Part-of-speech tagger for Slovene language implemented using convolutional and LSTM neural networks. Tagger uses character-level representation of sentences.
Neural Character-level Part-of-speech (POS) tagger of Slovene language. Neural model is based on character level language model proposed by by Kim et al. in Character-Aware neural language models (2015).
Our POS tagger has been presented in thesis Part of speech tagging of slovene language using deep neural networks. Thesis is accessible at University of Ljubljana
The thesis deals with part of speech tagging of Slovene language. Part of speech tagging is a process of matching sentences in natural language with a sequence of suitable tags, which contain information about parts of speech and morphological properties of words. Our solution uses character-level representation of words, which is different from typical solutions, which process input sentences as sequences of words. Our part of speech tagger is implemented using convolutional and recurrent neural networks. Unlike common approaches that address this problem as multi-class classification, our solution proposes a multi-label classification approach. In order to improve our results we implement an ensemble of three part of speech taggers. When comparing our solution with existing ones, we find that the proposed solution achieves the best results.
Our model has been trained on SSJ500k training corpus. This corpus is available at clarin.si: Training corpus SSJ500k 2.2 Our model has been trained using ssj500k-en.TEI.
Due to file size limitations of GitHub, trained model (model_weights.h5) is not available in this repository. It can be acquired at clarin.si repository: Part-of-speech tagger of Slovene language
Tagger loads Keras model configuration from model.json and model weights from model_weights.h5.
python tag.py input.xml output.xml
Tagging of text files requires Obeliks4J for segmentation and tokenisation. Path to Obeliks4J directory must be passed through --obelikspath parameter.
python tag.py input.txt output.xml --obelikspath /home/user/Obeliks4J/
-s, --slo Tagger predicts slovene tags.-f, --force Overwrite existing output file.
Training new models requires dataset in XML\TEI format. Output of training process consists of files model.json which contains Keras model configuration, model_weights.h5 containing weights of neural network and charset which contains a list of all character occuring in provided dataset. These three files are are used to make predictions with tag.py.
Number of training iterations (epochs) can be passed with -n parameter. Default number of epochs is 20.
python train.py ssj500k.xml outputdir -n 10
-n, --nepoch Number of training iterations.-s, --slo Indicates that tags in training set are in slovene language.-b, --beginning Can be used to train model on subset of input training set. Starting (inclusive) index of subset.-e, --end Can be used to train model on subset of input training set. Final (exclusive) index of subset.