Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -11,7 +11,7 @@ This dataset contains automatically aligned Romansh–German document pairs, ext
|
|
| 11 |
|
| 12 |
## Contents
|
| 13 |
|
| 14 |
-
The dataset consists of several alignment outputs based on different cosine similarity thresholds and sentence-length
|
| 15 |
|
| 16 |
threshold_matches/
|
| 17 |
|
|
@@ -47,7 +47,19 @@ Each `.jsonl` file contains entries with the following fields:
|
|
| 47 |
5. Only matches above a given threshold were retained.
|
| 48 |
|
| 49 |
**Thresholds tested:** `0.5 – 0.8`
|
| 50 |
-
**Final dataset:**
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 51 |
|
| 52 |
|
| 53 |
## Evaluation
|
|
@@ -65,5 +77,5 @@ Unit tests confirmed:
|
|
| 65 |
| File | Description |
|
| 66 |
| --------------------- | ---------------------------------- |
|
| 67 |
| `parallel_data/` | All alignment outputs by threshold |
|
| 68 |
-
|
|
| 69 |
| `embedded` | German and Romansh embeddings |
|
|
|
|
| 11 |
|
| 12 |
## Contents
|
| 13 |
|
| 14 |
+
The dataset consists of several alignment outputs based on different cosine similarity thresholds and sentence-length penalty:
|
| 15 |
|
| 16 |
threshold_matches/
|
| 17 |
|
|
|
|
| 47 |
5. Only matches above a given threshold were retained.
|
| 48 |
|
| 49 |
**Thresholds tested:** `0.5 – 0.8`
|
| 50 |
+
**Final dataset:** `/parallel_data/matched_0.60_0.005.jsonl` ("best" match)
|
| 51 |
+
|
| 52 |
+
### Dataset Quality Note
|
| 53 |
+
|
| 54 |
+
The dataset was aligned using a relatively low cosine similarity threshold (e.g. 0.60) to increase the number of matches. However, this likely reduces alignment quality and may introduce noisy or weakly related sentence pairs. Further filtering or manual inspection would be necessary to improve reliability.
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
## Threshold and Alpha Comparison
|
| 58 |
+
|
| 59 |
+
We compared cosine similarity thresholds (0.50–0.80) using two alpha settings: alpha=None (no length regularization) and alpha=0.005 (penalizing length mismatch). alpha=None produced more matches, especially at low thresholds, but with a steep quality drop as the threshold increased. alpha=0.005 yielded fewer but more consistent matches.
|
| 60 |
+
|
| 61 |
+
Conclusion:
|
| 62 |
+
For better balance between match quantity and alignment quality, we selected alpha=0.005 with a threshold of 0.60`.
|
| 63 |
|
| 64 |
|
| 65 |
## Evaluation
|
|
|
|
| 77 |
| File | Description |
|
| 78 |
| --------------------- | ---------------------------------- |
|
| 79 |
| `parallel_data/` | All alignment outputs by threshold |
|
| 80 |
+
| `/parallel_data/matched_0.60_0.005.jsonl` | "best match" |
|
| 81 |
| `embedded` | German and Romansh embeddings |
|