Access to this full-text is provided by Wiley.
Content available from Journal of Environmental and Public Health
This content is subject to copyright. Terms and conditions apply.
Research Article
A Novel English Translation Model in Complex Environments
Using Two-Stream Convolutional Neural Networks
Lijuan Zhang
Shanghai Jian Qiao University, Shanghai 201306, China
Correspondence should be addressed to Lijuan Zhang; 19215@gench.edu.cn
Received 15 July 2022; Revised 4 August 2022; Accepted 6 August 2022; Published 5 September 2022
Academic Editor: Zhao kaifa
Copyright ©2022 Lijuan Zhang. is is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Although translation is an essential component of learning English, it does not receive the attention it merits in the modern
English classroom. Teachers and students primarily emphasize listening, reading, and writing while neglecting the development of
translation skills. e English test in China now reflects the fact that there are now very specific requirements for students’
translation skills. As a result, we should emphasize developing students’ translation skills when teaching them English. e
following experimental data can be obtained following the study and experiment on the English translation simulation model
based on the two-stream convolutional neural network: English vocabulary and grammar have passing and excellent rates of 90
and 57 percent, respectively, while reading has passing and excellent rates of 69 and 8 percent, respectively. e ability of students
to translate into English has significantly improved after using the English translation simulation model based on the two-stream
convolutional neural network.
1. Introduction
A Chinese translation has been around for a while, and it has
gone through the following stages of development. e study
of Western translation theory began in the late Ming and
early Qing dynasties, and the translation process reached the
stage of in-depth research in the late Qing and early Ming
dynasties, from the Southern Song Dynasty to the Sui and
Tang Dynasties. After a considerable amount of time,
translators started to focus on the translation of new liter-
ature, and after the rise of New China, translation studies
expanded and became more thorough. Many translators
have given up translating great Tibetan works in favor of
other disciplines like mechanics, philosophy, biomedicine,
astronomy, and mathematics due to the need to comprehend
Western culture. Chinese translation advanced during the
late Qing Dynasty and the early Republic of China, and
numerous works were translated during this time. Nu-
merous patriots put in a lot of effort during a time of national
crisis to learn different Western translation theories and
disseminate cutting-edge Western scientific and cultural
knowledge, which not only helped to cultivate many Chinese
scientific and technological talents but also gave western
translation education a significant place in the history of
modern Chinese translation.
For English majors, translation is a required course.
China must develop outstanding translators who can adapt
to various levels immediately in the age of globalization and
information technology. e social responsibility of trans-
lation education cannot be ignored. e accumulation of
history and Western theories over the past 30 years has
greatly advanced translation education. Due to the unique
nature of Chinese students learning English, translation
remains a significant challenge in China. In order to ac-
complish this goal and raise the standard of translation
instruction in schools, teaching methods must be given
careful consideration in addition to student translation.
Many English instructors and students put all of their
attention into studying for exams and neglect translation.
Giving pupils the chance to gain proficiency in translating
can enhance their knowledge and influence learning in other
subjects. Exams are so stressful that only grammar, com-
position, and other practical aspects are highlighted, and
teachers’ instructional priorities prevent the translation from
Hindawi
Journal of Environmental and Public Health
Volume 2022, Article ID 8426460, 11 pages
https://doi.org/10.1155/2022/8426460
becoming a part of actual learning. Teachers are exam-fo-
cused and motivated to educate students that they need to
memorize vocabulary as well as numerous set phrases,
sentence patterns, and specific formation patterns. ey do
not even believe that translation will boost students’ English
scores. Students find it challenging to translate, and some
believe that translating does not help them to better their
English. e effect of translation on the acquisition of En-
glish is nuanced rather than immediate. Many teachers and
students do not think of or use translation in regular English
classes because they do not see it as a way to enhance English
learning. To address these issues, a two-stream convolutional
neural network-based English translation simulation model
is investigated in this research.
2. Related Work
is article studies some techniques of the English trans-
lation simulation model of the cloud dual-stream con-
volutional neural network, which can be fully applied to the
research in this field. Najjar et al. explored the use of hy-
perbole in the Qur’an and its English translation, and they
studied the morphological transformation of hyperbolic
patterns [1]. Fitriani studied the syntactic and lexical aspects
of the description and classification of grammatical errors
found in English-translated sentences [2]. Yuval and Avishai
described an optimization model based on dynamic pro-
gramming and a public transport simulation that validates
the benefits of such a model [3]. e elastic demand rela-
tionship that Shiaau and Wang embedded in the simulation
model is used to estimate the impact of changing service
time and lock-in during the simulation process [4]. ese
methods provide some references for our research, but they
have not been recognized by the public due to the short time
and small sample size of the relevant research.
