Morphological Tagging and Lemmatization of Albanian: A Manually Annotated Corpus and Neural Models

Nelda Kote, Marenglen Biba, Jenna Kanerva, Samuel Rönnqvist, Filip GinterarXiv 2019 · arXiv

Introduction

Albanian is an Indo-European language spoken by approximately 7 million people, yet it remains significantly "low-resource" in the field of Natural Language Processing (NLP). Its grammar is notoriously complex, representing an independent branch of the Indo-European language tree with a rich system of inflections, five cases, and ten parts of speech.

Historically, Albanian NLP tools have been rule-based, proprietary, or lacked the sentence-level context required for modern machine learning. The absence of a gold-standard, publicly available dataset following international standards like the Universal Dependencies (UD) schema has hindered the development of state-of-the-art neural models. This research addresses this gap by releasing the first manually annotated corpus for Albanian, designed for part-of-speech (POS) tagging, morphological analysis, and lemmatization.

Methodology

The researchers developed a multi-step pipeline to create both the data and the models:

  1. Data Collection and Annotation: Text was gathered from diverse sources, including fiction, grammar books, web crawls, and Wikipedia. The team adopted the Universal Dependencies (UD) schema to ensure cross-linguistic compatibility. Using a "seed" approach, they converted older datasets to UD format via scripts, followed by rigorous manual correction and expansion by native speakers.
  2. Neural Modeling: The Turku Neural Parser Pipeline was employed for modeling. This architecture utilizes UDPipe for segmentation, Bidirectional LSTMs for POS and morphological tagging, and sequence-to-sequence models for character-level lemmatization. Semantic context was provided by fastText word embeddings trained on Albanian Wikipedia.

Dataset Composition

The study introduces a benchmark corpus totaling approximately 185,000 tokens. A unique aspect of this dataset is the inclusion of "artificial" text—simple sentences generated from UniMorph verb lists—to ensure the model encounters a wide variety of morphological variations that are rare in standard prose.

Dataset statistics for training, development, and testing sets.

As shown in the statistics above, the artificial data was restricted to the training set to maintain the integrity and realism of the evaluation results on natural text.

Key Insights

  • The "Artificial" Boost: Synthetic sentences significantly improved the model's ability to learn rare inflections.
  • Handling Multi-word Tenses: Albanian often forms tenses using a verb and a particle (e.g., ka bërë for "has done"). Following UD standards, the authors tagged these individually to capture the underlying grammar of each component.
  • Standardization: Mapping Albanian's unique features (like its specific case system and definite/indefinite noun forms) to the global UD schema makes the language accessible to international research tools.

Results and Performance

The models were evaluated on a held-out test set, demonstrating strong performance across all primary NLP tasks.

Model accuracy across different NLP tasks.

While token and sentence segmentation reached near-perfect accuracy, morphological feature tagging (85.31%) proved more challenging, reflecting the inherent complexity of Albanian's inflectional system. Nevertheless, the 92.74% accuracy for POS tagging represents a significant milestone for the language.

Conclusion and Future Work

This work provides a much-needed open-source foundation for Albanian NLP. By releasing the corpus and models under an open license, the authors have enabled further research into this complex language. Future efforts will focus on expanding the corpus to include dependency parsing (syntactic relations) and exploring dialectal variations, such as the Gheg dialect, to move toward a more comprehensive representation of the Albanian language.

Summary updated

Abstract

In this paper, we present the first publicly available part-of-speech and morphologically tagged corpus for the Albanian language, as well as a neural morphological tagger and lemmatizer trained on it. There is currently a lack of available NLP resources for Albanian, and its complex grammar and morphology present challenges to their development. We have created an Albanian part-of-speech corpus based on the Universal Dependencies schema for morphological annotation, containing about 118,000 tokens of naturally occuring text collected from different text sources, with an addition of 67,000 tokens of artificially created simple sentences used only in training. On this corpus, we subsequently train and evaluate segmentation, morphological tagging and lemmatization models, using the Turku Neural Parser Pipeline. On the held-out evaluation set, the model achieves 92.74% accuracy on part-of-speech tagging, 85.31% on morphological tagging, and 89.95% on lemmatization. The manually annotated corpus, as well as the trained models are available under an open license.
BibTeX
@misc{kote2019morphological,
  author = {Nelda Kote and Marenglen Biba and Jenna Kanerva and Samuel Rönnqvist and Filip Ginter},
  title = {{Morphological Tagging and Lemmatization of Albanian: A Manually Annotated Corpus and Neural Models}},
  year = {2019},
  eprint = {1912.00991},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1912.00991},
}
Figure