Language training method and device, equipment and medium

Abstract

The invention belongs to the technical field of language learning, and discloses a language training method and device, equipment and a medium, and the method comprises the steps: constructing a candidate training unit pool; calculating a push priority score of each training unit in the candidate training unit pool; sorting all the training units in the candidate training unit pool from large to small according to the push priority scores, and selecting a preset number of training units; aiming at each selected training unit, adopting an interference strategy which corresponds to the training unit and contains grammar errors and / or cultural conflicts, generating training questions corresponding to the training unit based on a large language model, and forming a training question set by all the training questions; and performing language training according to the training question set. And the language anti-interference capability, the cross-culture sensitivity and the deep grammar grammar of the user can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of language learning, and in particular relates to a language training method, device, equipment and medium. Background Art

[0002] In language learning, effective communication relies not only on accurate vocabulary and grammar, but also on the appropriate use of language in a specific cultural context. However, in real-world interactions, language interference and cultural misunderstandings are key factors leading to communication failures.

[0003] However, the current core strategies for language learning often focus on "positive memory + repeated training", that is, reinforcing correct language forms through a large amount of input and imitation exercises, but fail to effectively integrate error recognition and cultural sensitivity, and the training effect still needs to be improved. Summary of the Invention

[0004] The object of the present invention is to provide a language training method, apparatus, device and storage medium that can improve language training effects by utilizing error recognition and cultural sensitivity.

[0005] A first aspect of the present invention discloses a language training method, comprising:

[0006] Build a pool of candidate training units;

[0007] Calculate the push priority score for each training unit in the candidate training unit pool;

[0008] Sort all training units in the candidate training unit pool according to their push priority scores from largest to smallest, and select a preset number of training units;

[0009] For each selected training unit, an interference strategy containing grammatical errors and / or cultural conflicts corresponding to the training unit is adopted, and training questions corresponding to the training unit are generated based on the large language model. All training questions constitute a training question set;

[0010] Language training is performed according to the training question set.

[0011] In some embodiments, calculating a push priority score for a training unit includes:

[0012] For the training unit, obtain the user's historical error records and the user's mastery status, and calculate the forgetting probability;

[0013] The push priority score is calculated according to the historical error records, the user's personalized subject preference, the mastery status, and the forgetting probability.

[0014] In some embodiments, calculating the forgetting probability includes:

[0015] The forgetting probability is calculated according to the time difference between the time when the user last correctly memorized the training unit and the current time and the user's forgetting rate.

[0016] In some embodiments, the employing of an interference strategy containing grammatical errors and / or cultural conflicts corresponding to the training unit, and generating training questions corresponding to the training unit based on the large language model, includes:

[0017] Determining the interference strategy based on the historical error records, the user's personalized subject preference, and the forgetting probability;

[0018] Obtaining an error template according to the interference strategy;

[0019] constructing a prompt word according to the training unit and the error template;

[0020] Based on the prompt words, a large language model is used to generate training questions corresponding to the training units.

[0021] In some embodiments, determining the interference strategy based on the historical error records, the user's personalized subject preference, and the forgetting probability includes:

[0022] When the number of historical error records is greater than a first preset threshold, setting the interference strategy to include a basic grammatical error strategy, wherein the basic grammatical error strategy is used to generate grammatical error questions;

[0023] When the personalized topic preference is greater than a second preset threshold, setting the interference strategy to include a cultural expression error example strategy, wherein the cultural expression error example strategy is used to generate a cultural expression error question;

[0024] When the forgetting probability is greater than a third preset threshold, the interference strategy is set to include a real error strategy, and the real error strategy is used to generate real wrong questions for the user.

[0025] In some embodiments, generating a training question corresponding to a training unit using a large language model based on the prompt word includes:

[0026] Calculate the error risk level of the training unit;

[0027] A large language model is used to generate training questions corresponding to the training unit according to the prompt word, the subject of the training unit and the error risk level.

