Selection and study of fuzzy semantics in machine translation

Abstract

In daily life, English is increasingly used in various scenarios, and the technology of translation using machines and others is gradually maturing, but there are still cases of inaccurate translation. To enhance translation’s accuracy, this study optimizes the method of fuzzy semantic selection, and then optimizes the method of domain analysis combined with neural networks to improve the accuracy of machine translation in different domains. The accuracy of the optimized neural network tends to be stable when the number of iterations is 15, the accuracy is 0.96, the accuracy of the traditional neural network is 0.91, and RNN is 0.82. Compared with the benchmark system, the bilingual mapping model has increased by 0.67% in the news field and 0.56% in the education field. The precision, recall and $F$ value of machine translation are 93%, 86% and 0.8 respectively. The comprehensive experimental results show that the selection method based on fuzzy semantics, combined with the domain analysis method of network neural, can remarkably enhance the accuracy.

Keywords

Machine translation fuzzy semantics neural network multisemantics domain analysis bilingual mapping

1. Introduction

Ambiguity is one of the characteristics of language that is prevalent in life. Regardless of the language expressions, many words have fuzzy nature and there is semantic uncertainty, which leads to the accuracy of translation work being affected. In the process of machine translation, the selection of fuzzy semantics is particularly important. The translation needs to be combined with the semantic environment in which the phrase or sentence is located to give a translation result that fits the context, otherwise, it will be out of the meaning of the translation and reduce the accuracy of the translation. Using the visual information or prior knowledge related to the text and the text itself can improve the accuracy of translation [1]. Due to the late introduction of fuzzy semantics and the small content of related literature, there are scholars in the existing literature who have investigated fuzzy semantics in machine translation. For example, rare words are combined by hierarchical clustering method and integrated into the codec in the proposed fuzzy semantic representation (FSR). In neural machine translation, this structure can solve the problem of data sparsity, capture the semantics of rare words in the case of fuzzy context, and compensate the semantic information of the source and target [2]. Adjacency can introduce fuzzy information into the syntax layer to correctly express meaning. Julian-Iranzo [3] introduced a unified algorithm to establish a new model, which can manage adjacency and adapt to weak SLD resolution rules. Approximate reasoning in logic programming framework can be realized by combining unified algorithm and proximity relationship. In order to be able to improve the accuracy of translation, the selection of fuzzy semantics needs to be further optimized, and the combination of deep learning such as neural networks can be considered to improve the performance of the method, as well as the combination of domain information to improve the accuracy of machine translation.

2. Related work

More efficient algorithms have emerged in natural language processing, allowing the use of computer science and technology for detailed analysis and synthesis of different languages, and speech translation. Machine translation automates the translation process and reduces human dependence. The common translation methods included in machine translation classification are heavily dependent on parallel corpora [4]. The quality of translation depends on the language and the method used, and Dugonik et al. [5] proposed a standard approach based on the well-known statistical machine translation, which is proposed by a newly proposed method for optimizing the weights of translation system components. The results showed that the method used in the experiments improved the translation quality using better system component weights. With the revival of neural networks, for most language pairs, the translation quality exceeds that obtained using statistical techniques. Mahmoud and Mengash [6] introduced an algorithm for preprocessing a bilingual corpus by proposing a Corpus-Trie (CT) data structure to respond to translation requests using CT. The number of phrases in CT system is proportional to BLEU value. Deep learning techniques have driven the development of neural network machine translation (NMT), and the adequacy of translation can be improved by using new methods to transfer semantic knowledge [7]. The literature shows that the translation output quality of NMT is better in the mutual translation of multiple languages [8]. Some scholars also use rule-based methods to translate Tunisian dialect (TD) texts into modern standard Arabic (MSA) machine translation systems [9]. A new machine translation model in related literature can effectively enhance the precision of neural machine translation by fusing the attentional information of words and characters. This study also showed that bidirectional selective pass recursive unit (GRU) networks can be used to automatically construct word-level information from input character sequences [10].

