Personalized Russian spoken language practice recommendation method and system based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence education, in particular to a Russian spoken language practice personalized recommendation method and system based on artificial intelligence, and the method comprises the steps: 1, outputting a phoneme sequence with a timestamp through Russian automatic voice recognition; 2, collecting an exercise interruption position and repeated read-after behavior data; 3, generating a dynamic learner portrait; 4, mapping high-frequency errors in the learner portrait into abnormal path weights of map nodes; 5, a lattice tail error option and a non-matching body verb interference item are injected; 6, when the voice fluency attenuation of the learner exceeds a dynamic threshold value, the sentence complexity is reduced; and 7, calculating an error rate descent gradient based on the exercise completion data, and dynamically adjusting the abnormal path weight of the knowledge graph. Through audio stream analysis and syntax tree construction, the system can accurately identify errors of the learner in grammar, pronunciation and other aspects, and the learning efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence education, and particularly relates to a Russian oral practice individualization recommendation method and system based on artificial intelligence. BACKGROUND

[0002] In the field of artificial intelligence-driven language learning, Russian oral practice systems face the following field-specific challenges:

[0003] Morphological complexity: Russian has highly inflectional features, including:

[0004] Semantic differences between verb aspect pairs (perfective / imperfective) (e.g., "писать / написать" represent continuous action and result action respectively)

[0005] Noun declension rules of six-case system (e.g., instrumental case suffix "-ом / -ем" needs to match preposition "с")

[0006] Coupling of sound change rules and grammatical structure (e.g., word-final clear consonants need to be voiced before vowels, but are subject to case changes) Such rules result in highly complex error patterns that require specialized analysis.

[0007] The defects of the prior art specifically include:

[0008] (1) Static models cannot adapt to dynamic ability changes: Traditional systems rely on historical data to generate fixed difficulty exercises (e.g., dividing the primary / intermediate question bank based on test scores).

[0009] However, the oral ability of Russian learners presents a non-linear fluctuation: The instrumental case mastery rate may drop from 75% to 52% in a single conversation due to changes in preposition collocation, and the frequency of verb aspect misuse fluctuates dramatically with changes in conversation topics (e.g., "describe habits" requires the imperfective aspect, and "report events" requires the perfective aspect), resulting in a mismatch between practice and real-time ability, with a disruption rate of more than 35% (actual data).

[0010] (2) Insufficient extraction of field-specific features: Mainstream systems use general language analysis frameworks: Simplify case change errors as "noun morphological errors", ignoring their strong association with preposition governing rules, independently detect speech fluency (e.g., speech rate) and grammatical accuracy, and do not establish the association between sound change rule violations and grammatical structure:

[0011] When the noun requires the second case suffix "-а", the word-final clear consonant must be weakened (e.g., );

[0012] Due to the fragmentation of analysis, the missed detection rate of such coupled errors is as high as 32%.

[0013] Therefore, there is an urgent need for an artificial intelligence-based Russian oral practice individual recommendation method and system to solve the above problems. SUMMARY

[0014] Based on the above purpose, the present application provides an artificial intelligence-based Russian oral practice individual recommendation method and system, wherein the artificial intelligence-based Russian oral practice individual recommendation method comprises:

[0015] Step 1: Extract the Russian phonetic frame sequence of the learner's Russian phonetic flow through phonetic spectrum analysis, and output the time-stamped phonetic sequence through Russian automatic phonetic recognition;

[0016] Step 2: Perform Russian grammar dependency analysis on the phonetic sequence to generate a grammar tree containing verb form types and noun case change forms, and collect practice interruption positions and repeated reading behavior data;

[0017] Step 3: Based on the grammar tree, calculate the noun case change error rate and verb form misuse frequency, and detect the number of sound change rule violations through phonetic spectrum energy analysis to generate a dynamic learner portrait;

[0018] Step 4: Construct a Russian knowledge graph containing verb form pairs, case change rule chains, and sound change trigger conditions, and map the high-frequency errors in the learner portrait to the abnormal path weight of the graph node;

[0019] Step 5: Generate a target grammar structure sentence template according to the subgraph with the highest abnormal path weight, and inject case tail error options and non-matching body verb interference items;

[0020] Step 6: When the learner's fluency decreases by more than the dynamic threshold, reduce the sentence complexity;

[0021] Step 7: Calculate the error rate reduction gradient based on the practice completion data and dynamically adjust the abnormal path weight of the knowledge graph.

[0022] Preferably, the method for calculating the noun case change error rate in step 3 comprises:

[0023] Extract the case form tags of all nouns from the grammar tree and compare them with the standard case tail rules of the variable case method to which the noun belongs in the knowledge graph;

[0024] Calculate the case form error judgment value for each noun: if the actual case tail does not conform to the case type or part of speech consistency rules required by the preposition, mark it as an error instance;

[0025] Calculate the proportion of error instances in the total number of nouns as the noun case change error rate, wherein the preposition requirement rules are obtained by querying the Russian preposition dominance relationship library, and the part of speech consistency rules are achieved by detecting the matching degree of the adjective suffix and the noun, number, and case.

[0026] Preferably, the method of detecting the number of violations of sound change rules in step 3 comprises:

[0027] When the grammar tree indicates that the ending of a word needs to perform the rule of palatalization, locate the time interval of the phoneme in the sequence of Russian phonetic frames;

[0028] Extract the speech spectrum of the interval, and calculate the energy ratio of the low-frequency energy band to the high-frequency energy band;

[0029] Compare the energy ratio with a dynamic threshold value: if it is lower than the threshold value, it is determined that the sound change rule is violated, and the number of violations is accumulated;

[0030] The dynamic threshold value is determined by analyzing the energy ratio distribution of the same position phonemes of the native speaker, specifically by taking the mean value of the energy ratio of the native speaker sample minus the standard deviation.

