Cross-prompt automatic composition scoring method based on data evaluation and meta-learning

Abstract

The invention discloses a cross-prompt automatic composition scoring method based on data evaluation and meta learning. The method comprises the steps of obtaining source prompt word data; performing value evaluation on the source cue word data to obtain a value probability of a cue word sample; obtaining a target value sample set according to the value probability; performing synonymous conversion enhancement on the target value sample set to obtain an enhanced sample set; processing the enhanced sample set to obtain enhanced training data; based on the enhanced training data, meta-learning cross-prompt automatic composition scoring model training is carried out, a target scoring model is obtained, and the target scoring model is used for carrying out automatic scoring on compositions.

Description

Technical Field

[0001] This invention belongs to the fields of natural language processing and reinforcement learning, and in particular relates to a cross-prompt automatic essay scoring method based on data evaluation and meta-learning. Background Technology

[0002] The core value of Automatic Essay Scoring (AES) lies in reducing teachers' workload and providing students with timely feedback, but current technologies still have many limitations. Early and mainstream research focused on specific prompting scenarios, relying on the assumption of independent and identically distributed data. Representative techniques include handcrafted feature extraction, CNN+LSTM+attention models, pre-trained model fine-tuning strategies, and pairwise contrastive regression. These techniques are only applicable to scenarios where the training and testing data use the same prompts and cannot generalize to new, unseen prompts, thus limiting their practicality. To improve adaptability, some research focuses on domain adaptation, such as constrained multi-task learning by Cummins et al., domain-adaptive models by Cao et al., and the "Skipflow" model by Tay et al. However, these methods all assume the existence of a certain amount of labeled target prompt data, while in actual cross-prompt scenarios, target prompt data is scarce, making it difficult to meet the preconditions.

[0003] For cross-prompt scoring requirements, existing technologies can be divided into two categories: feature-oriented and algorithm optimization-oriented. Feature-oriented technologies, such as PAES, pure feature-driven methods, PMAES / PLAES, and TDNN, can reduce distribution gaps by designing general features and contrastive learning. However, PAES and pure feature-driven methods rely on manual feature selection, which limits generalization. Moreover, none of these methods have solved the problems of noise interference and insufficient diversity in training data. Algorithm optimization-oriented technologies, such as the meta-learning framework MLCAES, can control distribution shifts through MMD, but they do not optimize for the quality and scale of training data. The shortcomings at the data level directly affect the model's generalization performance.

[0004] At the data level, existing data evaluation techniques have significant shortcomings: the LOO method tends to underestimate the value of samples, the Shapley method has an exponentially increasing computational cost, and the gradient Shapley (GS) method still maintains high computational overhead. Even though DVRL balances efficiency and accuracy, it is not combined with dataset augmentation and distribution consistency checks across cue AES, resulting in insufficient adaptability. In terms of data augmentation techniques, the back-translation process is complex, the effect of semantic bias mitigation is limited, random augmentation, MIXUP, and Transformer-based augmentation methods are prone to semantic and label bias, and quality control relies on human experience, making it difficult to uniformly adapt to essay data with different cuees. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention proposes a cross-prompt automatic essay scoring method based on data evaluation and meta-learning. This method reduces teachers' workload, provides students with timely feedback, and helps students improve their writing skills in a targeted manner.

[0006] To achieve the above objectives, this invention provides a cross-prompt automatic essay scoring method based on data evaluation and meta-learning, comprising:

[0007] Obtain source suggestion word data;

[0008] The source prompt word data is evaluated for value to obtain the value probability of the prompt word samples;

[0009] Based on the value probability, obtain the target value sample set;

[0010] The target value sample set is enhanced by synonym conversion to obtain an enhanced sample set;

[0011] The enhanced sample set is processed to obtain enhanced training data;

[0012] Based on the reinforcement training data, a meta-learning cross-prompt automatic essay scoring model is trained to obtain a target scoring model, wherein the target scoring model is used to automatically score essays.

[0013] Optionally, the source prompt word data is valued, and the probability of obtaining the value of the prompt word samples includes:

[0014] S1. Based on the semantic embedding and true score of the source prompt word data, calculate the initial value probability through a data value estimator, wherein the data value estimator is composed of a multilayer perceptron and a sigmoid activation function;

[0015] S2. Generate a sample selection indicator by Bernoulli sampling based on the initial value probability;

[0016] S3. Construct a weighted training loss based on the sample selection indicator, train the predictor to convergence using the weighted training loss, and obtain the current predictor parameters.

[0017] S4. Based on the current predictor parameters, calculate the gradient using the REINFORCE algorithm and update the data value estimator parameters;

[0018] S5. Repeat S1-S4 until the data value estimator converges and obtain the value probability of the prompt word sample.

[0019] Optionally, obtaining the target value sample set based on the value probability includes:

[0020] Based on the value probability, the source prompt word data is divided using a clustering algorithm to obtain the target value sample set.

[0021] Optionally, the target value sample set is enhanced by synonym transformation to obtain the enhanced sample set, including:

[0022] Based on each sample in the target value sample set, a synonym conversion is performed using a large language model to generate new samples with different expressions while retaining the rating labels. The rating labels include core semantics, writing style, and rating dimensions.

[0023] Based on the new samples, construct the enhanced sample set.

[0024] Optionally, processing the augmented sample set to obtain augmented training data includes:

[0025] The distribution correlation test is performed on the enhanced sample set and the target prompt test data in the original data to obtain the distribution consistency index;

[0026] Enhanced samples that meet the distribution requirements are selected based on the distribution consistency index, and the selection results are obtained;

[0027] The selection results are then optimized for diversity to obtain the reinforcement training data.

