Translation question correcting method and device, storage medium and computer equipment

Abstract

The invention discloses a translation question correcting method and device, a storage medium and computer equipment. The method comprises the steps that question statements of translation questions and answer statements for translating the translation questions are obtained; according to the question statement and the answer statement, multi-dimensional scoring is carried out on the answer condition of the translation question, and scoring dimensions comprise at least two of a semantic dimension, a grammar dimension, a tone dimension, a word dimension and a fluency dimension; and performing comprehensive score calculation on the answer statement according to the multi-dimensional score to obtain a total answer score of the answer statement. According to the method, the translation ability of the students can be comprehensively and objectively evaluated by performing multi-dimensional scoring on the answering statements of the translation questions, one-sidedness caused by single-dimensional or subjective judgment is avoided, learning characteristics and weaknesses of the students are more accurately known by evaluating the answering conditions in different dimensions, and the learning efficiency of the students is improved. Therefore, more targeted guidance and suggestions are provided, and personalized learning is promoted.

Description

Technical Field

[0001] The present application relates to the technical field of translation question marking, and in particular, to a method and device for marking translation questions, a storage medium, and a computer device. Background Art

[0002] With the acceleration of the globalization process, cross - language communication has become increasingly common and important. As an international common language, the learning demand for English is also constantly increasing. In foreign language learning, translation practice is an important part. Through translation practice, students can better understand language structures, vocabulary usage, and cultural differences. However, currently, the marking of most English translation questions on the market still relies on human teachers, which is not only time - consuming and laborious, but also may lead to inconsistent judgment criteria due to subjective factors, thus affecting students' learning effects.

[0003] Some automatic marking systems have been developed in the prior art, but these systems mainly focus on objective question types such as multiple - choice questions or fill - in - the - blank questions, and are inadequate for translation questions involving complex grammar and semantic analysis. Therefore, there is an urgent need for a method that can mark translation questions efficiently, accurately, and comprehensively to improve teaching efficiency and learning effects. Summary of the Invention

[0004] In view of this, the embodiments of the present application provide a method and device for marking translation questions, a storage medium, and a computer device, which can comprehensively and objectively evaluate students' translation abilities by performing multi - dimensional scoring on the answer sentences of translation questions.

[0005] According to one aspect of the present application, a method for marking translation questions is provided. The method includes:

[0006] Obtaining the question sentence of the translation question and the answer sentence for translating the translation question;

[0007] Performing multi - dimensional scoring on the answering situation of the translation question according to the question sentence and the answer sentence, where the scoring dimensions include at least two of the semantic dimension, the grammar dimension, the tone dimension, the word dimension, and the fluency dimension;

[0008] Performing a comprehensive scoring calculation on the answer sentence according to the multi - dimensional scoring to obtain the total answer score of the answer sentence.

[0009] According to another aspect of the present application, a device for marking translation questions is provided. The device includes:

[0010] An obtaining module, configured to obtain the question sentence of the translation question and the answer sentence for translating the translation question;

[0011] A multi-dimensional scoring module is used to perform multi-dimensional scoring on the answering situation of the translation question according to the question statement and the answering statement, where the scoring dimensions include at least two of a semantic dimension, a grammar dimension, a tone dimension, a word dimension, and a fluency dimension;

[0012] An overall score statistics module is used to perform an overall score calculation on the answering statement according to the multi-dimensional scoring to obtain the total answering score of the answering statement.

[0013] According to another aspect of the present application, there is provided a storage medium on which a computer program is stored, and when the program is executed by a processor, the above-mentioned method for correcting translation questions is implemented.

[0014] According to still another aspect of the present application, there is provided a computer device, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor. When the processor executes the program, the above-mentioned method for correcting translation questions is implemented.

[0015] By means of the above technical solutions, a method and device for correcting translation questions, a storage medium, and a computer device provided by the embodiments of the present application perform multi-dimensional scoring according to the question statement and the answering statement of the translation question, so as to realize the evaluation of the total answering score of the answering statement according to the multi-dimensional scoring. By performing multi-dimensional scoring on the answering statement of the translation question, the present application can comprehensively and objectively evaluate the translation ability of students, avoid the one-sidedness brought by single-dimensional or subjective judgment, and can more accurately understand the learning characteristics and weaknesses of students by evaluating the answering situation in different dimensions, so as to provide more targeted guidance and suggestions to promote personalized learning. Moreover, through this method, automatic correction of translation questions can be realized, improving the correction efficiency, reducing the correction burden of teachers, avoiding the subjectivity and inconsistency that may exist in manual correction, and ensuring the fairness and accuracy of the judgment.

[0016] The above description is only an overview of the technical solutions of the present application. In order to be able to understand the technical means of the present application more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features, and advantages of the present application more obvious and understandable, the following specifically illustrates the specific embodiments of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation to the present application. In the drawings:

[0018] Figure 1 A flowchart showing a method for correcting translation questions provided by an embodiment of the present application is shown;

[0019] Figure 2 It shows a schematic flowchart of another method for grading translation questions provided by an embodiment of the present application;

[0020] Figure 3A It shows a schematic flowchart of a method for constructing a semantic scoring large model provided by an embodiment of the present application;

[0021] Figure 3B It shows a schematic structural diagram of a semantic scoring large model provided by an embodiment of the present application;

[0022] Figure 3C It shows a schematic structural diagram of another semantic scoring large model provided by an embodiment of the present application;

[0023] Figure 4 It shows a schematic structural diagram of a device for grading translation questions provided by an embodiment of the present application;

[0024] Figure 5 It shows a schematic structural diagram of another device for grading translation questions provided by an embodiment of the present application. Detailed implementation manners

[0025] The present application will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments. It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0026] In this embodiment, a method for grading translation questions is provided. As Figure 1 shown, the method includes:

[0027] Step 101: Obtain the question statement of the translation question and the answer statement for translating the translation question.

[0028] Step 102: Perform multi-dimensional scoring on the answering situation of the translation question according to the question statement and the answer statement, where the scoring dimensions include at least two of the semantic dimension, the grammar dimension, the tone dimension, the word dimension, and the fluency dimension.

[0029] Step 103: Perform comprehensive scoring calculation on the answer statement according to the multi-dimensional scoring to obtain the total answer score of the answer statement.