[0031] Preferably, the method of generating abnormal path weights in step 4 comprises:

[0032] For each grammatical error pattern in the dynamic learner profile, calculate the base weight value of the corresponding node in the knowledge graph, which is proportional to the frequency of the error pattern in unit time;

[0033] Detect adjacent nodes that have a grammatical connection with the node: for verb case misuse nodes, find their commonly used collocation time adverbial nodes, for instrumental case error nodes, find their governing preposition nodes;

[0034] Calculate the associated weight gain value: if the error rate of the adjacent node significantly increases with the error rate of the current node, amplify the base weight value by the association strength coefficient;

[0035] The final abnormal path weight is the total weight after superimposing the associated weight gain value;

[0036] The method of calculating the association strength coefficient comprises:

[0037] Statistical high-frequency error node and adjacent node error rate historical data;

[0038] Calculate the covariance of the error rate of the adjacent node with the error rate fluctuation of the high-frequency node;

[0039] Compare the covariance value with the preset activation threshold value, and generate the association strength coefficient in proportion when the threshold value is exceeded.

[0040] Preferably, the injection rule of the non-matching body verb interference term in step 5 comprises:

[0041] Case ending error option generation:

[0042] Get all standard case endings of the target noun from the knowledge graph;

[0043] Excluding the correct case ending in the current context, screening the infrequent case endings from the remaining case endings to form the interference option set;

[0044] The criterion for determining the infrequent case ending is that the frequency of the case ending in the Russian corpus is lower than the average frequency of the belonging declension;

[0045] Non-matching body verb generation:

[0046] Locating the body-pair group of the target verb in the knowledge graph, and extracting the non-matching body verb candidate set;

[0047] Calculating the semantic similarity between the candidate verb and the target verb, and screening the verbs with similarity higher than the threshold to retain the tense form but replace the body type.

[0048] Preferably, the sentence complexity reduction method in step 6 includes:

[0049] Sentence structure simplification:

[0050] Identifying the modifying components of non-target grammatical structures in the sentence, and deleting the attributive clauses or adverbial clauses;

[0051] Splitting the complex sentence into two simple sentences, and only retaining the single sentence containing the target grammatical structure;

[0052] Vocabulary difficulty reduction:

[0053] Extracting the roots of non-target vocabulary in the sentence, and selecting alternative words from the basic vocabulary library of the same root;

[0054] The construction criterion of the basic vocabulary library is that the frequency of the vocabulary in the top 20% of Russian primary teaching materials.

[0055] Preferably, the dynamic learner profile in step 3 includes a cold start mechanism:

[0056] For users without historical data, load the preset common error pattern library of Russian native speakers as the initial profile;

[0057] Generate a calibration question set containing three detection dimensions:

[0058] Declension detection sentence: contains preposition collocation structure of different declension nouns;

[0059] Verb body detection sentence: contains tense contrast structure of perfective body / inperfective body;

[0060] Stress detection word: contains polysyllabic words with mobile stress rules;

[0061] Based on the distribution of user answer error types, weight and fuse the weights in the native speaker error pattern library.

[0062] Preferably, the method for calculating the error rate reduction gradient in step 7 comprises:

[0063] For each abnormal path weight corresponding to a syntax error type, calculate the difference between the current exercise error rate and the previous exercise error rate of the same type;

[0064] According to the time interval between the two exercises, calculate the error rate change per unit time as the reduction gradient;

[0065] For the coupled error type with error propagation link, if the associated node error rate is reduced synchronously, the gradient value is weighted and amplified according to the number of associated nodes.

[0066] Preferably, the way to dynamically adjust the abnormal path weight of the knowledge graph in step 7 comprises:

[0067] If the reduction gradient continuously exceeds the gradient threshold, the abnormal path weight of the corresponding node is attenuated to the basic value;

[0068] If the reduction gradient of the coupled error type continues to be lower than the threshold, increase the connection depth of the node and the associated phonetic rule node in the knowledge graph;

[0069] The gradient threshold is determined by the average gradient change rate of the effective progress users in the historical data.

[0070] Correspondingly, the embodiment of the application also provides a Russian oral practice individualized recommendation system based on artificial intelligence, comprising a memory configured to store instructions, a processor configured to call the instructions from the memory and capable of realizing the Russian oral practice individualized recommendation method based on artificial intelligence as any one of the embodiments of the application when executing the instructions.

[0071] The beneficial effects of the application are:

[0072] 1. The application adopts a dynamic learner profile and abnormal path weight mechanism, which can dynamically adjust the practice content according to the real-time performance of the learner in each exercise. This method discards the traditional fixed difficulty question bank based mode, and accurately matches the appropriate exercise task by monitoring the learner's oral ability fluctuation in real time. This real-time adaptability significantly reduces the error rate in the learning process and improves the learning efficiency.

[0073] 2、The present application accurately detects the errors of noun case changes, verb form misuse, and sound change rules violation in Russian oral speech by combining syntax tree analysis and speech spectrum energy analysis. In addition, through the sound change rule detection method, the system can detect the errors of word ending sound change caused by case changes, such as the failure to perform the rule of weakening the word ending clear consonant, and then generate targeted correction suggestions. The combination of these technologies effectively solves the problem that traditional systems cannot accurately identify domain-specific errors, greatly reduces the missed detection rate, and improves the accuracy and intelligence level of the system.

[0074] 3、By combining dynamic learner profile and abnormal path weight mechanism in knowledge graph, the system can generate personalized learning suggestions for different learner error patterns. Through this method, learners can correct high-frequency errors in time, instead of repeatedly practicing inefficient and repetitive content.

[0075] 4、The present application introduces a cold start mechanism for learners without historical data. When used for the first time, it generates a preliminary learning profile by loading the pre-set common error pattern library of Russian native speakers, and adjusts the learning content according to the real-time error distribution of the learner. This mechanism not only solves the problem of inaccurate recommendations for new users, but also enables even beginners to have personalized and targeted learning experience in the early stage.