[0028] Optionally, a distribution correlation test is performed on the enhanced sample set and the target cue test data to obtain a distribution consistency index, including:

[0029] Based on the enhanced sample set and the target cue test data, sample features are extracted and mapped to the regenerated Hilbert space to obtain target sample features;

[0030] Based on the target sample features, the similarity between sample pairs is calculated, and the kernel function value matrix is ​​obtained;

[0031] Based on the kernel function value matrix, calculate the mean kernel function of sample pairs within the augmented sample set, within the target cue test data, and across sets;

[0032] The maximum mean difference is calculated based on the combination of the kernel function mean values ​​to obtain the distribution consistency index.

[0033] Optionally, the selection results are optimized for diversity to obtain the reinforcement training data, including:

[0034] Based on the filtering results, semantic feature vectors and structural feature vectors are extracted;

[0035] Calculate the modified cosine distance between samples based on the semantic feature vector and the structural feature vector, and obtain the distance matrix;

[0036] A positive semi-definite kernel matrix is ​​constructed based on the distance matrix, wherein the elements of the kernel matrix are obtained by transforming the distance values ​​through an exponential function;

[0037] Based on the kernel matrix, a greedy approximation algorithm iteratively selects samples that maximize the increment of the subset determinant until a preset number is reached, thereby obtaining the reinforcement training data.

[0038] Optionally, training a meta-learning cross-prompt automatic essay scoring model based on the reinforcement training data to obtain the target scoring model includes:

[0039] Based on the enhanced training data, a virtual support set and a virtual query set are constructed, and real unseen prompt data is introduced to obtain the meta-learning training task.

[0040] Based on the meta-learning training task, the set of candidate meta-learner parameters is updated and generated through gradient descent.

[0041] Based on the feature distribution of the candidate meta learner parameter set and the real unseen prompt data, the distribution distance is calculated by the maximum mean difference.

[0042] Optimal candidate meta-learner parameters are selected based on the distribution distance;

[0043] Based on the optimal candidate meta-learner parameters, the global parameters of the model are optimized through a joint loss function, wherein the joint loss includes the meta-training loss of the virtual support set, the meta-test loss of the virtual query set, and the dynamic offset margin based on the distribution distance.

[0044] The target scoring model is obtained based on the optimized global parameters of the model.

[0045] Compared with the prior art, the present invention has the following advantages and technical effects:

[0046] (1) This invention verifies the effectiveness of meta-learning mechanisms in scoring this type of essay.

[0047] (2) This invention proposes a data augmentation method based on DVRL after data evaluation, and combines it with meta-learning to improve the effect of essay cross-prompt scoring.

[0048] (3) This invention designs a method for screening enhanced samples using DPP. By constructing a kernel matrix and greedy approximate sampling, the sample diversity is maximized while ensuring consistent data distribution, thus avoiding redundancy.

[0049] (4) The present invention uses the MMD two-stage test to ensure that the distribution of data in the enhancement stage is not significantly shifted. Attached Figure Description

[0050] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0051] Figure 1 This is a flowchart of a cross-prompt automatic essay scoring method based on data evaluation and meta-learning according to an embodiment of the present invention;

[0052] Figure 2 This is a flowchart of sample 1 of the present invention;

[0053] Figure 3 This is a schematic diagram of sample 2 of the present invention;

[0054] Figure 4 This is a schematic diagram comparing the enhanced sample and the original sample data in an embodiment of the present invention. Detailed Implementation

[0055] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0056] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0057] This embodiment uses large-scale models (such as GLM-4 and GPT series) with powerful semantic understanding and text generation capabilities to perform synonym transformation on high-quality essay data (i.e., samples with clear scoring labels, content that meets the prompt requirements, and logical and expressive norms) in the AES task. While preserving the core semantics and writing quality features of the original text, it generates new samples with diverse expressions to expand the scale of training data and improve the model's generalization ability. Specifically, the large-scale model can achieve synonym transformation based on the contextual information of high-quality data through synonym replacement, sentence reconstruction (such as conversion between active and passive sentences, rewriting of complex and simple sentences), and pragmatic expression adjustment. For example, for the high-quality sentence "Computers help people access information quickly", the large-scale model can generate synonyms such as "People can obtain information rapidly with the assistance of computers".

[0058] This embodiment proposes a cross-prompt automatic essay scoring method based on data evaluation and meta-learning, such as... Figure 1 As shown, the specific steps include:

[0059] Obtain source suggestion word data;

[0060] Value assessment of source prompt word data to obtain the value probability of prompt word samples;

[0061] Based on the value probability, obtain the target value sample set;

[0062] Perform synonym transformation enhancement on the target value sample set to obtain an enhanced sample set;

[0063] The augmented sample set is processed to obtain augmented training data;

[0064] A meta-learning cross-prompt automatic essay scoring model is trained based on reinforcement training data to obtain a target scoring model, which is used to automatically score essays.

