Toward Universal Dependencies for Shipibo-Konibo

Alonso Vasquez, Renzo Ego Aguirre, Candy Angulo, John Miller, Claudia Villanueva, Željko Agić, Roberto Zariquiey, Arturo OncevayEMNLP 2018

TL;DR

This paper introduces the first Universal Dependencies (UD) treebank for Shipibo-Konibo, an indigenous Amazonian language. The authors detail their novel approach to segmenting and annotating clitics as independent syntactic words, a key feature of the language, and present initial results on parsing and typological comparisons. This work is crucial for advancing NLP for under-resourced languages and understanding linguistic diversity.

What Problem Does This Paper Address?

Natural Language Processing (NLP) tools and research have historically focused on high-resource languages, leaving many of the world's languages with limited computational resources. This is particularly true for indigenous languages, especially those in regions like the Amazon. Shipibo-Konibo, an agglutinative Panoan language spoken in Peru, lacks a comprehensive UD treebank, hindering the development of NLP applications and linguistic analysis within the UD framework. Existing UD treebanks for South American indigenous languages are scarce, with the closest being for a Quechuan variant not fully compliant with UD guidelines. This paper aims to bridge this gap by creating a UD-compliant treebank for Shipibo-Konibo, enabling typological comparisons and facilitating future NLP research for this language.

Methods

The authors developed a multi-stage annotation process:

  1. Morpheme Segmentation: They utilized ChAnot for detailed morpheme segmentation (prefixes, roots, suffixes) and BRAT for graphical annotation.
  2. UD v2.0 Conversion: Segmented corpora were compiled into UD v2.0 format. A key methodological decision was to treat clitics (linguistic elements that are neither full words nor affixes) as independent syntactic words. This approach, while more time-consuming, better captures their phrasal nature and their ability to participate in different syntactic relations. This differs from some other UD treebanks that might label them differently.
  3. POS Tagging and Feature Annotation: They employed the standard UD POS tagset, using 'X' for onomatopoeia (a significant feature in Shipibo-Konibo) and 'PART' for clitics. They also introduced and defined several new morphological features to capture specific aspects of Shipibo-Konibo, such as verb tenses, aspects, and a special "chezative" case.
  4. Dependency Relation Annotation: They used 31 of the 37 standard UD dependency relations and introduced two language-specific subtype relations: aux:valid (for evidentiality and mood clitics) and compound:onom (for onomatopoeic compounds).

Datasets

The primary dataset introduced is the Shipibo-Konibo UD treebank, consisting of 407 sentences. This corpus was compiled from parallel Shipibo-Konibo and Spanish educational materials, storybooks, and elicited sentences. For evaluation, they used the Es-Ancora treebank for Spanish.

Key Insights and Ideas

  • Clitics as Independent Words: The decision to treat clitics as independent syntactic words is a significant methodological choice. It highlights how UD can be adapted to capture the unique grammatical structures of languages, even if it requires deviating from a strict interpretation of "word" in other linguistic traditions. This approach is crucial for languages with rich morphology like Shipibo-Konibo.
  • High Frequency of Evidentiality: The aux:valid relation, primarily used for evidentiality clitics, appears in 5.6% of words and nearly half of the sentences. This suggests that evidentiality is a highly salient and frequently expressed grammatical category in Shipibo-Konibo, inviting further linguistic investigation.
  • Abundance of Multiword Tokens (MWTs): Due to the clitic segmentation strategy, Shipibo-Konibo exhibits a proportionally high number of MWTs compared to languages like Spanish and Turkish, though less than Hebrew. This underscores the impact of annotation choices on corpus statistics.
  • Cross-Lingual Parsing Insights: The study found that languages with similar morphosyntactic features (like Kazakh and Japanese) can serve as good sources for delexicalized cross-lingual parsing, even if their WALS distance (a measure of linguistic similarity) is large. This suggests that feature similarity can sometimes outweigh broad typological distance.

Key Results

  • Treebank Size: The initial treebank contains 407 sentences, with 2706 orthographic tokens and 3148 syntactic words (after clitic segmentation).
  • POS Tag Frequencies: PUNCT and NOUN are the most frequent POS tags, followed by VERB and PART (which includes clitics). The high frequency of PART (14%) is notable.
  • Dependency Relation Frequencies: punct, root, and nsubj are the most common relations. aux:valid is also frequent (5.6%).
  • Monolingual Parsing: Using gold POS tags, the Parsito parser achieved 83.66% UAS and 77.81% LAS, while the Yara parser achieved 87.32% UAS and 81.25% LAS. Parsing raw text yielded significantly lower scores, highlighting the importance of accurate POS tagging.
  • Cross-Lingual Parsing: The best performance was achieved using the Kazakh treebank as a source, yielding a 66% UAS score. A negative correlation was observed between WALS distance and UAS scores, suggesting that linguistically similar source languages tend to perform better.

Limitations and Future Work

  • Treebank Size: The current treebank of 407 sentences is considered small, and the authors plan to increase its size.
  • Annotation Consistency: While efforts were made to ensure consistency, the authors acknowledge that some features were automatically inferred, and future work will involve manual annotation for greater accuracy.
  • Cross-Lingual Parsing Variance: The cross-validation results for monolingual parsing showed high variance, suggesting potential overfitting or scarcity issues. These results should be treated as preliminary.
  • WALS Feature Sparsity: The correlation analysis for cross-lingual parsing was affected by the sparsity of WALS features for some languages, which could be addressed by using more complete feature sets or inferring missing data.

Future work includes expanding the treebank, semi-supervised annotation of morphological features using an FSM-based analyzer, and extending UD tagsets and guidelines to related Panoan languages to foster language technology development for minority languages in Peru.

Summary updated

BibTeX
@inproceedings{vasquez2018toward,
  author = {Alonso Vasquez and Renzo Ego Aguirre and Candy Angulo and John Miller and Claudia Villanueva and Željko Agić and Roberto Zariquiey and Arturo Oncevay},
  title = {{Toward Universal Dependencies for Shipibo-Konibo}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2018},
}
Figure