Intelligent content pushing and interaction method and device

Abstract

The invention provides an intelligent content pushing and interaction method and device, and relates to the field of computers. The method comprises the following steps: constructing a target user portrait based on multi-dimensional user data; based on the target user portrait and the historical learning data, predicting a forgetting time point of each learning unit, and obtaining predicted memory data; wherein in the prediction process, an Ebbinghaus attenuation correction factor is introduced to dynamically adjust predicted data; the Ebbinghaus attenuation correction factor is generated based on the deviation between the actual memory retention rate of the consolidation test of the target user and the theoretical memory forgetting curve; based on the predicted memory data, a grey wolf algorithm is utilized to dynamically optimize multiple combined learning strategies, and an optimal combined learning strategy is determined; and based on the optimal combined learning strategy, establishing a pushing rule, and based on the pushing rule, pushing the corresponding learning content to a user terminal of the target user. According to the invention, intelligent content pushing can be realized, and the real-time performance and accuracy of pushing and the overall performance of the system are improved.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to an intelligent content push and interaction method and an intelligent content push and interaction device. Background Technology

[0002] Learning knowledge points, such as English vocabulary, is a fundamental and crucial step in mastering the English language. For a long time, English vocabulary learning methods have primarily relied on traditional printed media and classroom teaching. These traditional methods depend heavily on rote memorization, a process that is tedious and uninspiring, failing to maintain learning motivation, and resulting in poor practical application skills due to the disconnect from context.

[0003] In existing technologies, computer-aided learning schemes are commonly used, which improve efficiency to some extent by building personal vocabularies and simple interval repetition algorithms. However, these schemes mostly employ fixed learning strategies, have low dynamic adjustment capabilities to the learner's real-time state, and have limited global optimization capabilities for multiple combined learning strategies. This results in low accuracy of memory prediction models and an inability to generate optimal learning strategies, thus affecting the real-time performance, accuracy, and overall system performance of the push notifications. Summary of the Invention

[0004] To address the technical problems in existing technologies that employ fixed learning strategies, have limited dynamic adjustment and global optimization capabilities, resulting in low accuracy of memory prediction models and an inability to generate optimal learning strategies, thereby affecting the real-time performance, accuracy, and overall system performance of push notifications, this invention provides an intelligent content push and interaction method and an intelligent content push and interaction device. This intelligent content push and interaction method enables dynamic generation of learning strategies and intelligent content push, improving the real-time performance, accuracy, and overall system performance of push notifications.

[0005] To achieve the above objectives, the first aspect of the present invention provides an intelligent content push and interaction method, comprising: constructing a target user profile based on multi-dimensional user data; predicting the forgetting time point of each learning unit based on the target user profile and historical learning data, and obtaining predicted memory data; wherein, during the prediction process, an Ebbinghaus decay correction factor is introduced to dynamically adjust the predicted data; the Ebbinghaus decay correction factor is generated based on the deviation between the actual memory retention rate of the target user's consolidation test and the theoretical memory forgetting curve; based on the predicted memory data, using the Grey Wolf algorithm to dynamically optimize multiple combined learning strategies to determine the optimal combined learning strategy; based on the optimal combined learning strategy, establishing push rules, and based on the push rules, pushing the corresponding learning content to the target user's user terminal.

[0006] Furthermore, the step of predicting the forgetting time point of each learning unit and obtaining predicted memory data based on the target user profile and historical learning data includes: inputting the target user profile and the historical learning data into a pre-established bidirectional LSTM network to obtain initial predicted memory data; calculating the Ebbinghaus decay correction factor based on the deviation between the actual memory retention rate of the target user in the consolidation test and the theoretical memory forgetting curve; and dynamically adjusting the initial predicted memory data using the Ebbinghaus decay correction factor to determine the final predicted memory data.

[0007] Furthermore, the step of calculating the Ebbinghaus decay correction factor based on the deviation between the actual memory retention rate and the theoretical memory forgetting curve of the target user in the consolidation test includes: obtaining the theoretical memory retention rate corresponding to the test time point of the consolidation test on the theoretical memory forgetting curve; calculating the actual deviation value between the actual memory retention rate and the theoretical memory retention rate; and substituting the actual deviation value into a preset conversion function to determine the Ebbinghaus decay correction factor.

[0008] Furthermore, the step of dynamically adjusting the initial predicted memory data using the Ebbinghaus decay correction factor includes: multiplying the forgetting time points in the initial predicted memory data by the Ebbinghaus decay correction factor to obtain the dynamically adjusted final predicted memory data.

[0009] Furthermore, the method further includes: determining an initial feedback interval based on the target user profile; dynamically evaluating the real-time learning status of the target user based on the historical learning data and generating a status adjustment factor; calculating an optimal achievement feedback interval based on the initial feedback interval and the status adjustment factor; and pushing achievement feedback to the target user based on the optimal achievement feedback interval.

