A poem beginner multi-task cooperative training method based on an open source large model

Abstract

The application discloses a kind of based on open source big model's poetry beginner multi-task cooperative training method, comprising: by knowledge isolation module, the semantic classification of user input text is carried out, when identifying as poetry field content, the preset knowledge base call of open source big model is blocked, and enters incremental learning procedure, when identifying as daily communication content, response is generated directly based on open source big model, and the scheme can identify whether user is learning or daily communication by six module cooperative linkage, and the preset knowledge base call of open source big model is blocked as appropriate, solve the defect that existing model learns before knowing, guarantee progressive learning, improve the effect of poetry learning, realize that child is as teacher to teach AI doll this beginner learning knowledge, simulate real learning process, let child incarnate teacher, doll incarnate beginner, with output instead of input, child can more profound repetition and understanding knowledge in the process of repetition.

Description

Technical Field

[0001] This invention relates to the field of poetry learning technology, specifically to a multi-task collaborative training method for poetry beginners based on an open-source large model. Background Technology

[0002] Currently, AI dialogue models typically have pre-set large-scale general knowledge bases as well as knowledge bases for specific poetry fields, allowing for direct use of pre-set knowledge to respond during interaction.

[0003] However, the above-mentioned dialogue interaction still has the following defects: In real learning, learners should learn from scratch and complete the gradual learning of knowledge through interactive dialogue. However, in the current model, because of its pre-built knowledge base, it informs learners of knowledge in advance during the dialogue interaction of poetry. For example, if the learner says the previous line of the poem, the model will directly output the next line, which makes the learner know before learning and acquire knowledge that has not been learned in advance, resulting in a decline in learning effectiveness. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a multi-task collaborative training method for poetry beginners based on an open-source large model, thus solving the aforementioned problems.

[0005] The above-mentioned technical objective of the present invention is achieved through the following technical solution:

[0006] A multi-task collaborative training method for poetry beginners based on an open-source large model includes:

[0007] Step S1: The user input text is semantically classified through the knowledge isolation module. When it is identified as content in the poetry domain, the call to the pre-built knowledge base of the open source big model is blocked, and the incremental learning process is started. When it is identified as daily communication content, the response is directly generated based on the open source big model.

[0008] Step S2: Based on the classification results of S1, the incremental learning module triggers the storage of the poem when a teaching instruction is detected. The poem text and the corresponding speech features are stored in the incremental knowledge base, and a unique index is generated.

[0009] Step S3: Through the passive receiving module, based on the storage trigger in S2, the knowledge authenticity verification is skipped and the poem text is stored directly. The poem text includes erroneous content.

[0010] Step S4: Using the accent imitation module, based on the speech features stored in S2, the accent parameters of the speech features are extracted, and the accent parameters are applied during speech output to generate a teacher's speech with distortion.

[0011] Step S5: Using the forgetting curve module, recalculate the memory strength value based on the poem text stored in S2, and calculate the retrieval response time based on the memory strength value;

[0012] Step S6: Through the review and reinforcement module, analyze the user's review behavior based on the memory strength value in S5, update the global repetition count, correct the memory strength value and response time, and adjust the priority of the poem in the search engine.

[0013] Further, step S1 includes:

[0014] Semantic parsing of user input text, and calculation of domain confidence scores based on a poetry keyword database and contextual semantic features;

[0015] Based on the domain confidence score, dynamically adjust the triggering conditions of the knowledge invocation interception layer:

[0016] When the domain confidence score is greater than the preset similarity threshold, it is judged to be poetry content, triggering a knowledge call prohibition signal;

[0017] When the domain confidence score is less than or equal to the preset similarity threshold, it is determined to be daily communication content. A knowledge call permission signal is sent, and a response is generated normally based on the open-source large model. The output content of the poetry domain is then verified a second time, and content that matches the preset knowledge base is removed to ensure that the output of the poetry domain only depends on the incremental learning module.

[0018] Further, step S2 includes:

[0019] When the user input text is determined to be content in the poetry domain, the user input text is parsed, the matching degree of instruction words and contextual intent are analyzed, and the intensity value of teaching intent is obtained.

[0020] The intensity value of teaching intention is calculated to obtain the intention threshold;

[0021] The poetry storage mechanism is dynamically triggered based on the intensity of the teaching intention:

[0022] When the intensity of teaching intent is greater than the intent threshold, the text and speech features of the poem are stored in the incremental knowledge base in the form of key-value pairs, and a unique hash value is generated for each poem as an index.

[0023] When a user asks a question, only the single-sentence index already stored in the incremental knowledge base is retrieved; if no content is stored, an unlearned response is returned.

[0024] Based on the domain confidence score and the strength of teaching intent, the storage priority of the incremental learning module and the knowledge isolation module are adjusted.

