Initially, BraveNewWord focused on defining its computational framework, implementing computational characterizations of the three mechanisms of novel-word-driven meaning acquisition in the domain of vector space modelling. Such a view characterizes meaning as numerical vectors, which can in turn be seen as points in a multidimensional space. When such a semantic space is aligned with a linguistic space, populated by words and sublexical elements, one can predict which meaning will be evoked by novel, unfamiliar elements. The computational mechanisms allowing for such predictions are grounded in the three cognitive mechanisms described above: meaning induction from minimal linguistic context, semantic combination of morphological units, form-to-meaning mapping. To achieve its computational objectives, BraveNewWord applies both existing architectures and newly developed approaches.
The models developed by BraveNewWord naturally produce quantitative, empirically testable predictions about behaviour and neural activity. The project is testing such predictions with methodologies ranging from response times to neuroimaging to electrophysiology.
Concerning the impact of minimal linguistic context, we have observed a modulation of the N400 for novel words in context, after as few as two occurrences. The N400 is an electrophysiological response, measured via EEG, that indexes how surprising a word is within a given sentence or, from a different perspective, how difficult it is to integrate the encountered element in the previous context. This evidence indicates that novel words are rapidly assigned a meaning, which is routinely integrated with the previously presented familiar information. Crucially, the BraveNewWord computational approach can predict the N400 magnitude, and hence how well the novel word integrates with the preceding context.
Concerning morphology-induced meanings, we relied on and extended a model, CAOSS, previously proposed for novel compound words (e.g. rivercat). We adapted this architecture to other types of morphologically complex elements, such as prefixed (e.g. respeak) and suffixed words (e.g. quickify), and showed that the model predictions align with human responses in behavioral tasks across different languages. Furthermore, in a neuroimaging study, we observed distinct neural signatures for novel meanings induced by morphologically complex words.
Concerning form-meaning mapping, across a number of studies we investigated human intuitions about the possible meaning of completely unfamiliar linguistic strings (e.g. futmaw). Participants were shown to be able to produce consistent responses across a range of different semantic dimensions, up to actual definitions. Such responses were significantly predicted by the BraveNewWord computational approach. Moreover, uniquely within the BraveNewWord endeavour that typically moves from language to semantics, we designed a system that moves in the opposite direction, producing a new word on the basis of a desired meaning.
