Hugging Face's logo

undertheseanlp
/
sen-1

Text Classification
Scikit-learn
Vietnamese
vietnamese
tfidf
svm
Model card Files
xet
Community
sen-1 / references /index.html
Tiep's picture
Tiep
Add references folder and research skills
ef06968
raw
history blame contribute delete
57.9 kB
 <!DOCTYPE html>
<html lang="vi">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Sen-1 References - Vietnamese Text Classification</title>
<style>
:root {
--bg: #0d1117;
--surface: #161b22;
--surface2: #1c2333;
--border: #30363d;
--text: #e6edf3;
--text2: #8b949e;
--accent: #58a6ff;
--accent2: #3fb950;
--accent3: #d2a8ff;
--accent4: #f0883e;
--red: #f85149;
--tag-bg: #1f2937;
}
* { margin: 0; padding: 0; box-sizing: border-box; }
body {
font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Helvetica, Arial, sans-serif;
background: var(--bg);
color: var(--text);
line-height: 1.6;
}
a { color: var(--accent); text-decoration: none; }
a:hover { text-decoration: underline; }
/* Header */
.header {
background: var(--surface);
border-bottom: 1px solid var(--border);
padding: 1.5rem 2rem;
position: sticky;
top: 0;
z-index: 100;
}
.header-inner {
max-width: 1200px;
margin: 0 auto;
display: flex;
align-items: center;
justify-content: space-between;
gap: 1rem;
}
.logo {
display: flex;
align-items: center;
gap: 0.75rem;
}
.logo-icon {
width: 36px;
height: 36px;
background: linear-gradient(135deg, var(--accent), var(--accent3));
border-radius: 8px;
display: flex;
align-items: center;
justify-content: center;
font-size: 18px;
font-weight: 700;
color: #fff;
}
.logo h1 {
font-size: 1.2rem;
font-weight: 600;
}
.logo span { color: var(--text2); font-weight: 400; font-size: 0.9rem; }
.stats {
display: flex;
gap: 1.5rem;
font-size: 0.85rem;
color: var(--text2);
}
.stat-num { color: var(--text); font-weight: 600; font-size: 1.1rem; }
/* Tabs */
.tabs {
max-width: 1200px;
margin: 0 auto;
padding: 0 2rem;
display: flex;
gap: 0;
border-bottom: 1px solid var(--border);
overflow-x: auto;
}
.tab {
padding: 0.75rem 1.25rem;
cursor: pointer;
color: var(--text2);
font-size: 0.9rem;
border-bottom: 2px solid transparent;
white-space: nowrap;
transition: all 0.15s;
}
.tab:hover { color: var(--text); }
.tab.active {
color: var(--text);
border-bottom-color: var(--accent4);
font-weight: 500;
}
/* Content */
.content {
max-width: 1200px;
margin: 0 auto;
padding: 2rem;
}
.panel { display: none; }
.panel.active { display: block; }
/* Paper Cards */
.papers-grid {
display: grid;
gap: 1rem;
}
.paper-card {
background: var(--surface);
border: 1px solid var(--border);
border-radius: 8px;
padding: 1.25rem 1.5rem;
transition: border-color 0.15s;
}
.paper-card:hover { border-color: var(--accent); }
.paper-title {
font-size: 1.05rem;
font-weight: 600;
margin-bottom: 0.5rem;
display: flex;
align-items: flex-start;
gap: 0.5rem;
}
.paper-meta {
display: flex;
flex-wrap: wrap;
gap: 0.75rem;
font-size: 0.85rem;
color: var(--text2);
margin-bottom: 0.5rem;
}
.paper-abstract {
font-size: 0.88rem;
color: var(--text2);
line-height: 1.5;
margin-top: 0.75rem;
display: -webkit-box;
-webkit-line-clamp: 3;
-webkit-box-orient: vertical;
overflow: hidden;
}
.paper-abstract.expanded {
-webkit-line-clamp: unset;
}
.tag {
display: inline-block;
padding: 0.15rem 0.55rem;
border-radius: 12px;
font-size: 0.75rem;
font-weight: 500;
}
.tag-venue { background: #1a3a2a; color: var(--accent2); }
.tag-year { background: #2a1a3a; color: var(--accent3); }
.tag-file { background: var(--tag-bg); color: var(--text2); }
.tag-pdf { background: #3a1a1a; color: var(--red); }
.tag-tex { background: #1a2a3a; color: var(--accent); }
.tag-md { background: #2a2a1a; color: var(--accent4); }
.expand-btn {
background: none;
border: none;
color: var(--accent);
cursor: pointer;
font-size: 0.82rem;
margin-top: 0.3rem;
}
/* Category filters */
.filters {
display: flex;
flex-wrap: wrap;
gap: 0.5rem;
margin-bottom: 1.5rem;
}
.filter-btn {
padding: 0.4rem 0.9rem;
border-radius: 20px;
border: 1px solid var(--border);
background: var(--surface);
color: var(--text2);
cursor: pointer;
font-size: 0.82rem;
transition: all 0.15s;
}
.filter-btn:hover { border-color: var(--accent); color: var(--text); }
.filter-btn.active { background: var(--accent); color: #000; border-color: var(--accent); font-weight: 600; }
/* Tables */
table {
width: 100%;
border-collapse: collapse;
margin: 1rem 0;
font-size: 0.88rem;
}
th, td {
padding: 0.65rem 0.85rem;
text-align: left;
border-bottom: 1px solid var(--border);
}
th {
background: var(--surface2);
color: var(--text2);
font-weight: 600;
font-size: 0.8rem;
text-transform: uppercase;
letter-spacing: 0.03em;
position: sticky;
top: 0;
}
tr:hover td { background: var(--surface); }
.highlight { color: var(--accent2); font-weight: 600; }
/* Section headers */
.section-title {
font-size: 1.4rem;
font-weight: 700;
margin-bottom: 0.5rem;
}
.section-desc {
color: var(--text2);
margin-bottom: 1.5rem;
font-size: 0.92rem;
}
h3 {
font-size: 1.1rem;
margin: 2rem 0 0.75rem;
padding-bottom: 0.4rem;
border-bottom: 1px solid var(--border);
}
/* Citation network */
.network {
background: var(--surface);
border: 1px solid var(--border);
border-radius: 8px;
padding: 1.5rem;
font-family: 'SF Mono', 'Fira Code', 'Consolas', monospace;
font-size: 0.82rem;
line-height: 1.8;
overflow-x: auto;
white-space: pre;
color: var(--text2);
}
.network b { color: var(--accent); font-weight: 600; }
.network em { color: var(--accent2); font-style: normal; }
.network u { color: var(--accent3); text-decoration: none; }
/* Blockquote */
blockquote {
border-left: 3px solid var(--accent3);
padding: 0.75rem 1.25rem;
margin: 1rem 0;
background: var(--surface);
border-radius: 0 6px 6px 0;
font-style: italic;
color: var(--text2);
}
blockquote strong { color: var(--text); }
/* Cards row */
.cards-row {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(250px, 1fr));
gap: 1rem;
margin: 1rem 0;
}
.info-card {
background: var(--surface);
border: 1px solid var(--border);
border-radius: 8px;
padding: 1.25rem;
}
.info-card h4 {
font-size: 0.95rem;
margin-bottom: 0.5rem;
color: var(--accent);
}
.info-card p, .info-card li {
font-size: 0.85rem;
color: var(--text2);
line-height: 1.6;
}
.info-card ul { padding-left: 1.2rem; }
/* How to do science */
.principle {
background: var(--surface);
border: 1px solid var(--border);
border-radius: 8px;
padding: 1.25rem 1.5rem;
margin-bottom: 1rem;
}
.principle h4 {
color: var(--accent4);
margin-bottom: 0.5rem;
}
.principle p, .principle li {
font-size: 0.88rem;
color: var(--text2);
line-height: 1.6;
}
.principle ul { padding-left: 1.2rem; margin-top: 0.3rem; }
/* Leaderboard */
.lb-hero {
text-align: center;