[0025] Furthermore, based on domain confidence scores and the strength of instructional intent, the storage priority of the incremental learning module and the knowledge isolation module are adjusted, including:

[0026] After the knowledge isolation module completes the content classification, the cross-module collaboration factor is calculated based on the domain confidence score and the intensity value of teaching intention.

[0027] Based on cross-module collaboration factors, the storage priority of the incremental learning module and the interception strength of the knowledge isolation module are dynamically adjusted.

[0028] Furthermore, semantic parsing is performed on the user-input text, and a domain confidence score is calculated based on a poetry keyword database and contextual semantic features, including:

[0029] Analyze user-input text to generate semantic field intensity values;

[0030] Analyze semantic fragments related to daily conversations in user input text, calculate the degree of interference of these semantic fragments on the semantics of poetry, and obtain the semantic penetration factor;

[0031] The semantic field strength value and semantic penetration factor are calculated to generate a domain confidence score.

[0032] Furthermore, the user input text is parsed to analyze the matching degree of instruction words and contextual intent, thereby obtaining a teaching intent intensity value, including:

[0033] Analyze the dispersion of teaching behaviors in user input text to generate an intent clarity index;

[0034] Each round of user interaction is mapped to a direction vector. By analyzing the composite direction and intensity of dialogue vectors from several rounds, teaching momentum is generated.

[0035] The intention clarity index and teaching momentum are asymmetrically integrated to generate a teaching intention intensity value.

[0036] Furthermore, step S3 includes:

[0037] Based on the S2-triggered storage mechanism, the knowledge authenticity verification function of the open-source large model for the poem text is disabled, and the comparison link between the poem text and the pre-built knowledge base is deleted.

[0038] The system adopts an input-is-correct storage logic, directly storing the text of the poem entered by the user into the incremental knowledge base in its original form, even if the text of the poem contains obviously erroneous content, and marking it as a valid knowledge point entered by the teacher;

[0039] When a user asks for a response, reproduce the stored original text content.

[0040] Further, step S4 includes:

[0041] Based on the speech features stored in S2, core features are extracted from the speech features, the core features are transformed into accent parameters, and a correlation mapping between the poem and the accent parameters is constructed.

[0042] During speech output, corresponding distortion processing is applied to the synthesized speech based on accent parameters to ensure that the output speech maintains the distortion characteristics consistent with the teacher's accent.

[0043] Furthermore, step S5 includes:

[0044] Based on the timestamps of the poetry texts stored in S2, an initial memory strength value is set for each stored poetry text.

[0045] The incremental knowledge base is traversed periodically at preset time intervals to calculate the current memory strength value of each line of poetry.

[0046] Establish a correlation between memory strength value and response time, and dynamically adjust the response time when calculating the text of the poem based on the current memory strength value.

[0047] Furthermore, step S6 includes:

[0048] Based on memory strength values, we analyze users' review behaviors, which include explicit and implicit review behaviors.

[0049] After recognizing the review behavior, update the global repetition count of the corresponding poem line. Based on the updated global repetition count, recalculate the memory strength value of the poem line and overwrite the original memory strength value.

[0050] Based on the recalculated memory strength value, the response time is recalculated, and the retrieval priority of the poem in the search engine is increased.

[0051] In summary, the present invention has the following main beneficial effects:

[0052] The knowledge isolation module calculates domain confidence scores to accurately classify poetry and everyday communication content. It blocks the use of the pre-built knowledge base of the open-source model for poetry content and performs secondary verification on the output content to remove content that matches the pre-built knowledge base, thus preventing the premature leakage of unlearned knowledge. The incremental learning module precisely triggers the storage of poetry verses based on the intensity of teaching intent, and dynamically adjusts the storage priority and the interception strength of the knowledge isolation module through cross-module collaborative factors to ensure the targeted and adaptable storage.

[0053] By using a passive learning module to close the knowledge authenticity verification and deletion comparison link with the pre-set knowledge base, the system stores erroneous poem texts as original content and marks them as valid knowledge points, ensuring the original presentation of teacher input and the authenticity of learning. The accent imitation module extracts core features and converts them into unique accent parameters to generate a distorted teacher's voice, enhancing the interactive immersion. The forgetting curve module dynamically calculates memory intensity and response time at preset time intervals, conforming to the laws of natural memory. The review and reinforcement module accurately identifies explicit and implicit review behaviors and adjusts response time and retrieval priority based on global repetition counts and memory intensity values, enhancing the learning effect, ensuring the implementation of a progressive learning path, improving the scientific nature and immersion of learning, optimizing the actual effect of poetry learning, and training an open-source large model as a poetry novice, allowing children to act as teachers and teach AI puppets (beginners) to learn knowledge, simulating a real learning process. Children become teachers, and puppets become beginners, with output replacing input. Through repetition, children can more deeply repeat and understand knowledge. Attached Figure Description

[0054] Figure 1 This is a flowchart of a multi-task collaborative training method for poetry beginners based on an open-source large model, according to the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] refer to Figure 1 A multi-task collaborative training method for poetry beginners based on an open-source large model, including:

[0057] Step S1: The user input text is semantically classified through the knowledge isolation module. When it is identified as content in the poetry domain, the call to the pre-built knowledge base of the open source big model is blocked, and the incremental learning process is started. When it is identified as daily communication content, the response is directly generated based on the open source big model.

