Processing compounds: what frequency (alone) cannot explain

Observed elevation in typing latency for the initial letter of the second constituent of an English compound, compared with the typing time of the final letter of the first constituent (Gagné & Spalding 2016), suggests that both compounds ( snowball ) and pseudo-compounds ( carpet ) are decomposed but also that full form representations are available in the lexical store. To gain further insight into the lexical representations underlying typing, we used computational modelling. In particular, we used superpositional models of word memory, based on Self-Organising Recurrent Maps (TSOMs) (Ferro et al. 2016; Marzi et al. 2016), where both simple and compound words are processed (and stored) using the same pool of processing (and memory) resources, to model the elevation in typing time at the constituent boundary and the rate of typing. In addition, we also considered models based in the Compositional Distributional Semantics framework (CAOSS, Marelli et al. 2017), to simulate independent effects of semantic transparency on compound typing (Gagné & Spalding 2016). Due to co-activation and competition between compounds and their constituent words in TSOMs, levels of activation of processing nodes per letter positions appear to reflect degrees of context-sensitive predictability: the higher the level, the more expected the letter in that position. In English compounds, activation levels appeared to exhibit a characteristically U-shaped pattern, with min values centred on the constituent boundary. A similar pattern was found for pseudo-compounds, which nonetheless present a less pronounced U-shaped pattern and a higher activation value at the morpheme boundary than compounds do. The difference is in line with the higher speed-up rate in typing pseudo-compounds than compounds reported in Gagné and Spalding (2016). TSOMs were trained on letter-based representations, so computer experiments could simulate peripheral effects of serial processing of compound structure before lexical access. To investigate post-lexical issues, we also tested computational models of generation of the meanings of novel compounds based on CAOSS, which proved to be able to account for well-established relational effects in compound processing (Gagné 2001; Gagné & Shoben 1997) with an unsupervised data-driven framework (Marelli et al. 2017). We ran a mixed-effects regression analysis of the data in Gagné and Spalding (2016) using vector-semantics estimates and TSOM activation levels to predict typing time for the initial letter of the second constituent. There was a negative effect of TSOM letter activation levels: i.e. the more active a letter node is, the faster a subject is at typing the letter ( t =-2.7 p =.007). Also, there was a positive effect of CAOSS-based compositionality estimates: i.e. the more easily a compound's lexicalized meaning can be obtained through compositional operations on single constituent vectors, the slower participants were at typing the first letter of the second constituent ( t =2.4, p =.017). These results have interesting implications for an integrative computational architecture accounting for the whole range of experimental evidence reported by Gagné and Spalding (2016). In particular we will focus on evidence of a stronger competition (and longer typing time) in Transparent-Transparent and Transparent-Opaque compounds, vs. Opaque-Transparent compounds, which gives an indication of a non-trivial interaction between semantic compositionality and serial processing effects.