Exploring the many uses of word embeddings to calculate textual similarity

Presenter(s): Lubos Steskal (TV 2 Norway)

In the recent years it has become a norm in Natural Language Processing to represent words and their semantics as vectors (known as word embeddings) and using these vectors to solve various downstream tasks. This presentation reports on the progress made by Nowegian broadcaster TV 2, and their experiences with using word embeddings to calculate the similarity between texts. This work is being used in search and recommenders on the broadcasters platforms and incorporates online news articles, automated transcripts of various television content such as news and political debates, and movie subtitles.