Application of Neural Networks for Automatically Detecting Verbal Accents
The problem of automatically detecting verbal accents is solved. Word classes with one and two non-transitive accents, with transitive accents, and without accents are identified. An accent is determined in words in which it is not transitive. Words are grouped by the number of syllables. Each group is divided into word classes with the same numbers of accented syllables. Thus, the accents determination problem solved by means of neural networks boils down to word classification. The data array (training and test sets) is formed from A.A. Zaliznyak's Russian language grammatical dictionary, which contains word forms with placed accents. A word model comprises a list of syllables. In the data array, syllables are replaced by their numerical codes, for which syllable dictionaries are compiled. The numerical code of a syllable is its number in the syllable dictionary. The accents are searched in two stages. First, it is found out whether the word has non-transitive accents, and if yes, the word is transferred to the neural network that determines the accents. All neural networks designed in this study contain an Embedding layer which translates scalar representations of word syllables into vector ones. At its input, the neural network receives a vector with the numerical codes of word syllables, and at the output it yields the word class number, which in the case of one non-transitive accent coincides with the number of the accented syllable, and in the case of two non-transitive accents indicates the numbers of two accented syllables. The probabilities of correctly determining one and two non-transitive accents are estimated at 0.9474 and 0.9759, respectively.