[0076] 5、The present application can calculate the error rate reduction gradient of each error path by tracking the exercise completion data in real time, and dynamically adjust the weight of abnormal paths in the knowledge graph according to the gradient. When the learner's error rate continues to decline, the system will reduce the difficulty and provide simpler grammar structures; on the contrary, if the error rate does not decrease effectively, the system will increase the weight of the error path to ensure that the learner can focus on solving their main difficulties. This flexible adjustment mechanism effectively avoids stagnation or rapid progress of learning progress, ensuring that learners make continuous progress at an appropriate difficulty level. BRIEF DESCRIPTION OF DRAWINGS

[0077] In order to more clearly illustrate the technical solutions in the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without creative labor.

[0078] Fig. 1 The step flow chart of the method of the present application;

[0079] Fig. 2 The step flow chart of the error rate reduction gradient calculation method of the present application;

[0080] Fig. 3A step flowchart of a method for dynamically adjusting the weight of an abnormal path of a knowledge graph for the method of the present application. DETAILED DESCRIPTION

[0081] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted here that in order to make the embodiments more detailed, the following embodiments are the best, preferred embodiments, and other alternative ways can also be implemented by those skilled in the art for some known technologies; and the drawings are only used to more specifically describe the embodiments, and are not intended to specifically limit the present application.

[0082] Please refer to Figs. 1-3 The embodiment of the present application provides a Russian oral practice personalized recommendation method based on artificial intelligence. In step 1, the learner's Russian speech stream is first extracted through speech spectrum analysis and converted into a Russian phonetic frame sequence. Then, the speech stream is processed using Russian automatic speech recognition technology (ASR) to generate a phoneme sequence with a timestamp. The core of this step is to accurately convert the learner's speech content into a structured data format for subsequent syntax analysis and error detection. Through automatic speech recognition, the learner's pronunciation features and errors can be captured in real time, thereby laying the foundation for personalized recommendation.

[0083] In step 2, after obtaining the phoneme sequence, the system performs Russian syntax dependency analysis to construct a syntax tree containing verb case types and noun declension forms. This step can analyze the grammar structure of Russian oral speech in detail, especially the verb case and noun declension forms, to identify whether the learner has grammar errors (such as verb case misuse, noun case errors, etc.). In addition, the system also records the learner's interruption position and repeated reading behavior data during the practice process, further capturing the learner's learning state in actual practice, and evaluating their oral fluency and grammar mastery.

[0084] In step 3, based on the analysis results of the syntax tree, the system calculates the noun declension error rate and verb case misuse frequency, and combines the speech spectrum energy analysis to detect whether the sound change rule is violated (such as the word tail voicing is not properly pronounced). These data will be used to generate a dynamic learner profile that reflects the learner's error patterns and learning progress in real time, providing accurate basis for personalized recommendation. The learner profile not only reflects the immediate performance of a single practice, but also tracks long-term learning progress and high-frequency error areas.

[0085] In step 4, by counting the high-frequency errors in the learner's profile, the system maps these errors to the Russian knowledge graph containing verb case pairings, case change rule chains, and sound change trigger conditions. In the graph, different grammatical errors are mapped as abnormal path weights of graph nodes. This knowledge graph can concretize the learner's errors and adjust the practice content according to the learner's specific error situation, ensuring that the learner's practice content matches their weaknesses and improving learning efficiency.

[0086] In step 5, according to the abnormal path weights in the knowledge graph, the system identifies the sub-graph with the highest error frequency and generates target grammatical structure sentence templates based on these errors. The templates will inject case ending error options (such as correct and incorrect case ending changes) and non-matching body verb interference items (such as verbs that do not match the current tense or body). In this way, the learner will face customized practice questions that can both reinforce weaknesses and allow for timely correction during practice.

[0087] In step 6, when the system detects that the learner's speech fluency has declined (such as increased speech pauses or increased pronunciation errors) beyond a dynamic threshold, the system automatically reduces the complexity of the sentence, reducing the grammatical structure and vocabulary difficulty in the sentence to avoid frustration caused by excessive difficulty. This dynamic adjustment mechanism ensures that the learner can maintain high motivation and effectiveness during continuous practice.

[0088] In step 7, after each practice is completed, the system calculates the error rate decline gradient based on the practice completion data to evaluate the learner's progress during practice. According to the gradient change, the system dynamically adjusts the abnormal path weights in the knowledge graph. If the learner's error rate rapidly decreases, the system will appropriately increase the difficulty of the practice; otherwise, the system will reduce the difficulty to ensure that the learner continues to progress in appropriate challenges.

[0089] By introducing dynamic learner profiles, syntax dependency analysis, Russian knowledge graphs, and other technical features, the individualization and precision of Russian oral practice have been significantly improved. Compared with existing static systems, this method can adapt to the learner's ability changes in real time and dynamically adjust the practice content based on the learner's specific errors, avoiding the learner from prematurely entering inappropriate difficulty levels and reducing frustration during the learning process. At the same time, through the coupling detection of sound change rules and grammar rules, the system's recognition rate of complex grammatical errors is effectively improved, the learner's oral fluency and grammatical accuracy are improved, and the learning efficiency and learning experience are improved.

[0090] In one possible implementation, first, the system extracts all the case morphological tags of nouns from the syntax tree of a Russian sentence. The syntax tree is generated by Russian syntactic dependency analysis and can accurately identify the grammatical structure and relationship of each word in the sentence. On this basis, the system extracts the corresponding case morphological tags (such as nominative, dative, genitive, etc.) from the word form of each noun.

[0091] After extracting the case morphological tags of nouns, the system compares them with the standard case ending rules in the knowledge graph. The knowledge graph contains the declension rules of various types of nouns in Russian and their corresponding case ending patterns, such as providing standard case ending rules according to different nouns, parts of speech, gender, number, etc.

