Content recommendation method and device and electronic equipment

Abstract

The embodiment of the invention provides a content recommendation method and device and electronic equipment, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining the propaganda data, the first evaluation, the second evaluation and the current score of the first evaluation of a to-be-recommended first consumption target; calculating a similarity value between the text information and the image information of the first evaluation and a repetition value between the first evaluation and the second evaluation; correcting the current score of the first evaluation according to the similarity value and the repetition value to obtain a comprehensive score of the first evaluation; under the condition that the comprehensive score exceeds a preset score threshold value, judging the first evaluation as real evaluation data; calculating the similarity between the propaganda data and the true evaluation data of the first consumption target, and determining the propaganda trueness of the first consumption target; and recommending the first consumption target to the first consumer under the condition that the propaganda trueness exceeds a preset truth threshold. According to the embodiment of the invention, a more real consumption target can be recommended and propagandized to consumers.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a content recommendation method, device and electronic device. Background Art

[0002] With the development of the internet, promotional materials for consumer products are becoming increasingly exaggerated, making it difficult to distinguish authenticity. Consumers are increasingly relying on reviews to select consumer products. Search engine recommendation algorithms also rely heavily on these reviews when recommending consumer products. However, some unscrupulous businesses fabricate false positives to mislead consumers, while others even deliberately create false negative reviews to undermine their competitors and attract customers. This phenomenon not only affects consumer decision-making and the recommendation mechanisms of recommendation algorithms, but also undermines fair competition in the market, making it difficult to weed out false advertising from a wide range of consumer products and select genuine ones for recommendation. Summary of the Invention

[0003] The purpose of the embodiments of the present invention is to provide a content recommendation method, device, and electronic device to achieve more realistic consumer goals for consumer recommendations. The specific technical solution is as follows:

[0004] In a first aspect, an embodiment of the present invention provides a content recommendation method, comprising:

[0005] Obtaining promotional data, a first evaluation, a second evaluation, and a current rating of the first evaluation of a first consumer target to be recommended, wherein the first evaluation is an evaluation to be analyzed including text information and image information, and the second evaluation is an evaluation other than the first evaluation;

[0006] Calculating a similarity value between the text information and the image information of the first evaluation, and a repetition value between the first evaluation and the second evaluation, wherein the similarity value indicates a degree of association between the text information and the image information, and the repetition value indicates a degree of repetition between the first evaluation and the second evaluation;

[0007] Modifying the current score of the first evaluation according to the similarity value and the repetition value to obtain a comprehensive score of the first evaluation;

[0008] If the comprehensive score exceeds a preset score threshold, the first evaluation is determined to be true evaluation data;

[0009] Calculating the similarity between the promotional data and the real evaluation data of the first consumer target to determine the authenticity of the promotion of the first consumer target;

[0010] When the publicity authenticity exceeds a preset authenticity threshold, the first consumption target is recommended to the first consumer.

[0011] In one embodiment of the present invention, calculating the similarity between the text information and the image information of the first evaluation, and the repetition value between the first evaluation and the second evaluation, includes:

[0012] Inputting the text information and image information of the first evaluation into a pre-trained similarity judgment model to obtain a similarity value between the text information and the image information of the first evaluation, wherein the similarity judgment model is trained based on positive samples consisting of multiple sample images and corresponding text descriptions, and negative samples consisting of multiple sample images and non-corresponding text descriptions;

[0013] The first evaluation and the second evaluation are input into a pre-trained repeatability judgment model to obtain a repeatability value between the first evaluation and the second evaluation, wherein the repeatability judgment model is trained based on positive samples consisting of multiple repeat samples and negative samples consisting of multiple non-repeat samples.

[0014] In one embodiment of the present invention, inputting the text information and image information of the first evaluation into a pre-trained similarity judgment model to obtain a similarity value between the text information and image information of the first evaluation includes:

[0015] Inputting the text information of the first evaluation into the text extraction network of the similarity judgment model, and inputting the image information of the first evaluation into the image feature extraction network to obtain the text features and image features of the first evaluation;

[0016] The text features and image features of the first evaluation are input into a similarity measurement function to obtain a similarity value between the text information and the image information.

[0017] In one embodiment of the present invention, inputting the first evaluation and the second evaluation into a pre-trained repeatability judgment model to obtain a repeatability value between the first evaluation and the second evaluation includes:

[0018] Inputting the first evaluation and the second evaluation into the text extraction network and the image feature extraction network of the repeatability judgment model to obtain a single text feature and a single image feature of the first evaluation, as well as an overall text feature and an overall image feature of the first evaluation and the second evaluation;

[0019] Inputting the single text feature and the overall text feature into the text repetitiveness judgment module of the repetitiveness judgment model to obtain the text repetitiveness value of the first evaluation and the second evaluation;

[0020] Inputting the single image feature and the overall image feature into the image repeatability judgment module of the repeatability judgment model to obtain the image repeatability values of the first evaluation and the second evaluation;

[0021] Based on the text repetition value and the image repetition value, a repetition value between the first evaluation and the second evaluation is obtained.

[0022] In one embodiment of the present invention, the step of modifying the current score of the first evaluation based on the similarity value and the repetition value to obtain a comprehensive score of the first evaluation includes:

[0023] Calculating a malicious weight of the first evaluation according to the similarity value and the duplication value;

[0024] The current score of the first evaluation is modified based on the malicious weight to obtain a comprehensive score of the first evaluation.

[0025] In one embodiment of the present invention, calculating the malicious weight of the first evaluation based on the similarity value and the duplication value includes:

[0026] The malicious weight is calculated according to the following formula:

[0027]

[0028] Wherein, o is the malicious weight, m is the duplication value, and n is the similarity value;

[0029] The step of modifying the current score of the first evaluation based on the malicious weight to obtain a comprehensive score of the first evaluation includes:

[0030] The comprehensive score of the first evaluation is calculated according to the following formula:

[0031] p=h-(ah)×o

[0032] Wherein, p is the comprehensive score, h is the current score, a is half of the total score of the current score, and o is the malicious weight.

[0033] In one embodiment of the present invention, the calculating the similarity between the promotional data and the real evaluation data of the first consumer target and determining the authenticity of the promotion of the first consumer target includes:

[0034] Performing data cleaning on the promotional data and the real evaluation data to obtain first promotional data and first real evaluation data after removing duplicate data, invalid data, and noise data;

[0035] Performing word segmentation and useless word removal on the text data of the first promotional data and the first real evaluation data to obtain processed text data of the first promotional data and the first real evaluation data;

[0036] When video data exists in the first promotional data and the first real evaluation data, performing key frame extraction on the video data to obtain image data of the first promotional data and the first real evaluation data;

[0037] Inputting the image data of the first promotional data and the first real evaluation data into a feature extraction network to obtain deep features of the image data of the first promotional data and the first real evaluation data;

[0038] Based on the text data processed from the first promotional data and the first real evaluation data, and the depth features of the image data of the first promotional data and the first real evaluation data, the similarity between the promotional data and the real evaluation data of the first consumer target is calculated to determine the authenticity of the promotion of the first consumer target.

[0039] In one embodiment of the present invention, when the publicity authenticity exceeds a preset authenticity threshold, recommending the first consumption target to the first consumer includes:

[0040] When the authenticity of the promotion exceeds a preset authenticity threshold, obtaining a portrait attribute of the first consumer of the consumption target to be recommended, wherein the portrait attribute includes at least one of the basic information, consumption preferences, and consumption level of the first consumer;

[0041] Matching a consumer cluster corresponding to the first consumer based on the first consumer's profile attributes, wherein the consumer cluster is obtained by clustering each of the second consumers based on the evaluation attributes of the second consumers corresponding to the real evaluation data of multiple consumption targets, wherein the evaluation attributes include at least one of sentiment tendency, review topic, review keyword frequency, and rating distribution;

[0042] In a case where the consumer cluster indicates that the first consumer matches the first consumption goal, the first consumption goal is recommended to the first consumer.

