Method for generating and displaying product review keywords

Abstract

The application provides a commodity comment keyword generation and display method. By obtaining and preprocessing the comment dataset to form standard expression text, providing a data basis for subsequent analysis, and inputting the processed text into a large language model, dynamic calling of corresponding clustering weight configuration is carried out according to the target commodity category information for semantic clustering analysis, realizing adaptive keyword extraction for different emphases, and overcoming the defects of poor universality or uncontrollable output. Through multidimensional constraint filtering of the extracted initial keyword set, the data value of the finally generated keyword set is ensured. In addition, a label sequence is generated and dynamic display and interactive filtering are realized, and an incremental updating mechanism is introduced, ensuring that the label content can evolve in real time. The whole process of automatic generation and management from original comments to high-quality, interactive and real-time updating keyword labels is realized, and the accuracy, effectiveness and user experience of the e-commerce platform comment information summary are improved.

Description

Technical Field

[0001] This application relates to the field of e-commerce data processing, and in particular to a method for generating and displaying product review keywords. Background Technology

[0002] In e-commerce platforms, product review tags serve as a key interactive component for users to quickly obtain review summaries, playing a crucial role in improving information retrieval efficiency and transaction conversion. Especially with the rise of non-standardized products that possess randomness or are associated with specific groups, user reviews are becoming increasingly diverse, emotional, and personalized, posing significant challenges to traditional review tag generation methods.

[0003] Some feasible methods have limitations. On the one hand, methods that rely on manually preset fixed label templates can ensure the standardization of labels, but cannot flexibly adapt to the diverse evaluation dimensions of non-standard products, and the operation and maintenance costs are high. On the other hand, semantic clustering methods based on general large models can automatically extract keywords, but their output results are often limited by preset static attribute templates, or completely unconstrained, resulting in a large number of generalized words with low information density such as "very good" and "not bad", with inconsistent label quality and a lack of business compliance.

[0004] Therefore, there is a need for a method to generate and display product review keywords that can automatically adapt to the characteristics of different product categories while ensuring that the generated tags are accurate, positive, concise, and representative, in order to solve the problem of insufficient efficiency and effectiveness in information acquisition and display in the e-commerce field. Summary of the Invention

[0005] The purpose of this application is to address at least one of the aforementioned technical deficiencies, particularly the technical deficiency of insufficient efficiency and effectiveness in information acquisition and display in the e-commerce field in the prior art.

[0006] Firstly, this application provides a method for generating and displaying product review keywords, the method comprising:

[0007] Obtain a dataset of review information for the target product, and preprocess the dataset to obtain a set of standard description review texts;

[0008] The preprocessing includes data cleaning and semantic merging. Data cleaning is used to remove invalid characters or characters that deviate from semantics, and semantic merging is used to merge characters with the same or similar semantics.

[0009] The standard expression comment text set is input into a pre-trained large language model, and based on the category information associated with the target product, the corresponding scenario-based clustering weight configuration is invoked to perform semantic clustering analysis on the standard expression comment text set to extract an initial keyword set;

[0010] The initial keyword set is subjected to filtering based on preset filtering conditions to obtain a compliant keyword set;

[0011] The preset filtering conditions include at least one of keyword length constraints, sentiment constraints, semantic specificity constraints, and mention frequency constraints.

[0012] The keywords in the compliant keyword set are sorted according to preset sorting conditions to generate a keyword tag sequence;

[0013] The keyword tag sequence is dynamically displayed on the target product's display interface, and in response to a trigger operation on any keyword tag in the keyword tag sequence, comment information that semantically matches the keyword tag is filtered and displayed.

[0014] As an optional implementation method, the establishment of the scenario-based clustering weight configuration includes:

[0015] Multiple preset product categories are predefined, along with a weight configuration vector corresponding to each preset product category;

[0016] The weight configuration vector is used to indicate the attention weights of multiple predefined comment attribute dimensions during the semantic clustering analysis.

[0017] The predefined comment attribute dimensions include at least two of the following: emotion dimension, appearance dimension, function dimension, material dimension, operation dimension, and experience dimension.

[0018] Furthermore, based on the category information associated with the target product, the corresponding contextualized clustering weight configuration is invoked to perform semantic clustering analysis on the standard description comment text set, extracting an initial keyword set, including:

[0019] Based on the category information associated with the target product, the preset product category that matches the target product is determined, and the corresponding weight configuration vector is called to guide the large language model to adjust the degree of attention to each of the predefined comment attribute dimensions in the semantic clustering analysis according to the weight configuration vector, so as to obtain an initial keyword set.

[0020] As an optional implementation, the step of performing filtering processing on the initial keyword set based on preset filtering conditions to obtain a compliant keyword set includes:

[0021] Based on the keyword length constraint, a first candidate keyword with a character count not exceeding a first preset threshold is selected from the initial keyword set;

[0022] By invoking the sentiment analysis model interface or querying the preset sentiment dictionary, a second candidate keyword with a positive semantic sentiment tendency parameter is selected from the initial keyword set;

[0023] By using the semantic specificity constraint, words belonging to the preset generalized vocabulary list are excluded from the initial keyword set to obtain the third candidate keyword;

[0024] By using the mentioned frequency constraint, a fourth candidate keyword with a mention frequency of not less than the second preset threshold is selected from the initial keyword set in the standard expression comment text set;

[0025] A set of compliant keywords is obtained based on the first candidate keyword, the second candidate keyword, the third candidate keyword, and the fourth candidate keyword.

[0026] As an optional implementation, the step of filtering and displaying comment information that semantically matches the keyword tag in response to a triggering operation on any keyword tag in the keyword tag sequence includes:

[0027] Record the target keyword tags that are triggered;

[0028] From the set of standard expression comment texts, target comment information whose content semantics match the target keyword tags is selected;

[0029] In the preset comment display area of ​​the display interface, the target comment information is displayed in sorted order according to the mention frequency and the semantic sentiment tendency parameter.

[0030] As an optional implementation, the method further includes an incremental update step, specifically comprising:

[0031] Monitor new review data for the target product;

[0032] When the number of newly added comment information data reaches a preset incremental update threshold, an incremental keyword set is generated based on the newly added comment information data;

[0033] The incremental keyword set is merged and deduplicated with the currently displayed keyword tag sequence to update the keyword tag sequence.

[0034] As an optional implementation, the process of fusing and deduplicating the incremental keyword set with the currently displayed keyword tag sequence includes:

[0035] Calculate the semantic similarity between each incremental keyword in the incremental keyword set and each existing keyword in the current keyword tag sequence;

[0036] For keyword pairs whose semantic similarity exceeds a third preset threshold, a merging operation is performed, and the incremental keyword or the existing keyword is used as the merged expression based on the mention frequency and semantic sentiment tendency parameters of the incremental keyword and the existing keyword in the keyword pair.

[0037] For keyword pairs whose semantic similarity does not exceed the third preset threshold, the corresponding incremental keyword is added to the keyword tag sequence.

[0038] Secondly, this application provides a product review keyword generation and display device, comprising:

[0039] The processing module is used to acquire a dataset of review information for the target product and preprocess the dataset to obtain a set of standard description review texts.

[0040] The preprocessing includes data cleaning and semantic merging. Data cleaning is used to remove invalid characters or characters that deviate from semantics, and semantic merging is used to merge characters with the same or similar semantics.

[0041] The processing module is also used to input the standard expression comment text set into a pre-trained large language model, and based on the category information associated with the target product, call the corresponding scenario-based clustering weight configuration to perform semantic clustering analysis on the standard expression comment text set and extract an initial keyword set;

[0042] The processing module is further configured to perform filtering processing on the initial keyword set based on preset filtering conditions to obtain a compliant keyword set;

[0043] The preset filtering conditions include at least one of keyword length constraints, sentiment constraints, semantic specificity constraints, and mention frequency constraints.

[0044] The processing module is also used to sort each keyword in the compliant keyword set according to preset sorting conditions and generate a keyword tag sequence;

[0045] The processing module is also configured to dynamically display the keyword tag sequence on the display interface of the target product, and in response to the triggering operation of any keyword tag in the keyword tag sequence, filter and display comment information that semantically matches the keyword tag.

[0046] As an optional implementation, the specific method by which the processing module establishes the scenario-based clustering weight configuration includes:

[0047] Multiple preset product categories are predefined, along with a weight configuration vector corresponding to each preset product category;

[0048] The weight configuration vector is used to indicate the attention weights of multiple predefined comment attribute dimensions during the semantic clustering analysis.