[0028] In some embodiments, calculating the error risk level of the training unit includes:

[0029] Calculating a historical error rate based on the historical error records;

[0030] The error risk level is calculated according to the historical error rate and the forgetting probability.

[0031] A second aspect of the present invention discloses a language training device, comprising:

[0032] A candidate training unit pool construction module is used to construct a candidate training unit pool;

[0033] Scoring module, used to calculate the push priority score of each training unit in the candidate training unit pool;

[0034] The training question generation module is used to sort all training units in the candidate training unit pool from highest to lowest according to the push priority score, select a preset number of training units, and for each selected training unit, adopt the interference strategy containing grammatical errors and / or cultural conflicts corresponding to the training unit, and generate training questions corresponding to the training unit based on the large language model. All training questions constitute the training question set;

[0035] The training module is used to perform language training based on the training question set.

[0036] The third aspect of the present invention discloses an electronic device, comprising a memory storing executable program code and a processor coupled to the memory; the processor calls the executable program code stored in the memory to execute the language training method disclosed in the first aspect.

[0037] A fourth aspect of the present invention discloses a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the language training method disclosed in the first aspect.

[0038] The beneficial effect of the present invention is that, first, the push priority score of each training unit is calculated to determine the pushed training unit, and then, for the pushed training unit, training questions containing grammatical errors and / or cultural conflicts are generated, thereby constructing a training question set with more practical, error-correcting and memory-enhancing effects, breaking the traditional "positive memory + mechanical repetition" model, and adopting the "negative interference + erroneous reasoning" mechanism to effectively improve the user's language anti-interference ability, cross-cultural sensitivity and deep grammatical mastery. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings herein illustrate specific examples of the technical solutions described in the present invention, and together with the specific implementation methods constitute a part of the specification, and are used to explain the technical solutions, principles and effects of the present invention.

[0040] Unless otherwise specified or defined, the same reference numerals in different drawings represent the same or similar technical features, and the same or similar technical features may also be represented by different reference numerals.

[0041] Figure 1is a flow chart of a language training method disclosed in an embodiment of the present invention;

[0042] Figure 2 is a flowchart of calculating a push priority score according to an embodiment of the present invention;

[0043] Figure 3 is a flow chart of generating training questions according to an embodiment of the present invention;

[0044] Figure 4 is a flow chart of determining an interference strategy according to an embodiment of the present invention;

[0045] Figure 5 1 is a schematic structural diagram of a language training device according to an embodiment of the present invention;

[0046] Figure 6 It is a structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] Unless otherwise specified or defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art. In the context of combining the technical solution of the present invention with realistic scenarios, all technical and scientific terms used herein may also have meanings corresponding to the purpose of implementing the technical solution of the present invention. "First, second..." used herein is merely used to distinguish names and does not represent a specific quantity or order. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0048] It should be noted that when an element is considered to be "fixed to" another element, it can be directly fixed to the other element or there can be an intermediate element; when an element is considered to be "connected to" another element, it can be directly connected to the other element or there can be an intermediate element; when an element is considered to be "mounted on" another element, it can be directly mounted on the other element or there can be an intermediate element. When an element is considered to be "located on" another element, it can be directly located on the other element or there can be an intermediate element.

[0049] Unless otherwise specified or defined, the “said” and “the” used in this document refer to the technical features or technical contents mentioned or described before the corresponding position, and the technical features or technical contents may be the same as or similar to the technical features or technical contents mentioned therein. In addition, the terms “including” and “having” and any variations thereof used in this document are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or devices.

[0050] The present invention provides a language training method that promotes learning through mistakes, breaking the traditional "positive memory + mechanical repetition" model and adopting a "negative interference + incorrect reasoning" mechanism. By actively generating language expressions containing grammatical errors and / or cultural conflicts (i.e., incorrect reasoning), the method guides users to deduce correct rules (i.e., negative interference) in the process of "reverse error correction", thereby effectively improving users' language interference resistance, cross-cultural sensitivity and deep grammatical mastery.

[0051] To facilitate understanding of the present invention, specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings.

