Vocabulary exercise distractor generation method and electronic device

Abstract

The application discloses a vocabulary exercise distractor generation method and an electronic device. The method comprises the following steps: obtaining a target word and a correct answer to be set; determining a target sense of the target word, and recalling a candidate distractor set from a preset vocabulary based on the target sense; calculating a confusion degree score of the candidate word relative to the target sense, which combines a positive contribution of semantic closeness and a negative contribution of semantic equivalence penalty; determining a confusion degree threshold interval according to a target difficulty parameter, and screening out the candidate word falling into the interval as a target distractor; and combining the correct answer and the target distractor into an option set. The application can effectively solve the problems of difficulty in balancing the confusion and the uniqueness of the answer, rough difficulty control and poor context adaptability of polysemous words in the generation of distractors, and improve the quality stability and personalized adaptability of vocabulary test generation.

Description

Technical Field

[0001] This application relates to the field of intelligent education technology, and more specifically, to a method for generating vocabulary practice distractors and an electronic device. Background Technology

[0002] In the field of language learning and intelligent assessment, vocabulary practice is a fundamental step in building students' language proficiency, widely used in pre-class preparation, classroom training, and periodic assessments. Existing vocabulary test generation technologies primarily rely on manual question creation, rule-based template filling, and rudimentary automated recall mechanisms. Manual question creation involves professionals writing question stems and options by hand according to the teaching syllabus; while accurate, this method has low production efficiency. Rule-based automated question generation typically extracts target words from a pre-set vocabulary library, uses part-of-speech matching or simple string similarity to filter out distractors, and then fills them into a standard question template. Some natural language processing-based solutions attempt to use word vector models to calculate semantic similarity to assist in option generation, aiming to improve the scalability and response speed of question generation.

[0003] However, in the practical application of large-scale personalized teaching and high-precision assessment, existing technical solutions still face long-term challenges in terms of the controllability of the quality of generated distractors and the adaptability of difficulty. Specifically, the generated distractors often fail to maintain sufficient deceptiveness while strictly avoiding semantic equivalence with the correct answer, resulting in insufficient question differentiation or doubts about the uniqueness of the answer. At the same time, existing mechanisms are unable to make fine-grained adaptive adjustments to the difficulty of questions based on the dynamic changes in users' ability levels. This makes it easy for the generated test questions to have contextual mismatches or difficulty deviations from the expected target when dealing with the discrimination of polysemous words and complex semantic relationships, affecting the stability of automated question generation results and their applicability to teaching. Summary of the Invention

[0004] In view of the above problems, this application is proposed to provide a method and electronic device for generating vocabulary practice distractors, so as to overcome the problems of controllability of distractor generation quality and difficulty adaptability in the prior art. The specific solution is as follows:

[0005] Firstly, a method for generating vocabulary practice distractors is provided, including:

[0006] Obtain the target words for the questions to be generated and the correct answers corresponding to those target words in this question;

[0007] Determine the target meaning of the target word corresponding to the question, and recall a set of candidate interference items from a preset word library based on the target meaning;

[0008] For each candidate word in the candidate interference set, the confusion score of the candidate word relative to the target sense is calculated. The confusion score is obtained by weighted fusion of the positive contribution of the semantic proximity parameter and the negative contribution of the semantic equivalence penalty parameter. The semantic proximity parameter is used to characterize the semantic similarity between the candidate word and the target sense, and the semantic equivalence penalty parameter is used to characterize the risk of the candidate word and the target sense being synonymous.

[0009] Based on the target difficulty parameter of this question, the confusion threshold range is determined, and a set number of candidate words whose confusion scores fall within the confusion threshold range are selected from the candidate interference item set as target interference items.

[0010] The correct answer is combined with the target distractor to form a set of options for vocabulary practice.

[0011] In one possible design, in another implementation of the first aspect of the embodiments of this application, the process of obtaining the target words to be generated includes:

[0012] By combining the preset teaching objectives with the current user's ability data, the target words to be used in the questions are selected from the candidate word range of the preset word library.

[0013] In one possible design, in another implementation of the first aspect of the embodiments of this application, the process of filtering the target words to be generated includes:

[0014] Calculate the target word priority score for each candidate word, and select the target word according to the score ranking;

[0015] The target word priority score is calculated based on one or more of the following: word frequency weight, vocabulary level matching degree, topic matching degree, user's vocabulary weakness degree, review interval weight, and recent repeated question penalty.

[0016] In one possible design, in another implementation of the first aspect of this application embodiment, the process of determining the target meaning of the target word corresponding to the current question includes:

[0017] When the target word is a monosemous word, the unique meaning of the target word will be determined as the target meaning corresponding to this question.

[0018] When the target word is a polysemous word, the target meaning corresponding to this question is determined from all the meanings of the target word based on at least one of the following: the teaching objectives of this question, the context of the question stem, the topic tags of the target word, the part-of-speech requirements, and the user's historical knowledge.

[0019] In one possible design, in another implementation of the first aspect of the embodiments of this application, the semantic proximity parameter and the semantic equivalence penalty parameter are both calculated based on the semantic vector of the target sense and the semantic vector of each candidate word.

[0020] The semantic vector corresponding to the target meaning includes the target word base vector, the target meaning vector, and the question stem context vector.

[0021] In one possible design, in another implementation of the first aspect of the embodiments of this application, the confusion score is further obtained by weighted fusion based on the contribution of at least one of the following parameters:

[0022] The parameters are: word form similarity between the candidate word and the target meaning; part-of-speech consistency between the candidate word and the target word; topic consistency between the topic tag of the candidate word and the topic tag of the target word; and historical misselection rate of the candidate word.

[0023] The historical misselection rate parameter represents the percentage of times a candidate word was mistakenly selected as the correct answer by a user during historical vocabulary practice.

[0024] In one possible design, in another implementation of the first aspect of the embodiments of this application, the process of determining the confusion threshold range based on the target difficulty parameter of the question includes:

[0025] When the target difficulty parameter indicates a high difficulty setting, the confusion threshold range is determined to be within the set high score range of the confusion score;

[0026] When the target difficulty parameter indicates a low difficulty setting, the confusion threshold range is determined to be within the low confusion score range.

[0027] In one possible design, in another implementation of the first aspect of the embodiments of this application, after recalling the set of candidate distractors and before calculating the confusion score of the candidate words relative to the target sense, the method further includes:

[0028] The candidate interference set is filtered based on at least one of part-of-speech consistency, vocabulary level matching degree, and topic relevance to exclude synonyms, near-synonyms, hyponyms, and words that conflict with the context of the question stem.

[0029] In one possible design, in another implementation of the first aspect of the embodiments of this application, the target difficulty parameter is determined based on at least one of the teaching objectives, user ability parameters, and teacher configuration parameters;

[0030] The stated teaching objectives indicate the range of difficulty that this test aims to train students to achieve.

[0031] The user ability parameters include at least one of the following: the user's overall vocabulary ability score, mastery of a specified vocabulary or knowledge point, recent answer accuracy rate, continuous answer status, average answering time, and historical wrong answer records.

[0032] In one possible design, in another implementation of the first aspect of the embodiments of this application, after filtering out the target distractors and before combining the correct answer with the target distractors into a vocabulary practice option set, the method further includes:

[0033] The actual difficulty value of the test questions is calculated based on user ability parameters and test question characteristics, wherein the test question characteristics include at least one of the following: target word level, target word frequency, contextual complexity of the question stem, average confusion of distractors, and discriminability between distractors;

[0034] The actual difficulty value is compared with the target difficulty parameter. If the deviation exceeds a preset threshold, the test question features are adjusted until the actual difficulty value meets the requirements of the target difficulty parameter.

[0035] In one possible design, another implementation of the first aspect of the embodiments of this application further includes:

[0036] The set of options for the vocabulary exercise is subjected to AI quality review to obtain a review score. The review dimensions include at least one of the following: answer uniqueness, reasonableness of distractors, context matching degree, difficulty consistency, and option repetition degree.