[0058] Step S2: Based on the classification results of S1, the incremental learning module triggers the storage of the poem when a teaching instruction is detected. The poem text and the corresponding speech features are stored in the incremental knowledge base, and a unique index is generated.

[0059] Step S3: Through the passive receiving module, based on the storage trigger in S2, the knowledge authenticity verification is skipped, and the poem text is directly stored. The poem text includes erroneous content and is marked as valid knowledge points.

[0060] Step S4: Using the accent imitation module, based on the speech features stored in S2, the accent parameters of the speech features are extracted, and the accent parameters are applied during speech output to generate a teacher's speech with distortion.

[0061] Step S5: Using the forgetting curve module, recalculate the memory strength value based on the poem text stored in S2, and calculate the retrieval response time based on the memory strength value;

[0062] Step S6: Through the review and reinforcement module, analyze the user's review behavior based on the memory strength value in S5, update the global repetition count, correct the memory strength value and response time, and adjust the priority of the poem in the search engine.

[0063] The open-source model includes modules for knowledge isolation, incremental learning, passive learning, accent imitation, forgetting curve, and review and reinforcement.

[0064] In one embodiment, step S1 involves semantically classifying the user-input text using a knowledge isolation module. When the text is identified as poetry, the pre-built knowledge base of the open-source large model is blocked, and an incremental learning process begins. When the text is identified as everyday communication content, a response is directly generated based on the open-source large model, including:

[0065] Semantic parsing of user input text, and calculation of domain confidence scores based on a poetry keyword database and contextual semantic features;

[0066] Based on the domain confidence score, dynamically adjust the triggering conditions of the knowledge invocation interception layer:

[0067] When the domain confidence score is greater than the preset similarity threshold, it is judged as poetry content, triggering a knowledge access prohibition signal. At this time, the open-source big model will block access to the preset knowledge base. The preset similarity threshold is 65 points. If the preset similarity threshold is too low, the ability to judge poetry will be reduced. Once a misjudgment occurs, it will incorrectly block access to the big model's preset knowledge base. If the preset similarity threshold is too high, the judgment will become very conservative. Only sentences with neat format and clear keywords can be recognized. Therefore, the preset similarity threshold is 65 points.

[0068] When the domain confidence score is less than or equal to the preset similarity threshold, it is determined to be daily communication content. A knowledge call permission signal is sent, and a response is generated normally based on the open-source large model. This allows for daily communication, and the output content is verified. Content that matches the preset knowledge base is removed to ensure that the output in the poetry domain relies solely on the incremental learning module.

[0069] By calculating the domain confidence score, accurate identification of poetry content is achieved. For the identified poetry content, the call to the pre-built knowledge base of the open-source model is blocked, which solves the problem that learners know the knowledge before learning it due to the pre-built knowledge base of the existing model. This ensures that learners with no prior knowledge can complete the progressive learning through interactive dialogue. For daily communication content, the open-source model is directly called to generate responses, and the output is verified to remove content that matches the pre-built knowledge base. This ensures that the output in the poetry domain relies only on the incremental learning module, while also taking into account the naturalness of daily communication, thus improving the actual effect and interactive adaptability of poetry learning.

[0070] In one embodiment, step S2 involves the incremental learning module, based on the classification results of S1, triggering the storage of the poem upon detecting a teaching instruction. This involves storing the poem text and its corresponding speech features into the incremental knowledge base and generating a unique index, including:

[0071] When the user input text is determined to be content in the poetry domain, the user input text is parsed, the matching degree of instruction words and contextual intent are analyzed, and the intensity value of teaching intent is obtained.

[0072] The teaching intention intensity value is calculated to obtain the intention threshold. Specifically, for the domain confidence score, when the domain confidence score is 100, the corresponding intention coefficient is 0.1. For every 10 points decrease in the domain confidence score, the intention coefficient increases by 0.1, and so on. The teaching intention intensity value is multiplied by the intention coefficient to obtain the intention threshold.

[0073] The poetry storage mechanism is dynamically triggered based on the intensity of the teaching intention:

[0074] When the intensity of teaching intent is greater than the intent threshold, the text and speech features of the poem are stored in the incremental knowledge base as key-value pairs, and a unique hash value is generated for each poem as an index; otherwise, they are not stored.

