Weakly supervised analysis method and system for agricultural product repurchase intention based on multi-dimensional feature association

Abstract

The present invention belongs to the technical field of natural language processing, and specifically relates to a weakly supervised analysis method and system for agricultural product repurchase intention associated with multi-dimensional features. The method divides the agricultural product evaluation data set into a comment text list and a score list, mines out the factors affecting the repurchase intention and determines the fine-grained evaluation dimension; extracts triple information from the comment text, and processes it to obtain a fine-grained aspect item list and a fine-grained opinion item list; then generates a pseudo-label data set, and uses a pseudo-label correction strategy embedded with confidence learning and preliminary training of a multi-dimensional multi-label deep learning model to correct the pseudo-label training set; the contribution evaluation model quantifies the contribution of each repurchase intention influencing factor; and trains under a weakly supervised framework to mine the multi-dimensional emotions in the comment text. The present invention solves the problem of missing labels in each feature dimension, which helps to accurately grasp customer needs and further formulate targeted resource allocation strategies.

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing technology, and specifically relates to a weakly supervised analysis method and system for agricultural product repurchase intention associated with multi-dimensional features. Background Art

[0002] With the rapid development of e-commerce platforms and the advent of the big data era, online sales and purchases of agricultural products have become a widely accepted and increasingly popular transaction method. While this sales and purchase method provides convenience, it also raises several issues that require consideration and resolution. Consumer repurchase rates for agricultural products are often influenced by multiple factors, such as product quality, merchant service, and packaging appearance. Understanding customer needs and developing corresponding resource allocation strategies is crucial for merchants to increase product repurchase rates. Consumers also want to quickly select the agricultural products they desire from a wide variety of stores and products. To foster communication between merchants and customers, major e-commerce platforms have implemented review functions, allowing consumers to provide feedback on agricultural products and rate their shopping experience. Consumer ratings provide a direct reflection of their repurchase intention: higher scores indicate a stronger repurchase intention. The huge volume of reviews provides merchants with a more comprehensive and powerful basis for analyzing customer needs and consumer product selection. However, it is not easy for merchants to fully grasp the key factors that influence customers' willingness to repurchase. It is also difficult for customers to intuitively analyze how various products perform in the aspects they value. Traditional data analysis methods are often unable to quantify the content reflected in the review text. This is an urgent problem that needs to be solved for the development of agricultural product transactions on e-commerce platforms.

[0003] Against this backdrop, natural language processing tasks within machine learning and deep learning have been introduced to product review analysis, achieving remarkable results in recent years. Deep learning models can learn the patterns between semantic information and sentiment from large amounts of sentimentally labeled text data. These models can then be applied to unlabeled text to help people efficiently and intuitively analyze the emotional information contained within. By combining relevant corpora with sentiment analysis tools that offer capabilities such as part-of-speech tagging and sentiment intensity sensing, factors within reviews that correlate with repurchase intention can be captured.

[0004] However, the results obtained from this approach do not directly reflect the overall impact of various factors on repurchase intention, making it difficult for merchants to formulate reasonable resource allocation strategies. Furthermore, agricultural product evaluation datasets obtained from e-commerce platforms only include overall sentiment scores, not specific scores for each dimension. This lack of fine-grained labels hinders quantifying the contribution of individual influencing factors and building multi-dimensional, multi-label deep learning models. Summary of the Invention

[0005] Based on this, the present invention proposes a weakly supervised analysis method and system for agricultural product repurchase intention based on multi-dimensional feature association. A topic model is used to model short texts to capture the factors affecting repurchase intention and determine fine-grained evaluation dimensions accordingly. The sentiment triples in the review text are extracted to generate pseudo-labels on various coarse and fine granularity dimensions. A multi-task deep learning model is constructed and combined with confidence learning to correct the pseudo-labels to obtain multi-dimensional pseudo-labels with relatively high credibility. A global interpretation method based on game theory is applied to the machine model to construct a global contribution evaluation model, obtaining a quantitative representation of the contribution of each factor to repurchase intention. A multi-dimensional, multi-label deep learning model is then trained under a weakly supervised framework to mine fine-grained sentiment tendencies in the review text.

[0006] The present invention is achieved through the following technical solutions.

[0007] A weakly supervised analysis method for agricultural product repurchase intention based on multi-dimensional feature correlation is proposed. The steps are as follows:

[0008] Step 1: Obtain review texts and ratings from e-commerce platforms to construct an agricultural product evaluation dataset;

[0009] Step 2: Divide the agricultural product evaluation dataset into a review text list and a rating list. Create a review noun matrix and a noun dictionary based on the review text list. Then, convert the review text into a bag-of-words representation matrix based on the noun dictionary and review noun matrix. Then, perform topic modeling to obtain the topic-noun distribution, thereby mining the factors that influence repurchase intention and determining fine-grained evaluation dimensions.

[0010] Step 3: Use the aspect-sentiment triple extraction model based on the span-level interactive learning mechanism to extract triple information from the review text, including aspect terms, opinion terms, and sentiment polarity;

[0011] Step 4: Remove duplicates from the extracted aspect items and opinion items and classify them separately to obtain a list of fine-grained aspect items and a list of fine-grained opinion items;

[0012] Step 5: Generate a pseudo-labeled dataset based on the extracted triple information, fine-grained aspect item list, and fine-grained opinion item list. Extract data and manually correct it as a clean test set, and use the rest as a pseudo-labeled training set.

