Chinese and English mixed children English dialogue data set construction method and teaching system

Abstract

The invention relates to the technical field of artificial intelligence and natural language processing, and particularly discloses a Chinese and English mixed children English dialogue data set construction method and teaching system.The method comprises the steps that English learning levels are classified according to the basic language level of a target user, and a question and answer prompt template is configured for each English learning level; dialogue data is generated; inputting the dialogue data into a large language model to obtain a learning feedback result of the target user; performing reinforcement learning on the learning feedback result of the target user and the English learning level of the target user; evaluating the dialogue data and the learning feedback result according to a large language model for data quality evaluation; and storing the dialogue data of which the dialogue data evaluation result is greater than a preset threshold value to obtain a Chinese and English mixed children English dialogue data set. The data set constructed by the Chinese and English mixed children English dialogue data set construction method provided by the invention is high in quality and can meet the real requirements of children English learning.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and natural language processing, and in particular to a method for constructing a Chinese-English mixed children's English dialogue data set, a storage medium, and a Chinese-English mixed children's English teaching system. Background Art

[0002] In recent years, natural language processing (NLP) and computer-assisted language learning (CALL) have been increasingly applied in intelligent education. In particular, the deep integration of intelligent dialogue systems and language models has greatly enhanced the personalization and interactivity of instruction in children's English. However, currently available children's English corpora are still primarily monolingual and fail to fully consider the mixed Chinese and English environments found in actual language learning environments.

[0003] In real-world English teaching scenarios for children, interactions between teachers and students often involve a lot of alternating Chinese and English expressions, for example: "What is the English word for apple?" "The English word for apple is 'apple.' Did you know that 'An apple a day keeps the doctor away' is a famous English proverb?" This mixed language model not only aligns with learners' cognitive habits but also helps reduce learning difficulty and facilitate a natural transition to full English expression. However, there is currently a lack of high-quality datasets for mixed Chinese-English conversations on the market. Even constructing corpora through web crawling or manual annotation cannot meet the real needs of children's English learning.

[0004] Therefore, how to provide a high-quality dataset that meets the real needs of children's English learning has become a technical problem that needs to be solved urgently by those skilled in the art. Summary of the Invention

[0005] The present invention provides a method for constructing a Chinese-English mixed children's English dialogue dataset, a storage medium, and a Chinese-English mixed children's English teaching system, which solves the problem in the related art that the current Chinese-English mixed dialogue dataset cannot meet the actual needs of children's English learning.

[0006] As a first aspect of the present invention, a method for constructing a mixed Chinese-English children's English dialogue dataset is provided, which includes:

[0007] The English learning level is divided into multiple levels according to the basic language proficiency of the target users, and a corresponding question and answer prompt template is configured for each English learning level;

[0008] Generate conversation data based on each English learning level and the corresponding question-answer prompt template;

[0009] Inputting the conversation data into a large language model for comprehension ability assessment to obtain learning feedback results of the target user;

[0010] Performing reinforcement learning on the target user's learning feedback results and the target user's English learning level according to a reinforcement learning algorithm to dynamically adjust the target user's English learning level;

[0011] Repeat the above steps according to the dynamically adjusted English learning level of the target user to regenerate the conversation data and obtain the learning feedback results of the target user;

[0012] Evaluate the conversation data and the learning feedback result according to the large language model used for data quality assessment, and obtain a conversation data evaluation result;

[0013] The dialogue data whose dialogue data evaluation results are greater than a preset threshold are stored to obtain a mixed Chinese-English children's English dialogue data set.

[0014] Furthermore, the English learning level is divided into multiple levels according to the basic language level of the target user, including:

[0015] According to the basic language proficiency of the target users, the English learning levels are divided into elementary English level, intermediate English level and advanced English level. The basic language level corresponding to the elementary English level includes mastering basic vocabulary, the basic language level corresponding to the intermediate English level includes the ability to understand simple sentences and be able to ask preliminary questions, and the basic language level corresponding to the advanced English level includes the ability to understand sentences with grammatical structures and be able to conduct basic daily communication.

[0016] Furthermore, dialogue data is generated based on each English learning level and the corresponding question and answer prompt template, including:

[0017] Determine the corresponding question and answer prompt template based on the English learning level of the current target user;

[0018] According to the question-answer prompt template, the large language model used for question generation is called to obtain question content corresponding to the English learning level of the current target user;

[0019] Inputting the question content corresponding to the current target user's English learning level into the teacher model to obtain an pedagogical answer result, wherein the pedagogical answer result output by the teacher model at least includes: English word explanation, supplemented by Chinese scope, and introduction of cultural or background knowledge;

[0020] According to the actual teaching process, interactive logical question-answering based on continuous questioning, answering and follow-up questions is conducted to obtain dialogue data.

[0021] Furthermore, the conversation data is input into a large language model for comprehension ability assessment to obtain learning feedback results of the target user, including:

[0022] Inputting the conversation data into a large language model for comprehension ability assessment to obtain a target user comprehension ability assessment result;

[0023] A learning feedback result of the target user is generated according to a comparison result of the target user's comprehension ability evaluation result and a preset comprehension evaluation threshold.

[0024] Furthermore, reinforcement learning is performed on the learning feedback results of the target user and the English learning level of the target user according to a reinforcement learning algorithm, including:

[0025] Determine the state space of the reinforcement learning algorithm according to the English learning level of the target user;

[0026] Determining an action space of a reinforcement learning algorithm based on the learning feedback results of the target user;

[0027] The state space and the action space are updated and learned according to a preset reward function to dynamically adjust the English learning level of the target user according to the learning feedback result of the target user.