[0043] In one embodiment of the present invention, when the consumer cluster indicates that the first consumer matches the first consumption target, recommending the first consumption target to the first consumer includes:

[0044] When the consumer cluster indicates that the first consumer matches the first consumption target, obtaining historical consumption data and real-time recommendation data of the first consumer;

[0045] In a case where the historical consumption data and the real-time recommendation data indicate that the first consumer matches the first consumption goal, the first consumption goal is recommended to the first consumer.

[0046] In a second aspect, an embodiment of the present invention provides a content recommendation device, including:

[0047] a data acquisition module, configured to acquire promotional data, a first evaluation, a second evaluation, and a current rating of a first consumer target to be recommended, wherein the first evaluation is an evaluation to be analyzed including text information and image information, and the second evaluation is an evaluation other than the first evaluation;

[0048] a numerical calculation module, configured to calculate a similarity value between the text information and the image information of the first evaluation, and a repetition value between the first evaluation and the second evaluation, wherein the similarity value indicates a degree of association between the text information and the image information, and the repetition value indicates a degree of repetition between the first evaluation and the second evaluation;

[0049] a score correction module, configured to correct the current score of the first evaluation according to the similarity value and the repetition value to obtain a comprehensive score of the first evaluation;

[0050] An evaluation determination module, configured to determine the first evaluation as true evaluation data if the comprehensive score exceeds a preset score threshold;

[0051] a publicity authenticity determination module, configured to calculate the similarity between the publicity data and the real evaluation data of the first consumer target, and determine the publicity authenticity of the first consumer target;

[0052] The consumption target recommendation module is used to recommend the first consumption target to the first consumer when the authenticity of the publicity exceeds a preset authenticity threshold.

[0053] In one embodiment of the present invention, the numerical calculation module includes:

[0054] a similarity value calculation submodule, configured to input the text information and image information of the first evaluation into a pre-trained similarity judgment model to obtain a similarity value between the text information and image information of the first evaluation, wherein the similarity judgment model is trained based on positive samples consisting of multiple sample images and corresponding text descriptions, and negative samples consisting of multiple sample images and non-corresponding text descriptions;

[0055] The repeatability calculation submodule is used to input the first evaluation and the second evaluation into a pre-trained repeatability judgment model to obtain the repeatability value between the first evaluation and the second evaluation, wherein the repeatability judgment model is trained based on positive samples consisting of multiple repeat samples and negative samples consisting of multiple non-repeat samples.

[0056] In one embodiment of the present invention, the similarity value calculation submodule is specifically configured to:

[0057] Inputting the text information of the first evaluation into the text extraction network of the similarity judgment model, and inputting the image information of the first evaluation into the image feature extraction network to obtain the text features and image features of the first evaluation;

[0058] The text features and image features of the first evaluation are input into a similarity measurement function to obtain a similarity value between the text information and the image information.

[0059] In one embodiment of the present invention, the repetition value calculation submodule is specifically configured to:

[0060] Inputting the first evaluation and the second evaluation into the text extraction network and the image feature extraction network of the repeatability judgment model to obtain a single text feature and a single image feature of the first evaluation, as well as an overall text feature and an overall image feature of the first evaluation and the second evaluation;

[0061] Inputting the single text feature and the overall text feature into the text repetitiveness judgment module of the repetitiveness judgment model to obtain the text repetitiveness value of the first evaluation and the second evaluation;

[0062] Inputting the single image feature and the overall image feature into the image repeatability judgment module of the repeatability judgment model to obtain the image repeatability values of the first evaluation and the second evaluation;

[0063] Based on the text repetition value and the image repetition value, a repetition value between the first evaluation and the second evaluation is obtained.

[0064] In one embodiment of the present invention, the score correction module is specifically configured to:

[0065] Calculating a malicious weight of the first evaluation according to the similarity value and the duplication value;

[0066] The current score of the first evaluation is modified based on the malicious weight to obtain a comprehensive score of the first evaluation.

[0067] In one embodiment of the present invention, the advertisement authenticity determination module is specifically configured to:

[0068] Performing data cleaning on the promotional data and the real evaluation data to obtain first promotional data and first real evaluation data after removing duplicate data, invalid data, and noise data;

[0069] Performing word segmentation and useless word removal on the text data of the first promotional data and the first real evaluation data to obtain processed text data of the first promotional data and the first real evaluation data;

[0070] When video data exists in the first promotional data and the first real evaluation data, performing key frame extraction on the video data to obtain image data of the first promotional data and the first real evaluation data;

[0071] Inputting the image data of the first promotional data and the first real evaluation data into a feature extraction network to obtain deep features of the image data of the first promotional data and the first real evaluation data;

[0072] Based on the text data processed from the first promotional data and the first real evaluation data, and the depth features of the image data of the first promotional data and the first real evaluation data, the similarity between the promotional data and the real evaluation data of the first consumer target is calculated to determine the authenticity of the promotion of the first consumer target.

[0073] In one embodiment of the present invention, the consumption target recommendation module includes:

[0074] A portrait attribute acquisition submodule is configured to acquire, when the authenticity of the promotion exceeds a preset authenticity threshold, portrait attributes of the first consumer to be recommended as a consumption target, wherein the portrait attributes include at least one of the first consumer's basic information, consumption preferences, and consumption level;

[0075] A consumer cluster matching submodule is configured to match a consumer cluster corresponding to the first consumer based on the first consumer's profile attributes, wherein the consumer cluster is obtained by clustering each of the second consumers based on the evaluation attributes of the second consumers corresponding to the real evaluation data of multiple consumption targets, wherein the evaluation attributes include at least one of sentiment tendency, review topic, review keyword frequency, and rating distribution;

[0076] The consumption target recommendation submodule is configured to recommend the first consumption target to the first consumer when the consumer cluster indicates that the first consumer matches the first consumption target.

[0077] In one embodiment of the present invention, the consumption target recommendation submodule is specifically configured to:

[0078] When the consumer cluster indicates that the first consumer matches the first consumption target, obtaining historical consumption data and real-time recommendation data of the first consumer;

[0079] In a case where the historical consumption data and the real-time recommendation data indicate that the first consumer matches the first consumption goal, the first consumption goal is recommended to the first consumer.

[0080] In a third aspect, an embodiment of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0081] Memory for storing computer programs;

[0082] The processor is configured to implement any of the above-mentioned content recommendation methods when executing the program stored in the memory.

[0083] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements any of the above-mentioned content recommendation methods.

[0084] An embodiment of the present invention further provides a computer program product comprising instructions, which, when executed on a computer, enables the computer to execute any of the above-mentioned content recommendation methods.

[0085] Beneficial effects of the embodiments of the present invention:

[0086] The content recommendation method provided by the embodiment of the present invention obtains the promotional data, first evaluation, second evaluation, and current rating of the first evaluation of a first consumer target to be recommended; calculates the similarity between the text information and image information of the first evaluation to obtain the degree of association between the text information and image information in the first evaluation; and calculates the repetition value between the first evaluation and the second evaluation to obtain the degree of repetition between the first evaluation and the second evaluation; and modifies the current rating of the first evaluation based on the similarity and repetition values to obtain a comprehensive rating of the first evaluation. When the degree of association between the text information and the image information is low, or the degree of repetition with the second evaluation is high, the rating of the first evaluation is reduced; otherwise, the rating of the first evaluation is increased, thereby minimizing the rating of false evaluations and increasing the rating of true evaluations. When the comprehensive rating exceeds a preset rating threshold, the first evaluation is determined to be true evaluation data. The similarity between the promotional data and the true evaluation data of the first consumer target is calculated to determine the authenticity of the promotion of the first consumer target; when the authenticity of the promotion exceeds the preset authenticity threshold, it is considered that the degree of false promotion of the first consumer target is low, and the actual situation is more consistent with the promotion, and the first consumer target is recommended to the first consumer, thereby achieving the goal of recommending more authentic promotion to the consumer.

