An artificial intelligence-based multi-source brand data fusion analysis method and system

Abstract

The present application relates to the technical field of artificial intelligence and data fusion, and provides a multi-source brand data fusion analysis method and system based on artificial intelligence, which comprises the following steps: performing corresponding semantic bridging processing, mis-association risk elimination processing or sudden reinforcement processing according to the association problem type to obtain an association processing result; outputting the association result in layers and updating a brand object relationship graph according to the association processing result and the association confidence; obtaining incremental comment data, reviewing the to-be-confirmed association and the independent observation node based on the incremental comment data to obtain a review result, and adjusting the association state in the brand object relationship graph according to the review result; outputting the brand object relationship graph after the adjustment of the association state, and evaluating the brand visibility based on the association state between the brand model and the non-official indicator expression in the brand object relationship graph. The present application has the effect of improving the reliability of brand competition situation analysis.

Description

Technical Field

[0001] This invention relates to the technical field of artificial intelligence and data fusion, specifically to a method and system for multi-source brand data fusion analysis based on artificial intelligence. Background Technology

[0002] In multi-platform competitor monitoring, to integrate scattered brand basic information, competitor data, and historical monitoring samples to support visibility assessment, existing technologies typically employ AI-based multi-source brand data fusion techniques. This technology extracts and semantically encodes product models from the brand's own descriptive text, generating standard vector representations. It then processes competitor comments line by line, extracting and encoding referential terms. Based on vector similarity, it links referential terms in competitor comments to existing brand entity nodes, thereby establishing a dependency relationship between competitor user discussions and brand models.

[0003] However, in real-world scenarios involving the fusion of competitor reviews for mobile phone brands, e-commerce reviews are naturally filled with non-standardized colloquial expressions. Competitor users often use unofficial, colloquial names or pinyin abbreviations not included in the brand's own training data to refer to the brand's own models. When the fusion technology performs semantic encoding on these colloquial names and attempts to link them across source entities, the lack of contextual information about co-occurrence with the official names and the significant differences in their literal forms cause the fusion model to place the vectors of colloquial names far from the standard model vectors in the embedding space. This leads to the colloquial names being incorrectly identified as independent new entities or mistakenly associated with competitor models with similar pronunciations during the coreference resolution step.

[0004] Current AI data fusion technologies primarily rely on fixed vector distance thresholds and preset literal similarity for coreference alignment, making it difficult to autonomously perceive and bridge the significant semantic gap caused by the proliferation of colloquialisms. This coreference breakage causes dependency edges in competitor reviews actually discussing the brand / model in the object relationship graph to be connected to irrelevant entities or dangling nodes, resulting in a significant lack of competitor feedback dependency edges for that brand / model and potentially misassigning dependency relationships to competitor entities. Ultimately, dynamic monitoring models overlook large-scale competitor discussions when calculating brand visibility, severely distorting the brand competitive landscape. Therefore, in multi-platform competitor review fusion scenarios for mobile phone brands, when competitor users extensively use unofficial colloquialisms to refer to the brand's own models, the AI ​​multi-source data fusion technology cannot establish correct coreference mappings in the embedding space based on literal similarity and contextual information. This leads to a significant lack of competitor feedback dependency edges for that brand / model and potentially misassigning dependency relationships to competitor entities, directly impacting the completeness and accuracy of brand visibility assessment.

[0005] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0006] This application discloses an AI-based multi-source brand data fusion analysis method and system, aiming to solve the problem in the existing technology that, in the scenario of fusion of competitor reviews across multiple platforms for mobile phone brands, when a large number of competitor users use unofficial, colloquial terms to refer to the brand's own models, the AI ​​multi-source data fusion technology cannot establish a correct common reference mapping in the embedding space based on literal similarity and contextual information. This results in the brand model lacking a large number of competitor feedback dependency edges and may incorrectly assign dependency relationships to competitor entities, directly affecting the completeness and accuracy of brand visibility assessment.

[0007] The technical solution of this application is as follows: Firstly, this application discloses a multi-source brand data fusion analysis method based on artificial intelligence, including: Acquire and store benchmark information of brand and model, identify unofficial representatives to be associated from competitor review data from multiple platforms, and generate comprehensive feature information based on the literal features, contextual semantic features, comment subject role features, platform source features, and historical similar expression features of unofficial representatives; A candidate association set is generated based on comprehensive feature information and benchmark information. The candidate association set includes multiple candidate associations, which are unverified associations between an unofficial representative and a brand model. Association evidence and association confidence are calculated for the candidate associations in the candidate association set. Association evidence includes literal similarity, semantic similarity, historical co-occurrence trend, cross-comment consistency, competitive role matching degree, short-term occurrence rate, and cross-platform synchronous occurrence ratio. Based on the associated evidence and the confidence level of association, identify the types of association issues between unofficial representatives and brand models. These types of association issues include regular association with missing bridging information, risk of mis-association with competitors, or association with sudden missing bridging information. Based on the type of associated problem, perform corresponding semantic bridging processing, erroneous association risk mitigation processing, or sudden reinforcement processing to obtain the association processing result; Based on the association processing results and association confidence, the association results are output hierarchically and the brand object relationship graph is updated. The association results include high-confidence associations, associations to be confirmed, and independent observation nodes. Acquire incremental comment data, review the unconfirmed associations and independent observation nodes based on the incremental comment data, obtain the review results, and adjust the association status in the brand object relationship diagram according to the review results; Output the brand object relationship diagram after adjusting the association status, and evaluate brand visibility based on the association status between brand model and unofficial representative in the brand object relationship diagram.

[0008] Secondly, this application also discloses an artificial intelligence-based multi-source brand data fusion analysis system for performing artificial intelligence-based multi-source brand data fusion analysis, including: The comprehensive feature generation module is used to acquire and store the baseline information of the brand and model, identify the unofficial representative to be associated from the competitive product review data of multiple platforms, and generate comprehensive feature information based on the literal features, contextual semantic features, comment topic role features, platform source features and historical similar expression features of the unofficial representative. The candidate association generation module is used to generate a set of candidate associations based on comprehensive feature information and benchmark information. The set of candidate associations includes multiple candidate associations, which are unverified associations between an unofficial representative and a brand model. For the candidate associations in the set of candidate associations, association evidence and association confidence are calculated. Association evidence includes literal similarity, semantic similarity, historical co-occurrence trend, cross-comment consistency, competitive role matching degree, short-term occurrence rate, and cross-platform synchronous occurrence ratio. The issue type identification module is used to identify the type of association issues between unofficial representatives and brand models based on the associated evidence and association confidence level. The associated issue types include regular association with missing bridging information, risk of misassociation with competitors, or association with sudden missing bridging information. The processing result generation module is used to perform corresponding semantic bridging processing, false association risk mitigation processing, or sudden reinforcement processing according to the type of associated problem, and obtain the association processing result; The relationship graph update module is used to output the association results hierarchically and update the brand object relationship graph based on the association processing results and association confidence. The association results include high-confidence associations, associations to be confirmed, and independent observation nodes. The association status adjustment module is used to acquire incremental comment data, review the associations and independent observation nodes to be confirmed based on the incremental comment data, obtain the review results, and adjust the association status in the brand object relationship diagram according to the review results; The brand visibility assessment module is used to output a brand object relationship diagram after adjusting the association status, and to assess brand visibility based on the association status between brand models and unofficial representatives in the brand object relationship diagram.

[0009] This technical solution provides a system-level solution. Through modular design, it enables the fusion and analysis of multi-source brand data, effectively solving the problem of co-pointing fragmentation caused by unofficial endorsements, improving the accuracy and completeness of brand visibility assessment, and exhibiting good scalability and maintainability.

[0010] Beneficial Effects: This application discloses an AI-based multi-source brand data fusion analysis method. By acquiring and storing benchmark information of brand models, it identifies unofficial representatives to be associated from competitor review data across multiple platforms. Based on the literal features, contextual semantic features, comment subject role features, platform source features, and historical similarity expression features of the unofficial representatives, it generates comprehensive feature information. On this basis, it generates a candidate association set based on the comprehensive feature information and benchmark information, and calculates association evidence and association confidence for each candidate association in the set. Further, based on the association evidence and association confidence, it identifies the association problem type between unofficial representatives and brand models, and performs corresponding semantic bridging processing, false association risk mitigation processing, or sudden reinforcement processing according to the association problem type to obtain the association processing result. Finally, based on the association processing results and association confidence, the association results are output hierarchically and the brand object relationship graph is updated. The associations to be confirmed and independent observation nodes are reviewed through incremental comment data, and the association status in the brand object relationship graph is adjusted. Thus, the brand object relationship graph after the association status is adjusted is output. Based on the association status between brand models and unofficial representatives in the brand object relationship graph, brand visibility is evaluated.