[0049] The predefined comment attribute dimensions include at least two of the following: emotional dimension, appearance dimension, functional dimension, material dimension, operation dimension, and experience dimension.

[0050] Furthermore, based on the category information associated with the target product, the corresponding contextualized clustering weight configuration is invoked to perform semantic clustering analysis on the standard description comment text set, extracting an initial keyword set, including:

[0051] Based on the category information associated with the target product, the preset product category that matches the target product is determined, and the corresponding weight configuration vector is called to guide the large language model to adjust the degree of attention to each of the predefined comment attribute dimensions in the semantic clustering analysis according to the weight configuration vector, so as to obtain an initial keyword set.

[0052] Thirdly, this application provides a computer device including one or more processors and a memory storing computer-readable instructions that, when executed by the one or more processors, perform the steps of the method described in the first aspect.

[0053] Fourthly, this application provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the method described in the first aspect.

[0054] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0055] Based on any of the above embodiments, the method provided in this application acquires and preprocesses a comment dataset to form standard descriptive text, laying a high-quality data foundation for subsequent analysis. Then, the processed text is input into a pre-trained large language model, and semantic clustering analysis is performed by dynamically calling the corresponding scenario-based clustering weight configuration based on the target product category information. This achieves adaptive keyword extraction tailored to the focus of comments on different product categories, overcoming the shortcomings of traditional methods such as poor versatility or uncontrollable output. By applying hard constraints including length, sentiment, semantic specificity, and frequency of mention to the extracted initial keyword set, it ensures that the final generated keyword set simultaneously possesses business compliance, informational value, and user representativeness. Finally, a tag sequence is generated through sorting and dynamically displayed with interactive filtering. An incremental update mechanism is introduced to ensure that the tag content evolves in real time with new comments. This method achieves fully automated generation and management of high-quality, interactive, and real-time updatable keyword tags from original comments, significantly improving the accuracy, effectiveness, and user experience of e-commerce platform comment information summaries. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 A flowchart illustrating a method for generating and displaying product review keywords according to an embodiment of this application;

[0058] Figure 2 This is a schematic diagram of the process architecture based on a real-world application scenario for the product review keyword generation and display method provided in one embodiment of this application;

[0059] Figure 3 This is an internal structural diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0060] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0061] In the user shopping decision-making process on e-commerce platforms, the product review section is a core information scenario. Users judge the suitability of products based on other consumers' genuine reviews, and review keyword tags, as "review summaries," directly affect the efficiency of user information acquisition and purchase conversion. With the rise of non-standard products such as blind boxes and IP collaborations, user reviews are characterized by "personalization, emotionalization, and multiple segmentation dimensions" (e.g., blind boxes focus on "appearance, feel, and probability of hidden items," rather than the "quality and logistics" of traditional products). The traditional review tagging model can no longer meet the current user needs and business scenarios.

[0062] The core characteristics of the current application of comment tags are that user needs are upgrading. Users are no longer satisfied with "generalized tags" but want to quickly obtain evaluation dimensions that "precisely match product characteristics" and need to quickly locate corresponding comments through tags to reduce invalid browsing. Business scenarios are expanding. There are no fixed templates for evaluation dimensions of non-standard products. Manually preset tags cannot cover all personalized reviews, resulting in low tag coverage and delayed updates.

[0063] Specifically, for manually preset tags, operators pre-configure fixed general tags for different product categories (such as "good quality" and "fast logistics") and count the number of clicks mentioned accordingly. This approach cannot adapt to personalized reviews of popular non-standard products such as blind boxes and IP products, and the operating costs are too high for such products with the characteristics of "too many new products and too many SKUs".

[0064] For general large-scale model clustering, related functions can be implemented based on AI large-scale models. However, this requires pre-setting attribute templates for each category. The model only parses the keywords contained in the preset attribute templates. The essence of this implementation is to use a manually preset static attribute framework to constrain the scope of AI's comment summary. This leads to keyword generalization and cannot support refined operations, resulting in significant limitations. Alternatively, directly calling a general large-scale model to perform semantic clustering on all comments may extract keywords based on word frequency or simple models. This can easily produce neutral words with low information density, such as "very good" or "not bad," resulting in inconsistent tag quality. Keyword noise is high, and there is a lack of semantic normalization. Similar expressions are not merged, leading to tag redundancy.

[0065] To address the aforementioned shortcomings, this application provides an interactive method for semantic clustering and keyword extraction based on a large model. On one hand, it designs a "scenario-based adaptive tag generation mechanism" that uses an AI large model to accurately match tags with the evaluation dimensions of different product categories, meeting users' needs for "accurately obtaining personalized information" and completely replacing manual presets. At the same time, it achieves fully automatic tag generation and real-time updates, reducing operating costs while ensuring that tags are "accurate, positive, compliant, and concise," in line with the business goals of diversified and personalized operations.

[0066] In summary, the technical concept of this application lies in providing a method for generating and displaying product review keywords, aiming to systematically solve the problems of poor adaptability, low quality, high maintenance costs, and weak interactivity faced by e-commerce platforms, especially in the generation of non-standardized product review tags. Firstly, the original review information undergoes data cleaning and semantic merging preprocessing to effectively remove noise and normalize synonymous expressions, thereby obtaining a high-quality, standardized review text set, providing reliable data input for subsequent deep semantic analysis. Secondly, the innovation lies in the fact that when inputting the preprocessed text into a pre-trained large language model, instead of using an indiscriminate general analysis mode, it calls a predefined, scenario-based clustering weight configuration that matches the characteristics of the target product based on the specific category information. This configuration, in vector form, instructs the model to give differentiated attention to different review attribute dimensions (such as sentiment, appearance, function, material, etc.) in semantic clustering analysis, enabling keyword extraction to accurately focus on the evaluation dimensions most concerned by users of the current category. This achieves a leap from "general extraction" to "scenario-based adaptive extraction," which is the key to improving tag accuracy and category adaptability in this invention.

[0067] Furthermore, to address potential issues such as noise, generalization, or non-compliance with business requirements in the direct output of large language models, this application constructs a multi-dimensional hard constraint filtering mechanism that includes keyword length constraints, sentiment bias constraints, semantic specificity constraints, and mention frequency constraints. This mechanism rigorously filters the initially extracted keywords, retaining only words that are of reasonable length, have positive sentiment, specific semantics, and a certain level of user mention. This ensures that the generated keyword tags are concise, guiding, information-dense, and representative, effectively filtering out low-value or negative words and solving the problem of uncontrollable output from general-purpose models.

[0068] Based on this, the compliant keyword set is sorted according to preset sorting conditions (such as comprehensive mention frequency and positive sentiment), generating an ordered keyword tag sequence and dynamically displaying it on the product interface, enabling users to quickly obtain review summaries. Simultaneously, the system responds to user triggers on any tag, instantly filtering and displaying specific reviews that semantically match it, achieving seamless penetration from tag summaries to detailed content, greatly optimizing the user's information retrieval experience. Finally, the method designs an incremental update mechanism, which monitors new review data and automatically triggers a new round of clustering, filtering, and sorting processes when a threshold is reached, generating an incremental keyword set. This incremental set is then intelligently fused and deduplicated with the existing tag sequence by calculating semantic similarity, thereby achieving automated and real-time updates of tag content, ensuring that tags always reflect the latest product reputation dynamics. In summary, this application achieves high-quality, high-availability operation and maintenance of review keyword tags by connecting a closed loop of "data preprocessing - scenario-based adaptive clustering - hard-constraint business filtering - dynamic sorting and display - interactive penetration - incremental update".

[0069] The methods provided in this application will be described in detail below based on the corresponding implementation methods in some practical application scenarios.

[0070] Figure 1 This is a flowchart illustrating a method for generating and displaying product review keywords according to an embodiment of this application, as shown below. Figure 1 As shown, this application provides a method for generating and displaying product review keywords, the method comprising:

[0071] S101. Obtain the review information dataset of the target product, and preprocess the review information dataset to obtain a standard description review text set;

[0072] The preprocessing includes data cleaning and semantic merging. Data cleaning is used to remove invalid characters or characters that deviate from semantics, and semantic merging is used to merge characters with the same or similar semantics.