[0037] If the review score is higher than the preset pass threshold, the option set will be marked as available and added to the database;

[0038] If the review score is lower than the preset manual review threshold, the option set and question stem will be transferred to the manual review queue.

[0039] If the review score falls between the preset pass threshold and the preset manual review threshold, then some of the interfering items will be automatically replaced, the option set will be regenerated, and the AI ​​quality review will be performed again.

[0040] In a second aspect, an electronic device is provided, comprising: a memory and a processor;

[0041] The memory is used to store programs;

[0042] The processor is configured to execute the program to implement the steps of the vocabulary practice distractor generation method as described in any of the first aspects above.

[0043] By employing the aforementioned technical solution, this application introduces the target meaning of the target word as a semantic anchor, enabling the generation of distractors to accurately match the meaning of words in specific contexts, thus avoiding semantic confusion in polysemous word processing. Through a weighted fusion mechanism of semantic proximity parameters and semantic equivalence penalty parameters, it ensures that distractors have sufficient semantic relevance to produce a deceptive effect while effectively suppressing the risk of being synonymous with the correct answer, thereby ensuring the uniqueness of the answer. Furthermore, by dynamically adjusting the confusion threshold range using a target difficulty parameter, it achieves refined control over the deceptiveness of distractors, ensuring that the difficulty of the generated questions strictly matches the set target difficulty. Vocabulary practice questions generated using this application's solution are of more stable quality and better suited to teaching needs. Attached Figure Description

[0044] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0045] Figure 1 This application provides a schematic flowchart of a vocabulary practice distractor generation method.

[0046] Figure 2 A schematic diagram of another vocabulary practice distractor generation method provided in this application embodiment;

[0047] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0048] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0049] It is understood that before using the technical solutions disclosed in the various embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0050] The vocabulary practice distractor generation method of this application can be applied to a terminal or server. The terminal can be a mobile phone, tablet computer, learning machine, teaching large screen, laptop computer, ultra-mobile personal computer (UMPC), netbook, personal digital assistant (PDA), etc., and this application embodiment does not impose any limitations on this.

[0051] The method provided in this application can be applied to various scenarios requiring automatic vocabulary test question generation, including but not limited to: daily vocabulary practice modules in language learning applications, after-class reinforcement and unit assessment systems in online education platforms, automatic question filling and updating in intelligent question bank construction systems, and automatic question generation for the vocabulary section of standardized language proficiency tests. In these scenarios, the system needs to continuously generate a large number of vocabulary multiple-choice questions with different difficulty levels and question types for learners to practice independently or use in formal assessments. Each generated question's option set, in addition to the correct answer, needs to include several reasonably misleading distractors to ensure effective differentiation. The method in this application can be applied to the above scenarios, improving the quality stability, controllability, and adaptability to the target difficulty of the distractors in the generated vocabulary multiple-choice questions.

[0052] This application provides a method for generating vocabulary practice distractors, illustrated by applying the method to a computer device. (Refer to...) Figure 1 The method for generating distractors in this vocabulary practice specifically includes the following steps:

[0053] Step S100: Obtain the target word to be used in the question and the correct answer corresponding to the target word in this question.

[0054] Obtaining the target words for the questions is fundamental to generating vocabulary exercises. These target words typically come from pre-designed teaching syllabi, textbook unit vocabulary lists, or sets of vocabulary to be reviewed recommended based on user learning profiles. The correct answer is determined based on the specific question type. For example, in an English-to-Chinese translation question, the correct answer is the Chinese definition of the target word; in a fill-in-the-blank question, the correct answer is the target word itself.

[0055] Step S110: Determine the target meaning of the target word corresponding to this question, and recall the set of candidate interference items from the preset word library based on the target meaning.

[0056] In this step, in order to take into account that the target word may be a polysemous word, and in order to accurately determine the specific semantic meaning of the polysemous word in this question, it is necessary to determine the target meaning of the target word corresponding to this question, so as to avoid the subsequent recall of candidate interference items being out of touch with the context of the question stem or causing ambiguity.

[0057] The process of determining the target meaning can be based on teaching objectives, the context of the question stem, the topic tags of the target word, part-of-speech requirements, and the user's historical knowledge of the various meanings of the word. For example, when the target word is "bank" and the question stem involves a deposit and withdrawal context, the target meaning is determined to be "n. bank"; if the question stem involves a river context, the target meaning is determined to be "n. riverbank".

[0058] After identifying the target meaning, the system retrieves a set of candidate distractors from a pre-built lexicon. The pre-built lexicon is a structured vocabulary database that stores a vast amount of vocabulary information, including text, part-of-speech, definition, example sentences, vocabulary level, and topic tags. The retrieval process uses semantic retrieval or rule matching to initially filter out words potentially related to the target meaning, which then form the set of candidate distractors for further refined filtering. For example, for the target meaning "apple (fruit)," the retrieved set of candidate distractors might include related words like "pear" and "banana," or words with similar forms like "apply" and "ample," providing foundational data for calculating the confusion score.

[0059] Step S120: For each candidate word in the candidate interference set, calculate the confusion score of the candidate word relative to the target meaning. The confusion score is obtained by weighted fusion of the positive contribution of the semantic proximity parameter and the negative contribution of the semantic equivalence penalty parameter.

[0060] The semantic proximity parameter is used to characterize the semantic similarity between the candidate word and the target meaning, and the semantic equivalence penalty parameter is used to characterize the risk of the candidate word and the target meaning being synonymous.

[0061] Confusion score is a core indicator for quantifying the effectiveness of candidate words as distractors, and its design logic aims to balance deceptiveness and safety. This score is not a single-dimensional similarity calculation, but rather a weighted fusion method that comprehensively considers the positive contribution of semantic proximity parameters and the negative contribution of semantic equivalence penalty parameters.

[0062] The semantic proximity parameter represents the distance between a candidate word and the target meaning in the semantic space. The closer the distance, the higher the similarity, and the greater its positive contribution. This allows candidate words that are semantically related to the correct answer and are likely to cause hesitation among test takers to receive higher base scores, thereby enhancing their ability to cause confusion. The semantic equivalence penalty parameter, on the other hand, represents the degree of risk of a candidate word and the target meaning forming a synonymous or near-synonymous equivalent relationship. If a candidate word and the target meaning are too close or even interchangeable without changing the meaning of the sentence, then the candidate word will lead to multiple unique answers and is considered a high-risk option. In this case, the semantic equivalence penalty parameter has a significant negative contribution, greatly reducing the confusion score of the candidate word and causing it to be excluded during the screening stage.

[0063] This scoring mechanism can be formally represented as:

[0064] Score = w1 × semantic similarity - w2 × semantic equivalence penalty

[0065] Here, "Score" represents the confusion score of the candidate word, "semantic similarity" can be understood as a semantic proximity parameter, and "w1" and "w2" are the corresponding weight coefficients. Through positive and negative weighted fusion, the system can accurately filter out high-quality distractors that are both misleading and not equivalent to the correct answer.

[0066] Step S130: Determine the confusion threshold range based on the target difficulty parameter of this question, and select a set number of candidate words whose confusion scores fall within the confusion threshold range from the candidate interference item set as target interference items.

[0067] The target difficulty parameter reflects the desired level of difficulty for this question set, and its determination can be based on teaching objectives, user ability parameters, and teacher configuration parameters. The target difficulty parameter directly determines the strategy for setting the confusion threshold range, achieving an adaptive match between the difficulty of distractors and the overall difficulty of the questions.

[0068] When the target difficulty parameter indicates high difficulty, the system tends to select distractors that are semantically closer and more deceptive, and the determined confusion threshold range is located in the high segment of the confusion score distribution. Conversely, when the target difficulty parameter indicates low difficulty, the system tends to select distractors that are semantically farther away and easier to exclude, and the confusion threshold range is set in the low segment of the score distribution, allowing candidate words with lower scores to be selected.

