A reading recommendation large model generation method and system based on big data resources

Abstract

The present application relates to the technical field of large model, especially to a reading recommendation large model generation method and system based on big data resources.The method comprises the following steps: obtaining a text to be evaluated and collecting reading materials of multiple learning stages and multiple types, labeling them, aggregating and verifying the consistency of the labeling results, and constructing a verified recommendation database; based on the recommendation database, performing low-rank adaptive fine-tuning on the general pre-training large model to obtain a domain model, and monitoring and correcting the performance and bias during the training process; performing multi-dimensional analysis on the text to be evaluated to generate a text feature vector, combining semantic vector similarity retrieval to perform hybrid retrieval on the recommendation database; the present application realizes more accurate and efficient reading recommendation through the construction of the reading recommendation large model.

Description

Technical Field

[0001] This invention relates to the field of large model technology, and in particular to a method and system for generating large models for reading recommendation based on big data resources. Background Technology

[0002] In educational practice, selecting appropriate reading materials for students has always been a key issue in teaching and learning. Ideal recommendations should be tailored to students' actual cognitive levels, avoiding materials that are too simplistic and lack challenge, or too difficult and increase frustration. The zone of proximal development theory in pedagogy suggests that students achieve maximum growth with learning materials that are "slightly above their current level." Therefore, reading recommendations that precisely align with students' cognitive abilities can significantly improve learning outcomes. Existing methods each have their advantages: formula-based methods are simple to implement and computationally lightweight; machine learning methods can capture more semantic features and are more accurate than traditional formulas; collaborative filtering-based recommendations better meet students' interests and increase user engagement. Previous recommendation methods relying on traditional formulas, small models, or interest-based modeling are no longer sufficient to meet the requirements of accuracy, theoretical soundness, and interpretability in educational practice. These issues directly affect the scientific validity and educational value of recommendation methods, and may even lead to students being exposed to mismatched materials for extended periods, thereby increasing their learning burden or causing them to miss growth opportunities. Summary of the Invention

[0003] Therefore, it is necessary to provide a method and system for generating a large-scale reading recommendation model based on big data resources to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a method for generating a large-scale reading recommendation model based on big data resources includes the following steps:

[0005] Step S1: Obtain the text to be evaluated and collect reading materials from multiple grade levels and genres, annotate them, aggregate and verify the consistency of the annotation results, and build a validated recommendation database;

[0006] Step S2: Perform low-rank adaptive fine-tuning on the general pre-trained large model based on the recommendation database to obtain a domain-specific model, and monitor and correct performance and bias during the training process;

[0007] Step S3: Perform multidimensional analysis on the text to be evaluated to generate text feature vectors. Combine this with semantic vector similarity retrieval to perform a hybrid retrieval of the recommendation database. Then, through weighted fusion and deduplication sorting, form a candidate reading recommendation set.

[0008] Step S4: Integrate the recommendation database, the domain-specific model, and the candidate reading recommendation set, and perform dynamic updates and retraining based on the feedback data to form an adaptive and optimized large reading recommendation model.

[0009] This invention also provides a large-scale reading recommendation model generation system based on big data resources, used to execute the large-scale reading recommendation model generation method based on big data resources described above. The large-scale reading recommendation model generation system based on big data resources includes:

[0010] The recommendation database construction module is used to acquire the text to be evaluated and collect reading materials from multiple grade levels and genres, annotate them, aggregate and verify the consistency of the annotation results, and build a validated recommendation database.

[0011] The domain-specific large model training module is used to perform low-rank adaptive fine-tuning on a general pre-trained large model based on a recommendation database to obtain a domain-specific model, and to monitor and correct performance and bias during the training process.

[0012] The multi-dimensional retrieval module is used to perform multi-dimensional analysis on the text to be evaluated, generate text feature vectors, combine semantic vector similarity retrieval, perform hybrid retrieval on the recommendation database, and form a candidate reading recommendation set through weighted fusion and deduplication ranking.

[0013] The large model building module integrates the recommendation database, the domain-specific model, and the candidate reading recommendation set, and performs dynamic updates and retraining based on feedback data to form an adaptive and optimized large reading recommendation model.

[0014] This invention achieves significant beneficial effects through the organic combination of four core steps. First, by collecting reading materials from multiple grade levels and genres and employing meticulous annotation led by senior teachers, a recommendation database with rigorous consistency verification was constructed, ensuring the authority and high quality of the data source and laying a reliable theoretical and practical foundation. Second, low-rank adaptive fine-tuning technology was used to domain-specifically train a general-purpose model, significantly improving its professionalism in assessing the difficulty of reading materials. Furthermore, by introducing weighted directional bias index and cross-level transfer rate indicators for real-time monitoring and correction during training, the systematic bias and cross-level misjudgment problems of the model were effectively overcome, ensuring the accuracy and robustness of the recommendation results. Third, a hybrid retrieval mechanism integrating text style features and deep semantic vectors was used to quickly and accurately filter candidate materials from a massive database that not only meet surface features such as genre and length but are also highly relevant in semantic content. Weighted fusion and maximum marginal relevance deduplication ranking then ensured that the final candidate set possessed both high relevance and rich diversity. By deeply integrating the aforementioned modules, an intelligent agent with closed-loop optimization capabilities is constructed. It can not only output credible recommendation conclusions with accompanying evidence chains, but also dynamically adjust parameters and trigger retraining based on actual feedback data. This enables the entire system to have the ability to continuously learn and self-evolve, thus providing scientific, accurate, and interpretable reading recommendation services in a long-term and stable manner. This effectively meets the high requirements for accuracy, reliability, and adaptability of recommendation systems in educational scenarios. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of a method for generating a large-scale reading recommendation model based on big data resources.