[0073] S102. Input the standard expression comment text set into the pre-trained large language model, and based on the category information associated with the target product, call the corresponding scenario-based clustering weight configuration to perform semantic clustering analysis on the standard expression comment text set and extract the initial keyword set;

[0074] S103. Perform filtering processing on the initial keyword set based on preset filtering conditions to obtain a compliant keyword set;

[0075] The preset filtering conditions include at least one of keyword length constraints, sentiment constraints, semantic specificity constraints, and mention frequency constraints.

[0076] S104. Sort each keyword in the compliant keyword set according to the preset sorting conditions to generate a keyword tag sequence;

[0077] S105. Dynamically display the keyword tag sequence on the display interface of the target product, and in response to the triggering operation of any keyword tag in the keyword tag sequence, filter and display comment information that semantically matches the keyword tag.

[0078] Specifically, the method corresponding to this implementation provides a complete technical solution for automatically generating high-quality keyword tags from original product reviews and interactively displaying them. First, the method performs data acquisition and preprocessing steps. The system obtains all historical and real-time reviews of the target product from the e-commerce platform's backend database or real-time data stream interface, forming a review information dataset. This dataset typically contains metadata such as review text, ratings, user identifiers, and timestamps. Preprocessing involves data cleaning and semantic merging. Data cleaning aims to remove invalid or interfering information; for example, through regular expression rules or dictionary-based matching, it removes review content containing advertising links, phone numbers, garbled characters, completely meaningless strings (such as a series of punctuation marks), or reviews clearly unrelated to product evaluations. Semantic merging aims to solve the redundancy problem caused by the diversity of user expressions; its logic is to normalize different text fragments expressing the same or similar semantics into a standard expression. For example, for various descriptions such as "feels great," "great feel," and "comfortable to touch," semantic similarity is calculated using a pre-trained word vector model or a thesaurus. When the similarity exceeds a certain threshold, they are mapped to a unified semantic cluster, and the most representative standard expression (such as "comfortable to touch") is selected as the replacement text for all comments within that cluster in subsequent analysis. This process ultimately yields a set of standard expression comment texts that have eliminated noise and unified expression, laying a high-quality data foundation for subsequent accurate semantic analysis.

[0079] Next, the standard comment text set is input into a large language model pre-trained on large-scale general text data and potentially fine-tuned with data from the e-commerce comment domain. The key here is that instead of performing indiscriminate general semantic clustering on all comments, a "contextualized clustering weight configuration" is introduced. This method first requires establishing a category weight knowledge base, which predefines all product categories on the platform (such as "beauty and skincare," "home appliances," and "trendy blind boxes"), and assigns a multi-dimensional weight vector to each category. Each dimension of this vector corresponds to a predefined comment attribute dimension, such as sentiment, appearance, function, material, operation, and experience. The weight value represents the level of attention the model should give to the corresponding dimension when analyzing product reviews for that category. For example, for the "trendy blind box" category, the weight values ​​of the "sentiment" and "appearance" dimensions in its weight configuration vector will be significantly higher than those of the "function" dimension; while for the "home appliances" category, the weights of the "function" and "operation" dimensions will be even higher. In practice, the system retrieves the corresponding weight configuration vector from the knowledge base based on the category of the target product and inputs it into the large language model as part of the prompt words or as attention guidance parameters for the model. After receiving the weight guidance, the large language model focuses on dimensions with higher weights when performing semantic understanding, vectorization, and clustering analysis on the comment text, thereby extracting an initial keyword set that better matches the characteristics of the product category. For example, for blind box products, the model is more likely to cluster keywords such as "surprise," "cute," and "collectible value"; for rice cookers, it tends to cluster keywords such as "fast cooking," "easy to operate," and "easy to clean."

[0080] Then, the initial keyword set extracted in the above steps is filtered using multi-dimensional hard constraint rules. These preset filtering conditions work together to ensure that the final keywords meet the business display requirements. Keyword length constraints, by setting a maximum character limit, ensure that the generated tags can be displayed clearly and completely on the mobile interface, avoiding layout errors or information truncation caused by excessively long tags. Sentiment constraints, by calling an external sentiment analysis model interface or querying the built-in sentiment dictionary, score each keyword for sentiment, retaining only words with positive or neutral sentiment and eliminating negative words to create a positive product review atmosphere and guide user decision-making. Semantic specificity constraints rely on a predefined list of generalized words (such as those containing "good," "not bad," "okay," "average," etc.), excluding words with low information density from the candidate set, prompting the retained keywords to convey more specific and informative product attributes or experience descriptions. Mention frequency constraints require that the number of times the keyword appears in the standard description comment text set is no less than a preset threshold, ensuring that the generated tags have sufficient user consensus and representativeness, rather than being occasional mentions by individual users. Only keywords that simultaneously meet or pass multiple of the above criteria can constitute the final set of compliant keywords.

[0081] Subsequently, the set of compliant keywords is sorted according to preset sorting criteria. These criteria typically consider a weighted score combining keyword mention frequency and semantic sentiment parameters. For example, a comprehensive score could be the weighted sum of mention frequency and sentiment score. Keywords with high mention frequency and positive sentiment will receive higher ranking weights. The sequence generated based on this weighting is the keyword tag sequence. Finally, this keyword tag sequence is dynamically rendered on the product details page. Each tag is usually presented as a clickable button or link, and may include its mention count. When a user clicks any tag, the system immediately responds to the trigger, quickly filtering out all specific comments that semantically match the triggered tag from the standard comment text set. Semantic matching is achieved by calculating the cosine similarity between the comment text vector and the tag text vector and determining whether it exceeds a matching threshold. The filtered relevant comments are centrally displayed in the page's preset comment display area (such as a floating panel below the tag or a filtered comment list page), achieving seamless penetration from macro-level tag summaries to micro-level specific evaluations, greatly improving the efficiency of user browsing and information retrieval.

[0082] Please see Figure 2 , Figure 2 This is a schematic diagram of the process architecture based on a real-world application scenario for the product review keyword generation and display method provided in one embodiment of this application. It serves as an example to illustrate the implementation of this application in a specific application scenario. You can also refer to the detailed description of each implementation method below.

[0083] First, comment data preprocessing and filtering are performed. For example, after a user posts a comment, the backend cleans the data, identifies and removes advertisements and meaningless comments, and retains only valid comments. Through natural language processing technology, synonyms such as "super cute," "too cute," and "incredibly cute" are unified into "very cute," eliminating semantic redundancy, improving clustering accuracy, and making the labels more concise and consistent.

[0084] Meanwhile, based on a large language model, scenario-based semantic clustering and controllable keyword extraction are performed. By training the large model for shopping scenarios, the weight of category tag attributes is automatically adjusted, thereby automatically adjusting the tag extraction dimensions to make the tags fit the user's focus.

[0085] For example, for blind boxes or IP merchandise, the focus is on emotional appeal > appearance > craftsmanship, and tags such as "very cute," "good flocking feel," and "high aesthetic appeal" can be extracted. For beauty and skincare products, the focus is on skin feel > makeup lasting power > packaging, and tags such as "refreshing skin feel," "8-hour makeup lasting power," and "exquisite packaging" can be extracted. For home appliances and digital products, the focus is on function > operation > material, and tags such as "practical function," "easy to operate," and "durable material" can be extracted.

[0086] Next, the tag conditions are restricted through four constraints to eliminate words that do not meet the criteria. For example, the constraints could be: keyword length ≤ 10 characters (to meet mobile display space requirements and avoid excessively long tags with ellipses and line breaks); only positive semantics (to guide users to make positive decisions); avoid overly generalized words (such as "good" or "very good"); prioritize more specific promotional expressions; and mention the tag ≥ 3 times (to ensure the representativeness of the tag). For details, please refer to the relevant implementation methods.

[0087] It should be noted that this part can be executed through a large language model; for example, the relevant steps can be completed using the following task prompts:

[0088] String normalizeSystem="You are an e-commerce review analysis assistant. The input is a JSON object containing product_id, product_name, and terms."+

[0089] "terms is an array, and each element contains raw_term and sample_content."

[0090] Task: Normalize each raw_term into a unified core_content, rule: "+

[0091] "1) The length of core_content shall not exceed 10 characters, which must be positive and able to attract purchases;"+

[0092] "2) Words with the same or similar semantics within the same productid must be output as completely identical core_content (e.g., good / very good / easy to use / very good / very satisfied -> very easy to use; very soft / very soft and comfortable / soft and skin-friendly -> soft and skin-friendly);"+

[0093] "3) Prioritize the existing expressions in the knowledge base as the unified result; if there is no suitable one in the knowledge base, select a more specific expression from the current terms as the unified result;"+

[0094] "4) Avoid single-word generalized terms (such as "good"), which must be normalized to more specific terms;"+

[0095] "5) Strictly output a JSON array, each object only contains two fields: raw_term and core_content, no extra text shall be output.