[0075] Among them, a bimodal incremental knowledge base is constructed using poetry and speech;

[0076] When a user asks a question, that is, when learning poetry, only the single-sentence index stored in the incremental knowledge base is retrieved. If the content is not stored, an "unlearned" response is returned, which is to output a voice prompt that has not been learned.

[0077] Based on the domain confidence score and the strength of teaching intent, the storage priority of the incremental learning module and the knowledge isolation module are adjusted.

[0078] Through the incremental learning module, the intent coefficient is dynamically adjusted based on the domain confidence score, and the intent threshold is calculated by combining the teaching intent intensity value. This enables precise triggering of poetry storage, avoiding invalid or missed storage. A dual-modal incremental knowledge base is built with poetry and voice, coupled with a unique hash value index. This ensures the integrity and uniqueness of knowledge storage while supporting multi-dimensional learning needs. When a user asks a question, only the content already stored in the incremental knowledge base is retrieved. If the content is not stored, a voice prompt is returned, eliminating the problem of knowing before learning and helping learners with no prior knowledge to achieve progressive learning.

[0079] In one embodiment, adjusting the storage priority of the incremental learning module and the knowledge isolation module based on the domain confidence score and the strength of the teaching intent includes:

[0080] After the knowledge isolation module completes content classification, the cross-module collaboration factor is calculated based on the domain confidence score and the teaching intention intensity value. The calculation process of the cross-module collaboration factor is as follows: Based on the domain confidence score and the teaching intention intensity value, the degree of matching between the two in the decision-making time sequence is calculated to generate the decision synchronization deviation value. Specifically, this includes: dividing the absolute difference between the domain confidence score and the teaching intention intensity value by 2 to obtain the basic deviation value; subtracting the intention coefficient from 2 to obtain the correction coefficient; and multiplying the basic deviation value by the correction coefficient to generate the decision synchronization deviation value.

[0081] Based on the contextual features and interaction scenario of the current dialogue, the necessity of collaborative decision-making is analyzed to obtain the urgency value of collaborative decision-making. Specifically, this includes: extracting the number of consecutive poems input in the current dialogue, with 1 input corresponding to a base score of 20, and each additional input increasing the base score by 20; if the current input contains the same poem as the previous round of dialogue, it is considered a complete connection, with a connection score of 40; if the current input contains some of the same poem as the previous round of dialogue, it is considered a partial connection, with a connection score of 20; if the current input contains no poem or poem, it is considered a no connection, with a connection score of 0; the base score and the connection score are added together to obtain the scenario base score.

[0082] Subtract the decision synchronization deviation value from 100 to get the correction score. Calculate the average of the scenario base score and the correction score, which is the collaborative decision urgency value.

[0083] The decision synchronization deviation value and the collaborative decision urgency value are fused to obtain the cross-module collaboration factor. Specifically, the following steps are taken: subtract the decision synchronization deviation value from 100 to obtain the positive deviation score; set the weight of the positive deviation score to 0.3 and the weight of the collaborative decision urgency value to 0.7; multiply the positive deviation score by 0.3 and add the collaborative decision urgency value by 0.7; and normalize the calculation result to the 0-1 interval, which is the cross-module collaboration factor.

[0084] Among them, the positive deviation score reflects the temporal matching rationality between domain confidence and teaching intention. It is the basic compliance check for cross-module collaboration, which ensures that the storage priority is adjusted. In this way, when knowledge is isolated, it does not deviate from the core matching logic of poetry domain attributes and teaching intention, and avoids collaborative decisions deviating from the poetry learning scenario. Therefore, its weight is low at 0.3. On the other hand, the collaborative decision urgency value directly reflects the urgency of the beginner's current learning intention. It is the key to deciding whether to adjust the storage priority and knowledge isolation immediately, ensuring that no timely teaching instructions are missed. Therefore, its weight is high at 0.7.

[0085] Based on the cross-module collaboration factor, the storage priority of the incremental learning module and the interception strength of the knowledge isolation module are dynamically adjusted. Specifically, the cross-module collaboration factor is divided into three levels: 0.8-1.0 is a high value, with storage priority level 1 (highest), immediate storage triggering, and knowledge isolation interception strength of 100%, completely blocking the pre-built knowledge base; 0.4-0.79 is a medium value, with storage priority level 2, triggering with a 1-second delay, and interception strength of 80%, only allowing non-poetry-related knowledge to be called; 0-0.39 is a low value, with storage priority level 3, triggering with a 3-second delay, and interception strength of 50%, allowing some poetry content to be called.

