Can lossy-context surprisal capture both locality and anti-locality effects? A model evaluation using Russian, Hindi, and Persian data

Le résumé fourni par la source

A signature prediction of memory-based theories of sentence processing is that processing difficulty should increase with increased dependency distance. However, in addition to this so-called locality effect, an anti-locality effect has also been reported, where increased dependency distance leads to a reduction in processing difficulty. Anti-locality effects are a key prediction of expectation-based accounts; these accounts predict that increasing dependency distance can make the upcoming word more predictable, and therefore easier to process. A theory that has been argued to explain both locality and anti-locality effects within a unified framework is lossy-context (LC) surprisal (Futrell, R., Gibson, E., & Levy, R. P. (2021). Lossy-context surprisal: An information-theoretic model of memory effects in sentence processing. Cognitive Science, 44(3)). LC surprisal extends the expectation-based account with a memory component that lets memory for the linguistic context degrade and be reconstructed, and this affects the surprisal computation. Locality or anti-locality effects can arise depending on how much the context has degraded, and how it is reconstructed. Although this claim has been investigated by Futrell and colleagues using data from English and German, it has not been put to the test against other cross-linguistic experimental data in which both memory and expectation effects are in play. We present an evaluation of LC surprisal theory against three empirical studies involving Russian, Hindi, and Persian, which found evidence for both memory- and expectation-based effects. We show that the lossy-context surprisal model as laid out in Futrell et al. (2021) has only limited success in explaining the observed patterns in these data. Alternative implementations of LC surprisal need to be developed that align better with the observed data.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.