[0028] Furthermore, updating and learning the state space and the action space according to a preset reward function to dynamically adjust the English learning level of the target user according to the learning feedback result of the target user includes:

[0029] The state space and the dynamic space are updated and learned according to the Q-table update formula according to the preset reward function, wherein the expression of the preset reward function is:

[0030]

[0031] The expression of the Q table update formula is:

[0032]

[0033] Among them, α represents the learning rate, γ represents the discount factor, and s ′ Indicates the new state after the current state performs action a.

[0034] Furthermore, the conversation data and the learning feedback results are evaluated based on the large language model used for data quality assessment, and a conversation data evaluation result is obtained, including:

[0035] The dialogue data and the learning feedback results are input into the large language model for data quality assessment for data quality assessment, wherein the content of the data quality assessment at least includes: judging whether the questions and answers are pedagogically reasonable, whether they conform to the mixed Chinese and English standards, whether they have semantic coherence, and whether the language is natural and vivid.

[0036] Furthermore, storing the conversation data whose conversation data evaluation result is greater than a preset threshold includes:

[0037] Quantitatively processing the dialogue data evaluation result to obtain a quantitative evaluation result;

[0038] Comparing the quantitative evaluation results with a preset threshold;

[0039] If the quantitative evaluation result is greater than the preset threshold, the conversation data corresponding to the quantitative evaluation result is stored in a structured JSON file.

[0040] As another aspect of the present invention, a storage medium is provided, which is used to store computer instructions. When the computer instructions are loaded and executed by a processor, the method for constructing a mixed Chinese-English children's English dialogue dataset as described above is implemented.

[0041] As another aspect of the present invention, a Chinese-English mixed children's English teaching system is provided, which includes a memory and a processor, the memory is communicatively connected to the processor, the memory is used to store computer instructions, and the processor is used to load and execute the computer instructions to implement the Chinese-English mixed children's English dialogue dataset construction method described above.

[0042] The method for constructing a mixed Chinese-English children's English dialogue dataset provided by the present invention divides English learning levels based on the basic language proficiency of target users, and configures corresponding question and answer prompt templates for each English learning level. Then, dialogue data is generated through training. Through comprehension ability assessment and reinforcement learning mechanism, adaptive adjustment of data content difficulty is achieved, thereby constructing a personalized, semantically reasonable, and natural-style high-quality Chinese-English mixed teaching dataset to meet the real needs of children's English learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the present invention, but do not constitute a limitation of the present invention.

[0044] Figure 1 This is a flowchart of the method for constructing a mixed Chinese-English children's English dialogue dataset provided by the present invention.

[0045] Figure 2A flow chart of generating conversation data provided by the present invention.

[0046] Figure 3 This is a flowchart of the method for obtaining learning feedback results from target users provided by the present invention.

[0047] Figure 4 This is a flowchart of the reinforcement learning provided by the present invention.

[0048] Figure 5 This is a schematic diagram of the performance change trend of different models provided by the present invention under different experimental configurations.

[0049] Figure 6 This is a diagram of vocabulary richness provided by the present invention.

[0050] Figure 7 This is a distribution diagram of the number of conversation turns provided by the present invention.

[0051] Figure 8 This is a sentence length distribution diagram provided by the present invention. DETAILED DESCRIPTION

[0052] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention may be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0053] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0054] It should be noted that the terms "first," "second," and the like in the specification and claims of the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate for the embodiments of the present invention described herein. In addition, the terms "including," "having," and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatuses.

[0055] In this embodiment, a method for constructing a mixed Chinese-English children's English dialogue dataset is provided. Figure 1: is a flowchart of a method for constructing a mixed Chinese-English children's English dialogue dataset according to an embodiment of the present invention. Figure 1 Shown, including:

[0056] S100, grading the English learning level of the target user according to the basic language proficiency to obtain multiple English learning levels, and configuring a corresponding question and answer prompt template for each English learning level;

[0057] In an embodiment of the present invention, since the target users have different basic language levels, it is necessary to divide the English learning levels according to the basic language levels of the target users, and configure corresponding question and answer prompt templates based on the divided English learning levels, so as to ensure that the generated content meets the level of the target users.

[0058] S200, generating dialogue data according to each English learning level and the corresponding question and answer prompt template;

[0059] In an embodiment of the present invention, the English learning level corresponding to the target user is determined and a question-answer prompt template corresponding to the English learning level is trained to obtain dialogue data.

[0060] S300: Inputting the conversation data into a large language model for comprehension ability assessment to obtain learning feedback results of the target user;

[0061] Specifically, the conversation data is input into a large language model for comprehension ability assessment for training, so that learning feedback results for each target user can be obtained.

[0062] S400: Performing reinforcement learning on the target user's learning feedback result and the target user's English learning level according to a reinforcement learning algorithm to dynamically adjust the target user's English learning level;

[0063] The learning feedback results of the target user and the English learning level of the target user are reinforced by a reinforcement learning algorithm to dynamically adjust the English learning level of the target user.

[0064] It should be understood that the English learning level of the target user is not fixed. It needs to be dynamically adjusted according to the learning feedback results so as to provide the target user with a more suitable teaching method to improve the learning effect of the target user.