[0069] After determining the confusion threshold range, the system filters out a set number of candidate words from the candidate distractor set whose scores fall within that range. This set number is determined by the question type template; for example, multiple-choice questions typically require three distractors. The filtering process can be done by sorting the distractors by scores within the range or by randomly selecting from within the range, ensuring that the generated questions meet both difficulty requirements and possess randomness and diversity. The finally selected candidate words are identified as the target distractors and used to construct the final option set.

[0070] Step S140: Combine the correct answer with the target distractor to form a set of options for vocabulary practice.

[0071] After filtering the target distractors, the correct answer is combined with all the target distractors to form a set of options for vocabulary practice. This set contains the correct answer and multiple target distractors, and their order is usually randomly shuffled to avoid the cueing effect caused by positional patterns. Once the option set is generated, it can be filled into a preset question template and combined with the question stem text to generate a complete vocabulary practice test.

[0072] Through the above steps, precise recall based on the target meaning ensures semantic consistency between the distractors and the context of the question stem, resolving the semantic misalignment problem in polysemous word scenarios. When calculating the confusion score of candidate words, the positive contribution of semantic proximity and the negative contribution of semantic equivalence penalty are integrated, effectively mitigating the risk of non-unique answers while enhancing the deceptiveness of the distractors. The confusion threshold range is dynamically adjusted according to the target difficulty parameter, ensuring that the deceptiveness of the distractors accurately matches the user's current ability level and learning objectives. The various steps of the method in this embodiment collectively construct a high-quality, customizable vocabulary practice distractor generation scheme.

[0073] The method provided in this application supports multiple vocabulary practice question types within the same technical framework, including word meaning selection, contextual fill-in-the-blank, synonym discrimination, antonym discrimination, Chinese-English translation, and spelling discrimination. These question types share the same distractor generation logic: using the target meaning as the semantic benchmark, filtering distractors through confusion scores, and then combining them with the correct answer to form the option set. The differences between question types lie only in the form of the correct answer (e.g., the correct answer for word meaning selection is the definition text, the correct answer for Chinese-English translation is the target language vocabulary, and the correct answer for spelling discrimination is the correct spelling of the word) and the specific recall emphasis of the distractors (e.g., in addition to semantically similar words, spelling discrimination also recalls words with similar spellings). When it is necessary to expand to new vocabulary question types, only the rules for constructing the correct answer and the recall emphasis strategy for distractors under that question type need to be configured; the remaining algorithm logic can be reused, thus possessing good scalability and reusability.

[0074] In some embodiments of this application, an optional implementation method for obtaining the target words to be used in the question is described, including:

[0075] By combining the preset teaching objectives with the current user's ability data, the target words to be used in the questions are selected from the candidate word range of the preset word library.

[0076] Pre-set teaching objectives refer to pre-defined teaching requirements or learning standards used to guide this specific test task. These objectives can be a structured set of information encompassing multiple elements, including vocabulary level requirements stipulated in the syllabus, the coverage of a specific textbook unit, the types of knowledge points to be reinforced, and the expected level of language skills. Their construction typically occurs during the system configuration or curriculum design phase, set by teachers, curriculum researchers, or system administrators based on curriculum standards, examination syllabi, or textbook chapter headings. For example, in teaching the first unit of the third-year junior high school English textbook, the pre-set teaching objectives could be configured to cover the unit's key vocabulary, emphasize the distinction between nouns and verbs, and correspond to the CEFR A2 level vocabulary. Clearly defining these pre-set teaching objectives provides a macro-level constraint and directional guidance for subsequent target word selection.

[0077] User ability data is used to characterize the individualized characteristics of current learners in terms of vocabulary mastery, language proficiency, and learning status. Data sources include, but are not limited to, users' historical accuracy rates in past vocabulary exercises, time spent answering questions on specific words or parts of speech, statistics on consecutive correct or incorrect answers, distribution of incorrect answer records, and recent activity levels. This data undergoes systematic statistical analysis and is transformed into quantifiable ability parameters, such as overall vocabulary ability scores, mastery scores for each knowledge point, or markers of weak areas. This data can be continuously updated through a combination of real-time collection and offline batch processing, dynamically reflecting the user's learning progress and level fluctuations, thus providing data support for selecting target words that both meet teaching requirements and are tailored to the user's individual level.

[0078] When selecting target words for questions, we take into account both the pre-set teaching objectives and the user's ability data, aiming to accurately locate words from the candidate word range that meet both the requirements of the teaching syllabus and the user's current learning needs.

[0079] In one optional implementation, the process of filtering the target words to be used in the questions may include:

[0080] Calculate the target word priority score for each candidate word, and select the target word according to the score ranking.

[0081] Specifically, after determining the initial screening range based on teaching objectives and user ability data, a refined quantitative evaluation is performed on each candidate word within this range. Various feature indicators corresponding to each candidate word are extracted and converted into a single target word priority score according to a pre-set fusion algorithm. This score is a comprehensive adaptive quantification of the candidate word as the best candidate for question generation; the higher the score, the greater the value of prioritizing practice for that word in the current teaching scenario. After calculating the scores for all candidate words, they are sorted in descending order according to their target word priority scores. In practical applications, the word ranked first can be directly selected as the final target word, or a secondary extraction can be performed from a pre-set number of words (such as the top 3 or top 5) combined with a random strategy to ensure accurate screening while maintaining a certain degree of flexibility in question generation.

[0082] The target word priority score can be calculated based on one or more of the following: word frequency weight, vocabulary level matching degree, topic matching degree, user's vocabulary weakness, review interval weight, and recent repeated question penalty.

[0083] Among them, word frequency weight is used to characterize the basic importance of candidate words in actual language application. Its value can be obtained, for example, through statistical analysis of a large-scale corpus. High-frequency core words are usually given higher weights to ensure that users can prioritize mastering the most commonly used words.

[0084] The vocabulary level matching score reflects the degree of fit between the difficulty level of the candidate words and the vocabulary level required by the current teaching objectives. This value can be, for example, the inverse of the difference between the two levels. The closer the levels are, the higher the matching score.

[0085] Topic matching is used to measure the consistency between the topic tags (such as tourism, business, technology, etc.) of candidate words and the topics required by the current teaching objectives, ensuring that the generated questions can serve the current topic learning task.

[0086] A user's vocabulary weakness reflects their mastery of a candidate word. This metric can be calculated based on the user's historical answer records. If a user has a high error rate in questions containing the word or has never practiced it, their weakness score will increase significantly, thereby increasing the probability that the word will be selected as the target word.

[0087] The review interval weight is used to adapt to the forgetting curve. It is calculated based on the length of the interval between the user's last practice of the candidate word and the current time. The closer the interval is to the optimal review point indicated by the memory curve, the greater the weight value is assigned.

[0088] The penalty for repeated questions in recent questions is a negative adjustment parameter used to prevent users from frequently encountering the same words in a short period of time. Its value is usually calculated based on the number of times the word appears in recent historical questions, with an exponential decay. The more times it appears, the greater the penalty, thereby effectively reducing the priority score of the target word and preventing users from experiencing test fatigue.

[0089] In the specific calculation process, a weighted summation method can be used to aggregate the above features. For example, the target word priority score P can be formally represented as:

[0090] P = a1 × word frequency weight + a2 × vocabulary level matching degree + a3 × topic matching degree + a4 × user weakness degree + a5 × review interval weight - a6 × recent repetition penalty.

[0091] In the formula, a1 to a6 are the preset weight coefficients for the corresponding items. These coefficients can be flexibly configured according to different teaching focuses (for example, increasing the weight of word frequency in the basic reinforcement stage and increasing the weight of weakness in the pre-exam sprint stage). Through this multi-parameter collaborative calculation method, the system can scientifically and objectively select target words.