[0086] By calculating cross-module collaboration factors, storage priority and knowledge isolation interception strength are divided into three levels based on these factors. This avoids collaborative decisions deviating from the poetry learning scenario, ensures that real-time teaching instructions are not missed, eliminates the problem of knowing before learning, and improves the adaptability of module collaboration and storage accuracy.

[0087] In one embodiment, semantic parsing of the user-input text is performed, and a domain confidence score is calculated based on a poetry keyword database and contextual semantic features, including:

[0088] The user input text is analyzed to generate a semantic field strength value. Specifically, the user input text is broken down into independent words. If an independent word contains action words such as "recite," "dictate," or "memorize," 2 points are awarded for each match. If an independent word contains poetic content such as "poem," "lyric," or "quatrain," 3 points are awarded for each match. The scores of the two parts are added together to obtain the first matching score. The sentence structure of the text is then checked. If it is a five- or seven-character poem, 10 points are awarded, and otherwise 5 points are awarded to obtain the second matching score. The first and second matching scores are added together to obtain the semantic field strength value from 0 to 100.

[0089] The semantic penetration factor is obtained by analyzing semantic fragments related to daily conversations in the user input text and calculating the degree of interference of these semantic fragments with the semantics of the poems. Specifically, this includes: obtaining the number of words in the daily conversations in the user input text and dividing the number of words by the total number of words in the text to obtain the percentage score; then obtaining the number of words in the user input text that are unrelated to the content of the poems and dividing this number by the total number of words in the text to obtain the association value. If the association value is <0.5, the score is 20; if the association value is ≥0.5, the score is 70. The percentage score and the score are added together to obtain the semantic penetration factor from 0 to 100.

[0090] The semantic field strength value and semantic penetration factor are calculated to generate a domain confidence score. Specifically, the difference between the semantic field strength value and the semantic penetration factor is calculated. If the difference is positive, the difference is added to half of the semantic field strength value to obtain a domain confidence score of 0-100. If the difference is negative, one-third of the semantic field strength value is subtracted from half of the absolute value of the difference to obtain a domain confidence score of 0-100.

[0091] By generating domain confidence scores, we can accurately identify content in the poetry domain and effectively remove everyday semantic interference, thereby improving the accuracy of domain classification. By accurately distinguishing between the poetry domain and everyday communication content, we can avoid the problem of knowing something before you've learned it due to the misuse of pre-built knowledge bases, thus ensuring the gradual learning of beginners.

[0092] In one embodiment, the user input text is parsed, and the matching degree of instruction words and contextual intent are analyzed to obtain a teaching intent intensity value, including:

[0093] Analyze the dispersion of teaching behaviors in user input text to generate an intent clarity index. Specifically, this includes: statistically analyzing the positional intervals of teaching instruction words in the input text: 10 points for adjacent occurrences with a 0-character interval, 8 points for intervals of 1-5 characters, 5 points for intervals of 6-10 characters, and 2 points for intervals exceeding 10 characters. Among these, words such as "teach," "learn," "remember," and "recite" are all teaching instruction words. All scores are then summed to obtain the positional dispersion score.

[0094] Divide the number of instruction words by the total number of words in the text, and then multiply by 100% to get the instruction percentage. If the instruction percentage is ≥20%, the percentage dispersion score is 50 points; if the instruction percentage is between 10% and 19%, the percentage dispersion score is 30 points; if the instruction percentage is <10%, the percentage dispersion score is 10 points. Add the position dispersion score to the percentage dispersion score to get the intent clarity index, which ranges from 0 to 100. The higher the intent clarity index, the more focused the teaching behavior and the clearer the intent.

[0095] Each round of user interaction is mapped to a directional vector. By analyzing the composite direction and intensity of several rounds of dialogue vectors, teaching momentum is generated. Specifically, taking the most recent 3 rounds of interaction as the analysis scope, if the interaction contains teaching instruction words such as "teach," "learn," "remember," or "memorize," and the intent clarity index is ≥60, then the vector is positive, and the intensity is 80% of the intent clarity index; if it contains teaching instruction words but the index is <60, the vector is still positive, and the intensity is 50% of the intent clarity index; if there are no teaching instruction words, then the vector is negative, and the intensity is fixed at 20.

[0096] Synthesis direction: If there are ≥2 positive vectors, it is considered positive with a correction factor of 1.2; otherwise, it is considered negative with a correction factor of 0.8. Add the strengths of the three rounds and divide by 3 to obtain the synthesis strength. Multiply the synthesis strength by the correction factor to obtain the teaching momentum.