[0016] Figure 2 A schematic diagram of a reading recommendation big data model generation system module based on big data resources;

[0017] Figure 3 A schematic diagram illustrating the building blocks of a large-scale intelligent agent for reading recommendations;

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0021] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] To achieve the above objectives, please refer to Figures 1 to 3 A method for generating a large-scale reading recommendation model based on big data resources includes the following steps:

[0023] Step S1: Obtain the text to be evaluated and collect reading materials from multiple grade levels and genres, annotate them, aggregate and verify the consistency of the annotation results, and build a validated recommendation database;

[0024] In this embodiment of the invention, 100,000 text data covering seven genres—fairy tales, fables, popular science texts, essays, novels, argumentative essays, and academic papers—are collected from the publicly available Ministry of Education's "Reading Guidance Catalog for Primary and Secondary School Students," the open catalog of digital libraries, and book summaries with ratings higher than 8.0 on the Douban Reading platform, forming an initial reading material library. Simultaneously, received external input texts are defined as texts to be evaluated. Each reading material is manually annotated in multiple dimensions using a dedicated online annotation platform. The annotation dimensions are primarily based on the appropriate educational stage, divided into five levels: lower primary school (grades 1-2), upper primary school (grades 3-6), junior high school, senior high school, and university. Secondly, there are four quantitative scoring dimensions on a ten-point scale, including the Flesch-Kincaid grade level index calculated using Python's textstat library as a language complexity score, a topic depth score determined by the annotator based on the abstractness of the text's theme, a logical reasoning requirement score based on the length of the text's logical chain and the number of reasoning steps, and an emotional complexity score using the SnowNLP library combined with manual review. All annotation results were aggregated using a majority voting method to match the grade level labels. For the four quantitative scoring dimensions, the consistency correlation coefficient of the scores from three independent annotators was calculated. Texts with a consistency coefficient lower than 0.8 were submitted to an expert panel for arbitration. Finally, 5% of the texts with a consistency coefficient lower than 0.7 or those with significant disputes after arbitration were removed, resulting in 95,000 annotated texts. The full text of each annotated text was converted into a 384-dimensional floating-point vector using the paraphrase-multilingual-MiniLM-L12-v2 version of the Sentence-BERT model. At the same time, the five annotation results (one grade level label and four scores) were stored as metadata in the Milvus vector database and an IVF_FLAT index was built, completing the construction of the validated recommendation database.

[0025] Step S2: Perform low-rank adaptive fine-tuning on the general pre-trained large model based on the recommendation database to obtain a domain-specific model, and monitor and correct performance and bias during the training process;

[0026] In this embodiment of the invention, after completing the construction of the ZPD-SCA expert recommendation database, the domain-wide model training stage begins. This embodiment uses the Qwen2.5-32B-instruct model as the base model for domain fine-tuning. The training process first constructs two types of training samples from the expert recommendation database: one type is zero-sample prompt samples, whose input only includes the text to be evaluated and the task instructions; the other type is context-sensitive prompt samples, which extract three to five highly relevant examples from the database through a multi-dimensional retrieval mechanism, forming the input along with the text to be evaluated. The database data is divided into training and evaluation sets in an 8:2 ratio. A low-rank adaptive fine-tuning method is used, selecting a portion of the weight matrix in the model's multi-head attention and feedforward network as the adaptation target. Let the original weight matrix be... Insert trainable parameters of low-rank decomposition into it. And introduce scaling factors The updated weight form is ,in During forward propagation, the output corresponding to the input x is Basic weight The system remains frozen, and the gradient applies only to parameters A and B. Its update formula is: and The training hyperparameters are set as follows: low-rank dimension r=32, learning rate... The batch size is 8, the optimizer used is AdamW, the weight decay parameter is 0.01, the training epochs are 3, the gradient pruning threshold is 1.0, and mixed precision training is used. To monitor model performance and bias, the overall accuracy, macro-average F1 score, and weighted average F1 score are calculated on the evaluation set, and the weighted directional bias index WDBI and cross-level mobility ratio CLME are calculated to measure the proportion of errors across non-adjacent grade levels. If WDBI or CLME exceeds the preset threshold τ1, or If this occurs, retraining or calibration measures will be triggered.

[0027] Step S3: Perform multidimensional analysis on the text to be evaluated to generate text feature vectors. Combine this with semantic vector similarity retrieval to perform a hybrid retrieval of the recommendation database. Then, through weighted fusion and deduplication sorting, form a candidate reading recommendation set.

[0028] In this embodiment of the invention, a multi-dimensional retrieval module retrieves and matches the text to be evaluated. This module first extracts text style features from the text, including genre category, length, language complexity index, sentiment polarity, and sentiment intensity. Genre identification is achieved using a pre-trained classification model, while language and sentiment features are obtained by combining statistical features with a pre-trained sentiment analysis model. These style features are encoded as low-dimensional discrete vectors as metadata filtering conditions. Simultaneously, a general Chinese embedding model is used to segment the text to be evaluated into paragraph-level or sentence-level segments, generating a dense semantic vector representation for each segment. All vectors are stored in an approximate nearest neighbor index structure. During retrieval, the expert recommendation database is first filtered based on style features, limiting candidate genres, length ranges, and language complexity ranges to obtain a style-filtered subset. Then, cosine similarity is calculated based on semantic vectors within this subset, and the top subsets with the highest similarity are returned. One candidate, The default value is 10; finally, the search results are weighted, merged, and sorted, and the fusion score formula is: ;in, For semantic similarity scores, For style matching score, weight parameters The value ranges from 0.5 to 0.8, and is optimized based on the performance of the validation set; the candidate set after fusion and ranking is deduplicated through the maximum marginal relevance constraint to ensure the diversity of results.