[0013] Step 6: Build a multi-dimensional multi-label deep learning model based on convolutional neural networks and Transformer, and use a pseudo-label correction strategy that embeds confidence learning and preliminary training of the multi-dimensional multi-label deep learning model to correct the pseudo-label training set;

[0014] Step 7: Obtain the dataset required for the contribution evaluation model, use a histogram-based gradient boosting classifier, and further apply a feature contribution estimation method based on game theory to the histogram gradient boosting classifier. Use the contribution evaluation model to quantify the contribution of each repurchase intention influencing factor.

[0015] Step 8: Input the corrected pseudo-label dataset and train a multi-dimensional multi-label deep learning model under a weakly supervised framework to mine the multi-dimensional emotions in the review text.

[0016] Further preferably, in step 2, a latent Dirichlet distribution model is used for topic modeling.

[0017] Further preferably, the process of extracting triplet information from the review text using the aspect sentiment triplet extraction model based on the span-level interactive learning mechanism in step 3 is as follows:

[0018] Traverse each comment text, perform word segmentation encoding, use the pre-trained BERT model to capture the contextual representation of the comment text and select the last layer shape as the hidden state. Loop through the model to generate all possible spans in the comment text;

[0019] Classify spans: obtain the probability distribution of all spans belonging to different span categories and predict whether the span is a valid aspect item or opinion item;

[0020] The number of aspect items and opinion item spans that need to be retained for a given review text is determined based on the sequence length and pruning threshold, and then a dual-channel pruning strategy is adopted to select the top 100 in the probability distribution of span categories. The span index that is most likely to be an aspect item and the span index that is most likely to be an opinion item;

[0021] Construct a joint representation matrix for each pair of aspect items and opinion items in the review text;

[0022] The sentiment polarity is judged on the joint representation matrix of aspect item-opinion item pairs, and the probability distribution of all aspect item-opinion item pairs belonging to different sentiment polarity categories is obtained.

[0023] Further optimization, in step four, first, the aspect items and opinion items are deduplicated separately; then, the aspect items containing the same root word are clustered into one category; for the opinion items, after removing the adverbs indicating the function, word form restoration is performed and the items containing the same root word are classified into one category; further manual adjustment and screening are performed to generate a fine-grained aspect item list and a fine-grained opinion item list for each fine-grained evaluation dimension.

[0024] Further preferably, the specific process of step five is as follows:

[0025] First, set initial labels for each fine-grained evaluation dimension of each review;

[0026] Then define three types of pseudo labels and sentiment polarity-pseudo label conversion functions to convert sentiment polarity into corresponding pseudo labels;

[0027] Then, the fine-grained initial labels corresponding to all comment texts are modified based on the pseudo-label judgment method;

[0028] Finally, pseudo labels on coarse-grained dimensions are generated based on the triple information to obtain a pseudo label dataset;

[0029] Part of the data is extracted from the pseudo-label dataset and manually corrected as a pure test set, and the remaining data is used as a pseudo-label training set.

[0030] Further preferably, the specific process of step six is ​​as follows:

[0031] The pseudo-label training set is passed in, and a text classifier based on the Fasttext model is trained for each dimension separately. Confidence learning tools are then used to identify low-confidence samples.

[0032] A multi-dimensional and multi-label deep learning model based on convolutional neural networks and Transformer is constructed. Data is extracted from the pseudo-label training set for manual annotation and passed into the constructed multi-dimensional and multi-label deep learning model for training. A preliminary model is obtained based on the accuracy of the training set and the preliminary model is used to predict the low-confidence labels of low-confidence samples to obtain the model labels; the model labels of the low-confidence samples are then compared with the true labels recognized by the text classifier based on the Fasttext model for the low-confidence samples. It is stipulated that: if the model label is consistent with the true label recognized by the text classifier based on the Fasttext model, the corrected label is the model label. If the two are inconsistent, they are manually corrected; the samples corresponding to the corrected labels and the high-confidence samples together constitute the corrected pseudo-label training set.

[0033] Further preferably, the specific process of step seven is as follows:

[0034] Based on the rating lists, we generate labels for all review texts in the corrected pseudo-label training set and the clean test set on the repurchase intention dimension. We then transform the pseudo-labels on all fine-grained evaluation dimensions in the corrected pseudo-label training set and the clean test set to meet the data requirements of the contribution evaluation model. We then concatenate the datasets that form the contribution evaluation model and divide them into the training and test sets for the contribution evaluation model.

[0035] Build a contribution evaluation model, pass in the training set of the contribution evaluation model, use the pseudo labels on the fine-grained dimension as input features, use the label of the repurchase intention dimension as the target label, traverse the hyperparameter grid and evaluate the performance of the gradient boosting classifier under each parameter combination through five-fold cross-validation, and select the hyperparameter combination that makes the gradient boosting classifier perform best on multiple data subsets;

[0036] The training set is then fed back into the contribution evaluation model to quantify the contribution of each factor influencing repurchase intention.

[0037] Further preferably, in step eight, the manually corrected clean test set and the corrected pseudo-label training set together constitute a corrected pseudo-label dataset, which is used as the dataset of the multi-dimensional multi-label deep learning model.

[0038] The present invention also provides a weakly supervised analysis system for agricultural product repurchase intention associated with multi-dimensional features, which includes a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the various steps of the above-mentioned weakly supervised analysis method for agricultural product repurchase intention associated with multi-dimensional features.

[0039] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the various steps of the above-mentioned weak supervision analysis method of agricultural product repurchase intention associated with multi-dimensional features.