[0010] Further, the step of dynamically optimizing multiple combined learning strategies using the gray wolf algorithm based on the predicted memory data to determine the optimal combined learning strategy includes: initializing a population containing multiple gray wolf individuals; wherein, the position vector of each gray wolf individual represents a combined learning strategy; wherein, the combined learning strategy is generated based on the predicted memory data; determining a fitness function; wherein, the fitness function is used to: input the combined learning strategy represented by the position vector of any gray wolf individual into a bidirectional LSTM network used to acquire predicted memory data, generate simulated predicted memory data, and calculate the fitness value of the gray wolf individual based on the simulated predicted memory data; iteratively executing the following steps until the termination condition is met: determining the gray wolf individuals with the best, second best, and third best fitness in the current population based on the fitness function; updating the position vectors of the remaining gray wolf individuals in the population based on the position vectors of the best, second best, and third best gray wolf individuals; correcting the updated position vectors based on preset constraint processing conditions; and after the iteration terminates, determining the combined learning strategy represented by the position vector of the gray wolf individual with the best fitness as the optimal combined learning strategy.

[0011] Furthermore, the fitness function is a multi-objective fitness function, and the optimization objectives of the fitness function include at least: long-term memory retention rate, learning efficiency per unit time, and cognitive fatigue of the target user.

[0012] Furthermore, establishing push rules based on the optimal combination learning strategy includes: establishing push rules that include vocabulary push logic and a difficulty adaptive adjustment mechanism based on the optimal combination learning strategy; wherein, the vocabulary push logic is: pushing new words based on dual filtering of topic word similarity and word frequency level, and inserting a review task when the predicted memory data is lower than the review trigger threshold set by the optimal combination learning strategy; the difficulty adaptive adjustment mechanism is: generating staged example sentences with increasing complexity based on the example sentence complexity parameter in the optimal combination learning strategy, and inserting native language translation when the translation error rate is greater than a preset threshold.

[0013] Furthermore, the method also includes: acquiring error data of the target user in real time, and calculating the error rate related to a specific knowledge unit within a preset time period based on the error data; and shortening the review push interval for the specific knowledge unit when the error rate exceeds a preset error rate threshold.

[0014] A second aspect of this application provides an intelligent content push and interaction device, comprising: a learning profile construction module for constructing a target user profile based on multi-dimensional user data; a memory prediction module for predicting the forgetting time point of each learning unit based on the target user profile and historical learning data, and obtaining predicted memory data; wherein, an Ebbinghaus decay correction factor is introduced during the prediction process to dynamically adjust the predicted data; the Ebbinghaus decay correction factor is generated based on the deviation between the actual memory retention rate of the target user's consolidation test and the theoretical memory forgetting curve; a strategy determination module for dynamically optimizing multiple combined learning strategies using the Grey Wolf algorithm based on the predicted memory data to determine the optimal combined learning strategy; and a push module for establishing push rules based on the optimal combined learning strategy, and pushing the corresponding learning content to the target user's user terminal based on the push rules.

[0015] The present invention has at least the following technical effects through the technical solution provided by the present invention: The intelligent content push and interaction method of this invention constructs a target user profile based on multi-dimensional user data and combines it with a memory prediction model incorporating an Ebbinghaus forgetting curve correction factor to achieve accurate and dynamic prediction of the target user's memory state. Subsequently, the Grey Wolf algorithm is used to globally optimize multiple combined learning strategies to find the optimal strategy that balances long-term memory retention, learning efficiency, and cognitive fatigue, solving the problem of existing technologies having a single strategy and failing to achieve global optimization. Finally, push rules are established based on the optimal combined learning strategy, enabling intelligent new word push, review insertion, and difficulty adjustment. Combined with multimodal interaction and real-time feedback mechanisms, the push strategy is dynamically calibrated, improving the real-time performance and accuracy of content push, and optimizing the generation efficiency and quality of learning strategies. Therefore, the intelligent content push and interaction method provided in this application can achieve dynamic generation of learning strategies and intelligent content push, improving the real-time performance, accuracy, and overall system performance of the push.

[0016] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart of the intelligent content push and interaction method provided in the embodiments of the present invention; Figure 2 This is a structural diagram of the intelligent content push and interaction device provided in an embodiment of the present invention. Detailed Implementation

[0018] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0020] In this invention, unless otherwise stated, directional terms such as "upper," "lower," "top," and "bottom" are generally used to describe the relative positions of components in relation to the directions shown in the accompanying drawings or in relation to the vertical, perpendicular, or gravitational directions.

[0021] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] Please refer to Figure 1 The first aspect of this invention provides an intelligent content push and interaction method, comprising: S101: constructing a target user profile based on multi-dimensional user data; S102: predicting the forgetting time point of each learning unit based on the target user profile and historical learning data, and obtaining predicted memory data; wherein, an Ebbinghaus decay correction factor is introduced during the prediction process to dynamically adjust the predicted data; the Ebbinghaus decay correction factor is generated based on the deviation between the actual memory retention rate of the target user's consolidation test and the theoretical memory forgetting curve; S103: dynamically optimizing multiple combined learning strategies using the Grey Wolf algorithm based on the predicted memory data, and determining the optimal combined learning strategy; S104: establishing push rules based on the optimal combined learning strategy, and pushing the corresponding learning content to the target user's user terminal based on the push rules.

