TL;DR
This paper introduces the first Universal Dependencies (UD) treebank for Highland Puebla Nahuatl (HPN), a critically endangered indigenous language. This resource is crucial for advancing NLP for HPN and enabling comparative linguistic studies across Nahuatl dialects, paving the way for better digital language technology and representation.
What Problem Does This Paper Address?
Natural Language Processing (NLP) technologies, while powerful, often overlook or underrepresent endangered and marginalized languages. This gap hinders their inclusion in the digital world and limits the robustness of multilingual NLP models. Specifically, annotated linguistic corpora, like Universal Dependencies (UD) treebanks, are essential for training and evaluating NLP models. However, UD treebanks for indigenous languages of the Americas are scarce. This paper addresses this by creating a UD treebank for Highland Puebla Nahuatl (HPN), an Uto-Aztecan language spoken in central Mexico, which is at risk of being lost. This effort complements an existing treebank for another Nahuatl variant, enabling crucial comparative linguistic research and multi-dialectal NLP development.
Methods
The authors followed the Universal Dependencies (UD) framework to create the HPN treebank. They employed an automated pipeline for initial annotation, including tokenization, part-of-speech tagging, morphological analysis, and syntactic parsing, using a finite-state morphological analyzer. This was followed by a rigorous manual review and annotation process by two authors, with disagreements resolved through discussion. They also developed a consistent orthographic normalization strategy for lemmas, mapping them to a single norm derived from a specific Nahuatl course. For syntactic analysis, they utilized the UD Annotatrix tool and Arborator-Grew for tree updates. The paper also details specific annotation decisions for interesting syntactic constructions, such as the multi-functional word "in" and noun incorporation, aligning with UD guidelines and prior work on other Nahuatl variants.
Datasets
The HPN treebank was constructed from four sources:
- Axolotl Nahuatl-Spanish parallel corpus: Sampled texts from the HPN variant.
- Amith et al. (2019) audio corpus: Transcribed and translated monologues and dialogues.
- "Science outreach in indigenous languages" (SMF and INALI): Scientific texts translated into Nahuatl.
- Introductory HPN course materials: 32 example sentences.
The total corpus contains approximately 10,088 tokens across 1,261 trees. The paper also discusses a comparison with the Western Sierra Puebla Nahuatl (WSPN) treebank, which has a similar token count but a different genre distribution.
Key Insights and Ideas
- UD as a bridge for endangered languages: The creation of a UD treebank for HPN demonstrates the applicability of a standardized annotation framework to languages with limited digital resources, promoting their inclusion in NLP.
- Linguistic diversity within Nahuatl: Comparing the HPN treebank with the WSPN treebank reveals interesting dialectal and genre-driven differences in syntax, morphology, and sentence length distribution, highlighting the importance of dialect-specific resources.
- Challenges of polysynthetic languages: The paper sheds light on annotating complex phenomena like noun incorporation in polysynthetic languages, proposing a method to represent these structures within the UD framework.
- The multi-functional "in": The word "in" in HPN exhibits diverse roles, functioning as a determiner, pronoun, and subordinator, showcasing the flexibility of grammatical elements in the language.
- Relational nouns and adpositions: Nahuatl's use of relational nouns for spatial relations, similar to adpositions in other languages, is explored, along with the impact of Spanish contact on preposition usage.
Key Results
- A UD treebank for Highland Puebla Nahuatl has been created, comprising approximately 10,088 tokens.
- A neural NLP pipeline was trained on the HPN treebank, achieving promising results:
- Part-of-speech tagging: 94.5% accuracy (in-domain).
- Lemmatization: 89.8% accuracy (in-domain).
- Morphological features: 91.7% accuracy (in-domain).
- Universal Dependencies (UAS): 79.8% accuracy (in-domain).
- Labeled Dependency Arc (LAS): 72.7% accuracy (in-domain).
- Out-of-domain evaluation showed a drop in UPOS tagging but surprisingly better dependency parsing performance, attributed to the simpler sentence structures in the out-of-domain sample.
Limitations and Future Work
- Genre bias: The HPN treebank is biased towards botanical discussions due to its primary data sources, which may affect the generalizability of NLP models trained on it.
- Future NLP development: The authors suggest that detailed experimentation with NLP pipelines for Nahuatl, including optimizations and improvements, is a key area for future research.
- Comparative dialectology: The availability of both HPN and WSPN treebanks opens avenues for quantitative exploration of syntax across Nahuatl dialects.
- Syntactic phenomena: Further exploration of specific syntactic phenomena, such as incorporated nouns and their relation to external words, is designated for future work.
Summary updated