[0065] S500, repeating the above steps according to the dynamically adjusted English learning level of the target user to regenerate the conversation data and obtain the learning feedback result of the target user;

[0066] In the embodiment of the present invention, the above steps are repeated based on the dynamically adjusted English learning level of the target user, that is, the generated dialogue data and the learning feedback result of the target user are repeatedly obtained.

[0067] S600: Evaluate the conversation data and the learning feedback result according to the large language model for data quality assessment, and obtain a conversation data evaluation result;

[0068] In an embodiment of the present invention, the learning feedback results of the target user are input into a large language model for data quality assessment to obtain a dialogue data assessment result.

[0069] S700: Store the dialogue data whose dialogue data evaluation results are greater than a preset threshold to obtain a mixed Chinese-English children's English dialogue dataset.

[0070] In an embodiment of the present invention, a judgment is made based on the evaluation results of the dialogue data obtained above, and dialogue data that is greater than a preset threshold is stored to obtain a mixed Chinese-English children's English dialogue dataset.

[0071] In summary, the method for constructing a mixed Chinese-English children's English dialogue dataset provided by the present invention divides the English learning levels based on the basic language level of the target users, and configures a corresponding question and answer prompt template for each English learning level, and then generates dialogue data through training. Through comprehension ability assessment and reinforcement learning mechanism, adaptive adjustment of data content difficulty is achieved, thereby constructing a personalized, semantically reasonable, and natural-style high-quality Chinese-English mixed teaching dataset to meet the real needs of children's English learning.

[0072] In an embodiment of the present invention, the English learning level is divided into multiple levels according to the basic language level of the target user, including:

[0073] According to the basic language proficiency of the target users, the English learning levels are divided into elementary English level, intermediate English level and advanced English level. The basic language level corresponding to the elementary English level includes mastering basic vocabulary, the basic language level corresponding to the intermediate English level includes the ability to understand simple sentences and be able to ask preliminary questions, and the basic language level corresponding to the advanced English level includes the ability to understand sentences with grammatical structures and be able to conduct basic daily communication.

[0074] Specifically, the embodiment of the present invention defines the target users as child learners aged 3-12 who are in the primary stage of learning English. According to their cognitive ability and language level, they are divided into three levels: Beginner: mastering basic vocabulary (such as colors, animals, and fruit names); Intermediate: having the ability to understand simple sentences and can ask preliminary questions; Advanced: able to understand sentences with grammatical structures and having basic daily communication skills.

[0075] Each level corresponds to a set of prompt templates, which are used to control the complexity and language style of the question and answer content generated by the large language model (LLM) to ensure that the generated content fits the learner's level.

[0076] In the embodiment of the present invention, dialogue data is generated according to each English learning level and the corresponding question and answer prompt template, such as Figure 2 Shown, including:

[0077] S210, determining a corresponding question and answer prompt template according to the English learning level of the current target user;

[0078] S220, obtaining question content corresponding to the English learning level of the current target user according to the question and answer prompt template and calling the large language model for question generation;

[0079] In this embodiment of the present invention, an appropriate Prompt template is selected from a built-in template library based on the learner's current level, and a large language model (such as GPT-4) is used to automatically generate natural language questions. For example, Prompt (Beginner): "Please ask a question about fruit words in Chinese and English, suitable for beginners."

[0080] S230: Inputting the question content corresponding to the current target user's English learning level into the teacher model to obtain an pedagogical answer result, wherein the pedagogical answer result output by the teacher model at least includes: English word explanation, supplemented by Chinese scope, and introduction of cultural or background knowledge;

[0081] In this embodiment of the present invention, the generated questions are used as input, and the teacher model (GPT-4) is called to generate pedagogical responses. Output requirements are set for the teacher model's generated content, including providing English word explanations, supplemented by Chinese translations, and introducing cultural / background knowledge (at an advanced stage).

[0082] S240. Conduct interactive logical question-answering based on continuous questioning, answering, and follow-up questions according to the actual teaching process to obtain dialogue data.

[0083] In the embodiment of the present invention, multi-round dialogue generation is supported, and by maintaining the context, the interactive logic of continuous questioning-answering-follow-up questions in the actual teaching process is simulated.

[0084] In an embodiment of the present invention, the dialogue data is input into a large language model for comprehension ability assessment to obtain learning feedback results of the target user, such as Figure 3 Shown, including:

[0085] S310: Input the conversation data into a large language model for comprehension ability assessment to obtain a target user's comprehension ability assessment result;

[0086] It should be understood that in order to achieve personalized adjustment of content, the dialogue data is input into the comprehension ability assessment module. The process is as follows: first, the student's questions and the teacher's answers are input into GPT-4, a judgment prompt is constructed, and whether it is understood ("yes / no"); secondly, the model returns the understanding score (probability) as "environmental feedback" in reinforcement learning.

[0087] S320: Generate a learning feedback result for the target user according to a comparison result of the target user's comprehension ability evaluation result and a preset comprehension evaluation threshold.

[0088] It should be understood that if the probability of understanding is higher than a set threshold (e.g., 0.85), the student is considered to be able to accept the current difficulty level; otherwise, the difficulty level should be lowered. This approach simulates the instant feedback mechanism in a real classroom, making the system adaptive in teaching.

[0089] In an embodiment of the present invention, reinforcement learning is performed on the learning feedback results of the target user and the English learning level of the target user according to a reinforcement learning algorithm, such as Figure 4 Shown, including:

[0090] S410: Determine the state space of the reinforcement learning algorithm according to the English learning level of the target user;

[0091] The present invention introduces the Q-learning reinforcement learning algorithm to dynamically adjust the student's language level. Specifically, the state space definition (S) is: {Beginner, Intermediate, Advanced}.