[0096] After the screening is completed, keyword weight sorting and dynamic display are performed. For example, keywords can be sorted according to the weight of mention count + semantic positivity, and keywords with high mention count and high positivity are displayed in the front; each keyword is marked with the corresponding mention count, and the interaction conforms to the presentation form of "efficient browsing".

[0097] During the operation and maintenance process, closed-loop interactive control and real-time update are implemented, which responds to the user's click operation on the keyword tag, filters and displays the corresponding comments in real time, supports the user to click the "tag" to jump to the comment page, retains the current filtering conditions, and realizes data linkage display; after a new comment is submitted, incremental clustering and keyword update are automatically triggered without manual intervention to ensure the real-time performance of the tags.

[0098] The method described in this implementation first preprocesses the comment information dataset, including data cleaning and semantic merging, effectively purifying the original data and unifying semantic expressions to obtain a standardized comment text set, providing a high-quality data foundation for subsequent precise analysis. Then, the standardized text set is input into a pre-trained large language model, and semantic clustering analysis is performed by innovatively calling the corresponding scenario-based clustering weight configuration based on the category information associated with the target product. This allows the keyword extraction process to adapt to the comment focus of different product types; for example, for blind boxes, the focus is more on emotion and appearance, while for home appliances, the focus is more on function and operation, thus extracting an initial keyword set that better matches the product characteristics. Subsequently, this initial set is filtered by applying preset filtering conditions including keyword length, sentiment tendency, semantic specificity, and frequency of mention, ensuring that the final keyword set is concise and compliant in form, positive in sentiment, specific and clear in content, and representative of users. Finally, compliant keywords are sorted according to preset sorting conditions to generate a tag sequence, which is dynamically displayed and interactively responded to, enabling users to efficiently obtain comment summaries and accurately locate specific evaluation content of interest. This method, through the organic combination of the above steps, constructs a complete technical system from data preprocessing to intelligent analysis, strict filtering, orderly display, and real-time interaction, achieving the technical effect of high-quality, automated, and contextualized generation and efficient display of comment keywords.

[0099] As an optional implementation method, the establishment of the scenario-based clustering weight configuration includes:

[0100] Multiple preset product categories are predefined, along with a weight configuration vector corresponding to each preset product category;

[0101] The weight configuration vector is used to indicate the attention weights of multiple predefined comment attribute dimensions during the semantic clustering analysis.

[0102] The predefined comment attribute dimensions include at least two of the following: emotion dimension, appearance dimension, function dimension, material dimension, operation dimension, and experience dimension.

[0103] Furthermore, based on the category information associated with the target product, the corresponding contextualized clustering weight configuration is invoked to perform semantic clustering analysis on the standard description comment text set, extracting an initial keyword set, including:

[0104] Based on the category information associated with the target product, the preset product category that matches the target product is determined, and the corresponding weight configuration vector is called to guide the large language model to adjust the degree of attention to each of the predefined comment attribute dimensions in the semantic clustering analysis according to the weight configuration vector, so as to obtain an initial keyword set.

[0105] This implementation further elaborates on the construction and specific application logic of scenario-based clustering weight configuration. The establishment of the scenario-based clustering weight configuration is a process of transforming business knowledge into machine-understandable and executable quantitative guidance information. First, business experts or data analysis are needed to define a clear and mutually exclusive pre-defined product category system for the e-commerce platform's product system, such as "beauty and skincare," "personal care appliances," "trendy toys," "household goods," "digital accessories," and "clothing and footwear." For each pre-defined product category, a set of pre-defined review attribute dimensions that can summarize its user evaluation dimensions needs to be analyzed and defined. These dimensions typically include, but are not limited to: emotional dimensions (such as liking, satisfaction, surprise), appearance dimensions (such as aesthetics, color, design), functional dimensions (such as effect, performance, use), material dimensions (such as texture, materials, feel), operational dimensions (such as ease of use, installation complexity), and experience dimensions (such as comfort, durability), etc.

[0106] For each product category, a specific weight value is assigned to each predefined review attribute dimension, forming a weight configuration vector that corresponds one-to-one with that category. The weight allocation reflects the importance of that dimension in evaluating this type of product. For example, for the "trendy blind box" category, its weight configuration vector could be set as: [Emotional dimension: 0.4, Appearance dimension: 0.3, Material dimension: 0.2, Functional dimension: 0.05, Operational dimension: 0.05, Experience dimension: 0.0]. This means that when analyzing blind box reviews, the model should focus most on users' emotional expressions and descriptions of appearance. For the "electric toothbrush" category, its vector might be: [Functional dimension: 0.35, Operational dimension: 0.25, Experience dimension: 0.2, Material dimension: 0.15, Appearance dimension: 0.05, Emotional dimension: 0.0], emphasizing function, operation, and experience.

[0107] In practical applications, when the system needs to process reviews of a target product, it first parses the category information associated with the product (usually from a product category tree or product attribute tags). Then, it searches for a matching category in a pre-defined product category system. Upon successful matching, it calls the weight configuration vector corresponding to that category. This weight configuration vector is not directly used for arithmetic calculations, but rather serves as a high-level instruction or contextual information, inputting it along with a standard set of review text into a pre-trained large language model. Large language models (especially attention-based models) can understand the "emphasis" intention expressed by this weight vector. During semantic encoding, feature extraction, and clustering analysis within the model, the weight vector substantially influences the model's attention allocation to different semantic features. Specifically, when encoding review sentences, the model assigns higher attention scores to words or phrases related to high-weight dimensions, making these features more dominant in the generated text representation vector. When performing unsupervised or cue-based clustering analysis, the model tends to group reviews that express similarities along high-weight dimensions together and extracts the words or phrases that best represent those high-weight dimensions from these clusters as initial keywords. This approach achieves adaptive capability, dynamically adjusting the analytical focus according to the product category, ensuring that the extracted keywords accurately reflect the evaluation concerns of different product categories.

[0108] This implementation further defines the establishment and application of scenario-based clustering weight configuration. By predefining multiple product categories and their corresponding weight configuration vectors, this vector clarifies the differences in attention weights for different predefined comment attribute dimensions during semantic clustering analysis. In practical applications, the system matches and calls the corresponding weight configuration vector based on the specific category information of the target product, thereby guiding the large language model to selectively adjust the attention allocation for different dimensions (such as emotion, appearance, function, material, etc.) when analyzing comment text. This mechanism makes keyword extraction no longer a "one-size-fits-all" process, but rather allows for adaptive focusing based on product category characteristics. For example, for beauty products, the weights on dimensions such as "skin feel" and "makeup lasting power" can be increased, while for digital products, more attention is paid to dimensions such as "function" and "operation." Therefore, this implementation significantly improves the matching degree and accuracy between the extracted initial keyword set and the actual attributes of the product and the user's focus by introducing a configurable weight guidance mechanism strongly associated with the product category, providing a technical guarantee for achieving scenario-based and personalized comment tags.

[0109] As an optional implementation, the step of performing filtering processing on the initial keyword set based on preset filtering conditions to obtain a compliant keyword set includes:

[0110] Based on the keyword length constraint, a first candidate keyword with a character count not exceeding a first preset threshold is selected from the initial keyword set;

[0111] By invoking the sentiment analysis model interface or querying the preset sentiment dictionary, a second candidate keyword with a positive semantic sentiment tendency parameter is selected from the initial keyword set;

[0112] By using the semantic specificity constraint, words belonging to the preset generalized vocabulary list are excluded from the initial keyword set to obtain the third candidate keyword;

[0113] By using the mentioned frequency constraint, a fourth candidate keyword with a mention frequency of not less than the second preset threshold is selected from the initial keyword set in the standard expression comment text set;

[0114] A set of compliant keywords is obtained based on the first candidate keyword, the second candidate keyword, the third candidate keyword, and the fourth candidate keyword.

