High-quality live marketing text recognition method
By using attnGAN and EMD-LSTM models on e-commerce platforms to combine text, pictures and time features to identify high-quality evaluations, the problem of high-quality evaluation and identification in e-commerce platforms is solved, and the accuracy of consumer shopping decisions and product quality display is improved. Clarity.
Patent Information
- Application Number
- CN202411084142.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-08
- Publication Date
- 2024-11-29
AI Technical Summary
The existing technology is difficult to effectively identify and extract high-quality evaluations from e-commerce platforms, affecting consumers' shopping decisions and making it difficult for consumers to judge the true quality of the products.
Using high-quality live broadcast marketing text recognition method, the attention generation adversarial network (attnGAN) and EMD-LSTM model are used, and the product prediction image and prediction evaluation product are generated through text, pictures and time characteristics, and the picture-related matching reliability and evaluation are calculated. Add confidence, obtain the confidence of high-quality evaluation through weighted average, and identify high-quality evaluation.
It realizes more accurate identification and extraction of high-quality evaluations, helps consumers make correct consumption choices, and improves the clarity of product quality display and consumer trust.
Smart Images

Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] The present invention relates to the field of e-commerce commodity data mining, and in particular to a high-quality live broadcast marketing text recognition method. Background Art
[0002] In recent years, e-commerce has developed rapidly, and consumers can buy a variety of goods online. In the process of purchasing, reviews have a great impact on consumers' willingness to buy, especially high-quality reviews. According to a survey conducted by BrightLocal in 2019, about 90% of consumers said they would check online reviews before purchasing products or services. Moreover, 88% of consumers said they would trust online reviews just as they would trust the advice of friends or family. According to a study conducted by Spiegel Research Center in 2017, products with five-star high-quality reviews have nearly four times higher sales than products with no reviews or low reviews. According to a study conducted by Moz in 2018, products with high-quality reviews rank higher on search engine results pages than products without reviews. According to a 2015 Nielsen survey of 30,000 online shoppers in 22 countries around the world, 85% of consumers will check other consumers' reviews on the website before purchasing goods online, and 50% of consumers said they will only buy goods with high-quality reviews. Another survey shows that according to a questionnaire survey conducted by a survey company on 1,000 consumers at a certain time, 96% of consumers believe that high-quality reviews of e-commerce have an important impact on their shopping decisions, and 87% of consumers believe that the authenticity and credibility of reviews have a great impact on their shopping decisions. In addition, 82% of consumers will choose to buy products with high-quality reviews, and 63% of consumers will give up buying products with no reviews or poor reviews. These data show that whether the quality of e-commerce reviews is high has a vital impact on consumers' purchasing decisions. Low-quality reviews will interfere with consumers' purchasing decisions, making it difficult for them to judge the true quality of the goods, thereby causing consumers to choose inappropriate goods. High-quality reviews can more clearly show the actual parameters of the goods, allowing consumers to have a clearer understanding of the goods and then choose the goods they need. Therefore, it is necessary to develop an effective method to identify and extract high-quality reviews from product reviews on e-commerce platforms to help consumers make the right consumer choices. Summary of the invention
[0003] In order to solve the above problems, the present invention provides a high-quality live marketing text recognition method, which collects evaluation text, evaluation pictures, evaluation time, and product style and product detail parameter diagram of product evaluation; for evaluations within a long period of time, some high-quality evaluations are manually selected, and the comment texts and pictures marked as high-quality evaluations are divided into training sets and test sets; by using the evaluation text, product style and evaluation pictures in the training set, the attention generation adversarial network attnGAN is trained to obtain a model Q for generating product prediction pictures from the evaluation text, and the product prediction pictures generated by the evaluation text using the model Q and the actual evaluation pictures and product detail parameter diagrams are matched and extracted to obtain the picture-text correlation confidence w1 of the evaluation to be tested; the evaluation product (the product of the number of evaluations and the evaluation score) at each evaluation time in the evaluation data set is used for training through EMD-LSTM, and the evaluation product value that should be had at the evaluation time of the evaluation to be tested is predicted by fitting the evaluation product of the historical evaluation time, and the evaluation product confidence w2 of the evaluation to be tested is obtained by comparing the difference between the actual evaluation product and the predicted evaluation product. The confidence w1 is obtained by comparing the similarity between the image generated by text and the evaluation attached image and the product details parameter map, and the evaluation product confidence w2 is obtained by predicting the evaluation time-evaluation product curve through EMD-LSTM. The high-quality confidence w of the evaluation is obtained by weighted averaging the two confidences, including the following steps:
[0004] Based on the existing review information, we manually selected some high-quality reviews over a long period of time, annotated the features of the review texts and pictures identified as high-quality reviews, and divided them into training sets and test sets; the review time, review score, and number of existing review information were taken as part of the data set.
