Comment analysis system for commodity target users and implementation method thereof

Abstract

The invention relates to the technical field of digital data processing, and discloses a comment analysis system of a commodity target user and an implementation method of the comment analysis system, information files of commodities to be analyzed are received to form a folder of the commodities to be analyzed, so that corresponding comment analysis tasks can be created in the system according to the commodities to be analyzed and a network platform for obtaining a comment source, and the comment analysis task is used for analyzing the commodities to be analyzed. The method comprises the following steps: clearly analyzing a target and a data range, identifying a use scene and a user pain point mentioned in comments, summarizing and obtaining high-frequency common viewpoints, judging an emotional tendency, carrying out statistical classification on the common viewpoints, forming structured data, extracting common characteristics strongly related to a commodity, extracting common points concerned by a user, and finding out a core selling point of the product. And the analysis result can directly serve content creation, a data-driven closed loop is formed, the analysis result is displayed in forms of a chart statistical panel and the like, and users are supported to view commodity core selling points, user feedback distribution, high-frequency keywords and export reports or marketing suggestions.

Description

Technical Field

[0001] This invention relates to a product target user review analysis system and its implementation method, belonging to the field of digital data processing technology. Background Technology

[0002] With the increase in e-commerce on internet platforms, the amount of data generated by e-commerce transactions via the internet has also increased accordingly. Currently, mainstream e-commerce platforms use sampling and statistical methods to collect the core selling points of e-commerce products. However, this method is inefficient when dealing with massive amounts of reviews and can only provide simple statistics such as positive review rates. These sampled data are prone to noise, incompleteness, or errors, easily leading to distorted statistical results. For example, fake accounts can interfere with the analysis, causing misjudgments. Traditional e-commerce data sampling is easily interfered with by irrelevant reviews, making it difficult to identify core product characteristics from a vast amount of information, and thus cannot support precise marketing and product iteration. This makes it difficult for e-commerce platforms relying on traditional after-sales data statistics to accurately grasp the core selling points of products.

[0003] Therefore, it is necessary to propose a product review analysis system and its implementation method, addressing the issue that traditional e-commerce product after-sales data statistics are insufficient to reflect the core selling points of products. Summary of the Invention

[0004] This invention provides a product target user review analysis system and its implementation method, which can solve the problem that statistical results are difficult to reflect the core selling points of a product.

[0005] This invention provides a method for analyzing product reviews from target users, comprising: Receive information files of the products to be analyzed to form a folder of products to be analyzed; Select a product in the folder of products to be analyzed; Obtain the selected product data and derived product data to form the product search target; Based on the product search objectives and sales platform, establish a product search task; Using a product search task, retrieve review data from each sales platform; Perform data cleaning on each comment to obtain cleaned comment data; Incorporate the comment data into the trained semantic analysis model of product selling points; Obtain conclusions from product attribute and customer feedback analysis of product selling points; Return to the folder containing the products to be analyzed and select one product, repeating this process until all products have been selected.

[0006] This invention provides a product target user review analysis system, comprising: The server is used to execute the product target user review analysis method described above; The memory is connected in communication with the server.

[0007] This invention provides a product review analysis system and its implementation method for target users. By receiving information files of the products to be analyzed, a folder of products to be analyzed is formed. Based on the products to be analyzed, the network platforms from which the reviews originate can be obtained. Corresponding review analysis tasks are created in the system, and the analysis objectives and data scope are clearly defined. The information files of the products to be analyzed include the product name, the network platform on which the product is sold, and the network platform on which the product is promoted.

[0008] Because merchants may name the same product inconsistently across different platforms, identifying the product's intrinsic data and derivative product data to form a product retrieval target can improve the accuracy of core selling point analysis. This involves extracting genuine user reviews from various platforms into core selling points and transforming them into marketing content that precisely reaches user needs. Based on the product name and sales platform, a product retrieval task is established. According to the task configuration, review data from each sales platform is automatically retrieved, and product search and review retrieval are performed. The raw review data is persistently stored in the task database, which is beneficial for obtaining samples for the language model used to analyze reviews. With a large number of reviews, the language model analysis is highly efficient, allowing for the statistical analysis of core product feedback data (positive and negative reviews). This provides data support for in-depth analysis of core selling points, reduces interference from irrelevant fake reviews during the analysis process, and enables the identification of core product characteristics from review information, facilitating precise marketing and product iteration.