Based on the cloud dual-stream convolutional neural
network, we have read the following related materials to
optimize the English translation simulation model. Fang
et al. proposed a novel multimodal biometric recognition
system based on noncontact multispectral finger images,
aiming to address the limitations of single-modal biometric
recognition [5]. Li et al. proposed a real-time online gray-hot
tracking method via Laplacian sparse representation in a
Bayesian filtering framework [6]. Wu et al. proposed an
infrared human action recognition method based on the
spatiotemporal two-stream convolutional neural network
[7]. Li et al. argued that the representational power of
convolutional neural network (CNN) models for hyper-
spectral image (HSI) analysis is actually limited by the
available number of labeled samples, which is usually in-
sufficient to sustain deep networks with many parameters
[8]. Xu et al. introduced a new model, the Regional Con-
volutional 3D Network (R–C3D), which uses a 3D fully
convolutional network to encode the video stream, then
generates candidate temporal regions containing activities,
and finally classifies the selected regions as specific activity
[9]. Ko and Chen believed that neural network manipulation
systems were obtained by fitting data generated from op-
timal manipulation simulations [10]. Goh considered neural
networks to be information processing systems whose ar-
chitecture basically mimics the biological system of the brain
[11].
3. Method for English Translation Simulation
Model of Two-Stream Convolutional
Neural Network
Exchanges between English and Chinese have increased as
China’s national power has grown, drawing attention from
all around the world. Greetings are an essential part of
everyday language use and cannot be disregarded. e
foundation for the implementation of improved English
translation education is the research of the two-stream
convolutional neural network English translation simulation
model.
3.1. Algorithms of Convolutional Neural Networks
Convolution layer calculation: e convolution layer
calculation is to perform convolution multiplication
between the input feature map vector and the con-
volution kernel, and the result of the convolution
multiplication is transformed by the excitation function
to obtain a new feature map. Due to the different
convolution kernels, the feature maps obtained by
different convolution kernels are also different. e
feature map of each output is obtained by the combined
convolution operation of multiple feature maps in the
previous layer. e calculation of the convolutional
layer is as follows:
Ml
j�f
i∈Xj
Ml−1
i∗Kl
ij +bl
j
⎛
⎜
⎝⎞
⎟
⎠.(1)
In convolutional neural networks, the dominant ex-
citation function is the sigmoid function, and each
output function graph has a corresponding discrimi-
nant function. However, the convolution kernel for
each function graph is the same [12].
Calculate the gradient of the convolutional layer: e
convolutional layer lis usually followed by a sublayer
l+ 1, and the local feature of the convolutional layer
feature map corresponds to a point in the sublayer
feature map.
For each convolutional layer j, the error signal δl
jof the
feature map of the convolutional layer is given by:
δl
j�βl+1
jf′ul
j
·up δl+1
j
.(2)
Formula (2) represents the upsampling operation.
e bias basis gradient can be determined by summing
the error signals of layer lusing the above formula, as
follows:
zA
zKl
ij
�
u,v
δl
j
uv pl−1
i
uv.(3)
2Journal of Environmental and Public Health
Since there are many weights in common in convolu-
tional neural networks, the gradient of a given weight
must be summed over all weights associated with that
weight, and finally, its gradients are summed [13].
Calculation of downsampling layer: e principle of
downsampling is that the size of each output function
graph is a reduced version of the input function graph,
as shown in the following formula:
Ml
j�fβl
jdown Ml−1
j
+bl
j
.(4)
In the formula, down(·) represents the downsampling
function; βrepresents the multiplicative bias parame-
ter; brepresents the additive bias parameter.
e size of the downsampling window is y∗y, so the
graph of the output function is reduced by a factor of y.
e purpose of downsampling is to reduce the reso-
lution for scaling invariance. Each output graph has its
own multiplication and addition difference parameters.
Calculation of the downscaling gradient: To calculate
the error signal for the sensitivity map of the down-
scaled layer, one must first locate the layer that cor-
responds to the pixel in the sensitivity map of the
current layer. en, one must recursively calculate the
current error signal using the error signal of the sub-
sequent layer, as shown in the following formula:
δl
j�f′ul
j
·conv2 δl+1
j,rot180 Kl+1
j
,“full”
.(5)
Among them, full represents the full convolution
function [14].
“Full” is a full convolution function that processes and
analyzes convolution gates, replacing missing pixels with 0 s,
so βand bcan be determined. is can be illustrated by the
following formula:
zA
zbj
�
u,v
δl
j
uv,zA
zβj
�
u,v
δl
j·down Ml+1
j
uv.(6)
Neurons are connected through axons, transmit signals
through bulges, and adjust their states according to the “all
or nothing” principle, that is, there are only two states. When
the state of the neuron exceeds a certain threshold, it will be
in the excited state, otherwise it will be in the inhibitory state.
Artificial neuron simulates biological neuron generation
as the basic unit of convolutional network, which accepts
input signal and generates certain output according to the set
function. e basic structure of neurons is shown in Figure 1.
e neuron will receive binputs x�(x
1
,x
2
,. . .,x
n
), and
the input ycan be expressed as the following formula:
y�fωx+ω0
,(7)
brepresents the threshold, f(.) is the activation function, and
yis the output component. Here, net
j
represents the acti-
vation of the neuron unit, x
i
represents the input of the unit,
and a
ij
represents the weight corresponding to each input,
then the above relationship can be expressed by the fol-
lowing formula:
netj�
n
i�1
aijxi+b. (8)
Among them, ω�(ω
1
,ω
2
,. . .,ω
n
) is the n-dimensional
weight, and fis the activation function. In order to enhance
the function of the network, neurons need to introduce a
variety of transfer functions. e commonly used transfer
functions are:
f(s) �
r, s >r,
s, s ≤r,
−r, s <−r.