[0087] Of course, it is not necessary to achieve all of the advantages described above simultaneously in order to implement any product or method of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other embodiments can also be obtained based on these drawings.

[0089] Figure 1-1 A schematic diagram of a flow chart of a content recommendation method according to an embodiment of the present invention;

[0090] Figure 1-2 This is an example diagram of a repetition value calculation process according to an embodiment of the present invention;

[0091] Figure 2 This is a possible implementation of step S102 in the embodiment of the present invention;

[0092] Figure 3 This is a possible implementation of step S201 in the embodiment of the present invention;

[0093] Figure 4-1 This is a possible implementation of step S202 in the embodiment of the present invention;

[0094] Figure 4-2This is an example diagram of a comprehensive score calculation process according to an embodiment of the present invention;

[0095] Figure 5 This is a possible implementation of step S103 in the embodiment of the present invention;

[0096] Figure 6 This is a possible implementation of step S105 in the embodiment of the present invention;

[0097] Figure 7 This is a possible implementation of step S106 in the embodiment of the present invention;

[0098] Figure 8-1 This is a possible implementation of step S703 in the embodiment of the present invention;

[0099] Figure 8-2 This is an example diagram of a consumption target recommendation process according to an embodiment of the present invention;

[0100] Figure 9 A schematic structural diagram of a content recommendation device according to an embodiment of the present invention;

[0101] Figure 10 The figure is a structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0102] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field based on the present invention are within the scope of protection of the present invention.

[0103] In related technologies, the promotional copy for each consumer target is often exaggerated, and some reviews may be false. This makes it difficult to eliminate false promotions from a large number of consumer targets and select genuine promotions for consumer targets. To address this issue, embodiments of the present invention provide a content recommendation method, device, and electronic device.

[0104] The content recommendation method according to the embodiment of the present invention is described in detail below.

[0105] See also Figure 1-1 , Figure 1-1 A first flow chart of a content recommendation method provided by an embodiment of the present invention includes:

[0106] Step S101, obtaining the promotional data, first evaluation, second evaluation and current score of the first evaluation of a first consumer target to be recommended;

[0107] The first evaluation is an evaluation to be analyzed including text information and image information, and the second evaluation is an evaluation other than the first evaluation;

[0108] A consumption target refers to any item that needs to be consumed, such as hotels, restaurants, tourist attractions, and merchandise stores. Promotional data refers to the promotional content that the consumption target has posted online, including promotional copy, images, and videos. For example, promotional data can be collected from various internet channels using web crawlers. For example, if the consumption target is a tourist attraction, promotional data can be collected from travel websites and social media. If the consumption target is a hotel, it can be collected through hotel booking aggregation platforms, social media, and search engines.

[0109] The first review can be any review that includes text and image information in the review area of the consumer target on any Internet channel such as a purchase platform or social media. The second review can be any review other than the first review in the review area. The current score of the first review is the result of an initial scoring based on the information richness of the first review. For example, the score is assigned according to the number of images and the number of words in the first review. For example, if the first review has more than three images and more than 15 words, the score is 2. For each additional image or 15 words, the score increases by 1 until the upper limit of the score is reached. For example, the score range can be 1-5 points.

[0110] Step S102, calculating a similarity value between the text information and the image information of the first evaluation, and a repetition value between the first evaluation and the second evaluation;

[0111] Among them, the similarity value represents the degree of association between the text information and the image information. The higher the similarity value, the higher the degree of association, which indicates that the authenticity of the first evaluation is higher. The lower the similarity value, the lower the degree of association, which indicates that the first evaluation is more likely to be a forged false review and the lower the authenticity.

[0112] The repetition value indicates the degree of repetition between the first evaluation and the second evaluation. The higher the repetition value, the more likely the first evaluation is a false review copied from the second evaluation, and the lower its authenticity. The lower the repetition value, the more likely the first evaluation is a real review issued by the consumer, and the higher its authenticity.

[0113] Similarity values can be calculated using similarity algorithms, such as cosine similarity and Euclidean distance. Duplicate values can be calculated using data duplication matching algorithms, such as algorithms that convert data into hash values and then compare hash values, or algorithms that aggregate data and then determine based on aggregate characteristics.

[0114] In one embodiment of the present invention, Figure 2As shown, the above step S102 calculates the similarity value between the text information and the image information of the first evaluation, and the repetition value between the first evaluation and the second evaluation, including:

[0115] Step S201: inputting the text information and image information of the first evaluation into a pre-trained similarity judgment model to obtain a similarity value between the text information and image information of the first evaluation;

[0116] Among them, the similarity judgment model is obtained by training on positive samples consisting of multiple sample images and corresponding text descriptions, each positive sample includes a sample image and a text description corresponding to the image; and negative samples consisting of multiple sample images and non-corresponding text descriptions, each negative sample includes a sample image and a text description that does not correspond to the image.

[0117] Exemplarily, the training sample data set of the similarity judgment model can be one hundred dish names and corresponding dish images from Wikipedia, with each dish name and corresponding dish image being taken as a positive sample, for a total of one hundred positive samples; a dish name and any one of the 99 non-corresponding dish images being taken as a negative sample, for a total of one hundred negative samples.

[0118] Based on the above positive and negative samples, the similarity judgment model is trained comparatively. The training goal is to maximize the similarity measure between positive samples and minimize the similarity measure between negative sample pairs.

[0119] Both positive and negative samples are easy to obtain and have high authenticity, which reduces the difficulty of obtaining samples for model training and improves the efficiency of model training.

[0120] In one embodiment of the present invention, Figure 3 As shown, the above step S201 inputs the text information and image information of the first evaluation into a pre-trained similarity judgment model to obtain the similarity value between the text information and image information of the first evaluation, including:

[0121] Step S301: inputting the text information of the first evaluation into the text extraction network of the similarity judgment model, and inputting the image information of the first evaluation into the image feature extraction network to obtain the text features and image features of the first evaluation;

[0122] Step S302: Input the text features and image features of the first evaluation into a similarity measurement function to obtain a similarity value between the text information and the image information.

[0123] In one example, the text extraction network of the similarity judgment model can use a Sentence-Transformer (a natural language processing model based on the Transformer architecture) connected to a two-layer fully connected neural network to map text sentences into vector representations in a low-dimensional vector space, thereby calculating the semantic similarity between texts. The image feature extraction network can use a Vit (Vision Transformer) connected to a two-layer fully connected neural network. By treating the image as a series of patches and using these patches as input sequences, the encoder learns the image feature representation and extracts image features. Specifically, the output of both the text extraction network and the image feature extraction network in the similarity judgment model are 768-dimensional features. The similarity metric uses a sigmoid function (logistic function). For example, the features extracted by the extraction network are vectors. The distance between the vectors is calculated and input into the sigmoid function. The resulting output represents the similarity score between the two vectors, which ranges from 0 to 1, with values closer to 1 indicating greater similarity and closer to 0 indicating less similarity.

[0124] In an example, the loss function of the similarity judgment model is calculated according to the following formula:

[0125]

[0126] Among them, Loss is the current model loss, (p,q) is the image information and text information of a positive sample, (q,d k ) is the text information and image information of a negative sample, t is the preset smoothing coefficient, N is the total number of samples, and k represents the kth sample.

