Syntactic Belief Update as the Driver of Garden Path Processing Difficulty

Imagine you're trying to understand how people read sentences that can lead them to the wrong interpretation, like 'The horse raced past the barn fell.' Initially, readers might think one thing, but then a word comes along that flips their understanding. This is a common challenge in language processing, especially with what's called garden path sentences. Traditionally, researchers have relied on a measure called lexical surprisal to predict how difficult a sentence will be to process. However, this method often falls short for garden path sentences, leading to inaccurate predictions about how long it takes someone to read them. This is what's called a failure of lexical surprisal.

To address this, the authors propose a fresh perspective: instead of focusing solely on the words themselves, they suggest tracking a probability distribution over the possible syntactic structures of a sentence as it unfolds. This means that as each new word is read, the reader's understanding of the sentence's structure is updated. If the reader is misled by the initial part of the sentence, this update will be significant when they reach the critical word that changes everything. The authors measure how much this belief changes using a concept called generalized Rényi divergence, which looks at the differences in these syntactic beliefs.

What this means for builders and researchers is that by shifting the focus from lexical items to syntactic structures, we can gain a better understanding of how people process complex sentences. This could open up new avenues for research in psycholinguistics and improve models that aim to mimic human language understanding.

The paper introduces a new approach to understanding sentence processing that diverges from traditional lexical surprisal metrics.

The claims are supported by empirical data on human reading times, though further validation could strengthen the findings.