padding: 2.5rem 1rem 1.5rem;
}
.lb-hero h2 {
font-size: 2rem;
font-weight: 800;
background: linear-gradient(135deg, var(--accent), var(--accent3));
-webkit-background-clip: text;
-webkit-text-fill-color: transparent;
background-clip: text;
}
.lb-hero p {
color: var(--text2);
font-size: 0.92rem;
margin-top: 0.5rem;
}
.lb-updated {
color: var(--text2);
font-size: 0.75rem;
margin-top: 0.25rem;
opacity: 0.7;
}
.lb-spotlight {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(280px, 1fr));
gap: 1rem;
margin: 1.5rem 0;
}
.lb-spot-card {
background: var(--surface);
border: 1px solid var(--border);
border-radius: 10px;
padding: 1.25rem;
transition: border-color 0.15s;
}
.lb-spot-card:hover { border-color: var(--accent); }
.lb-spot-card h4 {
font-size: 0.78rem;
text-transform: uppercase;
letter-spacing: 0.05em;
color: var(--text2);
margin-bottom: 0.15rem;
}
.lb-spot-card .lb-bench-name {
font-size: 0.72rem;
color: var(--accent);
margin-bottom: 0.75rem;
}
.lb-rank-list { list-style: none; }
.lb-rank-list li {
display: flex;
align-items: center;
justify-content: space-between;
padding: 0.35rem 0;
font-size: 0.85rem;
border-bottom: 1px solid rgba(48,54,61,0.5);
}
.lb-rank-list li:last-child { border-bottom: none; }
.lb-rank-num {
display: inline-flex;
align-items: center;
justify-content: center;
width: 22px;
height: 22px;
border-radius: 50%;
font-size: 0.7rem;
font-weight: 700;
margin-right: 0.6rem;
flex-shrink: 0;
}
.lb-rank-1 { background: #fbbf24; color: #000; }
.lb-rank-2 { background: #94a3b8; color: #000; }
.lb-rank-3 { background: #b45309; color: #fff; }
.lb-rank-other { background: var(--tag-bg); color: var(--text2); }
.lb-model-name { flex: 1; }
.lb-model-name.lb-sen1 { color: var(--accent4); font-weight: 600; }
.lb-score {
font-weight: 600;
font-variant-numeric: tabular-nums;
color: var(--text);
}
.lb-score-top { color: var(--accent2); }
.lb-section-title {
font-size: 1.1rem;
font-weight: 700;
margin: 2.5rem 0 0.5rem;
padding-bottom: 0.4rem;
border-bottom: 1px solid var(--border);
}
.lb-section-desc {
color: var(--text2);
font-size: 0.85rem;
margin-bottom: 1rem;
}
/* Leaderboard table */
.lb-table-wrap {
overflow-x: auto;
border: 1px solid var(--border);
border-radius: 8px;
margin: 1rem 0;
}
.lb-table {
width: 100%;
border-collapse: collapse;
font-size: 0.84rem;
min-width: 900px;
}
.lb-table th {
background: var(--surface2);
color: var(--text2);
font-weight: 600;
font-size: 0.72rem;
text-transform: uppercase;
letter-spacing: 0.04em;
padding: 0.7rem 0.75rem;
border-bottom: 2px solid var(--border);
cursor: pointer;
white-space: nowrap;
user-select: none;
position: sticky;
top: 0;
}
.lb-table th:hover { color: var(--text); }
.lb-table th .sort-arrow { margin-left: 0.3rem; font-size: 0.65rem; color: var(--accent); }
.lb-table td {
padding: 0.6rem 0.75rem;
border-bottom: 1px solid rgba(48,54,61,0.5);
white-space: nowrap;
}
.lb-table tr:hover td { background: rgba(88,166,255,0.04); }
.lb-table .lb-row-sen1 { background: rgba(240,136,62,0.06); }
.lb-table .lb-row-sen1:hover td { background: rgba(240,136,62,0.1); }
.lb-table .model-type {
display: inline-block;
width: 8px;
height: 24px;
border-radius: 2px;
margin-right: 0.5rem;
vertical-align: middle;
}
.type-traditional { background: var(--accent4); }
.type-transformer { background: var(--accent); }
.type-multilingual { background: var(--accent3); }
.lb-table .cell-best { color: var(--accent2); font-weight: 700; }
.lb-table .cell-na { color: var(--text2); opacity: 0.5; }
.lb-table .model-cell {
display: flex;
align-items: center;
gap: 0.5rem;
}
/* Responsive */
@media (max-width: 768px) {
.header-inner { flex-direction: column; align-items: flex-start; }
.content { padding: 1rem; }
.cards-row { grid-template-columns: 1fr; }
.lb-spotlight { grid-template-columns: 1fr; }
.lb-hero h2 { font-size: 1.5rem; }
}
</style>
</head>
<body>
<div class="header">
<div class="header-inner">
<div class="logo">
<div class="logo-icon">S1</div>
<div>
<h1>Sen-1 References <span>Vietnamese Text Classification</span></h1>
</div>
</div>
<div class="stats">
<div><span class="stat-num" id="total-papers">12</span> papers</div>
<div><span class="stat-num">10</span> PDFs</div>
<div><span class="stat-num">7</span> LaTeX</div>
</div>
</div>
</div>
<div class="tabs" id="tabs">
<div class="tab active" data-tab="papers">Papers</div>
<div class="tab" data-tab="comparison">Benchmarks</div>
<div class="tab" data-tab="sota">SOTA</div>
<div class="tab" data-tab="network">Citation Network</div>
<div class="tab" data-tab="leaderboard">Leaderboard</div>
<div class="tab" data-tab="science">How to Do Science</div>
</div>
<!-- ===================== PAPERS ===================== -->
<div class="content">
<div class="panel active" id="panel-papers">
<div class="section-title">Paper Database</div>
<div class="section-desc">Research papers related to Vietnamese text classification, fetched from arXiv and ACL Anthology.</div>
<div class="filters">
<button class="filter-btn active" data-cat="all">All (12)</button>
<button class="filter-btn" data-cat="vn-cls">Vietnamese Classification</button>
<button class="filter-btn" data-cat="vn-pretrained">Vietnamese Pretrained</button>
<button class="filter-btn" data-cat="multilingual">Multilingual Models</button>
<button class="filter-btn" data-cat="datasets">Datasets</button>
</div>
<div class="papers-grid" id="papers-grid">
<!-- PhoBERT -->
<div class="paper-card" data-cats="vn-pretrained">
<div class="paper-title">
<a href="https://arxiv.org/abs/2003.00744" target="_blank">PhoBERT: Pre-trained language models for Vietnamese</a>
</div>
<div class="paper-meta">
<span class="tag tag-year">2020</span>
<span class="tag tag-venue">EMNLP Findings</span>
<span class="tag tag-pdf">PDF</span>
<span class="tag tag-tex">TEX</span>
<span class="tag tag-md">MD</span>
<span>Dat Quoc Nguyen, Anh Tuan Nguyen</span>
</div>
<div class="paper-abstract">We present PhoBERT with two versions, PhoBERT-base and PhoBERT-large, the first public large-scale monolingual language models pre-trained for Vietnamese. Experimental results show that PhoBERT consistently outperforms the recent best pre-trained multilingual model XLM-R and improves the state-of-the-art in multiple Vietnamese-specific NLP tasks including Part-of-speech tagging, Dependency parsing, Named-entity recognition and Natural language inference.</div>
<div style="font-size:0.78rem;color:var(--text2);margin-top:0.5rem">
<code>2020.arxiv.nguyen/</code> &middot; <code>2020.findings.anh/</code>
</div>
</div>
<!-- ViSoBERT -->
<div class="paper-card" data-cats="vn-pretrained">
<div class="paper-title">