[0127] By calculating the similarity between text information and image information using a similarity judgment model, the accuracy of similarity value calculation is improved.

[0128] Step S202, inputting the first evaluation and the second evaluation into a pre-trained repeatability judgment model to obtain a repeatability value between the first evaluation and the second evaluation;

[0129] Among them, the repeatability judgment model is based on positive samples composed of multiple repeated samples, and the positive samples can be multiple samples obtained by copying an evaluation including text and images; and negative samples composed of multiple non-repeated samples. The negative samples can be an evaluation including text and images, and multiple samples obtained by copying other evaluations with different text and images from the evaluation multiple times.

[0130] The training sample dataset for the repeatability judgment model can be a set of 100 dishes from Wikipedia and their corresponding images. One of the dishes and their corresponding images is treated as a single sample, replicated 100 times to obtain 100 positive samples. From the 100 samples, a completely different sample (different dish name and image) is randomly selected and replicated 100 times to obtain 100 negative samples.

[0131] Based on the above positive and negative samples, the repeatability judgment model is trained comparatively based on its training sample data set. The training goal is to maximize the similarity measure between positive samples and minimize the similarity measure between negative sample pairs.

[0132] In one example, the text extraction network of the repetitiveness detection model can use Roberta (an improved language model) to extract text features. The image feature extraction network can use Vit (Vision Transformer) connected to a two-layer fully connected neural network to extract image features. Specifically, the output of both the text extraction network and the image feature extraction network of the repetitiveness detection model is a 768-dimensional feature.

[0133] Both positive and negative samples are easy to obtain and have high authenticity, which reduces the difficulty of obtaining samples for model training and improves the efficiency of model training.

[0134] In one embodiment of the present invention, Figure 4-1 As shown, the above step S202 inputs the first evaluation and the second evaluation into a pre-trained repeatability judgment model to obtain the repeatability value between the first evaluation and the second evaluation, including:

[0135] Step S401: Inputting the first evaluation and the second evaluation into the text extraction network and the image feature extraction network of the repeatability judgment model to obtain a single text feature and a single image feature of the first evaluation, as well as an overall text feature and an overall image feature of the first evaluation and the second evaluation;

[0136] Step S402: inputting the single text feature and the overall text feature into the text repetitiveness judgment module of the repetitiveness judgment model to obtain the text repetitiveness value of the first evaluation and the second evaluation;

[0137] Step S403: inputting the single image feature and the overall image feature into the image repeatability judgment module of the repeatability judgment model to obtain the image repeatability values of the first evaluation and the second evaluation;

[0138] Step S404: obtaining a repetition value between the first evaluation and the second evaluation based on the text repetition value and the image repetition value.

[0139] The repeatability judgment model is used to judge the repeatability between a single sample and the entire sample. For example, the repeatability judgment process is shown in the figure below: Figure 4-2 As shown in the figure, overall feature extraction is performed on each sample in the overall sample. The features of each sample are then concatenated and fused. After aligning the overall features with the individual features of each sample, repeatability is evaluated. For example, feature alignment and repeatability evaluation are implemented using two fully connected layers (768*124 and 124*1) followed by a sigmoid function layer.

[0140] In this embodiment, the first evaluation is a single sample, and a single feature of the first evaluation, namely a single text feature and a single image feature, is extracted; the second evaluation is a plurality of evaluations other than the first evaluation, which is a whole sample. The text features and image features of the plurality of evaluations in the second evaluation are extracted, and the overall text features and overall image features are obtained after splicing them separately. The overall text features and the single text features are feature aligned through two fully connected layers of 768*124 plus 124*1 in the text repeatability judgment module, and the text repetition value is calculated through the sigmoid function of the text repetition judgment module; the overall image features and the single image features are feature aligned through two fully connected layers of 768*124 plus 124*1 in the image repetition judgment module. The image repetition value is calculated through the sigmoid function of the image repetition judgment module, and finally the repetition value between the first evaluation and the second evaluation is obtained.

[0141] In an example, the duplicate value between the first and second evaluations is calculated according to the following formula:

[0142] n=(t+p) / 2

[0143] Where n is the repetition value, t is the text repetition value, and p is the image repetition value.

[0144] In an example, the loss function of the repeatability judgment model is calculated according to the following formula:

[0145]

[0146] Among them, Loss is the current model loss, (p,q) is the image information and text information of a positive sample, (q,d k ) is the text information and image information of a negative sample, t is the preset smoothing coefficient, N is the total number of samples, and k represents the kth sample.

[0147] The accuracy of the repeated value calculation is improved by calculating the text repeated value and image repeated value of a single sample and the overall sample through the repeatability judgment model.

[0148] Step S103, modifying the current score of the first evaluation according to the similarity value and the repetition value to obtain a comprehensive score of the first evaluation;

[0149] According to the similarity value and the duplication value, the current score of the first evaluation is revised to obtain a revised and more realistic comprehensive score of the first evaluation.

[0150] Exemplarily, weights can be assigned to the current score based on the similarity value and the repetition value to calculate the comprehensive score of the first evaluation. For example, a positive weight can be assigned to the similarity value, and a negative weight can be assigned to the repetition value, so that the comprehensive score shows a correction trend that the larger the similarity value, the higher the score; the larger the repetition value, the lower the score.

[0151] Step S104: if the comprehensive score exceeds a preset score threshold, the first evaluation is determined to be true evaluation data;

[0152] When the comprehensive score exceeds a preset score threshold (a pre-set value), the first evaluation is determined to be real evaluation data, which means that the first evaluation is a real consumer evaluation, not a forged false evaluation.

[0153] If the comprehensive score does not exceed the preset score threshold, it means that the first evaluation may be a forged false evaluation, which can be manually reviewed and further judged; if the comprehensive score is lower than the preset low score threshold, it means that the first evaluation must be a forged false evaluation, and the first evaluation can be removed from the comment area.

[0154] In an example, when the first consumption target is a delicacy and the first evaluation is a delicacy review, the input includes a textual evaluation and an image evaluation of the delicacy. The comprehensive evaluation process example is as follows: Figure 1-2 As shown, the text processing network and image processing network of the similarity judgment module are used to realize similarity measurement, and the repeatability judgment module is used to perform text repeatability judgment and image repeatability judgment to realize repeatability judgment, and then the comprehensive evaluation module determines whether it is a true evaluation.

[0155] In one example, after determining that the first evaluation is real evaluation data, you can also capture information such as the number of views, number of likes, comment time, user ID who issued the evaluation, and public personal information (such as gender, age, region, etc.) corresponding to the first evaluation.

[0156] Step S105, calculating the similarity between the promotional data and the real evaluation data of the first consumer target, and determining the authenticity of the promotion of the first consumer target;

[0157] The similarity between the promotional data and the real evaluation data of the first consumer target is calculated to determine the publicity authenticity of the first consumer target. The publicity authenticity is the degree of publicity authenticity of the first consumer target. The higher the publicity authenticity, the lower the degree of false publicity that may exist in the first consumer target. Conversely, the higher the false publicity, the higher the degree of false publicity.

[0158] Exemplarily, the level of publicity authenticity may correspond to the value of similarity one to one. For example, when the similarity value is in a certain value range, the publicity authenticity is at a certain level. The higher the similarity, the higher the publicity authenticity.

[0159] Specifically, natural language processing technology can be used to extract key information between promotional data and real evaluation data. For example, when the consumption target is a tourist attraction, information such as the characteristics of the attraction, event arrangements, accommodation conditions, and service quality can be extracted. Semantic similarity calculation methods (such as cosine similarity and Jaccard (a similarity index coefficient) similarity) can be used to compare the consistency of promotional data with real evaluation data. For example, if the promotional copy mentions "the best beach" but the evaluation mentions "serious seawater pollution", the calculated similarity between the two texts is low, which can be determined to be false advertising and low authenticity of the promotion.

