Word Topic Prediction Model for Polysemous Words and Unknown Words Using a Topic Model

Abstract

A prediction method for word topics is proposed. Sentences are sometimes difficult to read if the meaning of one or more words is unknown. However, they become easily readable if the topics of the words are known. Therefore, a topic model in which the words are divided by topic and several words are chosen from the available topics was used to predict the topics. However, the model performed poorly when dividing polysemous and unknown words by topic. To address this problem, a new method was formulated by adding two new elements to the conventional method. One of the elements comprised separating polysemous words according to their examples. The other element involved associating unknown words with sentences composed of known words. Several experiments were conducted using the proposed method, which achieved accuracies of \(\sim \)60% and \(\sim \)80% for dividing polysemous and unknown words, respectively, by topic. Such results imply the superiority of the proposed method for these two tasks.