[0115] On one hand, keyword length constraint filtering is implemented. The system iterates through each keyword in the initial keyword set, counting the number of characters it contains (including Chinese characters, letters, numbers, and punctuation). A first preset threshold is set, which is usually determined based on the size of the label display area in the front-end UI design to ensure that the labels are clearly readable and aesthetically pleasing within a limited space. For example, this threshold can be set to the number of characters suitable for single-line display on mobile devices. Keywords with a character count not exceeding this threshold are filtered out, forming the first candidate keyword set. This step ensures the visual friendliness and layout standardization of the generated labels.

[0116] Simultaneously, sentiment bias filtering is implemented. For keywords that pass the length screening, the system needs to determine their sentiment tone. There are two ways to achieve this: one is to call a mature external sentiment analysis model application programming interface (API), taking the keyword as input and receiving the sentiment polarity (positive, neutral, negative) and confidence score returned by the model; the other is to query a pre-built sentiment dictionary optimized for the e-commerce review domain. This dictionary contains a large number of words and their sentiment bias tags and intensity values. The system retrieves the records of keywords in the dictionary, and if their sentiment bias parameters are marked as positive (or neutral and acceptable according to business rules), they are retained. Only keywords with positive sentiment bias parameters are included in the second candidate keyword set. This step ensures the positiveness of the tags and business compliance from a sentiment-oriented perspective, filtering out words that may evoke negative associations.

[0117] Furthermore, semantic specificity constraint filtering is performed. The system maintains a pre-defined list of generalized terms, which contains a series of generic evaluation terms with low information content and lacking specific references, such as "good," "not bad," "okay," "average," "not bad," and "satisfactory." The system compares the keywords to be filtered with this list, and any words appearing in the list are directly excluded. The keywords remaining after this filtering constitute a third set of candidate keywords. The purpose of this step is to improve the information density and distinguishability of the tags, forcing the system to output more specific and descriptive product features or user experience terms, thereby providing users with more valuable summary information.

[0118] Furthermore, frequency-based filtering is implemented. The system counts the frequency of each initial keyword within the preprocessed set of standard comment texts. A second preset threshold is set to measure whether a keyword has sufficient user consensus and representativeness, avoiding noise labels generated by individual users' specific expressions. Keywords with a frequency not lower than this threshold are selected as the fourth candidate keyword set. This step ensures that the final labels are based on the collective feedback of a certain number of users, enhancing the credibility and universality of the labels.

[0119] Finally, considering the results of the four filtering processes, the intersection of the results from any one or more of these processes is taken to obtain the compliant keyword set. After the four filtering steps generate the first, second, third, and fourth candidate keyword sets respectively, the system typically takes the intersection of these four sets. That is, a keyword must simultaneously meet four conditions: moderate length, positive sentiment, specific semantics, and frequent mention, to ultimately enter the compliant keyword set. In practice, a more flexible strategy can also be adopted, such as requiring keywords to pass all enabled constraints (at least one, usually all four), but allowing slight relaxation on one condition (such as setting a secondary threshold). Through this filtering mechanism, the final compliant keyword set meets high-quality standards for business presentation in terms of form, sentiment, content, and representativeness.

[0120] This implementation details the execution of filtering the initial keyword set based on multiple hard constraints. Keyword length constraints select keywords with a character count not exceeding a preset threshold, ensuring the readability and aesthetics of the generated tags within the limited display space of mobile devices. Sentiment constraints, utilizing sentiment analysis models or sentiment dictionaries, retain only keywords with positive semantic sentiment, filtering out neutral or negative words and guiding a positive consumer decision-making atmosphere. Semantic specificity constraints eliminate empty words from the preset generalized vocabulary list, avoiding low-information-density tags such as "good" or "not bad," thus enhancing the content value of the tags. Frequency constraints retain keywords frequently mentioned in comments, ensuring the representativeness and widespread acceptance of the tags. Through the synergistic effect of these four constraints, this implementation performs multi-dimensional and refined filtering of the initial keywords, resulting in a compliant keyword set that meets business requirements in terms of display format, sentiment guidance, information content, and user base, fundamentally solving the quality control and business compliance issues of automated tag generation.

[0121] As an optional implementation, the step of filtering and displaying comment information that semantically matches the keyword tag in response to a triggering operation on any keyword tag in the keyword tag sequence includes:

[0122] Record the target keyword tags that are triggered;

[0123] From the set of standard expression comment texts, target comment information whose content semantics match the target keyword tags is selected;

[0124] In the preset comment display area of ​​the display interface, the target comment information is displayed in sorted order according to the mention frequency and the semantic sentiment tendency parameter.

[0125] This implementation describes in detail the response process and display logic of the backend system after the user interacts with the keyword tags displayed on the front end, achieving precise penetration from tags to specific comment content.

[0126] When a user triggers a tag in the keyword tag sequence on the target product's display interface through mouse clicks, finger touches, or other means, the front-end captures this interaction event and sends the identifier of the triggered target keyword tag (such as tag text or a unique ID) to the back-end server. The back-end system first records this target keyword tag as the basis for this filtering operation.

[0127] Subsequently, the system performs semantic matching filtering from the stored set of standard expression comment texts. This matching is not a simple string inclusion relationship, but rather a calculation based on semantic similarity. The system utilizes natural language processing techniques to convert the target keyword tag and each standard expression comment into vector representations in a high-dimensional semantic space (e.g., using a sentence encoder derived from the aforementioned large language model). Next, it calculates the cosine similarity or other semantic similarity measures between the target keyword tag vector and the vector of each comment. A semantic matching threshold is set; comments with a similarity exceeding this threshold are considered target comment information whose "content semantics match the target keyword tag." For example, if a user clicks on the tag "long battery life," the system will filter out all comments expressing similar meanings, such as "the battery is very durable," "one charge lasts for several days," and "battery life exceeds expectations," even if they do not contain exactly the same words.

[0128] After filtering out the target comments, the system needs to sort and display them in a preset comment display area on the interface (such as a dynamically pop-up sidebar, an expanded panel, or a filtered view that refreshes and locates the comment list). The sorting is mainly based on two attributes of each target comment: First, the frequency of mentions of the target keyword tag (or its synonyms). This can be understood as how many users liked, found helpful, or replied to the comment itself, or the intensity of semantic similarity between the comment and the target tag. A more direct approach is to record the semantic similarity score between each comment and the target tag as a quantitative indicator of its relevance during semantic matching filtering. Second, the overall semantic sentiment tendency parameter of the comment, which can be obtained by performing sentiment analysis on the entire comment. The sorting algorithm can be designed as a comprehensive weighted sort, for example, comprehensive score = w1 * mention frequency or semantic similarity score + w2 * comment sentiment tendency score. In practice, other dimensions can also be added for sorting reference, such as the keyword length or semantic specificity mentioned above. The semantic similarity score ensures a high degree of relevance between comments and tag topics, while the sentiment score prioritizes more positive and uplifting reviews to optimize the user browsing experience. Finally, target comments are sorted from highest to lowest based on their calculated composite scores and dynamically rendered in the front-end display area. This sorting method not only quickly locates relevant comments but also further optimizes their presentation order, allowing users to see the most relevant and positive feedback first, significantly improving information retrieval efficiency and satisfaction.

[0129] This implementation specifically defines the response flow after a user interacts with a keyword tag. When a user triggers a keyword tag on the interface, the system records the target tag and filters out specific comment information that semantically matches the target tag from a pre-processed set of standard comment texts. Subsequently, these comments are displayed in a specific area of ​​the interface, sorted according to their frequency of mention and semantic sentiment parameters. This process achieves direct and precise penetration from abstract tags to specific comment content. Users do not need to manually browse and search through massive amounts of comments; they can quickly focus on all relevant comments that reflect a specific dimension of evaluation. Furthermore, the system's sorting mechanism prioritizes more popular and positive comments, further improving the efficiency of information acquisition and the quality of the reading experience. Therefore, this implementation, by achieving intelligent association and orderly presentation of tags and comment content, completes the interactive loop from keyword summaries to detailed browsing, greatly enhancing the system's practicality and user-friendliness.

[0130] As an optional implementation, the method further includes an incremental update step, specifically comprising:

[0131] Monitor new review data for the target product;

[0132] When the number of newly added comment information data reaches a preset incremental update threshold, an incremental keyword set is generated based on the newly added comment information data;

[0133] The incremental keyword set is merged and deduplicated with the currently displayed keyword tag sequence to update the keyword tag sequence.

[0134] This implementation also provides an incremental update step, which aims to enable keyword tags to evolve dynamically and automatically as new comments are generated, thus maintaining their timeliness and representativeness.

