Successful communication depends upon many cognitive processes. Speakers must create a signal that they think is the best for their recipients to infer the intended meaning. To achieve this, humans rely on words from their lexicon as a versatile tool for conveying meaning. However, the correspondence between the form of a word and its semantics is not entirely arbitrary, with some regularities in the mapping between word forms and corresponding meanings in natural languages. This systematicity can be extended beyond the boundaries of existing languages and applies even to non-existing letter strings, which lack conventional meaning (pseudowords).

Using distributional semantic models, we can produce quantitative representations for these novel words and investigate the extent to which people rely on their sub-lexical patterns to assess meaning. The present study builds on this logic: by presenting participants with novel, supposedly meaningless items in a reference game, we can tease apart how orthographic overlap and semantic similarity influence their intuitions and communicative behaviour. More specifically, we plan to recruit pairs of participants in which a “speaker” will have to select two pseudowords (“targets”) from a grid that, according to their intuition, go well together and generate an existing “clue” word that will allow a “recipient” to select the intended targets from a context of distractors.

We expect that both novel word processing and the production of existing words can be jointly modelled within a computational framework of word selection. The current study design will be presented (data yet to be collected).