<a href="https://arxiv.org/abs/2310.11166" target="_blank">ViSoBERT: A Pre-Trained Language Model for Vietnamese Social Media Text Processing</a>
</div>
<div class="paper-meta">
<span class="tag tag-year">2023</span>
<span class="tag tag-venue">EMNLP 2023</span>
<span class="tag tag-pdf">PDF</span>
<span class="tag tag-tex">TEX</span>
<span class="tag tag-md">MD</span>
<span>Quoc-Nam Nguyen, Thang Chau Phan, Duc-Vu Nguyen, Kiet Van Nguyen</span>
</div>
<div class="paper-abstract">We present the first monolingual pre-trained language model for Vietnamese social media texts, ViSoBERT, which is pre-trained on a large-scale corpus of high-quality and diverse Vietnamese social media texts using XLM-R architecture. ViSoBERT surpasses the previous state-of-the-art models on multiple Vietnamese social media tasks with far fewer parameters.</div>
<div style="font-size:0.78rem;color:var(--text2);margin-top:0.5rem">
<code>2023.arxiv.nguyen/</code> &middot; <code>2023.emnlp.kiet/</code>
</div>
</div>
<!-- vELECTRA -->
<div class="paper-card" data-cats="vn-pretrained">
<div class="paper-title">
<a href="https://arxiv.org/abs/2006.15994" target="_blank">Improving Sequence Tagging for Vietnamese Text Using Transformer-based Neural Models</a>
</div>
<div class="paper-meta">
<span class="tag tag-year">2020</span>
<span class="tag tag-venue">arXiv</span>
<span class="tag tag-pdf">PDF</span>
<span class="tag tag-tex">TEX</span>
<span class="tag tag-md">MD</span>
<span>Viet Bui The, Oanh Tran Thi, Phuong Le-Hong</span>
</div>
<div class="paper-abstract">Introduces viBERT (trained on 10GB) and vELECTRA (trained on 60GB) Vietnamese pretrained models. Strong performance on sequence tagging and text classification tasks. vELECTRA achieves 95.26% on ViOCD complaint classification in the SMTCE benchmark.</div>
<div style="font-size:0.78rem;color:var(--text2);margin-top:0.5rem"><code>2020.arxiv.the/</code></div>
</div>
<!-- SMTCE -->
<div class="paper-card" data-cats="vn-cls">
<div class="paper-title">
<a href="https://arxiv.org/abs/2209.10482" target="_blank">SMTCE: A Social Media Text Classification Evaluation Benchmark and BERTology Models for Vietnamese</a>
</div>
<div class="paper-meta">
<span class="tag tag-year">2022</span>
<span class="tag tag-venue">PACLIC 2022</span>
<span class="tag tag-pdf">PDF</span>
<span class="tag tag-tex">TEX</span>
<span class="tag tag-md">MD</span>
<span>Luan Thanh Nguyen, Kiet Van Nguyen, Ngan Luu-Thuy Nguyen</span>
</div>
<div class="paper-abstract">GLUE-inspired benchmark for Vietnamese social media text classification. Compares multilingual (mBERT, XLM-R, DistilmBERT) and monolingual (PhoBERT, viBERT, vELECTRA, viBERT4news) BERT models. Monolingual models consistently outperform multilingual for Vietnamese.</div>
<div style="font-size:0.78rem;color:var(--text2);margin-top:0.5rem">
<code>2022.arxiv.nguyen/</code> &middot; <code>2022.paclic.ngan/</code>
</div>
</div>
<!-- Vu et al 2007 -->
<div class="paper-card" data-cats="vn-cls">
<div class="paper-title">
<a href="https://ieeexplore.ieee.org/document/4223084/" target="_blank">A Comparative Study on Vietnamese Text Classification Methods</a>
</div>
<div class="paper-meta">
<span class="tag tag-year">2007</span>
<span class="tag tag-venue">IEEE RIVF</span>
<span class="tag tag-md">MD</span>
<span>Cong Duy Vu Hoang, Dien Dinh, Le Nguyen Nguyen, Quoc Hung Ngo</span>
</div>
<div class="paper-abstract">Seminal paper introducing VNTC corpus and comparing BOW and N-gram language model approaches for Vietnamese text classification. N-gram LM achieves 97.1% accuracy, SVM Multi achieves 93.4% on 10-topic news classification. The VNTC dataset remains the standard benchmark.</div>
<div style="font-size:0.78rem;color:var(--text2);margin-top:0.5rem"><code>2007.rivf.hoang/</code></div>
</div>
<!-- RoBERTa -->
<div class="paper-card" data-cats="multilingual">
<div class="paper-title">
<a href="https://arxiv.org/abs/1907.11692" target="_blank">RoBERTa: A Robustly Optimized BERT Pretraining Approach</a>
</div>
<div class="paper-meta">
<span class="tag tag-year">2019</span>
<span class="tag tag-venue">arXiv</span>
<span class="tag tag-pdf">PDF</span>
<span class="tag tag-tex">TEX</span>
<span class="tag tag-md">MD</span>
<span>Yinhan Liu, Myle Ott, Naman Goyal, ...</span>
</div>
<div class="paper-abstract">PhoBERT is based on the RoBERTa architecture. Key optimizations over BERT: dynamic masking, larger batches, more training data, removal of Next Sentence Prediction (NSP). Foundation for most Vietnamese pretrained models.</div>
<div style="font-size:0.78rem;color:var(--text2);margin-top:0.5rem"><code>2019.arxiv.liu/</code></div>
</div>
<!-- XLM-R -->
<div class="paper-card" data-cats="multilingual">
<div class="paper-title">
<a href="https://arxiv.org/abs/1911.02116" target="_blank">Unsupervised Cross-lingual Representation Learning at Scale (XLM-RoBERTa)</a>
</div>
<div class="paper-meta">
<span class="tag tag-year">2019</span>
<span class="tag tag-venue">ACL 2020</span>
<span class="tag tag-pdf">PDF</span>
<span class="tag tag-tex">TEX</span>
<span class="tag tag-md">MD</span>
<span>Alexis Conneau, Kartikay Khandelwal, Naman Goyal, ...</span>
</div>
<div class="paper-abstract">Multilingual pretrained model trained on 100 languages (2.5TB CC-100). Strong multilingual baseline for Vietnamese, but consistently outperformed by monolingual PhoBERT on Vietnamese-specific tasks.</div>
<div style="font-size:0.78rem;color:var(--text2);margin-top:0.5rem"><code>2019.arxiv.conneau/</code></div>
</div>
<!-- UIT-VSMEC -->
<div class="paper-card" data-cats="datasets">
<div class="paper-title">
<a href="https://arxiv.org/abs/1911.09339" target="_blank">Emotion Recognition for Vietnamese Social Media Text (UIT-VSMEC)</a>
</div>
<div class="paper-meta">
<span class="tag tag-year">2019</span>
<span class="tag tag-venue">CSoNet 2020</span>
<span class="tag tag-pdf">PDF</span>
<span class="tag tag-tex">TEX</span>
<span class="tag tag-md">MD</span>
<span>Vong Anh Ho, Duong Huynh-Cong Nguyen, Danh Hoang Nguyen, ...</span>
</div>
<div class="paper-abstract">Introduces UIT-VSMEC corpus: 6,927 emotion-annotated Vietnamese social media sentences with 7 labels (sadness, enjoyment, anger, disgust, fear, surprise, other). CNN baseline achieves 59.74% weighted F1.</div>
<div style="font-size:0.78rem;color:var(--text2);margin-top:0.5rem"><code>2019.arxiv.ho/</code></div>
</div>
<!-- UIT-VSFC -->
<div class="paper-card" data-cats="datasets">
<div class="paper-title">
<a href="https://ieeexplore.ieee.org/document/8573337/" target="_blank">UIT-VSFC: Vietnamese Students' Feedback Corpus for Sentiment Analysis</a>
</div>
<div class="paper-meta">
<span class="tag tag-year">2018</span>
