Temporal Semantic Analysis and Visualization of Words

Today there are many languages spoken in the world, among which English is the most popular one. However, words in English evolved a lot in history such that it is very difficult for contemporary people to read ancient English articles. There are many changes, such as the mutation of word itself, the migration of word usage from one context to another, etc. It is thus very interesting to understand the temporal evolution of word’s semantic across a long span of time. In this paper we look at two datasets: the New York Times and the National Geographic to study the temporal evolution of words. For this purpose a model that can embed word into vectors is needed. Word2Vec is such an neural network model that learns a vector representation for each word in a way that similar words are also similar in the vector space. By similar, I mean that words tends to co-occur in the same context. So, to obtain a temporal Word2Vec representation, a temporal Word2Vec model needs to be trained sequentially, training one individual Word2Vec model for a given dataset in each time period. The temporal Word2Vec model allows us to explore different visualisation techniques of word semantic evolution. Temporal Word cloud, Heatmap, t-distributed stochastic neighbour embedding are some of the techniques that makes the visualisation possible.