[0052] The present invention discloses a language training method for English learning that can be implemented through computer programming as a standalone language training system. The method can be implemented by an electronic device such as a computer, laptop, or tablet computer, or a control chip embedded in an electronic device, but the present invention is not limited thereto. It should be noted that the method is not limited to English training and can also be used for training in other languages.

[0053] like Figure 1 As shown, the method includes the following steps:

[0054] Step S100: constructing a candidate training unit pool;

[0055] The candidate training unit pool is a user-specific, personalized sentence pool, consisting of candidate words and / or sentences. Specifically, a training unit refers to the smallest linguistic unit used for push training. It broadly refers to all content units that can be used for push training, including not only words but also short sentences and grammatical structures. It includes the following types: 1. Words: such as "abandon" and "sustainable" in vocabulary; 2. Phrases or collocations: such as "go through the roof" and "as far as I know"; 3. Complete sentences: used for grammatical structures, listening and speaking training, or situational expression; 4. Questions: such as listening comprehension questions, cloze tests, and situational expression questions (sentences + distractors); 5. Semantic chunks: specifically, collections of expressions closely related to a skill point (such as "concession transition" and "emotional expression").

[0056] The candidate training unit pool can be constructed in the following ways: 1. Users actively add words and sentences, that is, users enter keywords, topic tags, or add historical wrong questions with one click to include words and sentences that they are interested in or have not yet mastered into the candidate training unit pool. For example, after entering keywords such as "study abroad oral English", "frequent interview questions", and "Chinese food menu", relevant words and sentences are automatically extracted and annotated with meta-information such as source, theme, context tags, etc., and added to the candidate training unit pool. 2. Wrong question set backtracking, that is, words and sentences that users have not fully mastered in historical training (wrong questions or repeatedly wrong items) will be included in the wrong question set, and the wrong question set will be used as a candidate training unit pool. 3. Fusion of high-frequency interference vocabulary, that is, based on the built-in culturally sensitive expression vocabulary and the corpus statistics mined by the NLP model (natural language processing model), common language misunderstandings, high-frequency incorrect expressions, and "easy-to-error points" in cross-cultural communication are collected to recommend expressions that users may not have noticed but are prone to errors in actual applications. For example, the following common grammatical difficulties: "advices" (uncountable error), "He don't know." (subject-verb inconsistency), "discuss about" (redundant preposition); the following Chinese expressions: "I'm convenient today." (literal translation error), "You are wrong." (too direct and rude); the following cultural taboos / inappropriate expressions: "How much do you earn?" (it is not appropriate to ask directly in Western culture).

[0057] Step S200: Calculate the push priority score of each training unit in the candidate training unit pool;

[0058] This embodiment designs an adversarial intelligent push model, in which the push priority score of each training unit in the candidate training unit pool is calculated to determine whether it should be pushed to the user first in the current training cycle.

[0059] The following key factors are considered when calculating the push priority score: 1. The user's historical error record (wrong words / sentences, distractor click history, etc.); 2. The user's mastery level; 3. Personalized topic preferences (such as "travel," "workplace," and "American drama" scenes); and 4. The probability of forgetting the training unit. A comprehensive recommendation score is calculated based on these multiple dimensions.

[0060] Error records are the user's history of incorrect answers for a training unit. For example, incorrectly answering a word three times or clicking on the wrong distractor in a question can signal "short-term forgetting / misjudgment."

[0061] Mastery status is used to comprehensively assess whether a user has truly mastered a particular knowledge point. It is related to factors such as the number of errors, the number of correct answers after review, and the time interval. For example, if a user missed a word but answered it correctly three times in the past, this is the key basis for determining whether further reinforcement is needed.

[0062] The forgetting probability of a training unit refers to the possibility that the user has forgotten the training unit at the current moment.

