Automatic consumption label analysis system and method based on multi-agent cooperation

Abstract

The invention provides an automatic consumption tag analysis system and method based on multi-agent collaboration, and the system comprises a data processing agent which is used for collecting data from a social media platform and carrying out the data preprocessing; the label identification intelligent agent is used for extracting consumption labels of different dimensions from the preprocessed data; the sentiment analysis agent is used for carrying out context modeling and sentiment tendency recognition and binding the recognized sentiment tendency to the corresponding consumption label; the label normalization agent is used for performing clustering and normalization processing on all consumption labels bound with emotional tendencies to generate a structured multi-layer label atlas; the central scheduling agent is used for scheduling other agents, generating a label analysis result by using the multi-layer label atlas and sending the label analysis result to the user; according to the invention, based on a multi-agent architecture, structured analysis is carried out on user tags, behavior attributes and consumption intentions in social media contents, so that high-precision and high-efficiency intelligent consumption insight is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an automated consumer tag analysis system and method based on multi-agent collaboration. Background Technology

[0002] With the rise of social media platforms, user-generated content (UGC) has become a crucial source of information for brands to obtain market feedback and understand consumer behavior. For example, in sectors such as FMCG, beauty, footwear, apparel, 3C electronics, and beverages, the text, images, and short videos posted by consumers on platforms like Weibo, Xiaohongshu, Douyin, and Zhihu contain a wealth of commentary information regarding product usage experiences, emotional preferences, functional needs, price sensitivity, and purchase motivations. This unstructured text data forms a key data foundation for brands to understand consumer trends, optimize product strategies, and improve user operation efficiency.

[0003] However, most social media data analysis methods still face the following core technical bottlenecks and shortcomings: 1) Reliance on manual rules and templates, resulting in low levels of intelligence: Traditional information extraction solutions are mostly based on keyword dictionaries, regular expressions, and rule templates, which cannot adapt to the rapidly evolving online language environment, such as popular slang, pinyin abbreviations, and cross-contextual expressions that are extremely common in social media. Especially when facing the diverse expression habits of users in categories such as beauty, clothing, and digital products, rule-based methods are inadequate, and it is difficult to guarantee both accuracy and recall.

[0004] 2) Coarse analysis granularity, unable to accurately identify multi-level tag systems: Currently, many data platforms' tag extraction systems can only identify first-level dimensions (such as price, appearance, brand), making it difficult to further explore deeper user needs or contextual emotions. For example, sentences like "high cost performance but the color makes it look dark" contain multiple evaluation dimensions, and cannot be broken down into multiple independent and non-conflicting analysis dimensions in terms of tag structure.

[0005] 3) Inability to achieve contextual semantic understanding and sentiment analysis: Social media corpora contain a large number of linguistic phenomena such as "parallel subject-verb agreement," "omission of subject," and "double negation." Taking "Xiaohongshu Notes" as an example, users may evaluate multiple brands and products in a single text. Without contextual modeling and core viewpoint identification capabilities, the system is prone to confusing the evaluated objects, leading to distorted conclusions.

[0006] 4) The system architecture lacks flexibility and is difficult to adapt to the analysis needs of different brands: Most solutions on the market are centralized monolithic architectures, which cannot perform functional decomposition, asynchronous task processing, resource optimization, etc. for analysis tasks. Especially when user needs change frequently or multiple brands are processed in parallel, the system stability and scalability are challenged.

[0007] 5) The task process is highly dependent on manual configuration and has a low level of automation: Most analysis links (such as corpus classification, tag extraction, sentiment judgment, normalization mapping, etc.) still require data analysts to manually define the process and filter the results, resulting in long time consumption and many errors, which greatly limits the universality and efficiency of the system on large-scale, multi-category, and multi-corpus platforms.

[0008] In summary, traditional social media consumer tagging analysis systems have significant bottlenecks in terms of intelligent architecture, semantic modeling capabilities, task scheduling efficiency, and cross-platform scalability, making it difficult to meet the actual business needs of brand clients for "multi-dimensional tags," "precise audience profiling," "multi-modal data fusion," "intelligent semantic parsing," and "real-time insight feedback."

[0009] Therefore, there is an urgent need to build an automated system with modular capabilities, intelligent scheduling mechanisms, tag semantic recognition, and consumer behavior modeling capabilities. Summary of the Invention

[0010] To overcome the technical problems of existing technologies in terms of intelligent architecture, semantic modeling capabilities, task scheduling efficiency, and cross-platform scalability, this invention provides an automated consumer tag analysis system and method based on multi-agent collaboration. Based on a multi-agent architecture, this invention performs structured analysis of user tags, behavioral attributes, and consumption intentions in social media content, thereby achieving high-precision and high-efficiency intelligent consumer insights.

[0011] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: An automated consumer tagging analysis system based on multi-agent collaboration includes: A data processing agent for collecting and preprocessing data from at least one social media platform; The tag recognition agent is equipped with a pre-trained natural language model and tag libraries from different domains, which are used to extract consumer tags of different dimensions from pre-processed data. A sentiment analysis agent is used to perform contextual modeling and sentiment identification on preprocessed data, and to bind the identified sentiment to the corresponding consumer tags. The tag normalization agent is used to cluster and normalize all consumer tags bound with emotional tendencies, generating a structured multi-layer tag map. The central scheduling agent is connected to the data processing agent, the tag recognition agent, the sentiment analysis agent, and the tag normalization agent, respectively, and is used to coordinate the scheduling of other agents, as well as to generate tag analysis results using the multi-layer tag graph and send them to the user.

[0012] Preferably, the data processing agent accesses at least one social media platform via an API interface or web crawler to collect content posted by users; The preprocessing includes: data cleaning and structural transformation of the collected data, followed by word segmentation, error correction, and sentence segmentation.

[0013] Preferably, in the tag recognition intelligent agent, the dimensions of the consumer tag include any one or more of the following: entity tag, attribute tag, behavior tag, and evaluation phrase.

