A Multimodal Sentiment Analysis-Based Intelligent Processing Method and System for E-commerce Reviews

Abstract

The present invention belongs to the technical field of computer information processing, and provides an intelligent processing method and system for e-commerce reviews with multi-modal sentiment analysis. The method is as follows: real-time collect and preprocess e-commerce review data to obtain the preprocessed review data; respectively process through a trained text sentiment analysis model and a trained image sentiment analysis model to obtain the text sentiment score and the image sentiment score of each review; dynamically calculate the text weight and the image weight, and calculate the comprehensive sentiment score of each review; determine whether the current review is an abnormal review and perform an abnormal review mark; based on the comprehensive sentiment score and / or the abnormal review mark, automatically trigger an e-commerce review response operation with cross-system linkage. The present invention can automatically process a large amount of e-commerce reviews, accurately identify the user sentiment conveyed in the reviews through multi-modal sentiment analysis, and can automatically perform cross-system linkage to achieve a processing closed-loop of front-end real-time analysis and back-end business execution.

Description

Technical Field

[0001] This invention relates to the field of computer information processing technology, and in particular to a method and system for intelligent processing of e-commerce reviews using multimodal sentiment analysis. Background Technology

[0002] With the rapid development of the internet, a large number of user reviews are generated on major e-commerce platforms. The emotional information contained in these reviews can not only help other users make better purchasing decisions, but also facilitate merchants in tracking and managing consumer feedback. Currently, e-commerce review processing methods typically rely on manual batch processing, which has the following drawbacks: First, the current processing methods for the massive amount of reviews on e-commerce platforms rely on manual labor, which is inefficient and cannot respond to customers in a timely manner. Second, existing processing methods lack automated multimodal sentiment analysis capabilities, making it impossible to fully understand the emotions expressed by users through images, thus leading to information omissions. Third, the sentiment analysis results of existing processing methods are isolated from the back-end business systems (such as customer service, promotion, and product systems), making it impossible to automatically trigger linked operations. This reliance on manual responses leads to low efficiency and high costs.

[0003] Therefore, we need to develop a multimodal sentiment analysis-based intelligent processing method and system for e-commerce reviews, capable of automatically processing massive amounts of e-commerce reviews. Through multimodal sentiment analysis, it can accurately identify the user emotions conveyed by the text or images in the reviews, and automatically link across systems to achieve a closed loop of real-time front-end analysis and back-end business execution. Summary of the Invention

[0004] The purpose of this invention is to provide a multimodal sentiment analysis-based intelligent processing method and system for e-commerce reviews, in order to solve the problems mentioned in the background art, such as low efficiency, inability to accurately understand user emotions in reviews, and inability to automatically trigger linked operations.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: According to one aspect of the present invention, a method for intelligent processing of e-commerce reviews based on multimodal sentiment analysis is provided, the method comprising the following steps: E-commerce review data is collected in real time, and the e-commerce review data is preprocessed to obtain preprocessed review data; wherein, the preprocessed review data includes text review data and / or image review data; Train a text sentiment analysis model, input the text comment data into the trained text sentiment analysis model for processing, and obtain the text sentiment score for each comment; Train an image sentiment analysis model, input the image comment data into the trained image sentiment analysis model for processing, and obtain the image sentiment score for each comment; Based on the validity of the preprocessed comment data, text weights and image weights are dynamically calculated, and a comprehensive sentiment score for each comment is calculated. For each comment in the preprocessed comment data, a sentiment consistency assessment is performed to determine whether the current comment is an anomalous comment, and the anomalous comments are marked. Based on the comprehensive sentiment score and / or abnormal comment markers, cross-system linkage e-commerce comment response operations are automatically triggered.

[0006] According to one aspect of the present invention, a multimodal sentiment analysis-based intelligent processing system for e-commerce reviews is provided. The system includes: a data preprocessing module, a text sentiment analysis module, an image sentiment analysis module, a comprehensive sentiment scoring module, an abnormal review marking module, and a business logic execution module; wherein: The aforementioned data preprocessing module is used to collect e-commerce review data in real time and preprocess the e-commerce review data to obtain preprocessed review data; wherein, the preprocessed review data includes text review data and / or image review data; The aforementioned text sentiment analysis module is used to train a text sentiment analysis model. The text comment data is input into the trained text sentiment analysis model for processing to obtain a text sentiment score for each comment. The aforementioned image sentiment analysis module is used to train an image sentiment analysis model. The image comment data is input into the trained image sentiment analysis model for processing to obtain an image sentiment score for each comment. The aforementioned comprehensive sentiment scoring module is used to dynamically calculate text weights and image weights based on the validity of the preprocessed comment data, and to calculate the comprehensive sentiment score for each comment. The aforementioned abnormal comment marking module is used to perform sentiment consistency assessment on each comment in the preprocessed comment data, determine whether the current comment is an abnormal comment, and mark the abnormal comment. The aforementioned business logic execution module is used to automatically trigger cross-system e-commerce comment response operations based on the comprehensive sentiment score and / or abnormal comment markers.