[0011] Through the above technical solution, this application effectively solves the problem of coreference breakage caused by unofficial and colloquial expressions in existing technologies, overcoming the limitations of traditional methods in bridging the significant semantic gap. Specifically, this application generates more comprehensive feature information by comprehensively considering the literal features, contextual semantic features, comment topic role features, platform source features, and historical similarity expression features of unofficial representatives, significantly improving the accuracy of identifying unofficial representatives. Furthermore, by introducing multiple types of association evidence (such as literal proximity, semantic proximity, historical co-occurrence trends, cross-comment consistency, competitor role matching, short-term occurrence rate, and cross-platform synchronous occurrence ratio) and association confidence calculation, it can more precisely assess the association strength between unofficial representatives and brand models. More importantly, this application can perform corresponding differentiated processing based on the identified association problem type (regular bridging information missing association, competitor mis-association risk, or sudden bridging information missing association), including semantic bridging processing, mis-association risk mitigation processing, or sudden reinforcement processing, thereby specifically solving different types of association problems. For example, regarding the risk of misattributing brand discussions to competitors, it effectively avoids incorrectly attributing brand discussions to competitors; for sudden missing bridging information, it can quickly respond and supplement the relevant information. Ultimately, by outputting association results in a layered manner and dynamically updating the brand object relationship graph, and by reviewing and adjusting it based on incremental comment data, the accuracy and timeliness of the brand object relationship graph are ensured. This makes brand visibility assessment more complete and accurate, effectively avoiding the problems of omitting large-scale competitor discussions or incorrectly assigning dependencies, and significantly improving the reliability of brand competitive landscape analysis. Attached Figure Description

[0012] Figure 1 This is a flowchart of a multi-source brand data fusion and analysis method based on artificial intelligence, as described in one embodiment of the present invention. Figure 2 This is a flowchart of a multi-source brand data fusion analysis method based on artificial intelligence, as described in another embodiment of the present invention. Figure 3 This is a system block diagram of a multi-source brand data fusion and analysis system based on artificial intelligence, according to another embodiment of the present invention. Explanation of reference numerals in the attached figures: 1. AI-based multi-source brand data fusion and analysis system; 11. Comprehensive feature generation module; 12. Candidate association generation module; 13. Problem type identification module; 14. Processing result generation module; 15. Relationship graph update module; 16. Association status adjustment module; 17. Brand visibility assessment module. Detailed Implementation

[0013] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0014] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0015] A multi-source brand data fusion and analysis method based on artificial intelligence, combined with Figure 1 As shown, it includes: S1. Acquire and store the baseline information of the brand and model, identify the unofficial representative to be associated from the competitor's review data from multiple platforms, and generate comprehensive feature information based on the literal features, contextual semantic features, comment subject role features, platform source features and historical similar expression features of the unofficial representative; S2, Generate a candidate association set based on comprehensive feature information and benchmark information; the candidate association set includes multiple candidate associations, which are unverified associations between an unofficial representative and a brand model; calculate association evidence and association confidence for the candidate associations in the candidate association set, and the association evidence includes literal similarity, semantic similarity, historical co-occurrence trend, cross-comment consistency, competitive role matching degree, short-term occurrence rate, and cross-platform synchronous occurrence ratio; S3. Based on the associated evidence and the confidence level of the association, identify the types of association issues between unofficial representatives and brand models. The types of association issues include the association of missing bridging information, the risk of mis-association with competitors, or the association of missing bridging information in a sudden situation. S4. Based on the type of association problem, perform corresponding semantic bridging processing, false association risk mitigation processing, or sudden reinforcement processing to obtain the association processing result. S5, based on the association processing results and association confidence, outputs association results in layers and updates the brand object relationship graph. The association results include high-confidence associations, associations to be confirmed, and independent observation nodes. S6, acquire incremental comment data, review the unconfirmed associations and independent observation nodes based on the incremental comment data, obtain the review results, and adjust the association status in the brand object relationship diagram according to the review results; S7 outputs the brand object relationship diagram after adjusting the association status, and evaluates brand visibility based on the association status between brand model and unofficial representative in the brand object relationship diagram.

[0016] "Basic information on brand models" refers to standardized information about product models officially released by the brand or confirmed by the brand, such as official names, aliases, product attributes, and historical discussion data. This basic information provides a foundation for identifying, comparing, and associating unofficial representatives. "Unofficial representatives" refers to brand model expressions used by users in comments that are not the brand's official standard names. These can manifest as colloquial terms, abbreviations, homophones, nicknames, misspellings, or popular platform abbreviations. "Comprehensive feature information" refers to the feature set obtained after multi-dimensional analysis of unofficial representatives, which can include literal features, contextual semantic features, comment topic / role features, platform source features, and historical similarity expression features. Among them, literal features can reflect the similarity of characters, pinyin, abbreviations or homophones between the unofficial representative and the brand model name; contextual semantic features can reflect the meaning of the unofficial representative in the comment context; comment topic role features can reflect the role of the unofficial representative in comment topics such as performance, price, appearance, and competitor comparison; platform source features can reflect that the comment comes from e-commerce platforms, social media or forum communities; historical similarity expression features can reflect expressions similar to the current unofficial representative in historical data and their related records.

[0017] The "candidate association set" refers to the set of combinations of unofficial brand representatives and brand models that the system initially identifies as potentially related. Each combination can be called a candidate association. "Association evidence" refers to the set of indicators used to measure the credibility of candidate associations, which may include literal similarity, semantic similarity, historical co-occurrence trend, cross-comment consistency, competitive role matching degree, short-term occurrence rate, and cross-platform synchronous occurrence ratio. Literal similarity indicates the degree of similarity between the unofficial brand representative and the brand model name at the character, pinyin, or abbreviation level; semantic similarity indicates the degree of proximity between the two in the semantic space; historical co-occurrence trend indicates the frequency and trend of their co-occurrence or associated occurrence in historical comment data; cross-comment consistency indicates the stability of the same unofficial brand representative pointing to the same brand model in different comments; competitive role matching degree indicates the degree of matching between the product role assumed by the unofficial brand representative in the comments and the competitive positioning of the brand model; short-term occurrence rate indicates the frequency of concentrated occurrence of unofficial brand representatives in a short period; and cross-platform synchronous occurrence ratio indicates the degree to which unofficial brand representatives appear synchronously across multiple platforms. "Association confidence level" refers to the result obtained by quantifying the credibility of the association between unofficial representatives and brand models after synthesizing association evidence.

[0018] "Association Issue Types" refers to the classification of issues existing in the association between unofficial representatives and brand models. These can include regular missing bridging information associations, competitor misassociation risks, and sudden missing bridging information associations. Regular missing bridging information associations refer to situations where there is insufficient direct evidence of a link between the unofficial representative and the brand model, making it difficult for the system to directly confirm their relationship. Competitor misassociation risks refer to situations where unofficial representatives may be incorrectly associated with competitor models due to literal similarity, semantic similarity, or mixed commentary context. Sudden missing bridging information associations refer to situations where a large number of unofficial representatives appear in a short period, but there are no corresponding association records in historical data, making it difficult for the system to establish a stable association temporarily. The "Brand Object Relationship Graph" is a graph structure built with brand models and unofficial representatives as nodes and the association relationships between them as edges. Edges can carry attributes such as association confidence, association issue type, association status, and update time. "Brand visibility" refers to an indicator used to measure the degree to which a brand or brand model is mentioned, identified, and associated in market reviews, user discussions, and competitor contexts. It can be evaluated based on the association status between brand models and unofficial representatives in the brand object relationship diagram.

[0019] The method proposed in this application first acquires and stores benchmark information for brand models. In practice, this can be achieved through manual input, official brand data interfaces, brand official websites, e-commerce platform product pages, or product databases, obtaining the official name, official aliases, product attributes, and historical discussion data of the brand model. Product attributes may include processor model, screen size, release date, price range, feature configuration, color version, or target audience. The system cleans, deduplicates, standardizes, and structures the acquired benchmark information, forming a benchmark information database for brand models. This database provides a standard reference for subsequent identification of unofficial expressions, enabling the system to compare colloquial expressions, abbreviations, or homophones in user comments with brand models.

[0020] Subsequently, the system identifies unofficial references to be associated from competitor review data across multiple platforms. Specifically, the system can obtain competitor review data from e-commerce platforms, social media, forums, short video comment sections, or review article comment sections, and perform word segmentation, cleaning, entity recognition, and phrase extraction on the review data. Through named entity recognition models, keyword extraction rules, phrase discovery algorithms, or user comment context analysis, the system can extract unofficial words or phrases that may refer to a brand or model from the reviews. For example, users may use abbreviations of brand or model names, initials of pinyin, nicknames based on appearance characteristics, price range names, or humorous aliases to refer to a product model. The system identifies these expressions that may refer to a brand or model but have not yet been officially standardized as unofficial references to be associated.