[0014] Preferably, in the sentiment analysis agent, context modeling includes: for preprocessed data, performing subject-object analysis and referential resolution in conjunction with the context to identify the evaluation subject and label attribution; and extracting viewpoints from complex texts with compound semantics, irony, or / and interrogative expressions in conjunction with the context. The sentiment tendency identification includes classifying the sentiment tendency expressed by the data after context modeling; the sentiment tendency includes at least any one of positive, negative and neutral.

[0015] Preferably, in the tag normalization agent, synonymous and / or near-synonymous consumer tags are identified and clustered to form multiple tag clusters; all consumer tags within the same tag cluster are normalized and mapped, and the multi-layer tag map is further constructed.

[0016] Preferably, when coordinating the scheduling of other intelligent agents, the central scheduling agent monitors the status of the task queue and other intelligent agents in real time, specifies the scheduling strategy according to the preset task priority and sends it to other intelligent agents; at the same time, it receives the execution feedback results of the scheduling strategy in real time, performs abnormal recovery according to preset rules and sends alarm information to the user when abnormal situations occur.

[0017] Preferably, in the central scheduling agent, the tag analysis results include at least one or more of the following: BI analysis results, population profile modeling results, consumer product feedback results, and consumer demand insight reports.

[0018] Preferably, the central scheduling agent is further configured to: convert the multi-layer label map into a standard JSON or Excel interface format and output it to the user in a visual manner; The visualization methods include at least one or more of the following: visualized label cloud, label matrix, and sentiment heatmap.

[0019] Preferably, the central scheduling agent, data processing agent, tag recognition agent, sentiment analysis agent, and tag normalization agent collaborate on tasks based on an event-driven model.

[0020] This invention also provides an automated consumer tag analysis method based on multi-agent collaboration, which, based on the above-mentioned system, includes the following steps: S1: The central scheduling agent schedules the data processing agent to collect data from at least one social media platform and perform data preprocessing; S2: The central scheduling agent schedules the tag recognition agent to extract consumer tags of different dimensions from the preprocessed data; S3: The central scheduling agent schedules the sentiment analysis agent to perform context modeling and sentiment tendency identification on the preprocessed data, and binds the identified sentiment tendency to the corresponding consumption tag. S4: The central scheduling agent schedules the tag normalization agent to cluster and normalize all consumer tags bound with sentiment tendencies, generating a structured multi-layer tag map. S5: The central scheduling agent uses the multi-layered tag map to generate tag analysis results and sends them to the user.

[0021] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention provides an automated consumer tag analysis system and method based on multi-agent collaboration. First, a central scheduling agent dispatches a data processing agent to collect data from at least one social media platform and perform data preprocessing. Second, the central scheduling agent dispatches a tag identification agent to extract consumer tags of different dimensions from the preprocessed data. Next, the central scheduling agent dispatches a sentiment analysis agent to perform contextual modeling and sentiment tendency identification on the preprocessed data, and binds the identified sentiment tendencies to the corresponding consumer tags. Then, the central scheduling agent dispatches a tag normalization agent to cluster and normalize all consumer tags bound to sentiment tendencies, generating a structured multi-layer tag graph. Finally, the central scheduling agent uses the multi-layer tag graph to generate tag analysis results and sends them to the user. This invention utilizes multi-agent collaborative modeling and intelligent scheduling technology to streamline the entire process from social media data access, text analysis, tag recognition to result output, constructing an automated tag analysis system for consumer insights. This system boasts technical advantages such as high accuracy, high scalability, and strong industry adaptability, and can be widely used in scenarios such as brand data analysis, user research, and product co-creation. Attached Figure Description

[0022] Figure 1 This is a structural diagram of an automated consumer tag analysis system based on multi-agent collaboration provided in Example 1.

[0023] Figure 2This is a flowchart of an automated consumer tag analysis method based on multi-agent collaboration, as provided in Example 3. Detailed Implementation

[0024] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this application. To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions; It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0025] The core technologies used in this invention include Natural Language Processing (NLP) and Multi-Agent System (MAS), primarily applied to the automated processing of large-scale unstructured user content data (such as product reviews, user testimonials, evaluation notes, and word-of-mouth comments) generated on social media platforms (such as Weibo and Xiaohongshu) in the consumer industry (e.g., FMCG, cosmetics, footwear, beverages, 3C digital products). By introducing language models (such as BERT and GPT), a tag prompt generation mechanism, and a multi-agent collaborative mechanism, the following tasks can be completed accurately and efficiently: 1) extraction and preprocessing of consumer text; 2) automatic generation and normalization of user preference and demand tags; 3) intelligent iteration and optimization of the tag system; 4) extraction and insight into the behavioral characteristics of user groups; and 5) automatic identification of context-related sentiment and semantic relationships.

[0026] This invention is widely applicable to consumer brands, market research companies, social media data service platforms, AI solution providers and other related fields. It is one of the core technology solutions with high adaptability, high scalability and high intelligence in future intelligent business data analysis.

[0027] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] Example 1 like Figure 1 As shown, this embodiment provides an automated consumer tagging analysis system based on multi-agent collaboration, including: A data processing agent 101 is used to collect data from at least one social media platform and perform data preprocessing. The tag recognition agent 102 is equipped with a pre-trained natural language model and tag libraries from different domains, which are used to extract consumer tags of different dimensions from pre-processed data. Sentiment analysis agent 103 is used to perform context modeling and sentiment tendency identification on preprocessed data, and bind the identified sentiment tendency to the corresponding consumption tag. The tag normalization agent 104 is used to cluster and normalize all consumer tags bound with emotional tendencies to generate a structured multi-layer tag map. The central scheduling agent 105 is connected to the data processing agent 101, the tag recognition agent 102, the sentiment analysis agent 103, and the tag normalization agent 104, respectively, and is used to coordinate the scheduling of other agents, and to generate tag analysis results using the multi-layer tag map and send them to the user.