[0030] In the embodiments of the present application, for translation questions that need to be corrected, such as Chinese-to-English translation questions. First, obtain the question statement of the translation question and the answer statement of the user's translation answer for the translation question. For example, for a Chinese-to-English translation question, the question statement is the Chinese sentence to be translated, and the answer statement is the English sentence translated by the user for the Chinese sentence. Then, perform multi-dimensional scoring on the user's answer statement, which can specifically include semantic dimension, grammar dimension, tone dimension, word dimension, and fluency dimension, etc. Among them, the scoring of each dimension mainly scores for the corresponding dimension, without paying too much attention to the answering situations of other dimensions. For example, the semantic dimension mainly examines whether the meaning expressed by the user's translated answer statement matches the question statement, and does not conduct too much examination on whether the answer statement is grammatically correct or the spelling of words is correct, etc. This helps to improve the comprehensiveness and objectivity of scoring, and also helps to better identify the knowledge dimensions lacking by oneself during the foreign language learning process, conduct targeted learning, and improve the foreign language learning effect. Finally, perform comprehensive calculation on the multi-dimensional scores to obtain the total answer score of the answer statement, so as to obtain the overall score of the answer statement and realize the automatic correction of translation questions. Specifically, the total answer score can be calculated by the method of weighted summation. For example, the semantic dimension score is S1, the grammar dimension score is S2, the tone dimension score is S3, the fluency dimension score is S4, the word dimension score is S5, and the total answer score is S, then S = w1*S1 + w2*S2 + w3*S3 + w4*S4 + w5*S5. In this embodiment, the sum of the weights of each score is equal to 1, that is, w1 + w2 + w3 + w4 + w5 = 1, where the semantic score weight is 0.3, the grammar score weight is 0.3, the word score weight is 0.2, the tone score weight is 0.1, and the fluency score weight is 0.1. In practical applications, the weights of each score can be adjusted according to the business scenario.

[0031] By applying the technical solution of this embodiment, perform multi-dimensional scoring based on the question statement and the answer statement of the translation question, so as to realize the evaluation of the total answer score of the answer statement according to the multi-dimensional scoring. Through multi-dimensional scoring of the answer statement of the translation question in this application, the translation ability of students can be comprehensively and objectively evaluated, avoiding the one-sidedness brought by a single dimension or subjective judgment. And by evaluating the answering situations in different dimensions, the learning characteristics and weaknesses of students can be more accurately understood, so as to provide more targeted guidance and suggestions to promote personalized learning. Moreover, through this method, the automatic correction of translation questions can be realized, improving the correction efficiency, reducing the correction burden of teachers, avoiding the subjectivity and inconsistency that may exist in manual correction, and ensuring the fairness and accuracy of judgment.

[0032] Further, as a refinement and extension of the specific implementation manner of the above embodiment, in order to completely illustrate the specific implementation process of this embodiment, another method for grading translation questions is provided to implement semantic dimension scoring for the answering situation of translation questions. For example, Figure 2 as shown, this method includes:

[0033] Step 201: Perform first-language keyword scoring and second-language keyword scoring respectively according to the question statement in the first language and the answer statement in the second language.

[0034] Step 202: If the first-language keyword scoring and the second-language keyword scoring meet the preset conditions, calculate the semantic similarity between the question statement and the answer statement through a semantic similarity scoring model to determine the semantic dimension score; wherein, the preset condition is that at least one of the first-language keyword scoring being greater than the first scoring threshold and the second-language keyword scoring being greater than the second scoring threshold is satisfied.

[0035] Step 203: If the first-language keyword scoring and the second-language keyword scoring do not meet the preset conditions, determine the semantic dimension score based on the question statement and the answer statement through a semantic scoring large model.

[0036] In the embodiment of the present application, when performing semantic dimension scoring, first, perform first-language keyword scoring and second-language keyword scoring according to the question statement in the first language and the answer statement in the second language. Taking the first language as Chinese and the second language as English as an example. Score the user's answering situation from the perspectives of Chinese and English according to the Chinese question statement and the English answer statement respectively. Then, determine whether at least one of the first-language keyword scoring being greater than the first scoring threshold and the second-language keyword scoring being greater than the second scoring threshold is satisfied. If at least one is satisfied, it indicates that the answer statement has shown a certain degree of accuracy or relevance at the keyword level. Then, use a semantic similarity scoring model to calculate the semantic similarity between the question statement and the answer statement, thereby determining the semantic dimension score. This model can quantify the semantic proximity between two statements, thereby providing an objective semantic dimension score. Otherwise, it means that there is a large deviation between the answer and the question requirements at the keyword level. At this time, use a more comprehensive semantic scoring large model to determine the semantic dimension score. This model may consider more language features, context information, and the complexity of semantic relationships.

[0037] In the embodiment of the present application, optionally, the first-language keyword scoring in step 201 according to the question statement in the first language and the answer statement in the second language includes:

[0038] Step 201-1: Translate each word in the response statement into the first language to obtain a list of first-language keywords, or translate the response statement into the first language and then perform word segmentation on the translated response statement in the first language to obtain a list of first-language keywords;

[0039] Step 201-2: Perform word segmentation on the question statement to obtain a list of question keywords;

[0040] Step 201-3: Match each keyword in the list of question keywords with the list of first-language keywords one by one, and determine the first-language keyword score according to the matching situation between each keyword in the list of question keywords and the list of first-language keywords.

[0041] In this embodiment, when determining the first-language keyword score, first generate a list of first-language keywords. Specifically, each word in the response statement can be directly translated into the first language and combined into a list of first-language keywords. It is also possible to first translate the entire response statement into the first language, and then perform word segmentation on the translated statement to obtain a list of first-language keywords. In actual operation, an appropriate method can be selected according to specific requirements and the language characteristics of the response statement. Secondly, perform word segmentation on the question statement to obtain a list of question keywords containing each word segment in the question statement. Finally, for each keyword in the list of question keywords, match it with the list of first-language keywords one by one. The matching can be an exact match or a semantic approximate match. Determine the first-language keyword score according to the matching situation. The score can be comprehensively calculated based on factors such as the proportion of matching keywords, the importance of keywords (such as part of speech, syntactic function, etc.), and the accuracy of matching (such as the weight difference between exact match and approximate match).

[0042] In an embodiment of the present application, optionally, in step 201, according to the question statement in the first language and the response statement in the second language, perform second-language keyword scoring, including:

[0043] Step 201-4: Perform word segmentation on the response statement to obtain a list of second-language keywords;

[0044] Step 201-5: Perform word segmentation on the correct response statement corresponding to the question statement to obtain a list of answer keywords;

[0045] Step 201-6: Match each keyword in the list of answer keywords with the list of second-language keywords one by one, and determine the second-language keyword score according to the matching situation between each keyword in the list of answer keywords and the list of second-language keywords.

[0046] In this embodiment, first, word segmentation is performed on the response statement to obtain a second-language keyword list containing each word segment in the response statement. Next, word segmentation is performed on the correct response statement corresponding to the question statement to obtain an answer keyword list, which represents the keywords in the correct answer expected by the question and is used to match the keywords in the response statement. After that, each keyword in the answer keyword list is matched with the second-language keyword list one by one, where the match can be an exact match or a semantic approximate match, depending on the requirements and implementation of the scoring system. Finally, the second-language keyword score is determined based on the matching situation. The score can be comprehensively calculated based on factors such as the proportion of the number of matching keywords, the importance of the keywords (such as part of speech, syntactic function, position in the answer, etc.), and the accuracy of the match (such as the weight difference between exact match and approximate match).

