
": capital-common-countries 81.82 79.45 -2.37 \n",
": capital-world 69.83 71.86 2.03 \n",
": currency 7.62 8.43 0.81 \n",
": city-in-state 37.37 43.66 6.28 \n",
": family 68.77 77.87 9.09 \n",
": gram1-adjective-to-adverb 12.70 19.15 6.45 \n",
": gram2-opposite 11.82 11.45 -0.37 \n",
": gram3-comparative 44.14 69.14 25.00 \n",
": gram4-superlative 17.02 37.61 20.59 \n",
": gram5-present-participle 29.92 40.81 10.89 \n",
": gram6-nationality-adjective 78.42 76.61 -1.81 \n",
": gram7-past-tense 44.29 46.03 1.73 \n",
": gram8-plural 38.59 49.55 10.96 \n",
": gram9-plural-verbs 37.82 48.16 10.34 \n",
"\n",
" noise (max-min of 4 seeds) beyond noise? \n",
"Overall 1.05 True \n",
"Semantic 1.80 True \n",
"Syntactic 1.26 True \n",
": capital-common-countries 2.77 False \n",
": capital-world 2.39 False \n",
": currency 1.50 False \n",
": city-in-state 1.66 True \n",
": family 4.74 True \n",
": gram1-adjective-to-adverb 3.02 True \n",
": gram2-opposite 2.22 False \n",
": gram3-comparative 4.73 True \n",
": gram4-superlative 2.76 True \n",
": gram5-present-participle 4.73 True \n",
": gram6-nationality-adjective 1.69 True \n",
": gram7-past-tense 2.44 False \n",
": gram8-plural 2.63 True \n",
": gram9-plural-verbs 2.07 True "
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"text/html": [
"| \n", " | Wiki 5% | \n", "Wiki 5% CBOW (sg=0) | \n", "diff | \n", "noise (max-min of 4 seeds) | \n", "beyond noise? | \n", "
|---|---|---|---|---|---|
| Overall | \n", "46.12 | \n", "52.58 | \n", "6.46 | \n", "1.05 | \n", "True | \n", "
| Semantic | \n", "55.35 | \n", "58.60 | \n", "3.25 | \n", "1.80 | \n", "True | \n", "
| Syntactic | \n", "38.45 | \n", "47.58 | \n", "9.12 | \n", "1.26 | \n", "True | \n", "
| : capital-common-countries | \n", "81.82 | \n", "79.45 | \n", "-2.37 | \n", "2.77 | \n", "False | \n", "
| : capital-world | \n", "69.83 | \n", "71.86 | \n", "2.03 | \n", "2.39 | \n", "False | \n", "
| : currency | \n", "7.62 | \n", "8.43 | \n", "0.81 | \n", "1.50 | \n", "False | \n", "
| : city-in-state | \n", "37.37 | \n", "43.66 | \n", "6.28 | \n", "1.66 | \n", "True | \n", "
| : family | \n", "68.77 | \n", "77.87 | \n", "9.09 | \n", "4.74 | \n", "True | \n", "
| : gram1-adjective-to-adverb | \n", "12.70 | \n", "19.15 | \n", "6.45 | \n", "3.02 | \n", "True | \n", "
| : gram2-opposite | \n", "11.82 | \n", "11.45 | \n", "-0.37 | \n", "2.22 | \n", "False | \n", "
| : gram3-comparative | \n", "44.14 | \n", "69.14 | \n", "25.00 | \n", "4.73 | \n", "True | \n", "
| : gram4-superlative | \n", "17.02 | \n", "37.61 | \n", "20.59 | \n", "2.76 | \n", "True | \n", "
| : gram5-present-participle | \n", "29.92 | \n", "40.81 | \n", "10.89 | \n", "4.73 | \n", "True | \n", "
| : gram6-nationality-adjective | \n", "78.42 | \n", "76.61 | \n", "-1.81 | \n", "1.69 | \n", "True | \n", "
| : gram7-past-tense | \n", "44.29 | \n", "46.03 | \n", "1.73 | \n", "2.44 | \n", "False | \n", "
| : gram8-plural | \n", "38.59 | \n", "49.55 | \n", "10.96 | \n", "2.63 | \n", "True | \n", "
| : gram9-plural-verbs | \n", "37.82 | \n", "48.16 | \n", "10.34 | \n", "2.07 | \n", "True | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Wiki 5% Wiki 5% window=10 diff \\\n",
"Overall 46.12 41.75 -4.37 \n",
"Semantic 55.35 51.22 -4.13 \n",
"Syntactic 38.45 33.87 -4.58 \n",
": capital-common-countries 81.82 72.53 -9.29 \n",
": capital-world 69.83 68.24 -1.59 \n",
": currency 7.62 7.85 0.23 \n",
": city-in-state 37.37 31.01 -6.36 \n",
": family 68.77 50.59 -18.18 \n",
": gram1-adjective-to-adverb 12.70 11.09 -1.61 \n",
": gram2-opposite 11.82 6.90 -4.93 \n",
": gram3-comparative 44.14 32.88 -11.26 \n",
": gram4-superlative 17.02 9.80 -7.22 \n",
": gram5-present-participle 29.92 27.46 -2.46 \n",
": gram6-nationality-adjective 78.42 78.80 0.38 \n",
": gram7-past-tense 44.29 39.49 -4.81 \n",
": gram8-plural 38.59 33.56 -5.03 \n",
": gram9-plural-verbs 37.82 33.22 -4.60 \n",
"\n",
" noise (max-min of 4 seeds) beyond noise? \n",
"Overall 1.05 True \n",
"Semantic 1.80 True \n",
"Syntactic 1.26 True \n",
": capital-common-countries 2.77 True \n",
": capital-world 2.39 False \n",
": currency 1.50 False \n",
": city-in-state 1.66 True \n",
": family 4.74 True \n",
": gram1-adjective-to-adverb 3.02 False \n",
": gram2-opposite 2.22 True \n",
": gram3-comparative 4.73 True \n",
": gram4-superlative 2.76 True \n",
": gram5-present-participle 4.73 False \n",
": gram6-nationality-adjective 1.69 False \n",
": gram7-past-tense 2.44 True \n",
": gram8-plural 2.63 True \n",
": gram9-plural-verbs 2.07 True "
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"text/html": [
