Detecting Counterpart Word Pairs across Time

Abstract

Most search engines require users to input specific words to obtain useful results from the Web. However, this requirement is sometimes challenging for users who want to search for information regarding unknown words. It may be especially difficult to learn about the past without knowing suitable words, such as answering the question "Who was the counterpart of the Prime Minister of the UK in the Ottoman Empire in 1900?" We propose a novel search framework for finding counterpart relationships represented by word pairs across time. This framework detects the counterparts by arithmetic operations as well as Word2Vec. To improve the accuracy, our framework groups news articles to develop context for words. After embedding the words into vector spaces, we map a given relationship to another vector space, then perform arithmetic operations. We show our algorithm outputs better results compared with simply applying Word2Vec.