[0047] In an embodiment of the present application, optionally, in step 202, semantic similarity calculation is performed on the question statement and the response statement through a semantic similarity scoring model to determine the semantic dimension score, including:

[0048] Step 202-1: Obtain a first translation statement obtained by translating the response statement into the first language;

[0049] Step 202-2: Input the question statement, the response statement, the correct response statement corresponding to the question statement, and the first translation statement into the semantic similarity scoring model to obtain the semantic dimension score. Among them, the semantic similarity scoring model calculates the first semantic similarity between the question statement and the first translation statement, the second semantic similarity between the question statement and the response statement, the third semantic similarity between the correct response statement and the response statement, and determines the semantic dimension score according to the first semantic similarity, the second semantic similarity, and the third semantic similarity.

[0050] In this embodiment, the response statement is translated into the first language (i.e., the language of the question statement) to obtain the first translated statement, so as to ensure that the content of the response statement can be compared with the question statement in the form of the first language, thereby calculating the semantic similarity. Then, the question statement, the response statement, the correct response statement corresponding to the question statement, and the first translated statement are input into a pre-trained semantic similarity scoring model. Inside the semantic similarity scoring model, the following three semantic similarities are calculated respectively: The first semantic similarity: the semantic similarity between the question statement and the first translated statement, which reflects the semantic proximity between the response statement after being translated back into the first language and the question statement. The second semantic similarity: the semantic similarity between the question statement and the response statement (in the original second language form), which directly evaluates the semantic matching degree between the response statement and the question statement in the original language form. The third semantic similarity: the semantic similarity between the correct response statement and the response statement, which measures the semantic consistency between the response statement and the expected correct answer. According to the calculated first semantic similarity, second semantic similarity, and third semantic similarity, combined with the preset weights or scoring rules inside the model, the semantic dimension score is comprehensively determined to comprehensively reflect the proximity of the response statement to the question statement and the correct response statement at the semantic level, thereby providing a more comprehensive and accurate semantic dimension score.

[0051] In an embodiment of the present application, optionally, after step 202, it further includes: if the semantic dimension score is less than the semantic score threshold, the semantic scoring large model is used to re-determine the semantic dimension score according to the question statement and the response statement; if the semantic dimension score is greater than or equal to the semantic score threshold, the semantic dimension score remains unchanged.

[0052] In this embodiment, after determining the semantic dimension score according to the semantic similarity scoring model, an additional judgment and processing process is introduced to decide whether to re-determine the score through the semantic scoring large model according to the value of the semantic dimension score. Specifically, the semantic score threshold is a value preset according to factors such as teaching requirements, question difficulty, and scoring criteria, and is used to judge the rationality and accuracy of the semantic dimension score. When the semantic dimension score is less than the semantic score threshold, this may mean that the initial semantic similarity calculation is not accurate or comprehensive enough, or there is a large deviation in semantics between the response statement and the question statement. In this case, the semantic scoring large model is used to re-perform semantic analysis according to the question statement and the response statement, and determine a new semantic dimension score. When the semantic dimension score is greater than or equal to the semantic score threshold, it indicates that the response statement is relatively close to the question statement in semantics. In this case, the semantic dimension score remains unchanged. Through the above processing process and optimization suggestions, the semantic scoring model can be more effectively used to evaluate the semantic consistency between the response statement and the question statement, while ensuring the accuracy and rationality of the scoring.

[0053] In an embodiment of the present application, optionally, in step 203, determining the semantic dimension score by the semantic scoring large model based on the topic statement and the response statement includes:

[0054] Step 203-1: Obtain a second translated statement obtained by translating the response statement into a first language.

[0055] Step 203-2: Input the topic statement, the response statement, the correct response statement corresponding to the topic statement, and the second translated statement into the semantic scoring large model to obtain the semantic dimension score, where the semantic scoring large model determines a first semantic score according to the response statement and the correct response statement, determines a second semantic score according to the second translated statement and the topic statement, determines a third semantic score according to the response statement and the topic statement, and determines the semantic dimension score according to the first semantic score, the second semantic score, and the third semantic score.

[0056] In this embodiment, in the process of determining the semantic dimension score by the semantic scoring large model based on the topic statement and the response statement, first, the response statement is translated into the first language (i.e., the language of the topic statement) to obtain a second translated statement, so as to ensure that the content of the response statement can be compared with the topic statement and other reference statements (such as the correct response statement) in the form of the first language, thereby more comprehensively evaluating its semantic consistency. Then, the topic statement, the response statement, the correct response statement corresponding to the topic statement, and the second translated statement are input into the pre-trained semantic scoring large model. The semantic scoring large model determines a first semantic score according to the semantic similarity or consistency between the response statement and the correct response statement to reflect the semantic proximity between the response statement and the expected correct answer. The second semantic score is determined according to the semantic similarity between the second translated statement and the topic statement to measure the semantic consistency between the response statement translated back into the first language and the topic statement. The third semantic score is determined according to the semantic similarity between the response statement (in the original second language form) and the topic statement to evaluate the semantic matching degree between the response statement and the topic statement in the original language form. Finally, the semantic scoring large model comprehensively determines the final semantic dimension score according to the first semantic score, the second semantic score, and the third semantic score, combined with the preset weights or scoring rules inside the model. This score can comprehensively reflect the semantic consistency of the response statement with the topic statement and the correct response statement in multiple dimensions. Thus, the semantic scoring large model is effectively used to evaluate the semantic consistency between the response statement and the topic statement and the correct response statement, thereby providing a more comprehensive and accurate semantic dimension score.

[0057] In an embodiment of the present application, optionally, as Figure 3A shown, the method for constructing a semantic scoring large model includes:

[0058] Step 301: Construct an initial semantic scoring large model, where the initial semantic scoring large model includes an input layer, a main branch, a first side branch, a second side branch, a fully connected network, and an output layer. The input layer is respectively connected to the main branch, the first side branch, and the second side branch. The fully connected network is respectively connected to the main branch, the first side branch, and the second side branch. The output layer is connected to the fully connected network. The main branch is a network structure based on a large model;

[0059] Step 302: Obtain semantic scoring large model training samples, where the semantic scoring large model training samples include question sentence samples, answer sentence samples, correct answer sentence samples corresponding to the question sentence samples, translation samples corresponding to the answer sentence samples, and total annotation scores;

[0060] Step 303: Input the question sentence samples, the answer sentence samples, the correct answer sentence samples, and the translation samples into the input layer. Calculate a first predicted score between the question sentence samples and the correct answer sentence samples through the main branch, calculate a second predicted score between the translation samples and the question sentence samples through the first side branch, calculate a third predicted score between the answer sentence samples and the question sentence samples through the second side branch, determine a total predicted score through the fully connected network according to the first predicted score, the second predicted score, and the third predicted score, and obtain the total predicted score output by the output layer;

[0061] Step 304: Optimize the parameters of the main branch, the first side branch, the second side branch, and the fully connected network according to the total predicted score and the total annotation score, and obtain the semantic scoring large model based on the input layer, the output layer, and the main branch, the first side branch, and the second side branch after parameter optimization.