[0007] Based on the aforementioned scheme, the step of inputting the text comment data into the trained text sentiment analysis model for processing specifically involves: The trained text sentiment analysis model extracts product attributes from the text review data, predicts the text sentiment category for each product attribute, and outputs an initial text score vector for each text sentiment category. Apply the softmax function to the initial score vector of the text to obtain the confidence score for each sentiment category of the text; A sentiment score is predefined for each of the text sentiment categories, and a weighted sum is performed with the confidence score to obtain the text sentiment score.

[0008] Based on the aforementioned scheme, the step of inputting the image comment data into the trained image sentiment analysis model for processing specifically involves: The image comment data is input into a trained image sentiment analysis model to extract image feature vectors; The image feature vectors are mapped to the image sentiment category space, and the initial image score vector for each image sentiment category is output. Apply the softmax function to the initial score vector of the image to obtain the confidence score of each image sentiment category; A predefined sentiment score is defined for each of the image sentiment categories, and a weighted sum is performed with the confidence score to obtain the image sentiment score.

[0009] Based on the aforementioned scheme, the dynamic calculation of text weights and image weights specifically involves: Iterate through each comment in the preprocessed comment data; If the current comment contains only text comments, then the text weight is set to 1 and the image weight is set to 0. If the current comment only contains image comments, then the text weight is set to 0 and the image weight is set to 1. If the current comment contains both text and image comments, then the text contribution score and image contribution score for the current comment are calculated separately. Based on the proportion of the text contribution score and the image contribution score in the total score, the text weight and the image weight are calculated respectively; specifically, the following steps are included: The number of valid characters in the current comment is counted, and the number of valid characters is matched with a preset score to obtain the text contribution score; Filter the valid images in the current comments and select the first K valid images; Each of the first K images is converted to a grayscale image, and the Laplacian variance of each image is calculated. The Laplacian variance is compared with a preset level threshold to obtain the sharpness level of each image; The sharpness level is mapped to a preset sharpness coefficient to obtain the sharpness coefficient for each image; The base score for each image is set to a uniform fixed value, and the base score is weighted and summed with the sharpness coefficient corresponding to each image to obtain the image contribution score; The sum of the text contribution score and the image contribution score is used to calculate the total score. Based on the proportion of the text contribution score and the image contribution score in the total score, the text weight and the image weight are obtained respectively.

[0010] Based on the aforementioned scheme, the comprehensive sentiment score is calculated using the text weight, the text sentiment score, the image weight, and the image sentiment score. The specific calculation formula is as follows: S = Wt*St + Wi*Si; Where S is the comprehensive sentiment score, and S∈[0,1], 0 represents extremely positive sentiment, and 1 represents extremely negative sentiment; Wt is the text weight, St is the text sentiment score, Wi is the image weight, and Si is the image sentiment score.

[0011] Based on the aforementioned solution, the determination of whether the current comment is an abnormal comment specifically involves: Based on the text sentiment score and the image sentiment score, calculate the sentiment consistency score between the text comment and the image comment in the current comment. Compare the sentiment consistency score with a preset threshold. If the sentiment consistency score is lower than the preset threshold, the current comment is determined to be a normal comment; otherwise, it is an abnormal comment. Based on the trained text sentiment analysis model, the system identifies whether there are N different sentiment categories for the same product attribute in the current comment. If so, the current comment is determined to be an abnormal comment.

[0012] Based on the aforementioned solution, the automatic triggering of cross-system linkage e-commerce comment response operations specifically includes: Based on the comprehensive sentiment score, a priority initial set is selected. The weight coefficient and activity status parameters of the product corresponding to each comment in the priority initial set are obtained. The processing priority of each comment is calculated, and a high-priority processing work order is automatically generated by calling the customer service system API interface. Specifically, the automatic calling of the customer service system API interface based on the comment's processing priority is as follows: The system presets the initial weight coefficients for all products and periodically adjusts the weight coefficients of the products based on their sales revenue. Calculate the activity status parameters of the product based on the quantity of the product participating in the promotional activity; Set a first priority threshold, filter out comments whose comprehensive sentiment score is higher than the first priority threshold, and obtain an initial set for priority processing; For each comment in the initial priority processing set, obtain the weight coefficient and activity status parameters of the product corresponding to each comment, and calculate the processing priority of each comment; Set a second priority threshold. When the processing priority is not lower than the second priority threshold, automatically call the customer service system API interface to generate a high-priority processing work order. For all comments marked with the aforementioned abnormal comment tag, the customer service system API interface will be automatically invoked to generate a high-priority processing work order; Based on the comprehensive sentiment score and the abnormal comment marker, determine whether the current comment is a high-quality positive review or a genuine negative review, automatically call the promotion system API interface to issue rewards to users who published the high-quality positive review or users who published the genuine negative review, and call the customer service system API to generate personalized automated replies; Monitor the positive ratio of high-quality positive reviews and the negative ratio of genuine negative reviews within a set time window, and compare them with preset control thresholds respectively. Automatically call the supply chain system API interface to increase the inventory of the product or automatically call the product system API interface to remove the relevant product from the shelves. The criteria for determining high-quality positive reviews and genuine negative reviews are as follows: There are preset positive and negative thresholds. If the overall sentiment score of the current comment is lower than the preset positive threshold and is not marked as abnormal, the current comment is considered a high-quality positive comment. If the overall sentiment score of the current comment is higher than the preset negative threshold and is not marked as abnormal, the current comment is considered a genuine negative comment.