In recent years, deep neural network models have shown excellent practicality in areas such as machine learning and thermal intelligence [11]. Translation program on the basis of neural network has become a mainstream approach in the field of machine translation, which means statistical machine translation is becoming to a faster and more reliable neural network translation. Comparing phrase-based SMT systems with NMT systems, Benkova et al. [12] evaluated the translation quality of English and Slovak using automatic metrics. The quality of neural machine translation was better regardless of the system used; based on all BLEU_n scores, SMT outperformed NMT with statistical significance. In Sanskrit translation work, the processing of initial data will affect the performance of the model and deep neural network was used to process Sanskrit raw data in a study. Better BLEU scores and word error rates were achieved with this model processing the Sanskrit text [13]. Researchers proposed a new DNN-based keyword recognition system that dynamically changes keywords and uses both trisyllabic and monophonic acoustic models. This model can effectively simplify the calculation difficulty and improve the generalization ability. The author tested on the FFMTIMIT corpus and found that the error rate of this model was more than 35% lower than that of the traditional model [14]. To extract text-related visual features, Kwon et al. [15] designed modulation networks based on visual information from a pre-trained CNN, applying a feature-based multiplicative transformation to obtain a modular trainable network model embedded in an existing multimodal translation model architecture. Through validation, it was confirmed that the model improves on text-based models and previous models. Liu and Chen [16] proposed a self-attention-based machine translation model. The model uses subword tokenization in corpus preprocessing to overcome rare words. In the encoder layer, a dual self-attention stack and fewer; In the decoder, the “decoder” stack is reduced to speed up training and inference while keeping the level of BLEU constant. Li et al. [17] proposed a deep learning model based on long-term and short-term memory (LSTM), which uses convolution neural network to deeply learn spatiotemporal features (DLSF-CNN) to measure the spatiotemporal correlation of time consuming on different routes, so as to accurately predict route travel time. Long short-term memory network has changed the machine learning and neural computing, and such models have improved Google’s speech recognition and greatly improved machine translation on Google [18]. In terms of recurrent neural network (RNN), Google neural machine translation (GNMT) is the most advanced language translation model using this network structure. Deep learning is applied to natural language processing, which can transform natural language description into source code to meet requirements [19]. A special RNN named as long short-term memory network (LSTM) is currently being used to analyze information from social media APPS for code-speaking and translating them phonetically into English [20].

In order to better study the research on the selection method of fuzzy semantics in machine translation, it is proposed in this research that the optimized neural network is used in the selection of fuzzy semantics, and the analysis is carried out in conjunction with the domain where the semantics are located to further improve the accuracy of translation, hoping to lay the foundation for further selection and optimization of fuzzy semantics under machine translation.

3. Optimization of selection method for fuzzy semantics in machine translation

3.1 Optimization method for fuzzy semantic best selection process

There are many problems that have not been solved in the daily translation process when using machines such as computers for translation [21]. For example, there are more meanings of the same textual content in Chinese or English, which can lead to semantic errors or contextual incompatibilities in the translation process, and it will reduce the accuracy and translation quality of the translation. To address these problems, a mathematical mapping method can be considered to label each semantic meaning in the fuzzy semantics, and then select the corresponding semantic labels according to the features of the text. In this study, a semantic model of citation translation in machine is constructed on this basis to optimize the selection of fuzzy semantics. First, the semantic content is characterized and extracted. A conceptual semantic constraint function $\alpha$ in machine translation is set to represent the set of real words matched by the analytic evaluation set $H$ of semantic information, which comes with two Mountain node labels with the range of values $[0,T]$ , as shown in Eq. (1).

$\displaystyle\Delta(\alpha)=\left\{{{\begin{array}[]{l}{h_{k},k=r(\alpha)}\\ {b_{k}=\alpha-k,b_{k}\in[-0.5,0.5)}\\ \end{array}}}\right.$ (1)

In Eq. (1), $\Delta$ denotes $[0,T]\to H\times[-0.5,0.5)$ , and $r$ denotes the semantic information rounding mapping operator, which can be used for automatic recognition of semantic information. Combined with the two tags with Mountain nodes in machine translation, the concept generalization relation processing is used to analyze the contextual language and improve the translation level. Then, a semantic ontology mapping model is established. Let a binary feature combination $(h_{k},b_{k})$ be used to modify the semantic target, then the definite language modification function for translation is shown in Eq. (2).

