1. bookAHEAD OF PRINT
Informacje o czasopiśmie
License
Format
Czasopismo
eISSN
2444-8656
Pierwsze wydanie
01 Jan 2016
Częstotliwość wydawania
2 razy w roku
Języki
Angielski
Otwarty dostęp

Query Translation Optimization and Mathematical Modeling for English-Chinese Cross-Language Information Retrieval

Data publikacji: 15 Jul 2022
Tom & Zeszyt: AHEAD OF PRINT
Zakres stron: -
Otrzymano: 19 Apr 2022
Przyjęty: 22 Jun 2022
Informacje o czasopiśmie
License
Format
Czasopismo
eISSN
2444-8656
Pierwsze wydanie
01 Jan 2016
Częstotliwość wydawania
2 razy w roku
Języki
Angielski
Introduction

Neural network algorithms have diverse characteristics, and as an essential learning method, a deep neural network is widely used in many fields, such as machine learning. Its most significant advantage is that it has a high degree of conformity with the human cognitive process. Its widely used neural network algorithm has potent functions. Its essence can be simplified into a multi-layer neural network. With the continuous development of science and technology, deep neural network algorithms have been favored by algorithm researchers in many fields such as natural language and applied in practice. The specific analysis algorithms included in the deep neural network include question answering analysis, dependency, and syntax analysis. This research uses the advantages of a deep neural network. It extends it to the mathematical modeling process of English-Chinese translation, which is used for English Chinese translation organization learning. So, specifically, this neural network contains the main structures of vocabulary, alignment, language, warping, translation model, etc. Deep neural networks are applied in natural language, and their processing is essentially embedded or represented in the form of words. The word embedding refers to the low-dimensional, high-density, and real-valued vectors to reflect the sentence meaning it contains. In this model, the deep neural network algorithm gives full play to its advantages and fully utilizes its benefits in transforming feature extraction methods. The traditional algorithm is mainly based on the manual extraction of features, while the deep neural network evolves it into automatic extraction. Therefore, compared with manual extraction, it breaks the inherent limitations and breaks through the shortcomings of incomplete information in manual feature extraction. In addition, after the eigenvalue extraction is completed, the deep neural network algorithm will participate in the eigenvalue classification or eigenvalue regression in the form of a classifier or regressor and ensure that the features after classification or regression exist in the same continuous spatial range. Therefore, it is beneficial to measure the relationship of the eigenvalues of this range space.

The mathematical model of English-Chinese translation constructed in this research, the neural network algorithm, is introduced in the training stage. Then the feature value extraction is performed to complete the decoding process to reduce the difficulty of prolonging the time caused by the complexity of decoding. This process is to divide the complexity effectively. And effectively reduce the complexity during the training phase. This research introduces a more optimized and robust functional model, that is, a translation model with hierarchical features to complete the English-Chinese translation mathematical model, and incorporates the feature differences between the original target etymology and the target language during the model building so that it can be Optimize the translation model. Secondly, the English-Chinese translation model featuring a hierarchical translation model has more substantial generalization power than general models, breaks through the limitations of available mathematical models, and can be widely used in various contexts, including classic phrases Or in the translation of words, and also in parenthetical transcription.

The deep recursive hierarchical machine translation model
System Framework

The English-Chinese translation mathematical model constructed in this subject has been built with componentized thinking. The advantage of this thinking is that it can measure the role of each component during the completion of the translation function from a better perspective[3]. When the model is trained, the standard training method can be used, and the typical training method can measure each component. The function that the part performs during translation (F, E) represents a sentence in the training corpus. (f, e) represents the phrase/rule extracted from the training corpus. Fi represents the phrase in the i position. fi represents the word at the place i.

Figure 1 shows a schematic diagram of the overall framework of deep neural network training. As can be seen from the figure, the data domain processing is divided into two processes single-statement and the double-statement preprocessing process. When data preprocessing is performed, several phrases or rules are selected. The trained neural network model is used to generate the corresponding word vector in this subject. The word vector is sent to the HRNN system as an input parameter to complete the learning and training of the process. It should be noted here that when performing the training process of the neural network, it must be ensured that the number of training layers has the same height as the number of derivations generated by the sentence.

