The representation of letter strings : psychological evidence and computational models

Abstract

Two ways of representing the spatial arrangement of letters in letter-strings are distinguished. In part-whole representations, the relationship of a letter to the letter-string as a whole is encoded. In part-part representations, the relationships of a letter to other letters in the string are encoded. Computational models of word perception typically use the former, but part-part representations are a very general feature of some neurocomputational models. Experiments ·are reported that examine for nonword and word wholes the representations used to encode their constituent parts; the first five experiments use measures of facilitation to infer encoding type, the next three primarily use error measures.

Experiment 1 shows that when a part of a recently learned letter-string is maintained in a briefly-presented test string, the test string is more accurately reported, showing perceptual transfer of training. No significant difference in the amount of transfer is found between maintaining the part in the same position (fixed-part) in the string and maintaining the part in a different position (moved-part) in the string. It is argued that this confirms part-part theories because transfer was obtained when only inter-letter relationships are maintained. Experiment 1 simulated on two implementations of part-whole theories shows that they fail to produce the obtained pattern of performance. This indicates that part-whole relational encoding is not a major part of the representations mediating these transfer effects. Experiment 2 replicates the fixed-part transfer and shows that it is restricted to parts made of adjacent letters. Experiments 3 and 4 use a prototype-extraction paradigm to show that novel parts made of adjacent letters are easier to learn than parts made of non-adjacent letters. Experiment 5 eplicates the moved-part transfer and shows that it is restricted to parts made of adjacent letters. These results show that the major inter-letter relationships encoded are between neighbouring letters. These first five results are taken as showing that pre-processing of the image to provide position-in-the-string information is not important for the representations that produce transfer. It is suggested that modelling the input to the graphemic input lexicon as the Primal Sketch of the image is more appropriate. In particular, realistic early vision algorithms such as MIRAGE appear to be potentially capable of modelling the results obtained.

Experiment 6 shows that reports of letters in nonwords have gradients of positional accuracy, with most positional errors occurring close to the correct position. Experiment 7 finds that migrations into the report of the second of two briefly-presented nonwords from the first nonword do not always maintain position though many do. Experiment 8 involved the presentation of mis-spelled words preceded by non words that either encouraged the detection of the mis-spelling or its lexicalisation. Lexicalisation responses involve the migration of a letter from the preceding string. These occur when primed by the lexicalisation letter in the same, but not in moved, positions in the first string, but only when presented in the context of neighbouring letters. Detection of mis-spelling shows both facilitation and inhibition. Facilitation is obtained with the part in moved positions in the source string but not in the same position, in which case inhibition is found. Facilitation is also obtained by prior presentation of the misspelled word or prior presentation of the correctly spelled word. These results are interpreted as showing that facilitation is obtained when the facilitating part of the preceding string either fully or minimally activates a representation of the word mis-spelled on second presentation. Partial activation of the word produces inhibition.

The results suggest that part-whole encoding is used for letters in familiar wholes, while part-part encoding is used for letters in unfamiliar wholes. This conclusion is used to motivate a model of the organisation and access of graphemic representations in which the ~ "» -. concept of scale plays an important role. The model is extended to other tasks involving visually presented words and nonwords and a brief account of the major findings attempted. Finally some extensions of the model to the domain of object perception are outlined.