[0023] Specifically, in this embodiment, multidimensional user data is first acquired to construct a target user profile. Taking vocabulary learning as an example, the multidimensional user data includes at least: vocabulary size data, assessment data, learning log data, and physiological signal data. When acquiring vocabulary size data, a tiered vocabulary test based on the CERF standard is used to test the target user and determine their current vocabulary size, for example, mastering 1200 C1-level words. When acquiring assessment data, the target user is evaluated based on a VARK model to determine their learning preferences, such as visual, auditory, reading / writing, or practical learning styles. When acquiring learning log data, learning behavior is analyzed by recording click hotspots, such as the frequency of clicks on skip buttons, and page dwell time, such as an average learning time of 35 seconds per word. When acquiring physiological signal data, the target user's facial expressions are captured to assess their fatigue and emotional state. By combining the above-mentioned vocabulary size data, assessment data, learning log data, and physiological signal data, a target user profile that comprehensively represents the target user's learning style, learning preferences, and learning ability is constructed.

[0024] Next, based on the target user profile and historical learning data, the forgetting time point for each learning unit is predicted, obtaining predicted memory data. The historical learning data stored in the server database and the constructed target user profile are input into a pre-trained memory prediction model for processing. During the prediction process, an Ebbinghaus forgetting curve correction factor is introduced to dynamically adjust the predicted data. For example, when the target user's actual memory performance in a consolidation test is better than the theoretical forgetting curve, such as when the test accuracy reaches 80%, this correction factor will dynamically extend the review interval; conversely, it will shorten it. During processing, the memory prediction model can also calculate the optimal achievement feedback interval for providing positive feedback incentives to the target user based on the target user's profile and historical data, such as triggering a progress broadcast every 5 words completed.

[0025] Then, based on the predictive memory data, the Gray Wolf Algorithm is used to dynamically optimize multiple combined learning strategies to determine the optimal combined learning strategy. Several different combined learning strategies are generated based on the predictive memory data and represented as position vectors in the Gray Wolf Algorithm. Dynamic wolf leader election is performed based on a multi-objective fitness function and the Gray Wolf Algorithm, with constraints set to simulate and evaluate the effects of different strategies. The optimal learning strategy is obtained through algorithm iteration.

[0026] Finally, based on the determined optimal learning strategy, a push rule is established, and based on this rule, the corresponding learning content is pushed to the target user's terminal. The push rule includes vocabulary push logic, adaptive difficulty adjustment, and anomaly handling mechanism, which guides the introduction of new words, the insertion of review tasks, the adjustment of example sentence difficulty, and the handling of learning anomalies.

[0027] To further enhance the learning experience, an interactive mechanism was established, including multimodal content presentation, gamified incentives, and instant feedback. Multimodal content presentation is dynamically constructed based on real-time target user profiles. For example, visually, the background color of word cards gradually changes according to memory intensity; auditorily, when physiological signal data analysis indicates that the target user is overly fatigued, the synthesized speech is replaced with example sentences spoken by a real person to maintain the target user's attention.

[0028] The method also includes real-time dynamic adjustments to the learning process. It acquires error data from target users at each stage, and if the error data shows a continuous upward trend, the push frequency for that stage is increased. Simultaneously, it analyzes example sentences saved by target users, extracts key themes, and uses these keywords to update the target user's profile, thereby pushing subsequent learning content that better aligns with their recent interests.

[0029] Furthermore, to ensure the continuous evolution of the entire learning system, a control group is established during off-peak traffic periods at night. A small portion (e.g., 5%) of user data is imported into a dedicated optimized neural network. This network retrains and corrects the parameters of the core memory prediction model and the Gray Wolf algorithm, such as updating the constraint parameters of the Gray Wolf algorithm. After correction, the updated model parameters are synchronized to all target users' terminals in real time via the cloud, ensuring prediction accuracy and push efficiency.

[0030] The intelligent content push and interaction method provided in this application can accurately and dynamically predict the memory state of target users by constructing a target user profile that integrates multi-dimensional data and combining it with a memory prediction model that incorporates an Ebbinghaus forgetting curve correction factor. By utilizing the Grey Wolf algorithm to globally optimize multiple combined learning strategies, the optimal strategy that balances long-term memory retention, learning efficiency, and cognitive fatigue is found, solving the problem of existing technologies having single strategies and failing to achieve global optimization. Based on the optimal strategy, push rules are established, enabling intelligent new word push, review insertion, and difficulty adjustment. Combined with multimodal interaction and real-time feedback mechanisms, the push strategy is dynamically calibrated, improving the real-time performance and accuracy of content push, thereby optimizing the generation efficiency and quality of learning strategies.

