Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -59,9 +59,9 @@ The texts in the dataset are in Catalan and Portuguese.
|
|
59 |
|
60 |
Two separated txt files are provided with the sentences sorted in the same order:
|
61 |
|
62 |
-
- ca-pt_2023_09_01_full.ca: contains
|
63 |
|
64 |
-
- ca-pt_2023_09_01_full.pt: contains
|
65 |
|
66 |
### Data Splits
|
67 |
|
@@ -73,25 +73,29 @@ The dataset contains a single split: `train`.
|
|
73 |
|
74 |
The dataset is a combination of the following datasets:
|
75 |
|
76 |
-
| Dataset | Sentences
|
77 |
-
|
78 |
-
| CCMatrix v1 |
|
79 |
-
| WikiMatrix
|
80 |
-
| GNOME
|
81 |
-
| KDE4
|
82 |
-
| QED
|
83 |
-
| TED2020 v1 |
|
84 |
-
| OpenSubtitles
|
85 |
-
| GlobalVoices|
|
86 |
-
| Tatoeba
|
87 |
-
| Europarl
|
88 |
-
| **Total** | **
|
89 |
-
|
90 |
-
All corpora except Europarl were collected from [Opus](https://opus.nlpl.eu/).
|
|
|
|
|
91 |
|
92 |
### Data preparation
|
93 |
|
94 |
-
All datasets are deduplicated and filtered to remove any sentence pairs with a cosine similarity of less than 0.75.
|
|
|
|
|
95 |
|
96 |
### Personal and Sensitive Information
|
97 |
|
@@ -105,7 +109,8 @@ The purpose of this dataset is to help develop Machines Translation tasks for lo
|
|
105 |
|
106 |
### Discussion of Biases
|
107 |
|
108 |
-
We are aware that since part of the data comes from unreliable web pages and non-curated texts, some biases may be present in the dataset.
|
|
|
109 |
|
110 |
### Other Known Limitations
|
111 |
|
|
|
59 |
|
60 |
Two separated txt files are provided with the sentences sorted in the same order:
|
61 |
|
62 |
+
- ca-pt_2023_09_01_full.ca: contains 9 892 953 Catalan sentences
|
63 |
|
64 |
+
- ca-pt_2023_09_01_full.pt: contains 9 892 953 Portuguese sentences
|
65 |
|
66 |
### Data Splits
|
67 |
|
|
|
73 |
|
74 |
The dataset is a combination of the following datasets:
|
75 |
|
76 |
+
| Dataset | Sentences |
|
77 |
+
|-------------------|-----------|
|
78 |
+
| CCMatrix v1 | 3.765.459 |
|
79 |
+
| WikiMatrix | 317.649 |
|
80 |
+
| GNOME | 1.752 |
|
81 |
+
| KDE4 | 117.828 |
|
82 |
+
| QED | 43.736 |
|
83 |
+
| TED2020 v1 | 41.461 |
|
84 |
+
| OpenSubtitles | 235.604 |
|
85 |
+
| GlobalVoices | 3.430 |
|
86 |
+
| Tatoeba | 723 |
|
87 |
+
| Europarl | 3.765.459 |
|
88 |
+
| **Total** | **6.159.631** |
|
89 |
+
|
90 |
+
All corpora except Europarl were collected from [Opus](https://opus.nlpl.eu/).
|
91 |
+
The Europarl corpus is a synthetic parallel corpus created from the original Spanish-Catalan corpus by [SoftCatalà](https://github.com/Softcatala/Europarl-catalan).
|
92 |
+
|
93 |
|
94 |
### Data preparation
|
95 |
|
96 |
+
All datasets are deduplicated and filtered to remove any sentence pairs with a cosine similarity of less than 0.75.
|
97 |
+
This is done using sentence embeddings calculated using [LaBSE](https://huggingface.co/sentence-transformers/LaBSE).
|
98 |
+
The filtered datasets are then concatenated to form a final corpus of 6.159.631 and before training the punctuation is normalized using a modified version of the join-single-file.py script from [SoftCatalà](https://github.com/Softcatala/nmt-models/blob/master/data-processing-tools/join-single-file.py)
|
99 |
|
100 |
### Personal and Sensitive Information
|
101 |
|
|
|
109 |
|
110 |
### Discussion of Biases
|
111 |
|
112 |
+
We are aware that since part of the data comes from unreliable web pages and non-curated texts, some biases may be present in the dataset.
|
113 |
+
Nonetheless, we have not applied any steps to reduce their impact.
|
114 |
|
115 |
### Other Known Limitations
|
116 |
|