[0009] By incorporating review data into a trained semantic analysis model of product selling points, we can obtain analytical conclusions about product attributes and customer feedback. This model can identify usage scenarios and user pain points mentioned in reviews, summarize and obtain high-frequency common viewpoints, and make sentiment judgments. Common viewpoints are statistically categorized to form structured data. Common characteristics strongly related to the product itself are extracted, and common points that users care about are refined to identify the core selling points of the product. Furthermore, the analysis results can directly influence content creation, forming a data-driven closed loop. The analysis results are displayed in the form of charts, statistical panels, etc., allowing users to view the core selling points of the product, the distribution of user feedback, and high-frequency keywords, and can be exported as reports or marketing suggestions. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a method for analyzing product reviews from target users, as described in one embodiment of the present invention. Figure 2 This is a structural connection diagram of a product target user review analysis system according to an embodiment of the present invention; Figure label: 100 - Server; 200 - Storage. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] like Figure 1 As shown, the present invention provides a method for analyzing product reviews from target users, comprising: S100 receives the information file of the product to be analyzed, and forms a folder of the product to be analyzed.

[0013] S200, Select a product in the product folder to be analyzed.

[0014] S300: Obtain the selected product data and derived product data to form the product search target.

[0015] S400 establishes product search tasks based on product search targets and sales platforms.

[0016] S500 uses product retrieval tasks to retrieve review data from each sales platform.

[0017] The S600 performs data cleaning on each comment to obtain cleaned comment data.

[0018] S700 incorporates review data into a trained semantic analysis model of product selling points.

[0019] S800 provides insights into product attributes and customer feedback regarding product selling points.

[0020] S900, return to the previous step and select a product in the product folder to be analyzed, until all products have been selected.

[0021] Specifically, the information file for the product to be analyzed contains relevant product information. For example, a seller might be selling toothbrushes, and on one online sales platform, this product is named "AE brand toothbrush." ​​Since toothbrushes can be categorized into children's toothbrushes, adult toothbrushes, toothbrushes with flat bristles, and toothbrushes with three-dimensional bristles, the information file for "AE brand toothbrush" will provide other names for "AE brand toothbrush," namely, "AE brand children's toothbrush," "AE brand adult toothbrush," "AE brand flat bristle toothbrush," and "AE brand three-dimensional bristle toothbrush." ​​In actual sales, more specific categories like "AE brand children's toothbrush" and "AE brand three-dimensional bristle toothbrush" will be placed on another online sales platform, which may cater to different consumer habits. The information file for the product to be analyzed will include the product's main name, derivative product names, and the different online sales platforms where the product is located. "AE brand toothbrush" is the main product name, while "AE brand children's toothbrush," "AE brand adult toothbrush," "AE brand flat bristle toothbrush," and "AE brand three-dimensional bristle toothbrush" are derivative product names.

[0022] By utilizing product ontology data and derived product data to form product search targets, the aim is to use AE brand children's toothbrushes, AE brand adult toothbrushes, AE brand flat bristle toothbrushes, and AE brand three-dimensional bristle toothbrushes as product search targets for AE brand toothbrushes.

[0023] Simply put, AE brand children's toothbrush, AE brand adult toothbrush, AE brand flat bristle toothbrush, and AE brand 3D bristle toothbrush are derivative product names of AE brand toothbrush. These derivative product names all belong to the products sold by the seller, namely toothbrushes. In order to prevent any derivative product name of AE brand toothbrush from being missed, it is necessary to use the product itself data and derivative product data to form the product search target.

[0024] Both the derived product name and the product itself will serve as the product search target. Furthermore, different online sales platforms have different system configurations during data retrieval. By configuring the interface between these platforms, it's possible to retrieve reviews from different online sales platforms. By using the product search target as the retrieval task and the platform interface configuration task as the product search task, review data from each sales platform can be retrieved efficiently.

[0025] Based on the product selling point semantic analysis model, we can obtain product selling point analysis conclusions about product attributes and customer feedback. By inputting the cleaned and valid reviews into the product selling point semantic analysis model in batches, we can identify the product attributes, usage scenarios, and user pain points mentioned in the reviews, summarize high-frequency common viewpoints, make sentiment judgments, and statistically classify the common viewpoints to form structured data.