[0092] For each noun, the system will determine whether the case morphology is correct through the preposition requirement rule and the part of speech consistency rule. If the actual case ending does not meet the case type required by the preposition (such as the preposition requires the genitive case), or the noun's ending is inconsistent with the gender, number, and case of the adjective it modifies, then the instance is marked as an error. Specifically, the preposition requirement rule is obtained by querying the Russian preposition domination relationship library to ensure that the preposition and noun are correctly matched; the part of speech consistency rule judges errors by detecting whether the adjective and noun endings are consistent.

[0093] By counting the number of instances marked as errors in all nouns and calculating the proportion of the total number of nouns, the system obtains the error rate of noun case declension. This error rate can reflect the learner's mastery of noun case declension and serve as a basis for subsequent personalized practice recommendations.

[0094] By comparing the case morphological tags of nouns with the standard case ending rules, the system can accurately identify the learner's errors in noun case declension, avoiding the rough judgment of grammatical errors in traditional teaching methods. This method not only focuses on the morphological changes of nouns, but also considers the collocation rules with prepositions and adjectives, improving the comprehensiveness and accuracy of identification.

[0095] The statistics of noun case declension error rate provide a scientific basis for dynamically adjusting learning content. If the learner's error rate is high, the system can generate more relevant practice content to help the learner focus on overcoming this difficulty, thereby improving learning efficiency.

[0096] Combined with the learner's case declension error rate, the system can customize personalized practice questions for each learner, improving the relevance and effectiveness of their learning. Through this personalized recommendation based on error data, the learner's weak points can be strengthened in a short period of time, avoiding repetitive and ineffective practice.

[0097] In one possible implementation, first, in the input Russian sentence, the system marks the word endings that need to perform the rule of palatalization of consonants by the process of syntax tree annotation. Palatalization of consonants is a common phonetic phenomenon in Russian, which occurs before or after certain consonants and can lead to weakening or change of pronunciation of phonemes. Syntax tree analysis can determine the location of these phonemes and related word endings.

[0098] Once the phonemes that need to perform the phonetic rule are marked, the system then locates the time interval of the phonemes according to the sequence of speech frames of the audio. Speech frame sequence refers to the division of continuous audio signal into small time periods for the analysis of each frequency component in the audio. This process helps to accurately obtain the pronunciation features of the phonemes, ensuring that the weakening of each phoneme can be analyzed independently.

[0099] After determining the time interval of the phonemes, the system extracts the speech spectrum of the interval, and the spectrum analysis can reveal the energy distribution of the audio signal at different frequency bands. The system is particularly interested in the energy ratio between the low frequency energy band (usually related to the clarity of consonant pronunciation) and the high frequency energy band (related to the clarity and sharpness). The phenomenon of palatalization of consonants usually manifests as a decrease in low frequency energy or an increase in high frequency energy.

[0100] Then, the system compares the calculated energy ratio with a dynamic threshold. The dynamic threshold is obtained by analyzing the energy ratio distribution of the same position phonemes of a large number of native Russian speakers. Specifically, when pronouncing normally, the ratio of low frequency and high frequency energy has a certain mean and standard deviation, and the system determines the dynamic threshold by taking the mean minus the standard deviation. If the learner's pronunciation is below this threshold, the system determines that the phoneme violates the rule of palatalization of consonants, and records the event as a violation.

[0101] The system accumulates the number of phonetic rule violations of the learner within a period of time, and this information will be used for personalized recommendations. High frequency of phonetic rule violations means that the learner has significant problems in palatalization of consonants, and the system will adjust the subsequent practice content according to this information to help the learner focus on improving this aspect.

[0102] The phonetic rule violation detection method of the present application, combined with the spectrum analysis of the speech signal and the dynamic threshold, is more accurate than the traditional error judgment method, effectively improving the personalization and intelligence level of Russian oral practice.

[0103] In one possible implementation, the dynamic learner profile is dynamically updated according to the error patterns generated by the learner during the learning process. The system calculates the base weight value of the corresponding node in the knowledge graph according to the frequency of a certain grammatical error pattern of the learner within a unit of time. The weight value is directly proportional to the error frequency, i.e. the more frequent the error pattern, the higher the weight. This process is marked and calculated through the nodes in the knowledge graph, ensuring that each error pattern has its own independent base weight.

[0104] After calculating the base weight value of the error pattern, the system detects the adjacent nodes that have a grammatical association with the node according to the grammatical rules. For example, for the verb form misuse node, the system will look for the time adverbial node that is commonly used with it; for the instrumental case error node, the system will look for the node that governs the preposition. These adjacent nodes represent other related parts of the grammatical structure, reflecting the propagation effect of the error.

[0105] The system generates the associated weight gain value by calculating the situation where the error rate of these adjacent nodes significantly increases with the error rate of the current node. Specifically, if the increase in the error rate of a node leads to a significant increase in the error rate of its associated nodes, it indicates that there is a strong association between these nodes. At this time, the system amplifies the base weight value by a certain association strength coefficient, so that the grammatical structure errors related to the error pattern are more strongly marked, thereby helping to generate personalized recommendations.

[0106] Finally, the system superimposes the base weight value and the associated weight gain value of the adjacent nodes to obtain the final abnormal path weight. This weight value not only reflects the severity of a single error pattern, but also takes into account its relationship with other grammatical parts, providing rich weight data for personalized recommendations.

[0107] To accurately measure the grammatical association between different nodes, the system first counts the error rate history data of high-frequency error nodes and their adjacent nodes. Then, the error rate fluctuation covariance between these nodes is calculated. Covariance can reveal whether the error rate trends of the nodes are correlated. If the covariance value exceeds the preset activation threshold, the system generates an association strength coefficient in proportion to adjust the weight gain of the adjacent nodes.

