--- license: apache-2.0 language: - ur base_model: - facebook/fasttext-km-vectors tags: - art --- --- # Card Metadata (Optional but Recommended) # You can fill these out directly in the Hugging Face UI or here. # Language: ur (Urdu) # Tasks: # - word-embeddings # Library: # - fasttext # Datasets: # - [Specify your dataset name here, e.g., your-dataset-name-on-hf, or just 'Custom Corpus'] # Tags: # - urdu # - word-vectors # - embeddings # - fasttext # - unsupervised # - urdu-nlp # License: [Specify your license here, e.g., mit, apache-2.0, cc-by-4.0] --- # Urdu Word Embeddings (fastText) ## Model Description This is an unsupervised word embedding model for the Urdu language, trained using the fastText library. It generates high-dimensional vectors for Urdu words, capturing semantic and syntactic relationships based on their context in the training data. Unlike traditional Word2Vec, this fastText model was trained with character n-grams (`minn=[Your minn]`, `maxn=[Your maxn]`), which is particularly beneficial for morphologically rich languages like Urdu. This allows the model to: - Learn representations for subword units. - Generate meaningful vectors for words it hasn't seen during training (Out-of-Vocabulary or OOV words) by composing vectors from their character n-grams. The model outputs vectors of dimension `[Your vector_size]`. ## Intended Use This model is intended for use in various Urdu Natural Language Processing (NLP) tasks, including: - Measuring semantic similarity between Urdu words. - Using word vectors as features for downstream tasks such as text classification, clustering, or named entity recognition. - Exploring word relationships and patterns within the vocabulary learned from the training corpus. - Obtaining vector representations for potentially unseen words based on their subword components. ## Training Data This model was trained on a custom text corpus of Urdu sentences. - **Dataset Source:** [Specify the source of your training data here. For example: "Collected from the COUNTER (COrpus of Urdu News TExt Reuse) dataset" or "A custom corpus gathered from [mention sources or domain]"]. - **Data Format:** The training data was processed into a single text file (`train.txt`) where each line represented a sentence or document, and words were separated by spaces. - **Preprocessing:** Basic preprocessing was applied, including replacing common punctuation marks with spaces and normalizing whitespace. [Mention any other specific preprocessing steps you performed, e.g., lowercasing (less common for Urdu), handling numbers, removing specific symbols]. [If your training data is publicly available or derived from a public source, provide a link or instructions on how others can access it.] [If the data is private, state that the data itself cannot be shared but the resulting model is being released.] ## Training Procedure The model was trained using the unsupervised capabilities of the fastText library. - **Algorithm:** Continuous Bag of Words (CBOW) model (`model=cbow`). [If you used `skipgram`, specify that instead and briefly explain why, e.g., "Skip-gram model (`model=skipgram`), often better for capturing representations of rare words."] - **Parameters:** The following parameters were used during training: - `dim`: `[Your vector_size]` (Vector dimensionality) - `ws`: `[Your window_size]` (Context window size) - `minCount`: `[Your min_word_count]` (Minimum word frequency to be included in vocabulary) - `epoch`: `[Your epochs]` (Number of training epochs) - `neg`: `[Your negative_samples]` (Number of negative samples) - `minn`: `[Your minn]` (Minimum character n-gram length) - `maxn`: `[Your maxn]` (Maximum character n-gram length) - `thread`: 4 (Number of threads used) - [List any other significant parameters you modified] - **Training Environment:** The training was performed in a Google Colab environment. ## How to Use You can load and use this model using the fastText Python library. First, make sure you have fastText installed: ```bash pip install fasttext import fasttext import numpy as np # For calculating cosine similarity # Path to the downloaded .bin model file model_path = "path/to/your/downloaded/urdu_fasttext.bin" # Load the fastText model try: model = fasttext.load_model(model_path) print("Model loaded successfully!") except ValueError as e: print(f"Error loading model: {e}") print("Ensure the file exists and is a valid fastText binary model.") model = None # Set model to None if loading fails if model: # --- Get Word Vector --- word = "پاکستان" # Example Urdu word print(f"\nVector for '{word}':") try: vector = model.get_word_vector(word) print(f"Shape: {vector.shape}") print(f"First 10 dimensions: {vector[:10]}") except ValueError as e: print(f"Error getting vector for '{word}': {e}. Word might be too short or have no valid subwords.") # --- Find Nearest Neighbors (Similar Words) --- word_for_neighbors = "اردو" # Example Urdu word print(f"\nWords similar to '{word_for_neighbors}':") try: # Get top 10 most similar words neighbors = model.get_nearest_neighbors(word_for_neighbors, k=10) if neighbors: print(neighbors) else: print(f"No similar words found for '{word_for_neighbors}'.") except ValueError as e: print(f"Error finding similar words for '{word_for_neighbors}': {e}. Word might not be valid.") # --- Calculate Similarity Between Two Words (Manual Cosine Similarity) --- word1 = "علم" # Example word 1 word2 = "روشنی" # Example word 2 print(f"\nSimilarity between '{word1}' and '{word2}':") try: vec1 = model.get_word_vector(word1) vec2 = model.get_word_vector(word2) # Calculate cosine similarity norm1 = np.linalg.norm(vec1) norm2 = np.linalg.norm(vec2) if norm1 > 0 and norm2 > 0: cosine_similarity = np.dot(vec1, vec2) / (norm1 * norm2) print(f"Cosine similarity: {cosine_similarity}") else: print("Cannot compute similarity: zero vector detected for one or both words.") except ValueError as e: print(f"Error calculating similarity between '{word1}' and '{word2}': {e}. One or both words might not be valid.") # --- Using the .vec file (Optional) --- # The .vec file contains just the word vectors for words in the vocabulary. # It can be loaded by other libraries like Gensim or spaCy. # Note: This method *does not* utilize fastText's subword capabilities for OOV words. # For fastText specific features, use the .bin file. # Example (using gensim - requires gensim installation): # from gensim.models import KeyedVectors # vec_file_path = "path/to/your/downloaded/urdu_fasttext.vec" # try: # # Load vectors in Word2Vec text format # word_vectors = KeyedVectors.load_word2vec_format(vec_file_path, binary=False) # print(f"\nLoaded {len(word_vectors.key_to_index)} vectors from .vec file using Gensim.") # # Example: Find similar words using Gensim # # print(word_vectors.most_similar("اردو")) # except Exception as e: # print(f"Error loading .vec file with Gensim: {e}") else: print("\nModel could not be loaded. Usage examples are skipped.") **Steps after creating the Model Card content:** 1. **Create a Model Repository on Hugging Face:** Go to huggingface.co, log in, click your profile picture -> "New model". 2. **Name your Model:** Choose a descriptive name (e.g., `urdu-fasttext-word-embeddings`). 3. **Set Visibility:** Choose Public or Private. 4. **Create Model:** This creates an empty repository. 5. **Upload Files:** Go to the "Files" tab of your new repository. You can either: * Click "Add file" and upload `urdu_fasttext.bin`, `urdu_fasttext.vec`, and your training script file. * Or, clone the repository locally and push the files using Git. 6. **Edit Model Card:** Go to the "Model card" tab. This is where you paste and format the content prepared above. You can edit it directly in the browser using Markdown. 7. **Fill in Placeholders:** Go through the content and replace all `[ ... ]` placeholders with your specific details (vector size, epochs, dataset source, license, your name, etc.). 8. **Format with Markdown:** Use the formatting options (headers, bold, code blocks) to make the card readable. 9. **Save Model Card:** Save the changes. Your model will then be available on Hugging Face with the documentation you've provided.