janPiteJansekeROILA
[...]
Vocabulary design
[...]
As a starting point for the first version of the vocabulary of ROILA we choose the artificial language Toki Pona [6] which caters for the expression of very simple concepts by just 115 words. Therefore this number formed the size of the ROILA vocabulary. In order to maintain a balance between our two research goals we set the word length to 4, 5 and 6 characters, with each word having 2 or 3 syllables rendering the following word types: CVCV, VCVC, VCCV, CVCVC, VCVCV, CVCVCV, VCCVCV, VCVCCV, where V refers to a vowel and C to a consonant from our pool of 16 phonemes. The 8 word types were simple extensions of words existing in Toki Pona based on the assumption that such words would be easy to learn and pronounce. To define the scalable representation of the words we utilized a genetic algorithm that would converge to a vocabulary of words that would have the lowest confusion amongst them and in theory be ideal for speech recognition. The genetic algorithm randomly initialized a vocabulary of N words, for P vocabularies, where each word was any one of the 8 afore-mentioned word types. The algorithm was then run for G generations with mutation and cross-over being the two primary offspring generating techniques. The fitness function was determined from data available in the form of a confusion matrix (from [7]), where the matrix provided the conditional probability of recognizing a phoneme pj by a speech recognizer when phoneme pi was said instead. Therefore, the confusion between any two words was determined by computing the probabilistic string edit distance, as suggested in [1]. The first ROILA vocabulary was generated by running the algorithm for P=G=200. In order to have a benchmark of English words to compare against we set the English vocabulary as the meanings of the 115 Toki Pona words.
[...]
http://www.bartneck.de/publications/2010/designAndEvaluationROILA/
jansegers
jansegers