Cross-lingual word and document embeddings

Lausanne Machine Learning Meetup, 31.05.2018 Martin Josifoski, EPFL Cross-lingual word and document embeddings Distributed word representations give us the great possibility to geometrically compare the semantic similarity between words in one language. The improvements they introduced to most of the downstream NLP tasks, inspired a lot of research that tries to similarly embed the semantical meaning across languages. Although successful to some extent, the limits of current work lie in the cost and the available amount of the data required (usually word translations or sentence translations between languages), as well as the scalability obtainable in training. We propose an approach that alleviates both of these issues by exploiting Wikipedia as a knowledge base and training by utilising only highly optimised linear algebra routines. The quality of the embeddings is validated in the cross lingual information retrieval setting. Organizers: Pawel Rosikiewicz (SwissAI.org, University of Lausanne) https://www.linkedin.com/in/pawel-ros... Prof Martin Jaggi (EPFL) https://www.linkedin.com/in/martinjaggi/ Clement Charollais (EPFL) Camera operator and editting https://www.linkedin.com/in/clément-c... For more information and future events please see: https://www.meetup.com/Lausanne-ML-TGIT/ Sposors: École Polytechnique Fédérale de Lausanne (EPFL) https://www.epfl.ch Unit8 http://unit8.co