luisarmando
commited on
Commit
•
1987119
1
Parent(s):
12811b8
Update README.md
Browse files
README.md
CHANGED
@@ -12,8 +12,8 @@ widget:
|
|
12 |
- text: 'translate spanish to nahuatl: México lindo y querido.'
|
13 |
---
|
14 |
|
15 |
-
#
|
16 |
-
Nahuatl is the most widely spoken indigenous language in Mexico
|
17 |
|
18 |
|
19 |
## Model description
|
@@ -38,7 +38,8 @@ outputs = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
|
|
38 |
|
39 |
## Approach
|
40 |
### Dataset
|
41 |
-
Since the Axolotl corpus contains misalignments,
|
|
|
42 |
|
43 |
| Axolotl best aligned books |
|
44 |
|:-----------------------------------------------------:|
|
@@ -57,10 +58,10 @@ Since the Axolotl corpus contains misalignments, we select the best samples (12,
|
|
57 |
| Una tortillita nomás - Se taxkaltsin saj |
|
58 |
| Vida económica de Tenochtitlan |
|
59 |
|
60 |
-
Also,
|
61 |
|
62 |
### Model and training
|
63 |
-
|
64 |
|
65 |
### Training-stage 1 (learning Spanish)
|
66 |
In training stage 1, we first introduce Spanish to the model. The goal is to learn a new language rich in data (Spanish) and not lose the previous knowledge. We use the English-Spanish [Anki](https://www.manythings.org/anki/) dataset, which consists of 118.964 text pairs. The model is trained till convergence, adding the prefix "Translate Spanish to English: "
|
@@ -73,15 +74,16 @@ We train the models on the same datasets for 660k steps using batch size = 16 an
|
|
73 |
|
74 |
|
75 |
## Evaluation results
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|:----------------------------:|:---------------:|:-----|-------:|
|
80 |
-
| False | 1.34 | 6.17 | 26.96 |
|
81 |
-
| True | 1.31 | 6.18 | 28.21 |
|
82 |
|
|
|
83 |
|
84 |
-
|
|
|
|
|
|
|
85 |
|
86 |
## References
|
87 |
- Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits
|
|
|
12 |
- text: 'translate spanish to nahuatl: México lindo y querido.'
|
13 |
---
|
14 |
|
15 |
+
# mt5-large-spanish-nahuatl
|
16 |
+
Nahuatl is the most widely spoken indigenous language in Mexico, yet training a neural network for machine translation presents significant challenges due to insufficient structured data. Popular datasets, like Axolot and the Bible corpus, contain only approximately 16,000 and 7,000 samples, respectively. Complicating matters further, Nahuatl has multiple dialects, and a single word in the Axolot dataset can appear in over three different forms. I conclude with evaluations of the model's performance using the Chrf and BLEU metrics.
|
17 |
|
18 |
|
19 |
## Model description
|
|
|
38 |
|
39 |
## Approach
|
40 |
### Dataset
|
41 |
+
Since the Axolotl corpus contains misalignments, the best samples were selected (12,207).
|
42 |
+
This in addition to the [bible-corpus](https://github.com/christos-c/bible-corpus) (7,821).
|
43 |
|
44 |
| Axolotl best aligned books |
|
45 |
|:-----------------------------------------------------:|
|
|
|
58 |
| Una tortillita nomás - Se taxkaltsin saj |
|
59 |
| Vida económica de Tenochtitlan |
|
60 |
|
61 |
+
Also, additional 30,000 samples were collected from the web to enhance the data.
|
62 |
|
63 |
### Model and training
|
64 |
+
The employed method uses a single training stage using the mt5. This model was leveraged given that it can handle different vocabularies and prefixes.
|
65 |
|
66 |
### Training-stage 1 (learning Spanish)
|
67 |
In training stage 1, we first introduce Spanish to the model. The goal is to learn a new language rich in data (Spanish) and not lose the previous knowledge. We use the English-Spanish [Anki](https://www.manythings.org/anki/) dataset, which consists of 118.964 text pairs. The model is trained till convergence, adding the prefix "Translate Spanish to English: "
|
|
|
74 |
|
75 |
|
76 |
## Evaluation results
|
77 |
+
The models are evaluated on 2 different datasets:
|
78 |
+
1. First on the test sentences similar to the evaluation ones.
|
79 |
+
2. Then, on zero-shot sentences obtained from the test sentences of AmericasNLP2021
|
|
|
|
|
|
|
80 |
|
81 |
+
The results are reported using CHRF++ and BLEU:
|
82 |
|
83 |
+
| Nahuatl-Spanish Bidirectional Training | Set | BLEU | CHRF++ |
|
84 |
+
|:----------------------------:|:---------------:|:-----|-------:|
|
85 |
+
| True | Test | 18.01 | 54.15 |
|
86 |
+
| True | Zero-shot | 5.24 | 25.7 |
|
87 |
|
88 |
## References
|
89 |
- Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits
|