[0040] The present invention combines four innovative strategies: a fine-grained pseudo-label generation method based on a cross-level sentiment triplet extraction model, a pseudo-label correction strategy embedded with confidence learning and preliminary training of a deep model, a contribution evaluation model with a global perspective, and a multi-dimensional multi-label deep learning model construction method under a weak supervision framework. These strategies solve the problem of missing labels in each feature dimension, and by mining and quantifying the contribution of each feature dimension in the global context, help merchants to efficiently and accurately grasp customer needs and further formulate targeted resource allocation strategies. The constructed multi-dimensional multi-label deep learning model helps consumers quickly select products by capturing the fine-grained sentiment tendencies in massive review texts. It also helps merchants discover potential problems in the product sales process, thereby promoting the development of e-commerce platforms in agricultural product transactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of a weakly supervised analysis method for agricultural product repurchase intention based on multi-dimensional feature association provided by the present invention;

[0042] Figure 2 This is a graph showing the overall label distribution of the crawled agricultural product evaluation dataset;

[0043] Figure 3 This is the overall framework diagram of the method;

[0044] Figure 4 It is a custom pseudo-label judgment method flowchart;

[0045] Figure 5 It is a quantitative result chart of the contribution of repurchase intention factors. DETAILED DESCRIPTION

[0046] The present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0047] Reference Figure 1 , a weakly supervised analysis method for agricultural product repurchase intention based on multi-dimensional feature association, the steps are as follows:

[0048] Step 1: Obtain review texts and ratings from e-commerce platforms to construct an agricultural product evaluation dataset;

[0049] Step 2: Divide the agricultural product evaluation dataset into a review text list and a rating list. Create a review noun matrix and a noun dictionary based on the review text list. Then, convert the review text into a bag-of-words representation matrix based on the noun dictionary and review noun matrix. Then, perform topic modeling to obtain the topic-noun distribution, thereby mining the factors that influence repurchase intention and determining fine-grained evaluation dimensions.

[0050] Step 3: Use the aspect-sentiment triple extraction model based on the span-level interactive learning mechanism to extract triple information from the review text, including aspect terms, opinion terms, and sentiment polarity;

[0051] Step 4: Remove duplicates from the extracted aspect items and opinion items and classify them separately to obtain a list of fine-grained aspect items and a list of fine-grained opinion items;

[0052] Step 5: Generate a pseudo-labeled dataset based on the extracted triple information, fine-grained aspect item list, and fine-grained opinion item list. Extract data and manually correct it as a clean test set, and use the rest as a pseudo-labeled training set.

[0053] Step 6: Build a multi-dimensional multi-label deep learning model based on convolutional neural network (CNN) and Transformer, and use the pseudo-label correction strategy embedded in confidence learning and preliminary training of the multi-dimensional multi-label deep learning model to correct the pseudo-label training set;

[0054] Step 7: Obtain the dataset required for the contribution evaluation model and use the histogram-based gradient boosting classifier HistGradientBoostingClassifier. Furthermore, we apply a game-theoretic feature contribution estimation method to the histogram gradient boosting classifier and use the contribution evaluation model to quantify the contribution of each repurchase intention influencing factor.

[0055] Step 8: Input the corrected pseudo-label dataset and train a multi-dimensional multi-label deep learning model under a weakly supervised framework to mine the multi-dimensional emotions in the review text.

[0056] In step 1 of this embodiment, automated crawler technology is used to obtain evaluation data of three types of agricultural products from multiple stores on three major e-commerce platforms to construct an agricultural product evaluation dataset (agricultural products include tomatoes, apples, and watermelons). The evaluation data consists of two parts: comment text and ratings. The ratings of all reviews are marked with five emotional levels from 1 to 5. The rating distribution is as follows: Figure 2 The original data set obtained first needs to be preprocessed, including removing duplicate data, removing empty data, and using regular expressions to remove special symbols contained in the comment text to meet the text requirements for data analysis.

[0057] In step 2 of this embodiment, the agricultural product evaluation dataset is divided into a comment text list and rating list ,in Respectively represent Comment text, Respectively represent The ratings corresponding to the review texts are then processed using the word segmentation tool and part-of-speech tagging tool of the nltk library to create a review noun matrix ,in Indicates the Comment text nouns and create a noun dictionary from them After filtering low-frequency and high-frequency words, all comment texts are converted into a bag-of-words representation matrix based on the noun dictionary and comment noun matrix. , the bag of words represents each element in the matrix Indicates the Comment text The number of times a noun appears, , Indicates the Nouns in the dictionary The corresponding mapping in Represents the frequency of occurrence. The Latent Dirichlet Allocation (LDA) model is then used for topic modeling. To improve the robustness of the LDA model to sparse data, the hyperparameters of the LDA model are optimized to adapt to the characteristics of short texts and improve the accuracy of topic modeling. Under each hyperparameter combination, the optimal number of topics under the hyperparameter configuration is determined by finding the maximum value of the LDA model consistency under all topic number settings. The calculation process is shown in Formula (1):

[0058] (1);

[0059] in Indicates in The number of topics, including Theme Number of review texts containing nouns; Indicates the Other nouns in the subject and The number of times a noun appears in the same comment text, is to adjust the parameters to avoid or When it is 0, the logarithmic function is undefined. Indicates the The number of nouns contained in the subject, Indicates the The total number of topics under the number of topics setting is obtained, and the LDA model is trained accordingly to obtain the topic-noun distribution under this hyperparameter configuration. Finally, the topic-noun distribution of the optimal LDA model is obtained according to the maximum consistency score under all parameter combinations, thereby excavating the factors affecting repurchase intention and determining the fine-grained evaluation dimension accordingly.

