Understanding bit by bit: information theory and the role of inflections in sentence processing

Abstract

What makes a sentence hard to process? Apart from the meanings of the words it contains, their number, and the way these words combine into constituents, words also contribute to processing difficulty on the basis of their accessibility in lexical retrieval. Apart from their frequency of use or their complexity in form, the words’ accessibility is also influenced by the number and roles of related forms in the structure in which they are stored. This is a factor that so far has not been sufficiently taken into account. As is experimentally shown in this dissertation, a measure for this is the inflectional entropy of the paradigm, an information theoretic measure that quantifies the support that a word can receive from its inflectional paradigm during activation. This study investigates how the interplay between the available processing resources, as predicted by the inflectional entropy, and the linguistic constraints, modulates the speed of processing within and between-sentences, as well as within- and between-languages. I follow the ACT-R model of sentence comprehension, proposed by Lewis and Vasishth (2005), which states that processing delays in working memory result from interference during retrieval induced by the similarity between the target and the already processed items. Assuming that interference effects are present across the board, I focus on the stage of lexical activation (the stage before interpretation) and on verbs, in particular. The experimental results reported in this dissertation demonstrate that a verb's accessibility in long term memory, what I call the Activation Potential (AP), is proportional to the inflectional entropy of its paradigm, determining the speed, and thus the effort, with which it is processed, and also, crucially,modulating the processing of the rest of the sentence. In fact, the processing speed of a reflexive object, like the Dutch zichzelf, and unlike a definite NP like Maria, depends on the way the main verb was processed, providing evidence that the reflexive’s interpretation requires an operation on the verb. The processing speed of dependent elements in a sentence follows the “easy-to-activate-hard-to-re-access” principle and can either be boosted or delayed by the amount of available processing resources: reflexives will be processed more slowly as an object of a quickly processed verb than as an object of a hardly processed verb, unless the processing system abounds in resources and can boost the slow processing of the former. This principle holds within a sentence but is more apparent at a between-language level. More precisely, languages with rich inflection, like Greek, although they have longer words and more complicated paradigms, nevertheless benefit in terms of processing effort, at least during theinitial processing stages. This is because they have verbs with higher inflectional entropy which require fewer resources during first activation when compared to languages with poor inflection, like Dutch. Saving on resources at the first stages of sentence processing can boost computations that are costly but can also lead to excessive processing effort in subsequent steps. I take this to be a reflection of the necessary trade-off between space and information for the sake of successful real-time computations.