Natural Language Processing for Associative Word Predictions

A potential of natural language processing to predict words based on their associative relations is assessed on a Serbian language dataset. The approach investigated is to use a number of different written materials, and see if a neural network can extract word associations just by reading regular texts. Evaluation is performed using ready made, publicly available games of association. The final result exceeds 70% certainty of a correct solution within 100 most probable words in a vocabulary of over 150,000. Results are evaluated using different dataset sizes and preparation techniques with the idea of identifying impact of a particular approach.