[0013] As can be seen from the above technical solution, compared with the prior art, the present invention has at least the following advantages and positive effects: (1) This invention uses text sentiment analysis model and image sentiment analysis model to analyze text and images in e-commerce reviews separately, and adopts dynamic weight fusion strategy to automatically perform multimodal sentiment analysis on e-commerce reviews, thereby accurately identifying user emotions in e-commerce reviews and significantly improving processing efficiency.

[0014] (2) The present invention constructs an automated abnormal comment identification mechanism. By evaluating the emotional consistency in the same comment, abnormal comments are screened out and targeted processing is carried out. It can effectively identify comments that are ironic or have conflicting emotional polarities, and further enhance the accuracy of the sentiment analysis results.

[0015] (3) This invention directly and automatically drives cross-system business execution such as customer service, products, orders, and promotions through sentiment analysis results. It can realize automated and differentiated business closed-loop response to positive and negative reviews, greatly reduce labor costs, and improve operational efficiency and user experience. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below only involve some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 The flowchart illustrates a method for intelligent processing of e-commerce reviews based on multimodal sentiment analysis, as provided in an embodiment of the present invention. Figure 2 The flowchart illustrates a method for dynamically calculating text weights and image weights provided in an embodiment of the present invention. Figure 3 The flowchart illustrates a method for calculating comment processing priority and generating high-priority processing work orders according to an embodiment of the present invention. Figure 4 This diagram illustrates the structure of an intelligent e-commerce review processing system based on multimodal sentiment analysis, as provided in an embodiment of the present invention. in, Figure 4 The annotations in the accompanying drawings are explained as follows: 400 - An intelligent e-commerce review processing system based on multimodal sentiment analysis; 401 - Data preprocessing module; 4011 - Data acquisition unit; 4012 - Data preprocessing unit; 402 - Text sentiment analysis module; 4021 - Text model training unit; 4022 - Text model inference unit; 403 - Image sentiment analysis module; 4031 - Image model training unit; 4032 - Image model inference unit; 404 - Comprehensive sentiment scoring module; 4041 - Weight allocation unit; 4042 - Comprehensive calculation unit; 405 - Abnormal review marking module; 4051 - Consistency assessment unit; 4052 - Abnormal marking unit; 406 - Business logic execution module; 4061 - Priority calculation unit; 4062 - Incentive compensation unit; 4063 - Intelligent control unit. Detailed Implementation

[0018] To more clearly explain the purpose, technical solutions, and advantages of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein. On the contrary, these embodiments are provided so that the present invention will be more comprehensive and complete, and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0019] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full understanding of embodiments of the invention. However, those skilled in the art will recognize that the technical solutions of the invention can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of the invention.

[0020] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0021] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0022] The present invention will now be described in detail with reference to specific embodiments: Example 1

[0023] like Figure 1 As shown in the figure, this embodiment of the invention provides a method for intelligent processing of e-commerce reviews based on multimodal sentiment analysis. The specific steps of this method are as follows: S1: Collect e-commerce review data in real time, and preprocess the e-commerce review data to obtain preprocessed review data; wherein, the preprocessed review data includes text review data and / or image review data; In this embodiment, the system collects massive amounts of comment data from various e-commerce platforms in real time and preprocesses each comment data item to obtain preprocessed comment data containing text comment data and / or image comment data. The preprocessing methods include: cleaning the text data to remove noise content such as emoticons, special characters, and HTML tags, and performing standardization processing such as unified formatting; and performing image data size standardization, watermark removal, and background separation. The specific preprocessing methods can be conventionally set according to actual needs, and this embodiment does not impose any restrictions.

[0024] S2: Train the text sentiment analysis model. Input the text comment data into the trained text sentiment analysis model for processing to obtain the text sentiment score for each comment. S3: Train the image sentiment analysis model. Input the image comment data into the trained image sentiment analysis model for processing to obtain the image sentiment score for each comment. Further, after obtaining the preprocessed comment data, a text sentiment analysis model and an image sentiment analysis model are trained. Preferably, in this embodiment, the text sentiment analysis model adopts the BERT model. Specifically, the BERT model uses a Transformer-based pre-trained encoder as a shared underlying layer and constructs a shared-dedicated multi-task learning framework. The main structural components of this framework include: a shared BERT encoder, used to encode the input token sequence and extract the general semantic representation of the text; an attribute classification head, connected to the output of the encoder, which is a fully connected layer, used to output a binary classification probability for each predefined product attribute to determine whether the attribute is mentioned; and a sentiment classification head, also connected to the output of the encoder, which is a fully connected layer, used to independently predict the sentiment polarity of each attribute identified by the attribute classification head based on the general semantic representation, and output a three-class (positive / neutral / negative) probability distribution.

