MOROCO: The Moldavian and Romanian Dialectal Corpus

Andrei M. Butnaru, Radu Tudor IonescuACL 2019 · DOI

Abstract

This paper introduces the Moldavian and Romanian Dialectal Corpus (MOROCO), a significant new resource for Natural Language Processing (NLP) focused on dialect identification (DID). While Romanian and Moldavian share a nearly identical literary standard, we demonstrate that machine learning models can distinguish between them with high precision. Using a combination of traditional string kernels and a novel "Squeeze-and-Excitation" CNN architecture, we achieve over 94% accuracy in dialect identification. Our findings suggest that regional stylistic preferences create a distinct "digital signature" for each dialect.

Introduction

Dialect Identification (DID) is the task of automatically determining the dialect of a given text. This task is particularly challenging for "low-distance" varieties like Romanian and Moldavian, which share the same literary standard and lack significant linguistic divisions (isoglosses) in professional writing. This research investigates whether computational models can discern differences between news texts from Bucharest and Chișinău, and whether these differences are rooted in deep linguistic patterns or merely geographical markers like named entities.

The MOROCO Dataset

The primary contribution of this work is the MOROCO dataset, which consists of 33,564 text samples totaling over 10 million tokens. The data was collected from the top five news websites in both Romania and the Republic of Moldova. Samples are categorized into six distinct topics: culture, finance, politics, science, sports, and tech.

Distribution of news samples across topics for Moldavian and Romanian dialects.

To ensure models were not relying solely on local names, the authors produced a version of the dataset where all named entities (people, places, organizations) were replaced with a generic $NE$ token.

Methodology

The study evaluates three primary computational approaches:

  1. String Kernels (Shallow Learning): Kernel Ridge Regression (KRR) based on character $n$-grams was employed to capture subtle stylistic and orthographic differences.
  2. Character-level CNN (Deep Learning): A deep neural network was used to process text character-by-character, enabling the learning of features without a predefined vocabulary.
  3. Squeeze-and-Excitation (SE) Blocks: Adapted from computer vision, SE blocks were integrated into the CNN to perform "feature recalibration," allowing the network to prioritize the most informative features for dialect identification.

Results and Discussion

The experimental results demonstrate that the dialects are highly distinguishable. The string kernel method (KRR) consistently outperformed deep learning models, reaching a peak accuracy of 94.13% for binary dialect classification.

Performance results for binary classification and topic categorization across different methods and validation/test sets.

Key findings include:

  • Robustness to Named Entity Removal: Accuracy remained high (~92.6%) even after removing named entities, indicating that the models learned genuine linguistic differences.
  • Linguistic Markers: A notable marker is the orthographic preference for 'î' vs. 'â'. Moldavian writers often use 'î' (e.g., cînd) where Romanian writers use 'â' (când).
  • Cross-Dialect Generalization: Models trained on topic categorization in one dialect showed significant performance degradation when tested on the other, highlighting the "dialect shift" challenge in NLP.

Conclusion

The MOROCO corpus provides a robust foundation for studying low-distance dialect identification. Our results show that even closely related dialects possess unique computational signatures. Future work will focus on shorter text snippets, such as social media posts, and exploring ensemble methods to further improve classification performance.

Summary updated

BibTeX
@inproceedings{butnaru2019moroco,
  author = {Andrei M. Butnaru and Radu Tudor Ionescu},
  title = {{MOROCO: The Moldavian and Romanian Dialectal Corpus}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
  doi = {10.18653/v1/p19-1068},
}
Figure