[0029] In its implementation, this system aims to address the following key technical issues in traditional consumer tag extraction and user demand analysis solutions within social media corpus-driven consumer insight scenarios: 1) Coarse-grained and simple-structured tag recognition: Most current social media-based tag extraction methods rely on keyword matching or traditional rule engines, making it difficult to extract multi-dimensional and multi-level tags. For example, in the FMCG or beauty industry, users often express multiple dimensions of evaluation information at the same time, such as "the packaging looks good but it is easy to spill" or "the makeup looks natural and is suitable for yellow skin". These texts contain multiple dimensions of tags such as emotional tendency, skin tone matching, and functional needs. Existing methods cannot effectively perform structured modeling and tag decomposition, resulting in distorted tag dimensions and coarse-grained analysis.

[0030] 2) Inability to effectively understand semantic relationships and referential structures in social media contexts: Social media texts often exhibit strong semantic jumps, irregular grammar, and frequently contain issues such as omissions, referential errors, and the mixing of slang and colloquialisms; for example, "This is much more natural than that, suitable for office workers."; if the system cannot combine context to restore the subject and determine the tag attribution, it is prone to mismatches, leading to confusion between the evaluation object and the tag attribution, rendering the analysis results unusable; existing systems lack the ability to perform deep semantic modeling for social media contexts and cannot correctly identify the semantic relationships between the real subject, opinion attribution, and evaluation dimensions involved in the statement.

[0031] 3) Lack of automation and task module decoupling in the processing: In current practice, consumer tag analysis systems mostly adopt a centralized pipeline processing architecture, lacking independence and flexibility between task modules. This leads to frequent adjustments to the system structure and reconfiguration of model processes when handling different platforms (such as Xiaohongshu and Weibo), different industries (such as beauty and 3C products), or different tasks (such as tag extraction and sentiment analysis), resulting in redundant development and maintenance burdens. In addition, this centralized architecture also has significant limitations on concurrent processing capabilities, fault tolerance, and model update capabilities, making it difficult to support high concurrency, high customization, and real-time business needs.

[0032] 4) The tagging system lacks scalability and struggles to adapt to industry and corpus evolution: Currently, brands have different understandings of consumer tags across industries. For example, in the beauty industry, tags such as "makeup staying power," "yellow skin friendly," and "oxidation and dullness" are unique; while in the 3C digital industry, tags such as "screen color," "battery anxiety," and "signal interference" appear. The existing tagging system is difficult to flexibly expand according to business changes, cannot support the custom maintenance of the "tag ontology library," and cannot form a unified tag mapping relationship (such as classifying "high cost performance" and "worth buying" as the same tag dimension). The lack of such tag generalization ability will lead to fragmented analysis results and prevent the formation of a storable and learnable tag knowledge system.

[0033] 5) Lack of unified scheduling and flexible management mechanism, resulting in low system intelligence: The current consumer insight process relies heavily on human analysts for task orchestration, model selection, and result verification, making it difficult for the system to achieve a truly fully automated processing chain. Faced with the data processing needs of multiple tasks, multiple models, and multiple platforms, the lack of a unified scheduling mechanism and intelligent routing control reduces system processing efficiency and increases the risk of model conflicts and result redundancy. Especially in large-scale analysis tasks, it is difficult for the system to dynamically allocate resources according to task priority, model load, or data type, resulting in low overall operating efficiency and intelligence level.

[0034] This system is designed for consumer demand mining scenarios driven by social media data. It adopts a modular, intelligent scheduling, and multi-task parallel processing design concept. Multiple functional intelligent agents work together to complete core tasks such as data parsing, tag recognition, sentiment understanding, tag normalization, and semantic structuring.

