How to Parse a Creole: When Martinican Creole Meets French

Proceedings of the 29th International Conference on Computational Linguistics, pages 4397–4406 October 12–17, 2022. 4397 How to Parse a Creole: When Martinican Creole Meets French Ludovic Mompelat Indiana University lmompela@iu.edu Daniel Dakota Indiana University ddakota@iu.edu Sandra Kübler Indiana University skuebler@indiana.edu Abstract We investigate methods to develop a parser for Martinican Creole, a highly under-resourced language, using a French treebank. We com- pare transfer learning and multi-task learning models and examine different input features and strategies to handle the massive size imbal- ance between the treebanks. Surprisingly, we find that a simple concatenated (French + Mar- tinican Creole) baseline yields optimal results even though it has access to only 80 Martinican Creole sentences. POS embeddings work better than lexical ones, but they suffer from negative transfer. 1 Introduction Syntactic analysis is an essential task for language documentation and language revitalization, as it al- lows a deeper understanding of languages. Under- resourced languages often suffer from the lack of annotated gold standard data available to de- velop and offer NLP solutions for the communi- ties speaking the language. Moreover, there is a low number of researchers trained in formal lin- guistics and/or linguistic annotations which causes additional challenges in the creation of language resources for such languages. In recent years, re- search on parsing has developed a focus on under- resourced languages (Agi´c et al., 2016; Vania et al., 2019; Meechan-Maddon and Nivre, 2019), but cre- oles have received less attention. In this study, we develop a dependency parsing model for Martinican Creole (MC), a French-based Creole, mostly spoken in Martinique, a French is- land in the Caribbean. Being a French territory, French and MC coexist in an unbalanced manner. The diglossic situation makes French the domi- nant language in many contexts, although in the past decades Martinican Creole has seen an ex- pansion of its communicative contexts (Bernabé, 2004; Véronique, 2020). This is due to the codifi- cation and standardization processes the language underwent, especially by the GEREC (1982), and to the linguistic policies developed in an effort to safeguard and revitalize this language. However, one aspect of this revitalization process that is cur- rently missing is the expansion of NLP tools and re- sources for MC (and other creole languages). Cre- ole languages based on the same lexifier language (in our case French) are extremely similar and in many cases, mutually intelligible. Thus, develop- ing NLP solutions for one creole language provides a basis to transfer knowledge to other related Cre- ole languages. The goal of this project is to investigate the best methods for developing a parser for an extremely low resource language when this language is a cre- ole language. In our case, the creole is Martinican Creole. The main question here is whether the lex- ifier language, i.e., French, is similar enough to serve as basis for training a parser without further modification. 2 Research Questions Our overarching research question is the following: Can we leverage a French treebank using transfer learning or multi-task learning approaches to …

Content truncated to 500 words