⎧
⎪
⎪
⎨
⎪
⎪
⎩(9)
In the neural network, rand −rare the thresholds. If sis
greater than r, the neuron outputs r. If s<−r, the neuron
outputs −r, and the other outputs are s.
Sigmoid function, Sigmoid is an S-shaped nonlinear
activation function, and its expression is as follows:
y�f(s)
�1
1+c−s.(10)
Sigmoid neurons and perceptron neurons are basically
the same, which can make the output results slightly change
when the weights change slightly. e input of the function
can be classified into the range of 0 to 1, and the middle area
is gained, and the two sides are suppressed. When sis a large
negative number, the output is approximately 0, and when s
is a large positive number, the output is approximately 1.
e hyperbolic tangent function, hyperbolic tangent
function can be regarded as a translated and enlarged Sig-
moid function, and its expression is as follows:
f(s) � cs−c−s
cs+c−s
�2∗Sigmod(2s) − 1.
(11)
Unlike the sigmoid function, the hyperbolic tangent
function has a mean of 0. e hyperbolic tangent function
has better performance in image processing applications.
In recent years, Convolutional Neural Networks (CNN)
have become the focus of machine learning research [15]. In
1986, the BP algorithm was introduced, which laid the
foundation for CNN. In 2012, CNN won the ILSVRC
competition. Since then, CNN has been applied in many
fields.
Currently, there are two main approaches to action
recognition using convolutional neural networks. One is a
3D Convolutional Neural Network, which uses ordered
video images as the input to the network. e other is a two-
stream convolutional neural network using post fusion,
which is based on the fusion of two separate recognition
streams, the spatial stream and the temporal stream [16].
Both streams are based on the combination of two-stream
structure and 3D convolution to propose a spatiotemporal
convolutional neural network structure, which is a break-
through achievement.
Journal of Environmental and Public Health 3
Two-stream network structure Each stream applies a
convolutional neural network. e basic structure of the
convolutional neural network is shown in Figure 2.
e output features of the two streams are fused and then
entered into the classifier for behavior classification and
recognition. e input of this network structure is a block of
stacked Lframes, and the input of the temporal stream and
the spatial stream are also different. e input to the spatial
stream is a single frame, and this stream effectively identifies
actions from static frames because some actions are closely
associated with specific objects. e input of the temporal
stream is the stacked optical flow displacement field of
several consecutive frames. Optical flow accurately extracts
dynamic features between frames, which makes actions
easier to recognize. e static and dynamic features of the
action are extracted by CNN and then combined in the
fusion layer. e action features extracted in this way have
both specific attributes on a single frame and temporal
associations, so that classification and recognition tasks will
have better results. And then these features are sent to the
classifier to obtain the classification result.
Fusion can be applied at any layer in the two networks,
which requires the feature maps of the two networks to have
the same spatial dimension at time t, which can be achieved
using up-convolution or upsampling. Based on the recog-
nition accuracy obtained from the split1 dataset of UCF101,
both spatial and temporal streams adopt the AlexNet
structure [17]. e effect of different fusion positions on the
accuracy is shown in Figure 3.
As can be seen from Figure 3, the recognition accuracy
obtained by fusion at different layers, it can be clearly seen
that the fusion accuracy after conv5 is higher. And since it is
fused before the fully connected layer, the number of pa-
rameters is twice less than that in the softmax layer. e effects
of different fusion methods on the accuracy are shown in
Table 1.
Table 1 shows the performance of different fusion
methods, which can make the convolution fusion method
work best, and the recognition rate of the fusion layer using
the combination of 3D convolution and 3D pooling is
improved. e reason why it is fused after the conv5 layer is
because of the correspondence between the spatial static
image features and the changing features of the action, and
the fusion of spatial and temporal information, while the
features of the fully connected layer lack such characteristics.
e results show that the fusion needs to learn fewer
parameters before the fully connected layer, and the model
performance is better. A fusion layer is added after the last
convolutional layer, and the input of the fusion layer is the
feature maps of the two streams. e network uses the
AlexNet network, which is currently widely used in behavior
recognition tasks to fuse at the ReLU5 layer (that is, the
activation layer after Conv5), and then input the fully
connected layer to obtain the loss, and then obtain the
gradient according to the loss function, backpropagation, so
as to optimize the parameters in the network [18].
3.2. English Translation. e accuracy and reliability of
translations are frequently in conflict due to the significant
differences between Chinese and English. Consequently, in
order to avoid awkward and rigid translations, it is essential
when translating to understand the differences in idioms and
semantics between the two languages.