[0025] Specifically, training the text sentiment analysis model (BERT model) involves: Multimodal e-commerce review data containing text and images is acquired and preprocessed to obtain an initial training set. The preprocessing method is the same as in step S1 and can be adjusted as needed, so it will not be elaborated here. Text review data is selected from the initial training set. For each review in the text review data, a combination of rule templates and unsupervised clustering is used to automatically label the product attributes involved in each review. Professional labelers then label the sentiment polarity (positive / neutral / negative) of each review, ultimately generating a structured labeled dataset. Each sample in the structured labeled dataset contains the review text, a list of related attributes, and sentiment labels for each attribute. The structured labeled dataset is divided into a training set and a validation set. The BERT tokenizer is used to segment and encode the review text in the training set, generating corresponding Token ID sequences, attention masks, and positional codes, which together constitute the standard input tensor of the model. The standard input tensor and the corresponding attributes and sentiment labels are input into the text sentiment analysis model (BERT model) for training. During forward propagation, after the text is encoded by the shared encoder, the attribute classification head calculates the attribute mention probability, and the sentiment classification head calculates the sentiment probability of each attribute. The total loss function L = α·L_attr + (1−α)·L_sent (where α is the dynamic weight coefficient) is calculated, and backpropagation is performed to optimize and update all parameters of the encoder and the two classification heads in the model. The performance of the trained model parameters is evaluated using a validation set, quantifying its metrics on attribute recognition and attribute-level sentiment analysis tasks. Finally, the trained text sentiment analysis model is encapsulated as a callable API service, completing training and deployment.

[0026] Preferably, in this embodiment, the image sentiment analysis model uses the ResNet model. Specifically, when training the image sentiment analysis model (ResNet), image comment data needs to be selected from the initial training set. The training and validation process adopts a conventional standard procedure, which will not be described in detail in this embodiment.

[0027] Furthermore, after training the text sentiment analysis model and the image sentiment analysis model, the preprocessed comment data obtained in step S1 is processed using the trained text sentiment analysis model and the trained image sentiment analysis model, respectively, to obtain the text sentiment score and the image sentiment score.

[0028] Specifically, obtaining a text sentiment score involves the following steps: The trained text sentiment analysis model extracts product attributes from the text review data, predicts the text sentiment category for each product attribute, and outputs an initial text score vector for each text sentiment category. Apply the softmax function to the initial score vector of the text to obtain the confidence score for each sentiment category of the text; A sentiment score is predefined for each of the text sentiment categories, and a weighted sum is performed with the confidence score to obtain the text sentiment score.

[0029] The sum of the confidence scores for all text sentiment categories is 1; the text sentiment score ranges from 0 to 1, where 0 represents extremely positive sentiment and 1 represents extremely negative sentiment.

[0030] Specifically, obtaining an image sentiment score involves the following steps: The image comment data is input into a trained image sentiment analysis model to extract image feature vectors; The image feature vectors are mapped to the image sentiment category space, and the initial image score vector for each image sentiment category is output; wherein, the sentiment category space is composed of a predefined set of sentiment categories; Apply the softmax function to the initial score vector of the image to obtain the confidence score of each image sentiment category; A predefined sentiment score is defined for each of the image sentiment categories, and a weighted sum is performed with the confidence score to obtain the image sentiment score.

[0031] The sum of the confidence scores for all image sentiment categories is 1; the range of the image sentiment score is 0 to 1, where 0 represents extremely positive sentiment and 1 represents extremely negative sentiment.

[0032] S4: Based on the validity of the preprocessed comment data, dynamically calculate the text weight and image weight, and calculate the comprehensive sentiment score for each comment; Further, after obtaining the text sentiment score and the image sentiment score, a comprehensive sentiment score is calculated for each comment. Specifically, the comprehensive sentiment score is calculated using text weight, the text sentiment score, image weight, and the image sentiment score, using the following formula: S = Wt*St + Wi*Si; Where S is the comprehensive sentiment score, and S∈[0,1], 0 represents extremely positive sentiment, and 1 represents extremely negative sentiment; Wt is the text weight, St is the text sentiment score, Wi is the image weight, and Si is the image sentiment score.

[0033] Preferably, in this embodiment, the text weights and image weights are dynamically calculated based on the validity of the preprocessed comment data obtained in step S1, such as... Figure 2 As shown, the specific steps include the following: S201: Traverse each comment in the preprocessed comment data and detect the components of each comment; S202: If the current comment only contains text comments, then directly set the text weight to 1 and the image weight to 0; S203: If the current comment only contains image comments, then directly set the text weight to 0 and the image weight to 1; S204: If the current comment contains both text comments and image comments, then calculate the text contribution score and image contribution score of the current comment respectively. Based on the proportion of the text contribution score and the image contribution score in the total score, calculate the text weight and the image weight respectively. S205: Complete the weight assignment for the current comment, start processing the next comment, and repeat the steps of traversing comments and weight assignment until the weight assignment for each comment in the preprocessed comment data has been completed.