[0060] The process of extracting triple information from the review text using the aspect sentiment triple extraction model based on the span-level interactive learning mechanism described in step 3 of this embodiment is as follows:

[0061] Traverse each comment text, perform word segmentation encoding and other processing to make it meet the input requirements of the BERT model, use the pre-trained BERT model to capture the context representation of the comment text and select the last layer shape The hidden state of Indicates the length of the encoded sequence, Represents the hidden layer dimension of the BERT model, ranging from 1 to Loop through and generate all possible spans in the comment text, which consists of a shape The span represents the tensor and a list of span index pairs Record span information, where N represents the total number of spans generated, is the dimension of the width embedding vector. Each span is represented by the concatenation of the tokens representing its start and end positions and a width embedding vector representing the span width. The span index is the element in the list. Respectively represent The position of the span in the original sequence of the comment text, , indicating the Span index information is a list consisting of two elements, Indicates the starting position index of the i-th span in the original sequence of the comment text, Indicates the ending position index of the i-th span in the original sequence of the comment text.

[0062] Next, we classify spans: spans represent tensors A feedforward neural network consisting of multiple linear layers and activation functions is passed in. The input feature dimension of the feedforward network is , after three layers of linear transformation, it will eventually be mapped to three feature dimensions, representing "invalid span", "aspect span" and "opinion span", thereby obtaining the span representation tensor after the output transformation , the specific process is shown in formula (2):

[0063] (2);

[0064] in Respectively represent The weight matrix of the linear layer, Respectively represent The bias vector of the linear layer, ReLU and softmax are the activation functions, and then normalization is used to obtain the probability distribution of all spans belonging to different span categories , predict whether the span is a valid aspect item or opinion item, as shown in formula (3):

[0065] (3);

[0066] in represents the probability of the i-th span in the c-th span category, , They represent the original scores of the i-th span in the c-th span category and the k-th span category after linear transformation, Indicates the number of span categories.

[0067] Then according to the sequence length and pruning threshold Determine the number of aspect and opinion spans to retain for a given review , and then adopt a dual-channel pruning strategy to select the span category probability distribution Center front The span index that is most likely to be a facet item and the span index that is most likely to be an opinion item ,in Respectively represent the aspects of spans in the span representation tensor The position index in , Respectively indicate the opinions of the spans in the span representation tensor The position index in .

[0068] Next, a joint representation matrix is ​​constructed for each pair of aspect items and opinion items in the review text. , , joint representation matrix Each element in Indicates that in the retained span, the Aspects and The joint expression of opinion items; Indicates the The terms in the span represent tensors The corresponding embedding representation in is, Indicates the Opinion items in The corresponding embedding representation in is, It contains Aspects and The embedding representation of the relative position information between opinion items is generated as shown in formula (4):

[0069] (4);

[0070] in and Respectively represent The starting and ending positions of each aspect item in the original sequence of the comment text, and Respectively represent The starting position and ending position of each opinion item in the original sequence of the comment text are taken as the minimum value and then the The operation maps it into a trainable embedding vector and returns a list of original position indices of candidate aspect item-opinion item pairs. , used to record the positions of aspect items and opinion items in all candidate aspect item-opinion item pairs in the original sequence of the comment text, Contains Aspects and The index of the starting and ending positions of each opinion item in the original sequence of the comment text, .

[0071] Then the joint representation matrix of aspect item-opinion item pair The input feature dimension of the feedforward neural network is similar to the previous one for sentiment polarity judgment. ,in is the distance embedding vector dimension, which is transformed into three layers of linear transformation and activation function After activation, it is mapped to four feature dimensions, representing the sentiment polarity of "invalid", "positive", "negative", and "neutral". After normalization, the probability distribution of all aspect item-opinion item pairs belonging to different sentiment polarity categories is finally obtained. : ;

[0072] in Each element in Respectively represent The aspect item-opinion item pair belongs to The probability value of the sentiment polarity category, , M is the number of aspect item-opinion item pairs.

[0073] Based on the original position index list of candidate aspect-opinion pairs and the probability distribution of sentiment polarity categories, we retrieve the text and sentiment polarity corresponding to each valid aspect-opinion pair in the original sentence. This completes the extraction of (aspect, opinion, sentiment polarity) triples for all agricultural product reviews.

[0074] In step 4 of this embodiment, the aspect items and opinion items are first deduplicated. Then, for aspect items, the aspect items containing the word roots "apple", "watermelon", and "tomato" are divided into three categories, and the remaining aspect items containing the same word roots are grouped into one category; for opinion items, after removing adverbs that express negation, modification, etc., word form restoration is performed and the word roots containing the same word roots are grouped into one category. After further manual adjustment and screening, the first A list of fine-grained aspect items for each fine-grained evaluation dimension and a list of fine-grained opinion items . Label the list consisting of all fine-grained aspect items as List 1, and the list consisting of all fine-grained opinion items as List 2.

[0075] The process of generating a pseudo-labeled dataset, a clean test set, and a pseudo-labeled training set based on the extracted triple information, the fine-grained aspect item list, and the fine-grained opinion item list in step 5 of this embodiment is as follows:

[0076] First, an initial label is set for each fine-grained evaluation dimension of each review, with the default value being 2.