[0035] Specifically, this system consists of the following core components: Data processing agent 101 is used to collect data from at least one social media platform and perform data preprocessing; it is responsible for accessing, cleaning, and structural transformation of social media data, is compatible with the content structure of multiple platforms (such as Xiaohongshu, Weibo, etc.), and performs basic NLP tasks such as word segmentation, error correction, and sentence segmentation. The label recognition agent 102 is equipped with a pre-trained natural language model and label libraries from different domains to extract consumer labels from pre-processed data in different dimensions. By combining models such as BERT, it can identify entity labels (such as product, brand), attribute labels (such as color, function), behavioral labels (such as trial, recommendation), evaluation phrases, etc. It can call label libraries and model parameters from different domains (such as recognition of differentiated expressions in beauty, 3C, etc.). Sentiment analysis agent 103 is used to perform context modeling and sentiment identification on preprocessed data, and bind the identified sentiment to the corresponding consumption tag; this agent is responsible for identifying the sentiment expressed in user text (positive / negative / neutral); combining context to perform subject-object analysis and referential resolution, attribution tags and evaluation subjects; and handling complex expressions such as compound semantics, irony, and questions; The tag normalization agent 104 is used to cluster and normalize all consumer tags bound with emotional tendencies, generating a structured multi-layer tag map (e.g., product dimension - user experience - refreshing feeling); this agent is responsible for normalizing and mapping synonym / near-synonymous tags, establishing the tag ontology dimension structure, and supporting brand-customized tag dimensions and tag tree updates. The central scheduling agent 105 is connected to the data processing agent 101, the tag recognition agent 102, the sentiment analysis agent 103, and the tag normalization agent 104, respectively. It is used to coordinate the scheduling of other agents and generate tag analysis results using the multi-layer tag graph and send them to the user. This agent is responsible for system resource coordination, task distribution, calling strategy formulation, managing task execution priority, model version scheduling, and anomaly recovery. At the same time, it is also responsible for outputting structured tags in a standard JSON / Excel interface format, supporting API or database integration for BI systems, crowd modeling, strategy configuration, and can output visual tag cloud maps, tag matrices, sentiment heat maps, and other functions. The entire system's tag analysis and processing flow is as follows: (1) Social media data access: Data processing intelligent agent 101 accesses user-posted content on platforms such as Weibo and Xiaohongshu through API or crawler; supports synchronous collection of text and image content, and performs URL cleaning, HTML parsing and text extraction. (2) Text preprocessing: Data processing agent 101 performs preprocessing operations such as sentence segmentation, noise reduction, keyword highlighting, and word segmentation error correction; it retains information such as emoticons and spoken expressions for subsequent emotion recognition auxiliary modeling. (3) Tag recognition: The tag recognition agent 102 calls the multi-task learning model to extract multiple tags for each text; the tag categories include: object tags (brand / product), dimension tags (user experience, appearance, etc.), emotion / motivation tags (expectation, aversion, liking, etc.); differentiated models can be loaded according to industry fields (such as beauty keyword recognition focuses more on skin feel dimension than 3C). (4) Semantic analysis and opinion attribution: The sentiment analysis agent 103 performs context modeling to identify the subject and object of the evaluation (e.g., "This is suitable for oily skin" is modeled as "the subject is the product and the user is an oily skin user"); then it performs sentiment tendency judgment and tone recognition (e.g., "not bad", "just so-so"); at the same time, it performs linkage judgment on pronouns in the sentence and the relationship between the preceding and following sentences to prevent label mismatch. (5) Tag standardization: The tag normalization agent 104 normalizes synonyms, colloquial expressions, etc.; for example, “worth the ticket price”, “not expensive”, and “cheap and plentiful” are uniformly mapped to the tag dimension of “high cost performance”; and a multi-layer tag map of different industries is constructed for archiving and tag statistics. (6) Structured output and service provision: The central scheduling agent 105 collects the results of other agents and integrates them into structured data; the tagged results can be used for BI analysis, user profiling, product feedback, insight report generation, etc.; the output supports API integration with brand CRM and public opinion systems. This system constructs a multi-agent architecture, using a central scheduling agent to uniformly manage multiple task module agents, achieving a high degree of decoupling and automation of the tag analysis process. Secondly, the system utilizes pre-trained language models (such as BERT) + Large Language Model (LLM) + domain tag ontology to improve the recognition accuracy, dimensional richness, and semantic expressive power of tags in social media text. Simultaneously, the system introduces tag normalization, opinion reference analysis, and sentiment discrimination mechanisms to solve semantic problems such as mixed multi-tag expressions and subject-object confusion. It supports multimodal unified processing and structured output of social media content from multiple platforms (such as Xiaohongshu and Weibo). Furthermore, the system constructs a sustainably expandable tag ontology management system, supporting personalized tag customization for brands and continuous model learning, thereby enabling brands to directly access tag insights and providing high-quality data support for product development, marketing strategies, audience profiling, and user segmentation.

[0036] Example 2 This embodiment provides an automated consumer tagging analysis system based on multi-agent collaboration, including: A data processing agent for collecting and preprocessing data from at least one social media platform; The tag recognition agent is equipped with a pre-trained natural language model and tag libraries from different domains, which are used to extract consumer tags of different dimensions from pre-processed data. A sentiment analysis agent is used to perform contextual modeling and sentiment identification on preprocessed data, and to bind the identified sentiment to the corresponding consumer tags. The tag normalization agent is used to cluster and normalize all consumer tags bound with emotional tendencies, generating a structured multi-layer tag map. The central scheduling agent is connected to the data processing agent, the tag recognition agent, the sentiment analysis agent, and the tag normalization agent, respectively, and is used to coordinate the scheduling of other agents, as well as to generate tag analysis results using the multi-layer tag graph and send them to the user; The data processing intelligent agent accesses at least one social media platform via API interface or web crawler and collects content posted by users. The preprocessing includes: data cleaning and structural transformation of the collected data, followed by word segmentation, error correction, and sentence segmentation; In the aforementioned tag recognition intelligent agent, the dimensions of consumer tags include any one or more of the following: entity tags, attribute tags, behavior tags, and evaluation phrases; In the sentiment analysis agent, context modeling includes: for preprocessed data, combining context to perform subject-object analysis and referential resolution, identifying the evaluator and label attribution; and combining context to extract viewpoints from complex texts with compound semantics, irony, or / and interrogative expressions. The sentiment tendency identification includes: classifying the sentiment tendency expressed by the data after context modeling; the sentiment tendency includes at least any one of positive, negative and neutral. In the tag normalization agent, synonymous and / or near-synonymous consumer tags are identified and clustered to form multiple tag clusters; all consumer tags within the same tag cluster are normalized and mapped, and the multi-layer tag map is further constructed. When coordinating the scheduling of other intelligent agents, the central scheduling agent monitors the status of the task queue and other intelligent agents in real time, specifies the scheduling strategy according to the preset task priority and sends it to other intelligent agents; at the same time, it receives the execution feedback results of the scheduling strategy in real time, performs abnormal recovery according to preset rules and sends alarm information to the user when abnormal situations occur. In the central scheduling intelligent agent, the tag analysis results include at least one or more of the following: BI analysis results, population profile modeling results, consumer product feedback results, and consumer demand insight reports. The central scheduling agent is also used to: convert the multi-layer label map into a standard JSON or Excel interface format and output it to the user in a visual manner; The visualization methods include at least one or more of the following: visualized label cloud map, label matrix, and sentiment heat map; The central scheduling agent, data processing agent, tag recognition agent, sentiment analysis agent, and tag normalization agent collaborate on tasks based on an event-driven model.

