Home // SEMAPRO 2015, The Ninth International Conference on Advances in Semantic Processing // View article

Deep Learning for Large-Scale Sentiment Analysis Using Distributed Representations

Authors:
Kazuhei Katoh
Takashi Ninomiya

Keywords: sentiment analysis; deep learning; distributed representations.

Abstract:
This paper presents the performance evaluations of deep learning classifiers for large-scale sentiment analysis using Rakuten Data. Many NLP theories and applications use 1-of-K representations for representing a word, but 1-of-K representations are difficult to use with many deep learners because they are vectors consisting of millions of dimensions. To reduce the number of dimensions of 1-of-K representations, we used distributed representations for words by using word2vec. Two experiments were conducted: (1) sentiment analysis using a small data set, the IMDB dataset, and (2) sentiment analysis using a large-scale data set, Rakuten Data. In the experiments, we observed that multi-layer neural networks did not work well for the small data set (i.e., neural networks without hidden layers achieved the best result), but multi-layer neural networks worked well for the large-scale data set. In the experiments using Rakuten Data, we tested the neural networks with 0-6 hidden layers, and neural networks with three hidden layers achieved the best result.

Pages: 92 to 96

Copyright: Copyright (c) IARIA, 2015

Publication date: July 19, 2015

Published in: conference

ISSN: 2308-4510

ISBN: 978-1-61208-420-6

Location: Nice, France

Dates: from July 19, 2015 to July 24, 2015