[0065] Specifically, to improve the training data quality and model generalization ability of meta-learning cross-prompt automatic essay scoring task, this paper first conducts a systematic data quality assessment on the original data corresponding to the source prompts. Specifically, it adopts the DVRL (Data Valuation using Reinforcement Learning) data assessment framework. This framework quantifies the "value" of data samples through reinforcement learning mechanisms, which can accurately identify high-quality data that makes a positive contribution to model training, as well as low-quality noisy data containing labeling errors, semantic ambiguity, or information redundancy (such as logically broken segments in essay samples, redundant expressions unrelated to the scoring criteria, etc.). Based on the assessment results, low-quality data is filtered and removed, laying a clean data foundation for subsequent training. After acquiring a high-quality subset of data, the GLM-4 large language model is introduced for data synonym conversion enhancement. Leveraging GLM-4's powerful semantic understanding and generation capabilities, multiple semantically equivalent but differently expressed samples are generated while strictly preserving the core semantics, writing style, and scoring dimensions (such as content completeness, logical coherence, and language standardization) of the original essay data. This expands the training data scale and avoids model overfitting due to data uniformity. To ensure that the enhanced data and the final test data used by the model maintain consistent distribution and avoid the impact of data distribution shift on scoring accuracy, the generated enhanced data and the test set data undergo an MMD (Maximum Mean Discrepancy) distribution test. MMD measures the distance between two data distributions in the Reproducing Kernel Hilbert Space (RKHS) to determine whether they come from the same distribution. Only when the test results show no significant difference in distribution between the enhanced data and the test set data does the next processing step begin. Subsequently, to further optimize data diversity and avoid potential sample redundancy or semantic repetition issues during the augmentation process, the Deterministic Point Process (DPP) was used to filter the data that passed the MMD test. DPP maximizes the differences between samples within the subset while ensuring consistency between the data subset and the original high-quality data distribution, effectively improving the coverage and representativeness of the training data. Finally, the high-quality data subset, filtered through DVRL evaluation, GLM-4 synonym augmentation, MMD distribution testing, and DPP diversity optimization, replaced the traditional original training set data for training the meta-learning cross-prompt automatic essay scoring model, thereby improving the model's scoring consistency and accuracy across different essay prompt scenarios. In the meta-learning stage: training, validation, and testing data were first organized according to specified topics. Then, each essay was converted into multi-source features, and a regression model was trained to fit the scores. The scoring performance of the model was then evaluated on the development and test sets.Specifically, the data phase reads samples of the question and their scores, scaling the scores to a suitable learning range. The feature phase performs hierarchical padding and flattening of the text's part-of-speech sequences, while extracting linguistic handcrafted features and readability features to form the model's multi-input vectors. The model phase uses a multi-input network (including attention modules) to learn the mapping between "text features and scores," using mean squared error as the target during training to make the predicted scores as close as possible to the true scores. The training process aligns samples according to the question's bin index, iteratively improving the fitting ability round by round. The evaluation phase runs the evaluator on the development / test set after each round, comparing the model's predicted scores with the true scores, outputting metrics such as error and relevance to measure the accuracy and stability of the automatic scoring. Finally, the model's prediction on the test set is the automatic scoring result for the essay on that question, and the evaluation metric QWK is used to determine whether the scoring effect has met expectations.

[0066] Furthermore, the source prompt word data is valued, and the probability of obtaining the value of the prompt word samples includes:

[0067] S1. Based on the semantic embedding and true score of the source prompt word data, calculate the initial value probability through the data value estimator, which consists of a multilayer perceptron and a sigmoid activation function.

[0068] S2. Generate a sample selection indicator based on the initial value probability using Bernoulli sampling;

[0069] S3. Construct a weighted training loss based on the sample selection indicator, train the predictor to convergence using the weighted training loss, and obtain the current predictor parameters.

[0070] S4. Based on the current predictor parameters, calculate the gradient using the REINFORCE algorithm and update the data value estimator parameters;

[0071] S5. Repeat S1-S4 until the data value estimator converges, and obtain the value probability of the prompt word sample.

[0072] Furthermore, based on the value probability, obtaining the target value sample set includes:

[0073] Based on the value probability, the source prompt word data is divided using a clustering algorithm to obtain a target value sample set.

[0074] Specifically, the mathematical framework of DVRL revolves around three key stages: "objective function of reinforcement learning," "data value calculation," and "parameter update." All formulas serve to enable the data value estimator to accurately determine whether a sample is useful to the AES model. First, data value calculation is fundamental. The data value estimator outputs the value probability of each sample (i.e., the likelihood of selecting that sample). This step corresponds to "action decision" in reinforcement learning, directly determining which samples will be selected to train the predictor. Second, the objective function of reinforcement learning is defined by "reward"—after training the predictor with the selected samples, the predictor's performance on the validation set (such as the quadratic weighted Kappa commonly used in AES tasks) is used as the reward. The core of the objective function is to maximize this reward. Finally, parameter update is the means to achieve the objective. Through the REINFORCE algorithm, combined with the previously obtained reward and the probability gradient corresponding to the data selection results, the parameters of the data value estimator are updated, making subsequent data value calculations more accurate (able to more accurately distinguish between high / low value samples), thereby driving the entire process to iteratively optimize and continuously improve the data evaluation effect.

[0075] Overall optimization goals:

[0076] The essence of DVRL is to train a "data value estimator" through reinforcement learning, so that the sample value it outputs can guide the "predictor" (AES model) to achieve optimal performance on the validation set. The optimization objective is as follows:

[0077] ;

[0078] in, It is a data value estimator The parameters (the estimator is a small MLP that outputs the "value score" of the sample), It is a predictor The parameters (the predictor is an AES model using DeBERTa+MLP, used to predict essay scores). / These are the corresponding training and validation sets. The output of the data value estimator (input is essay x and its true score y, output value) That is, the "value probability" of a sample. It is the output of the predictor (input is essay x, output is predicted score). ), It uses MSE as the loss function for the predictor. , It is the "reward function" of reinforcement learning (i.e., the reward function of the predictor on the validation set). The paper uses the QWK metric to measure performance (higher values ​​result in greater rewards). For datasets The expected value (i.e., the average over all samples). Part of the constraint is to first fix the parameters of the data value estimator. Find the predictor parameters that minimize the "weighted loss". The "weighted loss" here is... (Sample value) is a weighted loss for each training sample; samples with higher value ( Larger samples contribute more to the loss, so the predictor will focus more on these types of samples, while low-value noisy samples ( Small contributions are considered "weakened". Objective function, then adjust the parameters of the data value estimator. To make the "optimal predictor" Performance on the validation set "Maximize. The essence of this process is to teach the estimator to "identify which samples can help the predictor perform better on the validation set", that is, to accurately output the true value of the samples.