[0021] Next, comprehensive feature information is generated based on the literal features of the unofficial representative, contextual semantic features, comment subject role features, platform source features, and historical similar expression features. In practice, literal features can be extracted using string matching, edit distance, pinyin similarity, homophonic rules, or abbreviation matching to determine if there is a formal similarity between the unofficial representative and the official name or alias of the brand model. For contextual semantic features, word vector models, sentence vector models, or contextual semantic models can be used to extract adjacent words, modifiers, evaluation words, and comparison objects in the comments containing the unofficial representative, forming semantic vectors. For comment subject role features, topic classification models or comment intent recognition models can be used to determine if the comment mainly discusses product performance, price, appearance, battery life, screen, after-sales service, or competitor comparisons, further determining whether the unofficial representative serves as an evaluated object, a comparison object, or a substitute object in the comment. Platform source features can directly record the comment's source platform, comment time, comment location, and platform type. Historical similar expression features can be obtained by querying historical comment data and historical association records to obtain expressions similar to the current unofficial representative and their existing association results. The above-mentioned multidimensional features together form comprehensive feature information.

[0022] Furthermore, a candidate association set is generated based on comprehensive feature information and benchmark information of brand models. Specifically, the system compares the comprehensive feature information corresponding to unofficial product names with the benchmark information of brand models to filter out brand models that may have a relationship. For example, when an unofficial product name is similar to an alias of a brand model in terms of literal features, matches the product attributes of that model in terms of contextual semantics, and represents a performance evaluation of that model in terms of the comment subject role, the unofficial product name can be combined with that brand model to form a candidate association. The candidate association set can be generated through rule matching, machine learning classifiers, semantic retrieval models, or deep learning matching models. Through the candidate association set, the system can retain multiple possible results before formally determining the association relationship, so as to filter them later by combining association evidence and association confidence.

[0023] For candidate associations in the candidate association set, association evidence and association confidence are calculated. In specific implementation, the system can calculate association evidence such as literal similarity, semantic similarity, historical co-occurrence trend, cross-comment consistency, competitor role matching degree, short-term occurrence rate, and cross-platform synchronous occurrence ratio. Semantic similarity can be obtained by calculating the cosine similarity between the semantic vector of the unofficial representative's context and the semantic vector of the brand model's baseline information; historical co-occurrence trend can be obtained by statistically analyzing the frequency, time distribution, and growth trend of the co-occurrence of the unofficial representative and the brand model in historical comment data; cross-comment consistency can be determined by analyzing whether the same unofficial representative consistently points to the same brand model in multiple comments; competitor role matching degree can be determined by judging whether the product role of the unofficial representative in the comments is consistent with the brand model's positioning, price range, or competitor relationship; short-term occurrence rate can be obtained by monitoring the frequency changes of the unofficial representative within a day, a week, or other time windows; and cross-platform synchronous occurrence ratio can be obtained by statistically analyzing the degree to which the unofficial representative appears synchronously on multiple platforms. Based on the aforementioned evidence of association, the system performs comprehensive calculations using weighted summation, logistic regression models, gradient boosting models, or neural network models to obtain the association confidence level corresponding to the candidate association.

[0024] Subsequently, based on the association evidence and association confidence level, the types of association issues between unofficial representative products and brand models are identified. Specifically, when the association confidence level of a candidate association is low, and direct bridging evidence such as historical co-occurrence trends and cross-comment consistency is insufficient, it can be identified as a regular association with missing bridging information. If an unofficial representative product has a high degree of matching with a competitor's model in terms of literal similarity, semantic similarity, or comment topic role, but the association evidence with the target brand model is weak, it can be identified as a risk of mis-association with a competitor. If the frequency of unofficial representative products increases significantly in a short period of time and appears simultaneously on multiple platforms, but historical similarity features and historical co-occurrence trends are insufficient to support a stable association, it can be identified as a sudden association with missing bridging information. Identifying the types of association issues provides a basis for selecting corresponding handling methods.

[0025] Based on the type of association issue, corresponding semantic bridging processing, false association risk mitigation processing, or sudden reinforcement processing are performed to obtain the association processing result. In specific implementation, for associations with missing bridging information, semantic bridging relationships between unofficial representatives and brand models can be established through knowledge graph reasoning, similar expression expansion, comment context supplementation analysis, or manual annotation. For the risk of false association with competitors, competitor entity identification, competitor model benchmark information comparison, and semantic relationship analysis can be introduced to determine whether the unofficial representative actually points to a competitor model or whether it is used in comments to compare the target brand model with competitor models, thereby reducing the probability of false association. For associations with sudden missing bridging information, real-time monitoring and information reinforcement mechanisms can be activated, such as continuously collecting incremental comment data, tracking short-term dissemination paths, analyzing the initial platform and core context, and quickly supplementing the association evidence for the unofficial representative. After the above processing, the system obtains an association processing result indicating whether the association is valid, whether it needs review, or whether the association should be temporarily suspended.

[0026] Next, based on the association processing results and association confidence levels, the association results are output hierarchically, and the brand object relationship graph is updated. Specifically, when the association confidence level of a candidate association is higher than the preset high confidence threshold, and the association processing results support the establishment of the association, it can be marked as a high-confidence association, and an association edge between the unofficial representative node and the brand model node can be established or updated in the brand object relationship graph. When the association confidence level is in the middle range, or the association processing results show that there is still some uncertainty, it can be marked as a relationship to be confirmed, and an entry point for manual review or subsequent data verification can be retained. When the association confidence level is low, and the existing evidence is insufficient to support the establishment of a stable association with any brand model, the unofficial representative can be marked as an independent observation node, and a strong association with the brand model will not be established for the time being, but its frequency of occurrence, contextual changes, and cross-platform dissemination will continue to be monitored. Nodes in the brand object relationship graph can represent brand models and unofficial representatives, and edges can represent the association relationship between the two. The edge weights can be dynamically updated based on the association confidence level, the strength of association evidence, and the review results.

[0027] Subsequently, incremental comment data is acquired, and the relationships to be confirmed and independent observation nodes are reviewed based on this data to obtain review results. The relationship status in the brand object relationship graph is then adjusted according to these results. In practice, the system can periodically or in real-time acquire new comment data and re-extract contextual information, co-occurrence information, and cross-platform occurrence related to the relationships to be confirmed or independent observation nodes. For relationships to be confirmed, if the incremental comment data provides stronger literal evidence, semantic evidence, historical co-occurrence trends, or cross-comment consistency, it can be upgraded to a high-confidence relationship; if the incremental comment data indicates that it actually points to other competitor models, the confidence of the original relationship can be reduced or erroneous relationships can be removed. For independent observation nodes, if the incremental comment data shows clear brand model co-occurrence, stable semantic pointing, or synchronous propagation characteristics across platforms, new candidate relationships can be established, and the relationship evidence and confidence can be further calculated. Through this review mechanism, the brand object relationship graph can be continuously adjusted as comment data grows, avoiding the long-term stagnation of a one-time identification result.

[0028] Finally, the system outputs a brand object relationship diagram with adjusted association status, and assesses brand visibility based on the association status between brand models and unofficial representatives in the diagram. Specifically, the system can output the adjusted brand object relationship diagram in the form of a visualization, data table, interface data, or report, allowing brands to view the unofficial representatives, association confidence levels, associated question types, and dissemination platforms corresponding to each brand model. Brand visibility can be calculated based on indicators such as the number of unofficial representatives associated with a specific brand model, the frequency of unofficial representatives appearing in comments, the weighted sum of association confidence levels, the proportion of simultaneous appearance across platforms, and the growth trend of high-confidence associations. Through the above processing, this application can identify unofficial representatives from multi-platform competitor comment data, and perform hierarchical association, dynamic verification, and graph-based management with brand models, thereby improving the accuracy of brand sentiment analysis, competitor identification, and brand visibility assessment.

