AtlasNLP/analysis.html at main · MichiganNLP/AtlasNLP · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
<!DOCTYPE html>
<html lang="en">
<head>
  <meta charset="UTF-8">
  <meta name="viewport" content="width=device-width, initial-scale=1.0">
  <title>Analysis — AtlasNLP</title>
  <meta name="description" content="Key findings from AtlasNLP: geographic concentration, task sparsity, producer vs. content gaps, and more.">
  <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/@tabler/icons-webfont@2.44.0/dist/tabler-icons.min.css">
  <link rel="stylesheet" href="style.css">
  <style>
    .finding-chart-wrap { position: relative; height: 380px; }
    .finding-chart-wrap.tall { height: 460px; }
  </style>
</head>
<body class="pg">

<!-- NAV -->
<nav class="nav">
  <a href="index.html" class="logo">
    <div class="logo-mark"><i class="ti ti-world" aria-hidden="true"></i></div>
    AtlasNLP
  </a>
  <button class="hamburger" aria-label="Toggle menu">
    <span></span><span></span><span></span>
  </button>
  <div class="nav-links">
    <a href="index.html" class="nl">Home</a>
    <a href="datasets.html" class="nl">Datasets</a>
    <a href="analysis.html" class="nl">Analysis</a>
    <a href="visualizations.html" class="nl">Visualizations</a>
    <button class="nav-cta" onclick="window.location.href='datasets.html'">
      <i class="ti ti-download" aria-hidden="true"></i> Download
    </button>
  </div>
</nav>

<!-- HERO -->
<div class="hero-page">
  <span class="eyebrow">Key Findings</span>
  <h1>What the Data Reveals</h1>
  <p>Five findings from AtlasNLP — aggregated from 18,035 datasets across the Core set and 1,661 entries in the Gold set.</p>
</div>

<!-- FINDINGS -->
<div class="section" id="findings">

  <!-- Finding 1 -->
  <div class="finding-split">
    <div class="finding-prose">
      <span class="finding-tag">Finding 1</span>
      <h2>NLP Dataset Production Is Geographically Concentrated</h2>
      <div class="finding-stat-big">39%</div>
      <p>of all datasets are concentrated in the top 5 countries by content coverage.</p>
      <p>A small number of countries — primarily the United States, China, and a handful of European nations — dominate NLP dataset production. This concentration is visible across every task category.</p>
      <p>The chart shows the top 20 content countries by dataset count in AtlasNLP Core. The drop-off from the leading country to the rest is steep.</p>
    </div>
    <div>
      <div class="finding-chart-wrap tall" id="chart1-wrap">
        <div class="spinner-wrap"><div class="spinner"></div><span>Loading…</span></div>
      </div>
    </div>
  </div>

  <!-- Finding 2 -->
  <div class="finding-split flipped">
    <div class="finding-prose">
      <span class="finding-tag">Finding 2</span>
      <h2>Task Coverage Is Uneven Across the Atlas</h2>
      <p>Certain task categories — especially machine translation, language modeling, and named entity recognition — attract far more datasets than others. Many specialized tasks relevant to low-resource contexts are underrepresented.</p>
      <p>The chart shows the 15 most populated task categories in AtlasNLP Core, ranked by dataset count.</p>
    </div>
    <div>
      <div class="finding-chart-wrap tall" id="chart2-wrap">
        <div class="spinner-wrap"><div class="spinner"></div><span>Loading…</span></div>
      </div>
    </div>
  </div>

  <!-- Finding 3 -->
  <div class="finding-split">
    <div class="finding-prose">
      <span class="finding-tag">Finding 3</span>
      <h2>Producer and Content Countries Often Diverge</h2>
      <p>A dataset about a country's language is not necessarily produced by researchers from that country. AtlasNLP tracks both <em>content countries</em> (whose language or community is covered) and <em>producer countries</em> (where the authors are based).</p>
      <p>For many nations in Africa, Southeast Asia, and South America — content coverage exceeds producer presence, meaning their languages are studied largely by outsiders.</p>
    </div>
    <div>
      <div class="finding-chart-wrap" id="chart3-wrap">
        <div class="spinner-wrap"><div class="spinner"></div><span>Loading…</span></div>
      </div>
    </div>
  </div>