[0079] Data value calculation:

[0080] Data Value Estimator The output is the "value probability" of the sample. The Sigmoid activation function restricts the output to the interval [0,1], as shown in the following formula:

[0081] ;

[0082] in: It is an essay The pre-trained embeddings are extracted using DeBERTa to capture the semantic features of the essay; It is the main body of the data value estimator (5-layer fully connected network);

[0083] It is the Sigmoid function. ,make sure . This represents the probability that the sample will be selected into the training set. The closer it is to 1, the more the estimator judges the sample to be "high value and noise-free" and it should be used to train the predictor. The closer the value is to 0, the more "low-value and noisy" the sample is, and it should be excluded.

[0084] Sample selection indicator (Bernoulli sampling):

[0085] With value probability Subsequently, DVRL generates a "sample selection indicator" through "Bernoulli sampling". The formula for deciding whether to use the sample for predictor training is as follows:

[0086] ;

[0087] in, It is a Bernoulli distribution: The probability of =1 is (Sample selected) The probability of =1 is 1- (Samples are excluded). This step is a "randomized selection," not a "deterministic selection" (i.e., not a direct selection). The reason for this is that reinforcement learning needs to explore the impact of different sample combinations on the predictor's performance, avoiding the estimator from getting trapped in local optima (such as focusing only on a certain type of sample). For the AES task, this step can avoid "misjudging high-value samples": even if a truly high-value sample... Even at a slightly lower value (e.g., 0.45), there is still a 45% probability of being selected, ensuring the robustness of training.

[0088] Parameter update of data value estimator (REINFORCE algorithm):

[0089] DVRL uses the REINFORCE algorithm in reinforcement learning to update the parameters of the data value estimator. The core is "using rewards" gradient adjustment This leads to an overestimation of the value of samples that could improve validation set performance. The gradient update formula is as follows:

[0090] Core formula for gradient calculation:

[0091] ;

[0092] Likelihood function of sample selection :

[0093] ;

[0094] in, A set of selection indicators for all samples ; : Generate the current selection set under parameter $\phi$. The "likelihood probability" - that is, "selecting all" Samples with a value of 1 and exclude all. The joint probability of "samples with =0".

[0095] The gradient of the logarithm of the likelihood function (“log-likelihood gradient”) has the physical meaning of “adjusting”. To what extent can it change the probability of the current selection set?

[0096] Gradient update logic:

[0097] Calculate the log-likelihood gradient: Taking the logarithm and then the derivative (simplifies calculation; the logarithm does not change the monotonicity):

[0098] ;

[0099] right Find the gradient:

[0100] ;

[0101] The intuitive meaning of this gradient is: if a sample is selected ( And currently If the value is small (underestimated), the gradient is positive, and the update... Let Increase (correct underestimation); if the sample is excluded ( And currently If the value is too large (overestimated), the gradient is negative, and the update... Let Reduce (correct overestimation).

[0102] Combining reward with parameter updates: The final parameter update formula is (using the learning rate) (Control step size)

[0103] ;

[0104] If the current reward A higher gradient (the predictor performs well on the validation set) means the gradient is amplified. Larger update steps strengthen the "current sample selection logic" (i.e., make the estimator more inclined to give high value to this type of sample); if the reward If the gradient is low (the predictor performs poorly), then the gradient is reduced. The update step size is smaller—weakening the "current sample selection logic" and prompting the estimator to adjust its judgment criteria.

[0105] Complete DVRL algorithm flow:

[0106] The training process of DVRL is a loop of "iterative optimization of the predictor and data value estimator", consisting of 5 steps. Each step closely revolves around the above formula. The following is a detailed breakdown using the AES task (essay scoring):

[0107] Algorithm input and initialization:

[0108] Input: Training set for the AES task (Noisy essay samples), high-quality validation set (Manual confirmation of no noise), predictor structure (DeBERTa + 2-layer MLP), data value estimator structure (5-layer MLP).

[0109] Initialization: Randomly initialize the predictor parameters and data value estimator parameters Extract semantic embeddings of all essays using a DeBERTa pre-trained model. (Fixed, not used in DVRL training, only as feature input); Set hyperparameter: learning rate Number of iterations Predictor training rounds .

[0110] Iterative training steps:

[0111] Step 1: Calculate the value probability of each training sample :

[0112] For the training set Each sample in :enter (Embedded in the essay) and (True Score) to Data Value Estimator The value probability of a sample is obtained by using a 5-layer MLP and sigmoid activation. The estimator makes an initial judgment on whether "this essay is useful for training the scoring model," such as a high score. The lower scores likely correspond to essays with "clear structure and accurate score markings," while the lower scores correspond to noisy samples with "disorganized content and incorrect markings."

[0113] Step 2: Bernoulli sampling generates a sample selection indicator. :

[0114] For each sample Bernoulli distribution Sampling was performed in the sample to obtain... : The samples are selected into the "effective training set" for training the predictor; Samples are excluded and will not be used for training for the time being. Random sampling is used to "explore" the data, preventing the estimator from prematurely locking onto a particular type of sample and ensuring that subsequent parameter updates can cover more sample combinations.

[0115] Step 3: Train the predictor using the selected samples. :

[0116] Constructing a "weighted training loss": based on selected samples Calculate the weighted average of the losses:

[0117] ;

[0118] (The denominator is the number of selected samples, ensuring the loss is a weighted average): Minimize using the Adam optimizer. Update predictor parameters ,train The current optimal parameters are obtained after the round. The predictor prioritizes learning from "selected and high-value" samples (such as essays with "high scores and clear logic"), while giving low weight to "excluded or low-value" noisy samples (such as "narrative essays with incorrect score labeling") to reduce the interference of noise on the model.

[0119] Step 4: Calculate the reward And update the data value estimator parameters. :

[0120] The trained predictor In the validation set In the test, the QWK metric was used as the reward. (The higher the QWK value, the greater the reward); according to the formula Calculate the parameters of the data value estimator. The gradient; use gradient ascent (because the goal is to maximize). )renew : If the current predictor performs well on the validation set... (High), indicating "the current sample selection logic (by...)" If the decision is correct, then this logic is amplified (let...). More inclined to give high scores to this type of sample Conversely, adjust The criteria for judging sample value were revised.

