Introduction
The preservation and revitalization of Brazilian Indigenous Languages (BIL) face a significant technological hurdle: "data poverty." Linguistic resources such as dictionaries and treebanks are scarce, fragmented, and non-standardized. This lack of interoperability prevents the development of sophisticated educational tools. For instance, a dictionary for the Bororo language cannot easily interface with a syntax treebank for Guajajara. To address this, researchers have developed BILGraph, a unified knowledge graph designed to integrate these disparate resources into a semantic framework.
Graph Architecture and Methodology
The core of BILGraph is a knowledge extraction pipeline that transforms raw linguistic data into a structured graph using the Resource Description Framework (RDF). The architecture is defined by four hierarchical node types that bridge the gap between specific utterances and abstract meanings:
- Sentence: The top-level phrase extracted from source corpora.
- Token: Individual words as they appear in the sentence.
- Lemma: The base dictionary form of a word.
- Concept: High-level semantic abstractions (e.g.,
<FISH>or<EAT>) that serve as a universal linking layer.
The Entity Linking Challenge
Linking lemmas to concepts is the most critical and challenging stage. Due to varied spellings and limited data in indigenous languages, the authors employed a rule-based "neighborhood generation" approach. This includes character swap rules (e.g., "u" for "y" in Bororo) and Levenshtein distance to identify candidate matches in dictionaries. Accuracy is maintained by cross-referencing dictionary definitions with the original sentence context; a link is established only when semantic consistency is confirmed across sources.
Data Resources
The BILGraph prototype integrates several key linguistic resources:
- TuLeD (Tupian Lexical Database) and TuDeT (Tupian Dependency Treebanks).
- The Bororo Corpus & Dictionary, providing deep coverage for the Bororo language.
- Target Languages: The initial implementation focuses on Bororo and Tupian languages, including Guajajara, Munduruku, and Akuntsu.
The Semantic Bridge and Educational Impact
The introduction of the "Concept" node is the project's primary innovation. It enables cross-linguistic interoperability. A search for the concept <FISH> can simultaneously retrieve relevant sentences across different languages—such as the Bororo ure karo kowuje and the Guajajara u'u ipiratetea'u—despite their distinct linguistic origins.
This standardization offers significant scalability for education. Educators can query the graph for specific linguistic features (e.g., "sentences with possessive pronouns") or thematic clusters ("food-related vocabulary") across multiple languages, drastically accelerating the creation of pedagogical materials.
Results and Discussion
The authors successfully demonstrated a functional prototype of BILGraph capable of handling semi-structured data from diverse sources. The system effectively performs cross-linguistic retrieval, linking similar semantic themes across the targeted indigenous languages. Most importantly, the technology is already being applied to develop a curriculum for a Bororo language-learning application, validating the practical utility of the knowledge graph approach in real-world revitalization efforts.
Limitations and Future Work
Despite its success, BILGraph faces several limitations:
- Translation Coverage: Only four of the nine targeted languages currently include Portuguese translations, limiting broader educational accessibility.
- Lexical Ambiguity: Similarity-based linking can occasionally produce "hallucinated" links between phonetically similar but semantically unrelated words (e.g., Bororo apido "palm heart" vs. apodo "toucan").
- Community Involvement: The authors acknowledge that, with the exception of the Bororo community, the project has lacked direct involvement from indigenous speakers. Future work must prioritize community validation to ensure cultural and linguistic accuracy.
- Technical Evolution: Future iterations will explore probabilistic linking methods and pipelines that reduce dependency on pre-existing dictionaries.
Summary updated