<span class="tag tag-venue">KSE 2018</span>
<span class="tag tag-md">MD</span>
<span>Kiet Van Nguyen, Vu Duc Nguyen, Phu Xuan-Vinh Nguyen, ...</span>
</div>
<div class="paper-abstract">16,175 Vietnamese student feedback sentences annotated for sentiment (3 classes: positive, negative, neutral) and topic classification. Inter-annotator agreement: 91.20% for sentiment. MaxEnt baseline: 88% sentiment F1.</div>
<div style="font-size:0.78rem;color:var(--text2);margin-top:0.5rem"><code>2018.kse.nguyen/</code></div>
</div>
</div>
</div>
<!-- ===================== BENCHMARKS ===================== -->
<div class="panel" id="panel-comparison">
<div class="section-title">Benchmark Comparison</div>
<div class="section-desc">Vietnamese text classification results across datasets and models.</div>
<h3>VNTC Dataset (10-topic News Classification)</h3>
<table>
<tr><th>Model</th><th>Year</th><th>Accuracy</th><th>F1 (weighted)</th><th>Training</th><th>Inference</th><th>Size</th></tr>
<tr><td>N-gram LM (Vu et al.)</td><td>2007</td><td class="highlight">97.1%</td><td>-</td><td>~79 min</td><td>-</td><td>-</td></tr>
<tr><td>SVM Multi (Vu et al.)</td><td>2007</td><td>93.4%</td><td>-</td><td>~79 min</td><td>-</td><td>-</td></tr>
<tr><td>sonar_core_1 (SVC)</td><td>-</td><td>92.80%</td><td>92.0%</td><td>~54.6 min</td><td>-</td><td>~75MB</td></tr>
<tr style="background:var(--surface2)"><td><strong>Sen-1 (LinearSVC)</strong></td><td>2026</td><td>92.49%</td><td>92.40%</td><td class="highlight">37.6s</td><td class="highlight">66K/sec</td><td class="highlight">2.4MB</td></tr>
<tr><td>PhoBERT-base*</td><td>2020</td><td>~95-97%</td><td>~95%</td><td>Hours (GPU)</td><td>~20/sec</td><td>~400MB</td></tr>
</table>
<p style="font-size:0.78rem;color:var(--text2)">*PhoBERT not directly evaluated on VNTC; estimates from similar tasks.</p>
<h3>UTS2017_Bank Dataset (14-category Banking)</h3>
<table>
<tr><th>Model</th><th>Accuracy</th><th>F1 (weighted)</th><th>F1 (macro)</th><th>Training</th></tr>
<tr style="background:var(--surface2)"><td><strong>Sen-1</strong></td><td class="highlight">75.76%</td><td class="highlight">72.70%</td><td>36.18%</td><td class="highlight">0.13s</td></tr>
<tr><td>sonar_core_1</td><td>72.47%</td><td>66.0%</td><td>-</td><td>~5.3s</td></tr>
</table>
<h3>Vietnamese Pretrained Models</h3>
<table>
<tr><th>Model</th><th>Architecture</th><th>Pre-training Data</th><th>Languages</th><th>Vietnamese Tasks</th></tr>
<tr><td><strong>PhoBERT</strong></td><td>RoBERTa</td><td>20GB Vietnamese</td><td>1 (vi)</td><td>SOTA: POS, NER, NLI</td></tr>
<tr><td><strong>ViSoBERT</strong></td><td>XLM-R</td><td>Social media corpus</td><td>1 (vi)</td><td>SOTA: social media tasks</td></tr>
<tr><td><strong>vELECTRA</strong></td><td>ELECTRA</td><td>60GB Vietnamese</td><td>1 (vi)</td><td>Strong on classification</td></tr>
<tr><td>viBERT</td><td>BERT</td><td>10GB Vietnamese</td><td>1 (vi)</td><td>Baseline</td></tr>
<tr><td>XLM-R</td><td>RoBERTa</td><td>CC-100 (2.5TB)</td><td>100</td><td>Strong multilingual</td></tr>
<tr><td>mBERT</td><td>BERT</td><td>Wikipedia</td><td>104</td><td>Weakest on Vietnamese</td></tr>
</table>
<h3>SMTCE Benchmark (Best model per task)</h3>
<table>
<tr><th>Task</th><th>Best Model</th><th>Score</th><th>Runner-up</th></tr>
<tr><td>UIT-VSMEC (Emotion)</td><td class="highlight">PhoBERT</td><td>65.44% F1</td><td>viBERT4news</td></tr>
<tr><td>ViOCD (Complaint)</td><td class="highlight">vELECTRA</td><td>95.26% F1</td><td>PhoBERT</td></tr>
<tr><td>ViHSD (Hate Speech)</td><td class="highlight">PhoBERT</td><td>-</td><td>XLM-R</td></tr>
<tr><td>ViCTSD (Constructive)</td><td class="highlight">PhoBERT</td><td>-</td><td>vELECTRA</td></tr>
<tr><td>UIT-VSFC (Sentiment)</td><td class="highlight">PhoBERT</td><td>-</td><td>viBERT</td></tr>
</table>
<h3>Model Efficiency</h3>
<table>
<tr><th>Model</th><th>Size</th><th>VNTC Accuracy</th><th>Efficiency (Acc/MB)</th></tr>
<tr style="background:var(--surface2)"><td><strong>Sen-1</strong></td><td class="highlight">2.4 MB</td><td>92.49%</td><td class="highlight">38.5</td></tr>
<tr><td>PhoBERT-base</td><td>~400 MB</td><td>~95%</td><td>0.24</td></tr>
<tr><td>XLM-R-base</td><td>~1.1 GB</td><td>~93%</td><td>0.08</td></tr>
</table>
<p style="margin-top:0.75rem;font-size:0.9rem;color:var(--accent2)"><strong>Sen-1 is ~160x more efficient</strong> in accuracy-per-MB than PhoBERT.</p>
</div>
<!-- ===================== SOTA ===================== -->
<div class="panel" id="panel-sota">
<div class="section-title">State-of-the-Art</div>
<div class="section-desc">Current SOTA for Vietnamese text classification tasks (as of 2026).</div>
<table>
<tr><th>Task</th><th>Dataset</th><th>SOTA Model</th><th>Score</th><th>Paper</th></tr>
<tr><td>News Classification</td><td>VNTC</td><td>N-gram LM</td><td class="highlight">97.1% Acc</td><td>Vu et al. 2007</td></tr>
<tr><td>Emotion Recognition</td><td>UIT-VSMEC</td><td class="highlight">ViSoBERT</td><td>SOTA F1</td><td>Nguyen et al. 2023</td></tr>
<tr><td>Sentiment Analysis</td><td>UIT-VSFC</td><td class="highlight">PhoBERT</td><td>SOTA F1</td><td>SMTCE 2022</td></tr>
<tr><td>Hate Speech</td><td>ViHSD</td><td class="highlight">PhoBERT/ViSoBERT</td><td>SOTA F1</td><td>SMTCE/ViSoBERT</td></tr>
<tr><td>Complaint Detection</td><td>ViOCD</td><td class="highlight">vELECTRA</td><td>95.26% F1</td><td>SMTCE 2022</td></tr>
<tr><td>Spam Reviews</td><td>ViSpamReviews</td><td class="highlight">ViSoBERT</td><td>SOTA F1</td><td>Nguyen et al. 2023</td></tr>
</table>
<h3>Key Trends</h3>
<div class="cards-row">
<div class="info-card">
<h4>Monolingual > Multilingual</h4>
<p>PhoBERT, ViSoBERT, vELECTRA consistently outperform XLM-R, mBERT on Vietnamese tasks.</p>
</div>
<div class="info-card">
<h4>Domain-specific Pretraining</h4>
<p>ViSoBERT (social media) outperforms PhoBERT (general) on social media tasks.</p>
</div>
<div class="info-card">
<h4>Traditional ML Still Competitive</h4>
<p>TF-IDF + SVM achieves 92%+ on news classification with 160x less resources.</p>
</div>
<div class="info-card">
<h4>Word Segmentation Matters</h4>
<p>~5% accuracy gap between syllable-level (Sen-1) and word-level approaches.</p>
</div>
</div>
<h3>Sen-1 Position</h3>
<div class="network"><b> Accuracy</b>
High ^
| <b>PhoBERT</b> <u>ViSoBERT</u>
| * *
|
| <em>N-gram (2007)</em>
| *
| <b style="color:var(--accent4)">Sen-1</b>
| *
|
Low |
+-------------------------------&gt;
Fast Slow
Inference Speed</div>
<p style="margin-top:1rem;font-size:0.9rem;color:var(--text2)">
Sen-1 = <strong style="color:var(--accent4)">fast + lightweight</strong> quadrant: edge deployment, real-time batch processing, resource-constrained environments.
