Update README.md
Browse files
README.md
CHANGED
@@ -47,3 +47,38 @@ tags:
|
|
47 |
- parallel-corpus
|
48 |
- document-alignment
|
49 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
- parallel-corpus
|
48 |
- document-alignment
|
49 |
---
|
50 |
+
|
51 |
+
# Dataset Card for Pralekha
|
52 |
+
|
53 |
+
## Dataset Description
|
54 |
+
|
55 |
+
**Pralekha** is a large-scale benchmark for evaluating document-level alignment techniques, specifically designed for cross-lingual parallel data mining. The dataset contains over **2 million documents**, covering **11 Indic languages** and **English**, with a balanced mix of aligned and unaligned document pairs. It provides a foundational resource for exploring high-quality parallel data mining and supporting long-context cross-lingual applications like neural machine translation (NMT).
|
56 |
+
|
57 |
+
---
|
58 |
+
|
59 |
+
## Dataset Usage
|
60 |
+
|
61 |
+
Pralekha serves as a **benchmark dataset to evaluate document alignment approaches across various dimensions**, including embedding models, granularity levels (sentence, chunk, and document), and alignment algorithms. It enables the development of optimal strategies for mining high-quality parallel document pairs from large, noisy, web-crawled corpora, specifically for Indic languages and English.
|
62 |
+
|
63 |
+
---
|
64 |
+
|
65 |
+
## Dataset Statistics
|
66 |
+
|
67 |
+
| Split | Examples | Size (bytes) |
|
68 |
+
|-------------|----------|----------------|
|
69 |
+
| `aligned` | 1,566,404 | 10,274,361,211 |
|
70 |
+
| `unaligned` | 783,197 | 4,466,506,637 |
|
71 |
+
| **Total** | 2,349,601 | 14,740,867,848 |
|
72 |
+
|
73 |
+
---
|
74 |
+
|
75 |
+
## Dataset Structure
|
76 |
+
|
77 |
+
The dataset is split into `aligned` and `unaligned` subsets. Each entry follows the format:
|
78 |
+
```yaml
|
79 |
+
{
|
80 |
+
"n_id": "unique_parallel_id",
|
81 |
+
"doc_id": "unique_document_id",
|
82 |
+
"lang": "ISO-3_language_code",
|
83 |
+
"text": "textual_content"
|
84 |
+
}
|