[0062] In this embodiment, during the process of constructing the semantic scoring large model, first, construct an initial semantic scoring large model, where, as Figure 3BAs shown in the figure, the initial semantic scoring large model consists of an input layer, a main branch, a first side branch, a second side branch, a fully connected network, and an output layer. The input layer is responsible for receiving input information such as the question statement sample, the answer statement sample, the correct answer statement sample, and the translation sample. The main branch is a network structure based on the large model, which is used to calculate the first prediction score between the question statement sample and the correct answer statement sample. The first side branch is used to calculate the second prediction score between the translation sample (the translation of the answer statement sample) and the question statement sample. The second side branch is used to calculate the third prediction score between the answer statement sample and the question statement sample. The fully connected network comprehensively determines the total prediction score according to the first prediction score, the second prediction score, and the third prediction score. The output layer is used to output the total prediction score. Specifically, the main branch of the large model adopted in this embodiment is the large model in the prior art, such as the Qianwen 7B model. The output layer of the Qianwen 7B model is removed, and the feature vector of the penultimate layer is concatenated with the feature vectors of the penultimate layer of the two side branch models to form a large feature vector. An N-layer fully connected network is connected after this feature vector, and finally the binary classification result is output, that is, the confidence of similarity and the confidence of dissimilarity. The side branch 1 and side branch 2 of the large model can adopt the structBert model. The output layer is removed, and the neurons of the penultimate layer are combined with the neurons of the penultimate layer of the large model to form the input of the fully connected network. Other models can also be used in actual applications. During training, the parameters of the large model can be fixed, and the semantic understanding ability of the large model can be used to guide and train the parameters of the side branches and the fully connected network. The advantage of this design is to use the semantic understanding ability of the large model to extract the similar features between the user's answer and the answer in the main branch, and at the same time use the side branches to extract the similar features of Chinese sentence pairs and the similar features of Chinese-English sentence pairs. By comprehensively judging these three types of features, whether the sentence translated by the user is correct can be obtained with better results. During training, the parameters of the large model can also be not fixed, and it can be optimized together with the two side branch models and the fully connected network. Further, taking the case where the parameters of the large model are not fixed as an example, the training samples of the semantic scoring large model are obtained and the model is trained. The samples include the question statement sample, the answer statement sample, the correct answer statement sample corresponding to the question statement sample, the translation sample corresponding to the answer statement sample (that is, the translation of the answer statement into the language of the question statement), and the total annotation score (that is, the manually given semantic score). During model training, the question statement sample, the answer statement sample, the correct answer statement sample, and the translation sample are input into the input layer. The first prediction score is calculated through the main branch, the second prediction score is calculated through the first side branch, and the third prediction score is calculated through the second side branch. The fully connected network comprehensively determines the total prediction score according to the first prediction score, the second prediction score, and the third prediction score, and outputs the total prediction score through the output layer. According to the loss between the total prediction score and the total annotation score, the parameters of the main branch, the first side branch, the second side branch, and the fully connected network are tuned until the model meets the training conditions.Finally, based on the input layer, output layer, and the main branch, the first side branch, and the second side branch with optimized parameters, the final large semantic scoring model is constructed. Thus, an efficient and accurate large semantic scoring model is built to evaluate the semantic consistency between the response statement and the question statement and the correct response statement.

[0063] In a specific application scenario, after the training of the large semantic scoring model is completed, as Figure 3C shown, the question statement, the response statement, the correct response statement corresponding to the question statement, and the second translation statement are input into the input layer of the large semantic scoring model. The main branch determines the first semantic score based on the response statement and the correct response statement. The first side branch determines the second semantic score based on the second translation statement and the question statement. The second side branch determines the third semantic score based on the response statement and the question statement. The fully connected network determines the semantic dimension score based on the first semantic score, the second semantic score, and the third semantic score. The output layer outputs the semantic dimension score.

[0064] In an embodiment of the present application, optionally, step 302 includes:

[0065] Step 302-1: Obtain the question statement sample, the original response statement sample, the correct response statement sample, and the total annotation score;

[0066] Step 302-2: Select some words in the original response statement sample, and perform similar word replacement on the selected words to obtain the first simulated response statement corresponding to the original response statement sample, where the similar word replacement includes at least one of synonym replacement, part of speech replacement, singular and plural replacement, preposition replacement, and replacement of words with similar spellings but different meanings;

[0067] Step 302-3: Perform grammar rewriting on the original response statement sample to obtain the second simulated response statement corresponding to the original response statement sample, where the grammar rewriting includes at least one of tense rewriting, article rewriting, preposition rewriting, pronoun reference rewriting, sentence structure rewriting, negative form rewriting, singular and plural word usage rewriting, and fixed collocation usage rewriting;

[0068] Step 302-4: Use the original response statement sample, the first simulated response statement, and the second simulated response statement as the response statement sample, and construct the training sample of the large semantic scoring model according to the question statement sample, the response statement sample, the correct response statement sample, the translation sample corresponding to the response statement sample, and the total annotation score.

[0069] In this embodiment, the sample of the question sentence, the sample of the original answer sentence, the sample of the correct answer sentence, and the total annotation score corresponding to these samples are obtained, and these data are the basis for constructing the training sample. Then, the original answer sentence sample can be rewritten with words and grammar to construct more and richer samples. Further, the original answer sentence sample, the first simulated answer sentence obtained after the word rewriting, and the second simulated answer sentence obtained after the grammar rewriting can be used as the answer sentence sample for the construction of the semantic scoring large model training sample. Specifically, in the training sample constructed by the first simulated answer sentence and the second simulated answer sentence, the total annotation score can be the same as the total annotation score corresponding to the corresponding original answer sentence sample. It is also possible to perform a symbolic weighted penalty when calculating the similarity score for such inaccurately worded and grammatically inaccurate sentences according to the number of inappropriate words and grammar in the sentence, but still keep the similarity score greater than a specific threshold, and keep there is no excessive difference between the total annotation score of the corresponding original answer sentence sample. Through the above method, richer and more diverse semantic scoring large model training samples can be constructed, thereby improving the training effect and performance of the model.

[0070] Among them, when rewriting words, some words are selected from the original answer sentence sample, and these words are replaced with similar words to generate the first simulated answer sentence corresponding to the original answer sentence sample. Similar word replacement can include synonym replacement, part of speech replacement, singular and plural replacement, preposition replacement, and at least one method of word replacement with similar spelling but different meanings. This replacement is intended to simulate the lexical level variation that may occur in the answer. Specifically, this embodiment can randomly modify a word to a word with the same Chinese meaning but inaccurate wording according to the length of the sentence, such as randomly modifying one word for every five words. In actual application, the words in the sentence can be modified according to the specific scenario. In addition to this word replacement modification, that is, synonym modification, there are also word form modifications, such as verbs to nouns, nouns to adjectives, adjectives to nouns, etc. There are also tense modifications, such as changing the present tense to the past tense, and the past tense to the present tense. There are also singular to plural, plural to singular, etc. In the translation process, common forms of inaccurate wording also include the following, and various types of word rewriting can be performed with reference to the following inaccurate wording:

[0071] 1. Literal translation without context: Directly translating word for word without considering the customary expressions in the target language. For example, translating "他很牛" literally as "He is very cow", the correct expression should be "He is awesome".