[0029] Optionally, when the association issue type is a sudden missing bridging information association, sudden reinforcement processing is performed according to the association issue type to obtain the association processing result, including: Filter the core comments from the current comment stream and identify them as sources of supplementary information; Obtain the content quality assessment results of the supplementary information sources; Based on the candidate association set, identify the brand models that have candidate association relationships with unofficial representatives, and use the benchmark information corresponding to the brand models as known entity features; A temporary association network is constructed, which includes unofficial indicator delivery nodes, known entity nodes, and supplementary information source nodes. Connection edges are established based on the co-occurrence relationship or semantic orientation relationship between unofficial indicator delivery, known entity characteristics, and supplementary information sources. The node weights of the supplementary information source nodes are determined based on the content quality assessment results of the supplementary information source, and the edge weights are determined based on the co-occurrence strength or semantic pointing strength of the connecting edges. Based on node weights and edge weights, graph structure features are aggregated in the temporary association network. The features of known entity nodes are aggregated to the unofficial index delivery node through the reinforcement information source node to obtain the reinforced unofficial index delivery features. The authority of the supplementary information sources from different platforms and the credibility of the comment publishers were evaluated to obtain the platform credibility evaluation results; Based on the platform credibility assessment results, adjust the cross-validation weights of information sources from different platforms to strengthen the information source; Based on cross-validation weights, consistency verification is performed on evidence of the association between unofficial representatives and brand models in supplementary information sources from different platforms to obtain cross-platform consistency verification results. Based on the enhanced unofficial indicator delivery characteristics and cross-platform consistency verification results, the association confidence between the unofficial indicator delivery and the brand model is updated to obtain the corresponding association confidence. Based on the association confidence level, the corresponding association processing results are generated.

[0030] Specifically, when the system identifies a problem type as a sudden missing bridging information, it initiates a dedicated sudden reinforcement process. First, it filters comments with high informational value and representativeness from the current real-time or recent comment stream; these comments are identified as reinforcement information sources. These reinforcement information sources are key data used to fill in sudden information gaps; for example, they can be filtered based on comment interaction volume, information density, or keyword matching.

[0031] Furthermore, the content quality of these supplementary information sources needs to be evaluated to ensure their reliability. The content quality assessment results can reflect the objectivity, information density, emotional tone, and presence of misleading expressions in the comments, thus providing a basis for subsequent weighting.

[0032] Simultaneously, based on the existing candidate association set, the system will identify brand models that may have candidate association relationships with the unofficial representatives to be processed. The baseline information corresponding to these brand models, such as official names, aliases, attribute information, and semantic features of historical discussions, is used as known entity features and as anchor points for constructing the association network.

[0033] Based on this, a temporary association network is constructed. This network consists of unofficial representative nodes, known entity nodes, and supplementary information source nodes. The connections between nodes are established based on their co-occurrence relationships (e.g., appearing together in the same comment) or semantic pointing relationships (e.g., explicitly mentioned or implied in the comment content).

[0034] To quantify the importance of information in the network, the node weights of the supplementary information source nodes are determined based on the content quality assessment results, with higher-quality comments receiving higher weights. Simultaneously, the co-occurrence strength or semantic orientation strength of the connecting edges are also used to determine the edge weights, representing the tightness of the association.

[0035] Subsequently, through graph structure feature aggregation technology, the feature information of known entity nodes is transmitted and aggregated to the unofficial representative node via the reinforcement information source node. This process enables the unofficial representative to absorb contextual information from known brand models and high-quality reviews, thereby obtaining reinforced unofficial representative features, making its semantic representation richer and more accurate.

[0036] Given that comment data may originate from multiple platforms, to improve the reliability of the correlation, it is also necessary to assess the platform authority and comment publisher credibility of supplementary information sources from different platforms to obtain platform credibility assessment results. For example, official forums or well-known media platforms are generally more authoritative than personal blogs or anonymous forums, while active publishers with high-quality historical comments are more credible.

[0037] Based on the platform's credibility assessment results, the cross-validation weights of information sources from different platforms can be dynamically adjusted. This means that evidence from more credible platforms will receive higher weights to reflect its importance in association determination.

[0038] Based on the adjusted cross-validation weights, consistency verification is performed on evidence regarding the association between unofficial representatives and brand models from different platform sources. For example, if multiple highly authoritative platforms point to the same association, the consistency verification result will be higher, thereby enhancing the confidence of the association.

[0039] Finally, combining the enhanced unofficial representative characteristics and cross-platform consistency verification results, the confidence level of the association between unofficial representative and brand model is updated. This updated confidence level will more accurately reflect the reliability of the association between the two in the event of sudden information loss.

[0040] Based on the updated association confidence, the corresponding association processing result is generated, which will be used for subsequent brand object relationship graph updates and brand visibility assessment.

[0041] Optional, combined Figure 2 As shown, the steps for filtering core comments from the current comment stream and identifying them as supplementary information sources include: A1, perform content quality assessment on the comments in the current comment stream to obtain content quality assessment results. The content quality assessment includes identifying the emotional bias, misleading expressions, and information density in the comments; A2, retrieves the interaction count of comments in the current comment stream; A3: The comments are given a comprehensive score based on the evaluation results of interaction volume and content quality. A4: Based on the overall score, select the core comments from the current comment stream; A5 uses core comments as a source of supplementary information.

[0042] Specifically, content quality assessment aims to measure the validity and reliability of comments. Identifying the emotional bias in comments helps determine whether they are positive, negative, or neutral, thus assessing their potential bias or objectivity. Identifying misleading expressions focuses on whether comments contain false information, exaggerations, or misleading statements, which is crucial for ensuring the accuracy of supplementary information sources. Information density refers to the amount of effective information contained in a comment; comments with high information density typically provide richer context and stronger relevance evidence.

[0043] Furthermore, the level of interaction with a comment, such as the number of likes, shares, and comments, can reflect its level of attention and influence. Comments with high interaction levels often represent broader consensus or stronger dissemination power, and therefore have greater reference value when selecting core comments.

[0044] Therefore, the overall score is generated by weighting the content quality assessment results with the interaction volume. For example, one weight can be assigned to content quality, and another weight to interaction volume, and then the two are added together to obtain the final overall score. This scoring mechanism ensures that the selected core comments not only have high-quality content, but also have a certain social impact or attention.

[0045] Finally, by comprehensively scoring the comments, they can be ranked, and a threshold can be set or the top N comments can be selected as core comments. These core comments are considered the most representative, reliable, and informative, and are therefore identified as supplementary information sources for subsequent emergency reinforcement processing.

[0046] Optionally, when the association issue type is "competitor mis-association risk," the mis-association risk resolution process is performed according to the association issue type to obtain the association processing result, including: Identify the semantic relationships between unofficial statements and competing entities that co-occur in comments; Determine whether the semantic relationships include a contrastive, ironic, or contrasting relationship; Based on the co-occurrence frequency, text distance, and semantic intensity of unofficial statements and competing entities in comments, the direct referential association weights between unofficial statements and competing entities are determined. When there is a contrast, irony, or reverse comparison relationship, the direct referential association weight is reduced according to the semantic relationship to obtain the association weight after competitor elimination. Analyze the potential implications of unofficial statements regarding brand and model names in the comments; Based on the potential target intent, determine the brand target association weight between unofficial target expression and brand model; Based on the association weights after competitor elimination and the brand-pointing association weights, generate the association handling results for the risk of mis-association of competitors; When there is no contrast, irony, or reverse comparison relationship, the direct reference association weight is retained, and the corresponding candidate association is marked as a competitor association to be confirmed based on the direct reference association weight. Based on the competitor associations to be confirmed, generate the association handling results for the risk of erroneous association with competitors.

[0047] Specifically, when performing the risk mitigation of misassociations, the first step is to identify the semantic relationship between the unofficial representative and the competing entities that appear together in the review. This step aims to understand how the unofficial representative and the competing entities are related in the review text, such as whether they are simply listed side-by-side, compared, or have more complex semantic interactions. Further, it is determined whether the identified semantic relationships involve contrast, irony, or reverse framing. These rhetorical devices are common in consumer reviews and may cause the unofficial representative to appear superficially related to competing entities, but actually emphasize the advantages or disadvantages of the brand's model. For example, using phrases like "much better than XXX" to express contrast, "XXX is simply a waste of money" to express irony, or describing the shortcomings of competitors to highlight the advantages of the brand's model.

[0048] Based on this, the direct referential association weights between unofficial sources and competing entities are determined according to the co-occurrence frequency, textual distance, and semantic referential strength of unofficial sources and competing entities in comments. Co-occurrence frequency reflects how often the two appear together; textual distance measures their proximity in the text; and semantic referential strength analyzes the tightness of their semantic association through natural language processing techniques. These factors collectively quantify the likelihood that unofficial sources directly refer to competing entities.

[0049] When a contrasting, ironic, or indirect comparison relationship is identified, the direct referential association weight is reduced based on the identified semantic relationship to obtain the association weight after competitor deconstruction. This is because these rhetorical devices indicate that while the unofficial representative mentions competitors, their core intention is not to directly refer to competitors, but rather to indirectly express their opinion on the brand's model through competitors. Therefore, it is necessary to weaken this superficial direct association.