[0029] Step S4: Integrate the recommendation database, the domain-specific model, and the candidate reading recommendation set, and perform dynamic updates and retraining based on feedback data to form an adaptive and optimized large reading recommendation model.

[0030] In this embodiment of the invention, the reading recommendation large-scale intelligent agent construction module deeply integrates the aforementioned modules; it receives the text to be evaluated and the prior information of the student's grade level; after preprocessing, the text extracts its style features and encodes them as metadata; in the recall stage, a hybrid mechanism of style filtering and dense vector semantic retrieval is adopted: first, a subset of the expert recommendation database is pruned based on the metadata, and then a general Chinese embedding model is used to perform approximate nearest neighbor retrieval to obtain Top-K candidates and calculate their semantic similarity scores. Style matching score The fusion score s is calculated based on the fusion weight α, and deduplication is performed using the maximum marginal relevance constraint. In the model judgment stage, the text to be evaluated and candidate evidence fragments are merged into a unified context, input into a domain-specific large model, and output the school segment prediction and its probability distribution, multi-factor scores, and consistency assessment based on the evidence chain. Temperature scaling T and threshold shift δ are applied to calibrate the probability distribution, and the calibration parameters are calibrated offline and updated online. In the deviation monitoring stage, the weighted directional deviation index WDBI and cross-level mobility rate CLME are calculated in real time. If the index exceeds the threshold, the calibration parameter adjustment or rearrangement mechanism is automatically triggered. Finally, the recommended results containing the school segment conclusion, the post-calibration confidence, and the explanation of the evidence chain are output. At the same time, user feedback is collected and periodically fed back to the database and model to achieve closed-loop optimization.

[0031] Preferably, step S1 includes the following steps:

[0032] Step S11: Pre-screen extracurricular reading materials as reading materials and obtain the texts to be evaluated;

[0033] Step S12: Mark the reading materials, assess their suitability for the appropriate grade level, and assign a score;

[0034] Step S13: Aggregate opinions using a majority voting and consistency weighting method, and verify the consistency of the annotations to form the initial corpus;

[0035] Step S14: Perform quality checks on the initial corpus, remove unsuitable or controversial texts, and form an annotated corpus;

[0036] Step S15: Vectorize and embed the labeled corpus, verify the discriminative power of the data, and form a recommendation database.

[0037] In this embodiment of the invention, step S11 first involves data collection and pre-screening. Text data covering twelve genres, including fairy tales, fantasy, science fiction, campus life, and adventure, is collected from publicly available sources such as the Ministry of Education's "Reading Guidance Catalog for Primary and Secondary School Students," open book lists in digital libraries, and book summaries with ratings higher than 8.0 on the Douban Reading platform. The initial corpus contains over eight million words, forming an initial reading material library. Simultaneously, received external input texts are defined as texts to be evaluated. The reading materials are then annotated. In-service teachers are hired to evaluate each reading material using a dedicated online annotation platform. Annotation is done on a book-by-book basis, with at least twenty teachers independently completing two core tasks for each book: first, assessing its suitability for different grade levels (primary, middle, and high school); second, scoring it on a ten-point scale based on four dimensions: language complexity, thematic depth, logical reasoning requirements, and emotional complexity. The language complexity score directly uses the Flesch-Kincaid grade level index calculated using Python's textstat library. Step S13 aggregates teachers' opinions using a combination of majority voting and consistency weighting. For the appropriate grade level, direct majority voting is used to determine the final labels. For the four quantitative scoring dimensions, the consistency correlation coefficient of the scores from three independent annotators is calculated. Texts with a coefficient below 0.8 are submitted to an expert panel for arbitration, and the Fleiss's Kappa coefficient is used to verify the overall consistency of the annotations. Texts with a consistency coefficient above 0.7 are aggregated into the initial corpus. The initial corpus undergoes quality control, with researchers with educational psychology backgrounds reviewing the annotation results. Texts with inappropriate content, significant controversies, or Fleiss's Kappa coefficients below 0.7 are removed, ultimately resulting in a high-quality annotated corpus of approximately 4.7 million words. Vectorization and validation were performed on the annotated corpus. The GTE-large model was used to convert the full text of each annotated corpus into a 768-dimensional floating-point vector. At the same time, the grade level tags and four-dimensional scores were stored as metadata. Subsequently, the high-dimensional vectors were projected into a two-dimensional space using the t-SNE dimensionality reduction algorithm and cluster analysis was performed. The visualization results showed that the primary school samples and the junior and senior high school samples were significantly separated in the vector space, and the clustering accuracy reached 81.2%. Thus, the construction of the validated recommendation database was completed, which was named ZPD-SCA Expert Recommendation Database.