[0072] 2. Misuse of synonyms: Using the wrong synonyms or near-synonyms, resulting in a deviation in meaning. For example, translating "I feel embarrassed" as "I feel embarrassed" is correct, but if translated as "I feel awkward", although it can sometimes be used, it may be inappropriate in a specific context.

[0073] 3. Cultural differences: Ignoring the cultural differences between the two languages. For example, the Chinese greeting "Have you eaten?" when translated directly as "Have you eaten?" may confuse native English speakers because this is not a common greeting in English.

[0074] 4. Incorrect use of prepositions: The rules for using prepositions often vary between languages. For example, "rely on someone" should be translated as "rely on someone", not "rely to someone".

[0075] 5. Mistakes in fixed collocations: Ignoring the fixed collocations and idiomatic expressions in the target language. For example, translating "make a decision" as "make a decision" is correct, but translating it as "do a decision" is wrong.

[0076] 6. Misunderstanding of word parts of speech: Confusing the use of word parts of speech, such as using an adjective as a verb or vice versa. For example, the confusion between "interesting" and "interested": "I am interesting in this book" should be changed to "I am interested in this book".

[0077] 7. Words that are similar in spelling but different in meaning: For example, "advice" (advice) and "advise" (the verb form of advice), as well as "affect" (influence) and "effect" (effect).

[0078] Furthermore, the original response sentence samples can be rewritten grammatically to generate second simulated response sentence samples corresponding to the original response sentence samples. Grammatical rewriting can include at least one of the following ways: tense rewriting, article rewriting, preposition rewriting, pronoun reference rewriting, sentence structure rewriting, negative form rewriting, singular / plural word usage rewriting, fixed collocation phrase rewriting, etc. This rewriting aims to simulate the possible grammatical variations in the response. During the translation process, common grammar error sentences can be constructed: 1. Tense errors: Ignoring the tense changes in the target language. For example, directly translating "He went to Beijing yesterday" as "He goes to Beijing yesterday", and the correct one should be "He went to Beijing yesterday".

[0079] 2. Subject-verb inconsistency: The subject and predicate do not match in person and number. For example, "She likes apples" should be changed to "She likes apples".

[0080] 3. Incorrect use of articles: There are no articles in Chinese, but there are definite articles (the) and indefinite articles (a, an) in English. For example, "I saw a cat" should be translated as "I saw cat" instead of "I saw a cat".

[0081] 4. Preposition misuse: Different languages ​​may have different rules for the use of prepositions. For example, "rely on someone" is translated into "rely on someone" instead of "rely to someone".

[0082] 5. Unclear pronouns: In the translation process, if the pronouns do not refer to a clear object, it will lead to ambiguity or ambiguity. For example, "Zhang San told Li Si that he was happy" can be translated into "Zhang San told Li Si that he was happy", where "he" can refer to anyone.

[0083] 6. Confused sentence structure: Due to the different sentence structures of the two languages, direct word-for-word translation will result in confusing sentence structures. For example, if "because it was raining, I didn't go to the park" is directly translated into "Because raining, I didn't go park", the correct translation should be "Because it was raining, I didn't go to the park".

[0084] 7. Negative form errors: It is easy to make mistakes when dealing with negative sentences. For example, "He is never late" is translated as "Henever is late" and should be changed to "He is never late".

[0085] 8. Plural form error: Ignoring the singular and plural changes of nouns. For example, "many books" should be translated into "many books".

[0086] 9. Confusion between adjectives and adverbs: For example, “He runs quick” is translated as “He runs quick” and should be changed to “He runs quickly”.

[0087] 10. Collocation errors: Failure to use collocations in the target language accurately, for example, translating “做決定” into “do a decision” instead of “make a decision”.

[0088] By paying attention to these common problems, a large number of sentences with incorrect grammar or inaccurate word usage are added when constructing semantic training data to mimic the real answering situations of students. The addition of this kind of data enables the semantic scoring large model to give more accurate semantic scores semantically. This kind of grammar error and inaccurate word usage actually means that students are trying to express the meaning described in Chinese. Taking the example of Chinese-English translation, the semantic scoring large model only gives Chinese semantic scores based on the Chinese meaning corresponding to the English translated by the students, that is, how much Chinese meaning of the Chinese sentence to be translated is described, and whether the grammar and word usage of the description are accurate is not within the scope of evaluation of this model.

[0089] In an embodiment of the present application, optionally, a grammar dimension score is given to the answering situation of the translation question, including: inputting the answering sentence into a grammar analyzer, analyzing the grammar problems existing in the answering sentence through the grammar analyzer, and determining the grammar dimension score according to the grammar problems.

[0090] In this embodiment, first, the answering sentence needs to be input into a special grammar analyzer, and the grammar analyzer makes a detailed analysis of the input answering sentence. This analysis process aims to identify any grammar problems existing in the answering sentence. Grammar problems may include but are not limited to tense errors, subject-verb disagreement, lack of necessary punctuation, improper collocation of words, etc. After identifying the grammar problems in the answering sentence, a grammar dimension score is determined according to the nature and quantity of these problems. This score reflects the grammatical accuracy of the answering sentence.

[0091] In an embodiment of the present application, optionally, a tone dimension score is given to the answering situation of the translation question, including: respectively performing tone analysis on the question sentence and the answering sentence, and determining the tone dimension score based on the difference degree between the tone of the question sentence and the tone of the answering sentence.

[0092] In this embodiment, first, tone analysis is performed on the original sentence of the translation question (i.e., the question sentence), such as identifying tone features such as emotional color, formality, politeness, directness or indirectness in the sentence. Then, the same tone analysis is performed on the student's answering sentence. After completing the tone analysis, the system needs to calculate the tone difference degree between the question sentence and the answering sentence. Specifically, it can be achieved by comparing the similarity or difference degree of the two in tone features such as emotional color, formality, politeness, etc. Based on the calculated tone difference degree, the corresponding tone dimension score can be further determined to reflect the matching degree of the answering sentence with the question sentence in terms of tone. The higher the score, the closer the answering sentence is to the question sentence in terms of tone; the lower the score, the greater the difference between the two.

[0093] In an embodiment of the present application, optionally, a word - dimension score is given to the response to the translation question, including: obtaining a list of question keywords obtained by segmenting the question statement, and a list of second - language keywords obtained by segmenting the response statement; respectively verifying whether the first - language explanations corresponding to each keyword in the second - language keyword list match the meaning of any keyword in the question keyword list; if they match, determining that the spelling of the keyword in the second - language keyword list is correct, and if they do not match, checking whether there are spelling mistakes in the keyword in the second - language keyword list; determining the word - dimension score according to the spelling mistake situation of each keyword in the second - language keyword list.