[0050] Simultaneously, this analysis examines the potential intended meaning of unofficial representatives regarding brand models in their comments. This analysis aims to uncover the true target of these unofficial representatives from the overall context of the comments, even when they appear alongside competing entities.

[0051] Finally, based on the reduced association weights of competitors and the brand-related association weights, a result for handling the risk of mis-association with competitors is generated. This result comprehensively considers the weakened association between unofficial representatives and competitor entities, as well as their potential association with the brand's model, thus more accurately assessing their true referent.

[0052] As a supplement, when there is no contrast, irony, or reverse comparison relationship, the direct referential association weight is retained, and the corresponding candidate association is marked as a competitor association to be confirmed based on the direct referential association weight. This means that if there is no obvious rhetorical relationship between an unofficial expression and a competitor entity, its direct association weight is considered valid, but further confirmation is still required to prevent the omission of other forms of erroneous associations. Based on the competitor associations to be confirmed, the association processing results for the risk of erroneous competitor associations are generated.

[0053] Optionally, the steps for obtaining and storing the brand and model reference information include: Obtain the official name, aliases, attribute information, and historical discussion semantic features of the brand model to construct the benchmark information of the brand model. The benchmark information includes name features, attribute features, and historical semantic feature vectors. Store baseline information; Periodically perform time-decay weighted processing on the historical semantic feature vectors to obtain the decayed historical semantic feature vectors. Extract recent comments from incremental comment data; Extract semantic feature vectors from recent comments based on text data related to brand and model. Calculate the semantic offset between the semantic feature vector of recent comments and the baseline information; Based on the semantic offset and the number of recent comments, determine the update rate and update magnitude of the semantic features in the baseline information; Based on the update rate and update magnitude, the historical semantic feature vector after decay processing is fused with the recent comment semantic feature vector to update the benchmark information of the brand model.

[0054] Specifically, when constructing the baseline information for a brand model, the first step is to obtain its official name, common aliases, key attribute information (e.g., product category, core functions, release date), and semantic features extracted from historical comment data. This information is integrated to form the baseline information for the brand model, which includes name features, attribute features, and historical semantic feature vectors. Name features and attribute features provide a static, objective description of the brand model, while the historical semantic feature vectors capture the semantic patterns and trends of user discussions about the brand model over a past period. The baseline information is stored as a reference for subsequent correlation analysis.

[0055] To ensure the timeliness and accuracy of the baseline information, the system periodically performs time-decay weighting on historical semantic feature vectors. This means that earlier semantic features are assigned lower weights to reflect their potentially decreased relevance to current brand perception. The resulting decayed historical semantic feature vector better represents the cumulative semantic features of the brand and model prior to the current time point, but with the impact of outdated information appropriately mitigated.

[0056] Simultaneously, the system extracts recent comments from the continuously flowing incremental comment data. These recent comments represent the latest market feedback and user discussions regarding the brand and model. Based on the text data associated with the brand and model in the recent comments, a semantic feature vector of the recent comments is extracted. This vector reflects the semantic characteristics of the brand and model over a recent period, capturing the latest user concerns, trending terms, or product reviews.

[0057] Subsequently, the semantic offset between the semantic feature vector of recent comments and the baseline information (especially its semantic feature portion) is calculated. The semantic offset quantifies the degree of semantic difference between recent discussions and historical baselines, reflecting the magnitude of changes in brand perception or discussion focus. Simultaneously, combining this with the number of recent comments allows for a comprehensive assessment of the significance and reliability of this semantic offset. Based on the semantic offset and the number of recent comments, the update rate and magnitude of the semantic features in the baseline information are determined. For example, if the semantic offset is large and the number of recent comments is sufficient, it may be necessary to update the baseline information at a faster rate and with a larger magnitude.

[0058] Finally, based on the determined update rate and magnitude, the decayed historical semantic feature vectors are fused with the recent comment semantic feature vectors. This fusion process aims to balance historical stability with current dynamism, ensuring that the updated brand and model benchmark information retains its core identifying features while promptly reflecting the latest market changes and user perceptions.

[0059] Optionally, the steps of periodically performing time-decay weighted processing on historical semantic feature vectors to obtain decayed historical semantic feature vectors include: Obtain current market popularity information and lifecycle stage information for the brand and model. Current market popularity information includes market popularity value and market popularity fluctuation range. Based on the market popularity value and the fluctuation range of market popularity, the time decay control parameters are determined. The time decay control parameters include decay rate, recent comment time window and lower limit of historical semantic feature retention. Based on the lifecycle stage information, a time decay mechanism is selected, which includes a heat-sensitive time decay mechanism, a stable time decay mechanism, and an accelerated decay mechanism. When the fluctuation of market popularity exceeds the preset popularity fluctuation threshold, the popularity sensitive time decay mechanism is selected, and the decay rate of historical semantic feature vector is increased and the weight of historical semantic feature vector is decreased according to the time decay control parameters to obtain the popularity sensitive time decay weight. When a brand model is in a stable development stage, a steady-state time decay mechanism is selected, and the weights of the historical semantic feature vectors are set to decrease steadily according to the time decay control parameters to obtain the steady-state time decay weights. When a brand or model is in a decline phase, an accelerated decay mechanism is selected, and the decay rate of the historical semantic feature vector is increased according to the time decay control parameters to obtain the accelerated time decay weight. Based on the heat-sensitive time decay weight, the steady-state time decay weight, or the accelerated time decay weight, the historical semantic feature vector is subjected to time decay weighting to obtain the decayed historical semantic feature vector.

[0060] Specifically, when performing time-decay weighted processing on historical semantic feature vectors, it is first necessary to obtain the current market popularity information and lifecycle stage information of the brand and model. The current market popularity information can be understood as indicators such as the brand's attention and discussion volume within a specific time period, specifically including market popularity values ​​and fluctuations in market popularity, aiming to reflect the brand's activity level and trends in the market. Lifecycle stage information refers to the market stage the brand is in, such as the introduction stage, growth stage, stable development stage, or decline stage, aiming to provide a macro-level basis for selecting appropriate decay strategies.

[0061] Furthermore, based on the obtained market popularity value and the magnitude of market popularity fluctuations, time decay control parameters can be determined. These parameters include the decay rate, the recent comment time window, and the lower limit for retaining historical semantic features. The decay rate controls the speed at which the weights of historical semantic feature vectors decrease; the recent comment time window defines which comments are considered "recent," thus affecting the sensitivity of decay; and the lower limit for retaining historical semantic features ensures that even older but still valuable historical information can retain a certain weight, with the aim of finely adjusting the time decay strategy.

[0062] Based on this, an appropriate time decay mechanism is selected according to the brand model's lifecycle stage information. This application provides three main decay mechanisms: a popularity-sensitive time decay mechanism, a stable time decay mechanism, and an accelerated decay mechanism. These mechanisms aim to flexibly adjust the weight of historical semantic features according to the brand's actual situation.

[0063] Specifically, when market popularity fluctuations exceed a preset popularity fluctuation threshold, it indicates that the brand may be experiencing rapid changes or sudden events. In this case, a popularity-sensitive time decay mechanism is selected. Under this mechanism, based on time decay control parameters, the decay rate of historical semantic feature vectors is increased, while the weight of historical semantic feature vectors is decreased, thus obtaining the popularity-sensitive time decay weight. Its purpose is to quickly respond to market changes, ensuring that the semantic features of recent comments dominate, and guaranteeing that the brand's baseline information reflects the latest developments in a timely manner.

[0064] When a brand or model is in a stable development phase, its market performance is relatively stable, and a stable time decay mechanism is chosen. Under this mechanism, based on time decay control parameters, a steadily decreasing weight is set for the historical semantic feature vector, resulting in a stable time decay weight. The aim is to maintain the stability of historical information while gradually introducing new information to preserve the continuity and accuracy of the brand's semantic features.

[0065] When a brand or model is in decline, market attention gradually decreases, at which point an accelerated decay mechanism is chosen. Under this mechanism, based on time decay control parameters, the decay rate of historical semantic feature vectors is increased, resulting in accelerated time decay weights. The aim is to rapidly reduce the weights of historical semantic features that are no longer representative, preventing them from unnecessarily interfering with current brand evaluation.

[0066] Finally, based on the selected decay mechanism (i.e., heat-sensitive time decay weight, steady-state time decay weight, or accelerated time decay weight), the historical semantic feature vector is subjected to time decay weighting processing to obtain the decayed historical semantic feature vector.

