--- language: - ar - zh - da - en - fr - ja - ko - sv library_name: pytorch pipeline_tag: text-classification tags: - EDAN20 - language-identification - scikit-learn - pytorch --- # Nano Language Detector (EDAN20, Lab 4) A small language identifier inspired by Google's CLD3, built for the EDAN20 Language Technology course at Lund University. It recognises eight languages: Arabic (`ara`), Mandarin Chinese (`cmn`), Danish (`dan`), English (`eng`), French (`fra`), Japanese (`jpn`), Korean (`kor`), and Swedish (`swe`). ## Model Each sentence is lowercased and split into character unigrams, bigrams, and trigrams. The n-grams are hashed with MD5 into 521 + 1031 + 1031 = 2583 buckets, and the input vector contains the relative frequency of each bucket. The network has one hidden layer: ``` input (2583) -> Linear + ReLU (50) -> Linear (8) -> softmax ``` The same architecture was trained twice: with scikit-learn (`MLPClassifier`, 5 epochs) and with PyTorch (Adam, cross-entropy, batches of 32, 7 epochs). ## Files | File | Content | |---|---| | `nld.joblib` | scikit-learn `MLPClassifier` | | `nld.pth` | PyTorch weights (`state_dict`) | | `nld_vectorizer.joblib` | `DictVectorizer` mapping hash buckets to input columns (needed by both models) | | `nld_lang_codes.joblib` | Dictionary from class index to language code | | `predict.py` | Feature extraction and inference with both models | ## Usage ```bash pip install torch scikit-learn joblib huggingface_hub python predict.py "Hejsan grabbar!" "Salut les gars !" ``` The feature extraction in `predict.py` must be used as is: the models only work with the same hashing (MD5) and the same bucket sizes. ## Training data and results Trained on Tatoeba sentences: 25,000 per language (11,674 for Korean), 80 % for training and 20 % (37,335 sentences) for validation. | Model | Accuracy | Macro F1 | |---|---|---| | scikit-learn | 0.9912 | 0.9916 | | PyTorch | 0.9924 | 0.9927 | Most errors are between Swedish and Danish, and very short sentences can be misclassified.