[0094] In this embodiment, the question statement is segmented to obtain a list of question keywords. The response statement is segmented to obtain a list of second - language (i.e., the language used by the student in answering) keywords. These keywords reflect the vocabulary used by the student during the translation process. Further, each keyword in the second - language keyword list is traversed to find its corresponding first - language (i.e., the language used in the question statement) explanation. The found first - language explanation is matched with the keywords in the question keyword list to check whether the use of the keyword in the response statement is accurate. If the match is successful, it is considered that the spelling of the keyword in the response statement is correct. If the match is not successful, further check whether there are spelling mistakes in the keyword. For keywords that do not match and are determined to possibly have spelling mistakes, a count is made. The word - dimension score is determined according to the spelling mistake situation (the proportion of misspelled keywords) of each keyword in the second - language keyword list. Through this method, the accuracy and spelling of the words used by the student during the translation process can be evaluated more comprehensively and objectively.

[0095] In an embodiment of the present application, optionally, a fluency - dimension score is given to the response to the translation question, including: evaluating the coherence of the response statement through the text coherence evaluation tool, and determining the fluency - dimension score according to the coherence evaluation result, where the coherence evaluation includes sentence logical relationships and the use of conjunctions.

[0096] In this embodiment, a text coherence evaluation tool can be used. The student's answer statements are input into this evaluation tool for coherence evaluation. Specifically, the logical relationships between sentences can be evaluated, including whether the logical relationships between sentences in the answer statements are clear and reasonable. This includes judging causal, adversative, progressive, and other relationships between sentences. The use of conjunctive words can also be evaluated, including checking whether the conjunctive words in the answer statements are used appropriately and whether they can effectively connect the preceding and following sentences to make the overall text more coherent. According to the output results of the evaluation tool, the performance of the answer statements in terms of coherence is interpreted. Based on the coherence evaluation results, the score for the fluency dimension is determined.

[0097] Further, as Figure 1 a specific implementation of the method, an embodiment of the present application provides a correction device for translation questions, as Figure 4 shown. This device includes:

[0098] An acquisition module, configured to acquire the question statement of the translation question and the answer statement for translating the translation question;

[0099] A multi-dimensional scoring module, configured to perform multi-dimensional scoring on the answering situation of the translation question according to the question statement and the answer statement, where the scoring dimensions include at least two of the semantic dimension, the grammar dimension, the tone dimension, the word dimension, and the fluency dimension;

[0100] A comprehensive score statistics module, configured to perform comprehensive scoring calculation on the answer statement according to the multi-dimensional scoring to obtain the total answering score of the answer statement.

[0101] In an embodiment of the present application, optionally, as Figure 5 shown, the multi-dimensional scoring module includes: a semantic scoring unit, configured to:

[0102] Perform first-language keyword scoring and second-language keyword scoring according to the question statement in the first language and the answer statement in the second language;

[0103] Perform first-language keyword scoring and second-language keyword scoring according to the question statement in the first language and the answer statement in the second language;

[0104] If the first-language keyword scoring and the second-language keyword scoring meet the preset conditions, perform semantic similarity calculation on the question statement and the answer statement through a semantic similarity scoring model to determine the semantic dimension score; where the preset conditions are that at least one of the first-language keyword scoring is greater than the first scoring threshold and the second-language keyword scoring is greater than the second scoring threshold is satisfied;

[0105] If the first language keyword score and the second language keyword score do not meet the preset conditions, the semantic dimension score is determined by the semantic scoring model based on the title statement and the answer statement.

[0106] In an embodiment of the present application, optionally, the multi-dimensional scoring module includes: a query unit, configured to:

[0107] Determine whether the answer statement belongs to the correct answer statement corresponding to the title statement;

[0108] If not, perform a first language keyword score and a second language keyword score respectively based on the title statement in the first language and the answer statement in the second language;

[0109] If so, determine the total answer score of the answer statement according to the correct answer statement corresponding to the answer statement.

[0110] In an embodiment of the present application, optionally, the semantic scoring unit is configured to:

[0111] Translate each word in the answer statement into the first language to obtain a first language keyword list, or perform word segmentation on the answer statement in the first language after translating the answer statement into the first language to obtain a first language keyword list;

[0112] Perform word segmentation on the title statement to obtain a title keyword list;

[0113] Match each keyword in the title keyword list with the first language keyword list one by one, and determine the first language keyword score according to the matching situation of each keyword in the title keyword list and the first language keyword list.

[0114] In an embodiment of the present application, optionally, the semantic scoring unit is configured to:

[0115] Perform word segmentation on the answer statement to obtain a second language keyword list;

[0116] Perform word segmentation on the correct answer statement corresponding to the title statement to obtain an answer keyword list;

[0117] Match each keyword in the answer keyword list with the second language keyword list one by one, and determine the second language keyword score according to the matching situation of each keyword in the answer keyword list and the second language keyword list.

[0118] In an embodiment of the present application, optionally, the semantic scoring unit is configured to:

[0119] Obtain a first translation statement obtained by translating the answer statement into the first language;

[0120] Input the title statement, the answer statement, the correct answer statement corresponding to the title statement, and the first translation statement into a semantic similarity scoring model to obtain the semantic dimension score. Among them, the semantic similarity scoring model calculates the first semantic similarity between the title statement and the first translation statement, the second semantic similarity between the title statement and the answer statement, the third semantic similarity between the correct answer statement and the answer statement, and determines the semantic dimension score according to the first semantic similarity, the second semantic similarity, and the third semantic similarity.

[0121] In an embodiment of the present application, optionally, a semantic scoring unit is configured to:

[0122] If the semantic dimension score is less than the semantic scoring threshold, re-determine the semantic dimension score based on the title statement and the answer statement through a semantic scoring large model;

[0123] If the semantic dimension score is greater than or equal to the semantic scoring threshold, keep the semantic dimension score unchanged.

[0124] In an embodiment of the present application, optionally, a semantic scoring unit is configured to:

[0125] Obtain a second translation statement obtained by translating the answer statement into a first language;

[0126] Input the title statement, the answer statement, the correct answer statement corresponding to the title statement, and the second translation statement into a semantic scoring large model to obtain the semantic dimension score. Among them, the semantic scoring large model determines a first semantic score according to the answer statement and the correct answer statement, determines a second semantic score according to the second translation statement and the title statement, determines a third semantic score according to the answer statement and the title statement, and determines the semantic dimension score according to the first semantic score, the second semantic score, and the third semantic score.