[0067] Optionally, when the market popularity fluctuation exceeds a preset popularity fluctuation threshold, a popularity-sensitive time decay mechanism is selected. The steps to obtain the popularity-sensitive time decay weight include increasing the decay rate of the historical semantic feature vector and decreasing its weight based on the time decay control parameters: Update recent comments based on the recent comment time window in the time decay control parameters; The authenticity of the text content and the authenticity of the posting behavior are assessed in recent comments to obtain the results of the text content authenticity assessment and the posting behavior authenticity assessment. The credibility score of recent comments is determined by combining the interaction volume, text content authenticity assessment results, and posting behavior authenticity assessment results. Based on the credibility score, determine the credibility of the market popularity fluctuation corresponding to the fluctuation range; When the credibility of the heat fluctuation is higher than the preset credibility threshold, the decay rate of the historical semantic feature vector is increased and the weight of the historical semantic feature vector is reduced according to the decay rate and the lower limit of historical semantic feature retention in the time decay control parameters, thus obtaining the heat-sensitive time decay weight. When the credibility of the heat fluctuation is not higher than the preset credibility threshold, the increase in the decay rate of the historical semantic feature vector and the decrease in the weight of the retained historical semantic feature vector are limited according to the lower limit of historical semantic feature retention, so as to obtain the heat-sensitive time decay weight.

[0068] Specifically, the recent comment time window refers to the time range used to collect and analyze the latest comment data, such as comments from the past 24 hours, 7 days, or 30 days. Its purpose is to ensure the timeliness of the analyzed comment data and its ability to reflect the latest market dynamics of the brand and model. Text content authenticity assessment aims to identify whether the comment content itself contains false, exaggerated, or misleading information, for example, through semantic parsing and rhetorical structure analysis. Posting behavior authenticity assessment focuses on determining the authenticity and credibility of the comment's publisher and their posting behavior, such as identifying automated accounts, bulk account registration, or abnormal posting patterns. Its purpose is to comprehensively evaluate the reliability of recent comments from both content and behavioral dimensions. Interaction volume can be understood as user feedback data such as likes, shares, and comments received by the comment, reflecting the breadth and influence of the comment's dissemination. Credibility scoring is a quantitative evaluation of the overall quality of the comment, aiming to provide a basis for determining the credibility of subsequent fluctuations in popularity. The credibility of trending topics refers to the reliability of current market trends. It comprehensively considers the overall credibility of recent comments that constitute the trending topics, aiming to determine whether the current market trends truly reflect the market situation, rather than being driven by low-quality or malicious information. When market trends are considered genuine and credible, the system actively responds to this change, accelerating the decay of historical semantic features to allow brand benchmark information to adapt to new market dynamics more quickly. This indicates that when there is a credibility risk in market trends, the system will adopt a more cautious strategy to avoid excessive decay of historical semantic features, thereby preventing brand benchmark information from being misled by unreliable market signals.

[0069] Optionally, the steps for evaluating the text content authenticity of recent comments and obtaining the evaluation results include: Semantic analysis is performed on the text of recent comments to identify rhetorical structures in the recent comments, including irony, allusion, or analogy; Semantic analysis of recent commentaries involves using natural language processing (NLP) techniques to deeply analyze the vocabulary, syntax, and discourse structure of these texts to reveal their deeper meanings. This process pays particular attention to identifying rhetorical structures such as irony, allusion, and analogy. Irony typically manifests as the opposite of its intended meaning, allusion indirectly alludes to a specific object or event, and analogy clarifies a point by comparing different things. Identifying these rhetorical structures is a crucial step in understanding the true intent of a commentary, as they often carry information beyond the literal meaning of the words.

[0070] Perform intention analysis on rhetorical structures to determine whether they are used to express true opinions, to make ironies, to imply, or to use metaphors, and obtain the intention analysis results. Furthermore, after identifying the rhetorical structures, it is necessary to conduct intention analysis on these structures. Intention analysis aims to determine the purpose of the commentator's use of a particular rhetorical structure, such as expressing a true viewpoint, making an ironic statement, providing an allusion, or simply using it as a metaphor. For example, for an ironic comment that seems complimentary but is actually derogatory, intention analysis can reveal its true negative connotation. By conducting in-depth analysis of the intention of rhetorical structures, we can more accurately grasp the underlying meaning of the comment and avoid biases caused by literal interpretation.

[0071] Based on the intent analysis results and recent comments, identify whether there is false advertising, exaggeration, or malicious attack in the recent comments, and obtain the corresponding identification results; Based on this, by combining the results of intent analysis with the original text content of recent comments, it is possible to identify whether the comments contain false advertising, exaggeration, or malicious attacks. False advertising may manifest as unfounded boasting about product performance, exaggeration may involve over-emphasis on product experience descriptions, and malicious attacks may include defamatory statements against the brand or competitors. By comprehensively considering rhetorical intent and textual content, these misleading or damaging messages can be identified more effectively.

[0072] Based on the recognition results, an assessment of the authenticity of the text content is generated.

[0073] Finally, based on the above identification results, a text content authenticity assessment result is generated. This assessment result reflects the authenticity and credibility of recent comment text content, providing an important basis for subsequent credibility assessments of market trend fluctuations.

[0074] Optionally, the steps for conducting a posting behavior authenticity assessment on recent comments and obtaining the assessment results include: Verify the identity of the publishers of recent comments to identify whether they are automated accounts or accounts registered in bulk, and obtain the identity verification results; The content of recent comments is compared for similarity to identify whether there are similar comments and the content similarity comparison results are obtained. Perform pattern analysis on recent comment posting behavior to identify abnormal posting frequency, abnormal posting time, or abnormal posting source, and obtain the posting behavior pattern analysis results; Based on the identity verification results, content similarity comparison results, and posting behavior pattern analysis results, it is determined whether there are any abnormal posting behaviors in recent comments, and the judgment result is obtained; Based on the judgment results, an assessment of the authenticity of the publishing behavior is generated.

[0075] The process of verifying the identities of recent comment posters aims to determine the authenticity of the comment sources. Specifically, this involves analyzing the poster's registration information, historical behavior patterns, IP address, and other data to determine whether they are automated accounts controlled by a program, or accounts created through batch registration or collaborative operations. The goal is to eliminate interference from non-genuine or malicious users in verifying the authenticity of the comment data.

[0076] Furthermore, similarity comparisons are performed on recent comments to identify instances of plagiarism, duplication, or mass generation. Specifically, text similarity algorithms (such as cosine similarity and Jaccard similarity) can be used to compare comment texts pairwise or with known templates to identify highly similar or identical homogenized comment content. The aim is to uncover a large number of low-quality or fake comments generated through copying and pasting or templates.

[0077] Furthermore, pattern analysis of recent comment posting behavior aims to uncover abnormal posting patterns. For example, it can analyze posting frequency to determine if there is an abnormally high frequency of comments posted in a short period; analyze posting time to determine if there are abnormal posting times concentrated outside of normal working hours or specific time periods; and analyze posting source to determine if there are a large number of comments originating from the same IP address, the same device, or the same geographical location, thereby identifying abnormal posting sources. The purpose is to reveal potential unnatural posting behaviors such as comment manipulation and paid commenters.

[0078] Therefore, based on the identity verification results, content similarity comparison results, and posting behavior pattern analysis results, a comprehensive judgment can be made as to whether recent comments exhibit abnormal posting behavior. For example, if the poster is identified as an automated account, and its comment content is highly homogeneous while the posting frequency is abnormally high, then the comment can be judged to have abnormal posting behavior. Finally, based on this judgment result, a posting behavior authenticity assessment result is generated, which can be a binary judgment (abnormal behavior exists / does not exist) or a quantitative authenticity score.

[0079] This application also discloses an artificial intelligence-based multi-source brand data fusion analysis system, used to perform artificial intelligence-based multi-source brand data fusion analysis, combined with... Figure 3 As shown, the AI-based multi-source brand data fusion analysis system 1 includes: The comprehensive feature generation module 11 is used to acquire and store the benchmark information of the brand and model, identify the unofficial representative to be associated from the competitive product review data of multiple platforms, and generate comprehensive feature information based on the literal features, contextual semantic features, comment topic role features, platform source features and historical similar expression features of the unofficial representative; The candidate association generation module 12 is used to generate a candidate association set based on comprehensive feature information and benchmark information. The candidate association set includes multiple candidate associations, which are unverified associations between an unofficial representative and a brand model. The module calculates association evidence and association confidence for the candidate associations in the candidate association set. The association evidence includes literal similarity, semantic similarity, historical co-occurrence trend, cross-comment consistency, competitive role matching degree, short-term occurrence rate, and cross-platform synchronous occurrence ratio. Problem type identification module 13 is used to identify the type of association problem between unofficial representatives and brand models based on association evidence and association confidence. The association problem types include regular association with missing bridging information, risk of mis-association with competitors, or association with sudden missing bridging information. The processing result generation module 14 is used to perform corresponding semantic bridging processing, false association risk mitigation processing, or sudden reinforcement processing according to the type of associated problem, and obtain the association processing result; The relationship graph update module 15 is used to output the association results hierarchically and update the brand object relationship graph based on the association processing results and association confidence. The association results include high-confidence associations, associations to be confirmed, and independent observation nodes. The association status adjustment module 16 is used to obtain incremental comment data, review the associations and independent observation nodes to be confirmed based on the incremental comment data, obtain the review results, and adjust the association status in the brand object relationship diagram according to the review results; The brand visibility assessment module 17 is used to output the brand object relationship diagram after adjusting the association status, and to assess brand visibility based on the association status between brand models and unofficial representatives in the brand object relationship diagram.

