Files in this item

FilesDescriptionFormat

application/pdf

application/pdfECE499-Fa2018-liu.pdf (964kB)
(no description provided)PDF

Description

Title:World embedding models evaluation and word inference accelerator
Author(s):Liu, Zizhen
Contributor(s):Hwu, Wen-Mei
Subject(s):NLP
GPU
Word Embedding
Abstract:Many word2vec algorithms have been developed to represent vocabularies with more accurate number vectors. The most popular ones to date are word2vec, FastText, and GloVe. Despite their popularity, there has been a lack of published work that helps new users to understand how models trained by different algorithms are performing on words from different categories. Additionally, despite the great work being done in implementing word2vec training algorithms on GPUs, there has been a lack of published work in implementing word2vec inferencing algorithm on GPUs. Therefore, this thesis will first present our research on inference accuracy of models trained by word2vec, FastText and GloVe across multiple categories of datasets and explain why we see such differences in prediction accuracy. Secondly, the thesis will present an algorithm for performing batched word analogy queries running on GPUs, that can adapt to different models trained by any algorithm and thus can significantly reduce the efforts required in comparing models trained by different word vector generation algorithms.
Issue Date:2018-12
Genre:Other
Type:Text
Language:English
URI:http://hdl.handle.net/2142/105408
Date Available in IDEALS:2019-09-04


This item appears in the following Collection(s)

Item Statistics