[0077] Then define three types of pseudo labels (1 for positive, 0 for negative, and 2 for neutral or not mentioned) and a sentiment polarity-pseudo label conversion function to convert the sentiment polarities 'POS', 'NEG', and 'NEU' into corresponding pseudo labels.

[0078] Then, according to Figure 4 The custom pseudo-label judgment method shown modifies the fine-grained initial labels corresponding to all comment texts: traverse all comment texts, and for each comment text, traverse its corresponding triple list to process the triple information one by one. If the aspect item in a certain triple is in List 1, further determine which fine-grained evaluation dimension the aspect item represents. If the pseudo-label on this dimension is neither equal to 2 nor equal to the pseudo-label corresponding to the sentiment polarity, then set the pseudo-label of this dimension to 2, otherwise set it to the pseudo-label corresponding to the sentiment polarity. If the aspect item is not in List 1, check the opinion item. If the opinion item is in List 2, further determine which fine-grained evaluation dimension the opinion item represents. Similar to the aspect item, modify the pseudo-label of this dimension. If the aspect item of the triple is not in List 1 and the opinion is not in List 2, process the next tuple.

[0079] Finally, based on the triple information, we generate pseudo labels on a coarse-grained dimension: If the triple list corresponding to a review text contains only triplets indicating positive relationships, the pseudo label for that text on the coarse-grained dimension is 1; if it contains only triplets indicating negative relationships, the pseudo label on the coarse-grained dimension is 0; if there are triplets indicating both negative and positive relationships, the pseudo label on the coarse-grained dimension is 2, reflecting the customer's gambling mentality towards repurchasing products. This results in a pseudo-labeled dataset.

[0080] 525 data were extracted from the pseudo-labeled dataset and manually corrected as a clean test set. .in represents the pure test set comment text, Represents the clean test set comment text in clean labels on fine-grained dimensions, Represents the clean labels of the clean test set review text on a coarse-grained dimension. The remaining data serves as the pseudo-label training set.

[0081] The process of step six of this embodiment is as follows:

[0082] First, find the coarse and fine granular pseudo labels that need to be corrected. Enter the pseudo label training set, use the text classifier FirstColFtClassifier based on the Fasttext model to train each dimension separately, and then use the confidence learning tool Cleanlab to find low confidence samples. For each dimension, the comment text is used as the feature part of the training data, and the pseudo label on this dimension is used as the label part of the training data. After word mapping, the vector representation of all comment texts is obtained. ,in Respectively represent The vector representation of the comment text is obtained, and then the label category probability distribution of all comment texts in this fine-grained dimension is obtained through formula (5):

[0083] (5);

[0084] in Indicates the Comment texts are predicted to belong to The probability value of the class label, and Respectively , The first in the weight matrix of the class label elements, Indicates the The first comment in the text vector elements, and Respectively represent The bias term for the class label, is the number of label categories. Then, the confidence learning tool cleanlab constructs the joint distribution matrix of predicted labels and pseudo labels based on the obtained label category probability distribution , the joint distribution matrix Each element in Indicates that the given pseudo label category is Under the class label, the true label category is identified as The probability value of the class label is calculated as shown in formula (6).

[0085] (6);

[0086] in, represents the true label category, represents the pseudo label category, Indicates that the true label category is class labels, Indicates that the pseudo label category is Class labels, Indicates the The true label category of the samples, Indicates the The pseudo-label category given by the samples, represents the indicator function, is the total number of comment text samples, and the label quality score (between 0 and 1) of each sample is calculated based on this. The closer the quality score is to 0, the more likely the label is a noise label and the lower the credibility. The confidence learning tool CleanLab will select a threshold based on the quality score distribution to distinguish low-confidence labels from high-confidence labels. The samples corresponding to low-confidence labels and high-confidence labels are low-confidence samples and high-confidence samples, respectively.

[0087] Next, low-confidence labels are corrected and filtered. A multi-dimensional, multi-label deep learning model based on a convolutional neural network (CNN) and Transformer is constructed. This model consists of an embedding layer, a positional encoding layer, a convolutional layer, a Transformer encoder, and a fully connected layer. It aims to map the input sequence into multiple class predictions along multiple evaluation dimensions. 105 datasets were manually annotated from the pseudo-label training set and then trained on the constructed multi-dimensional, multi-label deep learning model. A preliminary model was generated based on the accuracy of the training set and used to predict low-confidence labels for low-confidence samples to obtain model labels. The model labels of low-confidence samples were then compared with the true labels determined by FirstColFtClassifier for the low-confidence samples. If the model labels match the true labels determined by FirstColFtClassifier, the corrected labels are the model labels. If they differ, manual correction is performed. In addition, there are inevitably vague labels in the evaluation dataset. Such labels may cause confusion during model training and affect the performance and generalization ability of multi-dimensional multi-label deep learning models. Therefore, it is necessary to remove some data with unclear label categories to improve the quality of the dataset and the robustness of the model, thereby obtaining the final corrected labels. The samples corresponding to the corrected labels and the high-confidence samples together constitute the corrected pseudo-label training set. ,in represents the pseudo-label training set comment text after correction, Respectively represent the pseudo-label training set comment text after correction in Pseudo labels on fine-grained dimensions, Represents the pseudo-labels of the review texts in the coarse-grained dimension of the pseudo-label training set after correction.