[0037] In the specific implementation process, in this system, each intelligent agent collaborates on tasks based on the Event-Driven Architecture (EDA); the central scheduling agent monitors the task queue and module status to achieve intelligent scheduling (such as prioritizing content with drastic fluctuations in sentiment tags); the scheduling strategy has the following characteristics: 1) Task decoupling: supports the plugging and replacing of any module; 2) Elastic scaling: supports multi-instance deployment to improve system concurrency performance; 3) Model switching: supports dynamically loading different model parameters according to industry dimensions; The label analysis and processing flow of this system is the same as that in Example 1; To further improve the recognition accuracy and semantic consistency of various tags (such as product entities, evaluation dimensions, subjective feelings, usage behaviors, etc.) in social media texts, this embodiment also introduces a multi-task learning (MTL) strategy into the tag recognition agent. This strategy aims to learn multiple related tasks simultaneously in a unified model by sharing a representation space, so as to make full use of the potential collaborative relationships between different tasks and avoid the performance bottleneck caused by isolated task modeling. Specifically, consumer tag analysis is first divided into the following four sub-tasks, and then uniformly incorporated into a multi-task framework: a) Entity Recognition Task (NER): Recognize named entities such as brands, product models, and competitors; b) Aspect Extraction: Extracting evaluation dimensions such as "concealing power" and "battery life". c) Subjective Aspect Classification: Judging the sentiment polarity of dimension labels (e.g., "positive", "negative", "neutral"); d) Using behavior tagging: extracting motivational behavior tags such as "repurchase", "trial", "transfer", and "intending to buy"; Next, pre-trained language models such as BERT or RoBERTa are used as a unified encoder to obtain context-dependent deep semantic representations. Meanwhile, the NER task adopts the BIO annotation mechanism + CRF structure; the sentiment classification task adopts the Softmax classifier; the behavior label adopts multi-label sigmoid output; all tasks share the underlying text representation, but have their own output decoding logic. During training, an additional semantic distance loss between label pairs is introduced to ensure that the label representation vectors learned by the model conform to the semantic similarity structure in the label ontology (e.g., "high cost performance" and "worth buying" are close to each other, while labels with opposite sentiments are kept far apart). For complex social media expressions such as contrastive expressions and subject omissions, the model uses BERT bidirectional context modeling combined with the in-sentence core word alignment mechanism (Coreference Alignment Attention) to ensure that the label is paired with the correct object, for example, pointing "it" to "lipstick" instead of "makeup remover" in the previous text; Compared to single-task training, multi-task learning improves label recognition accuracy by an average of 12%-18%. In complex emotional expression scenarios (such as "cheap but lacking quality"), it accurately identifies the "price-positive" and "material-negative" dual-label structures, which traditional models are prone to misclassification. At the same time, the system can be migrated to new industries (such as from cosmetics to beverages) without increasing labeling costs. By adjusting task weights and prompts, label accuracy can be stably maintained. Compared to independent training of multiple models, MTL reduces the overall training time by about 30%, and model parameter sharing brings a compression rate of up to 40%, significantly reducing deployment costs. In social media texts, user expressions often involve complex mapping relationships between multiple emotion holders and evaluation targets. Especially in contexts with multiple emotions in parallel, omitted subjects, and reversed word order, traditional positive / negative binary classification models cannot accurately identify "who expressed what emotion for what". To address this, this embodiment also introduces an Opinion Target Mining mechanism as a core component of the sentiment analysis agent to accurately align emotions with specific evaluation targets. Specifically, the following three types of components in a sentence are located using a finely tuned pre-trained language model: a) Opinion holder (subject): The entity that issues the evaluation (usually "user" by default); b) Opinion target: The object being evaluated (e.g., "lipstick", "signal", "packaging"); c) Opinion expression: Adjectives / verb phrases used to express opinions (such as "very good", "terrible", "acceptable"); Example: The text reads: "This lipstick has a very strong scent, but the packaging is quite luxurious." The identification results are: Target1 = taste, Opinion = negative (too strong); Target2 = outer packaging, Opinion = positive (quite high-end). By modeling the pairing relationship between each object and the emotional phrase, the system performs contextual semantic alignment judgment to identify whether the emotion points to the object and whether there is cross-sentence association (e.g., "It lasts much better than the previous one," where "it" refers to the foundation). If the subject is omitted in the expression (e.g., "It started to peel off after I bought it"), the system infers that the target is "product shell" through semantic completion. After completing the subject-object pairing, each (Target, Opinion) pair is categorized with fine-grained emotion tags. The tagging system may include, but is not limited to: positive: recommended, surprised, satisfied, good feeling...; negative: disappointed, difficult to use, terrible, complain...; neutral: so-so, indifferent, acceptable, needs improvement...; In practical applications, the opinion subject-object analysis method proposed in this embodiment significantly improves the matching accuracy and semantic clarity between emotions and evaluation objects in social media texts. Compared with traditional methods based on overall text emotion judgment, this method can identify the specific object or dimension corresponding to each emotion expression, thereby avoiding the problems of "emotion drift" or "emotion mismatch". In comparative tests, this method improves the accuracy of emotion-target matching in multi-emotion texts by approximately 19% to 35%, especially in user comments involving multiple evaluation dimensions or multiple product entities. For example, when faced with a combined expression such as "poor signal but good screen", the system can identify "signal" as a negative emotion target and "screen" as a positive emotion, effectively reducing the false attribution rate. In addition, this method shows strong adaptability to common problems of referential resolution and subject omission in social media contexts. The system can identify the entity objects corresponding to pronouns such as "this" and "it," thus reliably extracting effective information even in highly colloquial scenarios like "Xiaohongshu" and "Weibo." On relevant test sets, the accuracy