Aspects of the construction of a Universal Dependencies treebank for Nheengatu

Abstract

The alienation of natural language technologies adds up to the weakening of minority languages coexisting with majority languages. Especially younger speakers, who function as links in language transmission, tend to migrate to the language favored by these resources. Nheengatu, an endangered Brazilian indigenous language, has a digital support score of just 0.07 on the Digital Language Support (DLS) scale. This is significantly lower than the 0.97 score for Portuguese, to which Nheengatu has been continually losing speakers. The Nheengatu treebank of the Universal Dependencies collection aims to reduce this deficit by feeding the training of a neural parser. Initially released on 11/15/2023 with 196 sentences and 2,146 words, the latest version, as of 05/15/2024, comprises 1,470 sentences and 15,036 words from twenty publications spanning different historical phases of Nheengatu. This makes it the largest treebank for an Amerindian language in the collection. The use of an automatic analyzer facilitated the rapid expansion of the corpus, while human annotators reviewed each annotation to ensure a 100% validation rate, achieving a two-star rating, the highest for Amerindian language treebanks in the Universal Dependencies collection. The ongoing expansion and revision aim to include all public domain texts and achieve state-of-the-art parsing results.

Keywords:
Computational linguistics; Natural language processing; Tupinology; Treebank