[0121] Step 5: Repeat the iteration until convergence:

[0122] Repeat Steps 1 through 4. After several rounds, a well-trained data value estimator is finally obtained. At this point, input any training sample. , Output This refers to the "final data value" of the sample.

[0123] Furthermore, the target value sample set is enhanced by synonym transformation to obtain the enhanced sample set, including:

[0124] Based on each sample in the target value sample set, a large language model is used to perform synonym conversion. While retaining the rating labels, new samples with different expressions are generated. The rating labels include core semantics, writing style and rating dimensions.

[0125] Based on the new samples, construct an enhanced sample set.

[0126] Furthermore, the augmented sample set is processed to obtain augmented training data, including:

[0127] A distribution correlation test is performed on the enhanced sample set and the target cue test data in the original data to obtain the distribution consistency index. The original data is the existing data.

[0128] Enhanced samples that meet the distribution requirements are selected based on the distribution consistency index, and the selection results are obtained;

[0129] The selection results are optimized for diversity to obtain enhanced training data.

[0130] Specifically, the GLM-4 model is used to generate augmentation data:

[0131] use calculate Each sample Select high-value samples from the dataset. Use K-means (K-means) ) The samples were arranged according to Divided into "high-value clusters" ( Large) and "low-value clusters" Small sample These are considered high-value samples. GLM-4 Synonym Transformation: For the selected high-value samples, GLM-4 is used for synonym transformation. Let the GLM-4 transformation function be... It receives high-value samples. and output the synonym-converted text. ,Right now The transformed results are then integrated to form a new dataset. The original high-value samples... After GLM-4 conversion, the result is New dataset .

[0132] DPP diversity:

[0133] Feature extraction (constructing input for DPP):

[0134] Semantic and structural features are extracted from the transformed text to calculate the "similarity" between samples (the core input of DPP). Let the transformed data be a set. , ( For the sample size, (This is the synonym-transformed text generated by GLM-4), for each Extracting feature vectors ( (For feature dimensions).

[0135] DPP Diversified Screening:

[0136] DPP (Determinantal Point Processes) quantifies the "exclusivity" between samples using a kernel matrix (the lower the similarity, the more likely they are to be selected simultaneously), ultimately selecting the subset with the maximum diversity.

[0137] Kernel matrix construction:

[0138] The core of DPP is the kernel matrix. ,in Indicates sample and The "similarity" (the smaller the value, the more diverse the samples). Combining essay features, the kernel matrix formula is:

[0139] ;

[0140] in: : Corrected cosine distance (value) (The larger the value, the greater the difference in features between the two samples);

[0141] Bandwidth parameter (controls the sensitivity of similarity; recommended for essay data). (More sensitive to structural / lexical differences);

[0142] Convert distance to similarity (the greater the difference, the higher the similarity). The closer a sample is to 0, the more likely it is to be selected simultaneously.

[0143] kernel matrix It must satisfy positive semidefiniteness (to ensure that the probability distribution of DPP is effective).

[0144] DPP sampling (selecting diverse subsets):

[0145] DPP selects k samples from X using "determinant probability" (k is the preset selection number). The probability of selecting a subset S⊆X is:

[0146] ;

[0147] in: It is a submatrix of the corresponding subset S in the kernel matrix L; det(·) is the determinant operation (the larger the determinant value, the higher the diversity of the subset S); the denominator is the normalization term (to ensure that the sum of probabilities is 1).

[0148] Sampling algorithm (greedy approximation, suitable for large-scale data): Initialize an empty set S = ∅; for each unselected sample Calculate the increment of the determinant after adding S. ;choose The largest sample is added to S, and this process is repeated until |S|=k. This ultimately yields a diverse subset. .

[0149] Diversity index test:

[0150] The data value assessment constructs three core categories of indicators to comprehensively measure the effectiveness and rationality of the assessment results from multiple dimensions. The first category is quality relevance indicators, including the quality Pearson correlation coefficient and the quality Spearman correlation coefficient. The former measures the linear correlation between data value and text quality score by calculating pearsonr(data_values, quality_scores). The stronger the positive correlation, the higher the data value of high-quality text, directly reflecting the linear matching degree between the assessment results and text quality. The latter measures the monotonic correlation between the two by spearmanr(data_values, quality_scores). Its advantage lies in being more robust to outliers, focusing on ranking consistency rather than absolute values, and can supplement and verify the accuracy of the assessment results in text quality ranking. The second category is score correlation indicators, which include score Pearson correlation coefficient and score Spearman correlation coefficient. The former measures the linear correlation between data value and the original human score using pearsonr(data_values,y_train). The results in this paper show that it may be negatively correlated, which suggests that low-scoring samples may have higher model learning value and provide a reverse reference for data screening. The latter measures the monotonic consistency between the data value ranking and the human score ranking, reducing the evaluation bias caused by absolute numerical differences from the ranking dimension. The third category is the quality separation index, which includes the separation score, the average value of low-quality samples, and the average value of high-quality samples. The separation score is calculated by mean(high_quality_values) - mean(low_quality_values), with a positive expectation and the larger the value, the better. The larger the difference, the stronger the model's ability to distinguish between good and bad samples. The average value of low-quality samples is the average data value of the 10% of samples with the worst quality. The expectation is relatively low to verify the model's accurate identification of low-quality samples. The average value of high-quality samples is the average data value of the 90% of samples with the best quality. The expectation is significantly higher than that of low-quality samples. Together, they constitute the "value threshold" reference for quality differentiation, ensuring that the model can effectively assign reasonable values ​​to samples of different quality.

