Upload README.md with huggingface_hub

README.md

@@ -3,68 +3,154 @@ title: Romansh–German Parallel Dataset (FineWeb-based)
 colorFrom: gray
 colorTo: red
 sdk: static
-configs:
-- config_name: default
-  data_files: "parallel_data/matched_0.60_0.005.jsonl"
-  default: true
 ---
 
 # Romansh–German Parallel Dataset (FineWeb-Based)
 
 This dataset contains automatically aligned Romansh–German document pairs, extracted from the [Fineweb2](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2) using cosine similarity over OpenAI embeddings. It was created as part of a university programming project focused on document-level parallel data extraction.
 
-## Contents
+## Description
 
-The dataset consists of several alignment outputs based on different cosine similarity thresholds and sentence-length penalty:
+This project performs document-level alignment between Romansh and German web texts, which were extracted from the [Fineweb2](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2) dataset. It uses [OpenAI](https://platform.openai.com/docs/models/text-embedding-3-small) embeddings and cosine similarity to identify potential parallel texts.
+The full dataset is available on [Hugging Face](https://huggingface.co/datasets/Sudehsna/Romansh_German_Parallel_Data).
 
-threshold_matches/
 
-├── matched_{THRESHOLD}_{ALPHA}.jsonl 
+## Dataset
 
-├── matched_0.60_0.005.jsonl
+This project uses the Romansh and German partitions of the  [Fineweb2](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2) dataset. Both the original and the removed  versions of the dataset were used to improve alignment coverage.
 
-...
+Since the German dataset contained significantly more entries than the Romansh dataset, it was filtered by retaining only those entries whose domains matched those found in the Romansh dataset.
 
-With ALPHA being either 0.005 or None
 
+**Dataset Filtering**:
+We first used a brute-force approach by computing cosine similarities between all Romansh and German embeddings without filtering (see `data/embedded/original` and the initial outputs in `data/parallel_data_unfiltered/`on Huggingface). After evaluating the quality of these matches, we refined the Romansh embeddings by removing entries from long-tail domains (i.e., domains with fewer than 3 documents) and deduplicating (`data/embedded/filtered_notail_dedup`). After that we filtered this new dataset even more by keeping only entries whose domains end in `.ch`, `.org`, `.gov` and `.edu` since these are the most likely to have Romansh and German entries. Cosine similarities were then recomputed using this cleaner subset to improve alignment precision.
 
-## Format
+In total there were three different datasets used for alignment:
+- Original Fineweb dataset
+- Filtered dataset without long-tail domains and deduplication
+- More filtered dataset where top-level domains are `.ch`, `.org`, `.gov` and `.edu`
 
-Each `.jsonl` file contains entries with the following fields:
 
-```json
-{
-  "romansh_text": "Romansh document text",
-  "german_text": "Most similar German document text or null",
-  "similarity": 0.6123,              // Score after optional length penalty
-  "original_similarity": 0.6085       // Raw cosine similarity
-}
-```
+### Dataset statistics
+ **Original Dataset:**
+- Total aligned document pairs: ?
+- Total entries in embeddings:
+  - Romansh: 208321
+  - German: 333993 
+- Total tokens in embeddings:
+  - Romansh: 106,005,135
+  - German:  
+- Cosine similarity score distribution for original dataset:
+  - Mean: ?
+  - Median: ?
+  - Std Dev: ?
+  - Threshold range: 0.50 - 0.80
+
+- Top domain: www.rtr.ch (79603 documents)
+
+  **Filtered:**
+  - Total aligned document pairs: ?
+- Total entries in embeddings:
+  - Romansh: 208321
+  - German: 333993 
+- Total tokens in embeddings:
+  - Romansh: 101,263,110
+  - German:  
+- Number of total unique domains: 8106
+- Cosine similarity score distribution for filtered dataset (no long tail and deuplication):
+  - Mean: ?
+  - Median: ?
+  - Std Dev: ?
+  - Threshold range: 0.50 - 0.80
+- Long tail: 5582 domains with < 3 documents (removed)
+
+
+  **More Filtered:**
+  - Total aligned document pairs: ?
+- Total entries in embeddings:
+  - Romansh: 175195
+  - German: 333993 
+- Total tokens in embeddings:
+  - Romansh: 86,917,433
+  - German: 333993 (unfiltered)
+- Number of total unique domains: 8106
+- Cosine similarity score distribution for filtered dataset (only `.ch` and `.org`):
+  - Mean: ?
+  - Median: ?
+  - Std Dev: ?
+  - Threshold range: 0.50 - 0.80
+
+**General:**
+- Token length limits: Embeddings were truncated at 8192 tokens (OpenAI model constraint)
+- Sentence Length Penalties:
+  **Absolute penalty**: 
+`penalized_sim = cos_sim - α * abs(len_rm - len_de)`
+
+  **Relative penalty**:
+`penalized_sim = cos_sim - α * (abs(len_rm - len_de) / max(len_rm, len_de))`
+
+
+## Results from Dataset Analysis
+
+### Domain Distribution
+
+The domain distribution is highly imbalanced: the vast majority of URLs come from a small number of domains. Most notably, `www.rtr.ch` dominates the dataset with nearly 80,000 URLs, followed by `m.rtr.ch` and `www.gr.ch` with significantly fewer entries. All other domains appear less frequently, with fewer than 10,000 URLs each.
+
+This indicates that the Romansh dataset is heavily influenced by a few large sources, primarily public institutions and media outlets such as Radiotelevisiun Svizra Rumantscha (RTR) and cantonal websites.While this concentration can benefit consistency and linguistic quality, it may also limit the variety of content and styles represented in the dataset. This imbalance should be considered in any downstream tasks, as it could influence model performance and generalizability.
+
+### How alignable is the dataset?
+
+After manually examining the 50 randomly sampled URLs from above, I found that none of the 50 entries had a clearly alignable German counterpart. 
+To check alignments I relied on matching 1) named entities 2) dates and numbers 3) translations of phrases or words (with [Supertext](https://www.supertext.com/de-CH?gad_source=1&gad_campaignid=22600834922&gbraid=0AAAAAD-ejiqGvPoZBlBS3NZ_qVPUILwC0&gclid=CjwKCAjwp_LDBhBCEiwAK7FnkiyYnMO5qZN4ruGnqZ6qDR22rfVbOAtcFZoX4fkofquirt5hfXGLKBoCjVoQAvD_BwE)).
+
+This suggests that parallel data in the Romansh subset is extremely rare, and that future alignment should rely on automated methods with robust filtering. This set of 50 samples now serves as a gold standard for evaluating automatic alignment thresholds.
 
