Desktop version

Home arrow Language & Literature arrow Word and paradigm morphology


A 'learning-based' approach

The exploration of these issues is guided by two general hypotheses. The first is that form variation serves a fundamentally discriminative function, so that the function of a morphological exponent is best understood in terms of the forms that it distinguishes in a system, not what discrete meanings or properties it expresses in that system. The second is that the organization of a linguistic system is strongly influenced by its communicative function and by the constraints imposed by the process by which it is transmitted. A communicative perspective helps to clarify the role of irregular formations and other patterns that appear nonfunctional or even dysfunctional when language is regarded as a purely formal system. Taking learning into account likewise provides a positive function for regular patterns and offers a particularly natural interpretation of notions like ‘competition’ as a dimension of the learning process rather than as a component of the morphological system.

When carried to their logical conclusion, these hypotheses suggest a view of language as a complex adaptive system whose form strongly reflects communicative and learning constraints. Such a system is more aptly described in terms of

Word and Paradigm Morphology. First edition. James P. Blevins © James P Blevins 2016. First published 2016 by Oxford University Press dynamic communicative pressures and learning and processing principles than in terms of a deductive formal grammar. This general orientation is designated as a ‘learning-based’ perspective below, in part to emphasize the parallels with the communicative ‘usage-based’ tradition (Tomasello 2003), and in part to emphasize the role of a (discriminative) learning model.[1]

The hypotheses that guide a learning-based perspective are to some degree independent of the assumptions that underlie implicational WP approaches and it is possible to embed a WP model within a more conservative view of language. However, by combining a learning-based perspective with a WP approach, it is possible to clarify aspects of WP models that have remained largely unexamined. One cluster of such issues is summarized in (8.1).

(8.1) a. Why do morphological systems exhibit predictive dependencies?

b. How can the persistence of irregular formations be reconciled with the prevalence of regular patterns and what can their function be?

c. Why are notions of uncertainty and uncertainty reduction useful for measuring the variation and structure relevant to speakers?

It has long been known that the descriptive success of classical WP models reflects the interdependency of form variation, i.e., that “one inflection tends to predict another” (Matthews 1991:197). What is less well understood is why this should be the case. The usefulness of information theory for formalizing the classical WP model raises further questions about the probabilistic character of interdependencies. It is clear that these patterns are predominantly statistical but again less clear why this should be the case (particularly from the perspective of grammatical models based on categorical rule and constraint systems). The stable coexistence of regular and irregular formations in many languages raises a related issue. The models of morphological analysis developed within the formal grammar tradition can accommodate deviations from regular patterns but offer no insight into their function or resilience.

A learning-based approach attributes these patterns to factors that are very different in character from the formal constraints on units, representations or rule systems proposed within theoretical models. From a learning-based perspective, the organization of morphological systems is not anchored in a formal architecture or ‘innate language faculty’ but emerges mainly from the distributional biases of the forms in the system and the general-purpose learning strategies employed by speakers. Hence the first factor is just the structure of the input that speakers are exposed to. The second factor is the learning strategies that speakers employ when exposed to that input. The interaction of these factors suggests the answers in (8.2) to the questions in (8.1).

(8.2) a. Morphological systems exhibit predictive dependencies because, given the Zipfian structure of the input, speakers never encounter all the forms of a language and must be able to deduce new forms.

b. Irregular formations serve two useful communicative functions. As individual expressions, they are well discriminated. As exceptional members of larger sets of alternating elements, they emphasize contrasts that are less saliently marked in regular patterns.

c. Uncertainty reduction is relevant to speakers because learning a language involves the development of a predictive language model that reduces uncertainty about forms and distributions.

From a learning-based standpoint, the predictive dependencies exhibited by morphological systems are not due to abstract economy principles. Instead, interpredictability serves a very practical purpose; it is a prerequisite for the use and propagation of language, given the structure of the input that speakers encounter. Interpredictability is essentially a variety of regularity, and regularity aids generalization. Conversely, irregular forms are highly distinctive and communicatively useful but a much less reliable basis for extrapolation.

  • [1] There are also important points of contact with computational models that use temporal selforganizing maps (Chersi et al. 20Г4; Pirrelli et al. 2025), with a particular convergence at the level ofthe error-driven learning rules employed in these models.
Found a mistake? Please highlight the word and press Shift + Enter  
< Prev   CONTENTS   Next >

Related topics