[0005] We analyze the text and product styles in the high-quality review set, extract features based on the entity attributes of the text through transformer, and construct the review text feature set R1.
[0006] We analyze the images in the high-quality review set, perform object detection and feature extraction on the images based on the transformer, and construct the review image feature set R2.
[0007] The text feature set R1 and the image feature set R2 in the high-quality evaluation set are used to obtain the corresponding matching of each text feature and image feature in the high-quality data set using the attnGAN-based method. The generator and discriminator are trained to construct a function model Q for generating images based on text.
[0008] Specifically: Generator G is responsible for generating a product prediction map based on text features, while discriminator D is responsible for judging the quality of this image, that is, judging whether it is a real sample or a generated fake sample. Through gradual iteration, generator G can eventually generate more and more realistic images, while discriminator D can more accurately judge the authenticity of the image. When generator G is fixed, maximizing V(G,D) means that for real data x derived from the real data distribution Pdatas, the output of discriminator D(x) is close to 1; and for data G(z) generated by generator G, the output of discriminator D(G(z)) is close to 0. This means that by optimizing the parameters of discriminator D, it can effectively distinguish between real data and generated data. When discriminator D is fixed, the goal of minimizing V(G,D) is to make the distribution of data G(z) generated by generator G close to the real data distribution Pdatas, so that it is difficult for discriminator D to distinguish between generated data and real data. At this time, by optimizing the parameters of generator G, the data generated by it is statistically as close to the real data as possible.
[0009] Through this iterative process of mutual learning, the generator G can gradually learn to generate more realistic images, while the discriminator D can more accurately judge the authenticity of the image, thereby achieving the goal of mutual optimization between the generator and the discriminator and ultimately generating high-quality images.
[0010]
[0011]
[0012] Use the text-generated image model Q to analyze the comment text of each evaluation to be evaluated and generate image v1. Based on Faster-CNN, feature extraction is performed on the text-generated image v1, comment image v2, and product detail parameter image v3 to obtain the feature vector {t1, t2, t3}.
[0013] The obtained text is used to generate image features t1, comment image features t2, and product parameter image features t3. The DBSCAN density clustering algorithm is used to cluster the intra-family features t4 of the three images v1, v2, and v3. The cosine similarity is used to calculate the correlation degree matrix {u1, u2, u3} between t4 and t1, t2, and t3, respectively, and the image-text correlation matching confidence w1 of high-quality evaluation is obtained.
[0014] The confidence test of the evaluation product of the evaluation to be tested is based on EMD-LSTM. The Y-axis of the plane coordinate system is established with the evaluation product (number of evaluations * average score) in this time period, and the evaluation time is the X-axis. The evaluation product-evaluation time distribution function is established, and the evaluation product of the historical evaluation time is fitted through EMD-LSTM to obtain the evaluation time distribution prediction function P, and the evaluation product P(t) at the evaluation time of the evaluation to be tested is predicted.
[0015] Specifically: EMD-LSTM prediction can be divided into two parts, EMD decomposition and LSTM prediction. The evaluation product-evaluation time curve is decomposed into N modular functions (IMF1, IMF2, ..., IMFN) and RES through the EMD decomposition algorithm. LSTM models are established for each IMFi and res respectively, and then the prediction results are superimposed to obtain the final prediction result, and the root mean square error is used to evaluate the model prediction. The root mean square error expression is:
[0016]
[0017] The evaluation product confidence w2 is calculated by using the function M(m) formed by the difference m between the predicted evaluation product P(t) calculated by the evaluation time distribution prediction model at a certain moment and the actual evaluation product S(t) corresponding to the evaluation time period to be evaluated.
[0018] m=P(t)-S(t)(1)
[0019]
[0020] The confidence w of a certain consumer evaluation as a high-quality evaluation at that moment is obtained by the confidence w1 of the image-text correlation matching at a certain time point and the confidence w2 of the time series. By calculating the confidence of the high-quality evaluation data set, the confidence value of the lowest judgment as a high-quality evaluation is obtained.
[0021]
[0022] Among them, f1 is the weight corresponding to the confidence of image-text correlation matching, and f2 is the weight corresponding to the confidence of time series.