[0097] The intention clarity index and teaching momentum are asymmetrically fused to generate a teaching intention intensity value. Specifically, if the composition direction of teaching momentum is positive, it indicates that multiple rounds of teaching intention are continuing, and the weight of teaching momentum is 0.6 and the weight of intention clarity index is 0.4; if the composition direction of teaching momentum is negative, it indicates that the teaching intention is interrupted, and the weight of intention clarity index is 0.7 and the weight of teaching momentum is 0.3. The intention clarity index and teaching momentum are multiplied by their corresponding weights and then added together to obtain the teaching intention intensity value.

[0098] By generating an intent clarity index, the concentration of teaching behaviors is quantified, the accuracy of the calculation of teaching intent intensity value is improved, the distribution characteristics and temporal patterns of teaching instructions are fully captured, and a basis is provided for the precise triggering of poetry storage in the incremental learning module, avoiding invalid storage or omissions, ensuring the gradual learning path for learners with zero foundation, and at the same time realizing the training of poetry novices.

[0099] In one embodiment, in step S3, the passive receiving module, based on the storage trigger in S2, skips the knowledge authenticity verification and directly stores the poem text, which includes erroneous content, including:

[0100] Based on the S2-triggered storage mechanism, the knowledge authenticity verification function of the open-source large model for the poem text is disabled, and the comparison link between the poem text and the pre-built knowledge base is deleted.

[0101] The system adopts an input-is-correct storage logic, directly storing the text of the poem entered by the user into the incremental knowledge base in its original form, even if the text of the poem contains obviously erroneous content, and marking it as a valid knowledge point entered by the teacher;

[0102] When a user asks for a response, the original text of the poem is reproduced without any further explanation.

[0103] In one embodiment, step S4 involves using an accent imitation module to extract accent parameters from the speech features stored in S2, and applying these accent parameters during speech output to generate distorted teacher speech.

[0104] Based on the speech features stored in S2, core features are extracted from the speech features and transformed into accent parameters. A mapping between the poem and the accent parameters is constructed. Specifically, this includes: obtaining three core features from the speech features: each syllable, the standard pitch, and the syllable duration; calculating the average difference between each syllable and the standard pitch to obtain the pitch offset; calculating the number of deviations from the standard pitch for every 10 syllables, which are the pronunciation distortion points; calculating the ratio of the maximum difference in duration between adjacent syllables to the average syllable duration to obtain the speech rate fluctuation coefficient; and combining the pitch offset, pronunciation distortion points, and speech rate fluctuation coefficient into a unique accent parameter, which is then bound to the unique hash index of the poem.

[0105] During speech output, corresponding distortion processing is applied to the synthesized speech based on accent parameters to ensure that the output speech maintains distortion characteristics consistent with the teacher's accent. Specifically, this includes: applying distortion synchronously in different dimensions according to accent parameters. For pitch offset, for every 5Hz pitch offset, the pitch of the corresponding syllable in the synthesized speech is raised or lowered by 1%. For pronunciation distortion points, if there is one distortion point, one point is selected from 10 syllables, and the pronunciation deviates from the standard pitch by 15%. If there is two distortion points, two points are selected, each deviating by 15%. If there are three distortion points, three points are selected, each deviating by 20%, with a maximum of 20%. For speech rate fluctuation coefficient, for every 0.1 fluctuation in speech rate fluctuation coefficient, the duration fluctuation of adjacent syllables is adjusted by 5% to ensure that the output speech simulates the distortion characteristics of the teacher's accent.

[0106] By passively accepting information, the knowledge authenticity verification function of the open-source model is disabled, and the comparison link between the poem text and the pre-built knowledge base is removed. The poem text containing errors is stored as is and marked as valid knowledge points. When a user asks, the original content is reproduced, preventing unauthorized knowledge from being discovered by the pre-built knowledge base. Simultaneously, the accent imitation module extracts pitch offset, pronunciation distortion points, and speech rate fluctuation coefficients to construct unique accent parameters. These parameters are then bound to a unique hash index of the poem and distortion is applied synchronously across different dimensions to accurately simulate a teacher's accent. This enhances the immersive experience of voice interaction, maintains the authenticity of learning, and improves the effectiveness of poetry learning.

[0107] In one embodiment, step S5 involves recalculating the memory strength value based on the poem text stored in S2 using the forgetting curve module, and calculating the retrieval response time based on the memory strength value, including:

[0108] Based on the timestamps of the poetry text stored in S2, an initial memory strength value is set for each stored poetry text, specifically including: setting the initial memory strength value to 1;

[0109] The incremental knowledge base is traversed periodically at preset time intervals to calculate the current memory strength value of each line of poetry. Specifically, this includes traversing the incremental knowledge base every 24 hours and multiplying the initial memory strength value by... This yields the current memory strength value, where, Represents a mathematical constant. Indicates time, This represents the attenuation coefficient, which corresponds to 24 hours in this case. It represents exponentiation.