Figure 1

Overall framework of deep neural network training

Preprocessing is to generate word vectors after completing the extraction of phrases and rules (Fi, Ei), input the RTNN structure, obtain the starting word vector, and then introduce the vector into the hierarchical model after characterizing it. The model is recursive to get the final semantic vector and represent it zf, ze. The last step of the process is to describe the midpole semantic vector in the form of an inner product and complete, 〈zf, ze〉 the determination of the similarity between the phrase or rule.

Semantic Vectors
Phrase-Based Autoencoders

We use a total of 3 deeply recursive encoders. Recursive encoders for different target language structures are generated based on the word alignment results in the phrases [5]. The semantic vector recovered from the encoding part is denoted as f1 f_1^{'} . We add y2, y4, z2 and z4 nodes to the derivation tree. This process will also not affect the result of the push, and the different forms of the vector are also represented by the form of grayscale, and in different situations, whether the grayscale layer is added is also elegant. If a supervised component is introduced into the model, a grayscale layer must be added; otherwise, no grayscale layer needs to be added.

Training algorithm

The English-Chinese translation mathematical model built in this research introduces three encoders, among which the autoencoder has two characteristics, the first is a phrase-based word encoder, and the second is a rule-based word encoder, each part contains a lot of layers, so it is necessary to first train the word-phrase and rule-based word encoders separately, and then combine them to complete the joint training. For example, a phrase-based autoencoder is trained as follows

Unsupervised Self-Encoding

The hierarchical unsupervised pre-training stage needs to obtain word vectors. Word vectors exist in a matrix Ln ×|V| consisting of word vectors [6]. The matrix L has the following form: L=[f1,,f|V|] L = \left[ {{f_1}, \cdots ,{f_{\left| V \right|}}} \right]

n is the dimension of the word vector. |V| represents the dimension of the dictionary. L is the word vector matrix that has been obtained by RTNN. The subscript of each column vector in L is the number of words in dictionary V. The word embedding of the i word is extracted by the following formula: fi=Lein {f_i} = L{e_i} \in {\;^n}

ei is a unit column vector. It is 0 at all positions except the i position, which is 1.

The encoding layer and decoding layer use different matrices for linear transformation. The activation function of the neural network adopts the “sigmoid” function. The formula used by the function of the encoding layer is as follows: y1=sigmoid(Wf1[f2;f3]) {y_1} = sigmoid\left( {W_f^1\left[ {{f_2};{f_3}} \right]} \right)

The formula used by the decoding layer is as follows: [f2;f3]=sigmoid(Wf2y1) \left[ {f_2^\prime;f_3^\prime} \right] = sigmoid\left( {W_f^2{y_1}} \right)

f2 and f3 are word vectors. y1 is the generated partial phrase vector. [f2; f3] ∈ 2n×1 means concatenate n dimensional vectors f2 and f3 to form an 2n dimensional vector. Wf1n×2n W_f^1 \in {\;^{n \times 2n}} and Wf22n×n W_f^2 \in {\;^{2n \times n}} are the source language side projection matrices.

The English-Chinese translation mathematical model built in this research has the reconstruction error of self-encoding, so the norm of the vector is generally used to model the error. The following expressions (5) and (6) represent the functional expressions of word reconstruction and partial phrase reconstruction errors, respectively. The biggest difference between word reconstruction errors and phrase reconstruction errors is the length, because the difference in length will cause the reconstruction to affect the difference. Therefore, when using expression (6) to model short and reconstructed errors, it is necessary to obtain the weight of each reconstructed part based on the length. Erec([f2;f3]|Wf1,Wf2)=12[f2;f3][f2;f3]22 {E_{rec}}\left( {\left[ {{f_2};{f_3}} \right]|W_f^1,W_f^2} \right) = {1 \over 2}\left\| {\left[ {f_2^\prime;f_3^\prime} \right] - \left[ {{f_2};{f_3}} \right]} \right\|_2^2 Erec([f1;f2]|W1,W2)=|f1||f1|+|y2|[f1f1]22+|y2||f1|+|y2|y2y222 {E_{rec}}\left( {\left[ {{f_1};{f_2}} \right]|{W_1},{W_2}} \right) = {{\left| {{f_1}} \right|} \over {\left| {{f_1}} \right| + \left| {{y_2}} \right|}}\left\| {\left[ {f_1^\prime - {f_1}} \right]} \right\|_2^2 + {{\left| {{y_2}} \right|} \over {\left| {{f_1}} \right| + \left| {{y_2}} \right|}}\left\| {y_2^\prime - {y_2}} \right\|_2^2