[0160] Step S106: When the publicity authenticity exceeds a preset authenticity threshold, recommend the first consumption target to the first consumer.

[0161] When the authenticity of the publicity exceeds the preset true threshold, it means that the degree of false publicity of the first consumer target is low, and the first consumer target is actually more in line with its advertised content, and the first consumer target is recommended to the first consumer. When there are multiple first consumer targets, they are arranged into a recommendation list according to the level of publicity authenticity. The higher the publicity authenticity, the higher it is in the recommendation list. The preset true threshold can be set according to actual needs. For example, the publicity authenticity includes levels 1-5, and the preset true threshold is level 3. In the case where the publicity authenticity does not exceed the preset true threshold, it indicates that there may be false publicity, and its recommendation weight can be reduced so that it is at the back of the recommendation list. When the publicity authenticity is lower than the preset false threshold (for example, level 1), it is directly removed from the recommendation list, and the consumer target is not recommended to the consumer.

[0162] As can be seen from the above, the content recommendation method provided by the embodiment of the present invention obtains the promotional data, first evaluation, second evaluation and current score of the first evaluation of the first consumer target to be recommended; calculates the similarity between the text information and image information of the first evaluation to obtain the correlation between the text information and image information in the first evaluation; and the repetition value between the first evaluation and the second evaluation to obtain the repetition degree between the first evaluation and the second evaluation; based on the similarity and repetition values, modifies the current score of the first evaluation to obtain a comprehensive score of the first evaluation. When the correlation between the text information and the image information is low or the repetition degree with the second evaluation is high, the score of the first evaluation is reduced; otherwise, the score of the first evaluation is increased, thereby reducing the score of false evaluations as much as possible and making the score of true evaluations higher. When the comprehensive score exceeds a preset score threshold, the first evaluation is determined to be true evaluation data. The similarity between the promotional data and the true evaluation data of the first consumer target is calculated to determine the authenticity of the promotion of the first consumer target; when the authenticity of the promotion exceeds the preset authenticity threshold, it is considered that the degree of false promotion of the first consumer target is low and the actual situation is more consistent with the promotion situation. The first consumer target is recommended to the first consumer, thereby achieving the goal of recommending a more authentic promotion to the consumer.

[0163] In one embodiment of the present invention, Figure 5 As shown, the above step S103 modifies the current score of the first evaluation according to the similarity value and the repetition value to obtain a comprehensive score of the first evaluation, including:

[0164] Step S501, calculating the malicious weight of the first evaluation according to the similarity value and the duplication value;

[0165] Step S502: Modify the current score of the first evaluation based on the malicious weight to obtain a comprehensive score of the first evaluation.

[0166] The higher the malicious weight, the greater the reverse correction of the evaluation.

[0167] In one embodiment of the present invention, the calculation of the malicious weight of the first evaluation based on the similarity value and the duplication value includes:

[0168] The malicious weight is calculated according to the following formula:

[0169]

[0170] Wherein, o is the malicious weight, m is the duplication value, and n is the similarity value;

[0171] The step of modifying the current score of the first evaluation based on the malicious weight to obtain a comprehensive score of the first evaluation includes:

[0172] Calculate the comprehensive score of the first evaluation according to the following formula:

[0173] p = h - (a - h) × o

[0174] Where p is the comprehensive score, h is the current score, a is half of the total score of the current score (for example, when the score is 1 - 5, a is 2.5), and o is the malicious weight.

[0175] As can be seen from the above, the content recommendation method provided by the embodiments of the present invention calculates the malicious weight of the first evaluation according to the similarity value and the repetition value, modifies the current score of the first evaluation based on the malicious weight, and obtains the comprehensive score of the first evaluation based on the principle that the higher the malicious weight, the greater the reverse correction strength for the evaluation, which can more accurately reflect the authenticity and objectivity of the first evaluation.

[0176] In one embodiment of the present invention, as Figure 6 shown, the above step S105 calculates the similarity between the promotional data and the real evaluation data of the first consumption target, and determines the promotional authenticity of the first consumption target, including:

[0177] Step S601, perform data cleaning on the promotional data and the real evaluation data to obtain the first promotional data and the first real evaluation data after removing duplicate data, invalid data, and noise data;

[0178] Exemplarily, use regular expressions to remove HTML (HyperText Markup Language) tags, special characters, and advertising information.

[0179] Step S602, perform word segmentation processing and useless word removal processing on the text data of the first promotional data and the first real evaluation data to obtain the processed text data of the first promotional data and the first real evaluation data;

[0180] Exemplarily, perform word segmentation processing on the text data. For Chinese, the jieba (a Python Chinese word segmentation library) library can be used, and for English, the tokenizer in the NLTK (Natural Language Toolkit) library can be used. Then remove stop words (such as "de", "le", "zai", etc.) and punctuation marks and other useless words to facilitate subsequent text analysis.

[0181] Step S603, in the case where there is video data in the first promotional data and the first real evaluation data, extract key frames from the video data to obtain the image data of the first promotional data and the first real evaluation data;

[0182] For example, a key frame may be extracted from the video data every second and converted into image data, or the video may be transcoded, cropped, and the like by using a video processing library (OpenCV) to obtain image data of the key frame.

[0183] Step S604: inputting the image data of the first promotional data and the first real evaluation data into a feature extraction network to obtain deep features of the image data of the first promotional data and the first real evaluation data;

[0184] The original image data and the image data converted from the video data are formatted to adapt to the feature extraction network. For example, they can be converted to JPG format and resized to a uniform size. The feature extraction network can use an image analysis model, such as a pre-trained convolutional neural network (such as VGG or ResNet (common neural network models)), to extract deep features from the image data.

[0185] Step S605, based on the text data processed from the first promotional data and the first real evaluation data, and the depth features of the image data of the first promotional data and the first real evaluation data, calculate the similarity between the promotional data and the real evaluation data of the first consumer target, and determine the promotional authenticity of the first consumer target.

[0186] The similarity of the text data and the image data in the first promotional data and the first real evaluation data is calculated respectively by a similarity algorithm, thereby determining the promotional authenticity of the first consumer target.

[0187] From the above, it can be seen that the content recommendation method provided by the embodiment of the present invention also pre-processes the text data and image data separately before calculating the similarity between the first promotional data and the first real evaluation data, including removing useless information, adjusting the data format, etc., thereby improving the efficiency of subsequent data processing and determination of the authenticity of the promotion.

[0188] In one embodiment of the present invention, Figure 7 As shown, the above step S106 recommends the first consumption target to the first consumer when the authenticity of the promotion exceeds the preset authenticity threshold, including:

[0189] Step S701, when the authenticity of the promotion exceeds a preset authenticity threshold, obtaining a portrait attribute of the first consumer of the consumption target to be recommended;

[0190] The portrait attributes include at least one of the first consumer's basic information (such as age distribution and gender ratio), consumption preferences (such as favorite types of attractions and travel methods), and consumption level;

[0191] Step S702: matching a consumer cluster corresponding to the first consumer based on the portrait attributes of the first consumer;

[0192] The consumer clustering is obtained by clustering each of the second consumers according to the evaluation attributes of the second consumers corresponding to the real evaluation data of multiple consumption targets, wherein the evaluation attributes include at least one of sentiment tendency, review topic, review keyword frequency, and rating distribution;

[0193] Step S703 : When the consumer cluster indicates that the first consumer matches the first consumption goal, recommend the first consumption goal to the first consumer.