[0092] In some embodiments of this application, the process of determining the target meaning of the target word in this question is described, specifically:

[0093] When the target word is a monosemous word, the unique meaning of the target word is determined as the target meaning corresponding to this question.

[0094] When the target word is a polysemous word, the target meaning corresponding to this question is determined from all the meanings of the target word based on at least one of the following: the teaching objectives of this question, the context of the question stem, the topic tags of the target word, the part-of-speech requirements, and the user's historical knowledge.

[0095] In this embodiment, the semantic structure of the target word is identified. If the word contains only a single meaning in the preset thesaurus, it is determined to be a monosemous word. Since the semantic representation of a monosemous word is unique, its part-of-speech, topic association, and other attributes all point to this unique meaning. Therefore, there is no need to perform a complex meaning decision process. The unique meaning is directly locked as the target meaning for this question.

[0096] If a target word contains two or more independent meanings, it is considered a polysemous word. In this case, the target meaning corresponding to the question can be determined from all the meanings of the target word based on one or more dimensions of features.

[0097] For example, the teaching objectives of this question are collected and analyzed to clarify the specific vocabulary skills that the current exercise aims to test, such as distinguishing the application of the word in the financial field from its usage in daily life. Simultaneously, textual features of the question stem's context are extracted, and natural language processing techniques are used to analyze the contextual constraints on word meaning. Furthermore, the topic tag and part-of-speech requirement of the target word are used as strong filtering conditions to quickly eliminate meanings that do not match the current question type or theme. For example, when the topic tag is "computer" and the part-of-speech requirement is a noun, meanings corresponding to a computer mouse are prioritized over small animals. User's historical knowledge is used as a personalized adjustment factor; when a user's low level of mastery of a particular term or a history of errors is detected, that term is tended to be identified as the target term to strengthen the user's weak areas through practice.

[0098] Through the collaborative calculation of the above multidimensional features, the target meaning that best meets the current teaching needs, contextual constraints and user status is selected from all the meanings of the target word. This effectively avoids logical conflicts between the question stem and the options caused by mismatch of meanings in polysemous word questions, and improves the semantic rigor and teaching relevance of the test questions.

[0099] In some embodiments of this application, after recalling the candidate interference set in step S110 and before calculating the confusion score of the candidate words relative to the target meaning in step S120, a pre-filtering step may be added, specifically including:

[0100] The candidate interference set is filtered based on at least one of the following criteria: part-of-speech consistency, lexical level matching degree, and topic relevance, in order to exclude synonyms, near-synonyms, hyponyms, and words that conflict with the context of the question stem.

[0101] Specifically, a pre-filtering mechanism is introduced before performing complex confusion score calculations. This mechanism aims to remove words from the candidate distractor set that are obviously not reasonable for interference or may lead to multiple unique answers, thereby improving the processing efficiency of subsequent calculations and ensuring the rigor of the generated questions.

[0102] Among these, part-of-speech consistency requires that the part of speech of the candidate words match the part of speech of the target meaning. For example, when the target meaning is a noun, candidate words in the form of verbs or adjectives are eliminated. Vocabulary level matching requires that the difficulty level of the candidate words be appropriate to the vocabulary difficulty level specified in the teaching objectives to avoid words that are too difficult or too easy. Theme relevance is used to ensure that the candidate words and the target words belong to the same semantic theme domain. For example, when apple is used as the target word, words belonging to the category of fruit or food are retained first, while words belonging to the category of transportation or technology are excluded.

[0103] Synonyms are words that can be substituted for each other and are semantically equivalent in a specific context. In vocabulary exercises, if a distractor is synonymous with the correct answer, the answer will not be unique, rendering the question invalid. The system identifies candidate words that are synonyms with the target meaning by comparing them with predefined synonym mapping relationships in the vocabulary knowledge base or calculating semantic similarity scores, and removes them directly from the set.

[0104] Synonyms and equivalents refer to words that are semantically very similar and are easily misjudged as correct answers in non-strictly differentiated contexts. Although these words may have subtle semantic differences, they often pose a high risk of interference in practice questions lacking clear analytical context, and may even lead to ambiguity. By analyzing the semantic closeness parameter between candidate words and the target meaning, if this parameter exceeds a preset synonym risk threshold, the candidate word is identified as a synonym and equivalent and excluded. This process effectively avoids scoring disputes caused by semantically too similar distractors.

[0105] Hypernyms and hyponyms encompass the logical relationship of inclusion and being included, where a hypernym (e.g., fruit) includes a hyponym (e.g., apple). If the target meaning is a hyponym and the distractors are corresponding hypernyms, in some multiple-choice contexts, the hypernym might be logically considered correct or not incorrect due to its overly broad scope, thus interfering with the validity of the test. By utilizing a pre-defined hierarchical knowledge graph or classification tree structure, candidate words with a direct hypernym / hypernym inheritance relationship with the target meaning are identified and eliminated to ensure logical equivalence between options.

[0106] Contextual conflict in the question stem refers to situations where, although candidate words are related to the target word, they cannot naturally integrate into the specific question context, sentence structure, or pragmatic scenario, or even create logical contradictions. For example, the question stem might describe a mechanical operation process, while the candidate word is an abstract noun used only to describe an emotional state; even if they have the same part of speech, there is still a clear contextual conflict. By analyzing the compatibility between the context vector of the question stem and the semantic vector of the candidate words, or by using a pre-trained language model to judge the fluency and logical rationality of the candidate words after they are added to the question stem, conflicting words that lead to semantic breaks or logical paradoxes can be eliminated.

[0107] By employing the aforementioned multi-dimensional joint filtering, the quality of candidate words entering the subsequent confusion scoring stage is significantly optimized, reducing the consumption of ineffective computing resources. Simultaneously, it fundamentally avoids the risk of non-unique answers due to synonyms, near-synonyms, or hierarchical relationships, as well as comprehension obstacles caused by contextual conflicts, providing a solid data foundation for generating a high-quality option set.

[0108] In some embodiments of this application, the process of calculating the confusion score is further described.

[0109] The confusion score can be obtained by weighted fusion of the positive contribution of the semantic proximity parameter and the negative contribution of the semantic equivalence penalty parameter.

[0110] The semantic proximity parameter and the semantic equivalence penalty parameter can both be calculated based on the semantic vector of the target sense and the semantic vector of each candidate word.

[0111] In some optional implementations, the semantic vector corresponding to the target sense can adopt a high-dimensional vector structure constructed by multi-layer feature fusion, rather than a simple single-word representation. Specifically, the semantic vector corresponding to the target sense includes the target word base vector, the target sense vector, and the stem context vector.

[0112] The target word base vector is used to represent the basic semantic distribution information of the target word. This vector can be obtained by extracting the static or dynamic word embeddings of the target word through a pre-trained language model.

[0113] Target semantic vectors are used to fine-grainedly characterize the specific meaning of a target word in a polysemous context. When the target word is polysemous, the system obtains the explanatory text, example sentences, etc., corresponding to the determined unique target semantic term, and inputs them into the semantic model for encoding, thereby obtaining a semantic vector focused on the current specific meaning.

[0114] The context vector of the question stem is used to introduce the contextual constraints of this question. This vector is extracted from the semantic model by inputting the question stem text containing the target word. It encodes the specific usage of the target word under specific sentence structures, grammatical environments and collocations, enabling semantic computing to be aware of the actual application scenario.

[0115] The target word base vector, target meaning vector, and question stem context vector mentioned above can be fused through weighted concatenation or attention mechanism to form the meaning semantic vector corresponding to the target meaning.

[0116] The semantic vector of a candidate word can refer to the semantic representation of each candidate word in a pre-set lexicon. The generation method of this vector is consistent with the basic representation of the target word. For example, the output layer parameters of the same pre-trained model are used for mapping to ensure that the two vector spaces are comparable.