For sentence pattern selection, we must first understand
the differences between English and Chinese language
structures. Chinese prose emphasizes the physical and
spiritual gathering, while English prose emphasizes the
internal logical relationship between sentences. e Chinese
structure is loose and the English structure is strict. English
sentences have important connectives under them, while
Chinese sentences use few or no connectives but are still
fluent. For this reason, linguists often use bamboo syntax to
compare English sentences and running water syntax to
compare Chinese sentences. However, Chinese and English
sentences also have many similarities, which also provides
convenience and possibility for translation [19].
e difficulties posed by cultural differences must also be
taken into account when translating; it is not just a matter of
language. e domestication method and the transfer
method are two approaches to dealing with the cultural
aspects of the text, i.e., home to the culture of the source
language and home to the culture of the target language. e
various translation goals, the text’s type, the author’s
intended audience, and the chosen method must all be
considered by the translator.
Compared with literary English, technical English has
linguistic peculiarities, especially in terms of vocabulary and
syntax. e first is the use of a broad range of technical
terms, including purely scientific and technical terms,
general scientific and technical terms, and semi-technical
terms, i.e., terms that have a common meaning in themselves
but have different meanings in different disciplines and
fields. e second is to use abbreviations, and again, there are
more nouns or noun phrases. erefore, in the specific
translation process, for the processing of vocabulary, it is
necessary to consult a lot of materials and try to figure out
the original text, especially the professional connotation of
general vocabulary, so as to achieve lexical equivalence. In
terms of syntax, English for Science and Technology uses
more passive voice, complex structures, and various clauses
to reflect objective and accuracy and avoid subjective nar-
ration. e syntax of technical English can be said to be
“complicated in structure, many layers, indistinguishable,
ΣF (x)
x1
x2
x3
xn
b
Y
…
Figure 1: Basic structure of neurons.
4Journal of Environmental and Public Health
too many episodes, difficult to find context, reversed word
order, and distorted structure.” erefore, in the translation
process, it is not enough to achieve lexical equivalence. It is
also necessary to correctly understand the original text at the
syntactic level, clarify the sentence structure, and on this
basis, pay attention to morphological modifications and
adjust the word order to correctly convey information.
e main function of technical English is to convey in-
formation. In other words, technical English translation also
means conveying information and improving technical
communication. Some scholars define translation as “achieving
the best possible equivalence between text in the target lan-
guage and text in the source language, first semantically and
secondly stylistic equivalence.” erefore, this equivalence
must also be achieved in technical English translation [20].
According to the functional equivalence theory, the
dynamic relationship between the source and target lan-
guages is more significant to the translator than a direct
correspondence. In other words, the relationship between
the person receiving the translation and the information
being translated should, in theory, be the same as the re-
lationship between the person receiving the information
being translated and the information being originally
translated. Content is prioritized over form according to the
functional equivalence theory, which emphasizes that both
the original text’s author and the target language readers can
come to a mutual understanding and appreciation as well as
dynamic equilibrium. In this procedure, the translator is
crucial. Repeated thought and precise application of
translation techniques are required to accurately reproduce
the original author’s ideas. In technical English translation,
the lexical equivalence strategy is crucial. An English sen-
tence should typically have two main components. e two
main parts of a sentence are the subject and the predicate.
Nouns and verbs make up the majority of the vocabulary in a
sentence, and it’s crucial that they match up consistently.
e equivalence of the words that make up sentences, which
can express various logical relationships, is also significant.
3.3. Simulation Model. Eigen Trust is a global trust-based
model proposed in 2003. It is a typical reputation model, and
many later reputation models are improved and perfected on
the basis of it. In the Eigen Trust model, a node has a globally
unique reputation value, which is calculated iteratively
through the reputation evaluation of the entire network. e
calculation of reputation is based on the trust transfer be-
tween nodes, and its calculation method is as follows:
Tak �
b
Cab ∗Cbk.(12)
e meaning of this calculation method is that when
node a wants to know the reputation value of node k, it asks
its neighbor node bto get its evaluation of node k. en
according to its own evaluation of band b’s evaluation of kto
calculate its own evaluation of k.
In the above calculation formula, C
ab
represents the local
reputation value of node ato node b, and the calculation
method is:
Cab �max Sab,0
bmax Sab,0
.(13)
Sab represents the number of times that node ais sat-
isfied with node bduring the transaction between node a and
node b. In this model, it is assumed that there will always be a
fixed set of pretrusted nodes in the network. Due to the
setting of the set, it is always possible to converge in the
iterative calculation, and the nodes in the set all have high
reputation values. But how to select the node set itself is a
difficult problem to solve.
SVM is a binary classification model based on maximizing
intervals in the function space. Suppose someone is dealing
with a binary classification problem, and a linear machine with
reference vectors assumes that all elements in the input and
output spaces match. SVM transforms training samples into a
function space, generates a learned classifier, transforms test
samples in the same way during testing, and uses the learned
classifier to predict classes. Suppose there is a training dataset T:
T�m1, n1
, m2, n2
,. . . , mY, nY
.(14)
Among them,
mi∈M, ni∈N� +1,−1
{ }, i �1,2,..., Y, (15)
m
i
is the ith identification vector, n
i
is the category of mi, if
n
i
�+1, it is one category, otherwise it is another category.