$\displaystyle\Delta^{-1}:H\times[-0.5,0.5)\to[0,T]$ (2)

In Eq. (2), $h_{k}$ denotes the $k$ th element in $H$ and $b_{k}\in[-0.5,0.5)$ . In the process of machine translation there exists the case of dividing the definite clause in the scope of the subordinate clause into the scope of the main clause, assuming that there exist two binary semantics $(h_{k},b_{k})$ , $(h_{l},b_{l})$ , then an ontology mapping model for $(h_{k},b_{k})$ and $(h_{l},b_{l})$ can be obtained in the process of selecting the scope when $k<l$ , $(h_{k},b_{k})<(h_{l},b_{l})$ , and when $k=l$ , as shown in Eq. (3).

$\displaystyle\left\{{{\begin{array}[]{l}{b_{k}=b_{l}\Rightarrow(h_{k},b_{k})=(% h_{l},b_{l})}\\ {b_{k}<b_{l}\Rightarrow(h_{k},b_{k})<(h_{l},b_{l})}\\ {b_{k}>b_{l}\Rightarrow(h_{k},b_{k})>(h_{l},b_{l})}\\ \end{array}}}\right.$ (3)

On the basis of the previous two steps, the semantic selection method is improved and optimized for design, and the topic word attribute table is established as shown in Fig. 1.

Figure 1.

Fuzzy semantic subject word attribute table.

In Fig. 1, the semantic similarity is calculated by the gray correlation matching method, assuming that the translated semantic information of any direct superordinate word in the total evaluation set can be represented by the binary semantic group $(h_{k},b_{k})$ , $(h_{l},b_{l})$ , then its Euclidean distance to each component is shown in Eq. (4).

$\displaystyle d((h_{k},b_{k}),(h_{l},b_{l}))=\Delta(|{\Delta^{-1}(h_{k},b_{k})% -\Delta^{-1}(h_{l},b_{l})}|)$ (4)

The reliability factor is used to indicate the degree of reliability of the machine translation process using the search engine to translate adverbs based on ontologies, and the average reliability factor is defined as shown in Eq. (5).

$\displaystyle(\overline{h},\overline{b})=\Delta\left({\sum\limits_{j=1}^{n}{% \frac{1}{n}\Delta^{-1}(h_{j},b_{j})}}\right)$ (5)

A binary coordinate set $v_{i}=((w_{1},t_{1}),(w_{2},t_{2}))$ is set to the similarity between semantics. In the Tuscany region, the result set $R(X)$ of the word $X$ can be obtained using the difference of mapping relations, and the weight vector of $R(X)$ is $\omega=(\omega_{1},\omega_{2},\ldots,\omega_{n})^{T}$ , $\omega_{j}\in[0,1]$ , which gives the definition of semantic similarity as shown in Eq. (6).

$\displaystyle(\overline{h},\overline{b})=\Delta\left({\sum\limits_{j=1}^{n}{% \omega_{j}\Delta^{-1}(h_{j},b_{j})}}\right)$ (6)

In Eq. (6), $\sum\limits_{j=1}^{n}{\omega_{j}}=1$ , $\overline{h}\in H$ , $\overline{b}\in[-0.5,0.5]$ . The result of semantic similarity is calculated by Eq. (6), based on which the establishment of semantic mapping relationship can be realized, and then the calculation of the optimal solution is carried out. Using the circular stack control method for optimal matching degree search, the integrated weight similarity can be calculated. Let a set of binary semantic information $\{(h_{1},b_{1}),(h_{2},b_{2}),\ldots,(h_{n},b_{n})\}$ , let two similar words to be compared in the text $X$ , $Y$ , and calculate the extraction vector $\omega=((\omega_{1},b^{\prime}_{1}),(\omega_{2},b^{\prime}_{2}),\ldots,(\omega% _{n},b^{\prime}_{n}))^{T},\omega_{j}\in[0,1]$ of $Y$ , we can get the average operator $\varphi_{2}$ after weighting this binary semantic information, as shown in Eq. (7).

