Dynamics of language and cognition


An aspect of cognition that is receiving a growing level of attention from many researchers concerns the properties of the mind as a dynamical complex system, placed at the border between complete order and complete disorder. From the seminal work of Zipf (1949), it is known that linguistic units exhibit power-law frequency distributions. More recently it has been shown that, within a text, linguistic units also exhibit long range, fractal style, correlations (see, e.g., Ebeling & Pöschel, 1994). In recent work, I have explored in much detail the multi-fractal properties of written texts, and their implications for cognitive processing (Moscoso del Prado, submitted). Among other things, one finds that the frequency and informational spectra of text follow multiple fractal regimes, which correspond directly to traditional linguistic levels of description: lexicon, syntax, pragmatics, etc. Furthermore, the analyses of corpora make it clear that the traditional stationarity assumption is a dangereous one to make when dealing with human language; the statistical properties of language change along discourse. Importantly, this non-stationarity follows a regular multi-fractal pattern. In recent research (Moscoso del Prado and Dunn, in prep.), we have modelled the process of diversification and extinction of human languages along history. We find that the dynamical fractal patterns present in linguistic utterances, also arise in the statistics of the historical evolution of languages.

Interestingly, similar fractal properties have been found in the dynamics of human behavioral sequences (Gilden, Thornton, & Mallon, 1995). In recent research I have shown that reaction times also exhibit power-law properties, in a very similar way to language (Moscoso del Prado, 2009). This power-law property of reaction times is universal: Independently of the task subjects perform, the power-law tail is identical. This provides a direct link to the measures of cognitive cost mentioned above, cognitive processing is reflected in the transient deviation from the power-law distribution, which represents a sort of baseline or "resting state" of the system in its natural state.  In terms of cognitive models, the power-law is directly predicted by an extremely simple, yet very powerful model of human reaction times (Moscoso del Prado, submitted). In recent research I have shown how this theory outperforms most current theories of reaction time distribution.

The fractal properties of human behavior are not necesarily a generalized property. In a recent methodological paper, I have shown how information-theory can be of use in locating the sources of fractal scaling in concurrent sequences of behavioral responses (Moscoso del Prado, 2011). In addition, the information theoretical model of behavior can itself predict whether responses will or will not follow a fractal pattern: Fractal properties arise when the system is placed close to a phase-transition between different cognitive mechanisms (Moscoso del Prado, submitted).

Finally, I have also developed a statistical method (and the corresponding R library implementing it) to test detailed hypotheses on the fractal scaling properties of time series (which is used in the studies above). The method, the Bayesian Assessment of Scaling,  is based on an analytical Bayesian formulation, and does not require Monte Carlo simulations (Moscoso del Prado, 2011).
x