[0117] In some alternative embodiments, the confusion score can also be obtained by weighted fusion based on the contribution of at least one of the following parameters:

[0118] The parameters include the word form similarity between candidate words and target terms, the part-of-speech consistency between candidate words and target words, the topic consistency between candidate words and target words, and the historical misselection rate of candidate words.

[0119] The lemmatism similarity parameter characterizes the similarity between candidate words and their corresponding target meanings in terms of spelling. This parameter is not a simple character matching result, but a structured numerical value calculated based on edit distance, longest common subsequence, or glyphic features, reflecting the visual likelihood of confusion. In vocabulary practice scenarios, the lemmatism similarity parameter is used to filter out words with highly similar spellings but different meanings. For example, for the target word "apple," words with similar spellings such as "apply" and "ample" are recalled to specifically test the user's ability to accurately distinguish word spellings.

[0120] The part-of-speech consistency parameter characterizes the degree of agreement between the part-of-speech attributes of candidate words and target words. It is determined by comparing the part-of-speech tags (e.g., noun, verb, adjective) of candidate words in a pre-defined thesaurus with those of the target word. Maintaining part-of-speech consistency is a fundamental prerequisite for distractors to be effective in grammatically correct sentences or options; if the part-of-speech differences are too large, distractors are easily eliminated. Introducing the part-of-speech consistency parameter aims to impose grammatical constraints on the confusion calculation, ensuring that the selected distractors are at least grammatically reasonable potential options, and avoiding artificially lowering the difficulty of the questions due to part-of-speech conflicts.

[0121] The topic consistency parameter characterizes the relevance between the topic categories of candidate words and the target word. Based on the topic tags (e.g., fruit, finance, transportation) labeled for each word in a pre-defined thesaurus, the cosine similarity between the topic vectors of candidate words and target words is calculated, or it is determined whether the tags belong to the same hierarchical classification system, thus generating the topic consistency parameter. Building upon a high semantic similarity, the topic consistency parameter further limits the scope of distractors, ensuring they fall within the same semantic field or topic range as the target word. This approach increases the deceptiveness of the questions through semantic analysis of words within the same topic, effectively preventing irrelevant cross-topic words from being selected.

[0122] The historical misselection rate parameter represents the percentage of candidate words that users mistakenly selected as the correct answer in historical vocabulary practice. By retrieving past practice records from the backend database, the frequency of candidate words appearing as distractors and being incorrectly selected as the target word is statistically analyzed. This frequency is then divided by the total number of times the candidate word appeared as a distractor to obtain the historical misselection rate parameter. Introducing this parameter into the distractor generation process allows for the feedback of real user behavior data to the system, quantifying the actual confusing intensity of specific candidate words for different user groups. Words exhibiting high misselection rates in historical data often contain common cognitive biases; including these in the confusion scoring system enables the generated distractors to more accurately target learners' knowledge gaps.

[0123] The aforementioned parameters characterize the confusion level of candidate words from four dimensions: spelling, grammatical attributes, semantic theme, and empirical feedback. In calculating the confusion score, corresponding weight coefficients are assigned to the word form similarity parameter, part-of-speech consistency parameter, theme consistency parameter, and historical misselection rate parameter. The weighted values ​​of these parameters are then fused with the original semantic proximity parameter and semantic equivalence penalty parameter. This multi-parameter fusion mechanism generates interference items with superior overall performance, ensuring the rationality of the interference items at the linguistic knowledge level while leveraging historical data to enhance their effectiveness in addressing real learning difficulties.

[0124] In this embodiment, the confusion score of candidate words can be expressed as:

[0125] Score = w1 × semantic similarity + w3 × word form similarity + w4 × part-of-speech consistency + w5 × topic consistency + w6 × historical misselection rate - w2 × semantic equivalence penalty

[0126] Here, w1 to w6 are the corresponding weight coefficients. Through positive and negative weighted fusion, the system can accurately filter out high-quality distractors that are both deceptive and not equivalent to the correct answer.

[0127] In some embodiments of this application, a method for determining the target difficulty parameter of the question is described.

[0128] The target difficulty parameter can be determined based on at least one of the following: teaching objectives, user ability parameters, and teacher configuration parameters.

[0129] The learning objectives indicate the range of difficulty that this set of questions aims to train students in. As a macro-level guiding parameter, the learning objectives define the difficulty level of this set of questions, such as different levels set according to the teaching syllabus, such as basic consolidation, ability enhancement, or advanced challenge.

[0130] User ability parameters include at least one of the following: overall vocabulary ability score, mastery of specified vocabulary or knowledge points, recent answer accuracy rate, continuous answering status, average answering time, and historical wrong answer records.

[0131] User ability parameters serve as a basis for micro-adaptation, quantifying the user's current vocabulary cognition level and answering status to achieve dynamic matching between question difficulty and user ability.

[0132] A user's overall vocabulary proficiency score reflects their average mastery level across all tested vocabulary words, while the mastery level of specific vocabulary words or knowledge points represents the depth of their understanding of those words or points. Recent answer accuracy and historical error records reflect the distribution of accuracy and weaknesses in recent practice; continuous answering status reflects the user's current confidence and focus, for example, consecutive correct answers suggest increasing the difficulty, while consecutive incorrect answers suggest decreasing the difficulty; average answering time indirectly reflects the user's proficiency in vocabulary responses.

[0133] The teacher configuration parameters provide a manual intervention channel, allowing teachers to directly override or adjust the difficulty parameters automatically calculated by the system according to specific teaching scenarios or special student situations, so as to ensure the flexibility and controllability of teaching.

[0134] By integrating the aforementioned teaching objectives, user ability parameters, and teacher configuration parameters, a multi-dimensional and quantifiable difficulty decision model is constructed. This provides precise control input for subsequently determining the confusion threshold range based on the target difficulty parameters. This processing method ensures that the generated vocabulary practice questions not only meet the teaching progress requirements but also accurately match the user's actual ability level.

[0135] In some embodiments of this application, an optional implementation method is further described to determine the confusion threshold range based on the target difficulty parameter of the current question.

[0136] When the target difficulty parameter indicates that a high difficulty level is set, the confusion threshold range is determined to be within the set high score range of the confusion score.

[0137] When the target difficulty parameter indicates that the difficulty is set to low, the confusion threshold range is determined to be within the set low score range of the confusion score.

[0138] Setting high-score and low-score intervals is a way of dividing the numerical range of confusion scores, used to establish a mapping relationship between abstract teaching difficulty indicators and calculable algorithm parameters. A high-score interval can refer to a region within the confusion score range where the values ​​are relatively high, representing candidate words that are semantically very similar to or easily confused with the target meaning; a low-score interval can refer to a region within the confusion score range where the values ​​are relatively low, representing candidate words that are semantically significantly different from or relatively easy to distinguish with the target meaning.

[0139] When executing the threshold range determination logic, the target difficulty parameter configured for this question is first read. This target difficulty parameter is the expected difficulty level determined by the teaching objectives, the user's ability model, or the teacher's manual setting. For example, it can be a discrete level of high, medium, and low, or a continuous value between 0 and 1.

[0140] If the target difficulty parameter's value or level corresponds to a high difficulty range, it is determined that highly deceptive distractors need to be generated. In this case, a preset score-difficulty mapping table is retrieved, and the confusion threshold range is positioned within a set high-score range. For example, the set high-score range could be the top 20% of scores, i.e., [Score_80, Score_100], where Score_80 represents the 80th percentile of the score distribution. Candidate words selected within this range have higher semantic proximity parameters relative to the target meaning, indicating that these words are semantically closer to the correct answer or have more similar word forms, parts of speech, and thematic features, thereby increasing the cognitive load on the user to identify the correct answer.