[0034] The weighting of text and images in the comments is calculated based on their respective contributions, and the specific steps include the following: The number of valid characters in the current comment is counted, and the number of valid characters is matched with a preset score to obtain the text contribution score. Valid characters include Chinese, English, numbers, and other non-whitespace characters. The preset score can be set according to historical experience or operational needs, and includes multiple gradients, each gradient corresponding to a specific number of valid characters. Filter the valid images in the current comments and select the top K valid images; where invalid images include placeholder images, thumbnails automatically generated by the system, images with incorrect format or damage, etc. The specific definition of whether an image is valid can be set according to actual needs, and this embodiment does not impose any restrictions; Each of the first K images is converted to a grayscale image, and the Laplacian variance of each image is calculated. The Laplacian variance is compared with a preset level threshold to obtain the sharpness level of each image; the larger the Laplacian variance, the richer the edge details and the clearer the image is overall; conversely, the smaller the variance, the blurrier the image is. The preset threshold can be set according to historical experience or operational needs, and includes multiple gradients, each of which corresponds to a specific sharpness level. The sharpness level is mapped to a preset sharpness coefficient to obtain the sharpness coefficient for each image; The base score for each image is set to a uniform fixed value, and the base score is weighted and summed with the sharpness coefficient corresponding to each image to obtain the image contribution score; The sum of the text contribution score and the image contribution score is used to calculate the total score. Based on the proportion of the text contribution score and the image contribution score in the total score, the text weight and the image weight are obtained respectively.

[0035] S5: Perform sentiment consistency assessment on each comment in the preprocessed comment data, determine whether the current comment is an abnormal comment, and mark the abnormal comment; In this embodiment, after obtaining the text sentiment score and the image sentiment score, it is also necessary to identify and mark abnormal comments. Specifically, in this embodiment, determining whether the current comment is an abnormal comment includes two aspects: On one hand, based on the text sentiment score and the image sentiment score, the sentiment consistency score of the text comment and the image comment in the current comment is calculated; the specific calculation formula is: C = |St-Si|, where C is the sentiment consistency score, St is the text sentiment score, and Si is the image sentiment score; the sentiment consistency score is compared with a preset threshold. If the sentiment consistency score is lower than the preset threshold, the current comment is determined to be a normal comment; otherwise, it is an abnormal comment. The preset threshold can be set conventionally based on historical experience or operational needs, and this embodiment does not impose any restrictions.

[0036] On the other hand, based on the trained text sentiment analysis model, the system identifies whether there are N different sentiment categories for the same product attribute in the current comment. If so, the current comment is determined to be an abnormal comment. For example, a comment might read, "This phone has a really long standby time; it only lasts for 2 hours before needing to be charged." For this comment, the text sentiment analysis model will identify two completely opposite sentiment categories, so the system automatically marks this comment as an abnormal comment.

[0037] S6: Based on the comprehensive sentiment score and / or abnormal comment markers, automatically trigger cross-system linkage e-commerce comment response operations.

[0038] Furthermore, after obtaining a comprehensive sentiment score through step S4 and marking abnormal comments through step S5, cross-system collaborative business execution is carried out. Specifically, this automatically triggers cross-system collaborative e-commerce comment response operations, including: Based on the comprehensive sentiment score, a priority initial set is selected. The weight coefficient and activity status parameters of the product corresponding to each comment in the priority initial set are obtained. The processing priority of each comment is calculated, and a high-priority processing work order is automatically generated by calling the customer service system API interface. Wherein, for example... Figure 3 As shown, calculating comment processing priority and generating high-priority processing work orders includes the following steps: S301: Set the initial weight coefficients for all products in the system, and periodically adjust the weight coefficients of the products dynamically based on their sales revenue. Specifically, during system initialization, all products are configured as ordinary products by default, assuming the initial weight coefficient G is uniformly set to 1.0. Then, operations personnel can precisely adjust G according to marketing strategy needs, such as setting the G value of the main product to 1.5, the G value of the best-selling product to 2.0, etc. Furthermore, sales volume gradients can be set, for example: when the total sales amount of a product increases by 10,000 yuan, the G value automatically increases by 0.05, with no upper limit. The system will also automatically calculate the total sales amount of products and automatically adjust the G value on a regular basis (such as daily, weekly, etc.). By quantifying the correlation between sales data and weight coefficients, it ensures that high-sales products receive a better priority in work order priority calculation, achieving the goal of refined management of e-commerce operations.

[0039] S302: Calculate the activity status parameters of the product based on the quantity of the product participating in the promotional activity; Specifically, assuming the system divides promotional activities into three levels based on scale and sets corresponding activity coefficients for each level: general promotional activities (coefficient 1.1), medium-sized promotional activities (coefficient 1.2), and large-scale promotional activities (coefficient 1.3), let n1 be the number of general promotional activities a certain product is currently participating in, n2 be the number of medium-sized promotional activities, and n3 be the number of large-scale promotional activities. Then, the formula for calculating the activity status parameter P of the product is: P = 1.1*n1 + 1.2*n2 + 1.3*n3; the P value is automatically calculated every time the product is added to the product pool of a promotional activity.

[0040] S303: Set a first priority threshold, filter out comments whose comprehensive sentiment score is higher than the first priority threshold, and obtain an initial set for priority processing; Furthermore, a first priority threshold is set, which can be set according to actual operational needs; no restrictions are imposed in this embodiment. Initial screening is performed using a comprehensive sentiment score. The comprehensive sentiment score is compared with the set first priority threshold, and product review data that is higher than the first priority threshold is selected to obtain the initial set for priority processing.

[0041] S304: For each comment in the initial priority processing set, obtain the weight coefficient and activity status parameters of the product corresponding to each comment, and calculate the processing priority of each comment; Furthermore, the processing priority of each comment in the initial set is calculated using the following formula: Priority = S×G×P; where Priority is the processing priority, S is the overall sentiment score, G is the weight coefficient of the product, and P is the activity status parameter of the product.