[0092] S420: Determine an action space of a reinforcement learning algorithm according to the learning feedback result of the target user;

[0093] Specifically, the action space (A): {increase difficulty, reduce difficulty, maintain current level}.

[0094] S430: updating and learning the state space and the action space according to a preset reward function to dynamically adjust the English learning level of the target user according to the learning feedback result of the target user.

[0095] Specifically, updating and learning the state space and the action space according to a preset reward function to dynamically adjust the English learning level of the target user according to the learning feedback result of the target user includes:

[0096] The state space and the dynamic space are updated and learned according to the Q-table update formula according to the preset reward function, wherein the expression of the preset reward function is:

[0097]

[0098] Among them, L represents the current reward, P understand represents the probability of understanding at this time, and θ represents the threshold value of the probability of understanding.

[0099] The expression of the Q table update formula is:

[0100]

[0101] Among them, α represents the learning rate (such as 0.1), γ represents the discount factor (such as 0.9), and s ′ Indicates the new state after the current state performs action a.

[0102] It should be understood that through continuous iteration, the system learns the optimal learning path for students and achieves fine-grained control of students' ability levels.

[0103] In an embodiment of the present invention, the conversation data and the learning feedback results are evaluated based on the large language model used for data quality assessment, and the conversation data evaluation results are obtained, including:

[0104] The dialogue data and the learning feedback results are input into the large language model for data quality assessment for data quality assessment, wherein the content of the data quality assessment at least includes: judging whether the questions and answers are pedagogically reasonable, whether they conform to the mixed Chinese and English standards, whether they have semantic coherence, and whether the language is natural and vivid.

[0105] Specifically, storing the conversation data whose conversation data evaluation result is greater than a preset threshold includes:

[0106] 1) quantifying the evaluation results of the conversation data to obtain a quantitative evaluation result;

[0107] 2) Compare the quantitative evaluation results with the preset threshold;

[0108] 3) If the quantitative evaluation result is greater than the preset threshold, the conversation data corresponding to the quantitative evaluation result is stored in a structured JSON file.

[0109] It should be understood that after each round of question and answer generation, GPT-4 is used to conduct the following evaluations: It determines whether the question and answer are pedagogically sound; whether they conform to the standard of mixed Chinese and English; whether they are semantically coherent; and whether the language is natural and vivid. A scoring mechanism is used to quantify this; those below a set threshold (e.g., 0.8) are discarded. High-quality data is organized and stored in a structured JSON file. After quality control, the generated question and answer data is stored in a standard structure.

[0110] The following experiments are combined to verify the effectiveness of the dataset constructed by the method for constructing a Chinese-English mixed children's English dialogue dataset in an embodiment of the present invention.

[0111] Specifically, in order to solve the technical problems of the current children's English teaching system in the alternating context of Chinese and English, such as the lack of personalized corpus, the inability to adaptively adjust the difficulty of generated content, and the high cost of manual construction, the method for constructing a mixed Chinese-English children's English dialogue dataset proposed in the present invention is suitable for deployment in intelligent educational robots, children's English learning apps and voice dialogue platforms, providing the system with high-quality, layered difficulty, and semantically natural teaching question and answer data.

[0112] (1) Dataset, base model and training method.

[0113] The training process of reinforcement learning is as follows: First, generate the student's questions. Based on the current learning level s, GPT-4 generates questions Q that match their cognitive ability. s The second step is to generate the foreign teacher’s answer. GPT-4 generates the answer based on the student’s question Q s Generate foreign teacher answer A s . Student understanding will be assessed along the way: GPT-4 assesses students’ understanding of answering A s The student's understanding of the sentence is evaluated by GPT-4, and the student's understanding probability p is obtained. Then the reward R is calculated:

[0114]

[0115] Here, θ represents the understanding threshold.

[0116] At this point, action selection is performed, a=π(s) according to the ε-greedy strategy, and the selection is increase, decrease or stay. Then the execution state transition is performed: s ′ =f(s,a) where f(s,a) calculates the new state based on the current difficulty and action, and updates the Q-table:

[0117]

[0118] Training is then repeated until the Q-value converges. After training, the reinforcement learning algorithm can dynamically adjust the difficulty of English learning for the learner, keeping it within the most suitable challenge range. Compared to fixed rule methods, this method has many advantages. First, it can adjust the difficulty according to the individual learner's situation, rather than a fixed upgrade or downgrade. Second, through iterative optimization using Q-learning, it can find the difficulty adjustment method that maximizes learning results and avoid learning stagnation. Finally, the ε-greedy strategy avoids drastically adjusting the difficulty due to a single misassessment, but instead optimizes gradually.

[0119] To ensure fairness in the experiment, this embodiment of the present invention uses two datasets: the Baseline dataset, which is raw Chinese-English mixed teaching data generated directly from a large model without any optimization, and the Adaptive dataset, which uses reinforcement learning to adjust the difficulty to better match the learner's language level.

[0120] The data was divided into a training set (80% for model training) and a test set (20% for model evaluation). We ensured that all experimental configurations used the same training / test sets to eliminate the influence of data randomness.

[0121] Since the experiment of the embodiment of the present invention only evaluates the data set and does not involve the requirements of specific downstream tasks, in order to save resources, the T5-small, T5-base, and GPT-2 small models are finally selected for training, and the text is converted using T5Tokenizer to obtain input_ids and labels. During the training process, the AdamW optimizer is used for training for 3 rounds (Epochs). The loss value (Loss) changes of each Epoch are recorded. After the training is completed, the model weights are saved for evaluation.

