Research on grammatical error correction algorithm in English translation via deep learning

1 Introduction

With the deepening of globalization, the prominence of English as the world’s primary language is becoming increasingly evident. In cross-cultural communication, English translation plays a pivotal role [1]. While machine translation algorithms aid in English translation, the current algorithms, particularly for real-time translation, are still imperfect. Therefore, individuals are encouraged to master English as thoroughly as possible, using machine translation algorithms as aids in overcoming communication barriers arising from translation discrepancies. The process of learning English may involve grammatical errors in the translated text due to unfamiliarity with the language environment [2]. While these errors may not pose significant issues in daily spoken communication, they can introduce ambiguity in situations requiring precise information conveyance. Thus, it is imperative to correct grammatical errors during the process of acquiring the English language. Traditionally, error correction has been a teacher-supervised process. However, teachers have limited energy, making it difficult for them to provide feedback on all grammar errors [3], thereby reducing the efficiency of learning English. The emergence of intelligent algorithms offers a promising solution for grammar error correction. Recent studies have made notable contributions in this area. For instance, Zhang [4] developed an English grammar error correction model based on a sequence-to-sequence (seq2seq) approach, utilized edge computing methods to enhance performance, and verified the effectiveness of the improvement. Zhou and Liu [5] proposed an English grammatical error correction algorithm based on a classification model, analyzed the model architecture and optimizer of the algorithm, and verified its efficacy through simulation experiments. Lin et al. [6] treated grammatical error correction as a multi-classification task and integrated a multiple-language embedding model and a deep learning model to rectify various lexical errors in Indonesian text. Experimental results demonstrated that the word embedding-based long and short-term memory (LSTM) model delivered the most effective learning outcomes. Li et al. [7] endeavored to incorporate contextual information from pre-trained language models as a solution for the scarcity of annotation in multilingual contexts. The findings demonstrated that bidirectional encoder representations from transformers held significant promise when employed for grammar correction tasks. Additionally, Dai [8] introduced an optimized random forest model and utilized it for automated detection and rectification of pronunciation errors within English classrooms. They validated its efficacy through experiments. Wanni [9] combined a genetic algorithm with a k-nearest neighbor algorithm to construct an intelligent English grammar correction model. The effectiveness of this model was verified through experiments. The aforementioned studies have all discussed grammar correction and proposed corresponding solutions. However, this study considers the grammar correction issue in English translations as a special type of translation problem, aiming to translate translated texts with grammatical issues into ones without grammatical problems, thus achieving grammar correction and providing valuable references for correcting grammatical errors in translations. This study provides a brief introduction to a grammatical error correction algorithm based on an encoder-decoder machine translation structure and introduces an attention mechanism to enhance the algorithm’s performance. Subsequently, simulation experiments were performed to evaluate the grammatical error correction algorithm.

2 Grammatical error correction of translated text based on a deep learning algorithm

When employing deep learning techniques to rectify grammatical errors in English translations, there are two main approaches. The first approach involves error correction using classification models, while the second focuses on correction through translation models [10]. In the former approach, common grammatical errors are first identified and labeled. Following this, a classification model is trained using a training dataset. This model is then utilized to predict the locations in the source text where grammatical errors may occur. If the prediction matches the source text, it is considered error-free; if not, a grammatical error is identified in that position within the source text [11]. In the latter approach, the problem of correcting grammatical errors is viewed as a translation problem. The source text, which may contain grammatical errors, is translated into a grammatically correct “translation.” Subsequently, by comparing the source text with the “translation,” grammatical problems are identified, and the “translation” becomes the corrected text [12].

Classification model-based grammatical error correction algorithms offer the advantage of relatively easier access to training corpora containing grammatical errors. However, their performance is limited by the specific types of errors present in the training corpus. In contrast, the grammatical correction algorithm treats the task of error correction as a sequence-to-sequence transformation task based on a translation model, which demonstrates its generalizability without being limited to specific types of grammar errors [13]. Consequently, this study uses the translation model to rectify the grammar of English translations.

The basic principle of the translation model-based grammatical error correction algorithm is shown in Figure 1. The grammatical error correction algorithm has a basic structure similar to that of the machine translation algorithm because both refer to the principles of machine translation and consist of an encoder and decoder. When the overall algorithm carries out grammatical error correction for English translation [14], first, source text sequence X which may contain grammatical errors is input to the encoder for computation to get intermediate sequence C , and then C is input to the decoder for computation to get output sequence Y after error correction. The specific steps are shown in Figure 2.

Figure 1

Fundamentals of translation model-based grammatical error correction.

Figure 2

Steps of grammatical error correction.

(1) The English translation is preprocessed by word division and denoising, after which the English translation is text-vectored using Word2vec [15]. During preprocessing, word segmentation can be used to split abbreviation combinations and divide long sentences into shorter phrases, avoiding the processing of excessively long sequences by the encoder and decoder. Denoising refers to removing punctuation and special symbols from sentences to avoid any impact on subsequent processing.