SVM aims to utilize hyperplanes to separate input data
into two classes in feature space. In the general case of
linearly distributed training data, there are infinitely many
ways to split all the sample data. erefore, linear support
vector machines use the interval maximization criterion to
find the best segmentation method, where the selected
segmentation hyperplane is the only solution. e separation
of the hyperplane is shown in Figure 4.
As shown in Figure 4, His the optimal distribution
hypergraph of linearly separable samples, and the pentag-
onal blocks and triangular blocks represent different sample
data. H1 and H2 are the two distribution lines selected for
the samples corresponding to the pentagon block and the
samples corresponding to the triangles just divided, and the
optimal distribution line His located between H1 and H2. At
this time, the interval between H1 and H2 is called a clas-
sification interval. e selection criterion of His not only to
separate the two types of samples as much as possible, but
input
c
c
c
s
s
s
…
…
Input layer Convolutional
pooling layer
Fully connected
layer
Output layer
Figure 2: Basic structure of a convolutional neural network.
Journal of Environmental and Public Health 5
also to make the classification interval between H1 and H2 as
large as possible. e objective function to solve the optimal
separating hyperplane of SVM is:
min
w.r
1
2‖w‖2
s.t.lkwTxk+r
≥1, k �1,2,. . . , K.
(16)
e solution of the optimal separating hyperplane can be
solved by transforming formula (16) from the original
problem to the dual problem using the algorithm of the
Lagrangian dual problem. e objective function of formula
(16) is transformed into:
D(u) � sgn wTu
+
r
�sgn
K
k�1
bkykuku
+
r
⎧
⎨
⎩⎫
⎬
⎭.
(17)
Among them, sgn is the symbolic function, u u
i
is the dot
product of uand u
i
, and w,b
k
, and rare the variables that
share the hyperplane. In the case of linear inseparability,
formula (16) can be transformed into:
min
w.η
1
2‖w‖2+C
K
k�1
ηk
⎛
⎝⎞
⎠s.t.lkwTuk+r
−1+ηk≥0,
k�1,2,. . . , K.
(18)
Among them, C>0 is a constant to reduce the mis-
classification constraint and η
k
is a relaxation term. e
objective function in formula (18) is an expression of the
trade-off between the maximum classification interval and
the minimum number of misclassified samples.
e development of IDS simulation model is essentially a
special software development process. It is a typical network
security detection equipment. e research of the IDS
simulation model helps discuss and summarize the mod-
eling method of the detection mechanism simulation model.
erefore, the idea of Model-Driven Framework (MDA) is
introduced, and the modeling idea of the simulation model,
the adopted modeling method, the structure of simulation
model and the mapping relationship between models are
described.
e detection mechanism is an important link in the
model, which is the key to changing from static protection to
dynamic response. Taking IDS as an example, the estab-
lishment and verification of the detection mechanism
simulation model are studied. e corresponding simulation
model needs to meet the following requirements:
(1) To highlight functional modeling and simulation, the
simulation model should reflect the functional es-
sence of IDS. e simulation of the detection
mechanism mainly focuses on the simulation of the
function, and the focus is on whether the system can
complete the expected detection function according
to the set security policy during the simulation
process.
(2) e simulation model must be reusable, and the
purpose of the detection mechanism modeling is not
to meet a specific application but to adapt to the
different needs of network security simulation.
(3) e simulation model must be complete and effec-
tive, and a correct simulation model is a premise for
the credibility of the simulation results. erefore, in
the process of model establishment, it is necessary to
80
81
82
83
84
85
86
Accuracy
ReLU2 ReLU3somax
Ronghec
(a)
80
81
82
83
84
85
86
87
Accuracy
ReLU5 ReLU6ReLU4
Ronghec
(b)
Figure 3: e effect of different fusion positions on accuracy. (a) Fusion position accuracy 1. (b) Fusion position accuracy 2.
Table 1: Effects of different fusion methods on accuracy.
Integration methods Fusion layer Accuracy (%)
Max Conv5 82.60
Sum Conv5 85.10
2D conv Conv5 85.86
3D conv + 3D pooling Conv5 87.02
6Journal of Environmental and Public Health
ensure the completeness and effectiveness of its
functions by means of verification and other means.
(4) e simulation model should have a flexible con-
figuration interface to reflect the characteristics of
intelligence; in the modeling process, human factors
should be fully considered, and human reasoning
and decision-making should be simulated through
the detection rules to reflect its intelligent detection
process.
