Datasets:
Update datacard
Browse files
README.md
CHANGED
|
@@ -21,8 +21,8 @@ license: cc-by-nc-sa-4.0
|
|
| 21 |
|
| 22 |
### Dataset Summary
|
| 23 |
|
| 24 |
-
The CA-ZH Parallel Corpus is a Catalan-Chinese dataset of
|
| 25 |
-
Machine Translation.
|
| 26 |
|
| 27 |
### Supported Tasks and Leaderboards
|
| 28 |
|
|
@@ -63,23 +63,11 @@ This dataset is aimed at promoting the development of Machine Translation betwee
|
|
| 63 |
|
| 64 |
#### Initial Data Collection and Normalization
|
| 65 |
|
| 66 |
-
The
|
|
|
|
| 67 |
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
| WikiMatrix | 90.643 |
|
| 71 |
-
| XLENT | 535.803 |
|
| 72 |
-
| GNOME | 78|
|
| 73 |
-
| OpenSubtitles | 139.300 |
|
| 74 |
-
|
| 75 |
-
The 6.658.607 sentence pairs of synthetic parallel data were created from the following Spanish-Chinese datasets:
|
| 76 |
-
|
| 77 |
-
| Dataset | Sentences before cleaning |
|
| 78 |
-
|:------------------|---------------:|
|
| 79 |
-
| UNPC |17.599.223|
|
| 80 |
-
| CCMatrix | 24.051.233 |
|
| 81 |
-
| MultiParacrawl| 3.410.087|
|
| 82 |
-
| **Total** | **45.060.543** |
|
| 83 |
|
| 84 |
### Data preparation
|
| 85 |
|
|
@@ -87,7 +75,7 @@ The Chinese side of all datasets are passed through the [fastlangid](https://git
|
|
| 87 |
identified as simplified Chinese are discarded.
|
| 88 |
The datasets are then also deduplicated and filtered to remove any sentence pairs with a cosine similarity of less than 0.75.
|
| 89 |
This is done using sentence embeddings calculated using [LaBSE](https://huggingface.co/sentence-transformers/LaBSE).
|
| 90 |
-
The filtered datasets are then concatenated to form
|
| 91 |
|
| 92 |
#### Who are the source language producers?
|
| 93 |
|
|
|
|
| 21 |
|
| 22 |
### Dataset Summary
|
| 23 |
|
| 24 |
+
The CA-ZH Parallel Corpus is a Catalan-Chinese dataset of parallel sentences created to
|
| 25 |
+
support Catalan in NLP tasks, specifically Machine Translation.
|
| 26 |
|
| 27 |
### Supported Tasks and Leaderboards
|
| 28 |
|
|
|
|
| 63 |
|
| 64 |
#### Initial Data Collection and Normalization
|
| 65 |
|
| 66 |
+
The first portion of the corpus is a combination of the following original datasets collected from [Opus](https://opus.nlpl.eu/):
|
| 67 |
+
WikiMatrix, XLENT, GNOME, OpenSubtitles.
|
| 68 |
|
| 69 |
+
Additionally, the corpus contains synthetic parallel data generated from the following original Spanish-Chinese datasets collected from [Opus](https://opus.nlpl.eu/):
|
| 70 |
+
UNPC, CCMatrix, MultiParacrawl.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 71 |
|
| 72 |
### Data preparation
|
| 73 |
|
|
|
|
| 75 |
identified as simplified Chinese are discarded.
|
| 76 |
The datasets are then also deduplicated and filtered to remove any sentence pairs with a cosine similarity of less than 0.75.
|
| 77 |
This is done using sentence embeddings calculated using [LaBSE](https://huggingface.co/sentence-transformers/LaBSE).
|
| 78 |
+
The filtered datasets are then concatenated to form the final corpus.
|
| 79 |
|
| 80 |
#### Who are the source language producers?
|
| 81 |
|