[0127] In an embodiment of the present application, optionally, it further includes: a large model training module, configured to:

[0128] Construct an initial semantic scoring large model, where the initial semantic scoring large model includes an input layer, a main branch, a first side branch, a second side branch, a fully connected network, and an output layer. The input layer is respectively connected to the main branch, the first side branch, and the second side branch. The fully connected network is respectively connected to the main branch, the first side branch, and the second side branch. The output layer is connected to the fully connected network. The main branch is a network structure based on a large model;

[0129] Obtain training samples for the semantic scoring large model, where the training samples for the semantic scoring large model include question statement samples, answer statement samples, correct answer statement samples corresponding to the question statement samples, translation samples corresponding to the answer statement samples, and total annotation scores;

[0130] Input the question statement samples, the answer statement samples, the correct answer statement samples, and the translation samples into the input layer. Calculate the first predicted score between the question statement samples and the correct answer statement samples through the main branch, calculate the second predicted score between the translation samples and the question statement samples through the first side branch, calculate the third predicted score between the answer statement samples and the question statement samples through the second side branch, determine the total predicted score according to the first predicted score, the second predicted score, and the third predicted score through the fully connected network, and obtain the total predicted score output by the output layer;

[0131] Tune the parameters of the main branch, the first side branch, the second side branch, and the fully connected network according to the total predicted score and the total annotation score, and obtain the semantic scoring large model based on the input layer, the output layer, and the main branch, the first side branch, and the second side branch with tuned parameters.

[0132] In an embodiment of the present application, optionally, the large model training module is used for:

[0133] Obtain the question statement samples, the original answer statement samples, the correct answer statement samples, and the total annotation scores;

[0134] Select some words in the original answer statement samples, and replace the selected words with similar words to obtain the first simulated answer statement corresponding to the original answer statement samples, where the replacement of similar words includes at least one of synonym replacement, part-of-speech replacement, singular-plural replacement, preposition replacement, and replacement of words with similar spellings but different meanings;

[0135] Rewrite the grammar of the original answer statement samples to obtain the second simulated answer statement corresponding to the original answer statement samples, where the grammar rewrite includes at least one of tense rewrite, article rewrite, preposition rewrite, pronoun reference rewrite, sentence structure rewrite, negative form rewrite, singular-plural word rewrite, and fixed collocation rewrite;

[0136] Use the original answer statement samples, the first simulated answer statement, and the second simulated answer statement as the answer statement samples, and construct the training samples for the semantic scoring large model according to the question statement samples, the answer statement samples, the correct answer statement samples, the translation samples corresponding to the answer statement samples, and the total annotation scores.

[0137] In an embodiment of the present application, optionally, as Figure 5 shown, the multi-dimensional scoring module includes: a grammar scoring module, configured to:

[0138] Perform a grammar dimension score on the answer to the translation question, including: inputting the answer statement into a grammar analyzer, analyzing the grammar problems existing in the answer statement through the grammar analyzer, and determining the grammar dimension score according to the grammar problems.

[0139] In an embodiment of the present application, optionally, as Figure 5 shown, the multi-dimensional scoring module includes: a tone scoring module, configured to:

[0140] Perform a tone dimension score on the answer to the translation question, including: performing tone analysis on the question statement and the answer statement respectively, and determining the tone dimension score based on the difference degree between the tone of the question statement and the tone of the answer statement.

[0141] In an embodiment of the present application, optionally, as Figure 5 shown, the multi-dimensional scoring module includes: a word scoring module, configured to:

[0142] Perform a word dimension score on the answer to the translation question, including: obtaining a list of question keywords obtained by performing word segmentation on the question statement and a list of second-language keywords obtained by performing word segmentation on the answer statement; respectively verifying whether the first-language explanations corresponding to the keywords in the second-language keyword list match the meaning of any keyword in the question keyword list; if they match, determining that the keywords in the second-language keyword list are spelled correctly, and if they do not match, checking whether there are spelling mistakes in the keywords in the second-language keyword list; determining the word dimension score according to the spelling mistake situation of each keyword in the second-language keyword list.

[0143] In an embodiment of the present application, optionally, as Figure 5 shown, the multi-dimensional scoring module includes: a fluency scoring module, configured to:

[0144] Perform a fluency dimension score on the answer to the translation question, including: performing coherence evaluation on the answer statement through the text coherence evaluation tool, and determining the fluency dimension score according to the coherence evaluation result, where the coherence evaluation includes sentence logical relationship and the use of connecting words.

[0145] It should be noted that for other corresponding descriptions of each functional unit involved in a translation question marking device provided in an embodiment of the present application, reference can be made to Figures 1 to 3CThe corresponding description in the method will not be elaborated here.

[0146] The embodiments of the present application also provide a computer device, which can specifically be a personal computer, a server, a network device, etc. The computer device includes a bus, a processor, a memory, and a communication interface, and may also include an input / output interface and a display device. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store location information. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements the steps in the method embodiments.

[0147] Those skilled in the art can understand that the structure of the above computer device is only a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components, or combine certain components, or have different component arrangements.

[0148] In one embodiment, a computer-readable storage medium is provided. The computer-readable storage medium can be non-volatile or volatile, and stores a computer program. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0149] In one embodiment, a computer program product is provided, including a computer program. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0150] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data that have been authorized by the user or fully authorized by all parties.

[0151] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, graphics processors, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.

[0152] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0153] The above-described embodiments merely represent several implementation manners of the present application. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for grading translation questions, characterized in that, The method includes: Obtaining the topic statement of the translation topic and the answer statement for translating the translation topic; According to the topic statement and the answer statement, performing multi-dimensional scoring on the answering situation of the translation topic, where the scoring dimensions include at least two of the semantic dimension, the grammar dimension, the tone dimension, the word dimension, and the fluency dimension; Performing comprehensive scoring calculation on the answer statement according to the multi-dimensional scoring to obtain the total answer score of the answer statement.

2. The method according to claim 1, characterized in that, Performing semantic dimension scoring on the answering situation of the translation topic according to the topic statement in the first language and the answer statement in the second language, including: Performing first language keyword scoring and second language keyword scoring respectively according to the topic statement in the first language and the answer statement in the second language; If the first language keyword score and the second language keyword score meet the preset conditions, then performing semantic similarity calculation on the topic statement and the answer statement through a semantic similarity scoring model to determine the semantic dimension score; where the preset conditions are that at least one of the first language keyword score is greater than the first scoring threshold and the second language keyword score is greater than the second scoring threshold is satisfied; If the first language keyword score and the second language keyword score do not meet the preset conditions, then determining the semantic dimension score according to the topic statement and the answer statement by a semantic scoring large model.

3. The method according to claim 2, wherein Before performing first language keyword scoring and second language keyword scoring respectively according to the topic statement in the first language and the answer statement in the second language, it further includes: Judging whether the answer statement belongs to the correct answer statement corresponding to the topic statement; If not, performing first language keyword scoring and second language keyword scoring respectively according to the topic statement in the first language and the answer statement in the second language; If so, determining the total answer score of the answer statement according to the correct answer statement corresponding to the answer statement.