"| \n", " | Wiki 5% | \n", "Wiki 5% window=10 | \n", "diff | \n", "noise (max-min of 4 seeds) | \n", "beyond noise? | \n", "
|---|---|---|---|---|---|
| Overall | \n", "46.12 | \n", "41.75 | \n", "-4.37 | \n", "1.05 | \n", "True | \n", "
| Semantic | \n", "55.35 | \n", "51.22 | \n", "-4.13 | \n", "1.80 | \n", "True | \n", "
| Syntactic | \n", "38.45 | \n", "33.87 | \n", "-4.58 | \n", "1.26 | \n", "True | \n", "
| : capital-common-countries | \n", "81.82 | \n", "72.53 | \n", "-9.29 | \n", "2.77 | \n", "True | \n", "
| : capital-world | \n", "69.83 | \n", "68.24 | \n", "-1.59 | \n", "2.39 | \n", "False | \n", "
| : currency | \n", "7.62 | \n", "7.85 | \n", "0.23 | \n", "1.50 | \n", "False | \n", "
| : city-in-state | \n", "37.37 | \n", "31.01 | \n", "-6.36 | \n", "1.66 | \n", "True | \n", "
| : family | \n", "68.77 | \n", "50.59 | \n", "-18.18 | \n", "4.74 | \n", "True | \n", "
| : gram1-adjective-to-adverb | \n", "12.70 | \n", "11.09 | \n", "-1.61 | \n", "3.02 | \n", "False | \n", "
| : gram2-opposite | \n", "11.82 | \n", "6.90 | \n", "-4.93 | \n", "2.22 | \n", "True | \n", "
| : gram3-comparative | \n", "44.14 | \n", "32.88 | \n", "-11.26 | \n", "4.73 | \n", "True | \n", "
| : gram4-superlative | \n", "17.02 | \n", "9.80 | \n", "-7.22 | \n", "2.76 | \n", "True | \n", "
| : gram5-present-participle | \n", "29.92 | \n", "27.46 | \n", "-2.46 | \n", "4.73 | \n", "False | \n", "
| : gram6-nationality-adjective | \n", "78.42 | \n", "78.80 | \n", "0.38 | \n", "1.69 | \n", "False | \n", "
| : gram7-past-tense | \n", "44.29 | \n", "39.49 | \n", "-4.81 | \n", "2.44 | \n", "True | \n", "
| : gram8-plural | \n", "38.59 | \n", "33.56 | \n", "-5.03 | \n", "2.63 | \n", "True | \n", "
| : gram9-plural-verbs | \n", "37.82 | \n", "33.22 | \n", "-4.60 | \n", "2.07 | \n", "True | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Wiki 5% Wiki 5% vector_size=300 diff \\\n",
"Overall 46.12 57.44 11.32 \n",
"Semantic 55.35 68.54 13.19 \n",
"Syntactic 38.45 48.22 9.76 \n",
": capital-common-countries 81.82 95.45 13.64 \n",
": capital-world 69.83 79.55 9.73 \n",
": currency 7.62 7.62 0.00 \n",
": city-in-state 37.37 63.15 25.78 \n",
": family 68.77 73.72 4.94 \n",
": gram1-adjective-to-adverb 12.70 13.51 0.81 \n",
": gram2-opposite 11.82 14.90 3.08 \n",
": gram3-comparative 44.14 64.11 19.97 \n",
": gram4-superlative 17.02 26.74 9.71 \n",
": gram5-present-participle 29.92 38.35 8.43 \n",
": gram6-nationality-adjective 78.42 86.05 7.63 \n",
": gram7-past-tense 44.29 49.81 5.51 \n",
": gram8-plural 38.59 54.28 15.69 \n",
": gram9-plural-verbs 37.82 52.53 14.71 \n",
"\n",
" noise (max-min of 4 seeds) beyond noise? \n",
"Overall 1.05 True \n",
"Semantic 1.80 True \n",
"Syntactic 1.26 True \n",
": capital-common-countries 2.77 True \n",
": capital-world 2.39 True \n",
": currency 1.50 False \n",
": city-in-state 1.66 True \n",
": family 4.74 True \n",
": gram1-adjective-to-adverb 3.02 False \n",
": gram2-opposite 2.22 True \n",
": gram3-comparative 4.73 True \n",
": gram4-superlative 2.76 True \n",
": gram5-present-participle 4.73 True \n",
": gram6-nationality-adjective 1.69 True \n",
": gram7-past-tense 2.44 True \n",
": gram8-plural 2.63 True \n",
": gram9-plural-verbs 2.07 True "
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 5 題 (c)：超參數比較（都用 5% 維基，一次只改一個設定，其他跟主模型一樣）\n",
"variants = {\n",
" \"Wiki 5% CBOW (sg=0)\": dict(sg=0),\n",
" \"Wiki 5% window=10\": dict(window=10),\n",
" \"Wiki 5% vector_size=300\": dict(vector_size=300),\n",
"}\n",
"for name, change in variants.items():\n",
" t = time.time()\n",
" m = Word2Vec(corpus_file=\"wiki_sampled_5_clean.txt\", **{**W2V_PARAMS, **change})\n",
" print(f\"{name}: trained in {(time.time() - t) / 60:.1f} min\")\n",
" models[name] = m.wv\n",
" _ = evaluate(m.wv, name)\n",
"for name in variants:\n",
" display(compare(\"Wiki 5%\", name))\n"
]
},
{
"cell_type": "code",
"execution_count": 37,
"id": "e622dce1",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T14:49:47.191827Z",
"iopub.status.busy": "2026-10-05T14:49:47.191827Z",
"iopub.status.idle": "2026-10-05T14:49:47.225452Z",
"shell.execute_reply": "2026-10-05T14:49:47.225452Z"
}
},
"outputs": [
{
"data": {
"text/html": [
"| \n", " | Wiki 5% | \n", "Wiki 5% vector_size=300 | \n", "diff | \n", "noise (max-min of 4 seeds) | \n", "beyond noise? | \n", "
|---|---|---|---|---|---|
| Overall | \n", "46.12 | \n", "57.44 | \n", "11.32 | \n", "1.05 | \n", "True | \n", "
| Semantic | \n", "55.35 | \n", "68.54 | \n", "13.19 | \n", "1.80 | \n", "True | \n", "
| Syntactic | \n", "38.45 | \n", "48.22 | \n", "9.76 | \n", "1.26 | \n", "True | \n", "
| : capital-common-countries | \n", "81.82 | \n", "95.45 | \n", "13.64 | \n", "2.77 | \n", "True | \n", "