[0042] S305: Set a second priority threshold. When the processing priority is not lower than the second priority threshold, automatically call the customer service system API interface to generate a high-priority processing work order. Further, after calculating the processing priority of each comment, compare it with the preset second priority threshold, screen out the product comment data that is higher than or equal to the second priority threshold, and automatically call the work order API interface of the customer service system for these comments to generate corresponding high-priority processing work orders, and let the customer service personnel intervene and follow up, so as to prevent high-negative comments from not being appeased and processed in time, thereby forming an intelligent decision-making closed loop from emotion recognition to business response.

[0043] In addition, for all comments with the abnormal comment mark, also automatically call the customer service system API interface to generate a high-priority processing work order. Further, the cross-system linkage response further includes: based on the comprehensive sentiment score and the abnormal comment mark, judge whether the current comment is a high-quality positive comment or a real negative evaluation, automatically call the promotion system API interface to issue rewards to the user who posts the high-quality positive comment or the user who posts the real negative evaluation, and call the customer service system API to generate a personalized automated reply; wherein, the judgment basis for the high-quality positive comment and the real negative evaluation is: Preset a positive threshold and a negative threshold. If the comprehensive sentiment score corresponding to the current comment is lower than the preset positive threshold and is not marked as abnormal, then the current comment is a high-quality positive comment; if the comprehensive sentiment score corresponding to the current comment is higher than the preset negative threshold and is not marked as abnormal, then the current comment is a real negative comment.

[0044] The rewards issued include but are not limited to points, coupons, etc.; the personalized automated reply is obtained by matching from the reply templates pre-set in the system.

[0045] Monitor the positive proportion of the high-quality positive comments and the negative proportion of the real negative evaluations within the set time window, and compare them with the preset regulation thresholds respectively, and automatically call the supply chain system API interface to increase the inventory of the product or automatically call the product system API interface to remove the relevant product from the shelves. Specifically, the preset regulation thresholds here include a positive regulation threshold and a negative regulation threshold. When it is monitored that the positive proportion of the high-quality positive comments reaches or exceeds the positive regulation threshold, the supply chain system API interface will be automatically called to increase the inventory of the product; when it is monitored that the negative proportion of the real negative evaluations reaches or exceeds the negative regulation threshold, the product system API interface will be automatically called to remove the relevant product from the shelves; the preset regulation thresholds can be set regularly according to actual operation needs, and this embodiment does not make any restrictions.

[0046] The e-commerce review intelligent processing method based on multimodal sentiment analysis described in this invention can not only automatically perform multimodal sentiment analysis on e-commerce reviews, effectively identify ironic or emotionally conflicting reviews, and accurately identify user emotions in e-commerce reviews, significantly improving processing efficiency, but also directly and automatically drive cross-system business execution such as customer service, products, orders, and promotions. This enables automated and differentiated closed-loop responses to positive and negative reviews, greatly reducing labor costs and improving operational efficiency and user experience. Example 2

[0047] like Figure 4 As shown, this embodiment of the invention provides a multimodal sentiment analysis-based intelligent e-commerce review processing system 400. The system includes: a data preprocessing module 401, a text sentiment analysis module 402, an image sentiment analysis module 403, a comprehensive sentiment scoring module 404, an abnormal review marking module 405, and a business logic execution module 406; wherein: The data preprocessing module 401 is used to collect e-commerce review data in real time and preprocess the e-commerce review data to obtain preprocessed review data; wherein, the preprocessed review data includes text review data and / or image review data; The data preprocessing module 401 includes: a data acquisition unit 4011 and a data preprocessing unit 4012; wherein: The aforementioned data acquisition unit 4011 is configured to: collect e-commerce review data from various e-commerce platforms in real time; The aforementioned data preprocessing unit 4012 is configured to preprocess the e-commerce review data collected by the data acquisition unit 4011 to obtain preprocessed review data including text and / or images.

[0048] The text sentiment analysis module 402 is used to train the text sentiment analysis model. The text comment data is input into the trained text sentiment analysis model for processing to obtain the text sentiment score of each comment. The text sentiment analysis module 402 includes: a text model training unit 4021 and a text model inference unit 4022; wherein: The aforementioned text model training unit 4021 is configured to: pre-train a text sentiment analysis model using multimodal e-commerce review data; The aforementioned text model inference unit 4022 is configured to process the text comment data in the data preprocessing unit 4012 using a trained text sentiment analysis model to obtain a text sentiment score for each comment.

[0049] The image sentiment analysis module 403 is used to train the image sentiment analysis model. The image comment data is input into the trained image sentiment analysis model for processing to obtain the image sentiment score for each comment. Image sentiment analysis module 403 includes: image model training unit 4031 and image model inference unit 4032; wherein: The aforementioned image model training unit 4031 is configured to: pre-train an image sentiment analysis model using multimodal e-commerce review data; The aforementioned image model inference unit 4032 is configured to process the image comment data in the data preprocessing unit 4012 using a trained image sentiment analysis model to obtain an image sentiment score for each comment.