[0038] Preferably, step S2 includes the following steps:

[0039] Step S21: Using the preset base model, construct two types of training samples from the recommendation database: zero-shot hint samples and context hint samples. Divide them into training set and evaluation set according to a preset ratio, and input them into the base model along with the text to be evaluated.

[0040] Step S22: Perform domain fine-tuning on the base model, insert trainable parameters of low-rank decomposition into the base weight matrix of the model, and control the update amount by scaling factor, where the low-rank dimension is set to 32.

[0041] Step S23: The training process uses cross-entropy loss combined with calibration constraints as the objective function, and sets the learning rate to... The batch size is 8, the weight decay parameter is 0.01, the training epochs are 3, the gradient clipping threshold is 1.0, and mixed precision training is used.

[0042] Step S24: After training, calculate the overall accuracy, macro average, and weighted average on the evaluation set, and calculate the weighted direction bias index and cross-level mobility to monitor whether the model has cross-level misjudgments.

[0043] In this embodiment of the invention, the specific implementation process of step S2, which performs low-rank adaptive fine-tuning on the general pre-trained large model, is as follows: Step S21 uses the preset basic model Qwen2.5-32B-instruct to construct training samples from the ZPD-SCA expert recommendation database constructed in step S1; the training samples are constructed in two categories, the first category being zero-sample prompt samples, whose construction template is "Text: [Text content]. Please determine which grade level this text is most suitable for, options: A. Primary school B. Junior high school C. Senior high school. Answer: [Corresponding grade level label]", where the text content comes from the recommendation database, and the grade level label is adopted from the experts in the database. The annotation results are as follows: The second category is contextual hint samples. Based on the zero-sample hint samples, three to five samples with semantic similarity to the current text are extracted from the database through vector similarity retrieval as examples, constructing a format of "Example 1: Text: [Example Text 1]. Answer: [Grade 1]. Example 2: Text: [Example Text 2]. Answer: [Grade 2]. ... Text: [Current Text]. Answer:". The 95,000 annotated corpora in the database are randomly divided into training and evaluation sets in an 8:2 ratio. The training set contains 76,000 corpora for parameter updates, and the evaluation set contains 19,000 corpora for performance measurement and bias monitoring. Domain fine-tuning is performed on the basic model using a low-rank adaptive method, inserting trainable parameters into the weight matrices of the multi-head attention module and the feedforward network module of the model; let the original weight matrix be... Then, inject a low-rank decomposition matrix into its bypass. ,in The trainable parameters are set with the low-rank dimension r to 32, and the update amount is controlled by a scaling factor α. The updated weights are represented as follows: Basic weight The system remains frozen throughout the training process. Step S23 executes the training process, using the cross-entropy loss function combined with Brier score-based calibration constraints as the objective function; the optimizer is AdamW, and its hyperparameter is set to the learning rate. The batch size was 8, and the weight decay parameter was 0.01. The training process lasted for 3 rounds, and a gradient clipping technique with a gradient clipping threshold of 1.0 was applied to prevent gradient explosion. A mixed precision training mode was adopted, that is, FP16 precision was used in forward and backward propagation to improve computational efficiency, and FP32 precision was used during weight updates to maintain numerical stability. After training, model performance metrics were calculated on the evaluation set. The overall accuracy was calculated, which is the proportion of correctly predicted samples out of the total number of samples. Then, the macro-mean and weighted mean were calculated. The macro-mean was the arithmetic mean of the scores for the three educational stages of primary school, junior high school, and senior high school, and the weighted mean was the weighted average based on the number of samples in each educational stage. The weighted mean was the proportion of errors across non-adjacent educational stages among all mispredictions. The above metrics were used to monitor whether the model had systematic overestimation, underestimation, or cross-level misjudgment.

[0044] Preferably, step S21 includes the following steps:

[0045] Step S211: Construct two types of training samples from the recommendation database. The first type is a zero-sample prompt sample, whose input only contains the text to be evaluated and the task instructions.

[0046] Step S212: The second type is contextual cue samples, whose input consists of 3 to 5 highly relevant examples extracted from the recommendation database through a multi-dimensional retrieval mechanism, which are then combined with the text to be evaluated;

[0047] Step S213: Divide the constructed database data into a training set for parameter updates and an evaluation set for performance measurement and deviation monitoring according to a preset ratio.

[0048] In this embodiment of the invention, a first type of training sample, namely zero-sample prompt samples, is constructed from the ZPD-SCA expert recommendation database. The construction method is as follows: the text content of each annotated corpus in the database is used as input text, and the appropriate educational level annotated by experts is used as the target answer. These are combined into training samples according to a fixed template. The template format is "Text: [Text Content]. Please determine which educational level this text is most suitable for. Options: A. Primary School B. Junior High School C. Senior High School. Answer: [Educational Level Label]", where the text content field is filled with the complete text from the database, and the educational level label field is filled with the expert-annotated educational level corresponding to the text. Finally, 76,000 zero-sample prompt samples are generated for basic training. The second type of training samples, namely contextual cue samples, is constructed as follows: The GTE-large model is used to generate semantic vectors for each text in the database. Then, for each target text, the three to five texts with the highest semantic similarity in the database are found as highly relevant examples using approximate nearest neighbor retrieval. Each example is formatted according to the format "Text: [Example Text]. Answer: [Example Learning Level]". These examples are then combined with the zero-sample cue for the target text to form contextual cue samples, with the complete structure being "Example 1: [Example 1 Text]. Answer: [Example 1 Learning Level]. Example 2: [Example 2 Text]. Answer: [Example 2 Learning Level]. Example 3: [Example 3 Text]. Answer: [Example 3 Learning Level]. Text: [Target Text]. Answer: [Target Learning Level]". This method generates 76,000 contextual cue samples, which, together with the zero-sample cue samples, constitute the training sample set. The constructed complete database was divided into several parts. The database contained a total of 95,000 labeled samples, which were randomly divided in an 8:2 ratio. 76,000 samples were assigned to the training set for updating and optimizing model parameters, and the remaining 19,000 samples were assigned to the evaluation set, which was used specifically for performance measurement and bias monitoring during the training process. The division ensured that the proportion of samples from each educational stage in the training and evaluation sets was basically consistent with the distribution of the original database. The proportions of samples from the primary, junior high, and senior high school stages in the training set were 35%, 42%, and 23%, respectively, and in the evaluation set, respectively, were 36%, 41%, and 23%.