[0026] This application relates to a method for analyzing reviews from target users of a product. By receiving information files of the product to be analyzed, a folder of products to be analyzed is formed. Based on the product to be analyzed, the online platforms from which the reviews originate can be obtained. A corresponding review analysis task is created in the system, and the analysis objectives and data scope are clearly defined. The information files of the products to be analyzed include the product name, the online platforms on which the product is sold, and the online platforms on which the product is promoted.

[0027] Because merchants may name the same product inconsistently across different platforms, identifying the product's intrinsic data and derivative product data to form a product retrieval target can improve the accuracy of core selling point analysis. This involves extracting genuine user reviews from various platforms into core selling points and transforming them into marketing content that precisely reaches user needs. Based on the product name and sales platform, a product retrieval task is established. According to the task configuration, review data from each sales platform is automatically retrieved, and product search and review retrieval are performed. The raw review data is persistently stored in the task database, which is beneficial for obtaining samples for the language model used to analyze reviews. With a large number of reviews, the language model analysis is highly efficient, allowing for the statistical analysis of core product feedback data (positive and negative reviews). This provides data support for in-depth analysis of core selling points, reduces interference from irrelevant fake reviews during the analysis process, and enables the identification of core product characteristics from review information, facilitating precise marketing and product iteration.

[0028] By incorporating review data into a trained semantic analysis model of product selling points, we can obtain analytical conclusions about product attributes and customer feedback. This model can identify usage scenarios and user pain points mentioned in reviews, summarize and obtain high-frequency common viewpoints, and make sentiment judgments. Common viewpoints are statistically categorized to form structured data. Common characteristics strongly related to the product itself are extracted, and common points that users care about are refined to identify the core selling points of the product. Furthermore, the analysis results can directly influence content creation, forming a data-driven closed loop. The analysis results are displayed in the form of charts, statistical panels, etc., allowing users to view the core selling points of the product, the distribution of user feedback, and high-frequency keywords, and can be exported as reports or marketing suggestions.

[0029] In one embodiment of this application, S100 includes: S111, Receive an information file of the product to be analyzed.

[0030] S112, parse the information file of the product to be analyzed.

[0031] S113, Obtain the name of the product to be analyzed.

[0032] S114, determine whether the number of names of the product to be analyzed is greater than 1.

[0033] S115, if the number of names of the product to be analyzed is greater than 1, then it is determined that the product to be analyzed has derivative products, and a one-to-one mapping relationship is formed between each derivative product and the product itself.

[0034] S116, if the number of names of the product to be analyzed is equal to 1, then the product to be analyzed is the product itself.

[0035] S117, return to the previous step of receiving an information file of a product to be analyzed, until all information files of products to be analyzed have been selected.

[0036] Understandably, the information files of the products to be analyzed are generally transmitted in document or portable document format. Upon receiving the information files, the file format will be verified. When the information file is a structured file, its content can be read in chunks. When the information file is a semi-structured file, regular expressions can be used to optimize its content, and the optimized file can then be parsed to obtain the information file's contents.

[0037] By parsing the information file of the product to be analyzed, at least one name of the product can be obtained. When the number of names of the product to be analyzed is greater than one, it can be determined that the product to be analyzed has derivative products. When the number of names of the product to be analyzed is equal to one, it can be determined that the product to be analyzed is the product itself.

[0038] To ensure the uniqueness of a single review search, the product search target can be set to the product itself. When a product to be analyzed has derivative products, a mapping relationship will be established between the name of each derivative product and the name of the product itself, so that the review target is the product itself and to prevent reviews of the product itself from being left behind.

[0039] To establish a mapping relationship between the names of derivative products and the names of the product entity, the names of derivative products are generally defined as the source domain, i.e. the original data, and the names of the product entity are defined as the target domain, i.e. the space to which the data is expected to be mapped. An association key is constructed between the source domain and the target domain, and constraints between the source domain and the target domain are formed based on the transformation function.

[0040] By leveraging the mapping relationship between the names of derivative products and the names of the original products, we can reduce the probability of missing reviews, increase the utilization rate of reviews, improve the extraction of common features strongly related to the product itself, refine the common points that users care about, and find the core selling points of the product.

[0041] In one embodiment of this application, S100 further includes: S121, Select a name for the product itself.

[0042] S122, using the name of the product entity, find the name of the derivative product that has a mapping relationship with the product entity.

[0043] S123 uses the name of the product itself and the name of the derivative product as search criteria.

[0044] S124, in the information file, based on the search criteria, find the sales platform of the product itself.

[0045] S125 obtains communication link data from the sales platform.