[0031] Although this application uses English vocabulary learning as an example, its methodological framework is universal. Those skilled in the art can apply it to other language learning, professional skills training, and other fields, and make adaptive adjustments to the data model, optimization objectives, and push rules. Such variations all fall within the protection scope of this application.

[0032] Furthermore, the step of predicting the forgetting time point of each learning unit and obtaining predicted memory data based on the target user profile and historical learning data includes: inputting the target user profile and the historical learning data into a pre-established bidirectional LSTM network to obtain initial predicted memory data; calculating the Ebbinghaus decay correction factor based on the deviation between the actual memory retention rate of the target user in the consolidation test and the theoretical memory forgetting curve; and dynamically adjusting the initial predicted memory data using the Ebbinghaus decay correction factor to determine the final predicted memory data.

[0033] Furthermore, the step of calculating the Ebbinghaus decay correction factor based on the deviation between the actual memory retention rate and the theoretical memory forgetting curve of the target user in the consolidation test includes: obtaining the theoretical memory retention rate corresponding to the test time point of the consolidation test on the theoretical memory forgetting curve; calculating the actual deviation value between the actual memory retention rate and the theoretical memory retention rate; and substituting the actual deviation value into a preset conversion function to determine the Ebbinghaus decay correction factor.

[0034] Furthermore, the step of dynamically adjusting the initial predicted memory data using the Ebbinghaus decay correction factor includes: multiplying the forgetting time points in the initial predicted memory data by the Ebbinghaus decay correction factor to obtain the dynamically adjusted final predicted memory data.

[0035] Specifically, in this embodiment, during the data preparation stage, a target user profile that comprehensively represents the target user's learning style, preferences, and abilities is integrated and preprocessed with historical learning data stored in the server database. For example, non-numerical information such as learning style, preferences, and abilities in the target user profile is converted into numerical feature vectors. Historical learning data, such as time-series data like the duration of each word learning session, click behavior, test accuracy, and error type, is organized into fixed-length data vectors. These feature vectors and data vectors are concatenated in chronological order and input into a pre-established bidirectional LSTM network.

[0036] This application employs a bidirectional LSTM network, which not only learns patterns from past learning behaviors but also comprehensively assesses the potential impact of subsequent learning tasks on current word memorization, thus making more comprehensive predictions. The bidirectional LSTM network outputs two key results: initial predicted memory data for each word, for example, predicting that the memory retention rate of the word "ambiguity" will drop to 58% on the 3rd day after learning; and the optimal achievement feedback interval calculated based on the target user profile and historical data, for example, triggering a progress broadcast every 5 words completed.

[0037] Initial predicted memory data is based on group patterns learned during model training. To achieve true personalization, dynamic correction is needed. The correction process is triggered when the target user encounters the word "ambiguity" in a consolidation test. Assume this test is conducted on the third day after learning, and consists of four multiple-choice questions and one fill-in-the-blank question containing the word. First, the theoretical memory retention rate corresponding to the third day on the standard Ebbinghaus forgetting curve is obtained, assumed to be 60%. Then, the actual memory retention rate is calculated based on the target user's actual performance. For example, if each multiple-choice question has a weight of 0.2, and each fill-in-the-blank question has a weight of 0.2, and all four multiple-choice questions are answered correctly while the fill-in-the-blank question is answered incorrectly, the actual score is 0.8, meaning the actual memory retention rate is 80%.

[0038] Subsequently, the deviation between the actual retention rate of 80% and the theoretical retention rate of 60% is calculated, yielding an actual deviation value of 20%. This actual deviation value is then substituted into a preset conversion function to determine the Ebbinghaus decay correction factor. The conversion function can be a simple linear function, for example, correction factor = 1 + k * actual deviation value, where k is an adjustable system parameter, such as k = 1. Therefore, a 20% actual deviation value generates a correction factor of 1.2. The conversion function can avoid extreme corrections, for example, by limiting the deviation value to a certain range, ensuring the smoothness of the correction.

[0039] During the dynamic adjustment phase, the initial prediction data is adjusted using a correction factor of 1.2. Multiplying the initial predicted forgetting time of 3 days by the correction factor 1.2 yields 3.6 days. This can be rounded up to 4 days as the final predicted memory data for the word. The new review interval is immediately updated in the target user's schedule. Simultaneously, throughout the entire correction process, including actual test performance, calculated deviation values, and the final correction factor, all are stored as new historical learning data in the server database for use as prediction input for the next bidirectional LSTM network, thus forming a continuous learning and optimization intelligent closed loop. If actual performance is worse than the theoretical value, a correction factor less than 1 is generated, shortening the review interval and ensuring the target user receives timely reminders at the critical point of memory decay.