| : capital-world | \n", "69.83 | \n", "79.55 | \n", "9.73 | \n", "2.39 | \n", "True | \n", "
| : currency | \n", "7.62 | \n", "7.62 | \n", "0.00 | \n", "1.50 | \n", "False | \n", "
| : city-in-state | \n", "37.37 | \n", "63.15 | \n", "25.78 | \n", "1.66 | \n", "True | \n", "
| : family | \n", "68.77 | \n", "73.72 | \n", "4.94 | \n", "4.74 | \n", "True | \n", "
| : gram1-adjective-to-adverb | \n", "12.70 | \n", "13.51 | \n", "0.81 | \n", "3.02 | \n", "False | \n", "
| : gram2-opposite | \n", "11.82 | \n", "14.90 | \n", "3.08 | \n", "2.22 | \n", "True | \n", "
| : gram3-comparative | \n", "44.14 | \n", "64.11 | \n", "19.97 | \n", "4.73 | \n", "True | \n", "
| : gram4-superlative | \n", "17.02 | \n", "26.74 | \n", "9.71 | \n", "2.76 | \n", "True | \n", "
| : gram5-present-participle | \n", "29.92 | \n", "38.35 | \n", "8.43 | \n", "4.73 | \n", "True | \n", "
| : gram6-nationality-adjective | \n", "78.42 | \n", "86.05 | \n", "7.63 | \n", "1.69 | \n", "True | \n", "
| : gram7-past-tense | \n", "44.29 | \n", "49.81 | \n", "5.51 | \n", "2.44 | \n", "True | \n", "
| : gram8-plural | \n", "38.59 | \n", "54.28 | \n", "15.69 | \n", "2.63 | \n", "True | \n", "
| : gram9-plural-verbs | \n", "37.82 | \n", "52.53 | \n", "14.71 | \n", "2.07 | \n", "True | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" top-1 top-3 top-5 top-10\n",
"GloVe (pre-trained) 63.11 73.68 77.73 82.01\n",
"Wiki 20% 48.39 61.53 66.32 71.74"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 5 題 (d)：前 k 名有猜中就算對，正確率會變多少？（老師說 queen 可能不是第一名，而是第二、三名）\n",
"rows = {}\n",
"for name in [\"GloVe (pre-trained)\", \"Wiki 20%\"]:\n",
" ranks = details[name][\"ranks\"]\n",
" rows[name] = {f\"top-{k}\": np.mean([r is not None and r <= k for r in ranks]) * 100 for k in [1, 3, 5, 10]}\n",
"display(pd.DataFrame(rows).T.round(2))\n"
]
},
{
"cell_type": "code",
"execution_count": 38,
"id": "853f2bb5",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T14:49:47.301914Z",
"iopub.status.busy": "2026-10-05T14:49:47.301914Z",
"iopub.status.idle": "2026-10-05T14:49:47.486495Z",
"shell.execute_reply": "2026-10-05T14:49:47.486495Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"wrong answers: 10087 of 19544\n",
" OOV (question word not in vocab): 0\n",
" gold answer was 2nd-10th: 4563\n",
"\n",
": capital-world\n"
]
},
{
"data": {
"text/html": [
"| \n", " | top-1 | \n", "top-3 | \n", "top-5 | \n", "top-10 | \n", "
|---|---|---|---|---|
| GloVe (pre-trained) | \n", "63.11 | \n", "73.68 | \n", "77.73 | \n", "82.01 | \n", "
| Wiki 20% | \n", "48.39 | \n", "61.53 | \n", "66.32 | \n", "71.74 | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Question gold pred rank\n",
"508 Abuja Nigeria Amman Jordan jordan kuwait NaN\n",
"509 Abuja Nigeria Ankara Turkey turkey sakarya 3.0\n",
"510 Abuja Nigeria Antananarivo Madagascar madagascar senegal 8.0\n",
"511 Abuja Nigeria Apia Samoa samoa tonga 2.0\n",
"513 Abuja Nigeria Asmara Eritrea eritrea indonesia 2.0\n",
"516 Abuja Nigeria Baghdad Iraq iraq syria 2.0"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
": currency\n"
]
},
{
"data": {
"text/html": [
"| \n", " | Question | \n", "gold | \n", "pred | \n", "rank | \n", "
|---|---|---|---|---|
| 508 | \n", "Abuja Nigeria Amman Jordan | \n", "jordan | \n", "kuwait | \n", "NaN | \n", "
| 509 | \n", "Abuja Nigeria Ankara Turkey | \n", "turkey | \n", "sakarya | \n", "3.0 | \n", "
| 510 | \n", "Abuja Nigeria Antananarivo Madagascar | \n", "madagascar | \n", "senegal | \n", "8.0 | \n", "
| 511 | \n", "Abuja Nigeria Apia Samoa | \n", "samoa | \n", "tonga | \n", "2.0 | \n", "
| 513 | \n", "Abuja Nigeria Asmara Eritrea | \n", "eritrea | \n", "indonesia | \n", "2.0 | \n", "
| 516 | \n", "Abuja Nigeria Baghdad Iraq | \n", "iraq | \n", "syria | \n", "2.0 | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Question gold pred rank\n",
"5030 Algeria dinar Angola kwanza kwanza centavos NaN\n",
"5032 Algeria dinar Armenia dram dram hryvnia NaN\n",
"5033 Algeria dinar Brazil real real peso NaN\n",
"5034 Algeria dinar Bulgaria lev lev hryvnia NaN\n",
"5035 Algeria dinar Cambodia riel riel kyat NaN\n",
"5036 Algeria dinar Canada dollar dollar loonie NaN"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
": family\n"
]
},
{
"data": {
"text/html": [
"| \n", " | Question | \n", "gold | \n", "pred | \n", "rank | \n", "
|---|---|---|---|---|
| 5030 | \n", "Algeria dinar Angola kwanza | \n", "kwanza | \n", "centavos | \n", "NaN | \n", "
| 5032 | \n", "Algeria dinar Armenia dram | \n", "dram | \n", "hryvnia | \n", "NaN | \n", "
| 5033 | \n", "Algeria dinar Brazil real | \n", "real | \n", "peso | \n", "NaN | \n", "