In formula (5), a coefficient is added in front of the norm for the convenience of derivation. In formula (6) | f1 |,| y2 | represents the length of the word f1 and part of the phrase y2, respectively. According to the reconstruction error Erec(pf|Wf1,Wf2) {E_{rec}}\left( {{p_f}|W_f^1,W_f^2} \right) of the source phrase pf, the reconstruction error Erec(Cf|Wf1,Wf2) {E_{rec}}\left( {{C_f}|W_f^1,W_f^2} \right) of all the training source phrases Cf can be obtained: Erec(pf|Wf1,Wf2)=nT(pf)Erec([n.c1,n.c2]|Wf1,Wf2) {E_{rec}}\left( {{p_f}|W_f^1,W_f^2} \right) = \sum\limits_{n \in T\left( {{p_f}} \right)} {{E_{rec}}\left( {\left[ {n.{c_1},n.{c_2}} \right]|W_f^1,W_f^2} \right)} Erec(Cf|Wf1,Wf2)=1|C|pfCErec(pf|Wf1,Wf2)+λWf12Wf122+λWf22Wf222 {E_{rec}}\left( {{C_f}|W_f^1,W_f^2} \right) = {1 \over {\left| C \right|}}\sum\limits_{{p_f} \in C} {{E_{rec}}\left( {{p_f}|W_f^1,W_f^2} \right) + {{\lambda W_f^1} \over 2}\left\| {W_f^1} \right\|_2^2 + {{\lambda W_f^2} \over 2}\left\| {W_f^2} \right\|_2^2}

pf represents the source phrase. T(pf) represents the set of intermediate nodes of the spanning tree corresponding to the source phrase. n.c1 and n.c2 denote the semantic vectors of the left and right children of the intermediate node. λWf1 \lambda W_f^1 and λWf2 \lambda W_f^2 weights are set by manual tuning.

The reconstruction error of the source phrase is denoted as Erec(Cf|Wf1,Wf2) {E_{rec}}\left( {{C_f}|W_f^1,W_f^2} \right) . The reconstruction error of the target phrase is denoted as Erec(Ce|We1,We2) {E_{rec}}\left( {{C_e}|W_e^1,W_e^2} \right) . The refactoring error of the source rule is denoted as Erec(Rf|Wf1,Wf2) {E_{rec}}\left( {{R_f}|W_f^{1'},W_f^{2'}} \right) . The refactoring error of the target rule is denoted as Erec(Re|We1,We2) {E_{rec}}\left( {{R_e}|W_e^{1'},W_e^{2'}} \right) . We define the total refactoring error Jrecrec) as follows: Jrec(Θrec)=Erec(Cf|Wf1,Wf2)+Erec(Ce|We1,We2)+Erec(Rf|Wf1,Wf2)+Erec(Re|We1,We2) {J_{rec}}\left( {{\Theta _{rec}}} \right) = {E_{rec}}\left( {{C_f}|W_f^1,W_f^2} \right) + {E_{rec}}\left( {{C_e}|W_e^1,W_e^2} \right) + {E_{rec}}\left( {{R_f}|W_f^{1'},W_f^{2'}} \right) + {E_{rec}}\left( {{R_e}|W_e^1,W_e^2} \right)

We employ different encoding and decoding projection matrices to distinguish between phrases and rules and source and target sides. Θrec contains the projection matrix Wf1 W_f^1 , Wf2 W_f^2 , We1 W_e^1 , We2 W_e^2 , Wf1 W_f^{1'} , Wf2 W_f^{2'} , We1 W_e^{1'} , We2 W_e^{2'} of the source/target language phrases/rules.