[0135] This incremental update step is based on continuous monitoring of newly added review data for the target product. The system captures newly generated review data for the product in real-time or near real-time by listening to the e-commerce platform's review posting message queue, periodically polling the product review interface, or subscribing to database change logs. This newly added review data is temporarily stored in an incremental buffer or log.

[0136] The system presets an incremental update threshold, which can be an absolute number (e.g., 50 new comments), a relative proportion (e.g., 5% of the total existing comments), or a time period (e.g., triggered every 24 hours). When the number of new comment data detected reaches or exceeds the preset threshold, the system automatically triggers an incremental keyword generation process.

[0137] The specific process for generating the incremental keyword set can refer to the screening, extraction, and filtering procedures of any implementation method. The system takes the newly added comment information data as input and first performs preprocessing operations, including data cleaning and semantic merging, to obtain a subset of standard comment text based solely on the new comments. Then, this text subset is input into a pre-trained large language model, and based on the (potentially unchanged) category information associated with the target product, the corresponding contextualized clustering weight configuration is invoked to perform independent semantic clustering analysis on this new text subset, extracting an initial incremental keyword set. Next, the aforementioned preset filtering conditions (keyword length constraint, sentiment constraint, semantic specificity constraint, and mention frequency constraint) are applied to this initial incremental keyword set for filtering, resulting in a compliant incremental keyword set. Finally, this compliant incremental keyword set can be sorted according to preset sorting conditions to generate an incremental keyword sequence. Thus, an incremental keyword set based on the latest comment data that meets all quality requirements is generated.

[0138] After generating the incremental keyword set, the next step is to merge and deduplicatize it with the keyword tag sequence currently displayed on the product display interface. This process is not simply merging old and new keywords, but requires intelligent integration. The purpose of merging and deduplication is to incorporate new viewpoints and expressions emerging from new comments into the existing tag system, while merging semantically similar or identical expressions to avoid tag redundancy. It may also adjust the tag order or remove outdated tags based on new data. Through this step, the keyword tag sequence is ultimately updated, ensuring that the tags displayed to users continuously reflect the latest product reputation trends.

[0139] This implementation adds an incremental update step to the aforementioned keyword generation and display process. By continuously monitoring new review data for the target product, and automatically initiating a new round of keyword extraction based on the new data when the data volume reaches a preset threshold, an incremental keyword set is generated. Subsequently, this incremental set is merged and deduplicated with the currently displayed keyword tag sequence, thus completing the tag sequence update. This mechanism ensures that the keyword tags displayed for the product dynamically evolve with the continuous influx of user feedback, promptly reflecting the latest evaluation trends and changes in reputation, and avoiding the problem of tag content becoming outdated due to review updates. This implementation achieves automated and real-time maintenance of the keyword tag system, ensuring the synchronization of tag information with the current review ecosystem without manual intervention, further enhancing the system's dynamic adaptability and long-term effectiveness.

[0140] As an optional implementation, the process of fusing and deduplicating the incremental keyword set with the currently displayed keyword tag sequence includes:

[0141] Calculate the semantic similarity between each incremental keyword in the incremental keyword set and each existing keyword in the current keyword tag sequence;

[0142] For keyword pairs whose semantic similarity exceeds a third preset threshold, a merging operation is performed, and the incremental keyword or the existing keyword is used as the merged expression based on the mention frequency and semantic sentiment tendency parameters of the incremental keyword and the existing keyword in the keyword pair.

[0143] For keyword pairs whose semantic similarity does not exceed the third preset threshold, the corresponding incremental keyword is added to the keyword tag sequence.

[0144] This embodiment provides an intelligent merging and adding mechanism based on semantic similarity calculation.

[0145] First, the system needs to calculate the semantic similarity between each incremental keyword in the incremental keyword set and each existing keyword in the current keyword tag sequence. The calculation of semantic similarity typically relies on word embedding or sentence embedding techniques. The system can use pre-trained word embedding models (such as Word2Vec, a domain-adapted version of GloVe) or sentence encoding models (such as Transformer-based sentence representation models) to map each keyword (whether incremental or existing) into a real-valued vector in a high-dimensional space. Then, the degree of semantic similarity between the two vectors is quantified by calculating the cosine similarity. The value of cosine similarity typically ranges from -1 to 1, with values ​​closer to 1 indicating greater semantic similarity.

[0146] The system presets a third threshold as a critical point to determine whether two keywords are "semantically similar enough to need to be merged". For each pair of incremental keywords and existing keywords, the semantic similarity is calculated and the result is compared with the threshold.

[0147] For keyword pairs with semantic similarity exceeding a third preset threshold, the system determines that they express highly similar meanings and requires merging to avoid duplicate display. Merging is not arbitrary; the merged expression is determined based on certain rules. These rules typically consider both the frequency of mention and the semantic sentiment parameter of the two words in the keyword pair. For example, a priority strategy can be designed: the keyword with higher mention frequency is prioritized as the merged expression because it represents broader user consensus; if mention frequencies are similar, the one with a more positive semantic sentiment parameter (higher positive score) is selected; in some cases, if both words have their advantages and merging provides a more comprehensive expression (e.g., "good sound quality" and "clear voice"), the system can also generate a new expression (e.g., "clear and pleasant sound quality") as the merged result, but this new expression must ensure it meets all filtering constraints. After merging, the original existing keywords may be updated (e.g., their mention frequencies are updated to the sum of the two), while the corresponding incremental keywords are no longer added to the sequence separately.

[0148] For keyword pairs whose semantic similarity does not exceed the third preset threshold, it means that the incremental keyword is significantly different from any existing tag in semantics, representing a new evaluation perspective or expression method. The system then directly adds this incremental keyword to the current keyword tag sequence as a new tag item. After addition, the entire sequence usually needs to be reordered according to preset sorting conditions (such as combining the frequency of mention after the addition and sentiment tendency) to ensure that the display order of tags always reflects their comprehensive importance.

[0149] Through this refined processing based on semantic similarity, the incremental update process can intelligently absorb new information, merge similar items, and retain unique items, thereby achieving a smooth and orderly evolution of the label sequence. This avoids information redundancy and ensures the timeliness and richness of the label system.

[0150] This implementation further refines the deduplication logic for merging the incremental keyword set with the existing tag sequence. By calculating the semantic similarity between each incremental keyword and each existing keyword in the existing tag sequence, the system can intelligently identify the semantic relationship between new and old keywords. For keyword pairs with semantic similarity exceeding a preset threshold, a merging operation is performed, and the more suitable expression is selected as the representative of the merged pair by comprehensively considering parameters such as frequency of mention and sentiment, thus avoiding tag redundancy. For keywords with semantic similarity below the threshold, they are directly added to the sequence as new tags. This method not only achieves the fusion of new information but also ensures the simplicity and representativeness of the tag set through semantic deduplication and optimization. By introducing semantic similarity calculation and intelligent merging strategies, this implementation makes the incremental update process no longer a simple addition but an evolution with understanding capabilities, thereby continuously maintaining and optimizing the quality of the tag library during dynamic updates.

[0151] This application also provides a device for generating and displaying product review keywords, including:

[0152] The processing module is used to acquire a dataset of review information for the target product and preprocess the dataset to obtain a set of standard description review texts.

[0153] The preprocessing includes data cleaning and semantic merging. Data cleaning is used to remove invalid characters or characters that deviate from semantics, and semantic merging is used to merge characters with the same or similar semantics.

[0154] The processing module is also used to input the standard expression comment text set into a pre-trained large language model, and based on the category information associated with the target product, call the corresponding scenario-based clustering weight configuration to perform semantic clustering analysis on the standard expression comment text set and extract an initial keyword set;

[0155] The processing module is further configured to perform filtering processing on the initial keyword set based on preset filtering conditions to obtain a compliant keyword set;

[0156] The preset filtering conditions include at least one of keyword length constraints, sentiment constraints, semantic specificity constraints, and mention frequency constraints.

[0157] The processing module is also used to sort each keyword in the compliant keyword set according to preset sorting conditions and generate a keyword tag sequence;

[0158] The processing module is also configured to dynamically display the keyword tag sequence on the display interface of the target product, and in response to the triggering operation of any keyword tag in the keyword tag sequence, filter and display comment information that semantically matches the keyword tag.