of referential inference is improved by approximately 28% compared to traditional methods. This embodiment also introduces fine-grained emotion tags (such as "recommended," "surprise," "so-so," and "disappointing") to support a more discernible and operational emotion classification system. This allows brands to not only know whether users are "satisfied" or "dissatisfied," but also to specifically perceive whether they are "surprised" or "expecting improvement." This improved tag granularity helps brands achieve more precise product iteration and marketing copy optimization. In summary, the opinion subject-object analysis method not only solves the core pain point of "separation of emotion and object" in emotion recognition, but also significantly improves the system's intelligence level through semantic modeling and referential mechanisms. It is an important technological foundation supporting the "high-precision consumer tag insight" of this invention. To more clearly illustrate the technical concept and application process of this system, the following describes in detail how the automated consumer tag analysis system is implemented in actual work, using social media data application scenarios from several typical industries. This embodiment uses data from Xiaohongshu and Weibo platforms as sample input sources, covering the complete process of data collection, intelligent scheduling, tag recognition, and structured output. Scenario 1: Beauty Industry – Product Review Tag Identification and Reputation Analysis; A beauty brand wanted to understand user feedback and preferences on Xiaohongshu (Little Red Book) regarding its latest foundation product. The system's input data consisted of user notes related to the foundation, including text, user attributes, and interaction information. The execution process is as follows: (1) The data processing agent uses the platform's crawler interface or API to collect notes related to foundation, including keywords such as "foundation", "concealing", and "long-lasting makeup"; and performs noise reduction, word segmentation, and extraction of basic tags such as user gender, skin type, and makeup scene on the text data. (2) The central scheduling agent schedules the downstream label recognition agent, sentiment analysis agent and label normalization agent to work together according to the task content; (3) The label recognition intelligent agent (LLM / BERT driven) identifies typical dimension labels, such as: concealing power, lasting power, moisturizing power, and makeup weight; and extracts entity objects, such as: product name, usage scenario (date, work), and competitor brands; (4) The sentiment analysis agent analyzes user statements, such as “It started to oxidize and cake after three hours of use” and “This makeup lasts really well, I don’t need to touch up my makeup when I go out at night”; it identifies positive and negative emotions and subjective expressions and binds them to specific dimension labels, such as generating structural labels such as “coverage power - positive” and “makeup lasting power - negative”. (5) The label normalization agent clusters synonym / near-synonymous labels. For example, it aggregates semantics such as "natural makeup", "not fake", and "good fit" into the standard label "natural makeup". (6) Results Output and Visualization: The output of the central scheduling agent includes the Top 10 hot keywords of dimensions, positive and negative sentiment distribution, user preference map, typical expression examples, etc.; the data can be integrated into the brand BI system in JSON format for market placement adjustment. In this application scenario, the system extracted 15 categories of dimensional tags from 1,000+ user notes, with an accuracy approximately 34% higher than manual annotation, taking only 8 minutes. Compared to the traditional manual annotation process which takes 3-5 hours, this significantly improves analysis efficiency. Scenario 2: 3C Digital Industry - Automatic Summarization of User Purchase Pain Points; A mobile phone brand wants to analyze consumer feedback and purchase concerns regarding its latest flagship phone on Weibo. In this scenario, the system inputs comments and reposts in the format "#flagship model name# + user experience"; the execution process is as follows: (1) The data processing intelligent agent aggregates original text, forwarded and commented data with topic tags and brand keywords; removes advertising content and extracts real user expressions; (2) Label recognition intelligent agent extracts functional dimension labels: battery life, heat generation, signal strength, photography, system smoothness; extracts user background labels, such as photography enthusiasts, game users, office commuters, etc. (3) The sentiment analysis agent analyzes sentiment tendencies, such as resolving "severe frame drops while playing games" and "the camera is much better than my previous flagship" as "game performance - negative" and "imaging capabilities - positive"; at the same time, it identifies neutral suggestion sentiments such as "suggest improvement" and "needs optimization" for use in product development; (4) The tag-unifying agent unifies "battery not durable", "short battery life" and "three charges a day" into "poor battery life performance"; (5) Results output: Output structured analysis results such as “Consumer Attention Dimension Radar Chart”, “Pain Point Evaluation Cloud Chart”, and “Audience-Demand Matrix” to help brands accurately identify shortcomings and iterate products; In this application scenario, the system helped the brand quickly identify three main user complaint points (signal, overheating, and system update bugs) within the first week of product launch, optimize after-sales service strategies in a timely manner, and significantly reduce user churn rate; Scenario 3: Beverage Industry – Seasonal Taste Preference Trend Analysis; The brand wanted to understand consumers' preferences for different types of beverages during the summer. The input was user content from Xiaohongshu (Little Red Book) from July to August related to "refreshing drinks," "cooling," and "lightweight." The execution process was as follows: (1) Data processing intelligent agents aggregate content related to the keyword "cooling"; (2) Label recognition intelligent agent recognizes labels, such as: flavor (lemon, green grape), temperature (cold brew, chilled), health ingredients (low sugar, no preservatives); (3) The emotional analysis agent analyzes positive emotions, such as "quenches thirst without being greasy" and "does not leave a dry mouth after drinking", as well as negative emotions such as "too sweet" and "heavy artificial taste"; The system ultimately outputs "Top 5 Summer Beverage Preference Ingredients", "Health Label Correlation Graph", and "Consumer Sentiment Trend Chart", supporting new product development and advertising copy creation; The above application scenarios demonstrate that this system has excellent cross-industry adaptability and automated consumer tag processing capabilities. It can be quickly deployed in industries such as FMCG, beauty, beverage, and 3C, realizing a closed loop from raw social media data to structural insights, and significantly improving the efficiency of brand market insights and operational decision-making.

