Аннотации:
The article considers the problem of imageability ratings estimation of English words using artificial neural networks. To train and test the models, we use data of several freely available psycholinguistic databases. We compared two approaches based on different vector representations of words. The first approach uses pre-trained fastText vectors. The second one utilizes explicit word vectors built on the basis of co-occurrence statistics with the most frequent words extracted from the Google Books Ngram corpus. We employed the MRC Psycholinguistic Database to obtain the value of Spearman's correlation coefficient between imageability ratings and their estimations. The highest resulting value equaled 0.882. This significantly improves the results obtained in previous works. The approach proposed in this paper can be used to create large dictionaries with imageability ratings, which is important for many practical problems.