[0108] Through in-depth error pattern analysis, grammatical association detection, and weight gain calculation, highly personalized learning recommendations can be provided to the learner, improving the efficiency and quality of Russian oral learning.

[0109] In one possible implementation, first, the system identifies the target noun based on the learner's grammatical error and obtains all the standard case endings of the declension of the noun from the knowledge graph. In Russian, the declension of a noun involves different case endings, and the system needs to locate the correct case ending that the noun should use according to the learner's error.

[0110] When generating the interference options, the system excludes the correct case ending in the current context to ensure that the generated interference items are not consistent with the current context, thereby achieving the purpose of challenging the learner.

[0111] The system then selects the infrequently used case endings from the remaining case endings and forms a set of interference options. The criterion for determining infrequently used case endings is that the frequency of the case ending in the Russian corpus is lower than the average frequency of the declension to which it belongs. These infrequently used case endings are often prone to errors by learners and can effectively help learners distinguish the correct forms in different contexts as interference items.

[0112] Through the knowledge graph, the system locates the pair of body of the target verb. The body is a grammatical category of Russian verbs, referring to the completeness of the action or state of the verb (such as the perfective and imperfective bodies of the verb). The system first analyzes the verb used by the learner and finds out the body type that pairs with it.

[0113] The system extracts non-matching body verbs from the pair of body of the target verb, i.e. verbs that do not match the body type of the target verb. These verbs will be injected as interference items to help learners distinguish different body usage.

[0114] The system selects verbs with a semantic similarity to the target verb higher than a preset threshold by calculating the semantic similarity of these candidate verbs to the target verb and retains their tense forms, replacing only the body type. This ensures that the interference items are similar in semantics to the target verb but differ in body type, making them more challenging.

[0115] The non-matching body verb interference item injection rule in the present invention, through precise knowledge graph support and verb body pair analysis, can effectively generate interference items to help learners better master Russian grammar and improve the accuracy and fluency of oral expression.

[0116] In one possible implementation, first, the system analyzes the structure of the input sentence through natural language processing (NLP) techniques and identifies the modifying components in the sentence that are not related to the target grammatical structure. These modifying components include attributive clauses, adverbial clauses, etc. Through this process, the system can understand which parts of the sentence are unrelated to the target grammatical structure, thereby preparing for subsequent simplification.

[0117] In analyzing sentences, the system automatically identifies and removes attributive clauses or adverbial clauses. Attributive clauses and adverbial clauses often increase the complexity of a sentence, and for beginners, understanding these clauses can be confusing. By removing these modifiers, the system can ensure that the sentence becomes simpler, focusing on the target grammatical structure, thereby helping learners focus on the core of the grammar.

[0118] For complex sentences, the system splits them into two or more simple sentences, ensuring that each sentence maintains a simple and clear grammatical structure and only retains the single sentence that contains the target grammatical structure. This process helps to reduce the number of clauses and parallel structures in the sentence, making it easier for learners to understand the basic composition of the sentence.

[0119] During the simplification process, the system analyzes all the vocabulary in the sentence, identifies complex vocabulary that is not part of the target vocabulary, and extracts its root. Vocabulary complexity is often a significant obstacle for learners to understand and use the language, so the system focuses on processing these non-target vocabulary.

[0120] The system selects alternative words from the basic vocabulary library that correspond to the original root. The selection criteria for alternative words is that the vocabulary appears in the top 20% of frequency in Russian elementary textbooks. These basic vocabulary is usually the first vocabulary that beginners encounter in the learning process, easy to understand and use. Therefore, using these basic vocabulary can help learners to be more confident in practical application.

[0121] The basic vocabulary library is selected by analyzing the vocabulary in Russian elementary textbooks and selecting the most common and basic vocabulary. The selected vocabulary is usually the core vocabulary that constitutes simple sentences and basic grammatical structures, with high frequency of use and low learning difficulty.

[0122] The sentence complexity reduction method in the invention helps learners practice in a simpler language environment by simplifying grammatical structures and reducing vocabulary difficulty. Such personalized adjustments enable learners to gradually improve their language proficiency, thereby improving learning efficiency and confidence.

[0123] In one possible implementation, for users without historical learning data, the system will load a preset common error pattern library for Russian native speakers. This error pattern library is based on a large number of typical errors made by native speakers during the process of learning Russian, including but not limited to grammatical errors, improper use of vocabulary, pronunciation problems, etc. The error patterns in this library represent common difficulties in learning Russian, so for cold start users, it provides a more accurate starting learning profile, enabling them to start from common error types that affect all Russian learners, and effectively train the language.

[0124] To better personalize the assessment of learners, the system generates a calibration question set containing three detection dimensions to comprehensively detect learners' performance in Russian oral practice.

[0125] Inflection detection sentences: This type of sentence design contains preposition collocation structures with different inflection methods of nouns. Inflection is a complex grammatical element in Russian, and errors are common among beginners. Through inflection detection sentences, the system can detect learners' mastery of noun inflection and its collocation with prepositions.

[0126] Verb aspect detection sentences: This dimension focuses on learners' understanding and use of the tense contrast between perfective and imperfective aspects in Russian. The choice of verb aspect is crucial for the expression of time in a sentence, and incorrect use of perfective and imperfective aspects can lead to misunderstandings. This detection sentence can help the system assess learners' mastery of verb aspect.

[0127] Stress detection words: Russian polysyllabic words have stress rules, and the correct or incorrect stress directly affects the accuracy of the meaning and the fluency of spoken language. Through this dimension, the system can test learners' mastery of the rules of moving stress.