 ## Methodology
 
-1. Documents were embedded using [OpenAI](https://platform.openai.com/docs/models/text-embedding-3-small) `text-embedding-3-small`.
-2. Cosine similarity was computed between Romansh and German documents.
-3. A length penalty was applied:  
-   `penalized_score = cosine_similarity - alpha × |len_diff|`
-4. For each Romansh document, the best-scoring German document was selected.
-5. Only matches above a given threshold were retained.
+1. **Preprocessing**:
+    - German dataset filtered to only contain domains also appearing in Romansh dataset
+2. **Manual Evaluation**:
+   -   A sample of 50 documents was manually checked to estimate alignability (see `analysis/manual_alignment`)
+3. **Embedding**:
+   - Used [OpenAI](https://platform.openai.com/docs/models/text-embedding-3-small) `text-embedding-3-small` (with truncation at 8192 tokens)
+4. **Similarity Calculation**:
+   - Cosine similarity computed between Romansh and German document embeddings
+   - With or without penalizations for sentence length difference
 
-**Thresholds tested:** `0.5 – 0.8`  
-**Final dataset:** `/parallel_data/matched_0.60_0.005.jsonl` ("best" match)
 
 ## Threshold and Alpha Comparison
 
-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.
+We compared cosine similarity thresholds (0.50–0.80) using two penalty settings for the original dataset: Absolute and relative sentence length. 
+- `ALPHA = None` and `LENGTH_PENALTY_TYPE = None` (no length regularization)
+- `ALPHA=0.005` (absolute penalty)
+- `LENGTH_PENALTY_TYPE = 0.05` (relative penalty) 
+
+REVISE
+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.
+
+To determine a suitable _similarity threshold_ for aligning sentence pairs, we tested multiple values and observed that a threshold of **0.60** provided the best balance for this dataset (refer to `./analysis/threshold_analysis/threshold_comparisons.py`).
 
-Conclusion:
-For better balance between match quantity and alignment quality, we selected alpha=0.005 with a threshold of 0.60`.
 
 ### Dataset Quality Note
 
 As the dataset was aligned using a relatively low cosine similarity threshold (e.g. 0.60) to increase the number of matches it likely reduces alignment quality and may introduce noisy or weakly related sentence pairs. Further filtering or manual inspection would be necessary to improve reliability.
 
 
+## Format
+
+Each `.jsonl` file contains entries with the following fields:
+
+```json
+{
+  "romansh_text": " ... ",
+  "german_text": " ... ",
+  "romansh_url": " ...",
+  "german_url": " ...",
+  "similarity": float, 
+  "original_similarity": float
+}
+```
+
 ## Evaluation
 
 A negative gold standard was used to ensure quality: a set of Romansh documents known to have no valid German alignment. These were checked for false positives.
@@ -87,12 +173,33 @@ dataset = load_dataset(
 )
 ```
 
+
+## Contents
+
+The dataset consists of several alignment outputs based on different cosine similarity thresholds and sentence-length penalty:
+
+parallel_data_unfiltered/
+
+├── matched_{THRESHOLD}_{ALPHA}_{LENGTH_PENALTY_TYPE}.jsonl 
+
+├── matched_0.50_ 0.005_relative.jsonl
+├── matched_0.50_0.005_absolute.jsonl
+├── matched_0.50_none.jsonl
+...
+
+With ALPHA being `0.005` or `None` and LENGTH_PENALTY_TYPE can be `"relative"`, `"absolute"` or `None`.
+Same structure for `/parallel_data_filtered`.
+
 ## File Organization
 
 | File                  | Description                        |
 | --------------------- | ---------------------------------- |
-| `parallel_data/`  | All alignment outputs by threshold |
+| `parallel_data_unfiltered/`  | All alignment with unfiltered dataset|
+| `parallel_data_filtered/no_tail_dedup`  | All alignments with filtered dataset (removed long tail and deduplicated)|
+| `parallel_data_filtered/tld`  | All alignments with more filtered dataset (only few tlds)|
 | `/parallel_data/matched_0.60_0.005.jsonl` | "best match"     |
-| `embedded`           | German and Romansh embeddings  |
-| `data/`               | Original data 
+| `embedded/original`           | All German and Romansh embeddings  |
+| `embedded/filtered_notail_dedup`           | German and Romansh embeddings with removed long tail and deduplicated  |
+| `embedded/filtered_notail_dedup/filtered_topleveldomain`           | German and Romansh embeddings only domains with `.ch` and `.org`  |
+| `original/`               | Original data