| 5034 | \n", "Algeria dinar Bulgaria lev | \n", "lev | \n", "hryvnia | \n", "NaN | \n", "
| 5035 | \n", "Algeria dinar Cambodia riel | \n", "riel | \n", "kyat | \n", "NaN | \n", "
| 5036 | \n", "Algeria dinar Canada dollar | \n", "dollar | \n", "loonie | \n", "NaN | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Question gold pred rank\n",
"8366 boy girl father mother mother stepmother 2.0\n",
"8376 boy girl nephew niece niece granddaughter 6.0\n",
"8391 brother sister groom bride bride bridesmaid 2.0\n",
"8394 brother sister husband wife wife widowed 4.0\n",
"8396 brother sister man woman woman sexless 10.0\n",
"8397 brother sister nephew niece niece aunt NaN"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
": gram3-comparative\n"
]
},
{
"data": {
"text/html": [
"| \n", " | Question | \n", "gold | \n", "pred | \n", "rank | \n", "
|---|---|---|---|---|
| 8366 | \n", "boy girl father mother | \n", "mother | \n", "stepmother | \n", "2.0 | \n", "
| 8376 | \n", "boy girl nephew niece | \n", "niece | \n", "granddaughter | \n", "6.0 | \n", "
| 8391 | \n", "brother sister groom bride | \n", "bride | \n", "bridesmaid | \n", "2.0 | \n", "
| 8394 | \n", "brother sister husband wife | \n", "wife | \n", "widowed | \n", "4.0 | \n", "
| 8396 | \n", "brother sister man woman | \n", "woman | \n", "sexless | \n", "10.0 | \n", "
| 8397 | \n", "brother sister nephew niece | \n", "niece | \n", "aunt | \n", "NaN | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Question gold pred rank\n",
"10677 bad worse cool cooler cooler fresher 6.0\n",
"10682 bad worse great greater greater britain 7.0\n",
"10683 bad worse hard harder harder aisam 3.0\n",
"10686 bad worse hot hotter hotter peaking NaN\n",
"10687 bad worse large larger larger smaller 2.0\n",
"10688 bad worse long longer longer shorter 4.0"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
": gram6-nationality-adjective\n"
]
},
{
"data": {
"text/html": [
"| \n", " | Question | \n", "gold | \n", "pred | \n", "rank | \n", "
|---|---|---|---|---|
| 10677 | \n", "bad worse cool cooler | \n", "cooler | \n", "fresher | \n", "6.0 | \n", "
| 10682 | \n", "bad worse great greater | \n", "greater | \n", "britain | \n", "7.0 | \n", "
| 10683 | \n", "bad worse hard harder | \n", "harder | \n", "aisam | \n", "3.0 | \n", "
| 10686 | \n", "bad worse hot hotter | \n", "hotter | \n", "peaking | \n", "NaN | \n", "
| 10687 | \n", "bad worse large larger | \n", "larger | \n", "smaller | \n", "2.0 | \n", "
| 10688 | \n", "bad worse long longer | \n", "longer | \n", "shorter | \n", "4.0 | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Question gold pred \\\n",
"14183 Albania Albanian Argentina Argentinean argentinean argentinian \n",
"14186 Albania Albanian Belarus Belorussian belorussian belarusian \n",
"14199 Albania Albanian Greece Greek greek cypriot \n",
"14217 Albania Albanian Slovakia Slovakian slovakian slovak \n",
"14224 Argentina Argentinean Belarus Belorussian belorussian russian \n",
"14229 Argentina Argentinean China Chinese chinese guangdong \n",
"\n",
" rank \n",
"14183 8.0 \n",
"14186 NaN \n",
"14199 2.0 \n",
"14217 3.0 \n",
"14224 NaN \n",
"14229 2.0 "
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
": gram8-plural\n"
]
},
{
"data": {
"text/html": [
"| \n", " | Question | \n", "gold | \n", "pred | \n", "rank | \n", "
|---|---|---|---|---|
| 14183 | \n", "Albania Albanian Argentina Argentinean | \n", "argentinean | \n", "argentinian | \n", "8.0 | \n", "
| 14186 | \n", "Albania Albanian Belarus Belorussian | \n", "belorussian | \n", "belarusian | \n", "NaN | \n", "
| 14199 | \n", "Albania Albanian Greece Greek | \n", "greek | \n", "cypriot | \n", "2.0 | \n", "
| 14217 | \n", "Albania Albanian Slovakia Slovakian | \n", "slovakian | \n", "slovak | \n", "3.0 | \n", "
| 14224 | \n", "Argentina Argentinean Belarus Belorussian | \n", "belorussian | \n", "russian | \n", "NaN | \n", "
| 14229 | \n", "Argentina Argentinean China Chinese | \n", "chinese | \n", "guangdong | \n", "2.0 | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Question gold pred rank\n",
"17342 banana bananas bird birds birds waterfowl 8.0\n",
"17344 banana bananas building buildings buildings renovating 2.0\n",
"17345 banana bananas car cars cars truck 2.0\n",
"17346 banana bananas cat cats cats dogs 9.0\n",
"17347 banana bananas child children children childbearing 5.0\n",
"17348 banana bananas cloud clouds clouds lidar NaN"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 5 題 (e)：錯題分析——電腦答錯時，它到底猜了什麼？\n",
"d20 = details[\"Wiki 20%\"]\n",
"err = data.assign(pred=d20[\"preds\"], rank=d20[\"ranks\"],\n",