[0088] The process of step seven of this embodiment is as follows:

[0089] First, obtain the dataset required for the contribution evaluation model. Based on the score list, generate the labels of all review texts in the repurchase intention dimension in the corrected pseudo-label training set and the clean test set: Figure 3 As shown in the figure, if the score is 1 or 2, the label on the repurchase intention dimension is set to 0, if the score is 3 or 4, it is set to 1, and if the score is 5, it is set to 2. Then, the pseudo labels on all fine-grained evaluation dimensions in the corrected pseudo-label training set and the clean test set are transformed to meet the data requirements of the contribution evaluation model: pseudo labels indicating positive are transformed to 2, and pseudo labels indicating neutral or unmentioned are transformed to 1. The datasets that form the contribution evaluation model are spliced ​​together. .in Represents the pseudo-label training set and the clean test set after correction in The pseudo labels after transformation on a fine-grained dimension, Represents the labels on the repurchase intention dimension, thereby obtaining the training set and test set of the contribution evaluation model.

[0090] Next, a contribution assessment model was constructed. The model's training set was applied to the model for parameter optimization, and its performance was verified on the model's test set. A histogram-based gradient boosting classifier (HistGradientBoostingClassifier) ​​was employed, with various settings for the learning rate, L2 regularization strength, and maximum number of iterations. The model's training set was applied, with pseudo labels on the fine-grained dimension as input features and labels on the repurchase intention dimension as target labels. The model's performance was evaluated for each parameter combination using a five-fold cross-validation grid. The hyperparameter combination that yielded the best performance across multiple data subsets was selected. To further validate the reliability of the gradient boosting classifier, the trained gradient boosting classifier was then applied to the model's test set to predict repurchase intention. The classifier achieved an accuracy of 85.42%, demonstrating that it met the requirements for further development of the contribution assessment model.

[0091] Finally, the training set is fed back into the contribution assessment model to quantify the contribution of each repurchase intention influencing factor (each fine-grained evaluation dimension). The entire process will undergo a five-fold cross-validation to avoid misjudgments due to overfitting or underfitting of a single model and calculate the contribution value of each factor. During this process, the dataset will be divided into five subsets, with four subsets used in turn for training under the optimal hyperparameter combination, and the remaining subset used for validation. In each validation fold, the accuracy of the repurchase intention prediction on the current validation subset will be calculated. Then, by masking the features and observing the changes in model performance, we estimate the contribution of various factors affecting repurchase intention on the validation subset. ,in Respectively expressed in Fold the first The contribution value of the factors influencing repurchase intention is shown in formula (7):

[0092] (7);

[0093] in, Indicates in Fold the first The contribution value of the factors influencing repurchase intention, Indicates that all the covered The feature subset of the factors affecting repurchase intention is classified, and the cross entropy loss function is used to obtain the performance quantitative representation of a feature subset. And add the first Quantitative representation of the performance of factors influencing repurchase intention after characterization , by observing the difference, we can get The marginal contribution of class features to the performance of the contribution evaluation model, It is Characteristics of factors influencing repurchase intention The corresponding weight value, and Represents all feature numbers and The number of features in .

[0094] After the five-fold cross validation is completed, the validation set accuracy matrix is ​​constructed and contribution matrix ,in Respectively represent The accuracy of the validation subset, Represents the factors affecting repurchase intention in the first The contribution matrix on the validation subset is then normalized to obtain the matrix ,in, Respectively represent The relative importance value of the validation subset. Finally, the contribution value is weighted averaged and normalized to obtain the global feature contribution vector ,in Respectively represent The contribution of the factors influencing repurchase intention to consumers’ repurchase intention is quantitatively expressed. Each element in the vector The calculation process is shown in formula (8):

[0095] (8);

[0096] in, represents the relative importance value of the i-th validation subset, and Respectively expressed in On the validation subset The contribution value of factors influencing repurchase intention.

[0097] According to Figure 5 The quantitative analysis results shown in the figure give resource allocation suggestions to merchants based on the contribution value: Figure 5 As shown, there are three types of factors with contribution values ​​greater than the average, namely "taste", "freshness" and "appearance". Among them, the contribution value of "taste" is the largest, much higher than other topics. It can be seen that the taste of agricultural products is the most important factor affecting customers' repurchase intention. Agricultural product suppliers should focus their resources on improving the taste of agricultural products to increase customer repurchase rates. In second and third place are freshness and product appearance factors, respectively. Their contribution values ​​are slightly lower than the taste factor. These are also the two aspects that customers care about more. It is recommended that agricultural product sellers increase resource investment in transportation and improved preservation technology to ensure the freshness of agricultural products. At the same time, certain resources should also be allocated to optimizing the appearance of products. Secondly, in fourth place is the "texture" factor. It can be seen that the public is generally more tolerant of the product's powdery and glutinous texture or crispy texture, and whether the watermelon has seeds or no seeds. Although this factor does not play a decisive role in the repurchase intention, merchants should also maintain their advantages in this area and further optimize it under the premise of ensuring the "taste", "freshness" and "appearance" of the product. From Figure 5 It can be seen that the comments rarely involve opinions on the prices of agricultural products. Merchants may raise prices appropriately and set reasonable prices by observing changes in repurchase rates. Although the merchant's service attitude does not have a decisive impact on customers' repurchase intention, if resources permit, they can also consider improving customers' shopping experience and the convenience of delivery services. Otherwise, more resources can be invested in more important aspects.

