关键词: 字母位置编码模型, 词汇识别, 转置效应, 模型对比

Abstract: This article compares and analyzes six classic letter position encoding models, exploring the role of letter position in visual word recognition and the underlying cognitive mechanisms. The models discussed include the Overlap Model, the Open Bigrams Model, the Sequential Encoding Regulated by Inputs to Oscillations within Letter Units (the SERIOL Model), the Spatial-coding Model, the Bayesian Reader, and the Positional Ordering of N-Grams (the PONG). Each model is grounded on a distinct cognitive theory, offering unique perspectives on how letter positions are encoded and simulating various effects observed in word recognition.
A common feature of these models is the claim that letter position plays a crucial role in word recognition. All models effectively account for several experimental effects, such as letter transposition effects and the effects of letter repetition, insertion, and deletion. Nonetheless, significant differences emerge in these models’ theoretical frameworks and conceptualizations of letter position encoding. The Overlap Model and the Bayesian Reader suggest that letter position encoding is highly susceptible to noise. The Overlap Model tolerates position shifts through an overlapping encoding mechanism, while the Bayesian Reader interprets position adjustment through a probabilistic framework, dynamically adjusting based on context to fit various cognitive scenarios. In contrast, the SERIOL Model encodes letter position via temporal activation patterns and multi-level structures (e.g., retinal layer, feature layer, letter layer), emphasizing the sequential and temporal characteristics of letter positions. The Spatial Encoding Model employs spatial phase coding and a two-dimensional coordinate system to process letter positions, enhancing encoding precision through external letter banks and word length modules. The Open Bigrams Model proposes a flexible encoding mechanism, focusing on the relative positions of adjacent letter pairs rather than strict letter order, thus offering a moderately flexible approach. The PONG introduces a theory of N-grams position encoding, positing that the position of N-grams is more critical than individual letter positions.
Consequently, these models differ in their explanations of certain experimental effects of letter transposition. The Open Bigrams Model suggests that transposed nonwords overlap more with base words in terms of letter pairs, thereby triggering a stronger transposition effect. The Overlap Model posits that letter representations are distributed across sequential positions. Thus transposed nonwords, having more overlap with base words’ position distribution, are recognized more quickly than substituted nonwords. The Bayesian Reader, calculates the probability of each letter’s position and views transposed letters as having a smaller edit distance from the target word, thus enhancing the match score. The SERIOL Model incorporates the Open Bigrams framework and argues that the transposition conditions activate more matching letter pairs than the substitute conditions, further strengthening the transposition effect. The Spatial Encoding Model emphasizes phase encoding of letter positions, suggesting that transposed letters preserve more stable relative phase relationships, thus facilitating better word matching. Meanwhile, the PONG activates matching N-grams and contends that transposed nonwords engage more appropriate N-grams with similar hemispheric lateralization to the target word, resulting in better recognition compared to substituted nonwords.
Additionally, each model offers unique insights into specific phenomena. The PONG Model, for example, explains the flanker effect using N-grams, a phenomenon has not been addressed by other models. Both the SERIOL and PONG Models account for optimal fixation locations through hemispheric lateralization. The Open Bigrams Model explains the influence of reading direction by positing that letter detectors are based on the relative position of eye fixations along the horizontal line.
Based on the analysis and discussions of the above models, there are several key directions for the further development of letter position encoding models. First, there is a need to expand the scope of models to account for factors like prediction, word frequency, and transposition across word boundaries, which remain insufficiently explored. Second, integrating word recognition models with eye movement control models could improve ecological validity in understanding letter position encoding in reading. Third, research into Chinese character position processing should be further explored. Fourth, more attention should be paid to cross-language differences and the differential experimental effects observed across languages. Finally, with advances in EEG and neuroimaging technologies, integrating diverse sources of data would optimize existing models and address theoretical gaps.

Key words: computational models of letter position encoding, word recognition, transposition effect, model comparison