Supervised Autoencoder

We need to consider the correction of the phrase/rule semantic vector by the autoencoder between the source and target languages. The transformation matrices used for source/target language side phrases are WfL W_f^L and WeL W_e^L , respectively [7]. They are hierarchical structures used to obtain source/target language phrases. The transformation matrices used by the source/target language rules are WFL W_F^L and WEL W_E^L , respectively. Here is also an example of an autoencoder for bilingual phrases. The transformation matrix transformation function formula is expressed as follows: y4=sigmoid(WfLy3) {y_4} = sigmoid\left( {W_f^L{y_3}} \right) z4=sigmoid(WeLz3) {z_4} = sigmoid\left( {W_e^L{z_3}} \right)

The formula for the bilingual phrase/rule (f, e, a) similarity with alignment is as follows: hbiphr(f,e)=sim(f,e)=y4,z4 {h_{bi - phr}}\left( {f,e} \right) = sim\left( {f,e} \right) = \left\langle {{y_4},{z_4}} \right\rangle

The semantic vectors of the source language side and the target language side corresponding to the bilingual phrase (f, e) are y4 and z4 respectively [8]. The expected BLEU is based on the N-best list obtained by the minimum error rate training idea used in this paper. We use it as the training objective function. Its formula is as follows: Jexp(Θexp)=EGen(Fi)P(E|Fi)sBLEU(Ei,E)=EGen(Fi)exp(λTh(Fi,E))sBLEU(Ei,E)EGen(Fi)exp(λTh(Fi,E)) {J_{\exp }}\left( {{\Theta _{\exp }}} \right) = - \sum\limits_{E \in Gen\left( {{F_i}} \right)} {P\left( {E|{F_i}} \right)sBLEU\left( {{E_i},E} \right)} = - {{\sum\limits_{E \in Gen\left( {{F_i}} \right)} {\exp \left( {{\lambda ^T}h\left( {{F_i},E} \right)} \right)sBLEU\left( {{E_i},E} \right)} } \over {\sum\limits_{E \in Gen\left( {{F_i}} \right)} {\exp \left( {{\lambda ^T}h\left( {{F_i},E} \right)} \right)} }}

sBLEU (Ei, E) is the result of all possible candidate translations at the sentence level between the generation candidate translation sentence E and the reference translation sentence Ei. Where soft max is used to calculate the translation probability P(E | Fi). The formula for the total error Jtotal is as follows: Jtotal=αJrec(Θrec)+βJrtnn(Θrtnn)+(1αβ)Jexp(Θexp) {J_{total}} = \alpha {J_{rec}}\left( {{\Theta _{rec}}} \right) + \beta {J_{rtnn}}\left( {{\Theta _{rtnn}}} \right) + \left( {1 - \alpha - \beta } \right){J_{\rm exp}} \left( {{\Theta _{\rm exp}}} \right)

α and β (0 ≤ α, β ≤ 1) are weights. It characterizes the importance of each submodel. The three errors contain the training parameters of the deep network. We can obtain the compressed word vector through the objective function Jrtnnrtnn) in RTNN. Among them, sentences fi and ei are represented by an activation (oneshot) vector for each word. Θrtnn contains three projection matrices. We can generate semantic vectors of source/target language phrases/rules through Jrecrec). Θrec contains 8 projection matrices [9]. The source language side phrase projection matrices are denoted as Wf1 W_f^1 and Wf2 W_f^2 , the target language side phrase projection matrices are denoted as We1 W_e^1 and We2 W_e^2 , the source language side rule projection matrices are denoted as Wf1 W_f^{1'} and Wf2 W_f^{2'} , and the target language side rule projection matrices are denoted as We1 W_e^{1'} and We2 W_e^{2'} . We can rectify the monolingual phrase/rule semantic vector via Jexpexp). We use bilingual phrase/rule semantic vector representation. Θexp contains 4 projection matrices. The source language side phrase projection matrix is denoted as WfL W_f^L , the source language side rule projection matrix is denoted as WFL W_F^L , the target language side phrase projection matrix is denoted as WeL W_e^L , and the target language side rule projection matrix is denoted as WEL W_E^L .

Parameter estimation

The parameter estimation stage consists of three training parts. The first training stage is word vector training in an unsupervised context, the second stage is monolingual phrase training and rule training in an unsupervised context, and the third stage is supervised context Bilingual Phrase Training and Rule Training under[10]. The estimation of the parameter Θrtnn in the objective function Jrtnnrtnn) in the supervised word vector training phase adopts the traditional method. The features in Θrtnn are global parameters.

The Jrecrec) estimation of the parameter Θrec in the objective function Jrecrec) contains four reconstruction error parts. When in a supervised context, the word vector training process parameter estimation is traditionally done, which mainly acts on the parameters and, notably, is a global function. In addition, the parameter estimates cover four reconstruction errors, so each reconstruction error must be differentiated.