[0046] S126, establishes a mapping relationship between communication link data and the name of the product itself.

[0047] S127, return to the step of selecting a product name, until all product names have been selected.

[0048] S128, based on the name of each product entity, the name of the derivative product, and the communication link data of the sales platform, generate at least one product folder to be analyzed.

[0049] Understandably, because the same product has different subcategories, the same product may have different product names on different online sales platforms. In order to aggregate reviews from each online sales platform, the names of derivative products that have a mapping relationship with the original product name will be determined based on the original product name.

[0050] Based on the name of the product itself and the names of derivative products, the information file identifies each sales platform corresponding to the name of the product itself and the names of derivative products. For example, AE brand toothbrushes are sold on the first and second sales platforms; AE brand children's toothbrushes are sold on the second and third sales platforms; AE brand adult toothbrushes are sold only on the fourth sales platform; and AE brand flat bristle toothbrushes and AE brand three-dimensional bristle toothbrushes are sold on the fifth sales platform. Therefore, it can be determined that AE brand toothbrushes, as the name of the product itself, have the first, second, third, fourth, and fifth sales platforms as online sales platforms.

[0051] Online sales platforms need to establish communication links with their servers so that the servers can access or store the retrieved comment data. Each folder containing products to be analyzed stores the name of a single product and the names of any derivative products it may have, along with the communication link data between the sales platform and each product name and derivative product name.

[0052] In one embodiment of this application, S300 includes: S310, call the selected product entity in the product folder to be analyzed.

[0053] S320, search for the product attributes and usage scenarios of the product itself in the information file.

[0054] S330: Decompose the text content of the product attributes of the product ontology into vectors to obtain the first text association dataset of the product ontology.

[0055] S340, the text content of the product ontology's usage scenarios is vectorized to obtain the second text association dataset of the product ontology.

[0056] S350, based on the first text association dataset and the second text association dataset, establish the product retrieval target.

[0057] S360, return to the product body selected in the call, until all product bodies have been called.

[0058] Understandably, taking a toothbrush as an example, a toothbrush's product attributes include functional attributes, design attributes, technological attributes, and usage and experience attributes. Functional attributes include cleaning ability, removing plaque and food debris through the physical friction of the bristles. Gum protection includes the softness of the bristles, the smoothness of the brush head, and pressure sensing to prevent damage to the gums and tooth enamel. The bristle material is typically nylon or DuPont filament. Bristle hardness is generally categorized as soft, medium, and hard to meet the needs of different gum conditions. Bristle shapes include tapered bristles, diamond-shaped bristles, wavy bristles, and varying heights, designed to better clean between teeth and the gingival sulcus. Bristle density; the more numerous and denser the bristles, the higher the cleaning efficiency may be. Brush head shapes include small oval, diamond-shaped, and rectangular. The handle material includes PP plastic, TPE soft rubber, bamboo, wood, and metal.

[0059] The product attributes and usage scenario descriptions of the product itself can be decomposed into vectors. Based on the matching requirements of the comment text, a reference direction is selected as the projection target. The projection component of the original vector in the target direction is calculated through inner product operation. If the remaining component needs to be strictly orthogonal to the target direction, the scalar coefficient needs to be adjusted. The orthogonal component is obtained by subtracting the projection component from the original vector. The original text is transformed into a vector space, so that the product attributes and usage scenario descriptions can be machine learned.

[0060] By vectorizing the textual content of the product attributes and usage scenarios, we can obtain the product search target. This is beneficial for extracting common characteristics that are strongly related to the product itself from the comments, extracting common points that users care about, and identifying the core selling points of the product.

[0061] In one embodiment of this application, S400 includes: S410, Select a product search target.

[0062] S420: Obtain the communication link data of the product itself and the sales platform corresponding to the product search target.

[0063] S430, select a sales platform.

[0064] S440: Construct a product search task using the name of the product itself.

[0065] S450 determines the system configuration integration task of the selected sales platform based on the communication link data of the sales platform.

[0066] Specifically, the system configuration interface task is used to retrieve comment data from the selected sales platform.

[0067] S460, return to the step of selecting a sales platform until all sales platforms have been selected.

[0068] S470 integrates the product search task with the system configuration task to form a product retrieval task.

[0069] S480, return to the step of selecting a product search target until all product search targets have been selected.

