Dual-rout e Conn ec tioni s t Mod e l of Gr ee k Sp e lling Ioanna Katidioti (ikatidio@phs. uoa. gr) Graduate Program in Cognitive Science, Athens University Campus GR-15771 Athens, Greece Ian C . Simpson (is impson@ugr . es) , Campus de Cartuja s/n 18071 Granada, Spain Athanass ios Protopapas (aprotopapas @phs. uoa. gr  Department of Philosophy & History of Science, University of Athens GR-15771 Athens, Greece K e ywords : connectionist model; spelling; phoneme-grapheme mapping Introdu c tion We created a dual-route connectionist model of Gree k spelling. The model maps sequences of phonemes to corresponding sequences of graphemes, using a sublexical and a lexical route, i.e., phonographemic information and word knowledge, respectively. It is based on the model of Houghton and Zorzi (2003), but handles words up to 5 syllables long, with full connectivity between the syllables. Greek has 37 phonemes and 84 graphemes related via 118 mappings with 80,3% consistency (spelling) (Protopapas & Vlahou, 2009). Model architecture is as follows: Figure 1: Dual-route model of spelling Input-Output R e pr ese ntation The representation is syllabic and nucleus-centered. There are 4 consonant slots on each side of the vowel. The orthographic slots are occupied by graphemes, not letters. Figure 2: Input and output representation Training and parame t e r s To simulate spelling development using   we trained the model to a corpus of 30,391 words from elementary school books. The model was trained for 30 epochs, with learning rate 0.02 and no weight pruning. During spelling, feedback was set to a value of 0.2. R es ults Using both routes, the entire training set is spelled correctly. Using only the phonological route, 65.2% of the training set is spelled correctly and almost all errors are phonologically plausible. By adding a small contribution from the lexical route we were able to simulate Grade 3-4   of 48 words. In the simulation, 13 out of 14 mistakes were the same as those made by the children, and 11 of these were the most typical. Probl e ms The model made two kinds of phonologically implausible mistakes: it spelled /s/ inside 19 words with    only used word-finally) and it also omitted a grapheme in a few words. In addition, the model has two problems: (a) the number of cycles needed to compute the output  always correspond to the difficulty of the word and (b) certain palatal consonants were consistently misspelled  . Empiri c al validation Greek has a number of ambiguous phonemes, the alternative spellings of which appear with different frequency (Protopapas & Vlahou, 2009). For example, in our training corpus, the phoneme /o/ is spelled with the letter         Due to frequency-sensitive training the model usually spells the ambiguous phonemes with the highest-frequency grapheme. However, due to asymmetries in the distribution of consonant-vowel co-occurrences, this is not always the case. That is, the model will use the less frequent graphemic variant of a phoneme when more likely in the particular phonographemic context. If the model corresponds to human spelling performance, 315