IDS is a complex system. Specifically, its complexity is
mainly reflected in the following aspects. (1) e hierarchy of
the architecture; (2) e complexity of information pro-
cessing; (3) e uncertainty of system input and output
information; (4) e intelligence of system behavior.
e development of modeling and simulation systems is
essentially a special kind of software development process.
e simulation model is finally embodied in the form of
program code, and the establishment of the simulation
model is also a special software development process.
erefore, Model-Driven Architecture (MDA), as a new
software development framework, can also be used to guide
the establishment of simulation models. e modeling
process of the IDS simulation model is shown in Figure 5:
In Figure 5, there are five stages including requirements
analysis, functional modeling, object modeling, program
modeling, and model verification. (1) In the demand analysis
stage, the simulation target of IDS and the specific re-
quirements of the simulation model are mainly analyzed. (2)
In the functional modeling stage, the IDS is essentially
abstracted from the functional point of view, and the
functional composition of the IDS is determined, as well as
the restrictions and relationships between the various
functions. (3) In the object modeling stage, based on the
functional model, combined with the selected modeling
platform, the platform-related object model is established
through specific mapping rules. (4) In the program modeling
stage, the object model is used as the input, and combined
with the program modeling specification of the selected
platform, the program model is established by automatic
generation or manual writing. (5) Model verification stage.
Model verification activities run through the above four
stages, mainly to verify the correctness of the requirements
document, functional model, object model, and program
model generated in the modeling process to ensure the
validity of the IDS simulation model.
4. Experiment of English Translation
Simulation Model of Two-Stream
Convolutional Neural Network
4.1. Application Experiment of Convolutional Neural Network.
In terms of resilience, self-learning capacity, and associative
memory function, classical recurrent neural networks
continuously imitate various brain functions. Artificial
neural networks (ANNs) have a variety of network struc-
tures and functions. To ensure structural stability and de-
sign, it is necessary to integrate theoretical knowledge from
various fields and disciplines, such as nonlinear and dy-
namical systems. As a result, it is crucial to research the
structural design approaches used in recurrent neural net-
works. e block diagram of the recurrent convolutional
neural network is shown in Figure 6.
As can be seen from the frame diagram of the recursive
convolutional neural network, on the basis of the traditional
convolutional neural network, after the network passes
through the fully connected layer, it is not directed to the
classification layer of the convolutional network, but directly
to the added recurrent neural network layer. In this way, the
results output by the recurrent neural network can be better
applied in the classification layer.
e structure of the network is one more layer than the
traditional convolutional neural network, and the com-
plexity of the network is also improved. e learning al-
gorithm of the recurrent convolutional neural network is
mainly divided into two parts: one is the parameter ad-
justment in the convolutional neural network, and the other
is the parameter training in the recurrent network.
In order to better verify and evaluate the performance of
recurrent convolutional neural network, it is first applied to
Chinese license plate recognition. is data set collects
license plate figures from all over the country, divides them,
and finally organizes them into a license plate data set. e
network performance is tested on the commonly used
dataset MNIST handwriting dataset. e dataset contains
60,000 training samples and 10,000 test samples. e test
results of the license plate dataset are shown in Figure 7:
e three networks in Figure 7 are Convolutional Neural
Networks, Elman Convolutional Neural Networks, and
Elman-Jordan Convolutional Neural Networks. From the
results in Figure 7, it can be seen that the Elman-Jordan
convolutional neural network has a significantly lower error
rate than the other two types of network models. It can be
calculated from Figure 8 that the error rate of the recursive
convolutional neural network for license plate testing is
41.27% lower than that of the convolutional neural network,
and 11.24% lower than that of the Elman convolutional
neural network. e test results on MNIST handwritten data
are shown in Figure 8.
H1
H
H2
Figure 4: e optimal splitting hyperplane for the linearly sepa-
rable case.
Journal of Environmental and Public Health 7
model
validation
Requirements
document
functional
model object model program model
demand
analysis
functional
modeling
object
modeling
program
modeling
Figure 5: e modeling process of the IDS simulation model.
input
layer
Convol-
utional
layer
C1
Pooling
layer
S2
Convol-
utional
layer
C3
Pooling
layer
S4
fully
connect-
ed layer
recursi-
ve layer
classi-
cation
layer
Figure 6: Structure diagram of recursive convolutional neural network.
Dataset Test Results
Dataset Test Results
0
20000
40000
60000
80000
100000
120000
Number Of Iterations
Elman Elman-Jordanconvolution
Neural Networks
Elman Elman-Jordanconvolution
Neural Networks
0
0.5
1
1.5
2
2.5
Error Rate
Figure 7: Test results of license plate dataset.
Handwritten Data Test
Results
Handwritten Data Test
Results
0
50000
100000
150000
200000
250000
Number Of Iterations
Elman Elman-Jordanconvolution
Neural Networks
Elman Elman-Jordanconvolution
Neural Networks
0
0.2
0.4
0.6
0.8
1
Error Rate
Figure 8: Test results on MNIST handwritten data.
8Journal of Environmental and Public Health
From the data in Figure 8, it can be found that the error
rate of the recurrent convolutional neural network for
handwriting recognition has been reduced to a certain ex-
tent. Among them, it is 61.14% smaller than the convolu-
tional neural network, and 21.06% smaller than the Elman
convolutional neural network.
