Morpheme Segmentation in Multi- and Monolingual Wordlists

This package provides implementations for several algorithms by which words in a wordlist can be segmented into morphemes.

If you use this software package, please cite it accordingly:

Rubehn, A. and J.-M. List (2025). MorSeg: A Python package for morpheme segmentation in multi- and monolingual wordlists [Software Library, Version 0.1]. Chair for Multilingual Computational Linguistics, University of Passau.

Installation

This package can be conveniently installed using pip:

pip install morseg

Basic Usage

Loading data

Assuming your data is presented in a TSV file following the LingPy specifications (see /tests/test_data/german.tsv for an example), you can simply load your data with:

frommorseg.utils.wrappersimportWordlistWrapperwl=WordlistWrapper.from_file(YOUR_FILE)

This creates a wordlist wrapper object; a representation of a wordlist with three annotation levels: The predicted segmentations (by a model), the Gold standard segmentations, and the unsegmented form. The training of all models requires the data to be stored in this class!

Training a model

The Tokenizer class offers a unified interface for all models that are implemented in this library. For example, if you want to train a LSV (Letter Successor Variety) model, you can simply do so like that:

frommorseg.algorithms.tokenizerimportLSVTokenizermodel=LSVTokenizer()
model.train(wl)

The current release covers implementations of the following models:

• LSVTokenizer: Letter Successor Variety (Harris, 1955) with the following adaptations:
  • Letter Successor Entropy (Hafer and Weiss, 1974)
  • Letter Max-Drop Variety (Hammarström, 2009)
  • Normalized Letter Successor Variety (Çöltekin, 2010)
• LPVTokenizer: Letter Predecessor Variety (analogically to LSV, but processing the words backwards)
• LSPVTokenizer: A combination of Letter Successor Variety and Letter Predecessor Variety
• Morfessor: The Morfessor Baseline Model (Creutz and Lagus, 2002)
• SquareEntropyTokenizer (Méndez-Cruz et al., 2016)

Furthermore, some popular models for subword tokenization are implemented:

• PairEncoding: Byte-Pair Encoding (Sennrich et al., 2016)
• WordPiece (Schuster and Nakajima, 2012)
• UnigramSentencePiece (Kudo, 2018)

Obtain segmentations

You can obtain the predicted segmentations from your training data by calling:

forsegmented_wordinmodel.get_segmentations():
# do whatever

You can also try segmenting unseen words (depending on the model, this might work more or less well):

word= ["w", "o", "r", "d"]
segmented_word=model(word)