[0023] The beneficial effects of the present invention are as follows: the present invention adopts a high-quality live marketing text recognition method to collect evaluation text, evaluation pictures, evaluation time, product style and product detail pictures of product evaluation; manually select some high-quality evaluations, mark the comment texts and pictures identified as high-quality evaluations and the time of high-quality evaluations, and generate pictures from text using a text generation picture model Q trained by an attention generation adversarial network AttnGAN; calculate the confidence w1 by comparing the similarity of images generated by text and evaluation pictures and product detail pictures; and obtain the evaluation product confidence w2 by fitting a time-evaluation product curve through EMD-LSTM, predicting the difference between the predicted evaluation product P(t) and the actual evaluation product S(t) corresponding to the evaluation time of the evaluation to be tested, and by jointly acting on the multimodal features of the evaluation text, evaluation picture, evaluation time and product detail picture in the evaluation in the detection and extraction of high-quality evaluations, more reasonable high-quality evaluations can be better extracted. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 (1)(2) are the architecture diagram and flow chart of a high-quality live marketing text recognition method of this application.
[0026] Figure 2 This is a flow chart for calculating the confidence w1 of image-text matching in this application.
[0027] Figure 3 (1)(2) is the flow chart for calculating the evaluation product confidence w2 of this application. DETAILED DESCRIPTION
[0027] The specific steps of the implementation process are as follows: by using Octopus and Houyi Collector software to collect e-commerce product data and build a data set, introducing attn-GAN to build a text-to-image model and using the text and images in the data set for training, introducing EMD-LSTM to build a time-evaluation prediction model and using the evaluation time, number of evaluations, and evaluation scores in the data set for training, and using the trained model to calculate the high-quality evaluation confidence of the evaluation to be tested.
[0028] S1. First, from the actual e-commerce platform, use the third-party data collection software Houyi Collector and Octopus Collector to collect consumers' evaluation information on the product as well as the style and detailed picture information of the product.
[0029] S2: The collected product evaluation data and parameter data are screened out to select the evaluation text, evaluation pictures, product styles and parameter details pictures as data set 1 for analyzing and calculating the confidence of the pictures and texts, and the evaluation time, the number of evaluations at each time point and the evaluation score as data set 2 for analyzing and calculating the confidence of the evaluation product. Specifically, step S2 is as follows:
[0030] S3. Divide the image-text related dataset 1 and the evaluation time dataset 2 obtained in S2 into a training set and a test set.
[0031] Specifically, step S3 is as follows:
[0032] S301. According to step S2, the text and pictures of each consumer evaluation data collected are classified to ensure that they are the evaluation text and picture data of the same consumer. The score of each evaluation at each evaluation time is averaged (with the head and tail removed) to obtain the average evaluation score at each time point, and the number of evaluations at each time point is multiplied by the average evaluation score to obtain the evaluation product.
[0033] S302, 80% of data sets 1 and 2 are used as training sets for subsequent input into the constructed Attn-GAN model and EMD-LSTM to train the model and update the model parameters.
[0034] S303, the other 20% of the data set is used as a test set for subsequent input into the updated model for model testing. The purpose is to find the optimal model so that the model has better generalization ability.
[0035] S4, respectively, extract features from the text-generated images, evaluation drawings, and product parameter diagrams using transformers, and calculate the image-text correlation confidence w1 through k-means clustering and cosine similarity, and calculate the evaluation product confidence w2 based on the difference between the actual value of the evaluation time-evaluation product distribution curve and the predicted value obtained through EMD-LSTM. Specifically, step S4 is as follows:
[0036] S5. Use the data set obtained in S3 to input it into the model constructed by S3 and S4 for model training.
[0037] Specifically, step S5 is as follows:
[0038] S501: Input the training set in the data set into the constructed model to update the model parameters.
[0039] S502: After the model parameters are updated, the test set is input for testing to find the best model and evaluate the generalization ability of the model.
[0040] S503. Use the cross entropy loss function to evaluate the model. The calculation formula is as follows:
[0041]
[0042] In the formula represents the true feature classification probability of the data, Indicates the feature classification probability of data prediction. When the cross entropy of the calculation model is controlled above 80%, it means that the GAT model training is completed.
[0043] S6. Use the trained model to calculate the high-quality evaluation confidence of the evaluation to be tested, obtain the high-quality evaluation confidence of the evaluation, and compare the score with the confidence of the lowest high-quality evaluation in the data set to determine whether the evaluation is a high-quality evaluation.
[0044] Specifically, step S6 is as follows:
[0045] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0046] The above is only the most effective implementation scheme of the present invention. It should be pointed out that for ordinary technicians in this technical field, appropriate improvements and modifications can be made without departing from the working principle of the present invention. These improvements and modifications should also be regarded as protection of the present invention.