Equation (15) represents the partial derivative of the phrase reconstruction error on the original language side the parameter Wf1 W_f^1 is as follows: Erec(Cf|Wf1,Wf2)Wf1=1|C|pfCErec(pf|Wf1,Wf2)Wf1+λWf1Wf1 {{\partial {E_{rec}}\left( {{C_f}|W_f^1,W_f^2} \right)} \over {\partial W_f^1}} = {1 \over {\left| C \right|}}\sum\limits_{{p_f} \in C} {{{\partial {E_{rec}}\left( {{p_f}|W_f^1,W_f^2} \right)} \over {\partial W_f^1}} + {\lambda _{W_f^1}}W_f^1}

The estimation of the parameter Θexp in the objective function Jexpexp) is expressed as follows. We derive the formula for WfL W_f^L as follows: hbiphr(f,e)WfL=(y4)TWfLz4+(y4)Tz4WfL=y3(z4οsigmoid(y3))T+z3(y4οsigmoid(z3))T {{\partial {h_{bi - phr}}\left( {f,e} \right)} \over {\partial W_f^L}} = {{{{\left( {\partial {y_4}} \right)}^T}} \over {\partial W_f^L}}{z_4} + {\left( {{y_4}} \right)^T}{{\partial {z_4}} \over {\partial W_f^L}} = {y_3}{\left( {{z_4} \circ sigmoi{d^\prime}\left( {y_3^\prime} \right)} \right)^T} + {z_3}{\left( {{y_4} \circ sigmoi{d^\prime}\left( {z_3^\prime} \right)} \right)^T}

“○”is the hamard product of the element-wise multiplication of vectors. sigmoid(y3) sigmoi{d^\prime}\left( {y_3^\prime} \right) is the derivative of each element in the vector y3=WfLy3 y_3^\prime = W_f^L{y_3} . sigmoid(z3) sigmoi{d^\prime}\left( {z_3^\prime} \right) is the derivative of each vector component z3=WfLz3 z_3^\prime = W_f^L{z_3} .

Decoding integration and training of feature weights

Decoding integration is to integrate semantic features into the decoder. This paper sets the cooccurrence number c of each phrase pair to 1. The number of co-occurrences c of each rule pair is a fraction. The formula is as follows: zf=(e,f)(f,e,a)c×zf(e,f)(f,e,a)c {z_f} = {{\sum\limits_{\left( {e,f} \right) \in \left( {f,e,a} \right)} {c \times {z_f}} } \over {\sum\limits_{\left( {e,f} \right) \in \left( {f,e,a} \right)} c }} (f, e, a) represents a phrase/rule pair with aligned results. Here the inner product is used for calculation. The formula is as follows: hbiphr(f,e)=zf,ze {h_{bi - phr}}\left( {f,e} \right) = \left\langle {{z_f},{z_e}} \right\rangle

The two monolingual semantic features of the source language and the target language are denoted as hf−phr(f′, f) and he−phr(e′, e) [11]. This paper uses the standard entropy to define the source language side phrase/rule semantic sensitivity feature hf−sen(f, e). Therefore, phrases and rules need to be fed into the deep neural network to complete the training process. And when language and practices are used, they should be processed by information filtering to filter out regulations or phrases that contain their information: hfsen(f,e)=i=1|Zf|zfi×logzfi {h_{f - sen}}\left( {f,e} \right) = - \sum\limits_{i = 1}^{\left| {{Z_f}} \right|} {{z_{{f_i}}} \times \log {z_{{f_i}}}}

The evaluation of the target language side short sentence sensitivity feature and the target language rule semantic sensitivity feature has the same form as Equation (19). In this subject hf−sen(f, e), the object of the decoding process is two sentences, and it is trained. Therefore, phrases and rules need to be fed into the deep neural network to complete the training process. And when language and rules are used, they should be processed by information filtering to filter out phrases or rules that contain their information.

Influence of Semantic Features

Considering the influence of semantic features, this research adopts two similar baseline systems. First, there are baseline1 baseline systems without semantic features and deadline2 baseline systems without hierarchical model construction, mainly to measure the influence of translation performance by a semantic vector and evaluate the utility of the hierarchical neural network in this model. Bassline 2, which is not built with a hierarchical model, is introduced. The implementation process is similar to HRNN, except that the impact of sentence alignment is not considered too much during training, and it is applied to the phrase hierarchy and rule hierarchy of the original language and the target language. Process. The construction direction of the semantic vector on the original sentence side and the target language side used in this paper is from left to right. It is guaranteed that the baseline system baseline2 built without using the hierarchical model must contain the similarity of bilingual semantics.