$\displaystyle(\overline{h},\overline{b})=\Delta\left({\sum\limits_{j=1}^{n}{% \Delta^{-1}(\omega_{j},b^{\prime}_{j})\Delta^{-1}(h_{j},b_{j})}\left/\sum% \limits_{j=1}^{n}{\Delta^{-1}(\omega_{j},b^{\prime}_{j})}\right.}\right)=% \Delta\left({\sum\limits_{j=1}^{n}{\alpha_{j}\alpha^{\prime}_{j}\left/\sum% \limits_{j=1}^{n}{\alpha^{\prime}_{j}}\right.}}\right)$ (7)

In Eq. (7), $\sum\limits_{j=1}^{n}{\omega_{j}}=1$ , $\overline{h}\in H$ , $\overline{b}\in[-0.5,0.5]$ . The semantic centroid vector $C(Y)$ of $Y$ is calculated according to the position of the words divided from the phrases in the text, and the translation similarity of $X$ and $Y$ can be obtained, as shown in Eq. (8). The standardized processing of each component vector of $C(Y)$ yields the central vector $C(X)$ of $R(X)$ , as shown in Eq. (9).

$\displaystyle\textit{Sim}(X,Y)=\textit{Cos}(X,Y)\frac{C(X)\cdot C(Y)}{|{C(X)}|% \cdot|{C(Y)}|}$ (8) $\displaystyle C(X)=\frac{1}{n}\sum\limits_{i=1}^{n}{\frac{v_{i}}{|{v_{i}}|}}$ (9)

3.2 Optimization of fuzzy semantic selection method based on neural network

Domain-based translation models utilize neural networks to build models when processing information, and RNN, CNN, and R-CNN are often used for deep learning in existing studies. These deep learning neural networks have good performance and play an active role in different fields. For example, recurrent neural networks can be better adapted in modeling serialized information, training data autonomously and extracting image features, etc. [22]. Regional convolutional neural networks can combine expert prior knowledge into region definitions [23]. However, there are some limitations, such as recurrent neural networks cause inaccurate translation due to limited lexicon size and easy missing information. To address these problems, in this study, a bidirectional recurrent neural network (Bi RNN) is selected, which is a model widely used in machine translation, speech recognition and natural language processing [24]. Bi RNN can extract text features to complete word embedding, which can better understand contextual relationships and its word embedding effect is better. Using Bi RNN for deep learning of methods in machine translation, automatically extracted features can be obtained. On this basis, attention mechanism is used to enhance the effectiveness and precision of translation. Its structure schematic is shown in Fig. 2.

Figure 2.

Bi RNN network detailed structure.

The specific steps of the bidirectional recurrent neural network are represented in Fig. 2. First, for the text $T$ , the words in the text are represented by a unique thermal encoding vector, and then the embedding matrix is multiplied by each vector and the dimension is adjusted to 300. Then the embedding vectors of each word are learned using a bidirectional GRU with forward output and reverse output. The attention mechanism can enhance attention to key parts and highlight important information. The attention mechanism is introduced to enhance word-to-word dependency based on Bi RNN. The text translation model based on Bi RNN and attention mechanism is shown in Fig. 3.

Figure 3.

Text translation method based on Bi RNN and attention mechanism.

In Fig. 3, G stands fo GRU, and Bi-G stands fo Bi-GRU. The text translation is performed by first vectorizing the text entries and inputting the word vectors in the neural network. Then a word dictionary matrix is built, then a mapping relationship lookup table is built, and finally the words are converted into word embeddings by converting the ordinal numbers into word embeddings.