[0070] Understandably, a product search target corresponds to at least one sales platform. The server itself needs to establish a communication link with the sales platform based on the communication link data, i.e., the Uniform Resource Locator (URL). Establishing a communication link using the URL requires disassembling the URL structure, confirming the protocol and port, sequentially checking the browser cache, operating system cache, router cache, and ISP's DNS server, recursively querying until the target IP address is obtained, thus establishing a transport layer connection. Based on the transport layer connection, a request is sent, and the status code and response header returned by the sales platform's server are read, thus forming the communication link with the sales platform.

[0071] Each sales platform has a different URL structure, protocol and port, operating system, and router port communication protocol. By utilizing the communication link data of each sales platform, the system configuration integration task for the selected platform can be determined. Using the product's name and derived product names, the sales platform to be integrated is identified. Based on the sales platform's URL structure, protocol and port, operating system, and router port communication protocol, a system configuration integration task for that platform is generated. Combining the product search task and the system configuration integration task, a product retrieval task is formed, which allows the retrieval of reviews corresponding to the product's name.

[0072] In one embodiment of this application, S600 includes: S611, based on the received comment data, converts each comment data into a comment string.

[0073] S612, Select a comment string from the comment data.

[0074] S613, Select a character from the comment string.

[0075] S614 uses the character type to determine whether the selected character matches Chinese characters, numeric characters, or English characters.

[0076] S615, if the selected character does not match the Chinese character, numeric character, or English character, the selected character is removed, and one character of the selected comment string is returned, until all characters have been selected.

[0077] S616 If the selected character matches a Chinese character, a numeric character, or an English character, then return one character from the selected comment string until all characters have been selected.

[0078] S617, return the comment string of the selected comment data, until all comment data has been selected.

[0079] S618, obtain the preliminarily cleaned comment string about Chinese characters, numeric characters, and English characters.

[0080] It is understandable that there are irrelevant emoticons or symbols in the comment data. These irrelevant emoticons or symbols do not belong to Chinese characters, numbers, or English characters. The type of character can be used to clean up irrelevant emoticons or symbols in the comments.

[0081] In one embodiment of this application, S600 further includes: S621, Select a preliminary cleaned comment data.

[0082] S622, the comment strings of the initially cleaned comment data are decomposed into vectors to form comment vectors to be compared.

[0083] S623, compare the first text association dataset with the comment vector to be compared.

[0084] S624, Based on the comparison results of the first text association dataset, determine whether the first text association dataset and the comment vector to be compared match.

[0085] S625, if the first text association dataset and the comment vector to be compared do not match, then the selected comment data is determined to be meaningless text interaction between buyers, the selected comment data is removed, and the process of selecting a preliminary cleaned comment data is repeated until all the preliminary cleaned comment data has been selected.

[0086] S626, if the first text association dataset matches the comment vector to be compared, then compare the second text association dataset with the comment vector to be compared, and based on the comparison result of the second text association dataset, determine whether the second text association dataset matches the comment vector to be compared.

[0087] S627, if the second text association dataset and the comment vector to be compared do not match, then the selected comment data is determined to be comments interfered with by fake accounts, the selected comment data is removed, and the process of selecting preliminary cleaned comment data is repeated until all preliminary cleaned comment data is selected.

[0088] S628, if the second text association dataset matches the comment vector to be compared, then the cleaning of the selected comment data is completed, and the process of selecting a preliminary cleaned comment data is repeated until all preliminary cleaned comment data is selected.

[0089] Understandably, the first text association dataset of the product ontology consists of textual content about the product attributes of the product ontology. When the comment vector to be compared, formed by vector decomposition of the comment strings of the initially cleaned comment data, does not match the first text association dataset, it can be determined that the comment content is user dialogue, meaningless repetitive text, or descriptions unrelated to product attributes.

[0090] The second text association dataset of the product ontology is text content about the usage scenarios of the product ontology. When the comment vector to be compared, formed by vector decomposition of the comment strings of the initially cleaned comment data, does not match the second text association dataset, it can be determined that the comment content is a fake order comment or advertising content, i.e., a comment interfered by a fake account.

[0091] When the comment vector formed by vector decomposing the comment string of the initially cleaned comment data matches both the first text association dataset and the second text association dataset of the product ontology, the comment can be determined to be a valid comment and can be considered a cleaned comment.

[0092] In one embodiment of this application, before S700, the method further includes establishing a semantic analysis model of product selling points. Establishing a semantic analysis model of product selling points includes: S711 receives comment data with tagged content.