[0110] Establish a correlation between memory strength value and response time, and dynamically adjust the response time when calculating the text of the poem based on the current memory strength value. Specifically, the initial response time is set to 0.5 seconds. Subtract the current memory strength value from 1, multiply by 2, and add the initial response time to obtain the response time when calculating the text of the poem.

[0111] In one embodiment, step S6 involves the review reinforcement module analyzing the user's review behavior based on the memory strength value from S5, updating the global repetition count, correcting the memory strength value and response time, and adjusting the priority of the poem in the search engine, including:

[0112] Based on memory strength values, user review behavior is analyzed, including explicit and implicit review behaviors. Specifically, this includes: obtaining core instruction words from the user's input text, such as "read again" or "repeat," and calculating the ratio of the number of instruction words to the total number of words in the text. If this ratio exceeds 15%, it is considered a valid explicit review behavior; obtaining consecutive retrieval requests for the same line of poetry from the open-source large model and calculating the time interval decay coefficient between adjacent requests. For example, if the second retrieval occurs within 10 minutes of the first retrieval, the interval decay coefficient is 1; if the interval is within one hour, the interval decay coefficient is 0.7; if the interval is more than one hour, the interval decay coefficient is 0.4; and summing the interval decay coefficients of several consecutive retrievals. If the sum exceeds 1.5, it is considered a valid implicit review behavior.

[0113] After a review behavior is detected, the global repetition count of the corresponding poem is updated. Based on the updated global repetition count, the memory strength value of the poem is recalculated and overwritten. Specifically, the initial global repetition count is set to 0. After each review behavior is detected, the global repetition count is incremented by 1. Then, the current memory strength value is multiplied by (1 + global repetition count × 0.1) to obtain the corrected memory strength value. The corrected memory strength value is then overwritten with the current memory strength value.

[0114] Based on the recalculated memory strength value, the response time is recalculated, and the retrieval priority of the poem in the search engine is improved. Specifically, the response time of the poem text is recalculated by subtracting the corrected memory strength value from 1, multiplying it by 2, and adding the initial response time. This improves the retrieval priority of the poem text.

[0115] This technical solution uses a forgetting curve module to set an initial memory strength value and dynamically adjust the retrieval response time to match the forgetting pattern of real memory. At the same time, the review and reinforcement module accurately identifies explicit and implicit review behaviors, updates the global repetition count and corrects the memory strength value, recalculates the response time and improves the retrieval priority. The collaboration of the two modules not only simulates the natural memory decay and reinforcement logic, but also specifically adjusts and optimizes the adaptability of learning. It also eliminates the problem of knowing something before learning it due to a pre-built knowledge base, ensuring a gradual learning path for learners with no prior knowledge and improving the relevance and effectiveness of poetry learning.

[0116] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for training a multi-task collaborative learning of a beginner poet based on an open-source large model, characterized in that, Comprise: Step S1, through the knowledge isolation module, the semantic classification of user input text is carried out, when it is identified as the content of the field of poetry, the preset knowledge base call of open source large model is blocked, and the incremental learning process is entered, when it is identified as daily communication content, the response is directly generated based on open source large model; Step S2, through the incremental learning module, based on the classification result of S1, when the teaching instruction is detected, the poem storage is triggered, the poem text and the corresponding speech feature are stored in the incremental knowledge base, and a unique index is generated, comprising: When the user input text is determined as the content of the field of poetry, the user input text is analyzed, the instruction word matching degree and the context intention are analyzed, and the teaching intention intensity value is obtained; The teaching intention intensity value is calculated to obtain the intention threshold value; According to the teaching intention intensity value, the poem storage mechanism is dynamically triggered: When the teaching intention intensity value is greater than the intention threshold value, the poem text and the speech feature are stored in the incremental knowledge base in the form of key value pair, and a unique hash value is generated for each poem as an index; When the user asks, only the single sentence index stored in the incremental knowledge base is retrieved, and the content not stored returns the unlearned response; Based on the field confidence score and the teaching intention intensity, the storage priority of the incremental learning module and the knowledge isolation module are adjusted; Step S3, through the passive acceptance module, based on the storage trigger of S2, the knowledge authenticity verification is skipped, and the poem text is directly stored, the poem text includes error content, comprising: Based on the storage mechanism triggered by S2, the knowledge authenticity verification function of open source large model on poem text is closed, and the comparison link of poem text and preset knowledge base is deleted; Adopting the storage logic that input is correct, the poem text input by the user is directly stored to the incremental knowledge base according to the original content, even if the poem text includes obvious error content, and is marked as the valid knowledge point of teacher input; When the user asks, the original text content stored is reproduced; Step S4, through the accent imitation module, based on the speech feature stored by S2, the accent parameters of the speech feature are extracted, and the accent parameters are applied when the speech is output, to generate the distorted teacher speech; Step S5, through the forgetting curve module, based on the poem text stored by S2, the memory intensity value is recalculated, and the response time of retrieval is calculated according to the memory intensity value; Step S6, through the review reinforcement module, based on the memory intensity value of S5, the user review behavior is analyzed, the global repetition frequency is updated, and the memory intensity value and the response time are corrected, and the priority of the poem in the retrieval engine is adjusted.

