Light before heavy

(Waiting to be posted since early June, sigh.)

The abstract for a thesis defense in Stanford Linguistics, on June 9th:

Consistency in Variation
by Robin Melnick

Language scholars as far back as Behaghel (1909) have noted the tendency to produce short before long. What underlies this Principle of End Weight? Several accounts that develop an incremental view of production — speakers produce what they can as soon as they can — may be seen as necessarily depending on some kind of constrained capacity view of human cognitive faculties. Such resources — for example, working memory capacity — generally vary normally across the population, while the preference for end weight has been previously modeled as fixed across the population. I argue that if the Principle of End Weight is motivated by limits on an underlying resource that varies by individual, we should see significant variation by individual in strength of preference for end weight. I undertake both corpus and experimental studies that confirm significant individual variation in strength of such end-weight preferences.

Next, we see the Principle of End Weight at play to varying degree in several English syntactic-alternation constructions — dative shift, verb-particle placement, prepositional phrase ordering, heavy NP shift. Looking across these constructions, if individual preference for end weight is motivated by a limited, underlying cognitive resource, such a resource should itself still be a constant within each individual across constructions. So while we see individual variation in the Principle of End Weight, a given speaker’s strength of preference should remain relatively steady across constructions — i.e., “consistency in variation.”

Methodologically, such a proposition is challenging to test. Existing parsed corpora of natural speech are generally built “shallow” — many speakers but insufficient contributions per speaker to model individual weight preference. To address this, I develop a “deep” parsed corpus — 500K to 2.5M words per speaker, as well as novel statistical approaches for exploring correlation of per-speaker preferences across models of separate constructions. Multiple studies with this corpus and complementary experiments indicate a significant measure of consistency in each individual’s end-weight preference across constructions, as predicted.

These results certainly don’t prove constrained capacity; other compatible accounts include the possibility that we develop individual production preferences through exposure to prior distributions. If so, though, the results above introduce the provision that in developing our individual production preferences, each speaker synthesizes a single rule — or at least a substantially consistent preference — across different syntactic-alternation constructions. This may be seen as a syntactic “neighborhood” effect. Conversely, had the studies above failed to find any consistency in variation, it would certainly have challenged the constrained-capacity accounts.

Finally, though, prior works suggested that Australians may show greater strength of preference for end weight in the dative-shift alternation than Americans, an effect of group that is compatible with exposure to differing prior distributions but not at all predicted by constrained capacity. If preference is variable by individual but consistent across constructions — as results above indicate — might we expect the same at the varietal level? I conducted an experiment with US, UK, and AUS groups and across four different constructions, finding that the Australian preference for end weight is significantly stronger across 3 of 4 constructions (with the fourth statistically neutral). While constrained capacity may be at play in the variation by individual, the cross-varietal results support the role of learning by exposure to prior distribution, with syntactic neighborhood effects appearing cross-construction both by individual and by group.

Nice idea, that preferences (even if motivated by processing considerations) can vary from individual to individual, even from dialect to dialect.

Leave a Reply

%d bloggers like this: