Lexical is as lexical does: computational approaches to lexical representation

Abstract

In much of neuroimaging and neuropsychology, regions of the brain have been associated with 'lexical representation', with little consideration as to what this cognitive construct actually denotes. Within current computational models of word recognition, there are a number of different approaches to the representation of lexical knowledge. Structural lexical representations, found in original theories of word recognition, have been instantiated in modern localist models. However, such a representational scheme lacks neural plausibility in terms of economy and flexibility. Connectionist models have therefore adopted distributed representations of form and meaning. Semantic representations in connectionist models necessarily encode lexical knowledge. Yet when equipped with recurrent connections, connectionist models can also develop attractors for familiar forms that function as lexical representations. Current behavioural, neuropsychological and neuroimaging evidence shows a clear role for semantic information, but also suggests some modality- and task-specific lexical representations. A variety of connectionist architectures could implement these distributed functional representations, and further experimental and simulation work is required to discriminate between these alternatives. Future conceptualisations of lexical representations will therefore emerge from a synergy between modelling and neuroscience.

Keywords: computational models; lexical representation; orthography; phonology; semantics; word recognition.

Figure 1.. A schematic representation of activation of units encoding (a) localist representations and (b) distributed representations. In (a) representation of six words requires six units. In (b) representation of twelve words also requires six units. In (b) the representations are distributed at the word level but localist at the letter level for the purposes of exposition. A fully distributed scheme would have the capacity to represent many more words.

Figure 2.. A schematic diagram of a model of visual and auditory word recognition and production showing the location of hidden unit layers that could house distributed functional lexical representations in the form of attractors. Note bidirectional connections in all cases bar those from input. Additional hidden layers are shown in black. Within level connections are shown with U-shaped arrows.

Figure 3.. A version of the previous model of visual and auditory word recognition and production containing one large set of hidden units. Learning in the network occurs under a topographic bias that favours short connections. This allows graded modality specificity to emerge in the network, such that units close to a particular input or output participate more in tasks involving them, while units close to the centre are increasingly amodal.