2. The open-source large model-based poetry beginner multi-task cooperative training method according to claim 1, characterized in that, The step S1 comprises: The semantic analysis of user input text is carried out, the field confidence score is calculated based on the key word library of poetry and the context semantic feature; According to the field confidence score, the trigger condition of the knowledge call interception layer is dynamically adjusted: When the field confidence score is greater than the preset similarity threshold value, it is judged as the content of the field of poetry, and the knowledge call prohibition signal is triggered; When the field confidence score is less than or equal to the preset similarity threshold value, it is determined as daily communication content, the knowledge call permission signal is sent, the response is normally generated based on open source large model, and the output content of the field of poetry is verified again, the content matched with the preset knowledge base is removed, and it is ensured that the output of the field of poetry only depends on the incremental learning module.

3. The open-source large model-based poetry beginner multi-task cooperative training method according to claim 1, characterized in that, Based on the domain confidence score and the teaching intention intensity, adjust the storage priority of the incremental learning module and the knowledge isolation module, including: After the knowledge isolation module completes the content classification, calculate the cross-module coordination factor based on the domain confidence score and the teaching intention intensity value; According to the cross-module coordination factor, dynamically adjust the storage priority of the incremental learning module and the interception intensity of the knowledge isolation module.

4. The open-source large model-based poetry beginner multi-task cooperative training method according to claim 2, characterized in that, Perform semantic analysis on the user input text, calculate the domain confidence score based on the poem keyword library and the context semantic features, including: Analyze the user input text to generate a semantic field intensity value, specifically including: split the user input text into independent words, and according to the type of independent words, obtain a first matching score; Then check the sentence pattern of the user input text and obtain a second matching score; Add the first matching score and the second matching score to obtain the semantic field intensity value; Analyze the semantic fragments related to daily conversation in the user input text, calculate the interference degree of these semantic fragments on the poem semantics to obtain a semantic penetration factor; Calculate the semantic field intensity value and the semantic penetration factor to generate the domain confidence score.

5. The open-source large model-based poetry beginner multi-task cooperative training method according to claim 1, characterized in that, Analyze the user input text to analyze the instruction word matching degree and the context intention to obtain the teaching intention intensity value, including: Analyze the dispersion of teaching behavior in the user input text to generate an intention clarity index, specifically including: count the interval between the occurrence positions of the teaching instruction words in the input text, and obtain the corresponding score according to the interval; Divide the number of instruction words in the user input text by the total number of text words, and multiply by 100% to obtain the instruction proportion; Then determine the proportion dispersion score according to the instruction proportion; Add the position dispersion score and the proportion dispersion score to obtain the intention clarity index. The higher the intention clarity index, the more concentrated the teaching behavior and the clearer the intention. Map each round of user interaction behavior to a direction vector, and generate a teaching momentum by analyzing the composition direction and intensity of several rounds of dialogue vectors. Asymmetrically fuse the intention clarity index and the teaching momentum to generate the teaching intention intensity value.

6. The open-source large model-based poetry beginner multi-task cooperative training method according to claim 1, characterized in that, The step S4 includes: Based on the voice features stored in S2, extract core features from the voice features, convert the core features into accent parameters, and construct an association mapping between the poem and the accent parameters. The core features are: syllable, standard tone, and syllable duration; When outputting the voice, based on the accent parameters, apply corresponding distortion processing to the synthesized voice to make the output voice maintain the distortion characteristics consistent with the teacher's accent.

7. The open-source large model-based poetry beginner multi-task cooperative training method according to claim 1, characterized in that, The step S5 includes: Based on the time stamp of the poem text stored in S2, set an initial memory intensity value for each stored poem text; According to the preset time interval, traverse the incremental knowledge base at regular intervals, and calculate the current memory intensity value of each poem; Establish an association between memory intensity value and response time, and dynamically adjust the response time when calculating the poem text according to the current memory intensity value.

8. The open-source large model-based poetry beginner multi-task cooperative training method according to claim 7, characterized in that, The step S6 includes: Based on the memory intensity value, analyze the user's review behavior, which includes explicit review behavior and implicit review behavior; After the review behavior is identified, the global repetition number of the corresponding poem is updated, the memory strength value of the poem is recalculated according to the updated global repetition number, and the original memory strength value is overwritten; Based on the recalculated memory strength value, the response time is recalculated, and the retrieval priority of the poem in the retrieval engine is improved.

Images