[0151] Strengthen the correlation test between the data and the original data:

[0152] Two MMD checks are required to verify whether GLM-4 synonym transformation "preserves the semantic / structural features of the original data" (avoiding data distribution shift after transformation); and to verify whether DPP screening "destroys the diversity distribution of the original data" (ensuring that the screened data can still represent real writing scenarios).

[0153] Data preparation:

[0154] Data Preparation: Define the set of augmented data (or augmented data generated by GLM-4) after DPP screening and processing, and the set of validation data. Also define the feature vector form of the processed text. Feature Mapping: Map the data samples to the Regenerated Hilbert space (RKHS). This mapping is implicitly achieved through kernel functions, commonly such as the Gaussian kernel (RBF kernel).

[0155] Calculation formula:

[0156] MMD is used to measure the difference in means between two distributions P and Q in the RKHS. Given a dataset X sampled from P and a dataset Y sampled from Q, its empirical estimation formula is:

[0157] ;

[0158] Where k(·,·) is the kernel function, which must satisfy the properties of being positive definite and symmetric. The commonly used Gaussian kernel (RBF kernel) has the following form:

[0159] ;

[0160] Here, σ is the bandwidth parameter of the Gaussian kernel, controlling the smoothness of the kernel function. It needs to be selected appropriately based on the data (it can be determined through methods such as cross-validation). The first two items are for augmented datasets. Internal, validation dataset The average kernel function value of the internal sample pairs reflects the similarity within each dataset; the third term is the average kernel function value of the sample pairs between the two datasets, reflecting the similarity across datasets; finally, by combining these parts, we obtain an index that measures the difference between two distributions. The smaller the MMD value, the more similar the distributions of the two datasets are, and the higher the correlation.

[0161] After passing the MMD evaluation, we obtained reinforcement data with high similarity and diversity. .

[0162] Furthermore, a distribution correlation test is performed on the enhanced sample set and the target cue test data to obtain distribution consistency indices, including:

[0163] Based on the enhanced sample set and the target cue test data, sample features are extracted and mapped to the regenerated Hilbert space to obtain target sample features;

[0164] Based on the characteristics of the target samples, the similarity between sample pairs is calculated, and the kernel function value matrix is ​​obtained;

[0165] Based on the kernel function value matrix, calculate the mean kernel function of sample pairs within the augmented sample set, within the target cue test data, and across sets;

[0166] The maximum mean difference is calculated based on the combination of kernel function mean values ​​to obtain the distribution consistency index.

[0167] Furthermore, the selection results are optimized for diversity, and the resulting reinforcement training data includes:

[0168] Extract semantic feature vectors and structural feature vectors based on the screening results;

[0169] Calculate the modified cosine distance between samples based on semantic feature vectors and structural feature vectors, and obtain the distance matrix;

[0170] Construct a positive semi-definite kernel matrix based on the distance matrix, where the elements of the kernel matrix are obtained by transforming the distance values ​​using an exponential function;

[0171] Based on the kernel matrix, a greedy approximation algorithm is used to iteratively select samples that maximize the increment of the subset determinant until a preset number is reached, thereby obtaining reinforcement training data.

[0172] Furthermore, a meta-learning cross-prompt automatic essay scoring model is trained based on reinforcement training data to obtain the target scoring model, including:

[0173] Virtual support sets and virtual query sets are constructed based on reinforcement training data, and real unseen prompt data is introduced to obtain meta-learning training tasks;

[0174] Based on the meta-learning training task, the set of candidate meta-learner parameters is updated and generated through gradient descent.

[0175] Based on the feature distribution of the candidate meta-learner parameter set and the real unseen prompt data, the distribution distance is calculated by the maximum mean difference.

[0176] Optimal candidate meta-learner parameters are selected based on distribution distance;

[0177] Based on the optimal candidate meta-learner parameters, the global parameters of the model are optimized through a joint loss function, which includes the meta-training loss of the virtual support set, the meta-test loss of the virtual query set, and the dynamic offset margin based on the distribution distance.

[0178] The target scoring model is obtained based on the optimized global parameters of the model.

[0179] Specifically, meta-learning cross-prompt automatic essay scoring:

[0180] Using augmented data Replace the original training data The system performs cross-prompt automatic essay scoring by systematically solving the problem of "generalization failure caused by distribution offset" through a closed loop of "constructing virtual tasks, generating candidate models through multivariate training, selecting the optimal model through distribution guidance, and optimizing joint loss".

[0181] Meta-learning's "virtual task construction":

[0182] Cross-cue AES is essentially a "small sample generalization problem"—seen cue data is the "SupportSet," and unseen cue data is the "QuerySet." MLCAES first randomly constructs "virtual seen cue data" from the seen cue data. ")" and "virtual target prompts ( This involves introducing unlabeled data that was not explicitly mentioned. This simulates a cross-cue scenario during the "training-test" process. The core of this step is creating distributional differences: allowing the model to frequently encounter "seen" prompts during training. The distribution changes of "unseen" provide "generalization training materials" for subsequent meta-learning.

[0183] Generate candidate models that adapt to different prompts:

[0184] Faced with virtual tasks, MLCAES performs multi-metatraining: for each virtual task, it provides the necessary hints. Use gradient update Generate a "candidate meta-learner" If the virtual prompt is "Argumentative Essay Grading Task", then It will adapt to the semantic features of argumentative essays (such as higher weighting for arguments and evidence); if the virtual prompt is a "narrative task", then It will focus on narrative structure (such as timeline and feature extraction of character descriptions). In this way, the model generates a set of candidate meta-learners covering different cue distributions, providing a foundation for the subsequent "selection of the model that best fits the unseen cue".

[0185] Meta-learner selection:

[0186] Traditional meta-learning models that randomly select candidate models are prone to generalization failure due to "distribution mismatch". MLCAES introduces the Maximum Mean Difference (MMD) method. ), quantifying the difference between the "feature distribution of candidate meta learners" and the "feature distribution of true unseen prompts (Query)".