[0159] The method described in this implementation first preprocesses the comment information dataset, including data cleaning and semantic merging, effectively purifying the original data and unifying semantic expressions to obtain a standardized comment text set, providing a high-quality data foundation for subsequent precise analysis. Then, the standardized text set is input into a pre-trained large language model, and semantic clustering analysis is performed by innovatively calling the corresponding scenario-based clustering weight configuration based on the category information associated with the target product. This allows the keyword extraction process to adapt to the comment focus of different product types; for example, for blind boxes, the focus is more on emotion and appearance, while for home appliances, the focus is more on function and operation, thus extracting an initial keyword set that better matches the product characteristics. Subsequently, this initial set is filtered by applying preset filtering conditions including keyword length, sentiment tendency, semantic specificity, and frequency of mention, ensuring that the final keyword set is concise and compliant in form, positive in sentiment, specific and clear in content, and representative of users. Finally, compliant keywords are sorted according to preset sorting conditions to generate a tag sequence, which is dynamically displayed and interactively responded to, enabling users to efficiently obtain comment summaries and accurately locate specific evaluation content of interest. This method, through the organic combination of the above steps, constructs a complete technical system from data preprocessing to intelligent analysis, strict filtering, orderly display, and real-time interaction, achieving the technical effect of high-quality, automated, and contextualized generation and efficient display of comment keywords.

[0160] As an optional implementation, the specific method by which the processing module establishes the scenario-based clustering weight configuration includes:

[0161] Multiple preset product categories are predefined, along with a weight configuration vector corresponding to each preset product category;

[0162] The weight configuration vector is used to indicate the attention weights of multiple predefined comment attribute dimensions during the semantic clustering analysis.

[0163] The predefined comment attribute dimensions include at least two of the following: emotion dimension, appearance dimension, function dimension, material dimension, operation dimension, and experience dimension.

[0164] Furthermore, based on the category information associated with the target product, the corresponding contextualized clustering weight configuration is invoked to perform semantic clustering analysis on the standard description comment text set, extracting an initial keyword set, including:

[0165] Based on the category information associated with the target product, the preset product category that matches the target product is determined, and the corresponding weight configuration vector is called to guide the large language model to adjust the degree of attention to each of the predefined comment attribute dimensions in the semantic clustering analysis according to the weight configuration vector, so as to obtain an initial keyword set.

[0166] This implementation further defines the establishment and application of scenario-based clustering weight configuration. By predefining multiple product categories and their corresponding weight configuration vectors, this vector clarifies the differences in attention weights for different predefined comment attribute dimensions during semantic clustering analysis. In practical applications, the system matches and calls the corresponding weight configuration vector based on the specific category information of the target product, thereby guiding the large language model to selectively adjust the attention allocation for different dimensions (such as emotion, appearance, function, material, etc.) when analyzing comment text. This mechanism makes keyword extraction no longer a "one-size-fits-all" process, but rather allows for adaptive focusing based on product category characteristics. For example, for beauty products, the weights on dimensions such as "skin feel" and "makeup lasting power" can be increased, while for digital products, more attention is paid to dimensions such as "function" and "operation." Therefore, this implementation significantly improves the matching degree and accuracy between the extracted initial keyword set and the actual attributes of the product and the user's focus by introducing a configurable weight guidance mechanism strongly associated with the product category, providing a technical guarantee for achieving scenario-based and personalized comment tags.

[0167] As an optional implementation, the processing module performs filtering processing on the initial keyword set based on preset filtering conditions to obtain a compliant keyword set in the following specific ways:

[0168] Based on the keyword length constraint, a first candidate keyword with a character count not exceeding a first preset threshold is selected from the initial keyword set;

[0169] By invoking the sentiment analysis model interface or querying the preset sentiment dictionary, a second candidate keyword with a positive semantic sentiment tendency parameter is selected from the initial keyword set;

[0170] By using the semantic specificity constraint, words belonging to the preset generalized vocabulary list are excluded from the initial keyword set to obtain the third candidate keyword;

[0171] By using the mentioned frequency constraint, a fourth candidate keyword with a mention frequency of not less than the second preset threshold is selected from the initial keyword set in the standard expression comment text set;

[0172] A set of compliant keywords is obtained based on the first candidate keyword, the second candidate keyword, the third candidate keyword, and the fourth candidate keyword.

[0173] This implementation details the execution of filtering the initial keyword set based on multiple hard constraints. Keyword length constraints select keywords with a character count not exceeding a preset threshold, ensuring the readability and aesthetics of the generated tags within the limited display space of mobile devices. Sentiment constraints, utilizing sentiment analysis models or sentiment dictionaries, retain only keywords with positive semantic sentiment, filtering out neutral or negative words and guiding a positive consumer decision-making atmosphere. Semantic specificity constraints eliminate empty words from the preset generalized vocabulary list, avoiding low-information-density tags such as "good" or "not bad," thus enhancing the content value of the tags. Frequency constraints retain keywords frequently mentioned in comments, ensuring the representativeness and widespread acceptance of the tags. Through the synergistic effect of these four constraints, this implementation performs multi-dimensional and refined filtering of the initial keywords, resulting in a compliant keyword set that meets business requirements in terms of display format, sentiment guidance, information content, and user base, fundamentally solving the quality control and business compliance issues of automated tag generation.

[0174] As an optional implementation, the specific method by which the processing module filters and displays comment information semantically matching the keyword tag in response to a trigger operation on any keyword tag in the keyword tag sequence includes:

[0175] Record the target keyword tags that are triggered;

[0176] From the set of standard expression comment texts, target comment information whose content semantics match the target keyword tags is selected;

[0177] In the preset comment display area of ​​the display interface, the target comment information is displayed in sorted order according to the mention frequency and the semantic sentiment tendency parameter.

[0178] This implementation specifically defines the response flow after a user interacts with a keyword tag. When a user triggers a keyword tag on the interface, the system records the target tag and filters out specific comment information that semantically matches the target tag from a pre-processed set of standard comment texts. Subsequently, these comments are displayed in a specific area of ​​the interface, sorted according to their frequency of mention and semantic sentiment parameters. This process achieves direct and precise penetration from abstract tags to specific comment content. Users do not need to manually browse and search through massive amounts of comments; they can quickly focus on all relevant comments that reflect a specific dimension of evaluation. Furthermore, the system's sorting mechanism prioritizes more popular and positive comments, further improving the efficiency of information acquisition and the quality of the reading experience. Therefore, this implementation, by achieving intelligent association and orderly presentation of tags and comment content, completes the interactive loop from keyword summaries to detailed browsing, greatly enhancing the system's practicality and user-friendliness.

[0179] As an optional implementation, the processing module is also used for incremental updates, specifically for:

[0180] Monitor new review data for the target product;

[0181] When the number of newly added comment information data reaches a preset incremental update threshold, an incremental keyword set is generated based on the newly added comment information data;

[0182] The incremental keyword set is merged and deduplicated with the currently displayed keyword tag sequence to update the keyword tag sequence.

[0183] This implementation adds an incremental update step to the aforementioned keyword generation and display process. By continuously monitoring new review data for the target product, and automatically initiating a new round of keyword extraction based on the new data when the data volume reaches a preset threshold, an incremental keyword set is generated. Subsequently, this incremental set is merged and deduplicated with the currently displayed keyword tag sequence, thus completing the tag sequence update. This mechanism ensures that the keyword tags displayed for the product dynamically evolve with the continuous influx of user feedback, promptly reflecting the latest evaluation trends and changes in reputation, and avoiding the problem of tag content becoming outdated due to review updates. This implementation achieves automated and real-time maintenance of the keyword tag system, ensuring the synchronization of tag information with the current review ecosystem without manual intervention, further enhancing the system's dynamic adaptability and long-term effectiveness.

[0184] As an optional implementation, the specific method by which the processing module merges and deduplicates the incremental keyword set with the currently displayed keyword tag sequence includes:

[0185] Calculate the semantic similarity between each incremental keyword in the incremental keyword set and each existing keyword in the current keyword tag sequence;

[0186] For keyword pairs whose semantic similarity exceeds a third preset threshold, a merging operation is performed, and the incremental keyword or the existing keyword is used as the merged expression based on the mention frequency and semantic sentiment tendency parameters of the incremental keyword and the existing keyword in the keyword pair.

[0187] For keyword pairs whose semantic similarity does not exceed the third preset threshold, the corresponding incremental keyword is added to the keyword tag sequence.