</p>
<h3>Open Questions</h3>
<div class="cards-row">
<div class="info-card">
<h4>RQ1</h4>
<p>Can word segmentation close the gap between Sen-1 and PhoBERT?</p>
</div>
<div class="info-card">
<h4>RQ2</h4>
<p>How does Sen-1 perform on social media/informal text?</p>
</div>
<div class="info-card">
<h4>RQ3</h4>
<p>Can ensemble (Sen-1 + lightweight transformer) get speed + accuracy?</p>
</div>
<div class="info-card">
<h4>RQ4</h4>
<p>Minimum dataset size where PhoBERT outperforms TF-IDF+SVM?</p>
</div>
</div>
</div>
<!-- ===================== CITATION NETWORK ===================== -->
<div class="panel" id="panel-network">
<div class="section-title">Citation Network</div>
<div class="section-desc">How the papers in this collection relate to each other and to Sen-1.</div>
<div class="network"><b>Vu et al. 2007</b> (VNTC dataset)
|
+---> Vietnamese text classification research
| |
| <b>RoBERTa</b> (2019) ---> <b>PhoBERT</b> (2020) ---> <u>ViSoBERT</u> (2023)
| | |
| <b>XLM-R</b> (2019) ------> <em>vELECTRA</em> (2020) <em>SMTCE benchmark</em> (2022)
| | |
| <b style="color:var(--accent4)">Sen-2</b> (future) UIT-VSMEC, UIT-VSFC
|
+---> <b style="color:var(--accent4)">Sen-1</b> (TF-IDF + SVM baseline)
|
+---> 92.49% VNTC | 75.76% UTS2017_Bank
|
+---> Phase 2: word segmentation, PhoBERT comparison
|
+---> Phase 3: <b style="color:var(--accent4)">Sen-2</b> (PhoBERT-based)</div>
<h3>Available Datasets</h3>
<table>
<tr><th>Dataset</th><th>Task</th><th>Samples</th><th>Classes</th><th>Domain</th><th>Source</th></tr>
<tr><td><strong>VNTC</strong></td><td>Topic</td><td>84,132</td><td>10</td><td>News</td><td><a href="https://github.com/duyvuleo/VNTC">GitHub</a></td></tr>
<tr><td><strong>UTS2017_Bank</strong></td><td>Intent</td><td>1,977</td><td>14</td><td>Banking</td><td>HuggingFace</td></tr>
<tr><td><strong>UIT-VSMEC</strong></td><td>Emotion</td><td>6,927</td><td>7</td><td>Social media</td><td>UIT NLP</td></tr>
<tr><td><strong>UIT-VSFC</strong></td><td>Sentiment</td><td>16,175</td><td>3</td><td>Education</td><td><a href="https://huggingface.co/datasets/uitnlp/vietnamese_students_feedback">HuggingFace</a></td></tr>
<tr><td><strong>SMTCE</strong></td><td>Multi-task</td><td>Multiple</td><td>Various</td><td>Social media</td><td><a href="https://arxiv.org/abs/2209.10482">arXiv</a></td></tr>
</table>
<h3>Research Gaps</h3>
<div class="cards-row">
<div class="info-card">
<h4>Gap 1</h4>
<p>No comprehensive TF-IDF vs PhoBERT comparison on same Vietnamese benchmarks with controlled experiments.</p>
</div>
<div class="info-card">
<h4>Gap 2</h4>
<p>Limited edge/resource-constrained deployment studies. Most work focuses on accuracy, not efficiency.</p>
</div>
<div class="info-card">
<h4>Gap 3</h4>
<p>Class imbalance handling for Vietnamese datasets is under-explored.</p>
</div>
<div class="info-card">
<h4>Gap 4</h4>
<p>Cross-domain evaluation and ablation studies for Vietnamese features are rare.</p>
</div>
</div>
</div>
<!-- ===================== LEADERBOARD ===================== -->
<div class="panel" id="panel-leaderboard">
<div class="lb-hero">
<h2>Vietnamese Text Classification Leaderboard</h2>
<p>Comprehensive comparison of models across Vietnamese NLP benchmarks, speed, and efficiency.</p>
<div class="lb-updated">Updated: February 2026 &middot; Inspired by <a href="https://www.vellum.ai/llm-leaderboard" target="_blank">Vellum LLM Leaderboard</a></div>
</div>
<!-- ===== Spotlight: Quality Benchmarks ===== -->
<div class="lb-section-title">Quality Benchmarks</div>
<div class="lb-section-desc">Top models per dataset, ranked by primary metric.</div>
<div class="lb-spotlight">
<!-- VNTC -->
<div class="lb-spot-card">
<h4>News Classification</h4>
<div class="lb-bench-name">VNTC (10 topics, 84K samples)</div>
<ol class="lb-rank-list">
<li><span class="lb-rank-num lb-rank-1">1</span><span class="lb-model-name">N-gram LM</span><span class="lb-score lb-score-top">97.1%</span></li>
<li><span class="lb-rank-num lb-rank-2">2</span><span class="lb-model-name">PhoBERT-base*</span><span class="lb-score">~95%</span></li>
<li><span class="lb-rank-num lb-rank-3">3</span><span class="lb-model-name">SVM Multi</span><span class="lb-score">93.4%</span></li>
<li><span class="lb-rank-num lb-rank-other">4</span><span class="lb-model-name">sonar_core_1</span><span class="lb-score">92.80%</span></li>
<li><span class="lb-rank-num lb-rank-other">5</span><span class="lb-model-name lb-sen1">Sen-1</span><span class="lb-score">92.49%</span></li>
</ol>
</div>
<!-- UTS2017_Bank -->
<div class="lb-spot-card">
<h4>Banking Classification</h4>
<div class="lb-bench-name">UTS2017_Bank (14 categories, 1.9K samples)</div>
<ol class="lb-rank-list">
<li><span class="lb-rank-num lb-rank-1">1</span><span class="lb-model-name lb-sen1">Sen-1</span><span class="lb-score lb-score-top">75.76%</span></li>
<li><span class="lb-rank-num lb-rank-2">2</span><span class="lb-model-name">sonar_core_1</span><span class="lb-score">72.47%</span></li>
</ol>
</div>
<!-- Emotion -->
<div class="lb-spot-card">
<h4>Emotion Recognition</h4>
<div class="lb-bench-name">UIT-VSMEC (7 classes, 6.9K samples)</div>
<ol class="lb-rank-list">
<li><span class="lb-rank-num lb-rank-1">1</span><span class="lb-model-name">ViSoBERT</span><span class="lb-score lb-score-top">SOTA</span></li>
<li><span class="lb-rank-num lb-rank-2">2</span><span class="lb-model-name">PhoBERT</span><span class="lb-score">65.44%</span></li>
<li><span class="lb-rank-num lb-rank-3">3</span><span class="lb-model-name">viBERT4news</span><span class="lb-score">-</span></li>
<li><span class="lb-rank-num lb-rank-other">4</span><span class="lb-model-name">CNN baseline</span><span class="lb-score">59.74%</span></li>