This research pioneered the introduction of bilingual features, namely bass and bass, representing bilingual semantic similarity and sensitivity features. In addition, the monolingual features mssm and men are also considered in the model, which represent monosemantic similarity and sensitivity features, respectively.

Effects of different semantic features

Method NIST05 NIST06 NIST08 ALL
Baseline1 35.98 34.88 32.17 37.22
Baseline2 35.22 35.22 32.17 37.22
HRNN (bssm) 35.82 37.83 32.8 37.32
HRNN (bssm+bssn) 36.22 37.29 32.33 37.92
HRNN (bssm+bssn+mssm+mssn) 36.33 37.12 32.33 37.93

We design experiments where the dimension of the semantic vector takes different values n=50, 100, and 200. n=50 is the result in HRNN(bssm+bssn+mssm+mssn) in Table 1. Regardless of the importance of n, the HRNN model with bilingual and monolingual semantic features outperforms the baseline systems baseline1 and baseline2. At the same time, it achieves the best performance at n=100.

Effect of semantic vector dimension on translation performance

Method n NIST05 NIST06 NIST08 ALL
HRNN 100 37.55 34.64 31.66 37.22
200 37.25 34.52 31.64 37.08
The impact of rule binarization

When analyzing the influence of binary bifurcation of rules, the number of non-terminal symbols is limited to two to four. In processing, the number of differentiated non-terminal characters is normalized, and finally, two non-terminal marks are formed. Terminal symbol, and do the same projection matrix processing. Since all the rules in this model can be binarized, a small number of 3.24% binarized rules are filtered out, and the removed function is used to perform this function. And use binarization to complete the binary processing of the rules. The results are shown in Table 4. The effect of binary on the English-Chinese translation performance can be seen from the figure.

Method NIST05 NIST06 NIST08 ALL
Remove 37.32 37.12 32.08 35.82
Binarization 37.33 37.12 32.02 35.83
Conclusion and Outlook

In this research, a mathematical model of English-Chinese translation is built, and a deep learning algorithm is introduced to perform a hierarchical recursive analysis of the model to optimize the performance. The built model is more in line with the actual translation understanding process. The model incorporates global information Consideration of word vectors and bilingual alignment information is also included during neural network training. In addition, three training modules are introduced in the eigenvalue training process. Different training uses different objective functions better to reduce the impact of the interference balance module, and hierarchical pre-training is used in unsupervised situations to minimize the effect. Training time complexity improves training speed. And use bilingual features and monolingual features to filter the training data, in addition, to test the validity of the data, through example demonstration research found that the English-Chinese translation model built in this topic is superior to the classic baseline system, mainly in BLEU In terms of scores, it is about 1.84 BLEU scores higher than the baseline system.

Figure 1

Overall framework of deep neural network training
Overall framework of deep neural network training

j.amns.2022.2.0166.tab.003

Method NIST05 NIST06 NIST08 ALL
Remove 37.32 37.12 32.08 35.82
Binarization 37.33 37.12 32.02 35.83

Effects of different semantic features

Method NIST05 NIST06 NIST08 ALL
Baseline1 35.98 34.88 32.17 37.22
Baseline2 35.22 35.22 32.17 37.22
HRNN (bssm) 35.82 37.83 32.8 37.32
HRNN (bssm+bssn) 36.22 37.29 32.33 37.92
HRNN (bssm+bssn+mssm+mssn) 36.33 37.12 32.33 37.93

Effect of semantic vector dimension on translation performance

Method n NIST05 NIST06 NIST08 ALL
HRNN 100 37.55 34.64 31.66 37.22
200 37.25 34.52 31.64 37.08

Xia, Y. Research on statistical machine translation model based on deep neural network. Computing.,2020; 102(3): 643–661 XiaY. Research on statistical machine translation model based on deep neural network Computing 2020 102 3 643 661 10.1007/s00607-019-00752-1 Search in Google Scholar