  <!-- Finding 4 -->
  <div class="finding-split flipped">
    <div class="finding-prose">
      <span class="finding-tag">Finding 4</span>
      <h2>Most NLP Datasets Are Monolingual</h2>
      <div class="finding-stat-big">75%</div>
      <p>of Core datasets target a single language, with no cross-lingual scope.</p>
      <p>Despite growing interest in multilingual NLP, the field remains dominated by monolingual datasets. Multilingual datasets are disproportionately produced by high-resource countries and tend to include English as a primary language, reinforcing the geographic skew documented in other findings.</p>
    </div>
    <div>
      <div class="finding-chart-wrap" id="chart4-wrap">
        <div class="spinner-wrap"><div class="spinner"></div><span>Loading…</span></div>
      </div>
    </div>
  </div>

  <!-- Finding 5 -->
  <div class="finding-split">
    <div class="finding-prose">
      <span class="finding-tag">Finding 5</span>
      <h2>Country Attribution Relies on Multiple Evidence Tiers</h2>
      <div class="finding-stat-big">33.8%</div>
      <p>of datasets lack recoverable country attribution through any method.</p>
      <p>AtlasNLP uses a cascade of attribution methods: explicit metadata, language-to-country inference, author affiliation, and URL domain inference. A significant portion of datasets could not be attributed to any country even with all methods applied.</p>
    </div>
    <div>
      <div class="finding-chart-wrap tall" id="chart5-wrap">
        <div class="spinner-wrap"><div class="spinner"></div><span>Loading…</span></div>
      </div>
    </div>
  </div>


  <!-- Task Portfolio -->
  <div class="finding-split">
    <div class="finding-prose">
      <span class="finding-tag">Task Portfolio</span>
      <h2>Task Breadth by Country</h2>
      <p>How many distinct NLP task categories does each country appear in? A country concentrated in one or two tasks has a narrower portfolio than one with the same number of datasets spread across many.</p>
      <p>The table shows the top 20 countries by dataset count, their task breadth, and the combined share of their three most common tasks.</p>
    </div>
    <div>
      <div id="chart6-wrap" class="finding-chart-wrap tall">
        <div class="spinner-wrap"><div class="spinner"></div><span>Loading…</span></div>
      </div>
    </div>
  </div>

  <!-- Language Concentration -->
  <div class="finding-split flipped">
    <div class="finding-prose">
      <span class="finding-tag">Language Concentration</span>
      <h2>Which Countries Dominate Each Language?</h2>
      <p>For the top languages by dataset coverage, this chart shows what share of datasets come from the top countries. It reveals whether a language's NLP resources are concentrated in one place or distributed globally.</p>
    </div>
    <div>
      <div id="chart7-wrap" class="finding-chart-wrap tall">
        <div class="spinner-wrap"><div class="spinner"></div><span>Loading…</span></div>
      </div>
    </div>
  </div>

</div>

<!-- CTA -->
<div class="cta-section">
  <h2>Explore the Data Yourself</h2>
  <p>Browse the full dataset index, filter by country or language, and download subsets for your own research.</p>
  <div class="cta-btns" style="margin-top:1.25rem;">
    <a href="datasets.html" class="btn-p"><i class="ti ti-table" aria-hidden="true"></i> Explore Datasets</a>
    <a href="visualizations.html" class="btn-o"><i class="ti ti-chart-dots" aria-hidden="true"></i> Interactive Visualizations</a>
  </div>
</div>

<!-- FOOTER -->
<footer class="footer">
  <span class="ft">AtlasNLP · A country-aware atlas of NLP dataset representation</span>
  <div class="ft-links">
    <a href="index.html" class="ft-link">Home</a>
    <a href="datasets.html" class="ft-link">Datasets</a>
    <a href="analysis.html" class="ft-link">Analysis</a>
    <a href="visualizations.html" class="ft-link">Visualizations</a>
  </div>
</footer>

<script src="https://cdnjs.cloudflare.com/ajax/libs/PapaParse/5.4.1/papaparse.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/Chart.js/4.4.0/chart.umd.min.js"></script>
<script src="main.js"></script>
<script src="analysis.js"></script>
</body>
</html>