" gold=data[\"Question\"].str.lower().str.split().str[3])\n",
"wrong = err[err[\"pred\"] != err[\"gold\"]]\n",
"print(\"wrong answers:\", len(wrong), \"of\", len(err))\n",
"print(\" OOV (question word not in vocab):\", int(wrong[\"pred\"].isna().sum()))\n",
"print(\" gold answer was 2nd-10th:\", int(wrong[\"rank\"].notna().sum()))\n",
"for sub in [\": capital-world\", \": currency\", \": family\", \": gram3-comparative\", \": gram6-nationality-adjective\", \": gram8-plural\"]:\n",
" print(\"\\n\" + sub)\n",
" display(wrong[(wrong[\"SubCategory\"] == sub) & wrong[\"pred\"].notna()][[\"Question\", \"gold\", \"pred\", \"rank\"]].head(6))\n"
]
},
{
"cell_type": "code",
"execution_count": 39,
"id": "76df2075",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T14:49:47.560569Z",
"iopub.status.busy": "2026-10-05T14:49:47.560569Z",
"iopub.status.idle": "2026-10-05T14:55:08.176617Z",
"shell.execute_reply": "2026-10-05T14:55:08.175606Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Wiki 20% (3CosMul): overall 44.66% (OOV questions 0)\n"
]
},
{
"data": {
"text/html": [
"| \n", " | Question | \n", "gold | \n", "pred | \n", "rank | \n", "
|---|---|---|---|---|
| 17342 | \n", "banana bananas bird birds | \n", "birds | \n", "waterfowl | \n", "8.0 | \n", "
| 17344 | \n", "banana bananas building buildings | \n", "buildings | \n", "renovating | \n", "2.0 | \n", "
| 17345 | \n", "banana bananas car cars | \n", "cars | \n", "truck | \n", "2.0 | \n", "
| 17346 | \n", "banana bananas cat cats | \n", "cats | \n", "dogs | \n", "9.0 | \n", "
| 17347 | \n", "banana bananas child children | \n", "children | \n", "childbearing | \n", "5.0 | \n", "
| 17348 | \n", "banana bananas cloud clouds | \n", "clouds | \n", "lidar | \n", "NaN | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" Wiki 20% Wiki 20% (3CosMul)\n",
"Overall 48.39 44.66\n",
"Semantic 56.70 54.12\n",
"Syntactic 41.48 36.80\n",
": capital-common-countries 77.87 77.27\n",
": capital-world 72.81 69.30\n",
": currency 10.28 9.35\n",
": city-in-state 35.51 33.73\n",
": family 74.31 71.34\n",
": gram1-adjective-to-adverb 15.93 12.00\n",
": gram2-opposite 16.01 13.18\n",
": gram3-comparative 49.10 43.62\n",
": gram4-superlative 24.33 22.19\n",
": gram5-present-participle 31.34 24.43\n",
": gram6-nationality-adjective 80.49 76.05\n",
": gram7-past-tense 43.85 39.55\n",
": gram8-plural 41.67 34.23\n",
": gram9-plural-verbs 40.92 37.36\n",
"OOV questions 0.00 0.00"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 5 題 (f)：換一種「算答案」的公式。3CosAdd 是 b - a + c；3CosMul 用乘除。\n",
"# 出處：Levy & Goldberg (2014), Linguistic Regularities in Sparse and Explicit Word Representations, CoNLL. https://aclanthology.org/W14-1618/\n",
"# 同一份 Wiki 20% 詞向量，只換公式。\n",
"_ = evaluate(my_wv, \"Wiki 20% (3CosMul)\", method=\"cosmul\")\n",
"display(pd.DataFrame({k: results[k] for k in [\"Wiki 20%\", \"Wiki 20% (3CosMul)\"]}).round(2))\n"
]
},
{
"cell_type": "code",
"execution_count": 40,
"id": "2ed58697",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T14:55:08.272920Z",
"iopub.status.busy": "2026-10-05T14:55:08.272920Z",
"iopub.status.idle": "2026-10-05T14:55:09.052343Z",
"shell.execute_reply": "2026-10-05T14:55:09.050168Z"
}
},
"outputs": [
{
"data": {
"image/png": "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",
"text/plain": [
"| \n", " | Wiki 20% | \n", "Wiki 20% (3CosMul) | \n", "
|---|---|---|
| Overall | \n", "48.39 | \n", "44.66 | \n", "
| Semantic | \n", "56.70 | \n", "54.12 | \n", "
| Syntactic | \n", "41.48 | \n", "36.80 | \n", "
| : capital-common-countries | \n", "77.87 | \n", "77.27 | \n", "
| : capital-world | \n", "72.81 | \n", "69.30 | \n", "
| : currency | \n", "10.28 | \n", "9.35 | \n", "
| : city-in-state | \n", "35.51 | \n", "33.73 | \n", "
| : family | \n", "74.31 | \n", "71.34 | \n", "
| : gram1-adjective-to-adverb | \n", "15.93 | \n", "12.00 | \n", "
| : gram2-opposite | \n", "16.01 | \n", "13.18 | \n", "
| : gram3-comparative | \n", "49.10 | \n", "43.62 | \n", "
| : gram4-superlative | \n", "24.33 | \n", "22.19 | \n", "
| : gram5-present-participle | \n", "31.34 | \n", "24.43 | \n", "
| : gram6-nationality-adjective | \n", "80.49 | \n", "76.05 | \n", "
| : gram7-past-tense | \n", "43.85 | \n", "39.55 | \n", "
| : gram8-plural | \n", "41.67 | \n", "34.23 | \n", "
| : gram9-plural-verbs | \n", "40.92 | \n", "37.36 | \n", "
| OOV questions | \n", "0.00 | \n", "0.00 | \n", "
"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 5 題 (g)：同一批 family 字，用 PCA 和 t-SNE 兩種方法壓成 2 維，圖長得不一樣嗎？\n",
"from sklearn.decomposition import PCA\n",
"words = sorted({w for q in data[data[\"SubCategory\"] == \": family\"][\"Question\"] for w in q.lower().split()})\n",