[0093] S712 decomposes the sample comment data into vectors to obtain sample vectors.

[0094] S713, using the first text association dataset and the second text association dataset of the product ontology, obtain the target object vector.

[0095] S714, remove the target object vector from the sample vector to obtain the comment object vector.

[0096] S715 maps the target object vector and the evaluation object vector to each other in a one-to-one correspondence.

[0097] S716, obtain sample data for training the semantic analysis model of product selling points.

[0098] Understandably, the sample review data needs to be labeled beforehand. Product attributes and usage scenarios in the review data are used as target object vectors, and everything related to the target object vectors is related to the product itself. The target object vectors include the first text association dataset and the second text association dataset. In the valid review data, apart from the first and second text association datasets, the remaining review content is the comment object vector. The comment object vector is the substantive content of the review that identifies user pain points, frequently shared viewpoints, and emotional tendencies after product use.

[0099] Simply put, product attributes and usage scenarios in the sample review data can be labeled, and this content can be decomposed into vectors to form target object vectors.

[0100] The sample review data contains substantive content about product reviews, such as user pain points, frequently shared opinions, and reviews with emotional tendencies after product use. These can be decomposed into a vector of review objects.

[0101] By mapping the target object vector and the comment object vector to a one-to-one correspondence, sample data can be generated to train the semantic analysis model of product selling points.

[0102] In one embodiment of this application, establishing a semantic analysis model of product selling points further includes: S721, perform word segmentation on the received cleaned comment data to obtain the smallest semantic unit.

[0103] S722 digitizes the smallest semantic unit to obtain a numerical sequence.

[0104] S723 performs a high-dimensional vector transformation on the obtained numerical sequence to obtain an input word segmentation vector with positional encoding.

[0105] S724, Adjust the attention weights of the product selling point semantic analysis model.

[0106] S725 utilizes attention weights to perform layer stacking and residual connections on the input word segmentation vectors.

[0107] S726 uses the output vector of the product selling point semantic analysis model to generate mapping text about numerical sequences.

[0108] S727, obtain the semantic analysis model of product selling points to be trained.

[0109] S728, incorporate each sample data into the semantic analysis model of product selling points to be trained.

[0110] S729, sequentially obtain the output vectors of the semantic analysis model of product selling points.

[0111] S730, calculate the matching probability between the review object vector and the output vector.

[0112] S731, if the matching probability between the review object vector and the output vector is greater than or equal to the target probability threshold, output the trained semantic analysis model of product selling points.

[0113] S732, if the matching probability between the review object vector and the output vector is less than the target probability threshold, return to adjust the attention weight of the semantic analysis model of product selling points until the trained semantic analysis model of product selling points is output.

[0114] It can be understood that establishing the semantic analysis model of product selling points can include establishing the semantic analysis model of product selling points to be trained and training the semantic analysis model of product selling points to be trained.

[0115] The sub-word segmentation algorithm dynamically segments the text into the smallest semantic units. For example, "quite useful" is segmented into "quite / useful". Each smallest semantic unit is mapped to a unique numerical sequence through a predefined vocabulary, and the obtained numerical sequence is vectorized to obtain an input segmented vector with positional encoding, which is converted into a high-dimensional vector through a trainable embedding vector matrix. The vector spaces of words with similar semantics are close in distance. By adding rotational positional encoding or absolute positional encoding, an input segmented vector with positional encoding is obtained, solving the problem of word order dependence. The input vector generates query, key, and value matrices through a linear transformation, calculates the global correlation strength through a scaled dot product, aggregates the context information after normalization and weighting, and parallel multi-group attention heads capture different semantic dimensions, such as grammar and sentiment. The feed-forward network processing performs a non-linear transformation on the attention output to enhance the feature expression ability. The layer stacking and residual connection concatenate dozens to hundreds of transformation modules in series, and each layer stabilizes the training through residual connection and layer normalization. Finally, the final hidden state is mapped to the vocabulary dimension through a linear layer, and the numerical sequence is inversely mapped back to the text.

[0116] Call the sample data for training the semantic analysis model of product selling points, calculate the matching probability between the review object vector and the output vector. When the matching probability between the review object vector and the output vector is greater than or equal to the target probability threshold, output the trained semantic analysis model of product selling points. When the matching probability between the review object vector and the output vector is less than the target probability threshold, continue to train the semantic analysis model of product selling points.