[0194] The system obtains the real evaluation data of the second consumers corresponding to multiple consumption targets, then obtains all the evaluations issued by each second consumer, extracts the characteristics of these evaluations, and obtains the evaluation attributes of each second consumer. For example, it extracts the sentiment tendency of the review (for example, using sentiment analysis algorithms to analyze the sentiment polarity of each review and obtain positive, negative, and neutral classification results), the review topics (using topic modeling algorithms such as LDA (Latent Dirichlet Allocation) to extract the themes appearing in the reviews, such as natural landscapes, history and culture, food, accommodation, etc.), the frequency of keywords in the reviews, and the distribution of ratings.

[0195] A clustering algorithm (such as the k-means (an unsupervised learning algorithm) clustering algorithm) is used to cluster the second consumers according to their evaluation attributes. For example, each evaluation attribute can be used as a feature dimension to select the number of clusters (k value), and the optimal choice of k value is determined by the elbow rule, that is, the sum of squared clustering errors (SSE) under different k values is calculated, and the k value at the inflection point is selected as the optimal value. For example, for a group of tourist consumers, after k-means clustering, three clustering results are obtained: the first category is tourists who like natural scenery, and their comments frequently contain keywords such as "mountains and waters" and "scenery", and their emotional tendencies are mostly positive; the second category is tourists who pay attention to food, and their comments are mostly evaluations of local food, and the scores are relatively high; the third category is tourists who focus on accommodation experience, and their comments often mention hotel facilities, service levels, etc.

[0196] Analyze the group characteristics of each consumer cluster to generate corresponding profile attributes. For example, for nature lovers, the profile attributes may include an average age of 25-35, 60% male, primarily from coastal cities, and a preference for self-driving tours.

[0197] Obtain the portrait attributes of the first consumer to be recommended as the consumption target, and cluster consumers with similar portrait attributes as the matching results of the first consumer. For example, if the basic information of the first consumer is a 30-year-old male from a coastal city, and his consumption preference is to like self-driving tours, then the matching result can be a nature lover.

[0198] When the consumer cluster of the first consumer matches the first consumption target, that is, the consumer cluster of the first consumer matches the type of the consumption target, for example, the first consumer is a natural scenery enthusiast, and the promotional data and real evaluation data of the first consumption target both indicate that it is a tourist attraction with natural scenery, then the first consumption target is recommended to the first consumer.

[0199] In one example, when there are multiple first consumers and multiple first consumption targets, a collaborative filtering algorithm (such as consumer-based collaborative filtering, consumption target-based collaborative filtering) or a content-based recommendation algorithm can be used to recommend personalized consumption targets for different types of consumers.

[0200] From the above, it can be seen that the content recommendation method provided by the embodiment of the present invention, when the publicity authenticity of the first consumer target exceeds the preset authenticity threshold, also determines whether the first consumer and the first consumer target match through the portrait attributes of the first consumer. If they match, the first consumer target is recommended to the first consumer, thereby further improving the accuracy of the consumer target recommendation.

[0201] In one embodiment of the present invention, Figure 8-1 As shown, the above step S703, when the consumer cluster indicates that the first consumer matches the first consumption target, recommends the first consumption target to the first consumer, including:

[0202] Step S801: when the consumer cluster indicates that the first consumer matches the first consumption target, obtaining the first consumer's historical consumption data and real-time recommendation data;

[0203] Step S802 : recommending the first consumption target to the first consumer when the historical consumption data and the real-time recommendation data indicate that the first consumer matches the first consumption target.

[0204] Historical consumption data may include a consumer's past consumption records, browsing history, and favorites records, while real-time recommendation data may include a consumer's current search keyword, location information, and the like. For example, historical consumption data indicates a match if the first consumer's consumption records do not include the first consumer target, but the first consumer target is included in their favorites. Real-time recommendation data indicates a match if, for example, the current search keyword is consistent with the first consumer target's type, or the location information is relatively close to the first consumer target. If both the historical consumption data and the real-time recommendation data indicate a match between the first consumer and the first consumer target, the first consumer target will be recommended to the first consumer.

[0205] As can be seen from the above, the content recommendation method provided by the embodiment of the present invention further determines whether the first consumer matches the first consumption target based on historical consumption data and real-time recommendation data, thereby further improving the relevance and accuracy of consumption target recommendations.

[0206] In one example, it is also possible to collect consumer feedback data on the recommendation results, such as the click-through rate, conversion rate, favorable comment rate, satisfaction survey data of the first consumer for the first consumer target, and further adjust the consumer target recommendation based on the feedback. For example, if the click-through rate and conversion rate are low, the recommendation weight of the first consumer target in the recommendation list is reduced; if the favorable comment rate is low, the publicity authenticity of the first consumer target is reduced. For example, it is also possible to construct a false propaganda detection model based on historical data and machine learning algorithms (such as logistic regression, support vector machine, etc.), input the propaganda data and feedback data into the model, output the authenticity evaluation results of the consumer target (for example, different confidence levels for authenticity, uncertainty, and falsehood), and adjust the publicity authenticity of the first consumer target again based on the authenticity evaluation results.

[0207] In one example, the above content recommendation method can be executed by a consumption target recommendation system, as follows Figure 8-2 The process example diagram shown is implemented. For example, when recommending tourist attractions to tourists, the system first collects data such as the attraction's promotional copy and evaluation data. After data preprocessing, it evaluates the authenticity of the promotion and provides feedback on the authenticity. Furthermore, multiple tourists are clustered to generate user profiles. Personalized recommendations for tourist attractions are generated based on the profiles of the tourists to be recommended. Feedback from tourists on the recommended content, such as "not interested" or "not recommending in the future," can also be collected to further optimize the recommendations. For example, if a user selects "not interested," the number of recommendations can be reduced, and if a user selects "not recommending in the future," the content can be deleted from the recommendation queue.

[0208] See also Figure 9 , an embodiment of the present invention further provides a structural diagram of a content recommendation device, including:

[0209] Data acquisition module 901 is used to acquire promotional data, a first evaluation, a second evaluation, and a current rating of the first evaluation of a first consumer target to be recommended, wherein the first evaluation is the evaluation to be analyzed including text information and image information, and the second evaluation is the evaluation other than the first evaluation;

[0210] a numerical calculation module 902 for calculating a similarity value between the text information and the image information of the first evaluation, and a repetition value between the first evaluation and the second evaluation, wherein the similarity value indicates the degree of association between the text information and the image information, and the repetition value indicates the degree of repetition between the first evaluation and the second evaluation;

[0211] A score correction module 903 is configured to correct the current score of the first evaluation based on the similarity value and the repetition value to obtain a comprehensive score of the first evaluation;

[0212] An evaluation determination module 904 is configured to determine the first evaluation as true evaluation data if the comprehensive score exceeds a preset score threshold;

[0213] The publicity authenticity determination module 905 is used to calculate the similarity between the publicity data and the real evaluation data of the first consumer target, and determine the publicity authenticity of the first consumer target;

[0214] The consumption target recommendation module 906 is configured to recommend the first consumption target to the first consumer when the publicity authenticity exceeds a preset authenticity threshold.