(2) The vectorized English translation is input to the encoder for calculation, and the LSTM algorithm [16] is used within the encoder. The corresponding equations are:

where C ˜ t and C t are the temporary state and update state of the “cell” at the current moment, h t is the hidden state of the sequence data at the current moment, x t is the input at the current moment, f t , i t , o t are the outputs of the three gating cells at the current moment, namely, the forgetting, input, and output, ω f , ω i , and ω o are the weights in the corresponding gating cell, b f , b i , b o are biases in the corresponding gating cell.

(3) The intermediate sequence output from the encoder is fed into the decoder for decoding computation. The LSTM algorithm is also used in the decoder, and the equations are shown in the previous section. The decoder converts the intermediate sequence into another sequence and uses the converted sequence as the English-translated sequence after error correction. The traditional encoding-decoding structure assigns equal importance to each moment of data within the intermediate sequence when converting the sequence, which prevents focusing on linking moments of high relevance during decoding and leads to missing information. For this reason, the attention mechanism [17] is introduced for the purpose of assigning appropriate weights to each moment of the intermediate sequence, enabling the decoder to focus more on key moments. The calculation process is:

where e t , i is the attention score of hidden state h i in the encoder at moment t , g ( ) is the function to solve the attention score, s t − 1 is the hidden state of the decoder at the previous moment, α t , i is the attention weight of h i at moment t , and c t is the intermediate sequence at moment t .

(4) After the decoder performs LSTM forward computation on the intermediate sequence with attention weights, the sequence of character distribution probabilities is finally obtained. At this point, it needs to be decoded to obtain a definite sequence of characters. This study adopts the cluster search algorithm to decode the sequence of character distribution probabilities: the first k characters with the highest probability of distribution within each moment of the sequence are selected as a candidate output sequence, and then the sequence with the highest probability is selected [18].

3 Simulation experiment

4 Discussion

With the deepening of globalization, there has been an increase in cross-cultural communication. The importance of English as the world’s largest language is becoming increasingly prominent. For non-native English speakers, differences in language and cultural environments often lead to grammatical errors during translation. Serious grammar mistakes can directly affect the expression of meaning in sentences, so non-native learners need to pay attention to correcting grammar while learning English. In traditional learning processes, teachers guide students in grammar correction. However, on one hand, the effectiveness of correction is influenced by the teacher’s proficiency level; on the other hand, teachers have limited energy and cannot provide one-on-one guidance to all students. The emergence of intelligent algorithms has provided a new way for grammar correction in English translation. This article regards the grammar correction issues in translated texts as a translation problem and corrects incorrect translations to achieve the identification and correction of grammatical errors in translations. The “encoder-decoder” structure of a translation model was applied to grammar correction in translations. The LSTM algorithm was used in both the encoder and decoder, and the attention mechanism was incorporated. Subsequently, simulation experiments were conducted to compare the improved LSTM algorithm with two other error correction algorithms. The results obtained have been shown above. The LSTM translation model employed in this article, which incorporated an attention mechanism, outperformed the other two algorithms. The reasons behind the aforementioned results were analyzed. The classification model-based algorithm relies on a classification model to predict and correct errors. However, this approach is limited in its ability to predict only specific types of errors and is constrained to specific locations within the text. This restricted scope results in the worst error correction performance. It also deteriorates when applied to English translations by freshmen. The traditional translation-based algorithm leverages machine translation principles to convert potentially erroneous phrases into correct ones without needing to pinpoint error locations. It is not limited by error types, making it superior to the classification model-based algorithm in terms of error correction performance. The attention mechanism-combined algorithm also employs machine translation principles but introduces the attention mechanism to enhance traditional machine translation. This mechanism helps highlight key elements in the intermediate sequence, reducing decoder interference during decoding and minimizing information loss. As a result, this algorithm exhibits even better error correction performance.

5 Conclusion

In summary, this study provides a brief introduction to grammatical error correction algorithms based on an encoder-decoder machine translation structure. The grammatical error correction algorithm was combined with the attention mechanism to enhance the algorithm’s performance and conducted simulation experiments to assess its effectiveness. These experiments involved comparing the algorithm with error correction algorithms based on a classification model and a traditional translation model using both a public corpus and English translations from freshmen. The key findings are as follows: (1) In the grammar correction model used in this article, the best error correction performance is achieved when employing a sigmoid activation function and neural nodes of 256 (128) for the encoder (decoder). (2) The attention mechanism-combined algorithm was capable of correcting the source text more accurately. (3) When tested with a public corpus, the attention mechanism-combined algorithm outperformed the others, followed by the traditional translation-based algorithm, and the classification model-based algorithm performed the least effectively. (4) When applied to English translations written by freshmen, the performance ranking of the three algorithms remained consistent. However, all algorithms experienced a decrease in error correction performance, with the attention mechanism-combined algorithm showing a relatively smaller reduction. The contribution of this article lies in treating the grammar correction issue in English translations as a specific translation problem and correcting the translated text with grammar mistakes into one without any grammatical problems, thereby providing an effective reference for correcting the grammar of translated texts.