[0141] If the target difficulty parameter indicates a low difficulty setting, then the interference intensity of the distractors should be reduced. In this case, the confusion threshold range is set to the low score range. For example, the low score range could be the bottom 30% of the score rankings, i.e., [Score_0, Score_30]. Candidate words selected within this range have low semantic proximity parameters or are suppressed by strong semantic equivalence penalties, indicating that these words differ significantly from the correct answer in semantic features, word structure, or topic category. Users can easily eliminate incorrect options through semantic analysis or morphological comparison.

[0142] Through the above mapping mechanism, the target difficulty parameter is precisely transformed into specific numerical screening conditions, so that the confusion level of the target interference items selected from the candidate interference item set is consistent with the preset teaching difficulty.

[0143] It should be noted that the above only exemplifies one possible method for classifying the difficulty level of the target difficulty parameter, and the corresponding method for classifying the confusion threshold range. In addition, other methods for classifying the difficulty level of the target difficulty parameter can also be used, such as dividing it into three levels: low, medium, and high.

[0144] In some embodiments of this application, a difficulty adaptive adjustment mechanism is further provided. Specifically, after the target distractor items are selected in step S130 and before step S140 combines the correct answer with the target distractor items to form a set of options for vocabulary practice, the following steps may also be included:

[0145] S1. Calculate the actual difficulty value of the test questions based on the user's ability parameters and the test question characteristics, wherein the test question characteristics include at least one of the following: target word level, target word frequency, context complexity of the question stem, average confusion of distractors, and discrimination between distractors.

[0146] Specifically, after identifying the target distractors, the test difficulty verification stage begins. This stage aims to quantitatively assess the actual difficulty level of the generated test questions in terms of objective attributes. The calculation of the actual difficulty value of the test questions can comprehensively consider both static vocabulary attributes and the characteristics of dynamically generated distractors.

[0147] Target word level and target word frequency reflect the cognitive load of the vocabulary itself. Generally, the higher the vocabulary level and the lower the word frequency, the greater the corresponding difficulty contribution value. Contextual complexity of the question stem is used to characterize the grammatical complexity of the sentence structure and the degree of abstraction in semantic understanding. It can be measured by indicators such as the depth of the syntax tree, the number of clauses, or the perplexity of the pre-trained language model.

[0148] The average confusion level of distractors is a core indicator for measuring the overall deceptiveness of the answer choices. This parameter is obtained by statistically analyzing the central trend (such as the mean or median) of the distribution of the confusion scores of each candidate word calculated in the preceding steps. When the average confusion level of distractors is high, it means that most distractors are semantically or morphologically close to the correct answer, and the objective difficulty of the test question increases accordingly.

[0149] The discriminant strength between distractors is used to assess the dispersion within the option set. It is determined by calculating indicators such as semantic distance and word form difference between pairs of distractors, and taking their average negative value or similarity negative value. If the distractors are too semantically similar (e.g., multiple options are synonyms), the discriminant strength between distractors decreases, leading to an increase in the difficulty of the question.

[0150] User ability parameters serve as a benchmark in this calculation process. They map the objective numerical values ​​of test item features to a difficulty scale that references user ability. Specifically, a feature vector is constructed that includes target word level, reciprocal word frequency, question stem complexity score, average confusion level of distractors, and distractor discrimination level. This vector is then input into a pre-trained difficulty prediction regression model. This model, trained on massive historical user response data, outputs an actual test item difficulty value with the same dimensions as the user's ability parameters. For example, for a user with an English vocabulary proficiency score of 0.5 (normalized), if the test item feature vector shows a high target word level and extremely high distractor confusion level, the calculated actual test item difficulty value might be 0.8, indicating that the test item is highly challenging for the current user.

[0151] S2. Compare the actual difficulty value with the target difficulty parameter. If the deviation exceeds the preset threshold, adjust the test characteristics until the actual difficulty value meets the requirements of the target difficulty parameter.

[0152] After calculating the actual difficulty value of the test questions, it is compared with the pre-set target difficulty parameter to determine whether the currently generated test questions meet the expected difficulty of teaching or adaptive practice. The target difficulty parameter is a benchmark value determined comprehensively based on the expected training range indicated by the teaching objectives, user ability parameters (such as overall vocabulary ability value, recent answer accuracy rate), and teacher configuration parameters.

[0153] This comparison process establishes a closed-loop feedback control mechanism. The preset threshold serves as the tolerance boundary for determining whether the deviation is acceptable, and its setting is based on factors including the accuracy requirements of the teaching scenario. For example, for a refined adaptive learning scenario, the preset threshold can be set to 0.05 or less; while for a coarse-grained practice mode, the preset threshold can be appropriately relaxed to 0.1 or 0.15.

[0154] If the comparison results show that the difference between the actual difficulty value and the target difficulty parameter does not exceed the preset threshold, the current question characteristics are deemed reasonable and no adjustment is needed; the process can proceed directly to the subsequent option set combination step. If the deviation exceeds the preset threshold, for example, if the actual difficulty value is significantly higher than the target difficulty parameter, it indicates that the question is too difficult and exceeds the user's current problem-solving ability. Conversely, if the actual difficulty value is significantly lower than the target difficulty parameter, it indicates that the question is too easy and cannot provide effective training.

[0155] To address the aforementioned deviations, a dynamic adjustment process for test item features is initiated. Adjusting test item features specifically involves correcting key factors affecting difficulty. For example, when the actual difficulty value is too high, it can be adjusted by reducing the average confusion level of distractors. This can be achieved by replacing one or more distractors with the highest confusion scores with words that are semantically distant; or by reducing the complexity of the question stem context, such as rewriting long sentences containing multiple clauses into simpler sentences. Conversely, when the actual difficulty value is too low, the difficulty can be increased by raising the average confusion level of distractors (e.g., by introducing more similar-looking words or synonyms), increasing the target word level, or introducing more complex question stem contexts. Furthermore, for cases where the differentiation between distractors is low, optimization can be achieved by removing highly homogeneous distractors and adding significantly differentiated vocabulary. After the adjustment is completed, the actual difficulty value of the test items corresponding to the updated test item features is recalculated and compared again with the target difficulty parameter. Iterative adjustments are made until the deviation converges to a preset threshold range. This feedback-based closed-loop adjustment mechanism ensures that the final output test questions accurately match the user's subjective ability level and teaching objectives in terms of objective attributes.

[0156] In some embodiments of this application, a process for AI-based review of generated test questions is further provided, specifically:

[0157] The generated vocabulary exercise options set is subjected to AI quality review to obtain a review score. The review dimensions include at least one of the following: answer uniqueness, distractor rationality, context matching degree, difficulty consistency, and option repetition degree.

[0158] This application can utilize an AI quality review module to perform a comprehensive, automated quality assessment of the vocabulary practice options generated in the preceding steps. This assessment process is based on a composite judgment mechanism that combines deep semantic understanding with rule-based logic. In terms of answer uniqueness, semantic similarity calculation and logical reasoning are used to confirm whether there are any distractors in the option set that are synonymous or equivalent to the question stem, in addition to the correct answer, to avoid multiple correct answers being selected. In terms of distractor rationality, the reasonableness of distractors is assessed to determine whether they are valid English words, whether there are spelling errors, and whether they deviate too much from or are too closely related to the target word semantically (e.g., their hyponymous relationship is too close). In terms of context matching, the semantic coherence between the context of the question stem and each option (including the correct answer and distractors) is checked to ensure that the options are grammatically and logically compatible with the context constructed by the question stem. In terms of difficulty consistency, the actual difficulty of the test is estimated based on the target word level, question stem complexity, and distractor confusion, and compared with the target difficulty parameters set for this test to determine whether the two fall within the allowable error range. In terms of option repetition, the repetition, inclusion relationship, or excessive similarity between the texts of each option is checked to ensure that the options have sufficient differentiation.

[0159] The judgment results of the one or more review dimensions described above are comprehensively converted into a quantitative review score. For example, the scoring process may adopt a weighted summation algorithm, which weights and sums the scores of the one or more review dimensions to obtain a total review score.