[0050] The comprehensive sentiment scoring module 404 is used to dynamically calculate text weights and image weights based on the validity of the preprocessed comment data, and to calculate the comprehensive sentiment score for each comment. The comprehensive sentiment scoring module 404 includes: a weight allocation unit 4041 and a comprehensive calculation unit 4042; wherein: The aforementioned weight allocation unit 4041 is configured to dynamically calculate text weights and image weights based on the validity of the preprocessed comment data. The aforementioned comprehensive calculation unit 4042 is configured to calculate the comprehensive sentiment score for each comment based on the text weights and image weights assigned by the weight allocation unit 4041, the text sentiment score obtained in the text model inference unit 4022, and the image sentiment score obtained in the image model inference unit 4032.

[0051] The abnormal comment marking module 405 is used to perform sentiment consistency assessment on each comment in the preprocessed comment data, determine whether the current comment is an abnormal comment, and mark the abnormal comment. The anomaly comment marking module 405 includes: a consistency evaluation unit 4051 and an anomaly marking unit 4052; wherein: The aforementioned consistency assessment unit 4051 is configured to: perform sentiment consistency assessment on each comment in the preprocessed comment data, and calculate the sentiment consistency score for each comment; The aforementioned anomaly labeling unit 4052 is configured to: identify and label abnormal comments based on the sentiment consistency score obtained in the consistency evaluation unit 4051; and identify whether the same comment has N different sentiment categories for the same product attribute through the text sentiment analysis model trained in the text model training unit 4021, and then identify and label abnormal comments.

[0052] The business logic execution module 406 is used to automatically trigger cross-system linkage e-commerce comment response operations based on the comprehensive sentiment score and / or abnormal comment markers; The business logic execution module 406 includes: a priority calculation unit 4061, an incentive compensation unit 4062, and an intelligent control unit 4063; wherein: The aforementioned priority calculation unit 4061 is configured to: calculate the processing priority of each comment based on the comprehensive sentiment score obtained in the comprehensive calculation unit 4042, and automatically call the customer service system API interface to generate a high-priority processing work order; and automatically call the customer service system API interface to generate a high-priority processing work order for all comments with abnormal comment tags. The aforementioned incentive compensation unit 4062 is configured to: determine whether the current comment is a high-quality positive review or a genuine negative review based on the comprehensive sentiment score and abnormal comment markers, automatically call the promotion system API interface to issue rewards to the user who published the high-quality positive review or the user who published the genuine negative review, and call the customer service system API to generate personalized automated replies; The aforementioned intelligent control unit 4063 is configured to: monitor the positive ratio of high-quality positive reviews and the negative ratio of genuine negative reviews within a set time window, compare them with preset control thresholds respectively, and automatically call the supply chain system API interface to increase the inventory of the product or automatically call the product system API interface to remove the relevant product from the shelves.

[0053] In this embodiment, the text and images in e-commerce review data are processed separately by the text sentiment analysis module 402 and the image sentiment analysis module 403, and a dynamic weight fusion strategy is adopted to automatically perform multimodal sentiment analysis on e-commerce reviews, thereby accurately identifying user emotions in e-commerce reviews and significantly improving processing efficiency. The abnormal review marking module 405 effectively identifies ironic or emotionally conflicting reviews, which can further enhance the accuracy of sentiment analysis results. The business logic execution module 406 directly and automatically drives cross-system business execution such as customer service, products, orders, and promotions based on sentiment analysis results, which can realize automated and differentiated closed-loop business responses to positive and negative reviews, greatly reducing labor costs and improving operational efficiency and user experience.

[0054] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims. It should be understood that the invention is not limited to the precise structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for intelligent processing of e-commerce reviews based on multimodal sentiment analysis, characterized in that, Includes the following steps: E-commerce review data is collected in real time, and the e-commerce review data is preprocessed to obtain preprocessed review data; wherein, the preprocessed review data includes text review data and / or image review data; Train a text sentiment analysis model, input the text comment data into the trained text sentiment analysis model for processing, and obtain the text sentiment score for each comment; Train an image sentiment analysis model, input the image comment data into the trained image sentiment analysis model for processing, and obtain the image sentiment score for each comment; Based on the validity of the preprocessed comment data, text weights and image weights are dynamically calculated, and a comprehensive sentiment score for each comment is calculated. For each comment in the preprocessed comment data, a sentiment consistency assessment is performed to determine whether the current comment is an anomalous comment, and the anomalous comments are marked. Based on the comprehensive sentiment score and / or abnormal comment markers, cross-system linkage e-commerce comment response operations are automatically triggered.

2. The intelligent processing method according to claim 1, characterized in that, The process of inputting the text comment data into a trained text sentiment analysis model for processing specifically involves: The trained text sentiment analysis model extracts product attributes from the text review data, predicts the text sentiment category for each product attribute, and outputs an initial text score vector for each text sentiment category. Apply the softmax function to the initial score vector of the text to obtain the confidence score for each sentiment category of the text; A sentiment score is predefined for each of the text sentiment categories, and a weighted sum is performed with the confidence score to obtain the text sentiment score.