[0040] The intelligent content push and interaction method provided in this application combines deep learning models with mature cognitive science theories to achieve dynamic closed-loop correction, improve the accuracy of memory prediction, and adapt to individual dynamic changes. By capturing complex memory patterns through a bidirectional LSTM network and then dynamically calibrating using an Ebbinghaus decay correction factor calculated based on actual test performance, it leverages the powerful fitting ability of machine learning while ensuring the scientific validity and rationality of the prediction results. Furthermore, the correction factor is updated in real-time based on the results of each consolidation test, allowing the entire prediction process to continuously update in accordance with the real-time changes in the target user's state, achieving personalized adaptation. This ensures that review tasks are triggered when most needed, avoiding inefficiency caused by reviewing too early or too late, thereby optimizing the accuracy and personalized adaptability of the memory prediction model.

[0041] Furthermore, the method further includes: determining an initial feedback interval based on the target user profile; dynamically evaluating the real-time learning status of the target user based on the historical learning data and generating a status adjustment factor; calculating an optimal achievement feedback interval based on the initial feedback interval and the status adjustment factor; and pushing achievement feedback to the target user based on the optimal achievement feedback interval.

[0042] Specifically, in this embodiment, the initial feedback interval is determined based on the constructed target user profile. The target user profile comprehensively reflects the learning style, preferences, and abilities of the target users. For example, based on the personality trait tags in the profile, a shorter initial feedback interval is set for extroverted target users to provide frequent encouragement, while a longer initial interval is set for introverted target users to avoid excessive interference. Alternatively, for target users marked as highly motivated and highly focused, a trigger can be set after every 10 new words are completed; while for target users who are easily distracted and require continuous encouragement, a trigger can be set after every 5 new words are completed.

[0043] However, fixed intervals cannot adapt to the real-time fluctuations in the target user's state, thus requiring dynamic adjustment. Historical learning data is continuously analyzed to dynamically assess the target user's real-time learning state, generating a state adjustment factor that acts on the initial feedback interval. The evaluation process integrates multiple indicators, specifically analyzing recent answer accuracy, answering speed, and continuous learning duration: when the accuracy is higher than the first threshold or the number of consecutive correct answers exceeds the second threshold, the state adjustment factor is determined to be an incentive amplification value greater than 1; when the accuracy is lower than the third threshold, the number of consecutive incorrect answers exceeds the fourth threshold, or the answering speed is lower than the fifth threshold, it is determined to be an incentive compensation value less than 1; when the continuous learning duration exceeds the sixth threshold, a mandatory rest reminder is triggered, and the state adjustment factor is temporarily set to 0.

[0044] Finally, the initial feedback interval is multiplied by the state adjustment factor to calculate the optimal achievement feedback interval. For example, if the initial interval is 5 and the adjustment factor is 0.8, the optimal interval is 4. When the target user completes the required number of words at the dynamically calculated optimal interval, achievement feedback is pushed. Achievement feedback is multimodal, selecting from various preset feedback types such as text, badges, and sound effects based on the preferences in the target user profile. This could be an update to a progress bar, a pop-up of a new badge, a voice message of encouragement like "You're awesome!", or a cheerful sound effect, all designed to motivate the target user to continue making progress.

[0045] The intelligent content push and interaction method provided in this application can construct an intelligent feedback adjustment model by combining the target user profile and real-time learning status, dynamically adjust the feedback interval, effectively improve the learning experience and motivation, reduce the abandonment rate, and thus optimize the adaptive adjustment capability of the feedback incentive mechanism to maintain the stability of the user's learning process.

[0046] Further, the step of dynamically optimizing multiple combined learning strategies using the gray wolf algorithm based on the predicted memory data to determine the optimal combined learning strategy includes: initializing a population containing multiple gray wolf individuals; wherein, the position vector of each gray wolf individual represents a combined learning strategy; wherein, the combined learning strategy is generated based on the predicted memory data; determining a fitness function; wherein, the fitness function is used to: input the combined learning strategy represented by the position vector of any gray wolf individual into a bidirectional LSTM network used to acquire predicted memory data, generate simulated predicted memory data, and calculate the fitness value of the gray wolf individual based on the simulated predicted memory data; iteratively executing the following steps until the termination condition is met: determining the gray wolf individuals with the best, second best, and third best fitness in the current population based on the fitness function; updating the position vectors of the remaining gray wolf individuals in the population based on the position vectors of the best, second best, and third best gray wolf individuals; correcting the updated position vectors based on preset constraint processing conditions; and after the iteration terminates, determining the combined learning strategy represented by the position vector of the gray wolf individual with the best fitness as the optimal combined learning strategy.

[0047] Furthermore, the fitness function is a multi-objective fitness function, and the optimization objectives of the fitness function include at least: long-term memory retention rate, learning efficiency per unit time, and cognitive fatigue of the target user.

