PetroVec : desenvolvimento e avaliação de modelos vetoriais de palavras em português para o domínio de óleo e gás

Abstract

This work presents PetroVec, a set of word embedding models in Portuguese for the O&G domain. To make model training feasible, we created a specialized corpus composed of a vast collection of technical and scientific documents, published in Portuguese by Universities and major institutions from national petroleum-related industry. The specialized corpus comprises about 85 million tokens and it is currently the largest public textual resource ever reported for the O&G domain. Our specialized models are thoroughly evaluated, comprising quantitative methodologies based on intrinsic and extrinsic approaches, in addition to a series of qualitative analyses to explore linguistic properties encoded in the models’ semantic space. The intrinsic evaluation is performed by creating a semantic similarity dataset composed of 1500 pairs of terms labeled by experts in geosciences, whereas the extrinsic evaluation consists of a downstream task for named entity recognition in the Geology subdomain. Furthermore, we conducted a comprehensive analysis comparing our models and a pre-trained general-domain model in Portuguese. Our findings confirm that PetroVec consistently outperforms the generalcontext reference model, suggesting that our models were able to automatically capture syntactic and semantic vocabulary-specific properties from the specialized training corpus.