2017 Seventh International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW)
A Multi-granularity Data Augmentation based
Fusion Neural Network Model for Short Text Sentiment Analysis
Xiao Sun
School of Computer and Information
Hefei University of Technology
Hefei, Anhui, 230009
Email: sunx@hfut.edu.cn
Jiajin He
School of Computer and Information
Hefei University of Technology
Hefei, Anhui, 230009
Email: hejudgin@mail.hfut.edu.cn
Changqin Quan
Department of Computational Science
Kobe University
Kobe, Japan, 6578501
Email: quanchqin@gold.kobe-u.ac.jp
Abstract—Sentiment analysis is a challenging task in Natural
Language Processing due to the complexity of language struc-
ture, the semantic structure, and the relative scarcity of labeled
data and context information especially in the field of short-text
processing. To overcome data sparseness and the over-fitting
problem when adopting a deep learning model, we propose
multi-granularity text-oriented data augmentation technologies
to generate large amounts of data for neural network train-
ing. We propose a novel confused model (LSCNN) with the
proposed data argumentation technology that improves the
performance and outperforms other effective neural network
models. The proposed data augmentation method enhances
the generalization ability of the proposed model. We also show
that the proposed data augmentation method in combination
with the neural networks model can achieve astonishing per-
formance without any handcrafted features on cross-domain
sentiment analysis, which is a efficient technology for comments
sentiment detection.
1. Introduction
Sentiment analysis [1] is commonly used to mine the
user’s perception of a product and the user’s sentiment for
chat robots [2]. Effective sentiment analysis can obtain the
user’s subjective feelings towards some product and adjust
the service timely.
Recently, a neural network-based sentiment analysis
model became popular. A large amount of data is necessary
to train effective models. However, extremely deep neural
networks may lead to over-fitting. In order to solve the
problems, the idea of data argumentation is introduced into
the neural networks model. Neural network-based architec-
tures have achieved complete success in the field of Natural
Language Processing (NLP), such as Convolutional Neural
networks(CNN) [3] , Recurrent Neural networks(RNN) [4],
[5], and Long-Short Term Memory(LSTM) [6]; however,
these efforts were adversely affected by the lack of large-
scale data for training.
We introduce a data augmentation method to generate a
larger dataset for pre-training and training neural networks
for sentiment classification, which has been widely applied
in image processing [7], [8] and sound processing [11]. The
proposed data augmentation technology has been applied to
neural network-based models such as Convolutional Neu-
ral Networks [3], Long-Short Term Memory [6], [9] and
BOW-based SVMs model [10]. We show that the proposed
neural network models with data augmentation outperform
models without it and the BOW-based model. The crucial
contributions are as follows:
(1) We propose multi-granularity text-oriented data augmen-
tation technologies to automatically manufacture artificial
data to overcome the problem of data sparseness in NLP.
(2) We firstly proposed a data augmentation-oriented hybrid
neural network model called LSCNN and successfully apply
the proposed model to sentiment analysis and obtain signif-
icant improvements and enhance the generalization ability.
(3) The proposed LSCNN model is almost automatic and
independent of any manual features and other resources.
The remainder of the paper is organized as follows.
2 introduces the proposed data augmentation technologies.
Section 3 describes the proposed model. Section 4 reports
the experiments and evaluation results with and without the
data augmentation method. The conclusion and future work
are provided in the final section 5.
2. Data Augmentation
Data augmentation has been successfully applied to
image classification [7], [8]. The most convenient and
common way to avoid over-fitting when training a neu-
ral networks model is to automatically enlarge the dataset
using data augmentation technologies. There are many
methods for data augmentation for image data, such as
rotation/reflection, flip, zoom, shif t, changing the
scale and color
, contrast, and introduce noise. Our
approach involves using data augmentation technologies on
text sentiment analysis due to the great success achieved
with image classification. In this paper, we firstly propose
a multi-granularity data augmentation method, including
word-level, phrase-level, and sentence-level data augmen-
tation. We exploit some special ways for text data by
leveraging the characteristics of text as follows.
978-1-5386-0680-3/17/$31.00
c
2017 IEEE
12