Morphology in a Parallel, Distributed, Interactive Architecture of Language Production

Abstract

How do speakers produce novel words? This programmatic paper synthesizes research in linguistics and neuroscience to argue for a parallel distributed architecture of the language system, in which distributed semantic representations activate competing form chunks in parallel. This process accounts for both the synchronic phenomenon of paradigm uniformity and the diachronic process of paradigm leveling; i.e., the shaping or reshaping of relatively infrequent forms by semantically-related forms of higher frequency. However, it also raises the question of how leveling is avoided. A negative feedback cycle is argued to be responsible. The negative feedback cycle suppresses activated form chunks with unintended semantics or connotations and allows the speaker to decide when to begin speaking. The negative feedback cycle explains away much of the evidence for paradigmatic mappings, allowing more of the grammar to be described with only direct form-meaning mappings/constructions. However, there remains an important residue of cases for which paradigmatic mappings are necessary. I show that these cases can be accounted for by spreading activation down paradigmatic associations as the source of the activation is being inhibited by negative feedback. The negative feedback cycle provides a mechanistic explanation for several phenomena in language change that have so far eluded usage-based accounts. In particular, it provides a mechanism for degrammaticalization and affix liberation (e.g., the detachment of -holic from the context(s) in which it occurs), explaining how chunks can gain productivity despite occurring in a single fixed context. It also provides a novel perspective on paradigm gaps. Directions for future work are outlined.

Keywords: degrammaticalization; interactive activation; language production; negative feedback; paradigm leveling; paradigm uniformity; parallel processing; usage-based linguistics.

Figure 1

The initial, feedforward flow of activation. s = semantics. f = form. Input activation from the message shown by sourceless arrows at the top, which activate the semantic vector corresponding to the product form that is the best match to the intended message. The activated semantics are partly shared with the source form (s_{SourceProduct}) and partly not (s_Product). The shared semantics activate both forms in proportion to how well this vector predicts each form, with more activation reaching the frequent form compared to the less frequent product form (arrow width). This imbalance is what creates paradigm leveling. The additional activation received by f_Product from the meaning not shared with f_Source prevents paradigm leveling. Paradigm leveling occurs if f_Source is activated more than f_Product.

Figure 2

The second stage, negative feedback: the activated forms inhibit associated semantics (inhibition is shown by the circled minus signs). The amount of inhibition reaching a semantic feature from a form (and then available to spread back to the form) is proportional to the strength of the connection from the form to the semantic feature. In this example, f_Source is a strong cue to unintended semantics, S_Source, which is not part of the message. After feedback, the amount of activation spreading down from an S vector to a form is the difference between the top-down excitation S is receiving from the message (sourceless arrows) and the amount of inhibition it is receiving from the forms. Here, S_Product is receiving much more excitation than inhibition and will continue exciting f_Product. S_{SourceProduct} is receiving approximately equal excitation and inhibition and will no longer be a strong factor in form activation. The inhibition received by S_{SourceProduct} helps f_Product over f_Source because f_Source is favored by S_{SourceProduct} due to its higher frequency. Finally, S_Source is only receiving inhibition, and will begin inhibiting f_Source.

Figure 3

The use of negative feedback and paradigmatic associations to repair a plan that includes a form that does not fully match the intended meaning. The product form is associated with, and receives activation from, the distributed activation pattern specifying a source form, an intended meaning and an unintended meaning. It is also receiving the top-down activation from S_Product remaining after negative feedback. S_{SourceProduct} has been inhibited by the feedback and is no longer sending much activation. S_Source is sending inhibition down to f_Source.