International Core Journal of Engineering 2020-26 | Page 201
character feature information of text, a two-channel LSTM
neural network model is designed in this paper. The model
integrates word feature information and character feature
information of text, and can learn feature representation of
text in a deeper layer. The structure of the model is shown
in Figure 7.
Firstly, WordEmbedding is used to encode vocabulary,
and text is represented as word feature representation and
character feature representation respectively. Then bi-
directional LSTM is used to learn higher-dimensional
hidden features from two sets of feature representation.
h w
h c
can randomly disable some hidden nodes in the network
during model training.
Finally, the model completes the task of emotional
recognition of text through Sigmoid layer. The Sigmoid
layer accepts the output of the previous layer as input, and
outputs a one-dimensional vector of length 1. The value of
this vector is mapped between 0 and 1 by Sigmoid function,
which is used as the model to predict probability of whether
the text contains certain emotions.
IV. A NALYSIS OF E XPERIMENTAL R ESULTS
In the experiment, TensorFlow deep learning framework
is used for development and testing, and CUDA is used for
parallel acceleration of GPU training process. The
configuration parameters are shown in Table I.
LSTM w ( T w )
(2)
LSTM c ( T c )
Among them, h w and h c represent the features of word
hidden layer and character hidden layer learned by LSTM
respectively. In order to get the character fusion feature of
text, h w and h c are stitched and integrated together.
h w c
TABLE I. T HE E XPERIMENTAL P LATFORM P ARAMETER C ONFIGURATION .
Configuration
GPU
CPU
RAM
Operating System
Development Platform
Development Language
Framework
h w h c (3)
Among them h w c is the word fusion feature of the text,
and the symbol represents the stitching of vectors.
Parameter
NVIDIA GeForce GTX 1080Ti 8GB
GDDR5X
Intel(R) Core (TM) i7-8700 [email protected]
4.6GHz
64GB DDR4 2133MHz
Linux Ubuntu 14.0
Anaconda3
Python 3.5
TensorFlow 10.0
The network model shown in figure 7 was used to train
37617 texts and identify the six basic emotions-happiness,
sadness, fear, surprise, anger and jealousy. The curve of the
loss and recognition accuracy during training along with the
epoch was shown in Figure 8 and Figure 9.
Next is a full connectivity layer for learning deeper
representations of character-word fusion features. In order
to reduce the complexity of the model and prevent network
from over-fitting of training samples, Dropout layer is
added on top of the full connection layer. The Dropout layer
Figure 9. Training Accuracy Curve.
Figure 8. Training Loss Curve.
As can be seen from the figures, for the verification set,
it is possible to achieve a stable state of loss with very few
epochs. However, the recognition rate remained consistent
with the change curve of epoch and loss, and remained
stable after the two epochs, with the recognition rate
reaching 64%, which is a certain improvement compared
with the current 60% of the highest emotional recognition
of text, and can be rapidly converged.
emotional output. The experimental results show that the
multi-classification detection accuracy of the model in this
paper is up to 64.09%, and has a fast convergence rate,
which can effectively classify novels according to the six
basic emotional characteristics.
A CKNOWLEDGMENTS
This work was financially supported by Fablab.
V. C ONCLUSION
R EFERENCES
This paper proposes a Bi-directional Long Short-Term
Memory Language Model for novel classification task
based on text emotion recognition. In this paper, text is
presented as word feature presentation and character feature
presentation respectively by using Word Embedding
method to number vocabulary. Content words and
emotional function words are integrated to estimate the final
[1]
[2]
[3]
[4]
179
J. Cahn, “Generating Expression in Synthesized Speech,” Master’s
thesis, MIT, 1989.
Keikichi Hirose, Nobuaki Minematsu, etc, “Analytical and perceptual
study on the role of acoustic features in realizing emotional speech”,
ICSLP2000.
Donna Erickson,1 Arthur Abramson, “Articulatory characteristics of
emotional utterances in spoken English”, ICSLP2000.
Schröder M1, Cowie R2, Douglas-Cowie E2, “Acoustic Correlates of