[0038] Example 3 like Figure 2 As shown, this embodiment provides an automated consumer tag analysis method based on multi-agent collaboration, based on the system described in Embodiment 1 or 2, including the following steps: S1: The central scheduling agent schedules the data processing agent to collect data from at least one social media platform and perform data preprocessing; S2: The central scheduling agent schedules the tag recognition agent to extract consumer tags of different dimensions from the preprocessed data; S3: The central scheduling agent schedules the sentiment analysis agent to perform context modeling and sentiment tendency identification on the preprocessed data, and binds the identified sentiment tendency to the corresponding consumption tag. S4: The central scheduling agent schedules the tag normalization agent to cluster and normalize all consumer tags bound with sentiment tendencies, generating a structured multi-layer tag map. S5: The central scheduling agent uses the multi-layered tag map to generate tag analysis results and sends them to the user.

[0039] In the specific implementation process, the central scheduling agent first schedules the data processing agent to collect data from at least one social media platform and perform data preprocessing. Secondly, the central scheduling agent schedules the tag recognition agent to extract consumption tags of different dimensions from the preprocessed data; Next, the central scheduling agent schedules the sentiment analysis agent to perform context modeling and sentiment tendency identification on the preprocessed data, and binds the identified sentiment tendency to the corresponding consumption tag; Then, the central scheduling agent schedules the tag normalization agent to cluster and normalize all consumer tags bound with sentimental tendencies, generating a structured multi-layered tag map. Finally, the central scheduling agent uses the multi-layered label graph to generate label analysis results and sends them to the user; This method utilizes multi-agent collaborative modeling and intelligent scheduling technology to streamline the entire process from social media data access, text analysis, tag recognition to result output, constructing an automated tag analysis system for consumer insights. This system boasts technical advantages such as high accuracy, high scalability, and strong industry adaptability, and can be widely used in scenarios such as brand data analysis, user research, and product co-creation.

[0040] The core innovation of this invention lies in achieving fully automated identification and structured analysis of consumer tags through the combination of a multi-agent architecture and natural language processing technology. Compared to existing static algorithm processes or manual rule matching systems, this invention possesses stronger scalability, intelligence, and industry adaptability. The specific technical key points and the content to be protected include the following five aspects: (a) Multi-agent asynchronous cooperative architecture: This invention introduces a multi-agent design concept into a consumer tagging analysis system for the first time. It breaks down the entire analysis process into multiple agent modules with independent functional logic, including a central scheduling agent, a data processing agent, a tag recognition agent, a sentiment analysis agent, and a tag normalization agent. Through unified task allocation by the central scheduling agent, concurrent task processing, elastic resource scaling, and result fusion are achieved, significantly improving processing efficiency. This architecture breaks through the performance bottleneck of traditional single-process call logic and forms the foundation for the system's sustainable expansion and rapid cross-industry adaptation.

[0041] (II) Multidimensional Label Intelligent Recognition Mechanism Based on Large Language Model (BERT / LLM): This invention employs BERT and other large-scale pre-trained language models to propose a multi-dimensional tag recognition mechanism targeting the unstructured, multi-semantic, and emotional characteristics of social media text. By introducing a multi-task learning strategy, it simultaneously models multiple tag categories such as product entities, evaluation dimensions, subjective feelings, and usage behaviors, and combines semantic similarity and contextual dependencies to achieve fine-grained extraction. This solution is particularly suitable for complex social language expression scenarios, such as "fragrant and spicy, good with rice," capable of extracting "flavor tag = fragrant and spicy" and "eating method = with rice," improving tag accuracy and granularity.

[0042] (III) The introduction of subject-object analysis of emotion and the segmentation of emotion tags: Unlike traditional positive / negative classification methods, this invention introduces a subject-object viewpoint analysis approach to identify specific targets associated with emotions. Combined with a fine-grained emotion tag library (such as "recommend," "disappointed," "repurchase," "so-so"), it achieves deeper insights into user emotions, effectively solving the problem of "emotional expression detached from the evaluation object" in text. This helps brands identify emotional fluctuations triggered by specific products / dimensions, and is more instructive for marketing.

[0043] (iv) Tag normalization and ontology fusion mechanism: Considering the differences in user language habits and the diversity of industry expressions, this invention introduces a tag normalization intelligent agent, combining existing unsupervised techniques such as industry tag ontologies, thesaurus, word vectors, and KMeans clustering to achieve standardized tag processing. This mechanism not only improves the consistency and comparability of tag output but also supports unified recognition and comparison of tags across multiple languages ​​and platforms, effectively serving the tag data asset management needs of brands.

[0044] (v) Visualized and structured output and platform integration capabilities: At the final output layer of the system, this invention supports exporting structured tag results to standard formats (JSON, CSV, API integration) and automatically generates diverse charts such as hot word trends, tag clouds, and sentiment trends, facilitating users' quick insight into the analysis results. Simultaneously, the system possesses flexible interface capabilities, enabling integration with internal enterprise systems such as data warehouses, BI platforms, and CRM systems, thereby achieving a closed-loop application of consumer tag analysis from data collection to business decision-making.

[0045] This invention significantly outperforms existing consumer insight systems in terms of technical structure, application effects, and industrial value. By introducing a multi-agent mechanism and a natural language deep understanding model, the system demonstrates significant innovative advantages in multiple aspects, including analysis efficiency, tag quality, sentiment insight capabilities, industry adaptability, and platform scalability. Its beneficial effects can be summarized as follows: 1) This invention uses a central scheduling agent to uniformly manage the agents of each module, supporting asynchronous concurrent execution and intelligent scheduling of tasks. Especially when facing high-frequency social media data streams from multiple platforms, the system can automatically allocate resources and process tasks in parallel, significantly improving throughput and response speed. Compared with the traditional linear processing architecture, the processing efficiency can be improved by an average of more than 70%, meeting the brand's insight needs for "minute-level response". 2) This invention utilizes natural language models such as BERT / LLM and multi-task learning strategies. The system can simultaneously identify multiple categories of consumer tags, including product entities, functional dimensions, user behavior, and subjective evaluations. The tags are fine-grained and have clear semantic structures, making them adaptable to various consumer industry languages, including FMCG, beauty, footwear and apparel, beverages, and 3C products. Compared to static methods based on dictionaries or templates, the tag accuracy is improved by approximately 25% to 40%, making it particularly suitable for varied and ambiguous user expressions. 3) Through the emotional subject-object matching and emotion segmentation strategy, the system can accurately identify "what emotion the user expressed when evaluating which part of the content", which significantly avoids emotional misjudgment. For example, when faced with a complex expression with multiple emotions such as "the product tastes good but the packaging looks a bit cheap", the system can distinguish that "taste" corresponds to positive emotion and "packaging" corresponds to negative emotion. The granularity and accuracy of emotion recognition are significantly better than traditional methods, providing brands with more targeted feedback. 4) To address the differences in user expression across different platforms and times, this invention introduces tag ontology fusion and clustering normalization technology to semantically aggregate synonyms and polysemous words expressed by users. By unifying various expressions such as "high cost-performance ratio," "cheap and affordable," and "worth buying" into the "high cost-performance ratio" tag, the cleanliness of tag statistical results, the consistency of visual expression, and the comparability of subsequent analysis are significantly improved, enabling brands to build a long-term and stable tag asset system; 5) The final output structure of this invention supports multiple data interface formats (such as API, database synchronization, Excel export, etc.) and provides chart visualization interfaces. It can be directly integrated with the brand's existing CRM, CDP, public opinion system or BI platform to achieve a business closed loop of "analysis results as insights implemented". It avoids intermediate links such as manual transfer and data format conversion, greatly improves the enterprise's data utilization rate and saves labor costs. 6) Through the abstract and general "intelligent agent + tag ontology" architecture, this invention can be quickly migrated to new industries or fields. For example, only by updating the tag ontology library and making minor adjustments to the model, cross-industry semantic adaptation from cosmetics to footwear and apparel, and from beverages to 3C digital products can be completed. At the same time, it supports brands to quickly respond to market changes, monitor new product feedback, and compare competitor performance. It is one of the key infrastructures for future AI-driven brand growth strategies. 7) Traditional consumer tag analysis usually requires a lot of manual tag building, rule maintenance and post-operation. This invention achieves a large degree of automation through deep learning and intelligent agent automatic scheduling, which significantly reduces the dependence on data analysts. According to internal testing, under the same analysis scale, the system's human resource cost can be reduced by more than 60%, which is especially suitable for brands to carry out rapid insight and adjustment of advertising strategies during major promotional periods (such as Double 11 and 618) when operating costs are sensitive.