[0048] Specifically, in this implementation, it is necessary to first define what the combined learning strategy includes. A strategy is not a single instruction, but a set of learning parameters, such as the review interval ratio, the distribution ratio of different question types (multiple choice and fill-in-the-blank), the suggested learning time for a new word, and the maximum number of new words to learn per day. To find the optimal solution among numerous possible parameter combinations, this method introduces the Gray Wolf algorithm. At the start of the algorithm, a population containing multiple gray wolves is created. Each gray wolf represents a candidate combined learning strategy, and each number in the position vector directly corresponds to a specific parameter value within that combined learning strategy.

[0049] To evaluate the merits of each combined learning strategy, this application defines a multi-objective fitness function, which evaluates three dimensions: long-term memory retention, learning efficiency per unit time, and cognitive fatigue of the target user. These three objectives are often contradictory; for example, maximizing memory retention may require extensive repetitive review, reducing efficiency and causing user fatigue. The evaluation process involves inputting the strategy parameters represented by the individual gray wolf to be evaluated into a pre-trained bidirectional LSTM network. This network simulates the entire learning process of the target user according to the strategy and predicts the results, such as the retention rate of each word in the coming week and the total learning time. Finally, the fitness function calculates a comprehensive score based on these predictions using methods such as weighted summation or Pareto ranking, resulting in the fitness value. A higher score indicates that the strategy performs better in balancing memory, efficiency, and fatigue.

[0050] Then, the iterative optimization loop begins. In each loop, the three individuals with the highest fitness values ​​in the population are identified as the three alpha wolves—Alpha, Beta, and Delta—representing the best learning direction at present. The remaining gray wolves update their own positions based on the positions of these three alpha wolves using a specific mathematical formula, adjusting their policy parameters to move towards a better policy region. During the update process, they learn from the best while fine-tuning near known near-optimal solutions. After the position update, constraints are applied to ensure the generated strategy is reasonable and feasible. For example, any parameter representing time or quantity cannot be negative, the number of new words per day cannot exceed the memory limit, and the weights of all question types must sum to 100%. Parameter values ​​exceeding reasonable ranges are directly corrected to the effective boundary. When the number of loops reaches a preset value, or the optimal score no longer shows significant improvement, the entire optimization process ends. At this point, the combined learning strategy represented by the Alpha alpha wolf, after decoding, is determined to be the optimal combined learning strategy tailored to the current target user and is directly applied to the actual learning plan.

[0051] The intelligent content push and interaction method provided in this application can achieve a dynamic balance between multiple objectives, including memory retention, learning efficiency, and cognitive fatigue, ensuring the scientific nature of the learning strategy. The determined optimal strategy can efficiently consolidate knowledge while taking into account the user experience of the target audience and avoiding cognitive fatigue. It enhances the personalization and adaptability of the learning plan and strengthens learning stickiness. Ultimately, it achieves global optimization of the learning strategy under multi-objective constraints, improving the scientific nature and robustness of the strategy generation process.

[0052] Furthermore, establishing push rules based on the optimal combination learning strategy includes: establishing push rules that include vocabulary push logic and a difficulty adaptive adjustment mechanism based on the optimal combination learning strategy; wherein, the vocabulary push logic is: pushing new words based on dual filtering of topic word similarity and word frequency level, and inserting a review task when the predicted memory data is lower than the review trigger threshold set by the optimal combination learning strategy; the difficulty adaptive adjustment mechanism is: generating staged example sentences with increasing complexity based on the example sentence complexity parameter in the optimal combination learning strategy, and inserting native language translation when the translation error rate is greater than a preset threshold.

[0053] Specifically, in this embodiment, a push rule is established based on a determined optimal combination learning strategy, comprising vocabulary push logic and a difficulty adaptive adjustment mechanism. The vocabulary push logic is implemented as follows: a topic lexicon is constructed, and the similarity between new words and the target user's current learning topic is calculated using a word vector model. This is combined with word frequency levels for dual filtering, prioritizing the push of practical new words that are relevant to the topic and have high frequency. The predictive memory data output by the bidirectional LSTM network is continuously monitored. When the predictive retention rate of any word falls below the review trigger threshold set in the optimal combination learning strategy, a review task for that word is immediately inserted into subsequent learning tasks. The difficulty adaptive adjustment mechanism dynamically generates example sentences based on the example sentence complexity parameter in the optimal combination learning strategy: initially, simple, basic short sentences are pushed; as the target user's accuracy improves, the number of clauses, more complex tenses, and more abstract vocabulary are gradually increased, forming a phased example sentence with increasing complexity. During the translation practice session, the error rate of the target user is counted in real time. When the error rate for a specific sentence pattern or word exceeds a preset threshold, the native language translation is automatically inserted next to the question as a prompt to help the target user understand the difficulties. The prompt is then removed after the user has mastered the skills.

[0054] The intelligent content push and interaction method provided in this application enables dual adaptation of learning content and difficulty through refined push rules. It improves push efficiency by combining themes and word frequency, accurately inserts reviews based on predictive memory, and adjusts learning difficulty through incremental example sentences and conditional translation. This achieves an efficient and user-friendly learning experience, enhancing learning engagement and effectiveness.