[0063] In some embodiments, historical learning paths can also be considered when calculating push priority scores. Historical learning paths refer to the user's past learning behavior sequence, including multi-dimensional information such as time, content, and performance, which is used to predict their cognitive evolution trends. For example: 1. Training content records: words, sentences, questions, and grammar points that the user has learned in the past; 2. Performance data: whether the answer is correct, incorrect, or whether the wrong interference item is clicked; 3. Timeline data: the time interval between the last exposure to the knowledge point; 4. Growth path: the process of a certain type of vocabulary from not knowing → partially mastering → proficiently mastering. Example of historical learning path: User A repeatedly made mistakes in "by accident" and "on purpose" during systematic learning in March. His performance gradually improved in April and entered a stable mastery stage in May.

[0064] like Figure 2 As shown, the specific steps for calculating the push priority score in this embodiment are:

[0065] Step S210: For the training unit, obtain the user's historical error records and the user's mastery status, and calculate the forgetting probability;

[0066] The historical training records store the user's training records for each training unit. First, retrieve the training records corresponding to each training unit from the historical training records to obtain filtered training records. Based on these filtered training records, calculate the error rate to obtain the user's historical error record. Then, analyze these filtered training records to determine the user's number of errors, number of correct answers after review, and time interval between correct answers, thereby comprehensively assessing the user's mastery status. The forgetting probability is also calculated based on these filtered training records. The historical error record, mastery status, and forgetting probability are all numerical values ​​between 0 and 1.

[0067] Specifically, historical error records are represented by the historical error rate ErrRate(w,u): ErrRate(w,u) = num_errors / num_attempts, where w is the training unit, u is the user ID, num_errors is the number of errors, and num_attempts is the number of training attempts. For example, 0.6 indicates a 60% error rate for that word. The historical error rate is updated dynamically with each test. Mastery status can be determined based on pre-set rules. For example, if three correct answers were given within the past month, the mastery status is determined to be 0.8.

[0068] The forgetting probability is estimated by using the memory decay prediction model based on the time difference between the last time the user correctly memorized the training unit and the current time and the user's forgetting rate. The forgetting probability calculation formula is:

[0069]

[0070] Where w represents a training unit; u represents the user ID; t represents the timestamp of the current time converted to a value calculated by counting hours since January 1, 1970, such as 4000 hours; t_last(w,u) represents the timestamp of the last time user u correctly memorized (answered) a training unit, and this timestamp is converted to a value calculated by counting hours since January 1, 1970; λ u This parameter is a personalized parameter for the "user forgetting rate," with a value between 0 and 1, indicating how quickly a user forgets vocabulary. Initially, users can set this parameter based on whether they forget quickly or slowly. A larger value indicates faster forgetting. The default value is 0.3.

[0071] Step S220: Calculate the push priority score based on historical error records, the user's personalized topic preferences, mastery status, and forgetting probability.

[0072] The push priority score PushScore(w,u,t) is calculated by taking a weighted average of historical error records, the user's personalized topic preferences, mastery status, and forgetting probability.

[0073] The specific expression is:

[0074] PushScore(w,u,t)=α1*ForgetProb(w,u,t)+α2*ThemeMatch(w,u)

[0075] +α3*ErrRate(w,u)+α4*Newness(w,u).

[0076] ErrRate(w,u) represents the user's historical error rate for a particular word w, representing a historical error record. Newness(w,u) indicates whether the user has first encountered the word or phrase, or whether it appears in the learning pool for the first time. 1 = completely new, 0.5 = recently learned, and 0 = old. Together, Newness(w,u) and ErrRate(w,u) represent mastery status. Newness(w,u) can be designed as a discrete value or a time-decaying function (with new words receiving a higher weight). ThemeMatch(w,u) indicates the similarity (0-1) between the current word or phrase and the user's interests, indicating whether the current word or phrase belongs to the user's preferred topic, representing the user's personalized topic preference. This value is calculated based on the cosine similarity of keyword-topic vectors and ranges from 0 to 1. If the user prefers "study abroad in English," the matching value for this word is high. Coefficients α1 (forgetting priority), α2 (error priority), α3 (new word stimulation), and α4 (interest-related) determine the recommendation logic and can be adjusted based on user needs. The initial values ​​are: α1=0.4, α2=0.3, α3=0.2, α4=0.1.