Claims
1. A high-quality live marketing text recognition method, characterized in that: The process of extracting high-quality reviews is carried out by using the correlation between the text, pictures and product parameter description pictures of consumer product reviews and the abnormal suspicious points of the review time, including the following steps: 1) Based on the existing review information, manually select some high-quality reviews from a long period of time, annotate the features of the review texts and pictures that are identified as high-quality reviews, and divide them into training sets and test sets; the existing review information and review time, review score, and number of reviews are used as the data set and preprocessed; 2) For text analysis of high-quality reviews, extract the word features and sentence features of high-quality review texts, conditionally enhance the sentence features, and then merge them with the random noise vector as the text feature set R1; 3) Constructing a text-generated image network model based on Attn-GAN, the network model includes a pre-trained network and a multi-stage generation network, and the pre-trained network introduces a Transformer module; using the pre-trained Transformer network to extract features of the product image to obtain an image feature vector R2. 4) Input the text feature set R1 and the image feature set R2 in the high-quality evaluation set obtained in step 2 and step 3 into the AttnGAN model respectively to match each text feature and image feature in the high-quality data set, train the generator and the discriminator, and construct a function model Q based on text generation; 5) Use the text-to-image model Q obtained in step 4 to analyze the comment text of each evaluation to be evaluated, generate the comment prediction graph v1, and use the transformer model to extract features from the comment prediction graph v1, comment attachment v2, and product parameter graph v3 to obtain the feature vector {t1, t2, t3}. 6) Using the comment prediction graph feature t1, comment attachment feature t2, and product parameter graph feature t3 obtained in step 5, obtain the intra-family feature t4 of the three graphs v1, v2, and v3 through K-means clustering, and use cosine similarity to calculate the correlation degree matrix {u1, u2, u3} between t4 and t1, t2, and t3 respectively, and obtain the image-text correlation matching confidence w1 of high-quality evaluation; 7) Based on EMD-LSTM, the confidence test of the evaluation product of the evaluation to be tested is performed. The Y-axis of the plane coordinate system is established with the evaluation product (number of evaluations * average score) in this time period, and the evaluation time is the X-axis. The evaluation product-evaluation time distribution function is established, and the evaluation product of the historical evaluation time is fitted through EMD-LSTM to obtain the evaluation time distribution prediction function P, and the evaluation product P(t) at the evaluation time of the evaluation to be tested is predicted. 8) At a certain moment, calculate the function M(m) formed by the difference m between the predicted evaluation product P(t) calculated by the evaluation time distribution prediction model obtained in step 7) and the actual evaluation product S(t) at the evaluation time to be evaluated to calculate the evaluation product confidence w2. 9) The confidence w of a certain consumer evaluation at that moment being a high-quality evaluation is obtained by weighted average of the image-text matching confidence w1 and the time series confidence w2 at a certain time point. By calculating the confidence of the high-quality evaluation data set, the confidence value of the lowest judgment as a high-quality evaluation is obtained.
2. A high-quality live marketing text recognition method according to claim 1, characterized in that: The specific method for determining high-quality reviews in the data set of step 1) is: manually annotating the high-quality reviews based on the existing review information, annotating based on the text features and image features in the high-quality reviews, and dividing the data into training sets and test sets.
3. A high-quality live marketing text recognition method according to claim 1, characterized in that: Through the text-to-image model of step 4), a comment prediction graph can be generated from the comment text and product style information. In steps 5) and 6), the comment prediction graph generated from the comment text and product style text, the original evaluation image, and the product parameter graph are respectively subjected to transform feature extraction and intra-family feature extraction based on k-means clustering, and their correlation matrix is calculated through cosine similarity to obtain the image-text correlation confidence w1 of the comment.
4. A high-quality live marketing text recognition method according to claim 1, characterized in that: Through the step 6), the evaluation product-evaluation time curve is predicted using EMD-LSTM, and the evaluation product confidence w2 is calculated using the function M(m) formed by the difference m between the predicted evaluation product P(t) calculated by the evaluation time distribution prediction model and the actual evaluation product S(t) at the evaluation time of the evaluation to be tested.
5. A high-quality live marketing text recognition method, characterized in that: The total confidence w is the weighted average of the image-related confidence w1 and the evaluation product confidence w2. The multimodal features of the evaluation text, evaluation image, evaluation time and product details in the evaluation are used together to detect and extract high-quality evaluations, and the total confidence value is made between (0, 1) to facilitate intuitive detection and identification.
Citation Information
Patent Citations
Method, apparatus, electronic device and storage medium for analyzing evaluation data of store
CN109299460A
A semi-supervised e-commerce comment sentiment analysis method based on tripartite graphs and cluster analysis
CN109670039A
Online shopping assistant construction method for live marketing commodity quality perception analysis
CN116703509A
Method and apparatus for displaying product review information, electronic device and storage medium
WO2021036250A1