"words = [w for w in words if w in my_wv.key_to_index]\n",
"vecs = np.array([my_wv[w] for w in words])\n",
"fig, axes = plt.subplots(1, 2, figsize=(18, 8))\n",
"for ax, (title, pts) in zip(axes, [\n",
" (\"PCA\", PCA(n_components=2, random_state=42).fit_transform(vecs)),\n",
" (\"t-SNE\", TSNE(n_components=2, perplexity=15, random_state=42, init=\"pca\").fit_transform(vecs))]):\n",
" ax.scatter(pts[:, 0], pts[:, 1], s=12)\n",
" for (x, y), w in zip(pts, words):\n",
" ax.annotate(w, (x, y), fontsize=8)\n",
" ax.set_title(f\"{title} (Wiki 20%, family)\")\n",
"plt.show()\n"
]
},
{
"cell_type": "code",
"execution_count": 41,
"id": "e3304cd7",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T14:55:09.147791Z",
"iopub.status.busy": "2026-10-05T14:55:09.146791Z",
"iopub.status.idle": "2026-10-05T14:55:09.251897Z",
"shell.execute_reply": "2026-10-05T14:55:09.250891Z"
}
},
"outputs": [
{
"data": {
"text/html": [
"
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" GloVe Wiki 20%\n",
"SubCategory \n",
": capital-common-countries 100.0 100.0\n",
": capital-world 100.0 100.0\n",
": currency 100.0 100.0\n",
": city-in-state 100.0 100.0\n",
": family 100.0 100.0\n",
": gram1-adjective-to-adverb 100.0 100.0\n",
": gram2-opposite 100.0 100.0\n",
": gram3-comparative 100.0 100.0\n",
": gram4-superlative 100.0 100.0\n",
": gram5-present-participle 100.0 100.0\n",
": gram6-nationality-adjective 100.0 100.0\n",
": gram7-past-tense 100.0 100.0\n",
": gram8-plural 100.0 100.0\n",
": gram9-plural-verbs 100.0 100.0"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 第 5 題 (h)：各小類「四個字都在字典裡」的比例：GloVe vs Wiki 20%\n",
"display(pd.DataFrame({\"GloVe\": answer_coverage(model), \"Wiki 20%\": answer_coverage(my_wv)}))\n"
]
},
{
"cell_type": "markdown",
"id": "2e5cae87",
"metadata": {},
"source": [
"## Summary of all results"
]
},
{
"cell_type": "code",
"execution_count": 42,
"id": "146d3910",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T14:55:09.357529Z",
"iopub.status.busy": "2026-10-05T14:55:09.357529Z",
"iopub.status.idle": "2026-10-05T14:55:09.386767Z",
"shell.execute_reply": "2026-10-05T14:55:09.386254Z"
}
},
"outputs": [
{
"data": {
"text/html": [
"| \n", " | GloVe | \n", "Wiki 20% | \n", "
|---|---|---|
| SubCategory | \n", "\n", " | \n", " |
| : capital-common-countries | \n", "100.0 | \n", "100.0 | \n", "
| : capital-world | \n", "100.0 | \n", "100.0 | \n", "
| : currency | \n", "100.0 | \n", "100.0 | \n", "
| : city-in-state | \n", "100.0 | \n", "100.0 | \n", "
| : family | \n", "100.0 | \n", "100.0 | \n", "
| : gram1-adjective-to-adverb | \n", "100.0 | \n", "100.0 | \n", "
| : gram2-opposite | \n", "100.0 | \n", "100.0 | \n", "
| : gram3-comparative | \n", "100.0 | \n", "100.0 | \n", "
| : gram4-superlative | \n", "100.0 | \n", "100.0 | \n", "
| : gram5-present-participle | \n", "100.0 | \n", "100.0 | \n", "
| : gram6-nationality-adjective | \n", "100.0 | \n", "100.0 | \n", "
| : gram7-past-tense | \n", "100.0 | \n", "100.0 | \n", "
| : gram8-plural | \n", "100.0 | \n", "100.0 | \n", "
| : gram9-plural-verbs | \n", "100.0 | \n", "100.0 | \n", "
\n",
"\n",
"\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" GloVe (pre-trained) Wiki 20% \\\n",
"Overall 63.11 48.39 \n",
"Semantic 65.34 56.70 \n",
"Syntactic 61.26 41.48 \n",
": capital-common-countries 93.87 77.87 \n",
": capital-world 88.95 72.81 \n",
": currency 14.20 10.28 \n",
": city-in-state 30.81 35.51 \n",
": family 81.62 74.31 \n",
": gram1-adjective-to-adverb 24.40 15.93 \n",
": gram2-opposite 20.07 16.01 \n",
": gram3-comparative 79.13 49.10 \n",
": gram4-superlative 54.28 24.33 \n",
": gram5-present-participle 69.51 31.34 \n",
": gram6-nationality-adjective 87.87 80.49 \n",
": gram7-past-tense 55.45 43.85 \n",
": gram8-plural 72.00 41.67 \n",
": gram9-plural-verbs 58.39 40.92 \n",
"OOV questions 0.00 0.00 \n",
"\n",
" Forum (same size) Wiki control (same size) \\\n",
"Overall 25.78 41.46 \n",
"Semantic 12.84 45.77 \n",
"Syntactic 36.53 37.87 \n",
": capital-common-countries 33.60 76.48 \n",
": capital-world 12.51 54.00 \n",
": currency 1.27 6.93 \n",
": city-in-state 2.92 33.81 \n",
": family 63.24 66.21 \n",
": gram1-adjective-to-adverb 13.10 10.28 \n",
": gram2-opposite 14.04 8.87 \n",
": gram3-comparative 64.94 53.08 \n",
": gram4-superlative 40.55 17.20 \n",
": gram5-present-participle 51.33 31.82 \n",
": gram6-nationality-adjective 21.70 76.55 \n",
": gram7-past-tense 33.78 40.96 \n",
": gram8-plural 41.97 36.26 \n",
": gram9-plural-verbs 41.49 32.99 \n",
"OOV questions 3510.00 169.00 \n",
"\n",
" Wiki 5% Wiki 10% Wiki 5% seed=1 \\\n",