3.3 Optimization of fuzzy semantic domain analysis method based on neural network

For fuzzy semantics, in addition to matching the features of semantics, translation can also be performed by combining the domain in which the semantic information is located with the linguistic environment in which the target vocabulary is located for a more accurate translation of the semantically fuzzy target. Therefore, this experiment proposes to build a domain translation model and optimize it on the basis of semantic similarity, which can assign high scores to domain-related candidates by calculating the domain intertranslation degree of phrase pairs. This optimization method mainly consists of 2 elements, firstly, obtaining the semantic $k$ nearest neighbors of words, which can be obtained by implementing a bilingual mapping of word vectors in the target domain, and then obtaining the domain inter-translation probability of phrase pairs by calculating their inter-translation degree in the target domain. The specific framework of the model optimization method is shown in Fig. 4.

Figure 4.

System framework diagram of domain translation model optimization method based on semantic similarity.

In Fig. 4, the source language and target language vector of the target domain are trained firstly, and the word vectors are mapped to the two-dimensional space after dimensionality reduction. Then, the source language is mapped to the target language vector space $P$ point, and the semantic $k$ nearest neighbors are obtained with the point as the circle center. For the source language words, word vector representations of their semantic $k$ nearest neighbors and target candidate translations need to be obtained. In order to obtain the cross-language semantic representation of the target word, this experiment uses a neural network to construct the mapping relationship, and the specific mapping model framework is shown in Fig. 5.

Figure 5.

Bilingual mapping model of domain word vector based on Neural Network.

In Fig. 5, import the vector of the source language text in the input layer of the network model. After two nonlinear change, the mapping vector of the text enters the output layer. The error sum between the mapping result and the word vector can be reduced by training, and the formula of the input layer to the output layer and the loss function of the model are shown in Eqs (10) and (11).

$\displaystyle o=g(W_{2}(g(W_{1}x+b_{1})+b_{2})$ (10) $\displaystyle E=\frac{1}{2}\sum\limits_{d\in S}{(t_{d}-o_{d})^{2}}$ (11)

In Eqs (10) and (11), $g(x)$ is the activation function, the parameter matrix $W_{1}$ indicates the input layer to the hidden layer, and the parameter matrix $W_{2}$ indicates the hidden layer of the network structure to the input layer of the network structure. $b_{1}$ and $b_{2}$ are the bias terms, and the word vector and mapping result of the $d$ th target word in the training sample are represented by $t_{d}$ and $o_{d}$ . The source language words are mapped to the target language end by bilingualism, and then the semantic $k$ nearest neighbors and semantic relatedness are calculated, based on which the mutual translation degree of word pairs $t(e^{(k)}|{f_{i}})$ is calculated as shown in Eq. (12).

$\displaystyle t(e^{(k)}|{f_{i}})=\frac{\exp\{sim(e^{(k)}|{e^{\prime})}\}}{\sum% \limits_{e^{(k)}\in S(e^{\prime})}{\exp\{sim(e^{(k)}|{e^{\prime})}\}}}$ (12)

In Eq. (12), $e^{(k)}$ denotes the semantic $k$ nearest neighbors of the word $f_{i}$ , $e^{\prime}$ is the mapping vector of $f_{i}$ , and $S(e^{\prime})$ is the set of nearest neighbors of $f_{i}$ . Finally, the scores of word pair inter-translation degree $t(e^{(k)}|{f_{i}})$ and semantic relatedness $\textit{sim}(e_{j},e^{(k)})$ are fused to obtain the translation probability $p(e_{j}|{f_{i}})$ as shown in Eq. (13).

$\displaystyle p(e_{j}|{f_{i}})=\sum\limits_{e^{(k)}\in S(e^{\prime})}{\textit{% sim}(e_{j},e^{(k)})t(e^{(k)}|{f_{i}})}$ (13)

Using the weighting formula, the phrase pair intertranslation probabilities were estimated as shown in Eq. (14).

$\displaystyle\textit{Sim}(\bar{e}|\bar{f},a)=\mathop{\prod}\limits_{j=1}^{% \textit{length}(\bar{e})}\frac{1}{|{\{i}|(i,j)\in a\}}\sum\limits_{(i,j)\in a}% {p(e_{j}|{f_{i}}})$ (14)

To enable a better evaluation of the domain interpreter degree, the interpreter probability is calculated from the opposite direction, as shown in Eq. (15).

