Universal Dependencies for Saraiki

Meesum Alam, Francis Tyers, Emily Hanink, Sandra KüblerCOLING 2024

TL;DR

This paper introduces the first Universal Dependencies (UD) treebank for Saraiki, an under-resourced Indo-Aryan language spoken by 25 million people. By annotating 587 sentences, they provide crucial linguistic resources for NLP development and address specific syntactic phenomena like split ergativity and complex verb constructions.

What Problem Does This Paper Address?

The field of Natural Language Processing (NLP) heavily relies on linguistically annotated corpora, often in the form of treebanks, to train and evaluate parsing technologies and gain insights into language structures. However, Indo-Aryan languages have historically been underrepresented in these efforts. While some major Indo-Aryan languages like Hindi, Urdu, and Punjabi have UD treebanks, many others, like Saraiki, remain largely unaddressed. This lack of resources hinders NLP development for these languages, making them "neglected" in the digital sphere. This paper aims to bridge this gap by creating the first UD treebank for Saraiki, a language spoken by a significant population in Pakistan and India.

Methods

The authors followed the established Universal Dependencies (UD) annotation scheme to create the Saraiki treebank. They meticulously annotated 587 sentences, totaling 7,597 tokens, and made specific decisions for handling Saraiki's unique linguistic features. These decisions include:

  • Compounds: Annotating various types of compounds, including noun-noun compounds and reduplication, using the compound relation.
  • Verbal Structures: Differentiating between verbs and auxiliaries in light verb constructions (LVCs) based on their semantic contribution. They also annotated serial verb constructions (SVCs) by marking the last verb as the head and using compound:lvc relations for others, anticipating future refinements.
  • Split Ergativity: Addressing Saraiki's split ergative alignment, where the grammatical role of the agent (subject of a transitive verb) shifts between nominative and oblique cases depending on the verb's aspect (perfective vs. imperfective). They followed decisions made for Urdu and Hindi treebanks, annotating agents as nsubj and patients as obj.
  • Relative Clauses: Annotating both externally and internally headed relative clauses, as well as correlative forms. They used the relativizer jera (that, which) and made specific decisions on how to represent the head noun and its relation to the relative clause.

Datasets

The Saraiki treebank was built using sentences from three sources:

  • Saraiki Common Voice corpus: A collection of spoken language data.
  • Jhok Newspaper: A daily newspaper providing written text.
  • Generated Sentences: Sentences created specifically during the annotation process to clarify decisions on complex syntactic phenomena.

The paper notes that finding diverse digital texts for Saraiki was challenging due to its under-resourced status.

Key Insights and Ideas

  • UD as a Unifying Framework: The paper demonstrates the applicability of the Universal Dependencies framework to a previously unannotated Indo-Aryan language, highlighting its potential for cross-linguistic NLP research.
  • Linguistic Nuances of Saraiki: The creation of the treebank necessitated detailed analysis and decision-making for Saraiki-specific phenomena, such as its split ergativity and complex verb constructions, offering valuable insights into the language's grammar.
  • Challenges of Under-resourced Languages: The authors explicitly mention the difficulties in acquiring sufficient and diverse textual data for Saraiki, underscoring the ongoing need for resource creation for less-studied languages.
  • Annotation Decisions as a Collaborative Process: The paper emphasizes the iterative nature of annotation, involving native speakers, linguistics scholars, and consultation with experts from other UD projects to ensure accuracy and consistency.

Key Results

The paper presents the first UD treebank for Saraiki, comprising 587 annotated sentences and 7,597 tokens. The distribution of parts of speech tags is provided, showing that NOUNs and VERBs are the most frequent categories. The authors detail their annotation decisions for key syntactic phenomena, providing a foundation for future NLP research and tool development for Saraiki.

Limitations and Future Work

The authors acknowledge that the current treebank is a work in progress and will be extended to 1,000 sentences before official release. Specific areas for future work include:

  • Deeper analysis of relative clauses: Further investigation into the complexities of relative clause structures.
  • Automatic morphological feature annotation: Utilizing the Apertium morphological analyzer for Saraiki to automate the annotation of morphological features.
  • Syntactic parser development: Training a syntactic parser for Saraiki based on the created treebank.
  • Zero-shot techniques: Exploring the application of zero-shot learning to extend NLP capabilities to related languages like Punjabi, Hindko, and Khetrani.

Summary updated

BibTeX
@inproceedings{alam2024universal,
  author = {Meesum Alam and Francis Tyers and Emily Hanink and Sandra Kübler},
  title = {{Universal Dependencies for Saraiki}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2024},
}
Figure