"Overall 46.12 47.15 46.05 \n",
"Semantic 55.35 55.60 54.18 \n",
"Syntactic 38.45 40.13 39.30 \n",
": capital-common-countries 81.82 76.48 80.04 \n",
": capital-world 69.83 71.07 67.44 \n",
": currency 7.62 9.58 7.39 \n",
": city-in-state 37.37 35.47 38.27 \n",
": family 68.77 73.32 67.39 \n",
": gram1-adjective-to-adverb 12.70 10.99 11.29 \n",
": gram2-opposite 11.82 14.29 10.84 \n",
": gram3-comparative 44.14 46.92 46.10 \n",
": gram4-superlative 17.02 22.37 17.83 \n",
": gram5-present-participle 29.92 34.28 29.73 \n",
": gram6-nationality-adjective 78.42 76.11 80.11 \n",
": gram7-past-tense 44.29 44.87 44.81 \n",
": gram8-plural 38.59 40.77 40.62 \n",
": gram9-plural-verbs 37.82 41.49 39.77 \n",
"OOV questions 107.00 0.00 107.00 \n",
"\n",
" Wiki 5% seed=2 Wiki 5% seed=3 \\\n",
"Overall 47.10 46.34 \n",
"Semantic 55.98 54.72 \n",
"Syntactic 39.72 39.37 \n",
": capital-common-countries 82.81 81.82 \n",
": capital-world 69.43 68.21 \n",
": currency 8.89 8.31 \n",
": city-in-state 39.04 37.74 \n",
": family 72.13 69.17 \n",
": gram1-adjective-to-adverb 14.21 11.19 \n",
": gram2-opposite 11.08 13.05 \n",
": gram3-comparative 47.82 48.87 \n",
": gram4-superlative 18.98 19.79 \n",
": gram5-present-participle 31.53 34.47 \n",
": gram6-nationality-adjective 78.86 78.42 \n",
": gram7-past-tense 44.55 42.37 \n",
": gram8-plural 39.86 37.99 \n",
": gram9-plural-verbs 38.97 37.70 \n",
"OOV questions 107.00 107.00 \n",
"\n",
" Wiki 5% (stop words removed) \\\n",
"Overall 45.25 \n",
"Semantic 55.41 \n",
"Syntactic 36.81 \n",
": capital-common-countries 83.99 \n",
": capital-world 70.87 \n",
": currency 8.89 \n",
": city-in-state 37.82 \n",
": family 53.95 \n",
": gram1-adjective-to-adverb 12.80 \n",
": gram2-opposite 9.98 \n",
": gram3-comparative 37.01 \n",
": gram4-superlative 16.31 \n",
": gram5-present-participle 30.87 \n",
": gram6-nationality-adjective 80.61 \n",
": gram7-past-tense 41.09 \n",
": gram8-plural 41.89 \n",
": gram9-plural-verbs 26.55 \n",
"OOV questions 284.00 \n",
"\n",
" Wiki 5% CBOW (sg=0) Wiki 5% window=10 \\\n",
"Overall 52.58 41.75 \n",
"Semantic 58.60 51.22 \n",
"Syntactic 47.58 33.87 \n",
": capital-common-countries 79.45 72.53 \n",
": capital-world 71.86 68.24 \n",
": currency 8.43 7.85 \n",
": city-in-state 43.66 31.01 \n",
": family 77.87 50.59 \n",
": gram1-adjective-to-adverb 19.15 11.09 \n",
": gram2-opposite 11.45 6.90 \n",
": gram3-comparative 69.14 32.88 \n",
": gram4-superlative 37.61 9.80 \n",
": gram5-present-participle 40.81 27.46 \n",
": gram6-nationality-adjective 76.61 78.80 \n",
": gram7-past-tense 46.03 39.49 \n",
": gram8-plural 49.55 33.56 \n",
": gram9-plural-verbs 48.16 33.22 \n",
"OOV questions 107.00 107.00 \n",
"\n",
" Wiki 5% vector_size=300 Wiki 20% (3CosMul) \n",
"Overall 57.44 44.66 \n",
"Semantic 68.54 54.12 \n",
"Syntactic 48.22 36.80 \n",
": capital-common-countries 95.45 77.27 \n",
": capital-world 79.55 69.30 \n",
": currency 7.62 9.35 \n",
": city-in-state 63.15 33.73 \n",
": family 73.72 71.34 \n",
": gram1-adjective-to-adverb 13.51 12.00 \n",
": gram2-opposite 14.90 13.18 \n",
": gram3-comparative 64.11 43.62 \n",
": gram4-superlative 26.74 22.19 \n",
": gram5-present-participle 38.35 24.43 \n",
": gram6-nationality-adjective 86.05 76.05 \n",
": gram7-past-tense 49.81 39.55 \n",
": gram8-plural 54.28 34.23 \n",
": gram9-plural-verbs 52.53 37.36 \n",
"OOV questions 107.00 0.00 "
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"# 全部結果總表（報告引用的數字都來自這裡）\n",
"pd.set_option(\"display.max_columns\", 40)\n",
"display(pd.DataFrame(results).round(2))\n"
]
},
{
"cell_type": "markdown",
"id": "82cbba80",
"metadata": {},
"source": [
"## Running environment"
]
},
{
"cell_type": "code",
"execution_count": 43,
"id": "2f535ace",
"metadata": {
"execution": {
"iopub.execute_input": "2026-10-05T14:55:09.499777Z",
"iopub.status.busy": "2026-10-05T14:55:09.499777Z",
"iopub.status.idle": "2026-10-05T14:55:09.507837Z",
"shell.execute_reply": "2026-10-05T14:55:09.507837Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Python 3.12.3 (tags/v3.12.3:f6650f9, Apr 9 2024, 14:05:25) [MSC v.1938 64 bit (AMD64)]\n",
"Windows-11-10.0.26200-SP0 | Intel64 Family 6 Model 141 Stepping 1, GenuineIntel\n",
"gensim 4.4.0 | numpy 2.5.3 | pandas 3.0.6 | scikit-learn 1.9.1\n"
]
}
],
"source": [
"# [Generative AI] The code in this cell was generated by Claude (Anthropic).\n",
"import sys, platform, gensim, sklearn\n",
"print(\"Python\", sys.version)\n",
"print(platform.platform(), \"|\", platform.processor())\n",
"print(\"gensim\", gensim.__version__, \"| numpy\", np.__version__, \"| pandas\", pd.__version__, \"| scikit-learn\", sklearn.__version__)\n"
