International Core Journal of Engineering 2020-26 | Page 57
2019 International Conference on Artificial Intelligence and Advanced Manufacturing (AIAM)
Short-text Sentiment Classification based on Graph-
LSTM
Yuting Wan
School of Control and Computer Engineering
North China Electric Power University
Beijing, China
[email protected]
A. RNN and LSTM
The long and short memory neural network, often called
LSTM, is a special RNN that learns long dependencies. They
were introduced by Hochreiter & Schmidhuber and improved
and popularized by many people. They work very well on a
variety of issues and are now widely used LSTM is designed
to avoid long dependency problems. Remember that longer
historical information is actually their default behavior. All
recurrent neural networks have the form of a repeating module
chain of neural networks. In a standard RNN, the repeating
module will have a very simple structure.
Abstract—With the rapid development of social networks,
short texts make it easier for people to comment on hot events
and express their emotions. The information published by users
through short texts contains emotional features of different
trends, and it is of great significance to dig deep into these
features. However, the traditional RNN algorithm, SVM and
linear LSTM can only discriminate emotional sentiment because
of the short text grammar and the sparse data, which is far from
the purpose of opinion mining. Therefore, this paper proposes
to apply Graph LSTM to short text classification, mine deeper
information, and achieve good results. Finally, the paper
compares three different machine learning methods to achieve
fine-grained sentiment analysis.
h 0 h 1 h 2 h 3 h 4
A A A A A
X 0 X 1 X 2 X 3 X 4
Keywords—Graph-LSTM, Sentiment Classification, Short
text
I. I NTRODUCTION
With the rapid development of Internet technology and its
application, some new social media including Weibo, social
networking sites, instant messaging, etc. are fundamentally
changing the lives of human beings. However, due to the large
number of users and the flood of information, each Weibo has
a limited number of words, plus the use of many online
languages or symbols, how to discover social sensation, hot
topics and sensation about national security from Weibo.
Trends, how to find people's evaluation of certain events or
commercial products have become the focus of research.
Among them, sentiment analysis in short text natural language
processing is one of the key topics. The main purpose of short
text sentiment analysis is to identify subjective information
from short text information and to mine the opinions and
attitudes held by users on product, news, hot events and other
commentary information, provide a basis for enterprises to
understand the user's consumption concept and improve
product quality. Secondly, as a social manager or government,
relying on sentiment analysis technology to explore the
opinions and emotions of Weibo users on certain hot events or
certain policies, we can understand the focus of public opinion,
improve the government's policies and policies, and prevent
sudden incidents. Real-time analysis of information has
important reference value.
Fig. 1. Linear LSTM propagation
B. Graph LSTM
In Graph LSTM, the back propagation method is different
from the general linear structure LSTM, so the propagation
mode needs to be redefined. For linear LSTM, parameter
updates propagate from left to right or right to left along the
chain structure. For Graph LSTM, parameter updates are
propagated along the graph structure formed by the nodes.
Traditional LSTM is essentially a deep feedforward neural
network. For example, from left to right there is a hidden
linear vector corresponding to each word, which is generated
by the basic loop unit of the neural network, as the word
embedding of the given word and the hidden vector of the
previous word. In discriminative learning, these hidden
vectors update the gradient in the entire network by
backpropagation and finally input to the end classifier.
Extending this strategy to the graph, in essence, is
backpropagation according to the graph structure through the
feedforward neural network, and the calculation result of each
node is propagated to the adjacent node. The nodes on the map
are composed of one case. The relationship between the plot
of the similar case and the sentence of the trial requires a large
amount of data to be accurate, and the nodes are linked to form
a graph structure. Similar to the gradient backpropagation
method, there is Confidence Propagation (LBP). However, the
II. R ELATED RESEARCH
According to the different methods of classification
knowledge acquisition, the text automatic classification
system can be divided into two types: classification system
based on knowledge engineering and classification system
based on statistics.
978-1-7281-4691-1/19/$31.00 ©2019 IEEE
DOI 10.1109/AIAM48774.2019.00014
35