[0077] For example:

[0078]

[0079] As can be seen from the above table, although the forgetting probability of both is high, the error rate of "boarding pass" is high and the interest is more matched, so it is pushed first.

[0080] Step S300: sorting all training units in the candidate training unit pool from largest to smallest according to the push priority scores, and selecting a preset number of training units;

[0081] All training units in the candidate training unit pool are sorted from highest to lowest based on the calculated push priority scores. The top N (e.g., N = 10 or 20) ranked training units are selected and pushed as the training units for that day or period. The N value can be dynamically adjusted based on user configuration. For example, if a user is highly active or answers questions quickly, the number of pushed questions can be automatically increased.

[0082] Step S400: For each selected training unit, employing an interference strategy containing grammatical errors and / or cultural conflicts corresponding to the training unit, generating training questions corresponding to the training unit based on the large language model, and all training questions constitute a training question set;

[0083] After receiving the pushed training units, the corresponding interference strategy is determined for each training unit, and training questions are generated using the large language model based on the interference strategy. The interference strategy, also known as the generation strategy, is a set of purposeful methods designed to "create wrong questions" based on the user's current cognitive state, with the goal of enhancing memory intervention and language judgment. By constructing a "high-interference, low-repetitive" wrong question environment in learning tasks, the following goals are achieved: 1. Use countermeasures to stimulate attention and reflection, avoiding rote memorization. 2. Improve the ability to identify and correct errors, strengthen the memory of the "error recognition-correction" path, and reduce the probability of making errors next time.

[0084] In this embodiment, for each recommended training unit, an adversarial error generator (A-EG) is used to dynamically configure interference strategies for each word or sentence, generating personalized training questions with high interference and high variance. For example, each word or sentence can generate 2 to 3 different interference questions to avoid mechanical repetition and enhance the anti-interference memory effect. The adversarial error generator is a large language model that has been trained with corpus data such as grammatically incorrect sentences and cultural conflict sentences, such as the chatGPT model and the DeepSeek model.

[0085] like Figure 3 As shown, the specific steps of generating training questions in this embodiment include:

[0086] Step S410: Determine an interference strategy based on historical error records, the user's personalized topic preferences, and the probability of forgetting;

[0087] Historical error records, users' personalized topic preferences, and forgetfulness probability are factors in calculating push priority scores. The scores of these factors also represent the degree of influence on the recommended training unit. Therefore, by analyzing these factors, we derive the appropriate interference strategy, ensuring that the interference strategy is synchronized with the reasons for recommending the training unit.

[0088] The specific steps of this embodiment are as follows: Figure 4 Shown, including:

[0089] Step S411: When the number of historical error records is greater than a first preset threshold, setting an interference strategy including a basic grammatical error strategy, wherein the basic grammatical error strategy is used to generate grammatical error questions;

[0090] Historical error records are represented by ErrRate(w,u). When ErrRate(w,u) is greater than the first preset threshold, it means that the user often makes mistakes in the word / sentence. At this time, the recommended interference strategy is: basic grammatical error example strategy, which uses the basic grammatical error example strategy to generate grammatical error questions and reproduce typical errors.

[0091] Step S412: When the personalized topic preference is greater than a second preset threshold, setting an interference strategy to include a cultural expression error example strategy, wherein the cultural expression error example strategy is used to generate cultural expression error questions;

[0092] The personalized theme is represented by ThemeMatch(w,u). When ThemeMatch(w,u) is greater than the second preset threshold, it indicates that the training unit belongs to the user's preferred theme. In this case, the recommended interference strategy is the Cultural Expression Error Example Strategy. By using the Cultural Expression Error Example Strategy to generate cultural expression error questions, it is closer to the actual context and improves the realism of the interference.

[0093] Step S413: When the forgetting probability is greater than the third preset threshold, the interference strategy is set to include a real error strategy, wherein the real error strategy is used to generate real wrong questions of the user.

