Syntactic Node Count as Index of Predictability

Kohei Kajikawa, Shinnosuke Isono

Keywords: sentence processing; Node Count; reading time; predictability

TL;DR: We model reading time data and show that syntactic Node Count reflects predictability effects rather than processing complexity as usually assumed.

Abstract: Recent neurolinguistic studies of sentence comprehension often use Node Count as the linking hypothesis connecting syntactic structure building and neural activity. The assumption behind the use of Node Count is that syntactic structure building incurs cognitive effort. If so, larger Node Count should lead to increased processing difficulty. However, several studies using English and Japanese datasets have reported that higher dependency-based Node Count is associated with faster reading times. These findings challenge the interpretation of Node Count as an index of processing effort and instead point toward an alternative explanation in terms of predictability or anti-locality effects. Building on these findings, we ask the following questions: (i) Is the effect of Node Count on reading times negative across datasets and grammar formalisms? (ii) If so, do Node Counts based on different formalisms have independent effects? (iii) If syntactic Node Count indexes predictability, is it independent from large language model surprisal?

Submission Number: 7