Improvement of Reverse Dictionary by Tuning Word Vectors and Category Inference

Abstract

A reverse dictionary is a system that returns words based on user descriptions or definitions. OneLook Reverse Dictionary is a commercial reverse dictionary system constructed from existing dictionaries. Hill (2016) reported another reverse dictionary system was constructed from public dictionaries using word embeddings and that its performance was comparable to that of OneLook Reverse Dictionary at the time of the comparison. In this paper we report that, by selecting word vectors suitable for a reverse dictionary and combining Convolutional Neural Network text classification, we improved the reverse dictionary described by Hill. It is very significant that our model can automatically construct a reverse dictionary system from publicly available resources such that it obtains similar scores to those obtained with OneLook Reverse Dictionary in accuracy@100/1000. We also show that our model can be used as a filter to the OneLook Reverse Dictionary to improve its performance.