[0055] Furthermore, the push rules based on the optimal combination learning strategy also include an exception handling mechanism. This mechanism first sets a threshold for a sudden increase in the word test error rate. This threshold can be a specific numerical value, such as the error rate instantly exceeding 50%, or a dynamic statistical value based on historical data, such as the error rate exceeding twice the standard deviation of the recent average. When the real-time monitored error rate exceeds this threshold, a retrospective learning mode is immediately triggered. After the retrospective learning mode is activated, the error logs of the target user within the retrospective time window (e.g., the past three days) set by the optimal combination learning strategy are retrieved. The high-frequency error words with the highest number of errors are selected from the logs, and a special retrospective review task package is generated. This task package is pushed out before regular new word learning tasks, such as through spelling review, word meaning matching, and listening comprehension exercises to reinforce learning. The parameters of the error rate surge threshold and the retrospective time window are both pre-set by the optimal combination learning strategy based on the target user's historical performance and current status to ensure the timeliness and accuracy of the intervention. This application utilizes real-time response to sudden increases in error rate to conduct precise backtracking training on frequently misspelled words, thereby achieving accurate identification and rapid reinforcement of knowledge gaps.

[0056] Furthermore, the method also includes: acquiring error data of the target user in real time, and calculating the error rate related to a specific knowledge unit within a preset time period based on the error data; and shortening the review push interval for the specific knowledge unit when the error rate exceeds a preset error rate threshold.

[0057] Specifically, in this embodiment, error data generated by the target user during practice is recorded in real time. This error data is associated with one or more knowledge units, such as a specific English word or phrase. A preset time period, such as 24 hours, is set, and the error rate of a specific knowledge unit (e.g., the word "abandon") is calculated within that time period. When the calculated error rate exceeds a preset error rate threshold, such as above 30%, the system determines that the knowledge unit is a weak point. The review interval for that specific knowledge unit is shortened. For example, if the original review interval is 3 days, the system will adjust it to 1 day to increase the review frequency of that word.

[0058] Furthermore, this application periodically scans the example sentence library collected by the target user and uses natural language processing techniques, such as the Term Frequency-Inverse Document Frequency (TF-IDF) algorithm, to extract the highest-weighted thematic keywords from all collected example sentences. These keywords reflect the target user's current interests or learning focus. Subsequently, these keywords are used as new tags and integrated or updated into the target user's profile model, increasing the weight of the corresponding themes. When generating subsequent learning plans, learning materials related to high-weight keywords in the updated target user profile are prioritized. For example, if the weight of business negotiation and marketing increases in the profile, subsequent reading articles, listening materials, and vocabulary lists will tend to select content on these two themes.

[0059] The intelligent content push and interaction method provided in this application can respond to error data in real time and mine interests and preferences to achieve precise reinforcement and personalized content push, thereby realizing dynamic updates of user profiles and accuracy of content recommendations.

[0060] Please refer to Figure 2 The second aspect of this application provides an intelligent content push and interaction device, comprising: a learning profile construction module for constructing a target user profile based on multi-dimensional user data; a memory prediction module for predicting the forgetting time point of each learning unit based on the target user profile and historical learning data, and obtaining predicted memory data; wherein, an Ebbinghaus decay correction factor is introduced during the prediction process to dynamically adjust the predicted data; the Ebbinghaus decay correction factor is generated based on the deviation between the actual memory retention rate of the target user's consolidation test and the theoretical memory forgetting curve; a strategy determination module for dynamically optimizing multiple combined learning strategies using the Grey Wolf algorithm based on the predicted memory data to determine the optimal combined learning strategy; and a push module for establishing push rules based on the optimal combined learning strategy, and pushing the corresponding learning content to the target user's user terminal based on the push rules.

[0061] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various simple modifications can be made to the technical solution of the present invention, and these simple modifications all fall within the protection scope of the present invention.

[0062] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable way without contradiction. In order to avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

[0063] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed by the present invention.

Claims

1. An intelligent content push and interaction method, characterized in that, The method includes: Target user profiles are built based on multi-dimensional user data; Based on the target user profile and historical learning data, the forgetting time point of each learning unit is predicted to obtain predicted memory data; wherein, during the prediction process, an Ebbinghaus decay correction factor is introduced to dynamically adjust the predicted data; the Ebbinghaus decay correction factor is generated based on the deviation between the actual memory retention rate of the target user's consolidation test and the theoretical memory forgetting curve; Based on the predicted memory data, the Grey Wolf algorithm is used to dynamically optimize multiple combined learning strategies to determine the optimal combined learning strategy. Based on the optimal combination learning strategy, a push rule is established, and based on the push rule, the corresponding learning content is pushed to the target user's user terminal.