[0117] In fact, both the comment object vector and the output vector can be mapped back to text from numerical sequences through a vocabulary dimension. The predefined vocabularies used by the comment object vector and the output vector are consistent.

[0118] Language models offer high analytical efficiency, enabling the collection of core feedback data from both positive and negative reviews. This provides data support for in-depth analysis of a product's core selling points, reduces interference from irrelevant fake reviews during the analysis process, and allows for the identification of core product characteristics from review information, facilitating precise marketing and product iteration.

[0119] like Figure 2 As shown, the product target user review analysis system provided by the present invention includes a server 100 and a memory 200.

[0120] Server 100 is used to execute the product target user review analysis method.

[0121] The memory 200 is communicatively connected to the server 100.

[0122] This invention provides a product review analysis system for target users. The server 100 receives information files of the products to be analyzed from the storage 200, forming a folder of products to be analyzed. Based on the products to be analyzed, the system can obtain review data from the online platforms from which the reviews originate, store the review data in the storage 200, create corresponding review analysis tasks in the system, and clarify the analysis objectives and data scope. The information files of the products to be analyzed include the product name, the online platform on which the product is sold, and the online platform on which the product is promoted.

[0123] By incorporating the review data stored in memory 200 into the trained product selling point semantic analysis model, the system obtains product selling point analysis conclusions based on product attributes and customer feedback. It can identify usage scenarios and user pain points mentioned in reviews, summarize and obtain high-frequency common viewpoints, and perform sentiment analysis. Common viewpoints are statistically categorized to form structured data, extracting common characteristics strongly related to the product itself, refining common points of user concern, identifying the product's core selling points, and directly influencing content creation, forming a data-driven closed loop. Analysis results are displayed in the form of charts, statistical panels, etc., allowing users to view the product's core selling points, user feedback distribution, and high-frequency keywords, and can be exported as reports or marketing suggestions.

[0124] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for analyzing reviews from target users of a product, characterized in that, include: Receive information files of the products to be analyzed to form a folder of products to be analyzed; Select a product in the folder of products to be analyzed; Obtain the selected product data and derived product data to form the product search target; Based on the product search objectives and sales platform, establish a product search task; Using a product search task, retrieve review data from each sales platform; Perform data cleaning on each comment to obtain cleaned comment data; Incorporate the comment data into the trained semantic analysis model of product selling points; Obtain conclusions from product attribute and customer feedback analysis of product selling points; Return to the folder containing the products to be analyzed and select one product, repeating this process until all products have been selected.

2. The product target user review analysis method according to claim 1, characterized in that, The process of receiving information files of the products to be analyzed to form a folder of products to be analyzed includes: Receive an information file of the product to be analyzed; Parse the information file of the product to be analyzed; Obtain the name of the product to be analyzed; Determine if the number of names of the product to be analyzed is greater than 1; If the number of names of the product to be analyzed is greater than 1, it is determined that the product to be analyzed has derivative products, and a one-to-one mapping relationship is formed between each derivative product and the product itself. If the number of product names to be analyzed is equal to 1, then the product to be analyzed is the product itself. Return to the previous state to receive an information file of the product to be analyzed, until all information files of the products to be analyzed have been selected.

3. The product target user review analysis method according to claim 2, characterized in that, The step of receiving the information file of the product to be analyzed to form a folder of products to be analyzed also includes: Select a name for the product itself; Using the name of the product itself, find the names of derivative products that have a mapping relationship with the product itself; Use the name of the product itself and the names of derivative products as search criteria; Within the information file, based on the search criteria, locate the sales platform for the product itself; Obtain communication link data from the sales platform; Establish a mapping relationship between communication link data and the name of the product itself; Return to the previous step and select a name for a product entity until all product entity names have been selected. Based on the name of each product entity, the name of its derivative products, and the communication link data of the sales platform, at least one product folder to be analyzed is generated.

4. The product target user review analysis method according to claim 3, characterized in that, The process of obtaining the selected product ontology data and derived product data to form a product retrieval target includes: Call the selected product entity from the product folder to be analyzed; Find the product attributes and usage scenarios of the product itself in the information file; The text content of the product attributes of the product body is vectorized to obtain the first text association dataset of the product body; The text content of the product's usage scenarios is vectorized to obtain the second text association dataset of the product's main body; Based on the first and second text association datasets, establish product retrieval targets; Return to the selected product body in the call, until all product bodies have been called.