$\displaystyle\textit{Sim}(\bar{f}|\bar{e},a)=\mathop{\prod}\limits_{i=1}^{% \textit{length}(\bar{f})}\frac{1}{|{\{j}|(i,j)\in a\}}\sum\limits_{(i,j)\in a}% {p(f_{i}|{e_{j}}})$ (15)

In Eq. (15), $\bar{f}$ and $\bar{e}$ are the source language and the target phrase with the length of $\textit{length}(f)$ and $\textit{length}(e)$ , respectively, and the word alignment relationship in the phrase pair is represented by $a$ . $f_{i}$ and $e_{j}$ with the forward and reverse domain intertranslation probabilities of $p(e_{j}|{f_{i}})$ and $p(f_{i}|{e_{j}})$ , respectively.

4. Simulation results of fuzzy semantic selection in machine translation and neural network-based domain analysis

To verify the rationality and accuracy of the selection optimization method of fuzzy semantics, this experiment is simulated and analyzed on MATLAB, and the data set used is CWMT09. Using the repetition rate and the accuracy checking rate as indicators, and the translation results of different methods are shown in Fig. 6.

Figure 6.

Simulation analysis results of repetition rate and precision rate.

In Fig. 6 that the method used in this experiment matches better with the repetition rate of the standard method of manual translation when compared with the traditional translation method; and the accuracy checking rate of the method used in this experiment is better than the traditional method when compared with the traditional method.

To test the optimality and validity of the neural network optimization approach based on Bi RNN and attention mechanism, the traditional Bi RNN and the improved neural network are tested in the dataset. Before conducting the tests, both models are trained in the dataset to identify the best performance of RNN, traditional CRNN and improved CRNN methods until the end of convergence. The LOSS plots of these two methods are shown in Fig. 7.

Figure 7.

Comparison diagram of loss curve of two models.

From Fig. 7, it can be seen that the optimized method gradually reaches a steady state with a value of 0.09 after 15 iterations, indicating that the number of iterations when the method starts to converge is 15. The number of iterations when the traditional Bi RNN method starts to converge is 18, and the loss function’s value is 0.1. The number of iterations of RNN method is 20, and its loss function value is 0.12. The comparison of the LOSS curves shows that the optimized neural network loss function value is smaller than the value of the traditional method and has better performance. The accuracy of these two methods was verified in the validation set during the iteration process, and Fig. 8 displayed the results.

In Fig. 8, the accuracy of the optimized neural network tends to stabilize at the number of iterations of 15 with an accuracy of 0.96, the accuracy of the traditional Bi RNN at the time of stabilization has a value of 0.91, and RNN is 0.82, which are lower than that of the improved method, indicating that the optimized neural network has higher performance.

In this study, six machine translation systems in the fields of journalism and education are constructed, including Baseline based on a generic domain parallel corpus, Mikolov_Sim based on matrix mapping, NN_Sim based on neural networks and calculating the probability of domain inter-translation, NN_Sim 1 based on neural networks and calculating the probability of forward translation, NN_Sim 2 based on neural networks and calculating the probability of reverse translation, and Sennrich_2012, which fuses text translation from different domains. The effectiveness and generality of the bilingual mapping models are verified by testing the performance of these six systems on the measurement set, and the BLEU metric is used for evaluation.

Table 1

Machine translation system performance (BLEU%)

Translation system	Journalism domain		Education domain
	Development set	Test set	Development set	Test set
Baseline	26.73	21.46	24.5	25.13
Mikolov_Sim	27.14	22.01 ( $+$ 0.55)	24.77	25.36 ( $+$ 0.23)
NN_Sim	27.21	22.13 ( $+$ 0.67)	24.79	25.69 ( $+$ 0.56)
NN_Sim l	26.90	21.60 ( $+$ 0.14)	24.84	25.27 ( $+$ 0.14)
NN_Sim 2	26.92	21.72 ( $+$ 0.26)	24.69	25.32 ( $+$ 0.19)
Sennrich_2012	27.07	21.92 ( $+$ 0.46)	24.85	25.65 ( $+$ 0.52)

Figure 8.

