How hierarchical is language use?

[T. Florian Jaeger]

> -----funknet-bounces at mailman.rice.edu wrote: -----
> To: Funknet <funknet at mailman.rice.edu>
> From: "s.t. bischoff"
> Sent by: funknet-bounces at mailman.rice.edu
> Date: 30/10/2012 12:14
> Subject: [FUNKNET] How hierarchical is language use?
>
> Some may find this article of interest...
>
> How hierarchical is language
> use?<
> http://rspb.royalsocietypublishing.org/content/early/2012/09/05/rspb.2012.1741.full.pdf+html
> >
> Stefan L. Frank1,*, Rens Bod2 and Morten H. Christiansen3
> Proceedings of the Royal Society
>
>
> It is generally assumed that hierarchical phrase structure plays a central
> role in human language. However,
> considerations of simplicity and evolutionary continuity suggest that
> hierarchical structure should not
> be invoked too hastily. Indeed, recent neurophysiological, behavioural and
> computational studies show
> that sequential sentence structure has considerable explanatory power and
> that hierarchical processing
> is often not involved. In this paper, we review evidence from the recent
> literature supporting the hypothesis
> that sequential structure may be fundamental to the comprehension,
> production and acquisition of
> human language. Moreover, we provide a preliminary sketch outlining a
> non-hierarchical model of
> language use and discuss its implications and testable predictions. If
> linguistic phenomena can be
> explained by sequential rather than hierarchical structure, this will have
> considerable impact in a wide
> range of fields, such as linguistics, ethology, cognitive neuroscience,
> psychology and computer science.
>
>
> I thought I'd throw in my 2 cents about that paper and link to an
additional paper of interest by the same authors:

Frank and Bod 2011 in PsychScience "Insensitivity of the Human
Sentence-Processing System to Hierarchical Structure", preprint available
here:
http://www.academia.edu/509509/Insensitivity_of_the_Human_Sentence-Processing_System_to_Hierarchical_Structure


In that paper, the authors provide evidence that at least *some* measures
of processing difficulty during reading are more affected by sequential
rather than hierarchical features. They show that probabilistic phrase
structure grammars  (PCFGs) are better at predicting words than simple
recurrent networks (SRNs). Crucially though, the probabilities
(expectations) derived from SRNs provide a better fit against human reading
times than the probabilities derived from PCFGs (several versions of both
models are evaluated and compared; specifically, as I recall, the models
are evaluated under the surprisal linking hypothesis proposed by Hale,
2001; Levy, 2008 ). The authors take this to argue that humans rely on
representations that resemble those inherent to SRNs more closely than
PCFGs. The methodology employed by the authors is rather thorough, so that
I would think the results would replicate.

Three caveats: first, it should be clarified is that SRNs, of course,
*do* hierarchical
structure, though their latent structure is typically considerably flatter
than PCFGs (and less pre-determined, of course). Second, the quality of the
PCFGs depends on the quality of the Treebank they are trained on. While the
fact that the PCFGs predicted words better than SRNs did suggests that the
Treebank structures were decent, this does not mean that the study tested
whether human sentence comprehension involves *any* hierarchical structure
(of a different nature). The authors acknowledge something similar to that
in the discussion. Third, the data presented in the main text are analyzing
first past fixations (i.e., the measure in reading that is perhaps expected
to be most sensitive to the earliest pieces of information that become
available during processing). I haven't yet checked
the supplementary information, which contains analyses of additional
measures.

Finally, I felt that in the "How hierarchical is language use?" paper some
of the clarity that the previous paper (discussed above) had got lost. The
idea that we switch back and forth between different 'parallel' sequential
processing stream (see discussion) is just a different way of stating that
language *is* hierarchical. What is left is much less controversial (and
rather insightful, like the first paper): much of processing does not
require any or at least not particularly 'deep' hierarchical structure, if
we're willing to accept a rich mental lexicon that includes lexicalized
structures (which we need to anyway).

I am cc-ing Stefan Frank to correct any wrong details in my summary of his
paper.

Florian