5. The product target user review analysis method according to claim 4, characterized in that, The process of establishing a product search task based on the product search target and sales platform includes: Select a product search target; Obtain the communication link data of the product itself and the sales platform corresponding to the product search target; Choose a sales platform; Construct a product search task using the product's name; Based on the communication link data of the sales platform, determine the system configuration integration task of the selected sales platform; the system configuration integration task is used to call the comment data of the selected sales platform. Return to the previous step and select a sales platform until all sales platforms have been selected; The product search task and the system configuration task are integrated to form a product retrieval task; Return to the previous page and select a product search target until all product search targets have been selected.

6. The product target user review analysis method according to claim 5, characterized in that, The step of cleaning each comment data point to obtain cleaned comment data includes: Based on the received comment data, each comment data is converted into a comment string; Select a comment string from the comment data; Select a character from the comment string; Based on the character type, determine whether the selected character matches Chinese characters or numeric characters; If the selected character does not match a Chinese character or a number character, the selected character will be removed, and one character from the selected comment string will be returned, until all characters have been selected; If the selected character matches a Chinese character or a numeric character, then return one character from the selected comment string, until all characters have been selected; Return the comment string of the selected comment data, until all comment data has been selected; Obtain a preliminarily cleaned comment string containing Chinese characters and numeric characters.

7. The product target user review analysis method according to claim 6, characterized in that, The step of cleaning each comment data point to obtain cleaned comment data also includes: Select a preliminary cleaned comment data; The comment strings in the initially cleaned comment data are decomposed into vectors to form comment vectors for comparison. Compare the first text association dataset with the comment vectors to be compared; Based on the comparison results of the first text association dataset, determine whether the first text association dataset and the comment vector to be compared match; If the first text association dataset and the comment vector to be compared do not match, the selected comment data is determined to be meaningless text interaction between buyers. The selected comment data is then removed, and the process of selecting a preliminary cleaned comment data is repeated until all the preliminary cleaned comment data has been selected. If the first text association dataset matches the comment vector to be compared, then the second text association dataset and the comment vector to be compared are compared. Based on the comparison result of the second text association dataset, it is determined whether the second text association dataset and the comment vector to be compared match. If the second text association dataset and the comment vector to be compared do not match, the selected comment data is determined to be comments from fake accounts that interfere with the comparison. The selected comment data is then removed, and the process of selecting a preliminary cleaned comment data is repeated until all the preliminary cleaned comment data has been selected. If the second text association dataset matches the comment vector to be compared, the cleaning of the selected comment data is completed, and the process of selecting a preliminary cleaned comment data is repeated until all preliminary cleaned comment data has been selected.

8. The product target user review analysis method according to claim 7, characterized in that, Before incorporating the comment data into the trained product selling point semantic analysis model, the method includes establishing a product selling point semantic analysis model: Receive comment data with tagged content; The sample comment data is decomposed into vectors to obtain sample vectors; Using the first and second text association datasets of the product ontology, obtain the target object vector; Remove the target object vector from the sample vector to obtain the comment object vector; Map the target object vector and the comment object vector to each other in a one-to-one correspondence. Obtain sample data for training the semantic analysis model of product selling points.

9. The product target user review analysis method according to claim 8, characterized in that, The establishment of the product selling point semantic analysis model also includes: The received cleaned comment data is segmented into words to obtain the smallest semantic units; The smallest semantic unit is digitally mapped to obtain a numerical sequence; The obtained numerical sequence is transformed into a high-dimensional vector to obtain an input word segmentation vector with positional encoding; Adjust the attention weights of the product selling point semantic analysis model; Attention weights are used to stack layers and perform residual connections on the input word segmentation vectors; Using the output vector of the product selling point semantic analysis model, generate mapping text about numerical sequences; Obtain the semantic analysis model of product selling points to be trained; Each sample data point is incorporated into the product selling point semantic analysis model to be trained; The output vectors of the product selling point semantic analysis model are obtained sequentially; The probability of matching between the object vector and the output vector in the statistical evaluation; If the matching probability between the comment object vector and the output vector is greater than or equal to the target probability threshold, then the trained semantic analysis model of the product selling points is output. If the matching probability between the comment object vector and the output vector is less than the target probability threshold, then return to the adjustment of the attention weights of the product selling point semantic analysis model until the trained product selling point semantic analysis model is output.

10. A product target user review analysis system, characterized in that, include: A server is configured to execute the product target user review analysis method as described in any one of claims 1 to 9; The memory is connected in communication with the server.

Images