TL;DR
This paper presents a new, deep-learning-based morphological tagger for Bribri, an Indigenous language of Costa Rica, achieving a statistically significant improvement over previous methods. By leveraging Universal Dependencies' Universal Features, the authors created a more robust system that can better handle the complexities of Bribri morphology, paving the way for expanded NLP tools for this low-resource language.
What Problem Does This Paper Address?
Natural Language Processing (NLP) has made significant strides, but its reach into low-resource languages, particularly Indigenous ones, remains limited. These languages often present unique challenges due to limited data, orthographic variations, and complex morphosyntax. Morphological tagging, the process of identifying the grammatical properties of word parts, is a crucial step in building NLP tools. For Bribri, an Indigenous language spoken in Costa Rica with around 7000 speakers, existing NLP tools are scarce. Previous attempts at morphological tagging for Bribri relied on rule-based systems like Finite State Transducers (FSTs), which struggle with unseen words and require extensive manual rule creation. This paper aims to improve Bribri morphological tagging by developing a more accurate and robust system using modern deep learning techniques within the Universal Dependencies (UD) framework.
Methods
The authors employed a two-stage approach. First, they created a morphologically-tagged treebank for Bribri using a rule-based system based on regular expressions. This initial pass focused on verbs, nouns, pronouns, and copulas, leveraging linguistic knowledge of Bribri's ergative-absolutive case system and verb conjugation patterns. After manual correction of the rule-based output, the resulting treebank was used to train a UDPipe 2 parsing and tagging model. UDPipe 2 is a well-established tool for dependency parsing and morphological analysis, and by training it on the enriched Bribri data, the researchers aimed to achieve higher accuracy and better generalization than the previous FST-based tagger.
Datasets
The primary dataset used was an existing dependency treebank for Bribri, containing 315 sentences (1572 tokens). This treebank was originally tagged with Universal Dependencies v2. The authors expanded this treebank by applying their rule-based morphological tagging and then performing manual corrections, resulting in a dataset with 3051 morphological features. This enhanced treebank served as the training and evaluation data for the UDPipe 2 model.
Key Insights and Ideas
- UD Features as a Bridge: The Universal Dependencies (UD) framework, specifically its Universal Features (UFEATS) schema, provides a standardized way to represent morphological information across diverse languages. This standardization is crucial for low-resource languages like Bribri, allowing for better integration with broader NLP research and tools.
- Deep Learning Outperforms Rule-Based: The study demonstrates that a deep learning model (UDPipe 2) trained on a manually corrected, UD-tagged Bribri treebank significantly outperforms a traditional FST-based tagger. This highlights the power of modern machine learning for handling linguistic complexity, especially when provided with high-quality annotated data.
- Morphological Complexity is a Hurdle: Bribri exhibits significant morphological complexity, particularly in its verb system (tense, mood, aspect, voice) and its ergative-absolutive case marking. Accurately capturing these nuances is essential for effective NLP.
- Error Analysis Reveals Key Challenges: The error analysis of the UDPipe 2 model identified missing TAM (tense-mood-aspect) features and case markings as the most frequent errors. This provides valuable insights into the specific linguistic phenomena that remain challenging for automated systems.
Key Results
The trained UDPipe 2 model achieved a UFEATS precision of 80.5 ± 3.6%. This represents a statistically significant improvement over the previously available FOMA-based morphological tagger, which achieved a precision of 59.5 ± 4.2% on all tokens and 75.7 ± 4.7% on tokens with existing UFEATS in the gold standard. The authors' rule-based system initially struggled with approximately 24% of verbs, requiring manual correction. The final deep learning model, however, showed a substantial leap in performance.
Limitations and Future Work
The authors acknowledge several limitations and propose future directions:
- Written Bribri Focus: The current system is limited to written Bribri, while much of the language's data is oral. Developing a system that can handle oral narratives and diverse orthographic conventions is a key future goal.
- Expanding Tag Coverage: The current tagging scheme focused on verbs, nouns, pronouns, and copulas. Future work will expand coverage to other parts of speech, including numerical classifiers and determiners, which have complex morphological features not easily captured by current UD features.
- Handling Morphological Segmentation: The paper lays the groundwork for future work on automatic morphological segmentation, aiming to break down words into their constituent morphemes (e.g., root, suffix).
- Incorporating More Linguistic Nuances: Specific Bribri linguistic features, such as the deictic system (indicating distance and vertical position) and distinct past tense distinctions (hodiernal tense), require further development of UD features or separate annotation schemes.
Summary updated