International Core Journal of Engineering 2020-26 | Page 198
2019 International Conference on Artificial Intelligence and Advanced Manufacturing (AIAM)
A Novel Computer-Aided Emotion Recognition of
Text Method Based on WordEmbedding and Bi-
LSTM
Jia Zheng
The Harker School, San Jose, State of California 95112-7400, America.
Corresponding author: [email protected]
detection and multiple types of facial expression
classification based on support vector machine (SVM), and
used LBP histogram to obtain facial feature vector with
dual form. Literature [7] proposed a facial expression
algorithm based on convolutional neural network, which
uses EEG signals for classification. For example, literature
[8] proposed a convolutional neural network speech
emotion recognition based on recursive neural network
structure, and used 13 MFCC (mere frequency cestrum
coefficient) of speech signals as classification features.
Abstract—Emotional analysis in literary and artistic art is
a current research hotspot, while emotional-based fiction
classification is more complicated than other literary forms.
Emotional detection is usually done from the perspective of
voice or facial images. Most of the existing text-based research
uses typical binary or ternary classification methods.
Therefore, how to extract information of emotion from the
novel text and classify it is still a problem to be solved. In this
paper, a Bi-directional Long Short-Term Memory Language
Model (BiLSTM-LM) is proposed. The text sequence is
divided into six different emotional categories. Word
Embedding method is used to encode vocabulary, and the text
is represented as word feature representation and character
feature representation respectively. The final emotional output
is estimated by combining content words and emotional
function words. Experimental results show that the training of
the model in this paper can quickly converge, and the
detection accuracy is up to 64.09%, which is 4 percentage
points higher than the current best model, and basically can
effectively classify novels.
There is little research on extracting information of
emotion from text input. The literature [9] developed a
semantic network that extracts emotions from text content,
but the corpus supporting the results is rare. With the advent
of the era of big data, the scale of text data is also constantly
expanding. The deep learning method shows its advantages.
It is driven by data and can automatically obtain useful
feature knowledge from data without manual participation,
which avoid a lot of prior knowledge. Literature [10]
proposed a student feedback text academic emotion
recognition method based on LSTM model. Although the
recognition rate can reach 90%, it only uses binary
classification, namely positive emotion and negative
emotion. Literature [11] identified the basic six emotions by
analyzing the text, namely happiness, sadness, fear, surprise,
anger, and jealousy. However, the recognition rate is very
low.
Keywords—Novel Classification, Emotion Detection,
Language Model; Bi-directional Long Short-Term Memory
Network.
I. I NTRODUCTION
The transfer of emotion is an important role of literature
and art. As a member of a large family of literature, novels
also have the common character of literature - emotion. It is
a current research hotspot to classify novels according to
emotion. According to the emotional model theory [1], all
emotions can be divided into 3 categories, 6 groups and 22
types. Although there are many emotional theories, from
literature [2] study it is found that human beings have six
basic emotions, namely, happiness, sadness, fear, surprise,
anger and jealousy. Other emotions are combined with
these six basic emotions and belong to compound emotions.
For the text emotion analysis problem, this paper
proposes a Bi-directional Long Short-Term Memory (Bi-
LSTM), using a Language Model (LM), which is a
BiLSTM-LM. LSTM solves the problem of gradient
disappearance of RNN through a gate mechanism and can
effectively learn long-term relationships of dependency.
Combined with the language model, it acquires many
linguistic features of data, accelerates the convergence
speed of training, and improves the generalization ability of
the model. Through the comparative analysis of
experiments, the BiLSTM-LM model proposed in this paper
has significantly improved compared with other deep
learning models for the recognition of six basic emotions.
The research on emotion focuses on the mechanism of
emotion production, such as affecting factors of emotion,
emotional classification, emotional impact on people, while
the research on the judgment of human emotion is relatively
less [3]. Existing studies on emotional judgment are all
about expression, micro-expression and physical behavior,
which through the observation of other people's expressions
and behaviors, judge their emotions [4].
II. B ASIC T HEORY
A. Long Short-Term Memory
Long Short-Term Memory is a time recursive neural
network, which is suitable for processing and predicting
important events with relatively long interval and delay in
time series. LSTM is proposed to solve the problem of
“gradient disappearance” in the Recurrent Neural Network
(RNN) structure of RNN, which is a special kind of
At present, the research of emotion recognition
technology mainly focuses on three aspects: image, voice
and text. Literature [5] resolved the problem of
classification of smiles and neutral faces by using power
LBP features. Literature [6] proposed two types of emotion
978-1-7281-4691-1/19/$31.00 ©2019 IEEE
DOI 10.1109/AIAM48774.2019.00042
176