[0122] (2) Evaluation indicators: This embodiment of the present invention uses BERTScore and Distinct-n as dataset evaluation indicators.

[0123] BERTScore is a semantic matching evaluation metric based on BERT, suitable for free text generation tasks such as machine translation, text summarization, and open text generation. BERTScore calculates the similarity between the generated text and the reference text in the semantic space, rather than a simple n-gram match:

[0124]

[0125] Where P is the generated text, and R is the reference text. E(x) is the word embedding of Tokenx calculated using BERT. cosine(E(x),E(y)) calculates the cosine similarity between Tokenx and Tokeny in the reference text. For each generated Tokenx, find the most similar reference Tokeny for matching.

[0126] Distinct-n is an indicator used to evaluate the diversity of text generation. It mainly measures the proportion of unique n-grams (n consecutive words) in the text generated by the model to prevent the model from generating duplicate content. The main purpose of the embodiment of the present invention is to verify whether a large amount of duplicate content appears in the dataset and whether it affects the subsequent training of the model.

[0127] Distinct-n calculates the uniqueness of n-grams in the generated text. The formula is as follows:

[0128] In the formula,

[0129] Where: |Unique n-grams| is the number of unique n-grams in the generated text. |Total n-grams| is the total number of all n-grams in the generated text.

[0130] If the text has no repetitions at all, Distinct-n = 1 (i.e., all n-grams are unique). If the text is completely repetitive, Distinct-n = 0 (i.e., there is no new content in the n-gram).

[0131] (3) Experimental results

[0132] Table 1 Experimental results of T5-small model

[0133]

[0134] Table 1 shows that compared to the baseline, the Full RL configuration achieves comprehensive improvements on the T5-small model, with BERTScore increasing by 0.08 and Distinct-2 increasing by 0.13. This indicates that reinforcement learning, combined with probabilistic understanding and Q-learning strategies, effectively improves the semantic coherence and expressive diversity of generated data.

[0135] The UP-only (Understanding Probability Optimization) alone improves the BERTScore by 0.04, indicating that this method focuses more on the semantic quality of the text, while Q-only (Q-learning Optimization) focuses more on improving Distinct-2, increasing text diversity. This also shows that Q-learning alone, while increasing diversity, may not meet the actual understanding ability of students, and only by combining understanding probability can the optimal balance be achieved.

[0136] Table 2 T5-base model experimental results

[0137]

[0138] As can be seen from Table 2, the overall performance of T5-base model is higher than that of T5-small, which shows that increasing the model size helps to improve the generation effect, especially in semantic coherence (BERTScore), which performs more outstandingly.

[0139] The improvement effect of Full RL strategy is more significant, with a 0.09 BERTScore and a 0.14 Distinct-2 improvement over Baseline, indicating that the mechanism combined with reinforcement learning is more suitable for larger language models, effectively utilizing the stronger generalization ability of T5-base.

[0140] The difference between Q-only and UP-only on this model is smaller, indicating that larger models are more likely to balance quality and diversity, and the limitations of single strategies have been weakened.

[0141] Table 3 GPT-2 model experimental results

[0142]

[0143] As can be seen from Table 3, the GPT-2 model has the lowest initial performance among all models, with Distinct-2 only 0.50, indicating a lack of text diversity and a significant repetition problem in generated text. After optimization by reinforcement learning strategy, Distinct-2 is improved by 0.20, indicating that reinforcement learning is particularly effective in improving text diversity for smaller models.

[0144] The overall improvement effect of Full RL strategy on GPT-2 is the most prominent, indicating that GPT-2 is prone to pattern generation in pure data-driven training, and the addition of reinforcement learning effectively makes up for this deficiency.

[0145] Although the GPT-2 model itself is small in size and has fewer parameters, it can still effectively improve the quality of generated data after proper optimization, indicating that reinforcement learning is highly generalizable and adaptable.

[0146] As Figure 5 shown in the following figures, the trends of BERTScore and Distinct-2 for T5-small, T5-base and GPT-2 under different experimental configurations (Baseline, UP-only, Q-only and Full RL) are presented. From the figures, the following points can be clearly observed:

[0147] Firstly, in all three models, as the experimental configuration gradually transitions from Baseline to Full RL, both BERTScore and Distinct-2 indicators show a clear upward trend. This trend indicates that the reinforcement learning mechanism has a significant promoting effect on improving the generation quality and diversity of the dataset.

[0148] Secondly, further analysis of the performance of different models shows that the T5-base model outperforms the T5-small model and the GPT-2 model, and its indicators improve significantly after optimization by the reinforcement learning strategy (Baseline to Full RL, BERTScore improves by 0.09, and Distinct-2 improves by 0.14). This may be related to the fact that the larger model has better ability to capture language semantic details and generate diverse expressions.

[0149] At the same time, although the GPT-2 model has the lowest initial performance, the reinforcement learning strategy has the largest performance improvement, especially in the Distinct-2 indicator (from 0.50 in Baseline to 0.70 in Full RL, with an increase of 0.20). This shows that in smaller models, reinforcement learning is particularly effective in overcoming the phenomenon of patterned generation (single and repetitive generation results).