[0128] After the user has answered the above calibration question set, the system will analyze the user's answer errors and weight the error patterns in the native speaker's common error pattern library. Specifically, the system will dynamically adjust the weights in the error pattern library according to the user's error distribution in each detection dimension. For example, if the user frequently makes mistakes in inflection detection sentences, the system will increase the weight of inflection-related error patterns, so that the subsequent learning materials will focus more on the training of noun inflection and preposition collocation. This weighting fusion mechanism can accurately adjust the learning content and recommendation strategy according to the user's specific error types.

[0129] Through the cold start mechanism, the calibration question set of the three detection dimensions, and the weighting fusion based on error types, efficient personalized recommendation in Russian oral practice is achieved. This not only improves the relevance and efficiency of learning, but also enhances the learners' confidence and motivation.

[0130] In one possible implementation, after each practice, the system records the error rate of the learner in each grammar error type corresponding to the weight of the abnormal path. When calculating the gradient, the system compares the error rate of the current practice with the error rate of the same type of practice last time. This difference represents the change in the error rate of the user in this practice, which can reflect the improvement or regression of the learner's mastery of the grammar point.

[0131] For example, if a user's error rate for verb conjugation is 30% on the first practice and decreases to 20% on the second practice, the error rate difference is 10%. This difference represents the learner's progress or improvement in that grammar type.

[0132] To ensure that the trend of error rate change truly reflects the learner's learning progress, the system also calculates the error rate change per unit time based on the time interval between the two practices. Specifically, the system uses the following formula to calculate the decline gradient:

[0133] Decline gradient = Error rate difference / Time interval;

[0134] This calculation quantifies the improvement speed of error rate per unit time, helping the system assess the learner's learning effect in a specific time. For example, if the time interval between the two practices is long and the error rate decreases significantly, the system will consider that the learner has made significant progress in this period, and vice versa, which may reflect the learner's slow progress.

[0135] In Russian grammar, there are often some coupled error types, that is, there is a correlation between multiple error types, and the change of error rate may affect each other. For these coupled error types, if the error rates of related nodes decrease synchronously, the system will weight and amplify the gradient value according to the number of associated nodes. Specifically, the system will analyze the propagation link between error types, identify which error types are coupled, and amplify the learning progress of these error types by a certain weight.

[0136] For example, verb conjugation errors and tense selection errors may be coupled, and if the user has made significant progress in both types of errors, the system will increase the weight adjustment of these error types, thereby accelerating the recommendation and learning of this part of grammar knowledge.

[0137] Through the calculation method of error rate decline gradient, the progress of the learner at each grammar point can be accurately tracked, the personalized learning path is optimized, and the learning efficiency and overall satisfaction of the learner are improved through the weighted amplification of coupled error types.

[0138] In one possible implementation, first, the system continuously monitors the change of error rate of the learner on each grammar path and calculates the decline gradient of each error path. When a learner's decline gradient on a certain grammar error type continuously exceeds the preset gradient threshold, the system considers that this error type has been sufficiently improved, and then it does not need to recommend this path for reinforcement with excessive resources. Therefore, the system will decay the abnormal path weight of the node corresponding to this path to the base value, that is, it will reduce this grammar point from the recommendation list to avoid unnecessary repeated practice and learning.

[0139] For example, in an embodiment of the present application, the error rate of learners in verb conjugation is significantly reduced, and the gradient descent of the last three times exceeds the set threshold value, then the system will reduce the practice recommendation for this point and attenuate its weight to the base value, so that the system's recommendation is more intelligent and effective.

[0140] For some grammar error types that have coupling relationships, if the system detects that the descending gradient of the relevant node continues to be lower than the set threshold value, indicating that the learner has not made significant progress in these error types, the system will strengthen the recommended content by deepening the connection depth of the node with the associated phonetic variation rule node. Increasing the connection depth of the node means that the system will recommend more rules or exercises related to these error types to the learner in order to solve the root problem of the error type.

[0141] In an embodiment of the present application, when the system detects that the learner's performance in phonetic variation rules (such as verb tense phonetic variation rules) is poor and the descending gradient is at a low level for a long time, it will help the learner overcome this difficulty by increasing the learning nodes related to the phonetic variation rule (such as more detailed phonetic variation rules or supporting example sentences).

[0142] The setting of the gradient threshold value is determined based on the average gradient change rate of effective progress users (i.e. learners who have made significant progress) through statistical analysis of historical data. The system calculates the average gradient change rate of a large number of historical users and sets a suitable threshold value based on this rate. Paths with a gradient change rate higher than the threshold value are considered to have made significant progress, and paths with a gradient change rate lower than the threshold value are considered to have made slow progress.

[0143] In an embodiment of the present application, after the system has analyzed the learning data of 1000 effective progress users, it finds that the average gradient change rate of these users is 0.05, so the system will use this value as the gradient threshold value to ensure that only paths with a rate exceeding this rate will be considered to have made significant progress.

[0144] By dynamically adjusting the weight of abnormal paths in the knowledge graph, more accurate, efficient and personalized learning recommendations are achieved. This technical feature not only helps learners quickly master language knowledge, but also effectively improves the overall efficiency of learning and the satisfaction of learners.

[0145] Correspondingly, an embodiment of the present application also provides a Russian oral practice personalized recommendation system based on artificial intelligence, which includes a storage configured to store instructions, a processor configured to call the instructions from the storage and capable of implementing the personalized recommendation method for Russian oral practice based on artificial intelligence as described in any of the embodiments of the present application when executing the instructions.

[0146] The following will be described in detail through examples:

[0147] The embodiment describes a personalized recommendation system and method for Russian learning, especially for beginners' learning path recommendation. The system analyzes the learner's grammar mastery, pronunciation accuracy, and learning progress, and combines intelligent recommendation algorithms to generate personalized learning paths for learners, providing recommendations for grammar points, vocabulary, and pronunciation exercises.

[0148] Specifically, the learner enters the system, and the system obtains their pronunciation through voice recognition technology and analyzes their grammar errors through automated grammar checking tools.