</ol>
</div>
<!-- Sentiment -->
<div class="lb-spot-card">
<h4>Sentiment Analysis</h4>
<div class="lb-bench-name">UIT-VSFC (3 classes, 16K samples)</div>
<ol class="lb-rank-list">
<li><span class="lb-rank-num lb-rank-1">1</span><span class="lb-model-name">PhoBERT</span><span class="lb-score lb-score-top">SOTA</span></li>
<li><span class="lb-rank-num lb-rank-2">2</span><span class="lb-model-name">viBERT</span><span class="lb-score">-</span></li>
<li><span class="lb-rank-num lb-rank-3">3</span><span class="lb-model-name">MaxEnt baseline</span><span class="lb-score">88%</span></li>
</ol>
</div>
<!-- Complaint -->
<div class="lb-spot-card">
<h4>Complaint Detection</h4>
<div class="lb-bench-name">ViOCD (SMTCE benchmark)</div>
<ol class="lb-rank-list">
<li><span class="lb-rank-num lb-rank-1">1</span><span class="lb-model-name">vELECTRA</span><span class="lb-score lb-score-top">95.26%</span></li>
<li><span class="lb-rank-num lb-rank-2">2</span><span class="lb-model-name">PhoBERT</span><span class="lb-score">-</span></li>
<li><span class="lb-rank-num lb-rank-3">3</span><span class="lb-model-name">XLM-R</span><span class="lb-score">-</span></li>
</ol>
</div>
<!-- Hate Speech -->
<div class="lb-spot-card">
<h4>Hate Speech Detection</h4>
<div class="lb-bench-name">ViHSD (SMTCE benchmark)</div>
<ol class="lb-rank-list">
<li><span class="lb-rank-num lb-rank-1">1</span><span class="lb-model-name">PhoBERT</span><span class="lb-score lb-score-top">SOTA</span></li>
<li><span class="lb-rank-num lb-rank-2">2</span><span class="lb-model-name">ViSoBERT</span><span class="lb-score">-</span></li>
<li><span class="lb-rank-num lb-rank-3">3</span><span class="lb-model-name">XLM-R</span><span class="lb-score">-</span></li>
</ol>
</div>
</div>
<!-- ===== Spotlight: Performance ===== -->
<div class="lb-section-title">Performance Metrics</div>
<div class="lb-section-desc">Speed, latency, and efficiency rankings.</div>
<div class="lb-spotlight">
<!-- Fastest -->
<div class="lb-spot-card">
<h4>Fastest Inference</h4>
<div class="lb-bench-name">Batch throughput (samples/sec)</div>
<ol class="lb-rank-list">
<li><span class="lb-rank-num lb-rank-1">1</span><span class="lb-model-name lb-sen1">Sen-1</span><span class="lb-score lb-score-top">66,678/s</span></li>
<li><span class="lb-rank-num lb-rank-2">2</span><span class="lb-model-name">TF-IDF + SVM (sklearn)</span><span class="lb-score">~50K/s</span></li>
<li><span class="lb-rank-num lb-rank-3">3</span><span class="lb-model-name">PhoBERT (GPU)</span><span class="lb-score">~20/s</span></li>
</ol>
</div>
<!-- Smallest -->
<div class="lb-spot-card">
<h4>Smallest Model</h4>
<div class="lb-bench-name">Model file size</div>
<ol class="lb-rank-list">
<li><span class="lb-rank-num lb-rank-1">1</span><span class="lb-model-name lb-sen1">Sen-1</span><span class="lb-score lb-score-top">2.4 MB</span></li>
<li><span class="lb-rank-num lb-rank-2">2</span><span class="lb-model-name">sonar_core_1</span><span class="lb-score">~75 MB</span></li>
<li><span class="lb-rank-num lb-rank-3">3</span><span class="lb-model-name">PhoBERT-base</span><span class="lb-score">~400 MB</span></li>
<li><span class="lb-rank-num lb-rank-other">4</span><span class="lb-model-name">XLM-R-base</span><span class="lb-score">~1.1 GB</span></li>
</ol>
</div>
<!-- Most Efficient -->
<div class="lb-spot-card">
<h4>Most Efficient</h4>
<div class="lb-bench-name">Accuracy per MB (VNTC)</div>
<ol class="lb-rank-list">
<li><span class="lb-rank-num lb-rank-1">1</span><span class="lb-model-name lb-sen1">Sen-1</span><span class="lb-score lb-score-top">38.5</span></li>
<li><span class="lb-rank-num lb-rank-2">2</span><span class="lb-model-name">sonar_core_1</span><span class="lb-score">1.24</span></li>
<li><span class="lb-rank-num lb-rank-3">3</span><span class="lb-model-name">PhoBERT-base</span><span class="lb-score">0.24</span></li>
<li><span class="lb-rank-num lb-rank-other">4</span><span class="lb-model-name">XLM-R-base</span><span class="lb-score">0.08</span></li>
</ol>
</div>
<!-- Fastest Training -->
<div class="lb-spot-card">
<h4>Fastest Training</h4>
<div class="lb-bench-name">VNTC full training time</div>
<ol class="lb-rank-list">
<li><span class="lb-rank-num lb-rank-1">1</span><span class="lb-model-name lb-sen1">Sen-1 (Rust)</span><span class="lb-score lb-score-top">37.6s</span></li>
<li><span class="lb-rank-num lb-rank-2">2</span><span class="lb-model-name">TF-IDF+SVM (sklearn)</span><span class="lb-score">~2 min</span></li>
<li><span class="lb-rank-num lb-rank-3">3</span><span class="lb-model-name">sonar_core_1</span><span class="lb-score">54.6 min</span></li>
<li><span class="lb-rank-num lb-rank-other">4</span><span class="lb-model-name">N-gram LM</span><span class="lb-score">~79 min</span></li>
<li><span class="lb-rank-num lb-rank-other">5</span><span class="lb-model-name">PhoBERT fine-tune</span><span class="lb-score">Hours</span></li>
</ol>
</div>
</div>
<!-- ===== Comprehensive Table ===== -->
<div class="lb-section-title">Comprehensive Comparison</div>
<div class="lb-section-desc">All models with operational and benchmark metrics. Click column headers to sort.</div>
<div style="display:flex;gap:0.5rem;margin-bottom:1rem;flex-wrap:wrap">
<span style="display:inline-flex;align-items:center;gap:0.35rem;font-size:0.78rem;color:var(--text2)"><span class="model-type type-traditional" style="width:10px;height:10px;display:inline-block"></span> Traditional ML</span>
<span style="display:inline-flex;align-items:center;gap:0.35rem;font-size:0.78rem;color:var(--text2)"><span class="model-type type-transformer" style="width:10px;height:10px;display:inline-block"></span> Vietnamese Transformer</span>