[0150] Finally, by comparing the performance of UP-only (only understanding probability optimization) and Q-only (only Q-learning optimization), it is found that the UP-only strategy is more obvious in improving the BERTScore indicator (text semantic quality), while the Q-only strategy contributes more to the Distinct-2 indicator (text diversity). This shows that pure understanding probability adjustment mainly improves the semantic coherence of the text, while the Q-learning mechanism tends to encourage the model to explore diverse language expression methods.

[0151] In summary, the complete reinforcement learning strategy (Full RL) combining understanding probability and Q-learning shows the best comprehensive effect, which can improve both the semantic quality and the expression diversity of the text, and has significant advantages and high practical value in the construction of children's English teaching datasets.

[0152] (4) Data diversity analysis experiment. The experiment of the embodiment of the present invention is intended to verify the language diversity and coverage of the generated dialogues in the constructed Chinese-English mixed dataset, so as to ensure that the data can effectively support the model learning of rich expressions in real teaching scenarios. In the model training and evaluation stage, the embodiment of the present invention uses Distinct-2 as an indicator, which is based on the fact that this indicator is widely used in text generation tasks and is sufficient to reflect the diversity of model output. However, in the data diversity analysis experiment, we further used Distinct-3 and Distinct-4 indicators. This is because higher-order n-gram analysis can more deeply reveal the richness of the language expression of the dataset itself and the diversity of high-order language patterns, which helps to more comprehensively reflect the quality of the dataset.

[0153] The steps for data preprocessing are shown in Table 4 below.

[0154] Table 4 Data diversity analysis experiment preprocessing steps

[0155]

[0156] The embodiment of the present invention will calculate the word richness (VR), Distinct-3, and Distinct-4 of the dataset to ensure the language diversity and coverage of the dataset.

[0157] The calculation formula for word richness is:

[0158]

[0159] Among them, N uniquetokens Indicates the number of non-repeated words in the dataset, N toialtokens Represents the total number of all words in the dataset.

[0160] To focus more on the fine-grained features of the data itself, Distinct-3 and Distinct-4 can better reflect the repetition or patterning issues in longer sentences or conversations in text data, reflecting the real-world usage scenarios of the dataset and the richness of the text.

[0161] Table 5 Data diversity analysis experimental results

[0162]

[0163] Experimental results show that the dataset's lexical richness is 0.45, indicating that it contains a rich and diverse vocabulary, effectively avoiding the problem of text homogeneity. Distinct-3 is 0.65, and Distinct-4 is 0.52, indicating that the dataset's high-order n-gram expressions are highly diverse, enabling the model to learn a wider range of language patterns. Furthermore, the dataset has an average number of conversational turns of 4.5, and an average sentence length of 22.7 words, demonstrating that the dialogue scenarios are well-designed and the sentences are diverse, more realistically simulating actual teaching environments.

[0164] Figure 6 This chart shows the vocabulary richness distribution in the dataset. Approximately 45% of the vocabulary is unique, demonstrating the diversity of vocabulary usage in the dataset. A higher proportion of unique vocabulary helps improve the generalization ability of model learning and the richness of language expression, preventing the model from learning overly templated and repetitive content.

[0165] Figure 7 The figure shows the distribution of conversation turns. The histogram shows that most conversations are concentrated between 3 and 6 turns, indicating that the conversation lengths designed in the dataset are suitable for real-world teaching scenarios. This distribution of conversation lengths helps maintain learners' focus and interactive interest, and is consistent with the typical characteristics of teacher-student communication in real classroom teaching environments.

[0166] Figure 8 The following chart shows the sentence length distribution. Sentence lengths are primarily concentrated between 18 and 28 words, with an average length of 22.7 words, consistent with authentic teaching dialogue contexts. A moderate sentence length helps learners understand the meaning of sentences, avoiding the lack of information due to short sentences or the difficulty of understanding due to long sentences, thereby effectively improving learning outcomes.

[0167] This embodiment of the present invention presents the diversity of language expressions and structures in the dataset through in-depth analysis and visualization across multiple dimensions, comprehensively verifying the dataset's quality in terms of language richness, expression diversity, and structural rationality. This helps ensure that the dataset can better simulate the real-world English teaching environment for children, providing the model with richer and more authentic learning resources.

[0168] In summary, compared with the prior art, the present invention has the following significant advantages:

[0169] 1) Realize personalized and adaptive learning path control, breaking the "static bottleneck" of traditional data construction

[0170] Most of the existing methods for generating mixed Chinese-English corpus are "unified template, one-time generation", which lack the ability to dynamically adjust the content according to students' cognitive abilities, and are prone to teaching bottlenecks such as "content is too difficult / too easy" and "discontinuous understanding".

[0171] This invention introduces a learner modeling mechanism, categorizing students into three levels: beginner, intermediate, and advanced. Combined with language model prompt control, it ensures that the question and answer content accurately matches the student's current language level. More importantly, through a reinforcement learning (Q-learning) strategy, the learning level is adjusted in real time, increasing or decreasing the difficulty level based on the student's understanding, achieving true "teaching students in accordance with their aptitude" and "path adaptation."

[0172] This mechanism not only improves the teaching adaptability of data, but also provides a paradigm reference for the construction of "corpus generation with feedback mechanism" in the next generation of intelligent teaching systems.

[0173] 2) Significantly reduce manual labeling costs and improve data construction efficiency and scale expansion capabilities.

[0174] Traditional corpus construction often requires a large number of teaching experts to design question-and-answer content one by one. This is not only time-consuming and labor-intensive, but also prone to inconsistent data styles due to subjective bias. Although some studies have introduced LLM to generate corpora, they still rely on extensive manual post-processing and screening, which limits the scalability of the data and its engineering application capabilities.