3. The intelligent processing method according to claim 1, characterized in that, The step of inputting the image comment data into the trained image sentiment analysis model for processing specifically involves: The image comment data is input into a trained image sentiment analysis model to extract image feature vectors; The image feature vectors are mapped to the image sentiment category space, and the initial image score vector for each image sentiment category is output. Apply the softmax function to the initial score vector of the image to obtain the confidence score of each image sentiment category; A predefined sentiment score is defined for each of the image sentiment categories, and a weighted sum is performed with the confidence score to obtain the image sentiment score.

4. The intelligent processing method according to claim 1, characterized in that, The dynamic calculation of text weights and image weights specifically involves: Iterate through each comment in the preprocessed comment data; If the current comment contains only text comments, then the text weight is set to 1 and the image weight is set to 0. If the current comment only contains image comments, then the text weight is set to 0 and the image weight is set to 1. If the current comment contains both text and image comments, then the text contribution score and image contribution score in the current comment are calculated separately. Based on the proportion of the text contribution score and the image contribution score in the total score, the text weight and the image weight are calculated respectively.

5. The intelligent processing method according to claim 4, characterized in that, The text contribution score and image contribution score in the current comment are calculated separately. Based on the proportion of the text contribution score and image contribution score in the total score, the text weight and image weight are calculated respectively, specifically as follows: The number of valid characters in the current comment is counted, and the number of valid characters is matched with a preset score to obtain the text contribution score; Filter the valid images in the current comments, select the top K images from the valid images, calculate the clarity coefficient of each of the top K images, and obtain the image contribution score based on the clarity coefficient; The sum of the text contribution score and the image contribution score is used to calculate the total score. Based on the proportion of the text contribution score and the image contribution score in the total score, the text weight and the image weight are obtained respectively.

6. The intelligent processing method according to claim 5, characterized in that, The sharpness coefficient of each of the first K images is calculated, and the image contribution score is obtained based on the sharpness coefficient, specifically as follows: Each of the first K images is converted to a grayscale image, and the Laplacian variance of each image is calculated. The Laplacian variance is compared with a preset level threshold to obtain the sharpness level of each image; The sharpness level is mapped to a preset sharpness coefficient to obtain the sharpness coefficient for each image; The base score for each image is set to a uniform fixed value, and the base score is weighted and summed with the sharpness coefficient corresponding to each image to obtain the image contribution score.

7. The intelligent processing method according to claim 1, characterized in that, The comprehensive sentiment score is calculated using the following formula: S = Wt*St + Wi*Si; Where S is the comprehensive sentiment score, and S∈[0,1], 0 represents extremely positive sentiment, and 1 represents extremely negative sentiment; Wt is the text weight, St is the text sentiment score, Wi is the image weight, and Si is the image sentiment score.

8. The intelligent processing method according to claim 1, characterized in that, The determination of whether the current comment is an abnormal comment specifically involves: Based on the text sentiment score and the image sentiment score, calculate the sentiment consistency score between the text comment and the image comment in the current comment. Compare the sentiment consistency score with a preset threshold. If the sentiment consistency score is lower than the preset threshold, the current comment is determined to be a normal comment; otherwise, it is an abnormal comment. Based on the trained text sentiment analysis model, the system identifies whether there are N different sentiment categories for the same product attribute in the current comment. If so, the current comment is determined to be an abnormal comment.

9. The intelligent processing method according to claim 1, characterized in that, The automatically triggered cross-system e-commerce comment response operation specifically includes: Based on the comprehensive sentiment score, a priority initial set is selected, and the weight coefficient and activity status parameters of the product corresponding to each comment in the priority initial set are obtained. The processing priority of each comment is calculated, and the customer service system API interface is automatically called to generate a high-priority processing work order. For all comments marked with the aforementioned abnormal comment tag, the customer service system API interface will be automatically invoked to generate a high-priority processing work order; Based on the comprehensive sentiment score and the abnormal comment marker, determine whether the current comment is a high-quality positive review or a genuine negative review, automatically call the promotion system API interface to issue rewards to users who published the high-quality positive review or users who published the genuine negative review, and call the customer service system API to generate personalized automated replies; Monitor the positive ratio of high-quality positive reviews and the negative ratio of genuine negative reviews within a set time window, and compare them with preset control thresholds. Automatically call the supply chain system API interface to increase the inventory of the product or automatically call the product system API interface to remove the relevant product from the shelves.

10. A multimodal sentiment analysis-based intelligent processing system for e-commerce reviews, characterized in that, include: The data preprocessing module is used to collect e-commerce review data in real time and preprocess the e-commerce review data to obtain preprocessed review data; wherein, the preprocessed review data includes text review data and / or image review data; The text sentiment analysis module is used to train the text sentiment analysis model. The text comment data is input into the trained text sentiment analysis model for processing to obtain the text sentiment score of each comment. The image sentiment analysis module is used to train the image sentiment analysis model. The image comment data is input into the trained image sentiment analysis model for processing to obtain the image sentiment score for each comment. The comprehensive sentiment scoring module is used to dynamically calculate text weights and image weights based on the validity of the preprocessed comment data, and to calculate the comprehensive sentiment score for each comment. An abnormal comment marking module is used to perform sentiment consistency assessment on each comment in the preprocessed comment data, determine whether the current comment is an abnormal comment, and mark the abnormal comment. The business logic execution module is used to automatically trigger cross-system e-commerce comment response operations based on the comprehensive sentiment score and / or abnormal comment markers.

Images