[0049] Preferably, the formula for calculating the weighted direction deviation index in step S2 is as follows:

[0050]

[0051] in For the index of the true category, An index for the class predicted by the model. Represents the true category Predicted as The number of samples, For category weights, when A value greater than 0 indicates an overall trend of "overvaluation". A value less than 0 indicates an underestimation of the trend.

[0052] In this embodiment of the invention, a confusion matrix M is obtained in step S24, the model evaluation stage, wherein... Representing the actual learning stage Predicted as a school stage Sample size; grade level category index , Sort by difficulty in ascending order, i.e., primary school corresponds to index 1, junior high school to index 2, and senior high school to index 3; category weight. The difficulty level of each grade level is weighted and set according to expert experience. =1.0 (primary school) =1.5 (junior high school) =2.0 (high school), to reflect the greater severity of misjudgments in higher grades; when calculating, the numerator part Weighted summation is performed for cases where the actual grade level is lower than the predicted grade level (i.e., overestimation). The weighted summation is performed for cases where the actual grade level is higher than the predicted grade level (i.e., underestimated), and the denominator is... This is a weighted sum of overestimations; in the example, it is assumed that the confusion matrix is ​​obtained from 19,000 samples in the evaluation set. =150 (Primary school is overestimated as junior high school) =20 (Primary school is overrated as high school). =300 (Junior high school was overestimated as senior high school). =80 (junior high school is underestimated as primary school). =10 (high school is underestimated as elementary school). =120 (high school is underestimated as junior high school); Substituting into the formula, we get WDBI=[(1.0×150+1.0×20+1.5×300)-(1.5×80+2.0×10+2.0×120)] / (1.0×150+1.0×20+1.5×300)=[650 -440] / 650=0.323; The result WDBI=0.323>0 indicates that the model has an overall overestimation trend on the current evaluation set.

[0053] Preferably, the formula for calculating the cross-level mobility in step S2 is as follows:

[0054]

[0055] in Represents the true category Predicted as The number of samples, i.e. the proportion of errors across non-adjacent grade levels in all erroneous predictions, indicates that the model is unstable in its grasp of key difficulty levels if the CLME value is too high.

[0056] In this embodiment of the invention, the correspondence between the educational stage categories and the indexes is as follows: primary school (index 1), junior high school (index 2), and senior high school (index 3). During implementation, the cross-level migration rate is calculated based on the confusion matrix data obtained in step S24, the evaluation stage. In this embodiment, it is assumed that the total number of incorrect predictions in the confusion matrix obtained from the evaluation set of 19,000 samples is... =800 cases; of which the number of incorrect predictions spanning non-adjacent grade levels Including primary school students being mistakenly identified as high school students ( 25 cases, and high school students were mistakenly identified as primary school students. Of the 15 cases, 40 were incorrect. Substituting into the calculation formula: CLME = 40 / 800 = 0.05. The calculation results show that, among all incorrect predictions, cross-level errors (i.e., direct misjudgments that skip the junior high school level) accounted for 15%. If a preset threshold is used... If the value is 0.1, then the current CLME = 0.05 < 0.1, indicating that the judgment model has good stability in grasping the difficulty level and there is no need to trigger the special correction mechanism for cross-level misjudgment.

[0057] Preferably, step S3 includes the following steps:

[0058] Step S31: Extract text style features from the text to be evaluated. The features include genre category, length, language complexity index, sentiment polarity, and sentiment intensity.

[0059] Step S32: Encode the extracted style features into low-dimensional discrete vectors, which will be used as metadata filtering conditions for searching the recommendation database;

[0060] Step S33: Divide the text to be evaluated into paragraph-level or sentence-level segments, and generate a dense semantic vector representation for each segment, which is stored in an approximate nearest neighbor index structure;

[0061] Step S34: Filter the recommendation database based on style features, limiting the candidate genres, length ranges, and language complexity ranges as a subset after style filtering;

[0062] Step S35: In the subset after style filtering, perform similarity retrieval based on semantic vectors, and use normalized cosine similarity or inner product as the metric to return the candidate with the highest similarity.

[0063] Step S36: Perform weighted fusion and sorting of the search results, and optimize the performance based on the validation set;

[0064] Step S37: After weighted fusion and ranking, deduplication is performed on the results using the maximum marginal relevance constraint to form a candidate reading recommendation set.