]
}
],
"metadata": {
"colab": {
"provenance": []
},
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"name": "python"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
| \n", " | GloVe (pre-trained) | \n", "Wiki 20% | \n", "Forum (same size) | \n", "Wiki control (same size) | \n", "Wiki 5% | \n", "Wiki 10% | \n", "Wiki 5% seed=1 | \n", "Wiki 5% seed=2 | \n", "Wiki 5% seed=3 | \n", "Wiki 5% (stop words removed) | \n", "Wiki 5% CBOW (sg=0) | \n", "Wiki 5% window=10 | \n", "Wiki 5% vector_size=300 | \n", "Wiki 20% (3CosMul) | \n", "
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Overall | \n", "63.11 | \n", "48.39 | \n", "25.78 | \n", "41.46 | \n", "46.12 | \n", "47.15 | \n", "46.05 | \n", "47.10 | \n", "46.34 | \n", "45.25 | \n", "52.58 | \n", "41.75 | \n", "57.44 | \n", "44.66 | \n", "
| Semantic | \n", "65.34 | \n", "56.70 | \n", "12.84 | \n", "45.77 | \n", "55.35 | \n", "55.60 | \n", "54.18 | \n", "55.98 | \n", "54.72 | \n", "55.41 | \n", "58.60 | \n", "51.22 | \n", "68.54 | \n", "54.12 | \n", "
| Syntactic | \n", "61.26 | \n", "41.48 | \n", "36.53 | \n", "37.87 | \n", "38.45 | \n", "40.13 | \n", "39.30 | \n", "39.72 | \n", "39.37 | \n", "36.81 | \n", "47.58 | \n", "33.87 | \n", "48.22 | \n", "36.80 | \n", "
| : capital-common-countries | \n", "93.87 | \n", "77.87 | \n", "33.60 | \n", "76.48 | \n", "81.82 | \n", "76.48 | \n", "80.04 | \n", "82.81 | \n", "81.82 | \n", "83.99 | \n", "79.45 | \n", "72.53 | \n", "95.45 | \n", "77.27 | \n", "
| : capital-world | \n", "88.95 | \n", "72.81 | \n", "12.51 | \n", "54.00 | \n", "69.83 | \n", "71.07 | \n", "67.44 | \n", "69.43 | \n", "68.21 | \n", "70.87 | \n", "71.86 | \n", "68.24 | \n", "79.55 | \n", "69.30 | \n", "
| : currency | \n", "14.20 | \n", "10.28 | \n", "1.27 | \n", "6.93 | \n", "7.62 | \n", "9.58 | \n", "7.39 | \n", "8.89 | \n", "8.31 | \n", "8.89 | \n", "8.43 | \n", "7.85 | \n", "7.62 | \n", "9.35 | \n", "
| : city-in-state | \n", "30.81 | \n", "35.51 | \n", "2.92 | \n", "33.81 | \n", "37.37 | \n", "35.47 | \n", "38.27 | \n", "39.04 | \n", "37.74 | \n", "37.82 | \n", "43.66 | \n", "31.01 | \n", "63.15 | \n", "33.73 | \n", "
| : family | \n", "81.62 | \n", "74.31 | \n", "63.24 | \n", "66.21 | \n", "68.77 | \n", "73.32 | \n", "67.39 | \n", "72.13 | \n", "69.17 | \n", "53.95 | \n", "77.87 | \n", "50.59 | \n", "73.72 | \n", "71.34 | \n", "
| : gram1-adjective-to-adverb | \n", "24.40 | \n", "15.93 | \n", "13.10 | \n", "10.28 | \n", "12.70 | \n", "10.99 | \n", "11.29 | \n", "14.21 | \n", "11.19 | \n", "12.80 | \n", "19.15 | \n", "11.09 | \n", "13.51 | \n", "12.00 | \n", "
| : gram2-opposite | \n", "20.07 | \n", "16.01 | \n", "14.04 | \n", "8.87 | \n", "11.82 | \n", "14.29 | \n", "10.84 | \n", "11.08 | \n", "13.05 | \n", "9.98 | \n", "11.45 | \n", "6.90 | \n", "14.90 | \n", "13.18 | \n", "
| : gram3-comparative | \n", "79.13 | \n", "49.10 | \n", "64.94 | \n", "53.08 | \n", "44.14 | \n", "46.92 | \n", "46.10 | \n", "47.82 | \n", "48.87 | \n", "37.01 | \n", "69.14 | \n", "32.88 | \n", "64.11 | \n", "43.62 | \n", "
| : gram4-superlative | \n", "54.28 | \n", "24.33 | \n", "40.55 | \n", "17.20 | \n", "17.02 | \n", "22.37 | \n", "17.83 | \n", "18.98 | \n", "19.79 | \n", "16.31 | \n", "37.61 | \n", "9.80 | \n", "26.74 | \n", "22.19 | \n", "
| : gram5-present-participle | \n", "69.51 | \n", "31.34 | \n", "51.33 | \n", "31.82 | \n", "29.92 | \n", "34.28 | \n", "29.73 | \n", "31.53 | \n", "34.47 | \n", "30.87 | \n", "40.81 | \n", "27.46 | \n", "38.35 | \n", "24.43 | \n", "
| : gram6-nationality-adjective | \n", "87.87 | \n", "80.49 | \n", "21.70 | \n", "76.55 | \n", "78.42 | \n", "76.11 | \n", "80.11 | \n", "78.86 | \n", "78.42 | \n", "80.61 | \n", "76.61 | \n", "78.80 | \n", "86.05 | \n", "76.05 | \n", "
| : gram7-past-tense | \n", "55.45 | \n", "43.85 | \n", "33.78 | \n", "40.96 | \n", "44.29 | \n", "44.87 | \n", "44.81 | \n", "44.55 | \n", "42.37 | \n", "41.09 | \n", "46.03 | \n", "39.49 | \n", "49.81 | \n", "39.55 | \n", "
| : gram8-plural | \n", "72.00 | \n", "41.67 | \n", "41.97 | \n", "36.26 | \n", "38.59 | \n", "40.77 | \n", "40.62 | \n", "39.86 | \n", "37.99 | \n", "41.89 | \n", "49.55 | \n", "33.56 | \n", "54.28 | \n", "34.23 | \n", "
| : gram9-plural-verbs | \n", "58.39 | \n", "40.92 | \n", "41.49 | \n", "32.99 | \n", "37.82 | \n", "41.49 | \n", "39.77 | \n", "38.97 | \n", "37.70 | \n", "26.55 | \n", "48.16 | \n", "33.22 | \n", "52.53 | \n", "37.36 | \n", "
| OOV questions | \n", "0.00 | \n", "0.00 | \n", "3510.00 | \n", "169.00 | \n", "107.00 | \n", "0.00 | \n", "107.00 | \n", "107.00 | \n", "107.00 | \n", "284.00 | \n", "107.00 | \n", "107.00 | \n", "107.00 | \n", "0.00 | \n", "