[0187] It is the feature output of the candidate meta-learner on the meta-test sample (such as the semantic + structural embedding of the essay). It is a feature representation of real, unseen samples; The smaller the value, the closer the distribution of the candidate model is to the unseen prompt, and the more likely it is to generalize in real-world tasks. This step, guided by the "target distribution," selects the candidate meta-learners that have the "most potential to adapt to the unseen prompt," preventing the model from generalizing in an unreasonable direction.

[0188] Meta-testing and joint loss:

[0189] After selecting the optimal candidate meta-learners, MLCAES enters the meta-test phase, where the global parameters of the model are optimized using the "joint loss". );

[0190] Meta-training loss ( ): Ensure the model retains knowledge of previously seen prompts (such as essay scoring ability);

[0191] Meta-test loss ( ): Guide the model to adapt to the distribution of virtual target prompts;

[0192] Offset margin ( ): Dynamically control the weights of the meta-test loss—if the distribution of the observed target differs significantly from that of the virtual target ( (large), then Small, strengthen meta-test optimization; if the distribution is close to ( (small), then Larger, weaker meta-test optimization. Through this "dynamic balance," the model can both learn from seen prompts and generalize to unseen prompts, avoiding "overfitting to seen prompts" or "blindly adapting to unseen prompts."

[0193] The following is a detailed description of this embodiment with reference to the accompanying drawings:

[0194] This example uses the ASAP dataset, which (Table 1) contains eight groups of papers, each belonging to a different paper prompt.

[0195] Table 1

[0196]

[0197] To demonstrate the state-of-the-art nature of this study, the experimental results were compared with four competitive baselines:

[0198] CNN-CNN-MoT: The scoring model uses two CNN layers to extract the paper representation.

[0199] 2L-LSTM: Use a two-layer dual LSTM from GloVe (Alikaniotis et al., 2016) to obtain the semantic features of the paper.

[0200] BERT: The transformer-based pre-trained model has shown excellent performance on many NLP tasks (Devlin et al., 2019).

[0201] PAES: This model (Ridley et al., 2020) uses a neural model with handcrafted features to achieve cross-suggested paper scoring.

[0202] PMAES: The state-of-the-art and leading cross-cue single-total paper scoring model (Chen & Li, 2023) using a cue mapping contrastive learning approach.

[0203] Implementation details:

[0204] This embodiment uses DVRL as the scoring base library to achieve data value assessment. For the hyperparameters of DVRL tuning, QWK is used as the metric, with 1000 outer iterations, 100 inner iterations, a batch size of 100, a predicted batch size of 128, and a learning rate of 0.001. The hidden layer dimension is 200, and the combined layer dimension is 50. Subsequently, clustering is used to separate noise and non-noise data. After removing noise, the data is processed using a GLM-4 model for synonym transformation to enrich high-value data. After obtaining a large amount of data, the DPP method is used to sample the data to ensure data diversity while avoiding redundancy. The feature dimension is 50, and the regularization parameter is 1e-6. After another MMD test, meta-learning is performed. For the meta-learning stage, the Rmsprop optimizer is used in the multivariate training and meta-testing stages, with an initial learning rate set to 0.001. The total number of iterations is 50, and the mini-batch size is 5. For the cue-aware paper feature network, the BERT model is used. The maximum paper length for BERT is set to 512. The S-shaped layer dimension for obtaining Prompt-Essay compliance scores is 1. We use the length of the longest paper in the dataset as the padding length to ensure that paper information is preserved as much as possible. During training, the procedural scores of the papers are rescaled to the range [0, 1]. For evaluation, predictions are rescaled to the corresponding prompt score range. This method uses NVIDIA L40 for training.

[0205] The specific process involves first performing DVRL data evaluation on the source prompt data to filter out high-quality and low-quality data, removing low-quality noise data; secondly, performing GLM-4 data synonym transformation enhancement on the high-quality data, with the generated data inheriting the scores of the original data; next, performing MMD verification on the generated data and the test set data; and then, after passing the verification, performing DPP processing on the data again to ensure data diversity. Finally, the processed data replaces the training set data for meta-learning cross-prompt automatic essay scoring.

[0206] Table 2 shows the excellent results of this method. Our method was compared with various baselines, demonstrating its effectiveness. Experimental results show that the meta-learning model (META-LEARN) significantly outperforms other models compared to state-of-the-art cross-cue AES methods (PAES, PMAES), while BERT and CNN-LSTM-ATT are leading models for prompting relevant paper scores. Methods for prompting specific settings (CNN-CNN-MoT, 2L-LSTM, CNN-LSTM-ATT) significantly underperform BERT and cross-cue AES methods. The improved method, evaluated with data, further enhances the performance of meta-learning, fully demonstrating the feasibility and effectiveness of our approach, achieving an improvement of nearly 5.9% compared to state-of-the-art cross-cue AES methods.

[0207] Table 2

[0208]

[0209] Ablation experiment:

[0210] To further ensure the effectiveness and performance of our proposed method for the task, we compared the results of simple meta-learning with those of data evaluation augmentation, as shown in Table 3.

[0211] Table 3

[0212]

[0213] Our method improves the QWK score for each prompt, demonstrating its effectiveness in generalizing the model, particularly in prompts 1 and 7. Its mean score also surpasses that of meta-learning methods, indicating that our proposed method has a clear effect on further improving cross-prompt essay scoring.

[0214] Case Study:

[0215] Experimental studies were conducted on cases that might affect generalization ability in this research method. For example, among the five lowest-value samples identified in prompt2, sample 1... Figure 2 As shown, Figure 2 The least valuable sample information in prompt2 is as follows: first, the text ID; second, the essay text content; next, the normalized essay score and its resulting data value; and finally, the sample size. The data value of this sample is 0.6096.