[0094] The forgetting probability is represented by ForgetProb(w,u,t). When ForgetProb(w,u,t) is greater than the third preset threshold, it indicates that the user forgets quickly. In this case, the recommended interference strategy is the real error example strategy, which uses the real error example strategy to generate real user errors and strengthen the memory of existing errors.

[0095] For example, if ErrRate=0.8 and ForgetProb=0.6 of a certain training unit, the interference strategies to be determined include: basic grammatical error strategy and real error strategy.

[0096] In some embodiments, Newness(w,u) can also be used to determine the interference strategy. If Newness(w,u) exceeds a threshold, it indicates a new word. The recommended interference strategy is: a spelling error strategy, which uses spelling analysis or word order errors to increase the intensity of the first impression interference.

[0097] Step S420: Obtain an error template according to the interference strategy;

[0098] The error template is a prompt word template used for generating training questions for a large language model. The error template adapted to the interference strategy is retrieved from the error generation template library. Different interference strategies correspond to different error templates.

[0099] Specifically, the error templates for the basic grammatical error strategy are primarily used to simulate grammatical errors, such as altered tenses, subject-verb inconsistency, and incorrect preposition and article usage, generating incorrect sentences like "He buys an apple." The error templates for the cultural expression error strategy are primarily used to simulate cultural expression errors, such as imitating Chinese expressions (e.g., "Long time no see!"), impolite expressions, and illogical expressions. They generate incorrect sentences like "You go first time?" or transposing word order to construct English sentences that conform to Chinese logic but are incoherent, generating incorrect sentences like "I only eat vegetables can." The real error strategy simulates real errors, aggregating real user errors and selecting typical sentences through clustering or similarity. These real errors are used as interference options to enhance personalized training effectiveness, generating incorrect sentences like "He is married with her." The spelling error strategy simulates spelling / lexical errors, such as using similar spellings or similar words, for example, "affect" → "effect."

[0100] Step S430: constructing prompt words based on the training unit and the error template;

[0101] Combine the training unit and the error template to construct a prompt word. For example, the prompt word format is "Please generate three interference sentences that meet the following templates for sentence X. Error template 1: simulates Chinese expression. Error template 2: constructs English that conforms to Chinese logic but is not fluent."

[0102] Taking the training unit "She went to the park yesterday." as an example, the prompt is: Please generate three grammatically incorrect sentences for the sentence "She went to the park yesterday." Based on the following error template, the error template changes the past tense to the present tense, making the verb and the time adverbial inconsistent.

[0103] Step S440: Based on the prompt word, a large language model is used to generate training questions corresponding to the training unit.

[0104] After obtaining the prompt word, the prompt word is input into the large language model to generate training questions corresponding to the training unit. The training questions can be of various forms, including: multiple-choice questions (for identifying incorrect sentences), error correction questions (for free correction of errors), word selection questions (for testing grammar and semantics), and scenario simulation questions (for multiple rounds of incorrect sentences).

[0105] When this embodiment uses a large language model to generate training questions, it also calculates the error risk level of each training unit. Then, the prompt word, the prompt word corresponding to the topic of the training unit, and the prompt word corresponding to the error risk level are input into the large language model to generate training questions corresponding to the training unit. The generated personalized training questions have higher interference and higher differentiation.

[0106] The error risk level can be used to indicate the degree of error visibility in the generated training questions. The process of calculating the error risk level is as follows: first, the historical error rate is calculated based on the historical error records; then, based on the historical error rate and the forgetting probability, the error risk level RiskLevel(w,u) is calculated. The specific expression is:

[0107] RiskLevel(w,u)=λ1*ErrRate(w,u)+λ2*ForgetProb(w,u,t),

[0108] Where ErrRate(w,u) is the user's historical error rate for training unit w, ForgetProb(w,u,t) is the forgetting probability, and λ1 and λ2 are adjustable coefficients (e.g., λ1 = 0.6, λ2 = 0.4).