Overall, it can be said that the recurrent convolutional
neural network has a lower error rate than the conventional
convolutional neural network, which also demonstrates the
recurrent neural network’s superior classification capabil-
ities. e recurrent convolutional neural network effectively
utilizes this benefit of the recurrent network and improves
the network’s recognition capability. However, as the error
rate decreases, the network’s computation capacity grows as
well, leading to more network iterations and an increase in
the number of iterations overall, as well as an increase in the
network’s training time. is is due to the fact that the
recurrent network recursively passes the network’s output to
the hidden layer so that it can take part in the network
calculation. As a direct result, the network’s parameters
grow, which increases the network’s overall computation.
4.2. Experiments on English Translation. e research objects
are 33 full-time English translation students from a normal
college in 2008. Use the CATTI Level 3 English Translation
Comprehensive Ability Test to test the students’ translation
ability substantively. e translation method is English to
Chinese. e structure of the test paper is shown in Table 2.
Vocabulary and grammar, reading, and cloze are three
parts that make up this test paper. A maximum of 55 points
can be awarded for the reading section. is shows that
reading plays an important role in translation ability, it is the
first process of translation work, and it is also the most basic
prerequisite for doing a good job in translation work. e
test was conducted with the help of 33 students and their
teachers. e analysis of students’ papers focuses on lan-
guage skills (vocabulary, pragmatic skills, grammatical
knowledge), the ability to modify sentence structure, reading
skills and theory, and translation skills.
e overall average score for translation ability is 64.76,
with the highest being 78.5 and the lowest being 42. e
average score for translation management was 51.85, with
the highest score being 67 and the lowest being 35. Of the 33
students, 6 passed the exam. e pass rate is 18.18%. is
means that more than 80% of the students’ translation
skills do not meet the requirements of the CATTI syllabus.
e results of each part of the translation are shown in
Table 3.
It can be seen from the table that the excellent rate and
pass rate of the vocabulary and grammar part and the
reading part are higher than those of the cloze part. e
excellent and pass rates of the vocabulary and grammar
sections were 57% and 90%, respectively, followed by the
reading section, with a pass rate of 69%, but a lower excellent
rate of 8%. And the most unsatisfactory cloze part, the pass
rate is 32%, there is no excellent. Part of the reason for the
low results is that the test is subjective, which makes the test
more difficult. e results for each part are as follows:
(1) Vocabulary selection, emphasizing the use of idioms,
synonyms, analysis, and sentence comprehension.
(2) Word replacement requires students to choose words
or phrases that have similar meanings to the
underlined part, mainly to check whether students
understand the sentences and whether they are fa-
miliar with the concepts of “synonyms” and
“synonyms.”
(3) Proofread section covers the use of adjectives, nouns,
common collocations, and an understanding of the
meaning of major topics, while also examining
grammar skills.
e second part: reading comprehension, which is the
main part of translation work, consists of two reading ar-
ticles. e first reflects the life and social status of black
Americans, and the second is an article on popular science.
e first reading score is shown in Figure 9.
e first reading was on sociocultural studies, and the
results reflected in the papers showed that students had a
good grasp of this cultural aspect, with 64% of them getting
a distinction. e second reading score is shown in
Figure 10.
e second reading mainly introduces the organic
matter required for plant growth and its growth and de-
velopment. As can be seen from Figure 10, translators have
poor comprehension of technical articles, with only 10
students passing the test, 7 students getting good grades, and
6 received a failing grade. is suggests that liberal arts
students have low scientific knowledge. Translation involves
knowledge from all walks of life, and good translation re-
quires the translator to have extensive knowledge.
Table 2: e structure of the first volume of the CATTI English level 3 translator.
Number of items Scores Time needed
Section 1 Vocabulary and grammar 50 25 points
2 hoursSection 2 Reading comprehension 50 55 points
Section 3 Cloze 20 20 points
Table 3: Proportion of achievements in each part of translation.
Ingredient/category Vocabulary and grammar (%) Reading ability (%) Cloze ability (%)
Fail rate 10 31 68
Passing rate 90 69 32
Excellent rate 57 8 0
Journal of Environmental and Public Health 9
5. Conclusions
Language proficiency is a crucial component of commu-
nicative language proficiency. It refers to the four funda-
mental abilities of listening, speaking, reading, and writing
as well as the capacity to combine them. ese four fun-
damental abilities combine to form translation, which is in
essence a fundamental ability. Translation originally meant
converting a text from one language to another. Oral and
written translations between English and Chinese can be
classified into two categories. e written translation reflects
the students’ “reading” and “writing” skills, whereas the oral
translation reflects their “listening” and “speaking” skills.
For advancing the current development of English trans-
lation, the research on the two-stream convolutional neural
network-based English translation simulation model is also
very important.
Data Availability
e data used to support the findings of this study are in-
cluded within the article.
Conflicts of Interest
e author declares that there are no conflicts of interest.