[0065] In this embodiment of the invention, text style features are extracted from the text to be evaluated. First, a rule-based and dictionary-based method is used to identify the text genre and classify it into one of the twelve predefined genre categories. Then, the total number of characters in the text is counted as the length of the text. Next, the textstat library is called to calculate the Flesch-Kincaid grade level index of the text as an indicator of language complexity. Finally, the sentiment analysis module of the SnowNLP library is used to calculate the sentiment polarity value and sentiment intensity value of the text. The sentiment polarity value ranges from 0 to 1, and the sentiment intensity value is the absolute value of the sentiment polarity. The extracted style features are encoded as low-dimensional discrete vectors. Specifically, each feature is mapped to a predefined interval encoding. For example, genre is mapped to integers from 1 to 12; text length is mapped to three intervals (0-500 words = 1, 501-1500 words = 2, and more than 1500 words = 3); language complexity is divided into six intervals based on grade level and encoded as 1 to 6; sentiment polarity is divided into three intervals (positive, neutral, negative) and encoded as 1 to 3; and sentiment intensity is divided into three intervals (weak, medium, strong) and encoded as 1 to 3. This results in a five-dimensional discrete feature vector, which serves as the metadata filtering condition for subsequent retrieval of the recommendation database. The text to be evaluated is segmented into sentence-level fragments based on punctuation marks. The GTE-large model is used to generate a 768-dimensional dense semantic vector representation for each sentence fragment. All sentence vectors are stored in a flat index structure built on the Faiss library, which supports efficient approximate nearest neighbor search. The style feature vectors generated in step S32 are used to filter the recommendation database. Specifically, metadata filtering conditions are used to restrict the candidate texts to texts whose genre is completely consistent with the text to be evaluated, whose length is within the same encoding range, and whose language complexity is within the same or adjacent encoding range. These three filtering conditions are used to select a subset of style-matching texts from the database of 95,000 texts. This subset typically contains 2,000 to 5,000 candidate texts. Semantic similarity retrieval is then performed on this style-filtered subset. The maximum cosine similarity between all sentence vectors of the text to be evaluated and the sentence vectors of each text in the subset is calculated and used as the semantic similarity score for that candidate text. The texts are then sorted from highest to lowest score, and the top 100 candidate texts with the highest scores are returned. The search results are weighted and merged for ranking. The semantic similarity score and the generated style matching score are linearly weighted. The style matching score is the average of the matching scores of three dimensions: genre, length, and language complexity. Each dimension scores 1 point for a successful match and 0 points for a failure. The weighting formula is that the final score is equal to 0.7 times the semantic similarity score plus 0.3 times the style matching score. The weights of 0.7 and 0.3 are fixed values ​​determined by adjusting the ranking quality index on the validation set.The first 100 results after weighted fusion and ranking are deduplicated using the maximum marginal relevance algorithm. This algorithm selects the text with the lowest average similarity to the selected set in turn to obtain the final recommended text set.

[0066] Preferably, the fusion score formula in step S3 is as follows:

[0067] ;

[0068] in, For semantic similarity scores, Weight parameters for style matching score The value ranges from 0.5 to 0.8, and is optimized based on the performance of the validation set.

[0069] In this embodiment of the invention, semantic similarity score The style matching score is obtained by calculating the maximum cosine similarity between the sentence vectors of the text to be evaluated and the candidate texts, with a value of 0.85. The matching score is calculated based on three dimensions: genre, length, and language complexity. In this example, genre matching scores 1 point, length matching scores 1 point, and language complexity matching scores 0.5 points. The average score is... =0.83. Weight parameter The nDCG metric, after optimization based on the validation set, was determined to be 0.7. Substituting these values ​​into the formula, the fusion score s = 0.7 × 0.85 + (1 - 0.7) × 0.83 = 0.844 was calculated. This score is used for the final ranking of candidate results; a higher score indicates a higher recommendation priority.

[0070] Preferably, step S4 includes the following steps:

[0071] Step S41: Merge the text to be evaluated with the evidence fragments in the candidate reading recommendation set into a unified context, input it into the domain-specific model for reasoning, and obtain preliminary results including the prediction of the school segment and its probability distribution;

[0072] Step S42: Apply temperature scaling and threshold shifting to calibrate the confidence level of the probability distribution of the preliminary results;

[0073] Step S43: Based on the calibration results, calculate the weighted directional deviation index and cross-level mobility in real time to monitor the deviation; if any index exceeds the preset threshold, the calibration parameter adjustment or rearrangement mechanism will be automatically triggered to intervene.

[0074] Step S44: Output the recommendation results, which include the calibrated learning stage conclusions, confidence levels, and interpretations of the chain of evidence.