[0080] Specifically, the comprehensive feature generation module is configured to acquire and store benchmark information of the brand and model, identify unofficial representatives to be associated from competitor review data across multiple platforms, and generate comprehensive feature information based on the literal features, contextual semantic features, comment subject role features, platform source features, and historical similar expression features of the unofficial representatives. For the specific implementation and operational details of this module, please refer to the description of acquiring and storing benchmark information of the brand and model and generating comprehensive feature information in the above method embodiments, which will not be repeated here.

[0081] The candidate association generation module is configured to generate a set of candidate associations based on comprehensive feature information and benchmark information, and to calculate association evidence and association confidence for the candidate associations in the set. This module may include an association matching engine and a confidence calculation engine. The association matching engine is responsible for generating candidate associations based on various algorithms (e.g., rule-based, machine learning, or deep learning), while the confidence calculation engine comprehensively calculates the association confidence based on association evidence such as literal proximity, semantic proximity, historical co-occurrence trends, cross-comment consistency, matching degree of competing roles, short-term occurrence rate, and cross-platform synchronous occurrence ratio. For the specific implementation and operational details of this module, please refer to the description of generating the candidate association set and calculating association evidence and association confidence in the above method embodiments; further details will not be repeated here.

[0082] The issue type identification module is configured to identify the type of association issue between unofficial representatives and brand models based on association evidence and association confidence levels. This module can employ classification models, such as support vector machines or neural networks, to analyze the association evidence and confidence levels, thereby classifying association issues into categories such as regular associations lacking bridging information, competitor mis-association risks, or sudden associations lacking bridging information. For specific implementation details and operational nuances of this module, please refer to the description of association issue type identification in the above method embodiments; further details will not be repeated here.

[0083] The processing result generation module is configured to perform corresponding semantic bridging processing, erroneous association risk mitigation processing, or sudden reinforcement processing based on the type of association issue, thereby obtaining the association processing result. This module can integrate multiple processing algorithms. For example, for associations with missing bridging information, a knowledge graph inference unit can be invoked; for competitor erroneous association risks, a competitor semantic analysis unit can be invoked; and for sudden associations with missing bridging information, a real-time information reinforcement unit can be invoked. For the specific implementation and operational details of this module, please refer to the descriptions of performing corresponding semantic bridging processing, erroneous association risk mitigation processing, or sudden reinforcement processing in the above method embodiments, which will not be repeated here.

[0084] The relationship graph update module is configured to output association results hierarchically and update the brand object relationship graph based on the association processing results and association confidence levels. This module may include a graph database management system and a visualization interface for storing and managing the brand object relationship graph, and dynamically updating the graph structure based on different levels of association results, such as high-confidence associations, pending confirmation associations, and independent observation nodes. For the specific implementation and operational details of this module, please refer to the description of hierarchical output of association results and updating of the brand object relationship graph in the above method embodiment; it will not be repeated here.

[0085] The association status adjustment module is configured to acquire incremental comment data, review the associations to be confirmed and independent observation nodes based on the incremental comment data, obtain the review results, and adjust the association status in the brand object relationship diagram according to the review results. This module may include an incremental data processing unit and a review logic unit, used to continuously monitor new data and verify and adjust existing associations according to a preset review strategy. For the specific implementation and operational details of this module, please refer to the description of acquiring incremental comment data and reviewing the associations to be confirmed and independent observation nodes in the above method embodiment, which will not be repeated here.

[0086] The brand visibility assessment module is configured to output a brand object relationship graph with adjusted association states, and to assess brand visibility based on the association states between brand models and unofficial representatives in the graph. This module may include a report generator and an assessment metric calculator to generate a brand visibility assessment report based on metrics such as the number, strength, and activity of associations in the brand object relationship graph. For details on the specific implementation and operation of this module, please refer to the description of outputting the adjusted brand object relationship graph and assessing brand visibility in the above method embodiment; it will not be repeated here.

[0087] The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for multi-source brand data fusion and analysis based on artificial intelligence, characterized in that, include: Acquire and store the baseline information of the brand and model, identify the unofficial representative to be associated from the competitive product review data of multiple platforms, and generate comprehensive feature information based on the literal features, contextual semantic features, comment topic role features, platform source features and historical similar expression features of the unofficial representative; A candidate association set is generated based on the comprehensive feature information and the benchmark information; wherein, the candidate association set includes multiple candidate associations, and the candidate association is an unverified association relationship between an unofficial representative and a brand model; association evidence and association confidence are calculated for the candidate associations in the candidate association set, and the association evidence includes literal similarity, semantic similarity, historical co-occurrence trend, cross-comment consistency, competitive role matching degree, short-term occurrence rate, and cross-platform synchronous occurrence ratio; Based on the associated evidence and the associated confidence level, identify the types of association issues between the unofficial representative and the brand model. The types of association issues include regular association with missing bridging information, risk of misassociation with competitors, or association with sudden missing bridging information. Based on the type of associated problem, perform corresponding semantic bridging processing, false association risk mitigation processing, or sudden reinforcement processing to obtain the association processing result; Based on the association processing results and the association confidence level, the association results are output hierarchically and the brand object relationship graph is updated. The association results include high-confidence associations, associations to be confirmed, and independent observation nodes. Acquire incremental comment data, review the unconfirmed associations and the independent observation nodes based on the incremental comment data, obtain the review results, and adjust the association status in the brand object relationship diagram according to the review results; Output the brand object relationship graph after adjusting the association status, and evaluate the brand visibility based on the association status between brand models and unofficial representatives in the brand object relationship graph.

2. The method for multi-source brand data fusion and analysis based on artificial intelligence according to claim 1, characterized in that, When the association problem type is a sudden missing bridging information association, sudden reinforcement processing is performed according to the association problem type to obtain the association processing result, including: Filter core comments from the current comment stream and identify these core comments as sources of supplementary information. Obtain the content quality assessment results of the supplementary information source; Based on the candidate association set, determine the brand model that has a candidate association relationship with the unofficial representative, and use the benchmark information corresponding to the brand model as known entity features; A temporary association network is constructed, which includes unofficial indicator delivery nodes, known entity nodes, and supplementary information source nodes, and connection edges are established based on the co-occurrence relationship or semantic pointing relationship between the unofficial indicator delivery, the known entity features, and the supplementary information source. The node weight of the supplementary information source node is determined based on the content quality assessment result of the supplementary information source, and the edge weight is determined based on the co-occurrence strength or semantic pointing strength corresponding to the connection edge. According to the node weights and edge weights, the temporary association network is subjected to graph structure feature aggregation. The features of the known entity nodes are aggregated to the unofficial index delivery node via the reinforcement information source node to obtain the reinforced unofficial index delivery features. The authority of the supplementary information sources from different platforms and the credibility of the comment publishers were evaluated to obtain the platform credibility evaluation results; Based on the platform credibility assessment results, the cross-validation weights of the information sources from different platforms are adjusted. Based on the cross-validation weights, consistency verification is performed on the evidence of the association between the unofficial representative and the brand model in the supplementary information sources from different platforms to obtain cross-platform consistency verification results. Based on the enhanced unofficial indicator representation characteristics and the cross-platform consistency verification results, the association confidence between the unofficial indicator representation and the brand model is updated to obtain the corresponding association confidence. Based on the association confidence level, the corresponding association processing result is generated.

3. The method for multi-source brand data fusion analysis based on artificial intelligence according to claim 2, characterized in that, The step of filtering core comments from the current comment stream and identifying the core comments as supplementary information sources includes: The content quality of comments in the current comment stream is evaluated to obtain the content quality evaluation results. The content quality evaluation includes identifying the emotional tendency, misleading expressions, and information density in the comments. Get the interaction count of comments in the current comment stream; The comments are given a comprehensive score based on the interaction volume and the content quality assessment results. Based on the comprehensive score, select core comments from the current comment stream; The core comments were used as supplementary information sources.