[0216] It is clear that the essays that received low scores mainly contained many unreadable characters (???, etc.) and grammatical errors. For example, in sample 2... Figure 3 As shown, Figure 3 The lowest-valued sample information in prompt2, with a score of approximately 0.6149.

[0217] The main reasons for the low score of this sample are: logical jumps in the content, low relevance between personal experience and topic, lengthy expression without highlighting the key points, inclusion of an unreadable sample from the previous sample, and the content information of this sample was also taken into account. Secondly, the low original score is also one of the reasons for the low data value.

[0218] Data augmentation after data augmentation is more effective, increasing the amount of high-value text and improving the training results. For example... Figure 4 As shown, the data augmentation did not modify the original text semantics, but rather modified the text's logical structure and word choice. The data was expanded without breaking the original text, thereby improving the training effect. The above examples all demonstrate the analysis of how our proposed method improves experimental results.

[0219] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A cross-prompt automatic essay scoring method based on data evaluation and meta-learning, characterized in that, The method comprises the following steps: obtain source prompt word data; value assessment is performed on the source prompt word data to obtain a value probability of a prompt word sample; a target value sample set is obtained according to the value probability; synonym conversion enhancement is performed on the target value sample set to obtain an enhanced sample set; processing is performed on the enhanced sample set to obtain reinforcement training data; a meta-learning cross-prompt automatic composition scoring model is trained based on the reinforcement training data to obtain a target scoring model, wherein the target scoring model is used for automatically scoring compositions.

2. The cross-prompt automatic essay scoring method based on data evaluation and meta-learning according to claim 1, wherein, The value assessment performed on the source prompt word data to obtain the value probability of the prompt word sample comprises the following steps: S1. Calculate an initial value probability according to the semantic embedding and the true score of the source prompt word data by using a data value estimator, wherein the data value estimator is composed of a multi-layer perceptron and a Sigmoid activation function; S2. Generate a sample selection indicator by Bernoulli sampling according to the initial value probability; S3. Construct a weighted training loss according to the sample selection indicator, train a predictor to a convergent state by using the weighted training loss, and obtain current predictor parameters; S4. Calculate the gradient by using the REINFORCE algorithm based on the current predictor parameters, and update the data value estimator parameters; S5. Repeat S1-S4 until the data value estimator converges, and obtain the value probability of the prompt word sample.

3. The cross-prompt automatic essay scoring method based on data evaluation and meta-learning according to claim 1, wherein, The target value sample set is obtained according to the value probability by using a clustering algorithm. The synonym conversion enhancement performed on the target value sample set to obtain the enhanced sample set comprises the following steps:

4. The cross-prompt automatic essay scoring method based on data evaluation and meta-learning according to claim 1, wherein, According to each sample in the target value sample set, perform synonym conversion by using a large language model to generate new samples with different expressions while retaining the scoring labels, wherein the scoring labels include core semantics, writing style and scoring dimensions; The enhanced sample set is constructed according to the new samples. The processing performed on the enhanced sample set to obtain the reinforcement training data comprises the following steps:

5. The cross-prompt automatic essay scoring method based on data evaluation and meta-learning according to claim 1, wherein, Perform distribution correlation test on the enhanced sample set and the target prompt test data in the original data to obtain a distribution consistency index; According to the distribution consistency index, filter the enhanced samples that meet the distribution requirements to obtain a filtering result; The filtering result is subjected to diversity optimization to obtain the reinforcement training data. The distribution correlation test performed on the enhanced sample set and the target prompt test data to obtain the distribution consistency index comprises the following steps:

6. The cross-prompt automatic essay scoring method based on data evaluation and meta-learning according to claim 5, characterized in that, According to the enhanced sample set and the target prompt test data, sample features are extracted and mapped to a reproducing Hilbert space respectively to obtain target sample features; Based on the target sample features, the similarity between sample pairs is calculated to obtain a kernel function value matrix; According to the kernel function value matrix, the kernel function mean values of sample pairs within the enhanced sample set, within the target prompt test data and across sets are calculated; The maximum mean difference is calculated by combining the kernel function mean values to obtain the distribution consistency index. The diversity optimization performed on the filtering result to obtain the reinforcement training data comprises the following steps:

7. The cross-prompt automatic essay scoring method based on data evaluation and meta-learning according to claim 5, characterized in that, ​ According to the screening result, a semantic feature vector and a structure feature vector are extracted; According to the semantic feature vector and the structure feature vector, a modified cosine distance between samples is calculated to obtain a distance matrix; According to the distance matrix, a semi-positive definite kernel matrix is constructed, wherein a kernel matrix element is obtained by converting a distance value through an exponential function; According to the kernel matrix, a sample that maximizes a subset determinant increment is selected through a greedy approximation algorithm iteration until a preset number is reached, and the reinforced training data is obtained.

8. The cross-prompt automatic essay scoring method based on data evaluation and meta-learning according to claim 1, wherein, Based on the reinforced training data, a meta-learning cross-prompt automatic composition scoring model training is performed to obtain the target scoring model, including: Based on the reinforced training data, a virtual support set and a virtual query set are constructed, and real unseen prompt data is introduced to obtain a meta-learning training task; According to the meta-learning training task, a candidate meta-learner parameter set is generated by gradient descent update; According to the candidate meta-learner parameter set and the feature distribution of the real unseen prompt data, a distribution distance is calculated by maximum mean difference; According to the distribution distance, the optimal candidate meta-learner parameter is screened; According to the optimal candidate meta-learner parameter, the global parameters of the model are optimized through a joint loss function, wherein the joint loss includes a meta-training loss of the virtual support set, a meta-test loss of the virtual query set, and a dynamic offset allowance based on the distribution distance; According to the optimized model global parameters, the target scoring model is obtained.

Images