[0160] If the review score is higher than a preset passing threshold, the option set is marked as available and stored in the database.

[0161] The preset passing threshold is a dividing line for measuring whether the quality of a test question meets the direct release standard. When the calculated review score S satisfies S>T_pass (where T_pass is the preset passing threshold), it indicates that the test question has met the quality requirements set by the system in terms of answer accuracy, option logic, difficulty adaptability, compliance and other aspects. At this time, the status of the option set is marked as available or reviewed and passed, and the corresponding complete test question data (including the question stem, target word, correct answer, target distractors and analysis information) is persistently stored in the test question library. The test questions stored in the library can be directly called by the online learning system, practice platform or examination system, and pushed to end users for answering. By setting a reasonable preset passing threshold, the proportion of automatic storage into the library can be maximized on the premise of ensuring quality, and the frequency of manual intervention can be reduced.

[0162] If the review score is lower than a preset manual review threshold, the option set and the question stem are transferred to a manual review queue.

[0163] The preset manual review threshold is the bottom-line standard for identifying low-quality or suspected abnormal test questions. When the review score S satisfies S<T_manual (where T_manual is the preset manual review threshold, and T_manual<T_pass), it indicates that the test question has obvious quality problems, for example, the answer may not be unique, the distractor has a serious error or is completely disconnected from the question stem, and such problems are usually difficult to solve through simple automatic correction logic. At this time, a manual review task is generated, and the option set and the corresponding question stem are packaged and pushed into the manual review queue. Manual reviewers can view the detailed content of these test questions, the specific error items pointed out by AI review and the reasons for low scores through a dedicated workbench interface, and perform manual correction or direct discard. This mechanism ensures that even when the automated algorithm fails, the quality of test questions flowing to end users can still be guaranteed through manual checks, so as to avoid serious teaching accidents.

[0164] If the review score is between the preset passing threshold and the preset manual review threshold, part of the distractors are automatically replaced to regenerate the option set, and the AI quality review is performed again.

[0165] When the review score S satisfies T_manual ≤ S ≤ T_pass, it indicates that the test quality is in an intermediate state, acceptable but not excellent. These types of questions usually do not have hard errors, but there is still room for improvement in areas such as the deceptiveness of distractors, the accuracy of difficulty, or the differentiation of options. In this case, instead of discarding the question directly, an automatic optimization process should be initiated. The specific steps could be:

[0166] The original question stem and correct answer remain unchanged. The system removes target distractors that cause low scores from the candidate distractor set (e.g., distractors with excessively low or high confusion scores, or those semantically too similar to the correct answer). New candidate words are then recalled from the lexicon, confusion scores are calculated, and better distractors are selected to replace them, forming a new option set. After replacement, the system performs AI quality review on the newly generated option set. This closed-loop iterative mechanism of generation-review-feedback-correction significantly improves the final quality of the questions, bringing them closer to or reaching the preset passing threshold, thereby increasing the overall question entry rate and reducing the workload of manual review. To prevent the system from entering an infinite loop, a maximum retry limit can be set. If the number of retries exceeds the limit and the question still fails to pass review, it is transferred to the manual review queue.

[0167] Through the aforementioned three-tiered classification and processing strategy, an intelligent quality filtering system has been constructed. High-scoring questions are released directly, low-scoring questions are manually filled in, and questions of intermediate scores are automatically optimized. This tiered processing method ensures both the efficiency of question generation in a large-scale production environment and the reliability and pedagogical applicability of the output questions.

[0168] In some embodiments of this application, user answer result data can be further collected, and the user ability parameters of the corresponding user and the historical misselection rate parameters of each candidate word can be updated synchronously for subsequent interference item generation and difficulty adjustment.

[0169] Reference Figure 2 This provides an alternative implementation process for generating distractors in vocabulary practice.

[0170] First, obtain the initial configuration parameters, including the teaching objectives, user ability data, and question type configuration information for this question.

[0171] The teaching objectives are used to limit the vocabulary range, difficulty level, and thematic direction of the questions. User ability data is used to support personalized adaptive question generation. Question type configuration is used to determine the type of questions in the final output test.

[0172] Furthermore, candidate words are retrieved from the vocabulary and multi-layer semantic vectors are generated.

[0173] Based on the initial configuration range constraints, a set of candidate words is recalled from a pre-set vocabulary database. A pre-trained language model is invoked to generate corresponding basic word vectors, semantic vectors of senses, and semantic vectors of context for the candidate words, constructing a multi-granular semantic representation system to provide a semantic benchmark for subsequent target word selection and interference calculation.

[0174] Further, the target words and their corresponding target meanings are selected and identified.

[0175] By combining dimensions such as word frequency, vocabulary level, topic relevance, and user vocabulary mastery, the target word priority score for each candidate word is calculated, and the target word for this question is selected according to the score ranking. Taking into account teaching objectives, question context, topic tags, part-of-speech requirements, and user historical knowledge, the target meaning corresponding to this question is determined from all meanings of the target word, thus locking in the semantic benchmark for this test question.

[0176] Furthermore, candidate interference items are recalled and preliminary screening is performed.

[0177] Using the defined target meaning as the semantic benchmark, a set of candidate interference terms is recalled from the candidate word set. The set of candidate interference terms is pre-filtered according to rules such as part-of-speech consistency, word level matching degree, topic relevance, and synonym constraint, eliminating candidate words with mismatched parts of speech, level differences exceeding the preset range, topic mismatch, and obvious risk of synonym equivalence, thus narrowing the scope of subsequent calculations.

[0178] Furthermore, a confusion score is calculated and a target difficulty parameter is determined.

[0179] For each of the filtered candidate distractors, a confusion score is calculated. The confusion score integrates the positive contributions of multiple dimensions such as semantic proximity and word form similarity, and adds the negative contribution of semantic equivalence penalty, and obtains the final score through weighted fusion. At the same time, the target difficulty parameter for this question is determined by combining teaching objectives, user ability data, and teacher configuration parameters, which is used to guide the subsequent selection of distractors.

[0180] Furthermore, target distractors are filtered according to the target difficulty level, and the complete test questions are generated by rendering.

[0181] Based on the target difficulty parameter, a corresponding confusion threshold range is determined. A predetermined number of candidate words whose scores fall within this range are selected from the candidate distractor set and used as the final target distractors. The correct answer and the target distractors are combined into an option set. Using a preset template for the corresponding question type, a complete vocabulary test question containing the question stem, options, standard answer, and answer explanation is generated.

[0182] Furthermore, AI-based quality review and result triage processing are implemented.

[0183] The generated complete test questions undergo AI quality review, with review dimensions including answer uniqueness, reasonableness of distractors, context matching, and consistency of difficulty, and the corresponding review score is output.

[0184] If the review score reaches the preset passing threshold, the question is deemed qualified, stored in the question bank or marked as pending release, and the process ends.

[0185] If the review score does not meet the standard, a tiered processing will be performed based on the score result: if the score is in the middle range, the interference items or difficulty parameters will be automatically adjusted and the generation will be retried; if the score is lower than the manual review threshold, or the number of automatic retries exceeds the limit, it will be transferred to the manual review queue, and the process will end after the processing is completed.

[0186] This application provides a vocabulary practice distractor generation device, which includes some or all of the following modules:

[0187] The system comprises a vocabulary database, a semantic representation module, a target word selection module, a distractor generation module, an adaptive difficulty module, an AI quality review module, a human review module, and a question bank. These modules collaborate via data interfaces or service calls to generate vocabulary test questions.

[0188] Vocabulary database: Used to store basic data such as vocabulary text, part of speech, definition, example sentences, vocabulary level, topic tags, applicable grade level, and historical question records.

[0189] Semantic Representation Module: Used to call a pre-trained language model to convert words, senses, and context into high-dimensional semantic vectors.