Comparison diagram of accuracy curve of two models.

In Table 1, the bilingual mapping model improves the BLEU by 0.67% in the news domain and 0.56% in the education domain compared with the benchmark system, proving that the method proposed in this experiment can improve the translation performance in the news and education domains. Compared with Mikolov_Sim system, the Bleu index of NN_Sim system in the field of news and education has increased by 0.12% and 0.23% respectively, which proves that the translation performance of mapping methods in different fields is improved differently when testing, and the method based on neural network can obtain better performance. The NN_Sim system constructed by the bilingual mapping method improves its BLEU metrics by 0.21% in the news domain compared to Sennrich_2012, while the upper position in the education domain is significantly different. This may be because the size of the training corpus in the education domain is lower than that in the news domain, proving that the method in this experiment has a certain degree of dependence on the size of the corpus.

To evaluate more objectively the performance of the bilingual mapping model used in the method, training is performed in the dataset and algorithm’s overall performance is measured by the translation accuracy $P$ , recall $R$ and $F$ values, where $F=\frac{(\beta^{2}+1)\times P\times R}{\beta^{2}\times P+R}$ , the preference preset value $\beta=1$ .

In Fig. 9, the sample size is closely related to the new method’s performance, and the translation performance gradually improves as the sample size increases. When the sample size is small, the model used in this experiment has a weak learning ability in the training set, so there is a high accuracy rate and low recall and $F$ values. When the sample size increases, the learning ability of the method increases, and thus the recall and $F$ values improve simultaneously. To compare the translation effects of different translation software directly, an ancient Chinese poem is selected for comparison, as shown in Table 2.

Table 2

Translation examples

Original text	UTF8gbsn忽如一夜春风来，千树万树梨花开。
Google	Suddenly, like a spring breeze, thousands of trees and pears bloom.
Baidu	Suddenly, like a night of spring breeze, thousands of pear trees bloom.
Youdao	Suddenly like a night of spring breeze, thousands of trees pear blossom.
This paper	As if spring breeze had come back overnight suddenly, white snow like thousands of pear trees bloom.

Figure 9.

The relationship between accuracy, recall, $F$ value and the number of training samples.

In ancient poems, “UTF8gbsn千…万…” means “many” and several translation software have correct translation. “UTF8gbsn梨花” means white snow, which is described in the translation method and not reflected in other translation software. This shows that the fuzzy semantic translation method combined with domain analysis has certain accuracy and superiority.

5. Conclusion

The verification results of the optimization of fuzzy semantic selection methods in machine translation show that the repetition rate of the method used in this experiment is more consistent with the standard method of human translation; the precision of the method used in this experiment is better. The accuracy of the optimized neural network tends to be stable when the number of iterations is 15, the accuracy is 0.96, and the accuracy of the traditional Bi RNN is 0.91. Compared with the benchmark system, the bilingual mapping model has increased by 0.67% in the news field and 0.56% in the education field. The precision, recall and value of machine translation are 93%, 86% and 0.8 respectively. The performance of the machine translation model optimized by the fuzzy semantic selection method has been effectively improved, and the performance of this model is closely related to the sample size. In this study, we innovatively build a bilingual mapping model based on the optimization of fuzzy semantic selection, combined with neural networks, to achieve semantic domain analysis, so as to improve the accuracy and effectiveness of machine translation. The disadvantage is that, when validating the domain analysis of bilingual mapping, it is found that the bilingual mapping model has a certain dependence on the corpus size, and further optimization of the method will be considered in the later research to improve the translation performance.

Footnotes

Funding

The research is supported by The First-class Major Construction of Business English in Xi’an Fanyi University (No. 2021-2025).

