Using the first 36 WALS features as input, you can fine-tune RoBERTa to classify an unknown language's family (e.g., Indo-European vs. Sino-Tibetan) with high accuracy. The zip file provides balanced sets to prevent overfitting to dominant families.
clf = RandomForestClassifier() clf.fit(X, y) print("Accuracy on set1:", clf.score(X_test, y_test))
Tokenizing the language data using the RoBERTa tokenizer ( RobertaTokenizerFast ). WALS Roberta Sets 1-36.zip
unzip WALS_Roberta_Sets_1-36.zip -d ./wals_roberta/ cd wals_roberta conda create -n wals_roberta python=3.9 conda activate wals_roberta pip install transformers datasets numpy pandas scikit-learn
: Structured data points from the World Atlas of Language Structures. Using the first 36 WALS features as input,
This file name is a window into its structure and purpose. The name is composed of several key parts:
: Sets 1-36 may represent a partitioned dataset used to test how well a RoBERTa model trained on one set of languages performs on others based on their WALS features. Feature Extraction clf = RandomForestClassifier() clf
: Ensure you see folders for "Instruments" and "Samples." Add to Kontakt : Open Kontakt. Go to the Files tab. Browse to the "WALS Roberta" folder. Double-click an .nki file to load the instrument. 3. Managing Sets 1–36
“WALS Roberta Sets 1-36.zip is a pre-processed version of WALS 2020. Use sets 1-30 for training, sets 31-33 for validation, and sets 34-36 for testing. Each set contains 200 language varieties, balanced by genus.”