2. The intelligent content push and interaction method according to claim 1, characterized in that, The step of predicting the forgetting time point for each learning unit based on the target user profile and historical learning data, and obtaining predicted memory data, includes: The target user profile and the historical learning data are input into a pre-established bidirectional LSTM network to obtain initial predictive memory data; The Ebbinghaus decay correction factor is calculated based on the deviation between the actual memory retention rate of the target users in the consolidation test and the theoretical memory forgetting curve. The initial predictive memory data is dynamically adjusted using the Ebbinghaus decay correction factor to determine the final predictive memory data.

3. The intelligent content push and interaction method according to claim 2, characterized in that, The calculation of the Ebbinghaus forgetting curve correction factor based on the deviation between the target user's actual memory retention rate and the theoretical forgetting curve in the consolidation test includes: Obtain the theoretical memory retention rate corresponding to the test time point of the consolidation test on the theoretical memory forgetting curve; Calculate the actual deviation between the actual memory retention rate and the theoretical memory retention rate; The actual deviation value is substituted into a preset conversion function to determine the Ebbinghaus attenuation correction factor.

4. The intelligent content push and interaction method according to claim 2, characterized in that, The step of dynamically adjusting the initial prediction memory data using the Ebbinghaus decay correction factor includes: The forgetting time points in the initial predicted memory data are multiplied by the Ebbinghaus decay correction factor to obtain the dynamically adjusted final predicted memory data.

5. The intelligent content push and interaction method according to claim 1, characterized in that, The method further includes: Based on the target user profile, determine the initial feedback interval; Based on the historical learning data, the real-time learning status of the target user is dynamically evaluated, and a status adjustment factor is generated. Calculate the optimal achievement feedback interval based on the initial feedback interval and the state adjustment factor; Achievement feedback is pushed to the target user based on the optimal achievement feedback interval.

6. The intelligent content push and interaction method according to claim 1, characterized in that, The step of dynamically optimizing multiple combined learning strategies based on the predicted memory data using the Grey Wolf algorithm to determine the optimal combined learning strategy includes: A population containing multiple gray wolf individuals is initialized; wherein the position vector of each gray wolf individual represents a combined learning strategy; wherein the combined learning strategy is generated based on the predicted memory data; Determine the fitness function; wherein, the fitness function is used to: input the combined learning strategy represented by the position vector of any gray wolf individual into the bidirectional LSTM network used to obtain predictive memory data, generate simulated predictive memory data, and calculate the fitness value of the gray wolf individual based on the simulated predictive memory data; The following steps are executed iteratively until the termination condition is met: determine the best, second best, and third best gray wolf individuals in the current population based on the fitness function; update the position vectors of the remaining gray wolf individuals in the population based on the position vectors of the best, second best, and third best gray wolf individuals; and correct the updated position vectors based on preset constraint conditions. After the iteration terminates, the combined learning strategy represented by the position vector of the gray wolf individual with the best fitness is determined as the optimal combined learning strategy.

7. The intelligent content push and interaction method according to claim 6, characterized in that, The fitness function is a multi-objective fitness function, and the optimization objectives of the fitness function include at least: long-term memory retention rate, learning efficiency per unit time, and cognitive fatigue of the target user.

8. The intelligent content push and interaction method according to claim 1, characterized in that, The step of establishing push rules based on the optimal combination learning strategy includes: Based on the aforementioned optimal combination learning strategy, a push rule is established that includes vocabulary push logic and difficulty adaptive adjustment mechanism; The vocabulary push logic is as follows: new words are pushed based on dual filtering of topic word similarity and word frequency level, and a review task is inserted when the predicted memory data is lower than the review trigger threshold set by the optimal combination learning strategy. The difficulty adaptive adjustment mechanism is as follows: based on the example sentence complexity parameter in the optimal combination learning strategy, a phased example sentence with increasing complexity is generated, and native language translation is inserted when the translation error rate is greater than a preset threshold.

9. The intelligent content push and interaction method according to claim 1, characterized in that, The method further includes: The system acquires error data from target users in real time and calculates the error rate related to a specific knowledge unit within a preset time period based on the error data. If the error rate exceeds a preset error rate threshold, the review push interval for that specific knowledge unit will be shortened.

10. An intelligent content push and interaction device, characterized in that, The device includes: The profile building module is used to build target user profiles based on multi-dimensional user data. The memory prediction module is used to predict the forgetting time point of each learning unit based on the target user profile and historical learning data, and to obtain predicted memory data. In the prediction process, an Ebbinghaus decay correction factor is introduced to dynamically adjust the predicted data. The Ebbinghaus decay correction factor is generated based on the deviation between the actual memory retention rate of the target user's consolidation test and the theoretical memory forgetting curve. The strategy determination module is used to dynamically optimize multiple combined learning strategies based on the predicted memory data using the Grey Wolf algorithm, and determine the optimal combined learning strategy. The push module is used to establish push rules based on the optimal combination learning strategy, and push the corresponding learning content to the target user's user terminal based on the push rules.

Images