[0075] In this embodiment of the invention, a recommendation database, a domain-specific model, and a candidate reading recommendation set are integrated, and dynamic updates and retraining are performed based on feedback data to form an adaptively optimized large reading recommendation model. The specific implementation process is as follows: Step S41 merges the text to be evaluated with the evidence fragments in the candidate reading recommendation set into a unified context, which is then input into the domain-specific model for reasoning. Specifically, the full text of the text to be evaluated is concatenated with the title, first paragraph, and corresponding expert-annotated grade level information of each text in the candidate set, according to the fixed format "Text to be evaluated: [Text content]. Reference evidence 1: [Title and first paragraph of candidate text 1], Applicable grade level: [Grade level 1]. Reference evidence 2: [Title and first paragraph of candidate text 2], Applicable grade level: [Grade level 2]." to form a complete input context. This context is then input into the Qwen2.5-32B-instruct domain-specific model trained in step S2. The model outputs the prediction result of the appropriate grade level for the text to be evaluated. This result includes the probability distribution of three categories: primary school, junior high school, and senior high school. For example, primary school: 0.15, junior high school: 0.70, and senior high school: 0.15. Step S42 applies temperature scaling and threshold shifting to calibrate the confidence level of the probability distribution of the preliminary results. The temperature scaling parameter T is set to 0.9, and each value in the original probability distribution is normalized by taking the 1 / 0.9th power of the original value. The threshold shifting parameter δ is set to -0.1, and the decision threshold for the junior high school category is adjusted from 0.5 to 0.4. The calibrated probability distribution becomes: primary school: 0.12, junior high school: 0.75, senior high school: 0.13. The final recommended grade level is junior high school, with a confidence level of 0.75. Step S43 calculates the weighted directional bias index and cross-level migration rate in real time based on the calibrated results to monitor the bias. A statistical window is maintained that scrolls over time, recording the actual and predicted results of the most recent 1,000 predictions. The weighted directional bias index and cross-level migration rate are calculated every two hours. If the absolute value of the weighted directional bias index exceeds 0.1 for two consecutive times, or the cross-level migration rate exceeds 0.1 for 24 consecutive hours, the calibration parameter adjustment process is automatically triggered. The temperature scaling parameter T and threshold shifting parameter are refitted using the most recent 1,000 data points with feedback. The output includes the calibrated grade level conclusion (junior high), a calibrated confidence score of 0.75, and an explanation of the chain of evidence. The explanation of the chain of evidence lists the titles, sources, and appropriate grade levels of the three candidate texts most similar to the text to be evaluated, and explains the matching of the text to be evaluated with the junior high school grade level in terms of language complexity and topic depth. Simultaneously, all intermediate results and decision parameters of this recommendation are recorded, and user feedback signals on the recommendation results are collected. This feedback data is written to a distributed message queue and fed back in batches to the expert recommendation database and model training module every 24 hours to initiate a new round of incremental training, thereby achieving adaptive optimization of the large-scale reading recommendation model.

[0076] Please refer to [link / reference needed] for further information. Figure 3 By integrating original reading materials with professional annotations from top-tier teachers, a structured expert recommendation database (ZPD-SCA) is formed. This database serves as a high-quality training sample, driving the domain-wide model training module to optimize the parameters of the general pre-trained model and generate a domain model with professional cognitive judgment capabilities. When a new text to be evaluated is input, the reading recommendation model's intelligent agent construction module initiates a multi-dimensional collaborative analysis process. This module first invokes a multi-dimensional retrieval mechanism to filter candidate materials with high relevance in style features and semantic vectors from the knowledge base. The agent then integrates the retrieval results with the domain model's deep reasoning capabilities to conduct a comprehensive judgment and decision. The agent outputs recommendation results containing clear grade-level conclusions, quantified confidence levels, and traceable evidence chains. These results undergo human evaluation, and the resulting feedback data forms a feedback signal, triggering an internal dynamic optimization mechanism. This mechanism continuously calibrates and iterates the domain model parameters and retrieval strategies, thereby achieving self-evolution in recommendation accuracy and adaptability.

[0077] This invention also provides a large-scale reading recommendation model generation system based on big data resources, used to execute the large-scale reading recommendation model generation method based on big data resources described above. The large-scale reading recommendation model generation system based on big data resources includes:

[0078] The recommendation database construction module 101 is used to acquire the text to be evaluated and collect reading materials from multiple learning stages and genres, annotate them, aggregate and verify the consistency of the annotation results, and build a verified recommendation database.

[0079] Domain-specific large model training module 102 is used to perform low-rank adaptive fine-tuning on a general pre-trained large model based on a recommendation database to obtain a domain-specific model, and to monitor and correct performance and bias during the training process.

[0080] The multi-dimensional retrieval module 103 is used to perform multi-dimensional analysis on the text to be evaluated, generate text feature vectors, combine semantic vector similarity retrieval, perform hybrid retrieval on the recommendation database, and form a candidate reading recommendation set through weighted fusion and deduplication sorting.

[0081] The large model building module 104 is used to integrate the recommendation database, the domain-specific model and the candidate reading recommendation set, and to perform dynamic updates and retraining based on feedback data to form an adaptive and optimized large reading recommendation model.

[0082] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for generating a large-scale reading recommendation model based on big data resources, characterized in that, Includes the following steps: Step S1: Obtain the text to be evaluated and collect reading materials from multiple grade levels and genres, annotate them, aggregate and verify the consistency of the annotation results, and build a validated recommendation database; Step S2: Perform low-rank adaptive fine-tuning on the general pre-trained large model based on the recommendation database to obtain a domain-specific model, and monitor and correct performance and bias during training; Step S2 includes the following steps: Step S21: Using the preset base model, construct two types of training samples from the recommendation database: zero-shot hint samples and context hint samples. Divide them into a training set and an evaluation set according to a preset ratio, and input them into the base model along with the text to be evaluated. The zero-shot hint sample input only contains the text to be evaluated and the task instructions. The context hint sample input extracts 3 to 5 highly relevant examples from the recommendation database through a multi-dimensional retrieval mechanism and is used together with the text to be evaluated. The training set is used for parameter updates, and the evaluation set is used for performance measurement and bias monitoring. Step S22: Perform domain fine-tuning on the base model, insert trainable parameters of low-rank decomposition into the base weight matrix of the model, and control the update amount by scaling factor, where the low-rank dimension is set to 32. Step S23: The training process uses cross-entropy loss combined with calibration constraints as the objective function, and sets the learning rate to... The batch size is 8, the weight decay parameter is 0.01, the training epochs are 3, the gradient clipping threshold is 1.0, and mixed precision training is used. Step S24: After training, calculate the overall accuracy, macro average, and weighted average on the evaluation set, and calculate the weighted direction bias index and cross-level mobility to monitor whether the model has cross-level misjudgments. Step S3: Perform multidimensional analysis on the text to be evaluated, generate text feature vectors, combine semantic vector similarity retrieval, perform hybrid retrieval on the recommendation database, and form a candidate reading recommendation set through weighted fusion and deduplication ranking; Step S4: Integrate the recommendation database, the domain-specific model, and the candidate reading recommendation set, and perform dynamic updates and retraining based on the feedback data to form an adaptive and optimized large reading recommendation model.