[0098] In step eight of this embodiment, the manually corrected clean test set and the corrected pseudo-label training set together constitute the corrected pseudo-label dataset, which is used as the dataset for the multi-dimensional multi-label deep learning model;

[0099] After each comment text is segmented, encoded, and padded, it is fed into a multi-dimensional, multi-label deep learning model along with a mask that marks the padded position.

[0100] First, the embedding layer maps each word index in the text into a high-dimensional vector. The positional encoding layer then adds position information to each embedding vector, explicitly introducing position information in the sequence and helping multi-dimensional, multi-label deep learning models capture sequential relationships in the sequence.

[0101] Next, the model's convolutional layer uses one-dimensional convolution operations to capture local features in the input sequence. The ReLU activation function introduces nonlinearity, allowing the neural network to learn more complex, nonlinear features. Batch normalization is then performed on the convolutional features to facilitate model convergence and accelerate training. To prevent information loss after the convolution operation, the multi-dimensional, multi-label deep learning model saves the convolutional layer input information and adds it to the convolutional layer output to form a residual connection. This, along with a mask indicating the fill position, serves as the input to the transformer encoder.

[0102] The Transformer encoder consists of multiple Transformer Encoder layers, which process each time step in the text sequence in parallel. A multi-headed self-attention mechanism focuses on each position in the sequence to capture global dependencies, and a feedforward neural network (FFN) performs nonlinear transformations on the features at each position. This ultimately generates a hidden state that contains contextual information. The provided src_key_padding_mask mask causes the self-attention mechanism to ignore the padding in the sequence, preventing it from introducing noise into the calculation of the attention weights.

[0103] The hidden state at the last time step is extracted and passed to the shared connection layer, aiming to map the extracted features to multi-class labels across all fine-grained evaluation dimensions. Finally, a tensor operation is performed to present the prediction results as a three-dimensional tensor. Each element in the tensor represents the probability of a given text belonging to one of the three labels 0, 1, or 2 on a given evaluation dimension. Finally, the label with the maximum probability is taken as the model's prediction result.

[0104] During the training process, early stopping and dropout regularization techniques are used to prevent the model from overfitting. In order to enable the multi-dimensional multi-label deep learning model to be more finely adjusted in the later stage of training to find the optimal solution, the learning rate decay method is adopted. Due to the imbalance of the labels of the dataset used for the training of the multi-dimensional multi-label deep learning model, the cross entropy loss function is adopted and the loss weights of the three types of labels 0, 1, and 2 are set. During training, the mean loss of all dimensions is used as the loss backpropagation of a batch. Accuracy, precision, recall, and F1 value are used as evaluation indicators of the multi-dimensional multi-label deep learning model. The prediction effects of the overall and each fine-grained dimension are measured respectively, and the best multi-dimensional multi-label deep learning model is determined based on the comprehensive results of these four evaluation indicators on the entire test set. The accuracy measurement process is shown in formula (9):

[0105] (9);

[0106] in, Indicates the number of influencing factors, represents the number of test set samples, and Respectively represent The sample in The predicted results and true labels on the class influencing factors, Represents the indicator function.

[0107] The obtained optimal multi-dimensional multi-label deep learning model was further evaluated and compared with the four models. The results are shown in Table 1. The present invention has better performance.

[0108] Table 1

[0109]

[0110] Another embodiment of the present invention provides a weakly supervised analysis system for agricultural product repurchase intention associated with multi-dimensional features, the system comprising a memory, a processor, and a computer program stored in the memory, and when the computer program is executed by the processor, it implements the various steps of the weakly supervised analysis method for agricultural product repurchase intention associated with multi-dimensional features as described in the above embodiment.

[0111] Another embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the various steps of the weakly supervised analysis method of agricultural product repurchase intention associated with multi-dimensional features as described in the above embodiment.

[0112] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A weakly supervised analysis method for agricultural product repurchase intention based on multi-dimensional feature association, characterized in that: Here are the steps: Step 1: Obtain review texts and ratings from e-commerce platforms to construct an agricultural product evaluation dataset; Step 2: Divide the agricultural product evaluation dataset into a review text list and a rating list, create a review noun matrix and a noun dictionary based on the review text list, and then convert the review text into a bag-of-words representation matrix based on the noun dictionary and the review noun matrix; then perform topic modeling to obtain the topic-noun distribution, thereby mining the factors that affect repurchase intention and determining the fine-grained evaluation dimensions; Step 3: Use the aspect-sentiment triple extraction model based on span-level interactive learning mechanism to extract triple information from the review text, including aspect items, opinion items and sentiment polarity; Step 4: Remove duplicates from the extracted aspect items and opinion items and classify them separately to obtain a fine-grained aspect item list and a fine-grained opinion item list; Step 5: Generate a pseudo-labeled dataset based on the extracted triple information, fine-grained aspect item list, and fine-grained opinion item list. Extract data and manually correct it as a clean test set, and use the rest as a pseudo-labeled training set. Step 6: Build a multi-dimensional multi-label deep learning model based on convolutional neural network and Transformer, and use the pseudo-label correction strategy embedded with confidence learning and preliminary training of the multi-dimensional multi-label deep learning model to correct the pseudo-label training set; Step 7: Obtain the data set required for the contribution evaluation model, use the histogram-based gradient boosting classifier, and further apply the feature contribution estimation method based on game theory to the histogram gradient boosting classifier, and use the contribution evaluation model to quantify the contribution of each repurchase intention influencing factor; Step 8: Pass in the corrected pseudo-label dataset and train a multi-dimensional multi-label deep learning model under a weakly supervised framework to mine the multi-dimensional emotions in the comment text.

