Keyword Extraction from Parallel Abstracts of Scientific Publications

Abstract

In this paper, we study the keyword extraction from parallel abstracts of scientific publication in the Serbian and English languages. The keywords are extracted by a selectivity-based keyword extraction method. The method is based on the structural and statistical properties of text represented as a complex network. The constructed parallel corpus of scientific abstracts with annotated keywords allows a better comparison of the performance of the method across languages since we have the controlled experimental environment and data. The achieved keyword extraction results measured with an F1 score are 49.57% for English and 46.73% for the Serbian language, if we disregard keywords that are not present in the abstracts. In case that we evaluate against the whole keyword set, the F1 scores are 40.08% and 45.71% respectively. This work shows that SBKE can be easily ported to new a language, domain and type of text in the sense of its structure. Still, there are drawbacks – the method can extract only the words that appear in the text.