[0046] The same or similar labels correspond to the same or similar parts; The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this application. Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. An automated consumer tag analysis system based on multi-agent collaboration, characterized in that, include: A data processing agent for collecting and preprocessing data from at least one social media platform; The tag recognition agent is equipped with a pre-trained natural language model and tag libraries from different domains, which are used to extract consumer tags of different dimensions from pre-processed data. A sentiment analysis agent is used to perform contextual modeling and sentiment tendency identification on preprocessed data, and to bind the identified sentiment tendency to the corresponding consumption tag. The tag normalization agent is used to cluster and normalize all consumer tags bound with emotional biases, generating a structured multi-layered tag map. The central scheduling agent is connected to the data processing agent, the tag recognition agent, the sentiment analysis agent, and the tag normalization agent, respectively, and is used to coordinate the scheduling of other agents, as well as to generate tag analysis results using the multi-layer tag graph and send them to the user.

2. The automated consumer tagging analysis system based on multi-agent collaboration according to claim 1, characterized in that, The data processing intelligent agent accesses at least one social media platform via API interface or web crawler and collects content posted by users. The preprocessing includes: data cleaning and structural transformation of the collected data, followed by word segmentation, error correction, and sentence segmentation.

3. The automated consumer tagging analysis system based on multi-agent collaboration according to claim 1, characterized in that, In the label recognition intelligent agent, the dimensions of consumer labels include any one or more of the following: entity labels, attribute labels, behavior labels, and evaluation phrases.

4. The automated consumer tagging analysis system based on multi-agent collaboration according to claim 1, characterized in that, In the sentiment analysis agent, context modeling includes: for preprocessed data, combining context to perform subject-object analysis and referential resolution, identifying the evaluator and label attribution; and combining context to extract viewpoints from complex texts with compound semantics, irony, or / and interrogative expressions. The sentiment tendency identification includes classifying the sentiment tendency expressed by the data after context modeling; the sentiment tendency includes at least any one of positive, negative and neutral.

5. The automated consumer tag analysis system based on multi-agent collaboration according to claim 1, characterized in that, In the tag normalization agent, synonymous and / or near-synonymous consumer tags are identified and clustered to form multiple tag clusters; all consumer tags within the same tag cluster are normalized and mapped, and the multi-layer tag map is further constructed.

6. The automated consumer tagging analysis system based on multi-agent collaboration according to claim 1, characterized in that, When coordinating the scheduling of other intelligent agents, the central scheduling agent monitors the status of the task queue and other intelligent agents in real time, specifies the scheduling strategy according to the preset task priority and sends it to other intelligent agents; at the same time, it receives the execution feedback results of the scheduling strategy in real time, performs abnormal recovery according to preset rules and sends alarm information to the user when abnormal situations occur.

7. The automated consumer tagging analysis system based on multi-agent collaboration according to claim 1, characterized in that, In the central scheduling intelligent agent, the tag analysis results include at least one or more of the following: BI analysis results, population profile modeling results, consumer product feedback results, and consumer demand insight reports.

8. The automated consumer tagging analysis system based on multi-agent collaboration according to claim 1, characterized in that, The central scheduling agent is also used to: convert the multi-layer label map into a standard JSON or Excel interface format and output it to the user in a visual manner; The visualization methods include at least one or more of the following: visualized label cloud, label matrix, and sentiment heatmap.

9. An automated consumer tagging analysis system based on multi-agent collaboration as described in any one of claims 1 to 8, characterized in that, The central scheduling agent, data processing agent, tag recognition agent, sentiment analysis agent, and tag normalization agent collaborate on tasks based on an event-driven model.

10. An automated consumer tag analysis method based on multi-agent collaboration, based on the system described in any one of claims 1 to 9, characterized in that, Includes the following steps: S1: The central scheduling agent schedules the data processing agent to collect data from at least one social media platform and perform data preprocessing; S2: The central scheduling agent schedules the tag recognition agent to extract consumer tags of different dimensions from the preprocessed data; S3: The central scheduling agent schedules the sentiment analysis agent to perform context modeling and sentiment tendency identification on the preprocessed data, and binds the identified sentiment tendency to the corresponding consumption tag. S4: The central scheduling agent schedules the tag normalization agent to cluster and normalize all consumer tags bound with sentiment tendencies, generating a structured multi-layer tag map. S5: The central scheduling agent uses the multi-layered tag map to generate tag analysis results and sends them to the user.

Images