[0109] In this example, a RiskLevel(w,u) ≥ 0.7 indicates high risk, with the corresponding prompt being "The error is obvious." A RiskLevel(w,u) between 0.4 and 0.7 indicates medium risk, with the corresponding prompt being "The error is not easy to spot." A RiskLevel(w,u) < 0.4 indicates low risk, with the corresponding prompt being "The error is difficult to spot." Prompts that take error risk levels into account might look like: "Please generate three distractor sentences for sentence X that conform to the following template. Each error must be subtle but grammatically correct."

[0110] Step S500: Perform language training according to the training question set.

[0111] After the training set is generated, training tasks are generated based on the training questions in the set to conduct language training. Training tasks can be generated periodically or triggered based on user interaction, ensuring that the training rhythm matches the user's cognitive state, further improving memory retention and language interference resistance.

[0112] This embodiment also adopts differentiated push strategies for language training based on the source of training units and user learning preferences to maximize learning effects and user acceptance. There are two push methods: 1. When the user actively clicks "Today's Practice", the pushed phrase type is active learning phrase (ThemeMatch(w,u)=1), and the phrase source is the phrase or expression actively added or marked by the user. The push method is: no intervention, the user independently views and completes the practice through the "learning pool", there is no cycle limit, and the user can practice at will. 2. Active push based on forgetting prediction (ForgetProb>θ, θ is a threshold). That is, when the ForgetProb of a certain word is greater than θ, a training question pops up according to the learning time period set by the user. The pushed phrase type is passive recommendation phrase (ThemeMatch(w,u)<1), and the phrase source is the training question set. The push method is: regular reminder to complete the practice, combined with the forgetting probability ForgetProb(w,u,t) to control the recommendation frequency, and automatically add the forgetting probability above the threshold (such as >0.6) to the today's task. If it is below the threshold, push will be suspended to avoid learning interference and resource waste.

[0113] This embodiment also performs a structured analysis of the user's feedback data from the previous round of training, updates their personal memory model in real time, and provides a basis for more accurate content recommendations in the next round, thereby achieving highly adaptable and efficient dynamic training path adjustments. Specifically, key behavioral data for each training question is recorded, including but not limited to: answer result (correct / incorrect), answer time, number of prompt clicks, question redo frequency, and user preference for different question types. These behavioral data are used to update the following logic:

[0114]

[0115] The forgetting probability is then recalculated, the memory status of each training unit is dynamically evaluated, and used to update the push priority score of the next round of training tasks, further driving the sorting and personalized recommendation of training content.

[0116] In summary, the language training method of this embodiment combines the user's personalized memory status, wrong question record, topic preference and forgetting cycle, calculates the push priority score for each training unit such as a word or sentence, and dynamically determines the push priority score. In terms of training methods, it emphasizes replacing "rote memorization" with "reverse reasoning and error correction", integrating high-interference training mechanisms such as cultural misuse interference and common grammatical illusions, and constructing a more practical, error-correcting and memory-enhancing training question set to replace the traditional inefficient method of "blind review according to a time plan". Through "adversarial interference + erroneous reasoning + dynamic memory evaluation", a personalized, efficient and practical-oriented English learning experience is achieved.

[0117] like Figure 5As shown, based on the above-mentioned language training method, an embodiment of the present invention discloses a language training device, comprising:

[0118] A candidate training unit pool construction module 600 is used to construct a candidate training unit pool;

[0119] Scoring module 610, used to calculate the push priority score of each training unit in the candidate training unit pool;

[0120] The training question generation module 620 is configured to sort all training units in the candidate training unit pool from highest to lowest according to the push priority score, select a preset number of training units, and for each selected training unit, employ an interference strategy corresponding to the training unit that contains grammatical errors and / or cultural conflicts, and generate a training question corresponding to the training unit based on the large language model. All training questions constitute a training question set.

[0121] The training module 630 is configured to perform language training based on the training question set.