4. The method according to claim 2, wherein Performing first language keyword scoring according to the topic statement in the first language and the answer statement in the second language, including: Translating each word in the answer statement into the first language to obtain a first language keyword list, or performing word segmentation on the answer statement in the first language after translating the answer statement into the first language to obtain a first language keyword list; Performing word segmentation on the topic statement to obtain a topic keyword list; Matching each keyword in the topic keyword list with the first language keyword list one by one, and determining the first language keyword score according to the matching situation of each keyword in the topic keyword list and the first language keyword list.

5. The method according to claim 2, wherein Performing second language keyword scoring according to the topic statement in the first language and the answer statement in the second language, including: Performing word segmentation on the answer statement to obtain a second language keyword list; Performing word segmentation on the correct answer statement corresponding to the topic statement to obtain an answer keyword list; Match each keyword in the answer keyword list with the second language keyword list one by one, and determine the second language keyword score according to the matching situation of each keyword in the answer keyword list and the second language keyword list.

6. The method according to claim 2, characterized in that Calculate the semantic similarity between the question statement and the answer statement through a semantic similarity scoring model to determine the semantic dimension score, including: Obtain the first translation statement obtained by translating the answer statement into the first language; Input the question statement, the answer statement, the correct answer statement corresponding to the question statement, and the first translation statement into the semantic similarity scoring model to obtain the semantic dimension score. Among them, the semantic similarity scoring model calculates the first semantic similarity between the question statement and the first translation statement, the second semantic similarity between the question statement and the answer statement, the third semantic similarity between the correct answer statement and the answer statement, and determines the semantic dimension score according to the first semantic similarity, the second semantic similarity and the third semantic similarity.

7. The method according to claim 6, characterized in that, After determining the semantic dimension score, the method further includes: If the semantic dimension score is less than the semantic score threshold, re-determine the semantic dimension score according to the question statement and the answer statement through the semantic scoring large model; If the semantic dimension score is greater than or equal to the semantic score threshold, keep the semantic dimension score unchanged.

8. The method according to claim 2, wherein Determine the semantic dimension score according to the question statement and the answer statement through the semantic scoring large model, including: Obtain the second translation statement obtained by translating the answer statement into the first language; Input the question statement, the answer statement, the correct answer statement corresponding to the question statement, and the second translation statement into the semantic scoring large model to obtain the semantic dimension score. Among them, the semantic scoring large model determines the first semantic score according to the answer statement and the correct answer statement, determines the second semantic score according to the second translation statement and the question statement, determines the third semantic score according to the answer statement and the question statement, and determines the semantic dimension score according to the first semantic score, the second semantic score and the third semantic score.

9. The method according to claim 8, wherein Before inputting the question statement, the answer statement, the correct answer statement corresponding to the question statement, and the second translation statement into the semantic scoring large model, it further includes: Construct an initial semantic scoring large model, where the initial semantic scoring large model includes an input layer, a main branch, a first side branch, a second side branch, a fully connected network and an output layer. The input layer is respectively connected to the main branch, the first side branch and the second side branch. The fully connected network is respectively connected to the main branch, the first side branch and the second side branch. The output layer is connected to the fully connected network. The main branch is a network structure based on a large model. Obtain training samples for the semantic scoring large model, where the training samples for the semantic scoring large model include question statement samples, answer statement samples, correct answer statement samples corresponding to the question statement samples, translation samples corresponding to the answer statement samples, and total annotation scores; Input the question statement samples, the answer statement samples, the correct answer statement samples, and the translation samples into the input layer. Calculate the first predicted score between the question statement samples and the correct answer statement samples through the main branch, calculate the second predicted score between the translation samples and the question statement samples through the first side branch, calculate the third predicted score between the answer statement samples and the question statement samples through the second side branch, determine the total predicted score according to the first predicted score, the second predicted score, and the third predicted score through the fully connected network, and obtain the total predicted score output by the output layer; Optimize the parameters of the main branch, the first side branch, the second side branch, and the fully connected network according to the total predicted score and the total annotation score, and obtain the semantic scoring large model based on the input layer, the output layer, and the main branch, the first side branch, and the second side branch with optimized parameters.

10. The method according to claim 9, characterized in that, Before obtaining the training samples for the semantic scoring large model, it further includes: Obtain the question statement samples, the original answer statement samples, the correct answer statement samples, and the total annotation scores; Select some words in the original answer statement samples, and replace the selected words with similar words to obtain the first simulated answer statement corresponding to the original answer statement samples, where the similar word replacement includes at least one of synonym replacement, part-of-speech replacement, singular-plural replacement, preposition replacement, and replacement of words with similar spellings but different meanings; Rewrite the grammar of the original answer statement samples to obtain the second simulated answer statement corresponding to the original answer statement samples, where the grammar rewrite includes at least one of tense rewrite, article rewrite, preposition rewrite, pronoun reference rewrite, sentence structure rewrite, negative form rewrite, singular-plural word rewrite, and fixed collocation rewrite; Use the original answer statement samples, the first simulated answer statement, and the second simulated answer statement as the answer statement samples, and construct the training samples for the semantic scoring large model according to the question statement samples, the answer statement samples, the correct answer statement samples, the translation samples corresponding to the answer statement samples, and the total annotation scores.

11. The method according to any one of claims 1 to 10, characterized in that Perform a grammar dimension score on the answering situation of the translation questions, including: input the answer statement into a grammar analyzer, analyze the grammar problems existing in the answer statement through the grammar analyzer, and determine the grammar dimension score according to the grammar problems; and / or Perform a tone dimension scoring on the answering situation of the translation question, including: respectively perform tone analysis on the question statement and the answering statement, and determine the tone dimension scoring based on the difference degree between the tone of the question statement and the tone of the answering statement; and / or, Perform a word dimension scoring on the answering situation of the translation question, including: obtain a list of question keywords obtained by performing word segmentation on the question statement, and a list of second-language keywords obtained by performing word segmentation on the answering statement; respectively verify whether the first-language explanations corresponding to the keywords in the second-language keyword list match the meaning of any keyword in the question keyword list; if they match, determine that the spelling of the keyword in the second-language keyword list is correct, if they do not match, check whether there are spelling mistakes in the keywords in the second-language keyword list; determine the word dimension scoring according to the spelling mistake situation of each keyword in the second-language keyword list; and / or, Perform a fluency dimension scoring on the answering situation of the translation question, including: perform a coherence evaluation on the answering statement through the text coherence evaluation tool, and determine the fluency dimension scoring according to the coherence evaluation result, where the coherence evaluation includes sentence logical relationship and the use of connecting words.

12. A correction device for translation questions, characterized in that The device includes: An acquisition module, configured to acquire a question statement of a translation question and an answering statement for translating the translation question; A multi-dimension scoring module, configured to perform multi-dimension scoring on the answering situation of the translation question according to the question statement and the answering statement, where the scoring dimensions include at least two of a semantic dimension, a grammar dimension, a tone dimension, a word dimension, and a fluency dimension; A comprehensive score statistics module, configured to perform a comprehensive score calculation on the answering statement according to the multi-dimension scoring to obtain the total answering score of the answering statement.

13. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 11.

14. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 11.

Images