<span style="display:inline-flex;align-items:center;gap:0.35rem;font-size:0.78rem;color:var(--text2)"><span class="model-type type-multilingual" style="width:10px;height:10px;display:inline-block"></span> Multilingual</span>
</div>
<div class="lb-table-wrap">
<table class="lb-table" id="lb-main-table">
<thead>
<tr>
<th data-col="0" data-type="num">#</th>
<th data-col="1" data-type="str">Model</th>
<th data-col="2" data-type="str">Type</th>
<th data-col="3" data-type="str">Architecture</th>
<th data-col="4" data-type="num">Size</th>
<th data-col="5" data-type="num">VNTC<br>Acc %</th>
<th data-col="6" data-type="num">UTS2017<br>Acc %</th>
<th data-col="7" data-type="num">UIT-VSMEC<br>F1 %</th>
<th data-col="8" data-type="num">ViOCD<br>F1 %</th>
<th data-col="9" data-type="num">Training</th>
<th data-col="10" data-type="num">Inference<br>/sec</th>
<th data-col="11" data-type="num">Eff.<br>Acc/MB</th>
</tr>
</thead>
<tbody>
<tr><td>1</td><td><div class="model-cell"><span class="model-type type-traditional"></span>N-gram LM</div></td><td>Traditional</td><td>N-gram Language Model</td><td class="cell-na">n/a</td><td class="cell-best">97.1</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td>~79 min</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td></tr>
<tr><td>2</td><td><div class="model-cell"><span class="model-type type-transformer"></span>PhoBERT-base</div></td><td>Transformer</td><td>RoBERTa (20GB vi)</td><td>~400 MB</td><td>~95</td><td class="cell-na">n/a</td><td>65.44</td><td class="cell-na">n/a</td><td>Hours (GPU)</td><td>~20</td><td>0.24</td></tr>
<tr><td>3</td><td><div class="model-cell"><span class="model-type type-transformer"></span>ViSoBERT</div></td><td>Transformer</td><td>XLM-R (social media)</td><td>~400 MB</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-best">SOTA</td><td class="cell-na">n/a</td><td>Hours (GPU)</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td></tr>
<tr><td>4</td><td><div class="model-cell"><span class="model-type type-transformer"></span>vELECTRA</div></td><td>Transformer</td><td>ELECTRA (60GB vi)</td><td>~400 MB</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-best">95.26</td><td>Hours (GPU)</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td></tr>
<tr><td>5</td><td><div class="model-cell"><span class="model-type type-traditional"></span>SVM Multi</div></td><td>Traditional</td><td>SVM + BOW</td><td class="cell-na">n/a</td><td>93.4</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td>~79 min</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td></tr>
<tr><td>6</td><td><div class="model-cell"><span class="model-type type-traditional"></span>sonar_core_1</div></td><td>Traditional</td><td>TF-IDF + SVC (RBF)</td><td>~75 MB</td><td>92.80</td><td>72.47</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td>54.6 min</td><td class="cell-na">n/a</td><td>1.24</td></tr>
<tr class="lb-row-sen1"><td>7</td><td><div class="model-cell"><span class="model-type type-traditional"></span><strong style="color:var(--accent4)">Sen-1</strong></div></td><td>Traditional</td><td>TF-IDF + LinearSVC (Rust)</td><td class="cell-best">2.4 MB</td><td>92.49</td><td class="cell-best">75.76</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-best">37.6s</td><td class="cell-best">66,678</td><td class="cell-best">38.5</td></tr>
<tr><td>8</td><td><div class="model-cell"><span class="model-type type-multilingual"></span>XLM-R-base</div></td><td>Multilingual</td><td>RoBERTa (100 langs)</td><td>~1.1 GB</td><td>~93</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td>Hours (GPU)</td><td class="cell-na">n/a</td><td>0.08</td></tr>
<tr><td>9</td><td><div class="model-cell"><span class="model-type type-multilingual"></span>mBERT</div></td><td>Multilingual</td><td>BERT (104 langs)</td><td>~700 MB</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td>Hours (GPU)</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td></tr>
<tr><td>10</td><td><div class="model-cell"><span class="model-type type-transformer"></span>viBERT</div></td><td>Transformer</td><td>BERT (10GB vi)</td><td>~400 MB</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td>Hours (GPU)</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td></tr>
<tr><td>11</td><td><div class="model-cell"><span class="model-type type-transformer"></span>viBERT4news</div></td><td>Transformer</td><td>BERT (news domain)</td><td>~400 MB</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td>Hours (GPU)</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td></tr>
<tr><td>12</td><td><div class="model-cell"><span class="model-type type-traditional"></span>MaxEnt baseline</div></td><td>Traditional</td><td>Maximum Entropy</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td></tr>
<tr><td>13</td><td><div class="model-cell"><span class="model-type type-traditional"></span>CNN baseline</div></td><td>Traditional</td><td>Convolutional NN</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td>59.74</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td></tr>
<tr><td>14</td><td><div class="model-cell"><span class="model-type type-multilingual"></span>DistilmBERT</div></td><td>Multilingual</td><td>DistilBERT (multilingual)</td><td>~260 MB</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td><td>Hours (GPU)</td><td class="cell-na">n/a</td><td class="cell-na">n/a</td></tr>
</tbody>
</table>
</div>
<p style="font-size:0.75rem;color:var(--text2);margin-top:0.5rem">
* PhoBERT VNTC estimate based on similar Vietnamese classification tasks. Blank cells indicate benchmark not evaluated.
<br>Efficiency = VNTC Accuracy / Model Size in MB. Higher is better.