[0188] This implementation further refines the deduplication logic for merging the incremental keyword set with the existing tag sequence. By calculating the semantic similarity between each incremental keyword and each existing keyword in the existing tag sequence, the system can intelligently identify the semantic relationship between new and old keywords. For keyword pairs with semantic similarity exceeding a preset threshold, a merging operation is performed, and the more suitable expression is selected as the representative of the merged pair by comprehensively considering parameters such as frequency of mention and sentiment, thus avoiding tag redundancy. For keywords with semantic similarity below the threshold, they are directly added to the sequence as new tags. This method not only achieves the fusion of new information but also ensures the simplicity and representativeness of the tag set through semantic deduplication and optimization. By introducing semantic similarity calculation and intelligent merging strategies, this implementation makes the incremental update process no longer a simple addition but an evolution with understanding capabilities, thereby continuously maintaining and optimizing the quality of the tag library during dynamic updates.

[0189] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls; they can be fully implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, a processing module can be a separate processing element, or it can be integrated into a chip within the device. Alternatively, it can be stored as program code in the device's memory, and its functions can be called and executed by a processing element. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0190] Indicatively, such as Figure 3 As shown, Figure 3 This is a schematic diagram of the internal structure of a computer device 300 provided in an embodiment of this application. The computer device 300 can be provided as a server. (Refer to...) Figure 3 The computer device 300 includes a processing component 302, which further includes one or more processors, and memory resources represented by memory 301 for storing instructions, such as application programs, that can be executed by the processing component 302. The application programs stored in memory 301 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 302 is configured to execute instructions to perform the methods of any of the embodiments described above.

[0191] The computer device 300 may also include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate on an operating system stored in memory 301, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0192] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0193] This application provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the method provided in any embodiment.

[0194] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0195] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0196] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating and displaying product review keywords, characterized in that, include: Obtain a dataset of review information for the target product, and preprocess the dataset to obtain a set of standard description review texts; The preprocessing includes data cleaning and semantic merging. Data cleaning is used to remove invalid characters or characters that deviate from semantics, and semantic merging is used to merge characters with the same or similar semantics. The standard expression comment text set is input into a pre-trained large language model, and based on the category information associated with the target product, the corresponding scenario-based clustering weight configuration is invoked to perform semantic clustering analysis on the standard expression comment text set to extract an initial keyword set; The initial keyword set is subjected to filtering based on preset filtering conditions to obtain a compliant keyword set; The preset filtering conditions include at least one of keyword length constraints, sentiment constraints, semantic specificity constraints, and mention frequency constraints. The keywords in the compliant keyword set are sorted according to preset sorting conditions to generate a keyword tag sequence; The keyword tag sequence is dynamically displayed on the target product's display interface, and in response to a trigger operation on any keyword tag in the keyword tag sequence, comment information that semantically matches the keyword tag is filtered and displayed.

2. The method according to claim 1, characterized in that, The methods for establishing the scenario-based clustering weight configuration include: Multiple preset product categories are predefined, along with a weight configuration vector corresponding to each preset product category; The weight configuration vector is used to indicate the attention weights of multiple predefined comment attribute dimensions during the semantic clustering analysis. The predefined comment attribute dimensions include at least two of the following: emotional dimension, appearance dimension, functional dimension, material dimension, operation dimension, and experience dimension. Furthermore, based on the category information associated with the target product, the corresponding contextualized clustering weight configuration is invoked to perform semantic clustering analysis on the standard description comment text set, extracting an initial keyword set, including: Based on the category information associated with the target product, the preset product category that matches the target product is determined, and the corresponding weight configuration vector is called to guide the large language model to adjust the degree of attention to each of the predefined comment attribute dimensions in the semantic clustering analysis according to the weight configuration vector, so as to obtain an initial keyword set.

3. The method according to claim 1, characterized in that, The step of performing filtering processing on the initial keyword set based on preset filtering conditions to obtain a compliant keyword set includes: Based on the keyword length constraint, a first candidate keyword with a character count not exceeding a first preset threshold is selected from the initial keyword set; By invoking the sentiment analysis model interface or querying the preset sentiment dictionary, a second candidate keyword with a positive semantic sentiment tendency parameter is selected from the initial keyword set; By using the semantic specificity constraint, words belonging to the preset generalized vocabulary list are excluded from the initial keyword set to obtain the third candidate keyword; By using the mentioned frequency constraint, a fourth candidate keyword with a mention frequency of not less than the second preset threshold is selected from the initial keyword set in the standard expression comment text set; A set of compliant keywords is obtained based on the first candidate keyword, the second candidate keyword, the third candidate keyword, and the fourth candidate keyword.

4. The method according to claim 3, characterized in that, The step of responding to a trigger operation on any keyword tag in the keyword tag sequence, filtering and displaying comment information that semantically matches the keyword tag, includes: Record the target keyword tags that are triggered; From the set of standard expression comment texts, target comment information whose content semantics match the target keyword tags is selected; In the preset comment display area of ​​the display interface, the target comment information is displayed in sorted order according to the mention frequency and the semantic sentiment tendency parameter.

5. The method according to any one of claims 1-4, characterized in that, The method also includes an incremental update step, specifically including: Monitor new review data for the target product; When the number of newly added comment information data reaches a preset incremental update threshold, an incremental keyword set is generated based on the newly added comment information data; The incremental keyword set is merged and deduplicated with the currently displayed keyword tag sequence to update the keyword tag sequence.

6. The method according to claim 5, characterized in that, The process of fusing and deduplicating the incremental keyword set with the currently displayed keyword tag sequence includes: Calculate the semantic similarity between each incremental keyword in the incremental keyword set and each existing keyword in the current keyword tag sequence; For keyword pairs whose semantic similarity exceeds a third preset threshold, a merging operation is performed, and the incremental keyword or the existing keyword is used as the merged expression based on the mention frequency and semantic sentiment tendency parameters of the incremental keyword and the existing keyword in the keyword pair. For keyword pairs whose semantic similarity does not exceed the third preset threshold, the corresponding incremental keyword is added to the keyword tag sequence.

7. A device for generating and displaying product review keywords, characterized in that, include: The processing module is used to acquire a dataset of review information for the target product and preprocess the dataset to obtain a set of standard description review texts. The preprocessing includes data cleaning and semantic merging. Data cleaning is used to remove invalid characters or characters that deviate from semantics, and semantic merging is used to merge characters with the same or similar semantics. The processing module is also used to input the standard expression comment text set into a pre-trained large language model, and based on the category information associated with the target product, call the corresponding scenario-based clustering weight configuration to perform semantic clustering analysis on the standard expression comment text set and extract an initial keyword set; The processing module is further configured to perform filtering processing on the initial keyword set based on preset filtering conditions to obtain a compliant keyword set; The preset filtering conditions include at least one of keyword length constraints, sentiment constraints, semantic specificity constraints, and mention frequency constraints. The processing module is also used to sort each keyword in the compliant keyword set according to preset sorting conditions and generate a keyword tag sequence; The processing module is also configured to dynamically display the keyword tag sequence on the display interface of the target product, and in response to the triggering operation of any keyword tag in the keyword tag sequence, filter and display comment information that semantically matches the keyword tag.

8. The apparatus according to claim 7, characterized in that, The specific methods by which the processing module establishes scenario-based clustering weight configurations include: Multiple preset product categories are predefined, along with a weight configuration vector corresponding to each preset product category; The weight configuration vector is used to indicate the attention weights of multiple predefined comment attribute dimensions during the semantic clustering analysis. The predefined comment attribute dimensions include at least two of the following: emotion dimension, appearance dimension, function dimension, material dimension, operation dimension, and experience dimension. Furthermore, based on the category information associated with the target product, the corresponding contextualized clustering weight configuration is invoked to perform semantic clustering analysis on the standard description comment text set, extracting an initial keyword set, including: Based on the category information associated with the target product, the preset product category that matches the target product is determined, and the corresponding weight configuration vector is called to guide the large language model to adjust the degree of attention to each of the predefined comment attribute dimensions in the semantic clustering analysis according to the weight configuration vector, so as to obtain an initial keyword set.

9. A computer device, characterized in that, The method includes one or more processors and a memory storing computer-readable instructions that, when executed by the one or more processors, perform the steps of the method as described in any one of claims 1-6.

10. A storage medium, characterized in that, The storage medium stores computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the method as described in any one of claims 1-6.

Images