[0175] This invention achieves the automated and batch construction of high-quality Chinese and English teaching materials through precise prompt template guidance, multiple rounds of dynamic context generation, automatic evaluation module quality control, and reinforcement learning to automatically select content trends. It basically eliminates manual dependence and supports the construction of millions of Chinese-English dialogue pairs. In actual experiments, it saves more than 90% of labor costs and greatly improves the efficiency of corpus development.

[0176] 3) Fill the gap in authentic teaching materials in the “Chinese-English mixed context” and meet the needs of cross-language alternating expression.

[0177] In real teaching scenarios, Chinese and English appear alternately frequently, especially in the basic education stage (K12). Teachers often use Chinese to explain English vocabulary and guide students to naturally transition to pure English expression.

[0178] Most of the existing corpus is original English or manually translated, and lacks real context data that conforms to the laws of language mixing. This makes it difficult for the trained model to handle switching conversations between Chinese and English, affecting its adaptability in actual educational scenarios.

[0179] Starting from the natural scenarios of teaching language, this invention designs prompts and model generation strategies to automatically generate question-and-answer content containing a "Chinese-English-Chinese" structure, such as: "Do you know what 'apple' means? Apple is a sweet fruit." This simulates the natural classroom language style, improves the semantic transition of data and teaching practicality, and can be widely used in real scenarios such as bilingual assistants and children's AI teachers.

[0180] 4) Reinforcement learning drives iterative optimization of data to form a high-quality closed-loop feedback mechanism.

[0181] Traditionally, data generation has been triggered by static templates, lacking the ability to learn and optimize. Even with the introduction of automatic language model generation, it remains a linear process of "one-time generation followed by manual screening," making it difficult to adapt to changes in student cognition or improve content strategies.

[0182] This method uses the Q-learning reinforcement learning algorithm to construct a "state-action-reward" structure, using the student's probability of understanding as feedback to optimize the generation path and difficulty adjustment strategy. Through multiple rounds of training, the model can self-learn how to select the most appropriate content, maintaining a dynamic balance between generated content and student ability.

[0183] This mechanism can be seen as a generation system with "teaching planning capabilities", providing a more intelligent, data-driven optimization method for educational data generation.

[0184] 5) Multiple evaluation mechanisms ensure data quality and ensure that the questions and answers are accurate, natural and coherent.

[0185] After data generation, the present invention introduces the GPT-4 evaluation module to perform quality scoring on each set of questions and answers based on the following dimensions: whether the question and answer are logically matched, whether the answer covers core knowledge points, whether the expression is natural, whether the style is consistent with the teaching scenario, whether it is consistent with the current learning level, and whether it contains high-quality Chinese-English switching expressions.

[0186] The system sets a scoring threshold, automatically removing data with a plausibility score of less than 0.8, ensuring that only high-quality data remains in the corpus. The evaluation process can be automated, supporting large-scale data screening and improving the accuracy and teaching practicality of the overall dataset.

[0187] 6) It has good scalability and implementation potential, and is suitable for multi-language, multi-task teaching AI systems.

[0188] This method can not only be used for children's English teaching, but can also be transferred to the following fields: generation of mixed teaching materials of Chinese and Japanese / Chinese and Korean, alternating training of mathematical logic and spoken English, generation of mixed-style multi-round materials for fine-tuning of dialogue systems / large models, and construction of multilingual scenarios for cross-language customer service / question and answer / guided tours.

[0189] At the same time, the present invention designs a standardized JSON storage structure and a modular pluggable architecture, which facilitates rapid integration into existing teaching robots, dialogue systems, and educational model fine-tuning processes. It has strong reusability and engineering controllability, and has the ability to be implemented in the industry and the prospects for commercial promotion.

[0190] As another embodiment of the present application, a storage medium is provided, wherein a computer instruction is stored, and when the computer instruction is loaded and executed by a processor, the above-mentioned method for constructing a mixed English and Chinese dialogue data set for children is implemented.

[0191] In the embodiments of the present application, a non-transitory computer readable storage medium is provided, which stores computer executable instructions, and the computer executable instructions are used to execute the method for constructing a mixed English and Chinese dialogue data set for children in any of the above-mentioned embodiments. The storage medium can be a disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD), etc. The storage medium can also include a combination of the above-mentioned types of memories.

[0192] As another embodiment of the present application, a mixed English and Chinese teaching system for children is provided, which comprises a memory and a processor, the memory is in communication connection with the processor, the memory is used to store computer instructions, and the processor is used to load and execute the computer instructions to implement the above-mentioned method for constructing a mixed English and Chinese dialogue data set for children.

[0193] The mixed English and Chinese teaching system for children provided by the present application can divide the English learning levels based on the basic language level of the target user by executing the above-mentioned method for constructing a mixed English and Chinese dialogue data set for children through the computer instructions stored in the memory, and can configure the corresponding question and answer prompt templates for each English learning level, and then generate dialogue data through training, and realize the self-adaptive adjustment of the difficulty of the data content through the understanding ability evaluation and reinforcement learning mechanism, so as to construct a personalized, semantically reasonable and naturally styled mixed English and Chinese teaching data set, thereby meeting the real needs of children's English learning.