[0190] Target word selection module: Used to obtain target words for the questions to be generated. For example, it combines word frequency, vocabulary level, topic tags, teaching scope and user mastery to filter target words suitable for generating test questions, obtain the correct answer corresponding to the target word for this question, and determine the target meaning of the target word for this question.

[0191] The distractor generation module is used to recall a set of candidate distractors from the vocabulary based on the target meaning, calculate the confusion score of the candidate words relative to the target meaning, determine the confusion threshold range according to the target difficulty parameter of this question, and select a set number of candidate words whose confusion scores fall within the confusion threshold range as target distractors from the set of candidate distractors.

[0192] The difficulty adaptive module is used to calculate the actual difficulty value of the test questions based on the user's ability parameters and test question features. The test question features include at least one of the following: target word level, target word frequency, context complexity of the question stem, average confusion of distractors, and discrimination between distractors. The actual difficulty value is compared with the target difficulty parameter. If the deviation exceeds the preset threshold, the test question features are adjusted until the actual difficulty value meets the requirements of the target difficulty parameter.

[0193] Question Rendering Module: This module combines the correct answer with the target distractor to form a set of options for vocabulary practice. It then fills the set of options into a preset question template and combines it with the question text to generate a complete vocabulary practice test.

[0194] AI quality review module: used to automatically review the uniqueness of answers, the rationality of distractors, the consistency of difficulty, and the quality of question expression in vocabulary practice questions.

[0195] Manual review module: Used to receive questions whose AI review scores are below the threshold or which fail after multiple automatic retries.

[0196] This application also provides an electronic device in its embodiments. (See reference...) Figure 3 The diagram illustrates a structural schematic suitable for implementing the electronic device in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, tablets, learning machines, large-screen teaching displays, desktop computers, etc. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0197] like Figure 3 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 1, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 2 or a program loaded from a storage device 8 into a random access memory (RAM) 3, to implement the vocabulary practice interference item generation method of the foregoing embodiments of this application. When the electronic device is powered on, the RAM 3 also stores various programs and data required for the operation of the electronic device. The processing unit 1, ROM 2, and RAM 3 are interconnected via a bus 4. An input / output (I / O) interface 5 is also connected to the bus 4.

[0198] Typically, the following devices can be connected to I / O interface 5: input devices 6 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 7 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 8 including, for example, memory cards, hard drives, etc.; and communication devices 9. Communication device 9 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have instead.

[0199] This application also provides a computer program product, including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the vocabulary practice distractor generation methods provided in this application.

[0200] This application also provides a computer-readable storage medium that carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the vocabulary practice distractor generation methods provided in this application.

[0201] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0202] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0203] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0204] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0205] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

Claims

1. A method for generating vocabulary practice distractors, characterized in that, include: Obtain the target words for the questions to be generated and the correct answers corresponding to those target words in this question; Determine the target meaning of the target word corresponding to the question, and recall a set of candidate interference items from a preset word library based on the target meaning; For each candidate word in the candidate interference set, the confusion score of the candidate word relative to the target sense is calculated. The confusion score is obtained by weighted fusion of the positive contribution of the semantic proximity parameter and the negative contribution of the semantic equivalence penalty parameter. The semantic proximity parameter is used to characterize the semantic similarity between the candidate word and the target sense, and the semantic equivalence penalty parameter is used to characterize the risk of the candidate word and the target sense being synonymous. Based on the target difficulty parameter of this question, the confusion threshold range is determined, and a set number of candidate words whose confusion scores fall within the confusion threshold range are selected from the candidate interference item set as target interference items. The correct answer is combined with the target distractor to form a set of options for vocabulary practice.

2. The method according to claim 1, characterized in that, The process of obtaining the target words for the questions includes: By combining the preset teaching objectives with the current user's ability data, the target words to be used in the questions are selected from the candidate word range of the preset word library.

3. The method according to claim 2, characterized in that, The process of selecting target words for questions includes: Calculate the target word priority score for each candidate word, and select the target word according to the score ranking; The target word priority score is calculated based on one or more of the following: word frequency weight, vocabulary level matching degree, topic matching degree, user's vocabulary weakness degree, review interval weight, and recent repeated question penalty.

4. The method according to claim 1, characterized in that, The process of determining the target meaning of the target word corresponding to the target meaning of this question includes: When the target word is a monosemous word, the unique meaning of the target word will be determined as the target meaning corresponding to this question. When the target word is a polysemous word, the target meaning corresponding to this question is determined from all the meanings of the target word based on at least one of the following: the teaching objectives of this question, the context of the question stem, the topic tags of the target word, the part-of-speech requirements, and the user's historical knowledge.

5. The method according to claim 1, characterized in that, The semantic proximity parameter and the semantic equivalence penalty parameter are both calculated based on the semantic vector of the target sense and the semantic vector of each candidate word. The semantic vector corresponding to the target meaning includes the target word base vector, the target meaning vector, and the question stem context vector.

6. The method according to claim 1, characterized in that, The confusion score is also obtained by weighted fusion based on the contribution of at least one of the following parameters: The parameters are: word form similarity between the candidate word and the target meaning; part-of-speech consistency between the candidate word and the target word; topic consistency between the topic tag of the candidate word and the topic tag of the target word; and historical misselection rate of the candidate word. The historical misselection rate parameter represents the percentage of times a candidate word was mistakenly selected as the correct answer by a user during historical vocabulary practice.

7. The method according to claim 1, characterized in that, The process of determining the confusion threshold range based on the target difficulty parameters of this question includes: When the target difficulty parameter indicates a high difficulty setting, the confusion threshold range is determined to be within the set high score range of the confusion score; When the target difficulty parameter indicates a low difficulty setting, the confusion threshold range is determined to be within the low confusion score range.

8. The method according to claim 1, characterized in that, After recalling the candidate distractor set and before calculating the confusion score of the candidate word relative to the target sense, the method further includes: The candidate interference set is filtered based on at least one of part-of-speech consistency, vocabulary level matching degree, and topic relevance to exclude synonyms, near-synonyms, hyponyms, and words that conflict with the context of the question stem.

9. The method according to claim 1, characterized in that, The target difficulty parameter is determined based on at least one of the teaching objectives, user ability parameters, and teacher configuration parameters. The stated teaching objectives indicate the range of difficulty that this test aims to train students to achieve. The user ability parameters include at least one of the following: the user's overall vocabulary ability score, mastery of a specified vocabulary or knowledge point, recent answer accuracy rate, continuous answer status, average answering time, and historical wrong answer records.

10. The method according to claim 1, characterized in that, After filtering out the target distractors, and before combining the correct answer with the target distractors to form the vocabulary practice option set, the process also includes: The actual difficulty value of the test questions is calculated based on user ability parameters and test question characteristics, wherein the test question characteristics include at least one of the following: target word level, target word frequency, contextual complexity of the question stem, average confusion of distractors, and discriminability between distractors; The actual difficulty value is compared with the target difficulty parameter. If the deviation exceeds a preset threshold, the test question features are adjusted until the actual difficulty value meets the requirements of the target difficulty parameter.

11. The method according to any one of claims 1-10, characterized in that, Also includes: The set of options for the vocabulary exercise is subjected to AI quality review to obtain a review score. The review dimensions include at least one of the following: answer uniqueness, reasonableness of distractors, context matching degree, difficulty consistency, and option repetition degree. If the review score is higher than the preset pass threshold, the option set will be marked as available and added to the database; If the review score is lower than the preset manual review threshold, the set of options and the question stem will be transferred to the manual review queue. If the review score falls between the preset pass threshold and the preset manual review threshold, then some of the interfering items will be automatically replaced, the option set will be regenerated, and the AI ​​quality review will be performed again.

12. An electronic device, characterized in that, include: Memory and processor; The memory is used to store programs; The processor is configured to execute the program to implement each step of the vocabulary practice distractor generation method as described in any one of claims 1 to 11.

Images