[0122] like Figure 6 As shown, an embodiment of the present invention discloses an electronic device, including a memory 401 storing executable program code and a processor 402 coupled to the memory 401;

[0123] The processor 402 calls the executable program code stored in the memory 401 to execute the language training method described in the above embodiments.

[0124] An embodiment of the present invention further discloses a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the language training method described in the above embodiments.

[0125] The purpose of the above embodiments is to exemplify and deduce the technical solution of the present invention, and to fully describe the technical solution, purpose and effect of the present invention. Its purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosed content of the present invention, and it does not limit the scope of protection of the present invention.

[0126] The above embodiments are not exhaustive and may include many other embodiments not listed above. Any replacements and improvements made without violating the concept of the present invention are within the scope of protection of the present invention.

Claims

1. A language training method, characterized in that: include: Build a pool of candidate training units; Calculate the push priority score for each training unit in the candidate training unit pool; Sort all training units in the candidate training unit pool according to their push priority scores from largest to smallest, and select a preset number of training units; For each selected training unit, an interference strategy containing grammatical errors and / or cultural conflicts corresponding to the training unit is adopted, and training questions corresponding to the training unit are generated based on the large language model. All training questions constitute a training question set; Language training is performed according to the training question set.

2. The language training method according to claim 1, wherein: Calculate the push priority score for the training unit, including: For the training unit, obtain the user's historical error records and the user's mastery status, and calculate the forgetting probability; The push priority score is calculated according to the historical error records, the user's personalized subject preference, the mastery status, and the forgetting probability.

3. The language training method according to claim 2, wherein: Calculate the forgetting probability, including: The forgetting probability is calculated according to the time difference between the time when the user last correctly memorized the training unit and the current time and the user's forgetting rate.

4. The language training method according to claim 2, wherein: The method adopts an interference strategy containing grammatical errors and / or cultural conflicts corresponding to the training unit and generates training questions corresponding to the training unit based on the large language model, including: Determining the interference strategy based on the historical error records, the user's personalized subject preference, and the forgetting probability; Obtaining an error template according to the interference strategy; constructing a prompt word according to the training unit and the error template; Based on the prompt words, a large language model is used to generate training questions corresponding to the training units.

5. The language training method according to claim 4, wherein: Determining the interference strategy based on the historical error records, the user's personalized subject preference, and the forgetting probability includes: When the number of historical error records is greater than a first preset threshold, setting the interference strategy to include a basic grammatical error strategy, wherein the basic grammatical error strategy is used to generate grammatical error questions; When the personalized topic preference is greater than a second preset threshold, setting the interference strategy to include a cultural expression error example strategy, wherein the cultural expression error example strategy is used to generate a cultural expression error question; When the forgetting probability is greater than a third preset threshold, the interference strategy is set to include a real error strategy, and the real error strategy is used to generate real wrong questions for the user.

6. The language training method according to claim 4, wherein: The step of generating training questions corresponding to the training units using a large language model based on the prompt words includes: Calculate the error risk level of the training unit; A large language model is used to generate training questions corresponding to the training unit according to the prompt word, the subject of the training unit and the error risk level.

7. The language training method according to claim 6, wherein: The calculating the error risk level of the training unit includes: Calculating a historical error rate based on the historical error records; The error risk level is calculated according to the historical error rate and the forgetting probability.

8. A language training device, characterized in that: include: A candidate training unit pool construction module is used to construct a candidate training unit pool; Scoring module, used to calculate the push priority score of each training unit in the candidate training unit pool; The training question generation module is used to sort all training units in the candidate training unit pool from highest to lowest according to the push priority score, select a preset number of training units, and for each selected training unit, adopt the interference strategy containing grammatical errors and / or cultural conflicts corresponding to the training unit, and generate training questions corresponding to the training unit based on the large language model. All training questions constitute the training question set; The training module is used to perform language training based on the training question set.

9. An electronic device, characterized in that: The method comprises a memory storing executable program code and a processor coupled to the memory; the processor calls the executable program code stored in the memory to execute the language training method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, wherein the computer program enables a computer to execute the language training method according to any one of claims 1 to 7.

Images