[0194] It can be understood that the above-mentioned embodiments are only exemplary embodiments for illustrating the principles of the present application, and the present application is not limited thereto. Various modifications and improvements can be made by those skilled in the art without departing from the spirit and essence of the present application, and these modifications and improvements are also considered to be within the protection scope of the present application.

Claims

1. A method for constructing a mixed Chinese-English children's English dialogue dataset, characterized by: include: The English learning level is divided into multiple levels according to the basic language proficiency of the target users, and a corresponding question and answer prompt template is configured for each English learning level; Generate conversation data based on each English learning level and the corresponding question-answer prompt template; Inputting the conversation data into a large language model for comprehension ability assessment to obtain learning feedback results of the target user; Performing reinforcement learning on the target user's learning feedback results and the target user's English learning level according to a reinforcement learning algorithm to dynamically adjust the target user's English learning level; Repeat the above steps according to the dynamically adjusted English learning level of the target user to regenerate the conversation data and obtain the learning feedback results of the target user; Evaluate the conversation data and the learning feedback result according to the large language model used for data quality assessment, and obtain a conversation data evaluation result; The dialogue data whose dialogue data evaluation results are greater than a preset threshold are stored to obtain a mixed Chinese-English children's English dialogue data set.

2. The method for constructing a mixed Chinese-English children's English dialogue dataset according to claim 1 is characterized in that: English learning levels are divided into multiple levels according to the basic language level of the target users, including: According to the basic language proficiency of the target users, the English learning levels are divided into elementary English level, intermediate English level and advanced English level. The basic language level corresponding to the elementary English level includes mastering basic vocabulary, the basic language level corresponding to the intermediate English level includes the ability to understand simple sentences and be able to ask preliminary questions, and the basic language level corresponding to the advanced English level includes the ability to understand sentences with grammatical structures and be able to conduct basic daily communication.

3. The method for constructing a mixed Chinese-English children's English dialogue dataset according to claim 1, characterized in that: Generate conversation data based on each English learning level and the corresponding question and answer prompt template, including: Determine the corresponding question and answer prompt template based on the English learning level of the current target user; According to the question-answer prompt template, the large language model used for question generation is called to obtain question content corresponding to the English learning level of the current target user; Inputting the question content corresponding to the current target user's English learning level into the teacher model to obtain an pedagogical answer result, wherein the pedagogical answer result output by the teacher model at least includes: English word explanation, supplemented by Chinese scope, and introduction of cultural or background knowledge; According to the actual teaching process, interactive logical question-answering based on continuous questioning, answering and follow-up questions is conducted to obtain dialogue data.

4. The method for constructing a mixed Chinese-English children's English dialogue dataset according to claim 1, characterized in that: Inputting the conversation data into a large language model for comprehension assessment to obtain learning feedback results from the target user, including: Inputting the conversation data into a large language model for comprehension ability assessment to obtain a target user comprehension ability assessment result; A learning feedback result of the target user is generated according to a comparison result of the target user's comprehension ability evaluation result and a preset comprehension evaluation threshold.

5. The method for constructing a mixed Chinese-English children's English dialogue dataset according to claim 1, characterized in that: Performing reinforcement learning on the target user's learning feedback results and the target user's English learning level according to a reinforcement learning algorithm, including: Determine the state space of the reinforcement learning algorithm according to the English learning level of the target user; Determining an action space of a reinforcement learning algorithm based on the learning feedback results of the target user; The state space and the action space are updated and learned according to a preset reward function to dynamically adjust the English learning level of the target user according to the learning feedback result of the target user.

6. The method for constructing a mixed Chinese-English children's English dialogue dataset according to claim 5, characterized in that: The state space and the action space are updated and learned according to a preset reward function to dynamically adjust the English learning level of the target user according to the learning feedback result of the target user, including: The state space and the dynamic space are updated and learned according to the Q-table update formula according to the preset reward function, wherein the expression of the preset reward function is: The expression of the Q table update formula is: Among them, α represents the learning rate, γ represents the discount factor, and s ′ Indicates the new state after the current state performs action a.

7. The method for constructing a mixed Chinese-English children's English dialogue dataset according to claim 1, characterized in that: The conversation data and the learning feedback results are evaluated according to the large language model used for data quality assessment, and a conversation data evaluation result is obtained, including: The dialogue data and the learning feedback results are input into the large language model for data quality assessment for data quality assessment, wherein the content of the data quality assessment at least includes: judging whether the questions and answers are teaching reasonable, whether they comply with the mixed Chinese and English standards, whether they have semantic coherence, and whether the language is natural and vivid.

8. The method for constructing a mixed Chinese-English children's English dialogue dataset according to claim 7, characterized in that: Storing the conversation data whose conversation data evaluation result is greater than a preset threshold includes: Quantitatively processing the dialogue data evaluation result to obtain a quantitative evaluation result; Comparing the quantitative evaluation results with a preset threshold; If the quantitative evaluation result is greater than the preset threshold, the conversation data corresponding to the quantitative evaluation result is stored in a structured JSON file.

9. A storage medium, characterized in that: Used to store computer instructions, when the computer instructions are loaded and executed by a processor to implement the method for constructing a Chinese-English mixed children's English dialogue dataset as described in any one of claims 1 to 8.

10. A Chinese-English mixed children's English teaching system, characterized by: The method comprises a memory and a processor, wherein the memory is communicatively connected to the processor, the memory is used to store computer instructions, and the processor is used to load and execute the computer instructions to implement the method for constructing a Chinese-English mixed children's English dialogue dataset as described in any one of claims 1 to 8.

Images