[0215] As can be seen from the above, the content recommendation device provided by the embodiment of the present invention obtains the promotional data, first evaluation, second evaluation and current score of the first evaluation of the first consumer target to be recommended; calculates the similarity between the text information and image information of the first evaluation to obtain the correlation between the text information and image information in the first evaluation; and the repetition value between the first evaluation and the second evaluation to obtain the repetition degree between the first evaluation and the second evaluation; based on the similarity and repetition values, modifies the current score of the first evaluation to obtain a comprehensive score of the first evaluation. When the correlation between the text information and the image information is low or the repetition degree with the second evaluation is high, the score of the first evaluation is reduced; otherwise, the score of the first evaluation is increased, thereby minimizing the score of false evaluations and increasing the score of true evaluations. When the comprehensive score exceeds a preset score threshold, the first evaluation is determined to be true evaluation data. The similarity between the promotional data and the true evaluation data of the first consumer target is calculated to determine the authenticity of the promotion of the first consumer target; when the authenticity of the promotion exceeds the preset authenticity threshold, it is considered that the false promotion of the first consumer target is low and the actual situation is more consistent with the promotion. The first consumer target is recommended to the first consumer, thereby achieving the goal of recommending a more authentic promotion to the consumer.

[0216] In one embodiment of the present invention, the numerical calculation module 902 includes:

[0217] a similarity value calculation submodule, configured to input the text information and image information of the first evaluation into a pre-trained similarity judgment model to obtain a similarity value between the text information and image information of the first evaluation, wherein the similarity judgment model is trained based on positive samples consisting of multiple sample images and corresponding text descriptions, and negative samples consisting of multiple sample images and non-corresponding text descriptions;

[0218] The repeatability calculation submodule is used to input the first evaluation and the second evaluation into a pre-trained repeatability judgment model to obtain the repeatability value between the first evaluation and the second evaluation, wherein the repeatability judgment model is trained based on positive samples consisting of multiple repeat samples and negative samples consisting of multiple non-repeat samples.

[0219] In one embodiment of the present invention, the similarity value calculation submodule is specifically configured to:

[0220] Inputting the text information of the first evaluation into the text extraction network of the similarity judgment model, and inputting the image information of the first evaluation into the image feature extraction network to obtain the text features and image features of the first evaluation;

[0221] The text features and image features of the first evaluation are input into a similarity measurement function to obtain a similarity value between the text information and the image information.

[0222] In one embodiment of the present invention, the repetition value calculation submodule is specifically configured to:

[0223] Inputting the first evaluation and the second evaluation into the text extraction network and the image feature extraction network of the repeatability judgment model to obtain a single text feature and a single image feature of the first evaluation, as well as an overall text feature and an overall image feature of the first evaluation and the second evaluation;

[0224] Inputting the single text feature and the overall text feature into the text repetitiveness judgment module of the repetitiveness judgment model to obtain the text repetitiveness value of the first evaluation and the second evaluation;

[0225] Inputting the single image feature and the overall image feature into the image repeatability judgment module of the repeatability judgment model to obtain the image repeatability values of the first evaluation and the second evaluation;

[0226] Based on the text repetition value and the image repetition value, a repetition value between the first evaluation and the second evaluation is obtained.

[0227] In one embodiment of the present invention, the score correction module 903 is specifically configured to:

[0228] Calculating a malicious weight of the first evaluation according to the similarity value and the duplication value;

[0229] The current score of the first evaluation is modified based on the malicious weight to obtain a comprehensive score of the first evaluation.

[0230] From the above, it can be seen that the content recommendation device provided by the embodiment of the present invention calculates the malicious weight of the first evaluation based on the similarity value and the duplication value, and corrects the current score of the first evaluation based on the malicious weight. Based on the principle that the higher the malicious weight, the greater the reverse correction force of the evaluation, the comprehensive score of the first evaluation is obtained, which can more accurately reflect the authenticity and objectivity of the first evaluation.

[0231] In one embodiment of the present invention, the advertisement authenticity determination module 905 is specifically configured to:

[0232] Performing data cleaning on the promotional data and the real evaluation data to obtain first promotional data and first real evaluation data after removing duplicate data, invalid data, and noise data;

[0233] Performing word segmentation and useless word removal on the text data of the first promotional data and the first real evaluation data to obtain processed text data of the first promotional data and the first real evaluation data;

[0234] When video data exists in the first promotional data and the first real evaluation data, performing key frame extraction on the video data to obtain image data of the first promotional data and the first real evaluation data;

[0235] Inputting the image data of the first promotional data and the first real evaluation data into a feature extraction network to obtain deep features of the image data of the first promotional data and the first real evaluation data;

[0236] Based on the text data processed from the first promotional data and the first real evaluation data, and the depth features of the image data of the first promotional data and the first real evaluation data, the similarity between the promotional data and the real evaluation data of the first consumer target is calculated to determine the authenticity of the promotion of the first consumer target.

[0237] From the above, it can be seen that the content recommendation device provided by the embodiment of the present invention also pre-processes the text data and image data separately before calculating the similarity between the first promotional data and the first real evaluation data, including removing useless information, adjusting the data format, etc., thereby improving the efficiency of subsequent data processing and determination of the authenticity of the promotion.

[0238] In one embodiment of the present invention, the consumption target recommendation module 906 includes:

[0239] A portrait attribute acquisition submodule is configured to acquire, when the authenticity of the promotion exceeds a preset authenticity threshold, portrait attributes of the first consumer to be recommended as a consumption target, wherein the portrait attributes include at least one of the first consumer's basic information, consumption preferences, and consumption level;

[0240] A consumer cluster matching submodule is configured to match a consumer cluster corresponding to the first consumer based on the first consumer's profile attributes, wherein the consumer cluster is obtained by clustering each of the second consumers based on the evaluation attributes of the second consumers corresponding to the real evaluation data of multiple consumption targets, wherein the evaluation attributes include at least one of sentiment tendency, review topic, review keyword frequency, and rating distribution;

[0241] The consumption target recommendation submodule is configured to recommend the first consumption target to the first consumer when the consumer cluster indicates that the first consumer matches the first consumption target.

[0242] From the above, it can be seen that the content recommendation device provided by the embodiment of the present invention, when the publicity authenticity of the first consumer target exceeds the preset authenticity threshold, also determines whether the first consumer and the first consumer target match through the portrait attributes of the first consumer. If they match, the first consumer target is recommended to the first consumer, thereby further improving the accuracy of the consumer target recommendation.

[0243] In one embodiment of the present invention, the consumption target recommendation submodule is specifically configured to:

[0244] When the consumer cluster indicates that the first consumer matches the first consumption target, obtaining historical consumption data and real-time recommendation data of the first consumer;

[0245] In a case where the historical consumption data and the real-time recommendation data indicate that the first consumer matches the first consumption goal, the first consumption goal is recommended to the first consumer.

[0246] As can be seen from the above, the content recommendation device provided by the embodiment of the present invention further determines whether the first consumer matches the first consumption target based on historical consumption data and real-time recommendation data, thereby further improving the relevance and accuracy of consumption target recommendations.

[0247] The embodiment of the present invention further provides an electronic device, such as Figure 10 As shown, it includes a processor 1001 , a communication interface 1002 , a memory 1003 and a communication bus 1004 , wherein the processor 1001 , the communication interface 1002 , and the memory 1003 communicate with each other via the communication bus 1004 .

[0248] Memory 1003, used for storing computer programs;

[0249] The processor 1001 is configured to implement any of the above-mentioned content recommendation methods when executing the program stored in the memory 1003 .

[0250] The communication bus mentioned in the electronic device mentioned above may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. This communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, only one thick line is used in the figure, but this does not mean that there is only one bus or only one type of bus.

[0251] The communication interface is used for communication between the above electronic device and other devices.

[0252] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage. Alternatively, the memory may be at least one storage device located away from the processor.

[0253] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0254] In another embodiment of the present invention, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the steps of any of the above-mentioned content recommendation methods are implemented.

[0255] In another embodiment of the present invention, a computer program product including instructions is provided. When the computer program product is executed on a computer, the computer is enabled to execute any one of the content recommendation methods in the above embodiments.