Rani, S., & Kumar, P. Deep learning based sentiment analysis using convolution neural network. Arabian Journal for Science and Engineering.,2019; 44(4): 3305–3314 RaniS. KumarP. Deep learning based sentiment analysis using convolution neural network Arabian Journal for Science and Engineering 2019 44 4 3305 3314 10.1007/s13369-018-3500-z Search in Google Scholar

Phan-Vu, H. H., Tran, V. T., Dang, H. V., & Do, P. T. Neural machine translation between Vietnamese and English: an empirical study. Journal of Computer Science and Cybernetics.,2019; 35(2): 147–166 Phan-VuH. H. TranV. T. DangH. V. DoP. T. Neural machine translation between Vietnamese and English: an empirical study Journal of Computer Science and Cybernetics 2019 35 2 147 166 10.15625/1813-9663/35/2/13233 Search in Google Scholar

Wang, X., Chen, C., & Xing, Z. Domain-specific machine translation with recurrent neural network for software localization. Empirical Software Engineering., 2019;24(6): 3514–3545 WangX. ChenC. XingZ. Domain-specific machine translation with recurrent neural network for software localization Empirical Software Engineering 2019 24 6 3514 3545 10.1007/s10664-019-09702-z Search in Google Scholar

Weng, T. RETRACTED ARTICLE: Non-point source pollution in river basin based on Bayesian network and intelligent translation system of English books. Arabian Journal of Geosciences.,2021; 14(16): 1–11 WengT. RETRACTED ARTICLE: Non-point source pollution in river basin based on Bayesian network and intelligent translation system of English books Arabian Journal of Geosciences 2021 14 16 1 11 10.1007/s12517-021-07928-0 Search in Google Scholar

Eria, K., & Jayabalan, M. Neural machine translation: A review of the approaches. Journal of Computational and Theoretical Nanoscience.,2019; 16(8): 3596–3602 EriaK. JayabalanM. Neural machine translation: A review of the approaches Journal of Computational and Theoretical Nanoscience 2019 16 8 3596 3602 10.1166/jctn.2019.8331 Search in Google Scholar

Bi, S. Intelligent system for English translation using automated knowledge base. Journal of Intelligent & Fuzzy Systems.,2020; 39(4): 5057–5066 BiS. Intelligent system for English translation using automated knowledge base Journal of Intelligent & Fuzzy Systems 2020 39 4 5057 5066 10.3233/JIFS-179991 Search in Google Scholar

Khan, F. A., & Abubakar, A. Machine translation in natural language processing by implementing artificial neural network modelling techniques: An analysis. International Journal on Perceptive and Cognitive Computing.,2020; 6(1): 9–18 KhanF. A. AbubakarA. Machine translation in natural language processing by implementing artificial neural network modelling techniques: An analysis International Journal on Perceptive and Cognitive Computing 2020 6 1 9 18 Search in Google Scholar

Mi, C., Xie, L., & Zhang, Y. Improving adversarial neural machine translation for morphologically rich language. IEEE Transactions on Emerging Topics in Computational Intelligence.,2020; 4(4): 417–426 MiC. XieL. ZhangY. Improving adversarial neural machine translation for morphologically rich language IEEE Transactions on Emerging Topics in Computational Intelligence 2020 4 4 417 426 10.1109/TETCI.2019.2960546 Search in Google Scholar

Gençoğlu, M. T., & Agarwal, P. Use of quantum differential equations in sonic processes. Applied Mathematics and Nonlinear Sciences.,2021; 6(1): 21–28 GençoğluM. T. AgarwalP. Use of quantum differential equations in sonic processes Applied Mathematics and Nonlinear Sciences 2021 6 1 21 28 10.2478/amns.2020.2.00003 Search in Google Scholar

Eskitaşçıoğlu, E. İ., Aktaş, M. B., & Baskonus, H. M. New complex and hyperbolic forms for Ablowitz–Kaup–Newell–Segur wave equation with fourth order. Applied Mathematics and Nonlinear Sciences.,2019; 4(1): 93–100 EskitaşçıoğluE. İ. AktaşM. B. BaskonusH. M. New complex and hyperbolic forms for Ablowitz–Kaup–Newell–Segur wave equation with fourth order Applied Mathematics and Nonlinear Sciences 2019 4 1 93 100 10.2478/AMNS.2019.1.00010 Search in Google Scholar

Polecane artykuły z Trend MD

Zaplanuj zdalną konferencję ze Sciendo