2. The method for generating a large-scale reading recommendation model based on big data resources according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Pre-screen extracurricular reading materials as reading materials and obtain the texts to be evaluated; Step S12: Mark the reading materials, assess their suitability for the appropriate grade level, and assign a score; Step S13: Aggregate opinions using a majority voting and consistency weighting method, and verify the consistency of the annotations to form the initial corpus; Step S14: Perform quality checks on the initial corpus, remove unsuitable or controversial texts, and form an annotated corpus; Step S15: Vectorize and embed the labeled corpus, verify the discriminative power of the data, and form a recommendation database.

3. The method for generating a large-scale reading recommendation model based on big data resources according to claim 1, characterized in that, The specific formula for calculating the weighted direction deviation index in step S2 is as follows: ； in For the index of the true category, An index for the class predicted by the model. Represents the true category Predicted as The number of samples, For category weights, when A value greater than 0 indicates an overall trend of "overvaluation". A value less than 0 indicates an underestimation of the trend.

4. The method for generating a large-scale reading recommendation model based on big data resources according to claim 1, characterized in that, The specific formula for calculating the cross-level mobility in step S2 is as follows: ； in Represents the true category Predicted as The sample size, i.e., the proportion of errors across non-adjacent grade levels in all erroneous predictions, indicates that the model is unstable in its grasp of key difficulty levels if the CLME value is too high.

5. The method for generating a large-scale reading recommendation model based on big data resources according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Extract text style features from the text to be evaluated. The features include genre category, length, language complexity index, sentiment polarity, and sentiment intensity. Step S32: Encode the extracted style features into low-dimensional discrete vectors, which will be used as metadata filtering conditions for searching the recommendation database; Step S33: Divide the text to be evaluated into paragraph-level or sentence-level segments, and generate a dense semantic vector representation for each segment, which is stored in an approximate nearest neighbor index structure; Step S34: Filter the recommendation database based on style features, limiting the candidate genres, length ranges, and language complexity ranges as a subset after style filtering; Step S35: In the subset after style filtering, perform similarity retrieval based on semantic vectors, and use normalized cosine similarity or inner product as the metric to return the candidate with the highest similarity. Step S36: Perform weighted fusion and sorting of the search results, and optimize the performance based on the validation set; Step S37: After weighted fusion and ranking, deduplication is performed on the results using the maximum marginal relevance constraint to form a candidate reading recommendation set.

6. The method for generating a large-scale reading recommendation model based on big data resources according to claim 5, characterized in that, The specific formula for the fusion score in step S3 is as follows: ； in, For semantic similarity scores, Weight parameters for style matching score The value ranges from 0.5 to 0.8, and is optimized based on the performance of the validation set.

7. The method for generating a large-scale reading recommendation model based on big data resources according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Merge the text to be evaluated with the evidence fragments in the candidate reading recommendation set into a unified context, input it into the domain-specific model for reasoning, and obtain preliminary results including the prediction of the school segment and its probability distribution; Step S42: Apply temperature scaling and threshold shifting to calibrate the confidence level of the probability distribution of the preliminary results; Step S43: Based on the calibration results, calculate the weighted directional deviation index and cross-level mobility in real time to monitor the deviation; if any index exceeds the preset threshold, the calibration parameter adjustment or rearrangement mechanism will be automatically triggered to intervene. Step S44: Output the recommendation results, which include the calibrated learning stage conclusions, confidence levels, and interpretations of the chain of evidence.

8. A reading recommendation model generation system based on big data resources, characterized in that, For executing the method for generating a large-scale reading recommendation model based on big data resources as described in claim 1, the system for generating a large-scale reading recommendation model based on big data resources includes: The recommendation database construction module is used to acquire the text to be evaluated and collect reading materials from multiple grade levels and genres, annotate them, aggregate and verify the consistency of the annotation results, and build a validated recommendation database. The domain-specific large model training module is used to perform low-rank adaptive fine-tuning on a general pre-trained large model based on a recommendation database to obtain a domain-specific model, and to monitor and correct performance and bias during the training process. The multi-dimensional retrieval module is used to perform multi-dimensional analysis on the text to be evaluated, generate text feature vectors, combine semantic vector similarity retrieval, perform hybrid retrieval on the recommendation database, and form a candidate reading recommendation set through weighted fusion and deduplication ranking. The large model building module integrates the recommendation database, the domain-specific model, and the candidate reading recommendation set, and performs dynamic updates and retraining based on feedback data to form an adaptive and optimized large reading recommendation model.

Images