4. The method for multi-source brand data fusion analysis based on artificial intelligence according to claim 1, characterized in that, When the association issue type is the risk of mis-association with competitors, the mis-association risk resolution process is performed according to the association issue type to obtain the association processing result, including: Identify the semantic relationships between the unofficial statements and competing entities that co-occur in the comments; Determine whether the semantic relationship exists as a contrast, irony, or antithesis. Based on the co-occurrence frequency, text distance, and semantic pointing strength of the unofficial reference expression and the competing entity in the comments, the direct referential association weight between the unofficial reference expression and the competing entity is determined; When the aforementioned contrast relationship, the aforementioned ironic relationship, or the aforementioned reverse contrast relationship exists, the direct referential association weight is reduced according to the semantic relationship to obtain the association weight after competitor elimination. The analysis examines the potential intended meaning of the unofficial references in the comments regarding the aforementioned brand and model. Based on the potential intended meaning, determine the brand-related association weight between the unofficial representation and the brand model; Based on the association weights after competitor elimination and the brand-pointing association weights, the association processing result for the competitor mis-association risk is generated; When the aforementioned comparison relationship, ironic relationship, and reverse contrast relationship do not exist, the direct referential association weight is retained, and the corresponding candidate association is marked as a competitor association to be confirmed based on the direct referential association weight. Based on the competitor associations to be confirmed, generate the association processing result for the risk of misassociation of the competitors.

5. The method for multi-source brand data fusion analysis based on artificial intelligence according to claim 1, characterized in that, The steps for acquiring and storing the brand and model reference information include: Obtain the official name, aliases, attribute information, and historical discussion semantic features of the brand model, and construct the baseline information of the brand model, which includes name features, attribute features, and historical semantic feature vectors; Store the reference information; The historical semantic feature vector is periodically subjected to time decay weighting processing to obtain the decayed historical semantic feature vector; Extract recent comments from the incremental comment data; Extract semantic feature vectors of recent comments based on the text data associated with the brand and model in the recent comments; Calculate the semantic offset between the recent comment semantic feature vector and the baseline information; Based on the semantic offset and the number of recent comments, determine the update rate and update magnitude of the semantic features in the benchmark information; Based on the update rate and the update magnitude, the historical semantic feature vector after decay processing is fused with the recent comment semantic feature vector to update the baseline information of the brand model.

6. The method for multi-source brand data fusion analysis based on artificial intelligence according to claim 5, characterized in that, The step of periodically performing time-decay weighted processing on the historical semantic feature vector to obtain the decayed historical semantic feature vector includes: Obtain the current market popularity information and life cycle stage information of the brand model, wherein the current market popularity information includes the market popularity value and the market popularity fluctuation range; Based on the market popularity value and the market popularity fluctuation range, time decay control parameters are determined, including decay rate, recent comment time window, and lower limit for retaining historical semantic features; Based on the lifecycle stage information, a time decay mechanism is selected, including a heat-sensitive time decay mechanism, a stable time decay mechanism, and an accelerated decay mechanism. When the fluctuation of market popularity exceeds the preset popularity fluctuation threshold, a popularity-sensitive time decay mechanism is selected, and the decay rate of the historical semantic feature vector is increased and the weight of the historical semantic feature vector is decreased according to the time decay control parameters to obtain the popularity-sensitive time decay weight. When the brand model is in a stable development stage, a steady-state time decay mechanism is selected, and a steadily decreasing weight is set on the historical semantic feature vector according to the time decay control parameters to obtain the steady-state time decay weight. When the brand model is in the decline stage, an accelerated decay mechanism is selected, and the decay rate of the historical semantic feature vector is increased according to the time decay control parameter to obtain the accelerated time decay weight. Based on the heat-sensitive time decay weight, the stable time decay weight, or the accelerated time decay weight, the historical semantic feature vector is subjected to time decay weighting processing to obtain the decayed historical semantic feature vector.

7. The method for multi-source brand data fusion analysis based on artificial intelligence according to claim 6, characterized in that, The step of selecting a heat-sensitive time decay mechanism when the market heat fluctuation exceeds a preset heat fluctuation threshold, and increasing the decay rate of the historical semantic feature vector and decreasing the weight of the historical semantic feature vector according to the time decay control parameters to obtain the heat-sensitive time decay weight includes: The recent comments are updated according to the recent comment time window in the time decay control parameters; The recent comments are evaluated for both text content authenticity and posting behavior authenticity, resulting in text content authenticity evaluation results and posting behavior authenticity evaluation results. The credibility score of the recent comments is determined by combining the interaction volume of the recent comments, the authenticity assessment results of the text content, and the authenticity assessment results of the posting behavior. Based on the credibility score, the credibility of the market popularity fluctuation corresponding to the market popularity fluctuation range is determined; When the credibility of the heat fluctuation is higher than the preset credibility threshold, the decay rate of the historical semantic feature vector is increased and the weight of the historical semantic feature vector is decreased according to the decay rate and the lower limit of historical semantic feature retention in the time decay control parameters, so as to obtain the heat-sensitive time decay weight. When the credibility of the heat fluctuation is not higher than the preset credibility threshold, the increase in the decay rate of the historical semantic feature vector and the decrease in the weight of the retained historical semantic feature vector are limited according to the lower limit of the historical semantic feature retention, so as to obtain the heat-sensitive time decay weight.

8. The method for multi-source brand data fusion analysis based on artificial intelligence according to claim 7, characterized in that, The steps for evaluating the text content authenticity of the recent comments and obtaining the text content authenticity evaluation results include: Semantic analysis is performed on the text of the recent comments to identify the rhetorical structures in the recent comments, including irony, allusion, or analogy; The intention analysis is performed on the rhetorical structure to determine whether the rhetorical structure is used to express a true viewpoint, to make irony, to imply, or to metaphor, and the intention analysis results are obtained. Based on the intent analysis results and the recent comments, identify whether there is false advertising, exaggeration, or malicious attack in the recent comments, and obtain the corresponding identification results; Based on the recognition results, a text content authenticity assessment result is generated.

9. The multi-source brand data fusion analysis method based on artificial intelligence according to claim 7, characterized in that, The steps for evaluating the authenticity of recent comments and obtaining the evaluation results include: The publishers of the recent comments are identified to determine whether they are automated accounts or bulk-registered accounts, and the identification results are obtained. The content of the recent comments is compared for similarity to identify whether there are any homogeneous comments, and the content similarity comparison results are obtained. The recent comment posting behavior is analyzed to identify whether there are abnormal posting frequencies, abnormal posting times, or abnormal posting sources, and the posting behavior pattern analysis results are obtained. Based on the identity verification result, the content similarity comparison result, and the posting behavior pattern analysis result, it is determined whether the recent comments exhibit abnormal posting behavior, and a judgment result is obtained; Based on the judgment results, an evaluation result on the authenticity of the publishing behavior is generated.

10. A multi-source brand data fusion analysis system based on artificial intelligence, used to perform multi-source brand data fusion analysis based on artificial intelligence, characterized in that, include: The comprehensive feature generation module is used to acquire and store the baseline information of the brand and model, identify the unofficial representative to be associated from the competitive product review data of multiple platforms, and generate comprehensive feature information based on the literal features, contextual semantic features, comment topic role features, platform source features and historical similar expression features of the unofficial representative; The candidate association generation module is used to generate a candidate association set based on the comprehensive feature information and the benchmark information; wherein, the candidate association set includes multiple candidate associations, and the candidate association is an unverified association relationship between an unofficial representative and a brand model; for the candidate associations in the candidate association set, association evidence and association confidence are calculated, and the association evidence includes literal similarity, semantic similarity, historical co-occurrence trend, cross-comment consistency, competitive role matching degree, short-term occurrence rate, and cross-platform synchronous occurrence ratio; The problem type identification module is used to identify the type of association problem between the unofficial representative and the brand model based on the association evidence and the association confidence level. The association problem types include regular bridging information missing association, competitor mis-association risk, or sudden bridging information missing association. The processing result generation module is used to perform corresponding semantic bridging processing, false association risk mitigation processing, or sudden reinforcement processing according to the type of the associated problem, and obtain the association processing result; The relationship graph update module is used to output the association results hierarchically and update the brand object relationship graph based on the association processing results and the association confidence. The association results include high-confidence associations, associations to be confirmed, and independent observation nodes. The association status adjustment module is used to acquire incremental comment data, review the association to be confirmed and the independent observation node based on the incremental comment data, obtain the review result, and adjust the association status in the brand object relationship diagram according to the review result; The brand visibility assessment module is used to output a brand object relationship diagram after adjusting the association status, and to assess brand visibility based on the association status between brand models and unofficial representatives in the brand object relationship diagram.

Images