Seminars

Download as iCal file
By Ira Leviant
Location Bloomfield 152, Faculty of Industrial Engineering and Management
Advisor(s): Roi Reichart
PhD
Academic Program: Please choose
 
From Wednesday 01 January 2020 -  11:30
To Monday 06 January 2020 - 12:30
Abstract:
State-of-the-art word embeddings provide high quality representations of nouns compared to verbs and adjectives. In this work we show that using naive iterative algorithm improves significantly word2vec verb and adjective similarity performance. Our algorithm iteratively extracts both symmetric and asymmetric patterns from the corpus. These fully unsupervised pattern contexts are superior to previous bag-of-words context based representations. Our naive approach is easily applicable in other languages and leads to superior performance also in a multilingual setup.
Bio:
Ira Leviant is a PhD student at the Industrial Engineering and Management faculty of the Technion, working with Professor Roi Reichart. Ira's research focuses on multilingual semantic vector space models. In her research, Ira released two prominent data sets: Multilingual WS353 and Multilingual SimLex999. These data sets are widely used among the Natural Language Processing (NLP) community for the evaluation of multilingual vector space models.