References

Heo

Kang

Yoo

. Multimodal neural machine translation with weakly labeled images. IEEE Access. 2019; 7(4): 54042-54053.

Yang

Liu

Chen

Zhang

Zhao

. A hierarchical clustering approach to fuzzy semantic representation of rare words in neural machine translation. IEEE Trans Fuzzy Syst. 2020; 28(5): 992-1002.

Julian-Iranzo

Saenz-Perez

. Proximity-based unification: An efficient implementation method. IEEE Trans Fuzzy Syst. 2020; 29(5): 1238-1251.

Qin

Hua

. Research on Uyghur-Chinese neural machine translation based on the transformer at multistrategy segmentation granularity. Mobile Information Systems; 2021.

Dugonik

Bokovi

Brest

Mauec

. Improving statistical machine translation quality using differential evolution. Inf. 2019; 30(4): 629-645.

Mahmoud

Mengash

. Machine translation utilizing the frequent-item set concept. Sens. 2021; 21(4): 1493-1513.

Shi

Huang

Wang

Jian

Kang

. Improving neural machine translation with sentence alignment learning. Neurocomput. 2021; 420(8): 15-26.

Sanchez-Gijon

Moorkens

Way

. Post-editing neural machine translation versus translation memory segments. Mach Transl. 2019; 33(1): 31-59.

Xie

Chen

. A position weighted information based word embedding model for machine translation. International Journal on Artificial Intelligence Tools; 2020.

10.

Wang

Chen

Yang

. Hybrid attention for Chinese character-level neural machine translation. Neurocomput. 2019; 358(8): 44-52.

11.

Kriegeskorte

Golan

. Neural network models and deep learning. Curr Biol. 2019; 29(7): 225-240.

12.

Benkova

Munkova

Benko

Munk

. Evaluation of English-Slovak neural and statistical machine translation. Appl Sci. 2021; 11(7): 2948-2964.

13.

Singh

Kumar

Chana

. Corpus based machine translation system with deep neural network for Sanskrit to Hindi translation. Proc Comput Sci. 2020; 167(1): 2534-2544.

14.

Leem

Yoo

Yook

. Multitask learning of deep neural network based keyword spotting for IoT devices. IEEE Trans Consum Electr. 2019; 65(2): 188-194.

15.

Kwon

Lee

. A text-based visual context modulation neural model for multimodal machine translation. Patt Recognit Lett. 2020; 136(8): 212-218.

16.

Liu

Chen

. Re-transformer: A self-attention based model for machine translation. Proc Comput Sci. 2021; 189(8): 3-10.

17.

Wang

Sun

. Spatiotemporal features-extracted travel time prediction leveraging deep-learning-enabled graph convolutional neural network model. Sustainability. 2021; 13(3): 1253-1267.

18.

Houdt

Mosquera

Nápoles

. A review on the long short-term memory model. Artif Intell Rev. 2020; 53(8): 5929-5955.

19.

Liu

Shen

Zhu

Niu

Zhang

. Deep learning based program generation from requirements text: Are we there yet? IEEE Trans Software Eng. 2020; 48(4): 1268-1289.

20.

Vathsala

Holi

. RNN based machine translation and transliteration for Twitter data. Int J Speech Technol. 2020; 23(3): 499-504.

21.

Balagovic

Daugherty

Entova-Aizenbud

Halacheva

Hennig

Seong Im

Letzter

Norton

Serganova

Stroppel

. Translation functors and decomposition numbers for the periplectic Lie superalgebra p(n). Math Res Lett. 2019; 26(3): 643-710.

22.

Zhao

Sun

. Mask attention-guided graph convolution layer for weakly supervised temporal action detection. Multimedia Tools Appl. 2021; 81(3): 4323-4340.

23.

Chen

Yong

. AU R-CNN: Encoding expert prior knowledge into R-CNN for action unit detection. Neurocomput. 2019; 355(25): 35-47.

24.

Ren

Wang

Cai

. A new hybrid-parameter recurrent neural network for online handwritten Chinese character recognition. Patt Recognit Lett. 2019; 128(8): 400-406.