[0149] In the embodiment of the application, the learner inputs the sentence "I am a student", and the system identifies "I" and "student" as the basic vocabulary of subject and object.

[0150] The system constructs a corresponding grammar tree based on the content input by the learner, for example, identifies the subject, verb, and object structures, and maps them to the corresponding knowledge graph nodes.

[0151] For example, "I" and "is" are mapped to the "subject + copula" structure, and "student" is mapped to the "noun" structure.

[0152] The system selects a recommendation algorithm (such as a deep neural network based on collaborative filtering) to generate a personalized learning path based on the learner's performance.

[0153] In the embodiment of the application, the recommended learning path includes:

[0154] Grammar point recommendation: verb conjugation, noun number, etc.

[0155] Pronunciation practice recommendation: stress, sound change rules, etc.

[0156] In the embodiment of the application, the system uses a gradient-based optimization method to update the recommended path based on the learner's error performance (e.g., pronunciation errors or grammar errors).

[0157] The system sets the gradient threshold to 0.05, indicating that the learner's performance fluctuation exceeds the threshold, and the recommended path needs to be adjusted.

[0158] For example, if the learner frequently makes mistakes in the "verb conjugation" section, the system will increase the training proportion of this section.

[0159] Gradient update formula:

[0160] △W = -η▽L(W);

[0161] Where: ΔW is the weight update amount; η is the learning rate, taking a value of 0.01; is the gradient of the loss function, indicating the size of the learner's error.

[0162] In an embodiment of the invention, if a learner has been struggling with a certain grammar point (e.g. verb conjugation) for a long time, the system will decay the learning path weight for that section, for example with a decay rate of 0.8.

[0163] Decay formula:

[0164] W new = W old × decay_rate;

[0165] Where: W old is the original path weight; W new is the decayed path weight; decay_rate is the decay rate, taking a value of 0.8.

[0166] The system will dynamically adjust the recommended path based on the learner's feedback and real-time performance (e.g. error rate, exercise completion rate, etc.).

[0167] In an embodiment of the invention, if a learner performs poorly in the "verb conjugation" section, the system will optimize the learning path by dynamically adjusting parameters (e.g. reducing the number of recommended times for that grammar point).

[0168] Specific parameter settings:

[0169] Learning rate: 0.01, determines the step size in the model training process.

[0170] Gradient threshold: 0.05, indicates that when the learner's performance fluctuates more than this threshold, the system will trigger the adjustment of the learning path.

[0171] Decay rate: 0.8, used to control the weight decay of the path.

[0172] Maximum length of recommended path: 5, the system will recommend a maximum of 5 learning paths to avoid information overload.

[0173] Number of knowledge graph nodes: 50, the number of knowledge graph nodes for each learner is fixed to ensure the computational efficiency of the system.

[0174] In order to verify the effect of the invention, a comparative experiment with the traditional learning system was designed. In an embodiment of the invention, two groups of learners, the experimental group uses the personalized recommendation system of the invention, and the control group uses the traditional static learning path.

[0175] Experimental group: through the personalized recommended learning path, the system continuously optimizes the recommended content according to the learner's feedback, focusing on recommending weak grammar points and pronunciation rules for the learner.

[0176] Control group: use fixed learning path, all learners follow the same learning steps, no personalized adjustment.

[0177] Experimental results:

[0178] Learning progress speed: The learning progress speed of the experimental group increased by an average of 25%, and the grammar mastery and pronunciation accuracy of the learners improved significantly.

[0179] Learning completion: The learning completion rate of the experimental group was 90%, while that of the control group was 75%.

[0180] Error rate reduction: The grammar and pronunciation error rates of the experimental group were 20% lower than those of the control group.

[0181] From the above experimental results, it can be seen that the personalized recommendation system of the present application has obvious advantages in learning progress and learning quality.

[0182] This embodiment clearly shows how the present application realizes personalized learning path recommendation through specific application scenarios, algorithm processes, parameter values and formulas, and verifies its outstanding effect in practical application through comparative experiments. This process not only ensures the practical operability of the system, but also ensures the realizability and innovativeness of the technology.

[0183] The present application covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present application. In order to enable the public to have a thorough understanding of the present application, specific details are described in the following preferred embodiments of the present application, and the present application can also be fully understood without these details by those skilled in the art. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits, etc. are not described in detail.

[0184] The above is only the preferred embodiment of the present application, and it should be noted that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can also be made, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A personalized recommendation method for Russian oral practice based on artificial intelligence, characterized in that, include: Step 1: Extract the Russian speech frame sequence of the learner's Russian speech stream through speech spectrum analysis, and output the time-stamped phoneme sequence through Russian automatic speech recognition; Step 2: Perform Russian grammatical dependency analysis on the phoneme sequence to generate a grammar tree that includes verb aspect type and noun case inflection, and collect data on practice interruption positions and repeated reading behavior. Step 3: Based on the syntax tree, the error rate of noun case changes and the frequency of verb aspect misuse are statistically analyzed. Combined with speech spectrum energy analysis, the number of violations of sound change rules is detected, and a dynamic learner profile is generated. Step 4: Construct a Russian knowledge graph that includes verb aspect matching, case inflection rule chains, and sound change triggering conditions, and map high-frequency errors in learner profiles to abnormal path weights of graph nodes; Step 5: Generate the target grammatical structure sentence template based on the subgraph with the highest weight of the abnormal path, and inject case tail error options and non-matching aspect verb interference items; Step 6: When the learner's fluency decays beyond the dynamic threshold, reduce the sentence complexity; Step 7: Calculate the error rate descent gradient based on the completed practice data, and dynamically adjust the abnormal path weights of the knowledge graph.