</p>
</div>
<!-- ===================== HOW TO DO SCIENCE ===================== -->
<div class="panel" id="panel-science">
<div class="section-title">How to Do Science</div>
<div class="section-desc">Research methodology guide compiled from Hamming, Schulman, Marek Rei, and Microsoft Research Asia.</div>
<h3>1. Choosing Important Problems</h3>
<blockquote>"If you do not work on an important problem, it's unlikely you'll do important work." <strong>-- Richard Hamming</strong></blockquote>
<div class="principle">
<h4>Hamming's Principles</h4>
<ul>
<li>Maintain a list of <strong>10-20 important problems</strong> in your field</li>
<li>A problem becomes important when you have a <strong>reasonable attack</strong></li>
<li>Dedicate deep thinking time (Friday "Great Thoughts Time")</li>
<li>Have <strong>courage</strong> to pursue unconventional ideas</li>
</ul>
</div>
<div class="principle">
<h4>Schulman's Framework (OpenAI)</h4>
<ul>
<li><strong>Work on the right problems</strong></li>
<li><strong>Make continual progress</strong></li>
<li><strong>Achieve continual personal growth</strong></li>
</ul>
<p style="margin-top:0.5rem">Develop "research taste" by reading broadly, collaborating widely, and asking: "If this succeeds, how big is the impact?"</p>
</div>
<div class="principle">
<h4>Microsoft Research Asia (Dr. Ming Zhou)</h4>
<ul>
<li>Read recent <strong>ACL proceedings</strong> to find your field</li>
<li>Target <strong>"blue ocean"</strong> areas - new fields with less competition</li>
<li>Verify 3 prerequisites: math/ML framework, standard datasets, active research teams</li>
<li>Find <strong>gaps</strong>: what can be improved, combined, or inverted</li>
</ul>
</div>
<h3>2. Reading Papers</h3>
<div class="principle">
<h4>Effective Reading Strategy</h4>
<ul>
<li><strong>Read broadly</strong>: Not just NLP - also cognitive science, neuroscience, linguistics, vision</li>
<li><strong>Read deeply</strong>: Become the "world-leading expert" on your narrow question</li>
<li><strong>Read textbooks</strong>: More knowledge-dense than papers</li>
<li><strong>Follow citation chains</strong> via Google Scholar, Semantic Scholar</li>
<li>Use <strong>PRISMA</strong> methodology for systematic reviews</li>
</ul>
</div>
<h3>3. Running Experiments</h3>
<div class="principle">
<h4>Step 1: Reproduce baselines first</h4>
<p>"Reimplement existing state-of-the-art work first to validate your setup." -- Marek Rei</p>
<ul>
<li>Choose open source project, compile, run demo, <strong>match results</strong></li>
<li>Understand the algorithm deeply, then reimplement</li>
<li>Test on standard test set until results align</li>
</ul>
</div>
<div class="principle">
<h4>Step 2: Simple baseline (1-2 weeks)</h4>
<p>Implement the simplest approach before building complex architectures. Verify your setup works.</p>
</div>
<div class="principle">
<h4>Step 3: Rigorous experimentation</h4>
<ul>
<li><strong>Debug</strong>: Don't assume bug-free code. Test with toy examples. Add assertions.</li>
<li><strong>Evaluate</strong>: Separate train/dev/test. Run 10+ times. Report mean + std.</li>
<li><strong>Ablate</strong>: Significance tests and ablation studies for every novel component.</li>
<li><strong>Avoid</strong>: Single-run results, weak-only baselines, blind trend-following.</li>
</ul>
</div>
<h3>4. Writing Papers</h3>
<div class="principle">
<h4>ACL Paper Structure (Dr. Ming Zhou)</h4>
<ul>
<li><strong>Title</strong>: Specific, no generic words</li>
<li><strong>Abstract</strong>: Problem + Method + Advantage + Achievement</li>
<li><strong>Introduction</strong>: Background &rarr; existing &rarr; limitations &rarr; contribution (&le;3 points)</li>
<li><strong>Related Work</strong>: Organized by <strong>theme</strong>, not chronology</li>
<li><strong>Methodology</strong>: Problem definition &rarr; notation &rarr; formulas</li>
<li><strong>Experiments</strong>: Purpose &rarr; data &rarr; parameters &rarr; reproducibility</li>
<li><strong>Limitations</strong>: Required by ACL - honest assessment</li>
</ul>
<p style="margin-top:0.5rem"><strong>Revision</strong>: 3 passes - self review &rarr; team review &rarr; outsider review.</p>
</div>
<h3>5. Mindset & Habits</h3>
<table>
<tr><th>Principle</th><th>Lesson</th></tr>
<tr><td><strong>Open doors</strong></td><td>Stay connected to the community; know emerging problems</td></tr>
<tr><td><strong>Preparation</strong></td><td>"Luck favors the prepared mind" (Pasteur)</td></tr>
<tr><td><strong>Constraints</strong></td><td>Difficult conditions often lead to breakthroughs</td></tr>
<tr><td><strong>Commitment</strong></td><td>Deep immersion activates subconscious problem-solving</td></tr>
<tr><td><strong>Selling work</strong></td><td>Presentation matters - great work needs effective communication</td></tr>
</table>
<h3>Essential Reading</h3>
<div class="papers-grid" style="margin-top:0.75rem">
<div class="paper-card">
<div class="paper-title"><a href="https://www.cs.virginia.edu/~robins/YouAndYourResearch.html" target="_blank">You and Your Research</a></div>
<div class="paper-meta"><span>Richard Hamming</span> &middot; <span>Choosing important problems, mindset</span></div>
</div>
<div class="paper-card">
<div class="paper-title"><a href="http://joschu.net/blog/opinionated-guide-ml-research.html" target="_blank">An Opinionated Guide to ML Research</a></div>
<div class="paper-meta"><span>John Schulman (OpenAI)</span> &middot; <span>Problem selection, progress, growth</span></div>
</div>
<div class="paper-card">
<div class="paper-title"><a href="https://www.marekrei.com/blog/ml-nlp-research-project-advice/" target="_blank">ML/NLP Research Project Advice</a></div>
<div class="paper-meta"><span>Marek Rei</span> &middot; <span>Practical experiment workflow</span></div>
</div>
<div class="paper-card">
<div class="paper-title"><a href="https://microsoft.com/en-us/research/lab/microsoft-research-asia/articles/make-first-accomplishment-nlp-field/" target="_blank">How to Make First Accomplishment in NLP</a></div>
<div class="paper-meta"><span>Dr. Ming Zhou (MSRA)</span> &middot; <span>NLP research methodology</span></div>
</div>
</div>
</div>
</div><!-- /content -->
<script>
// Tab switching
document.querySelectorAll('.tab').forEach(tab => {
tab.addEventListener('click', () => {
document.querySelectorAll('.tab').forEach(t => t.classList.remove('active'));
document.querySelectorAll('.panel').forEach(p => p.classList.remove('active'));
tab.classList.add('active');
document.getElementById('panel-' + tab.dataset.tab).classList.add('active');
});
});
// Category filters
document.querySelectorAll('.filter-btn').forEach(btn => {
btn.addEventListener('click', () => {
document.querySelectorAll('.filter-btn').forEach(b => b.classList.remove('active'));
btn.classList.add('active');
const cat = btn.dataset.cat;
document.querySelectorAll('.paper-card[data-cats]').forEach(card => {
if (cat === 'all' || card.dataset.cats === cat) {
card.style.display = '';
} else {
card.style.display = 'none';
}
});
});
});
// Abstract expand/collapse
document.querySelectorAll('.paper-abstract').forEach(abs => {
abs.addEventListener('click', () => {
abs.classList.toggle('expanded');
});
abs.style.cursor = 'pointer';
});
// Leaderboard table sorting
(function() {
const table = document.getElementById('lb-main-table');
if (!table) return;
const thead = table.querySelector('thead');
const tbody = table.querySelector('tbody');
let sortCol = -1, sortAsc = true;
function parseVal(td) {
let txt = td.textContent.trim();
if (txt === 'n/a' || txt === '-' || txt === 'SOTA') return -Infinity;
txt = txt.replace(/[~,/s]/g, '').replace('MB', '').replace('GB', function() { return ''; });
// Handle GB -> convert
let raw = td.textContent.trim();
if (raw.includes('GB')) {
let n = parseFloat(raw.replace(/[~]/g, ''));
return n * 1000;
}
if (raw.includes('MB')) return parseFloat(raw.replace(/[~]/g, ''));
if (raw.includes('min')) return parseFloat(raw) * 60;
if (raw.includes('Hours')) return 36000;
if (raw.endsWith('s') && !raw.includes('/s')) return parseFloat(raw);
let num = parseFloat(txt.replace(/[^0-9.\-]/g, ''));
return isNaN(num) ? -Infinity : num;
}
thead.querySelectorAll('th').forEach((th, i) => {
th.addEventListener('click', () => {
if (sortCol === i) { sortAsc = !sortAsc; }
else { sortCol = i; sortAsc = false; }
const rows = Array.from(tbody.querySelectorAll('tr'));
rows.sort((a, b) => {
let va = parseVal(a.children[i]);
let vb = parseVal(b.children[i]);
if (th.dataset.type === 'str') {
va = a.children[i].textContent.trim().toLowerCase();
vb = b.children[i].textContent.trim().toLowerCase();
return sortAsc ? va.localeCompare(vb) : vb.localeCompare(va);
}
return sortAsc ? va - vb : vb - va;
});
// Re-number
rows.forEach((r, idx) => { r.children[0].textContent = idx + 1; tbody.appendChild(r); });
// Update arrows
thead.querySelectorAll('.sort-arrow').forEach(a => a.remove());
const arrow = document.createElement('span');
arrow.className = 'sort-arrow';
arrow.textContent = sortAsc ? '\u25B2' : '\u25BC';
th.appendChild(arrow);
});
});
})();
</script>
</body>
</html>