[0256] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0257] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0258] Each embodiment in this specification is described in a related manner. Similar portions between the embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences from other embodiments. In particular, the device embodiments are generally similar to the method embodiments, so their description is relatively simple. For related portions, refer to the description of the method embodiments.

[0259] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A content recommendation method, characterized in that: include: Obtaining promotional data, a first evaluation, a second evaluation, and a current rating of the first evaluation of a first consumer target to be recommended, wherein the first evaluation is an evaluation to be analyzed including text information and image information, and the second evaluation is an evaluation other than the first evaluation; Calculating a similarity value between the text information and the image information of the first evaluation, and a repetition value between the first evaluation and the second evaluation, wherein the similarity value indicates a degree of association between the text information and the image information, and the repetition value indicates a degree of repetition between the first evaluation and the second evaluation; Modifying the current score of the first evaluation according to the similarity value and the repetition value to obtain a comprehensive score of the first evaluation; If the comprehensive score exceeds a preset score threshold, the first evaluation is determined to be true evaluation data; Calculating the similarity between the promotional data and the real evaluation data of the first consumer target to determine the authenticity of the promotion of the first consumer target; When the publicity authenticity exceeds a preset authenticity threshold, the first consumption target is recommended to the first consumer.

2. The method according to claim 1, characterized in that The calculating of the similarity between the text information and the image information of the first evaluation, and the repetition value between the first evaluation and the second evaluation, includes: Inputting the text information and image information of the first evaluation into a pre-trained similarity judgment model to obtain a similarity value between the text information and the image information of the first evaluation, wherein the similarity judgment model is trained based on positive samples consisting of multiple sample images and corresponding text descriptions, and negative samples consisting of multiple sample images and non-corresponding text descriptions; The first evaluation and the second evaluation are input into a pre-trained repeatability judgment model to obtain a repeatability value between the first evaluation and the second evaluation, wherein the repeatability judgment model is trained based on positive samples consisting of multiple repeat samples and negative samples consisting of multiple non-repeat samples.

3. The method according to claim 2, characterized in that Inputting the text information and image information of the first evaluation into a pre-trained similarity judgment model to obtain a similarity value between the text information and image information of the first evaluation includes: Inputting the text information of the first evaluation into the text extraction network of the similarity judgment model, and inputting the image information of the first evaluation into the image feature extraction network to obtain the text features and image features of the first evaluation; The text features and image features of the first evaluation are input into a similarity measurement function to obtain a similarity value between the text information and the image information.

4. The method according to claim 2, characterized in that The step of inputting the first evaluation and the second evaluation into a pre-trained repeatability judgment model to obtain a repeatability value between the first evaluation and the second evaluation includes: Inputting the first evaluation and the second evaluation into the text extraction network and the image feature extraction network of the repeatability judgment model to obtain a single text feature and a single image feature of the first evaluation, as well as an overall text feature and an overall image feature of the first evaluation and the second evaluation; Inputting the single text feature and the overall text feature into the text repetitiveness judgment module of the repetitiveness judgment model to obtain the text repetitiveness value of the first evaluation and the second evaluation; Inputting the single image feature and the overall image feature into the image repeatability judgment module of the repeatability judgment model to obtain the image repeatability values of the first evaluation and the second evaluation; Based on the text repetition value and the image repetition value, a repetition value between the first evaluation and the second evaluation is obtained.

5. The method according to claim 1, wherein The step of modifying the current score of the first evaluation according to the similarity value and the repetition value to obtain a comprehensive score of the first evaluation includes: Calculating a malicious weight of the first evaluation according to the similarity value and the duplication value; The current score of the first evaluation is modified based on the malicious weight to obtain a comprehensive score of the first evaluation.

6. The method according to claim 1, characterized in that Calculating the malicious weight of the first evaluation according to the similarity value and the duplication value includes: The malicious weight is calculated according to the following formula: Wherein, o is the malicious weight, m is the duplication value, and n is the similarity value; The step of modifying the current score of the first evaluation based on the malicious weight to obtain a comprehensive score of the first evaluation includes: The comprehensive score of the first evaluation is calculated according to the following formula: p=h-(ah)×o Wherein, p is the comprehensive score, h is the current score, a is half of the total score of the current score, and o is the malicious weight.

7. The method according to claim 1, characterized in that The calculating the similarity between the promotional data and the real evaluation data of the first consumer target and determining the authenticity of the promotion of the first consumer target includes: Performing data cleaning on the promotional data and the real evaluation data to obtain first promotional data and first real evaluation data after removing duplicate data, invalid data, and noise data; Performing word segmentation and useless word removal on the text data of the first promotional data and the first real evaluation data to obtain processed text data of the first promotional data and the first real evaluation data; When video data exists in the first promotional data and the first real evaluation data, performing key frame extraction on the video data to obtain image data of the first promotional data and the first real evaluation data; Inputting the image data of the first promotional data and the first real evaluation data into a feature extraction network to obtain deep features of the image data of the first promotional data and the first real evaluation data; Based on the text data processed from the first promotional data and the first real evaluation data, and the depth features of the image data of the first promotional data and the first real evaluation data, the similarity between the promotional data and the real evaluation data of the first consumer target is calculated to determine the authenticity of the promotion of the first consumer target.

8. The method according to claim 1, characterized in that When the publicity authenticity exceeds a preset authenticity threshold, recommending the first consumption target to the first consumer includes: When the authenticity of the promotion exceeds a preset authenticity threshold, obtaining a portrait attribute of the first consumer of the consumption target to be recommended, wherein the portrait attribute includes at least one of the basic information, consumption preferences, and consumption level of the first consumer; Matching a consumer cluster corresponding to the first consumer based on the first consumer's profile attributes, wherein the consumer cluster is obtained by clustering each of the second consumers based on the evaluation attributes of the second consumers corresponding to the real evaluation data of multiple consumption targets, wherein the evaluation attributes include at least one of sentiment tendency, review topic, review keyword frequency, and rating distribution; In a case where the consumer cluster indicates that the first consumer matches the first consumption goal, the first consumption goal is recommended to the first consumer.

9. The method according to claim 8, characterized in that When the consumer cluster indicates that the first consumer matches the first consumption target, recommending the first consumption target to the first consumer includes: When the consumer cluster indicates that the first consumer matches the first consumption target, obtaining historical consumption data and real-time recommendation data of the first consumer; In a case where the historical consumption data and the real-time recommendation data indicate that the first consumer matches the first consumption goal, the first consumption goal is recommended to the first consumer.

10. A content recommendation device, characterized in that: include: a data acquisition module, configured to acquire promotional data, a first evaluation, a second evaluation, and a current rating of a first consumer target to be recommended, wherein the first evaluation is an evaluation to be analyzed including text information and image information, and the second evaluation is an evaluation other than the first evaluation; a numerical calculation module, configured to calculate a similarity value between the text information and the image information of the first evaluation, and a repetition value between the first evaluation and the second evaluation, wherein the similarity value indicates a degree of association between the text information and the image information, and the repetition value indicates a degree of repetition between the first evaluation and the second evaluation; a score correction module, configured to correct the current score of the first evaluation according to the similarity value and the repetition value to obtain a comprehensive score of the first evaluation; An evaluation determination module, configured to determine the first evaluation as true evaluation data if the comprehensive score exceeds a preset score threshold; a publicity authenticity determination module, configured to calculate the similarity between the publicity data and the real evaluation data of the first consumer target, and determine the publicity authenticity of the first consumer target; The consumption target recommendation module is used to recommend the first consumption target to the first consumer when the authenticity of the publicity exceeds a preset authenticity threshold.

11. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; Memory for storing computer programs; A processor, configured to implement the method steps described in any one of claims 1 to 9 when executing a program stored in a memory.

12. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method steps of any one of claims 1 to 9 are implemented.

Images