A NER Model for Science Fiction’s Lexical Innovation

Speculative fiction, including science fiction, is distinguished by its ability to generate an invented lexicon composed of terms called novums. However, these lexical units are absent from Named Entity Recognition (NER) models applied to computational literary studies, mainly because of their complexity in detection (highly varied syntactic distributions, strong semantic ambiguity). We present a CamemBERT-based NER model, adapted to the analysis of science fiction narratives and enriched with a dedicated class (NOV for novum) to identify these lexical inventions. Evaluated on a specific annotated corpus, our model achieves an F1-score of 63% for the detection of the very rare NOV class, with a macro F1-score of 79% across all entities. This model thus provides a tool for studying how these lexical creations reflect or even anticipate the technological and social imaginaries of their time. The resources presented here, including the annotated corpus and code, are available online.