References
[1] I. Najjar, K. A. Kadhim, and S. Al-Heeh, “Morphological shift
of hyperbolic patterns in the Quran, with reference to English
translation,” Forum, vol. 19, no. 2, pp. 148–163, 2021.
[2] E. Fitriani, “Errors found on the English translation sentences
of the third-year students of English at IKIP budi utomo.
Journey,” Journeys, vol. 3, no. 1, pp. 36–45, 2020.
[3] H. Yuval and Avishai, “Public transit simulation model for
optimal synchronized transfers,” Transportation Research
Record, vol. 2063, no. 1, pp. 52–59, 2018.
[4] L. Shiaau and P. Wang, “Demand elasticity and benefit
measurement in a waterway simulation model,” Trans-
portation Research Record, vol. 2033, no. 1, pp. 53–61, 2018.
[5] Y. Fang, Q. Wu, and W. Kang, “A novel finger vein verifi-
cation system based on two-stream convolutional network
learning,” Neurocomputing, vol. 290, no. 17, pp. 100–107,
2018.
[6] C. Li, X. Wu, N. Zhao, X. Cao, and J. Tang, “Fusing two-
stream convolutional neural networks for RGB-T object
tracking,” Neurocomputing, vol. 281, no. 15, pp. 78–85, 2018.
[7] X. Wu, S. Sun, and J. Li, “Infrared behavior recognition based
on spatio-temporal two-stream convolutional neural net-
works,” Journal of Applied Optics, vol. 39, no. 5, pp. 743–750,
2018.
[8] X. Li, M. Ding, and A. Pizurica, “Deep feature fusion via two-
stream convolutional neural network for hyperspectral image
classification,” IEEE Transactions on Geoscience and Remote
Sensing, vol. 58, no. 4, pp. 2615–2629, 2020.
[9] H. Xu, A. Das, and K. Saenko, “Two-stream region con-
volutional 3D network for temporal activity detection,” IEEE
Transactions on Pattern Analysis and Machine Intelligence,
vol. 41, no. 10, pp. 2319–2332, 2019.
[10] C. H. Ko and J. K. Chen, “Grasping force based manipulation
for multifingered hand-arm Robot using neural networks,”
Numerical Algebra, Control and Optimization, vol. 4, no. 1,
pp. 59–74, 2014.
[11] A. T. C. Goh, “Seismic liquefaction potential assessed by
neural networks,” Environmental Earth Sciences, vol. 76, no. 9,
pp. 1467–1480, 2017.
[12] Y. Ding, Z. Zhang, and X. Zhao, “Multi-feature fusion: graph
neural network and CNN combining for hyperspectral image
classification,” Neurocomputing, vol. 501, 2022.
[13] M. Zhao, C. H. Chang, W. Xie, Z. Xie, and J. Hu, “Cloud shape
classification system based on multi-channel cnn and im-
proved fdm,” IEEE Access, vol. 8, pp. 44111–44124, 2020.
[14] Y. Tan, X. Qiao, and Y. Ma, “A study on Chinese-English
translation of international publicity of the northern shaanxi
tourism in China from the perspective of reception aes-
thetics,” Open Journal of Modern Linguistics, vol. 11, no. 3,
pp. 415–423, 2021.
[15] W. Cai, “Hierarchical domain adaptation projective dictio-
nary pair learning model for EEG classification in IoMT
systems,” IEEE Transactions on Computational Social Systems,
Early Access.
[16] D. D. Sokoloff, M. S. Nuraliev, A. A. Oskolski, and
M. V Remizowa, “Gynoecium evolution in angiosperms:
monomery, pseudomonomery, and mixomery,” Moscow
University Biological Sciences Bulletin, vol. 72, no. 3, pp. 97–
108, 2017.
[17] H. Zhao and D. Nan, “Dynamic analysis of stochastic
Cohen–Grossberg neural networks with time delays,” Applied
Mathematics and Computation, vol. 183, no. 1, pp. 464–470,
2017.
First Reading
0
5
10
15
20
Number Of People
Pass excellentfailed
Grade Classification
Figure 9: e first reading score.
Second Reading
0
5
10
15
20
Number Of People
Pass excellentfailed
Grade Classification
Figure 10: e second reading score.
10 Journal of Environmental and Public Health
[18] L. Wang, Q. Song, Y. Liu, Z. Zhao, and F. E. Alsaadi, “Finite-
time stability analysis of fractional-order complex-valued
memristor-based neural networks with both leakage and
time-varying delays,” Neurocomputing, vol. 245, no. 5,
pp. 86–101, 2017.
[19] H. He, F. Gang, and C. Jinde, “Robust state estimation for
uncertain neural networks with time-varying delay,” Journal
of Jishou University (Natural Sciences Edition), vol. 19, no. 8,
pp. 1329–1339, 2019.
[20] A. Y. Alanis and Y. Alma, “Electricity prices forecasting using
artificial neural networks,” IEEE Latin America Transactions,
vol. 16, no. 1, pp. 105–111, 2018.
Journal of Environmental and Public Health 11
Available via license: CC BY
Content may be subject to copyright.