2. The weakly supervised analysis method of agricultural product repurchase intention based on multi-dimensional feature association according to claim 1 is characterized in that: In step 2, the latent Dirichlet allocation model is used for topic modeling.

3. The weakly supervised analysis method of agricultural product repurchase intention based on multi-dimensional feature association according to claim 1 is characterized in that: The process of extracting triple information from the review text using the aspect sentiment triple extraction model based on the span-level interactive learning mechanism described in step 3 is as follows: Traverse each comment text, perform word segmentation encoding, use the pre-trained BERT model to capture the contextual representation of the comment text and select the last layer shape as the hidden state, loop through, and generate all possible spans in the comment text; Classify spans: obtain the probability distribution of all spans belonging to different span categories and predict whether the span is a valid aspect item or opinion item; The number of aspect items and opinion item spans that need to be retained for a given review text is determined based on the sequence length and pruning threshold, and then a dual-channel pruning strategy is adopted to select the top 10 aspect items and opinion item spans in the probability distribution of span categories. The span index that is most likely to be an aspect item and the span index that is most likely to be an opinion item; Construct a joint representation matrix for each pair of aspect terms and opinion terms in the review text; The sentiment polarity is judged on the joint representation matrix of aspect item-opinion item pairs, and the probability distribution of all aspect item-opinion item pairs belonging to different sentiment polarity categories is obtained.

4. The weakly supervised analysis method of agricultural product repurchase intention based on multi-dimensional feature association according to claim 1 is characterized in that: In step 4, first, the aspect items and opinion items are deduplicated separately; then, the aspect items with the same root word are grouped into one category; for the opinion items, after removing the adverbs indicating the function, the word form is restored and the items with the same root word are grouped into one category; further, after manual adjustment and screening, the fine-grained aspect item list and fine-grained opinion item list of each fine-grained evaluation dimension are generated.

5. The weakly supervised analysis method of agricultural product repurchase intention based on multi-dimensional feature association according to claim 1 is characterized in that: The specific process of step five is as follows: First, set initial labels for each fine-grained evaluation dimension of each review; Then define three types of pseudo labels and sentiment polarity-pseudo label conversion functions to convert sentiment polarity into corresponding pseudo labels; Next, the fine-grained initial labels corresponding to all comment texts are modified based on the pseudo-label judgment method; Finally, pseudo labels on coarse-grained dimensions are generated based on the triple information to obtain a pseudo label dataset; Some data are extracted from the pseudo-label dataset and manually corrected as a pure test set, and the remaining data is used as a pseudo-label training set.

6. The weakly supervised analysis method of agricultural product repurchase intention based on multi-dimensional feature association according to claim 1 is characterized in that: The specific process of step six is ​​as follows: The pseudo-label training set is passed in, and the text classifier based on the Fasttext model is trained for each dimension separately, and then the confidence learning tool is used to find low-confidence samples; A multi-dimensional and multi-label deep learning model based on convolutional neural networks and Transformer is constructed. Data is extracted from the pseudo-label training set for manual annotation and passed into the constructed multi-dimensional and multi-label deep learning model for training. A preliminary model is obtained according to the accuracy of the training set and the preliminary model is used to predict the low-confidence labels of low-confidence samples to obtain the model labels. The model labels of the low-confidence samples are then compared with the true labels recognized by the text classifier based on the Fasttext model for the low-confidence samples. It is stipulated that: if the model label is consistent with the true label recognized by the text classifier based on the Fasttext model, the corrected label is the model label. If the two are inconsistent, they are corrected manually. The samples corresponding to the corrected labels and the high-confidence samples together constitute the corrected pseudo-label training set.

7. The weakly supervised analysis method of agricultural product repurchase intention based on multi-dimensional feature association according to claim 1 is characterized in that: The specific process of step seven is as follows: Generate labels for all review texts in the corrected pseudo-label training set and the clean test set on the repurchase intention dimension based on the score list; then transform the pseudo-labels on all fine-grained evaluation dimensions in the corrected pseudo-label training set and the clean test set to meet the data requirements of the contribution evaluation model; The data sets constituting the contribution evaluation model are concatenated and divided into a training set and a test set of the contribution evaluation model; Construct a contribution evaluation model, pass in the training set of the contribution evaluation model, use the pseudo labels on the fine-grained dimension as input features, use the labels on the repurchase intention dimension as target labels, traverse the hyperparameter grid and evaluate the performance of the gradient boosting classifier under each parameter combination through five-fold cross validation, and select the hyperparameter combination that makes the gradient boosting classifier perform best on multiple data subsets; The training set is then fed back into the contribution evaluation model to quantify the contribution of each factor influencing repurchase intention.

8. The weakly supervised analysis method of agricultural product repurchase intention based on multi-dimensional feature association according to claim 1 is characterized in that: In step eight, the manually corrected pure test set and the corrected pseudo-label training set together constitute the corrected pseudo-label dataset, which is used as the dataset for the multi-dimensional and multi-label deep learning model.

9. A weakly supervised analysis system for agricultural product repurchase intention based on multi-dimensional feature association, characterized in that: It includes a memory, a processor and a computer program stored in the memory, and when the computer program is executed by the processor, it implements the various steps of the weakly supervised analysis method of agricultural product repurchase intention associated with multi-dimensional features as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, each step of the weakly supervised analysis method of agricultural product repurchase intention associated with multi-dimensional features as described in any one of claims 1 to 8 is implemented.

Images