31
relations. We can think of various types of grammars as graded in a hierarchy of increasing
computational power, namely the power to generate syntactic structures. In this hierarchy, finite-
state grammars are found at the bottom. Various types of phrase-structure grammars exhibiting
greater generative power and requiring more sophisticated computations are found in a graded
hierarchy above the bottom level (Chomsky, 1956).
Learning the rules of a finite-state grammar requires the ability to make certain kinds of
statistical inference. Specifically, it requires the ability to compute transitional probabilities:
Given the occurrence of one symbol, one needs to be able to compute the probability of the
symbol which is likely to appear next. For example, given that “The” was written, one can
compute the probability of its being followed by “boy” or by “planet” or whatever.
W. Tecumseh Fitch and Marc Hauser have recently gathered evidence showing that some
non-human primates can compute transitional probabilities (2004). Fitch and Hauser tested adult
human and tamarin monkeys for their ability to learn finite-state and phrase-structure grammars,
specifically their ability to distinguish strings of nonsense syllables which follow a pertinent rule
from those violating it. Humans and tamarins could both master finite-state grammars, but only
humans could grasp grammars that employ recursive embedding. There are, in fact, many data
suggesting that non-human primates are limited in their abilities to master the hierarchical
structures necessary for recursive embedding (Spinozzi et al., 1999; Conway and Christiansen,
2001). The matter is still controversial (Bergman et al., 2003; McGonigle et al., 2003) as is the
interpretation of Fitch and Hauser’s data (Liberman, 2004), but there is at least a suggestive hint
here as to why human intelligence seems qualitatively different from the forms of intelligence
found in other species. More work must be done.
At most, the studies performed by Fitch and Hauser show that tamarins cannot employ
recursive embedding in learning the formation rules for strings of syllables. The studies do not