2. The personalized recommendation method for Russian oral practice based on artificial intelligence according to claim 1, characterized in that, The statistical methods for calculating the error rate of noun case changes in step 3 include: Extract the case morphology labels of all nouns from the syntax tree and compare them with the standard case tail rules of the declension rules of the nouns in the knowledge graph; Calculate a case morphology error value for each noun: if the actual case ending does not conform to the case type or part-of-speech consistency rule required by the preposition, mark it as an error instance; The proportion of statistically incorrect instances to the total number of nouns is used as the noun case change error rate. The preposition requirement rule is obtained by querying the Russian preposition dominion relation database, and the part-of-speech consistency rule is achieved by detecting the matching degree between the adjective ending and the noun's gender, number, and case.

3. The personalized recommendation method for Russian oral practice based on artificial intelligence according to claim 1, characterized in that, The methods for detecting the number of violations of the tone change rules in step 3 include: When the syntax tree annotation requires the execution of voiceless consonant weakening rules for word endings, locate the time interval of the phoneme in the Russian speech frame sequence; Extract the speech spectrum of this interval and calculate the energy ratio of the low-frequency energy band to the high-frequency energy band; Compare the energy ratio with a dynamic threshold: if it is lower than the threshold, it is determined that the sound change rule is violated, and the number of violations is accumulated. The dynamic threshold is determined by analyzing the energy ratio distribution of co-positional phonemes among native Russian speakers, specifically by subtracting the standard deviation from the mean energy ratio of the native speaker sample.

4. The personalized recommendation method for Russian oral practice based on artificial intelligence according to claim 1, characterized in that, The methods for generating abnormal path weights in step 4 include: For each grammatical error pattern in the dynamic learner profile, calculate the basic weight value of its corresponding node in the knowledge graph. This value is proportional to the frequency of occurrence of the error pattern per unit time. Detect adjacent nodes that are grammatically related to this node: for nodes with verb misuse, find the time adverbs that are commonly used in their collocations; for nodes with incorrect instrumental cases, find the prepositions that govern them. Calculate the association weight gain value: If the error rate of adjacent nodes increases significantly with the error rate of the current node, amplify the basic weight value according to the association strength coefficient; The final abnormal path weight is the total weight after adding the associated weight gain value; The methods for calculating the correlation strength coefficient include: Statistical analysis of historical error rates between high-frequency error nodes and their adjacent nodes; Calculate the covariance of the error rate of adjacent nodes as the error rate of high-frequency nodes fluctuates; The covariance value is compared with a preset activation threshold. If the value exceeds the threshold, a correlation strength coefficient is generated proportionally.

5. The method for personalized recommendation of Russian oral practice based on artificial intelligence according to claim 1, characterized in that, The injection rules for non-matching verb distractors in step 5 include: Generate tail error options: Obtain the set of all standard case endings of the declension to which the target noun belongs from the knowledge graph; Eliminate case endings that are correct in the current context, and select uncommon case endings from the remaining case endings to form a set of distracting options; The criterion for determining an uncommon case ending is: the frequency of the case ending in the Russian corpus is lower than the average frequency of the declension of that case. Non-matching aspect verb generation: Locate the aspect pairings of the target verb in the knowledge graph and extract the candidate set of non-matching aspect verbs; Calculate the semantic similarity between candidate verbs and target verbs, and filter verbs with similarity higher than the threshold to retain their tense form but replace the verb type.

6. The personalized recommendation method for Russian oral practice based on artificial intelligence according to claim 1, characterized in that, The methods for reducing sentence complexity in step 6 include: Sentence structure simplification: Identify modifiers in sentences that are not part of the target grammatical structure, and delete relative clauses or adverbial clauses; Break the complex sentence into two simple sentences, and retain only the single sentence containing the target grammatical structure; Vocabulary difficulty reduced: Extract the root words of non-target words in the sentence and select replacement words from the basic vocabulary of the same root word; The standard for constructing the basic vocabulary database is that the vocabulary must rank in the top 20% of the frequency of occurrence in Russian elementary textbooks.

7. The personalized recommendation method for Russian oral practice based on artificial intelligence according to claim 1, characterized in that, Step 3, the dynamic learner profile includes a cold start mechanism: For users without historical data, a pre-built library of common error patterns of native Russian speakers is loaded as an initial profile. Generate a calibration problem set containing three detection dimensions: Declension probe sentences: prepositional phrases containing nouns with different declension cases; Verb aspect detection sentences: containing tense contrast structures of perfective / imperfective aspects; Stress detection words: Multisyllabic words that include rules for moving stress; Based on the distribution of user answer error types, the weights in the native speaker error pattern library are weighted and fused.

8. The personalized recommendation method for Russian oral practice based on artificial intelligence according to claim 1, characterized in that, The calculation method for the error rate descent gradient in step 7 includes: For each abnormal path weight corresponding to a syntax error type, calculate the difference between the error rate of this exercise and the error rate of the same type of exercise in the previous exercise; The change in error rate per unit time is calculated as the descent gradient based on the time interval between two practice sessions. For coupled error types with error propagation links, if the error rates of associated nodes decrease synchronously, the gradient values ​​are weighted and amplified according to the number of associated nodes.

9. The method for personalized recommendation of Russian oral practice based on artificial intelligence according to claim 1, characterized in that, The methods for dynamically adjusting the weights of abnormal paths in the knowledge graph in step 7 include: If the descent gradient exceeds the gradient threshold consecutively, the abnormal path weight of the corresponding node will be decayed to the base value. If the descent gradient of the coupling error type remains below the threshold, increase the connection depth between the node and the associated sound change rule node in the knowledge graph; The gradient threshold is determined by the average gradient change rate of users who have made effective progress in statistical historical data.

10. A personalized recommendation system for Russian oral practice based on artificial intelligence, characterized in that, It includes a memory configured to store instructions, a processor configured to retrieve the instructions from the memory, and, when executing the instructions, to implement an AI-based personalized recommendation method for Russian oral practice as described in any one of claims 1-9.

Images