An AI product comparison platform tailored to enterprise needs

By constructing a knowledge graph of the enterprise service domain and introducing a feedback and iterative optimization mechanism, the problems of imprecise demand analysis and single comparison strategies in existing technologies have been solved. This has enabled a precise understanding and multi-dimensional comparison of the personalized needs of enterprises, generating fine-grained comparison and analysis reports, and improving the efficiency and accuracy of enterprise product selection.

CN123071052APending Publication Date: 2026-11-10SHANGHAI ECOSTAN NETWORK TECHNOLOGY CO LTD
View PDF 3 Cites -1 Cited by

Patent Information

Application Number
CN202611236367.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-14
Publication Date
2026-11-10

Smart Images

  • Figure CN123071052A_ABST
    Figure CN123071052A_ABST
Patent Text Reader

Abstract

This invention discloses an AI product comparison platform tailored to enterprise needs, belonging to the interdisciplinary field of artificial intelligence and intelligent recommendation technology. It includes a needs analysis module, a product data collection module, a knowledge graph construction module, a candidate product recall module, an intelligent comparison engine, a feedback iterative optimization module, and an output module. The needs analysis module transforms user natural language needs into semantic representations of those needs. The candidate product recall module retrieves a set of candidate products based on the product knowledge graph. The intelligent comparison engine comprehensively scores and ranks candidate products from three dimensions: semantic matching, structured feature similarity, and knowledge graph relevance, and automatically generates a structured comparison analysis report containing multi-dimensional matching scores, a horizontal comparison matrix, and a suitability assessment. This invention enables intelligent, multi-dimensional, and interpretable comparison for enterprise product selection, improving the efficiency of selection decisions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and intelligent recommendation technology, and specifically relates to an AI product comparison platform for enterprise needs. Background Technology

[0002] As enterprises deepen their digital transformation, the product selection decisions they face during the IT infrastructure development process are becoming increasingly complex. Taking enterprise resource planning (ERP) systems, customer relationship management (CRM) systems, and supply chain management (SRM) systems as examples, the market is flooded with products that overlap in functionality, have varying pricing strategies, and different service models. Enterprise users often need to spend significant manpower on information gathering, needs analysis, and product research to make informed purchasing decisions. Traditional product comparison methods rely heavily on manually consulting product manuals, consulting suppliers, or referring to third-party evaluation reports, which is not only inefficient but also fails to cover the unique needs of individual enterprises. In recent years, some e-commerce platforms and enterprise service marketplaces have begun offering product recommendation functions, but their core mechanisms are mostly based on keyword matching or simple rule filtering, lacking a deep semantic understanding of the requirements. The comparison dimensions are often limited to explicit parameters such as price and feature lists, making it difficult to effectively match the implicit needs of enterprises in terms of industry attributes, enterprise size, business scenarios, and integration capabilities.

[0003] Currently, some patent applications involve the field of intelligent matching and recommendation. For example, Chinese patent application CN120353921A discloses an intelligent matching method and system for an intelligent supply chain platform, which achieves feature matching between supply and demand through the collaboration of knowledge graphs and natural language processing; Chinese patent application CN121685066A discloses a big data intelligent recommendation and matching platform for enterprise services, which uses multi-dimensional feature modeling to recommend enterprise services; and Chinese patent application CN122288817A discloses a product recommendation method and related equipment, which uses a large language model to extract user demand features and perform semantic matching with product information. However, the above-mentioned existing technologies mainly focus on recommendation scenarios, that is, selecting a number of candidate products from a product library and presenting them to the user in a list format. Their core goal is to improve the accuracy and recall rate of recommendations.

[0004] The existing technologies mentioned above still have the following significant shortcomings: First, their demand analysis capabilities are limited. Most solutions only process user demands at the level of keyword extraction or simple text classification, lacking the ability to construct a semantic network of demands that incorporates knowledge of the enterprise service domain. This makes it difficult to accurately understand complex demand expressions that include industry attributes, enterprise stages, and scenario constraints. Second, their comparison mechanisms are simplistic. Existing solutions typically employ only one method—semantic similarity calculation or structured feature encoding matching—failing to form a multi-level, multi-dimensional, progressive comparison strategy, resulting in insufficient comprehensiveness and accuracy of the comparison results. Third, their output formats are simple. Existing recommendation systems only return a product ranking list or basic rating to users, lacking fine-grained analysis of the matching status of candidate products across various demand dimensions. Users cannot intuitively understand the advantages and disadvantages of each product, its suitability risks, and its horizontal differences with competitors, making it difficult to directly translate recommendation results into purchasing decision-making basis. Fourth, they lack adaptive optimization capabilities. The matching models of existing solutions are mostly static architectures, unable to dynamically adjust based on user behavior feedback such as clicks, favorites, comparison dwell time, and final transactions. The recommendation effect is difficult to continuously improve with data accumulation. Summary of the Invention

[0005] This invention aims to solve the technical problems of existing enterprise product recommendation and comparison systems, such as insufficient refinement of demand semantic analysis and single comparison strategy. It provides an AI product comparison platform that can achieve accurate demand understanding and multi-dimensional progressive comprehensive comparison tailored to enterprise needs.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: an AI product comparison platform for enterprise needs, comprising: The requirement parsing module is used to receive natural language requirement text input by enterprise users, perform word segmentation, entity recognition and semantic vectorization processing on the natural language requirement text, and generate a requirement semantic representation containing industry tags, scenario tags and constraints. The product data acquisition module is used to collect product data from multiple data sources, including structured attribute information and unstructured description information of the product. The knowledge graph construction module is used to construct a product knowledge graph based on the product data, which includes enterprise entities, product entities, service provider entities, and relationship edges between entities. The candidate product recall module is used to retrieve a set of matching candidate products from the product knowledge graph based on the demand semantic representation. The intelligent comparison engine is used to calculate the semantic matching degree, structured feature similarity, and knowledge graph-based correlation degree between the semantic representation of the demand and each candidate product. The engine weights and fuses the above three scores according to the preset fusion weight to generate a comprehensive comparison score for each candidate product. The candidate products are ranked according to the comprehensive comparison score. For the top-ranked candidate products, the engine automatically extracts the structured attribute information and unstructured description information of each product to generate a structured comparison analysis report containing multi-dimensional matching scores, horizontal comparison matrix, and suitability assessment. The feedback iterative optimization module is used to collect user interaction data on the structured comparison analysis report and iteratively update the preset fusion weights based on a reinforcement learning algorithm. The output module is used to output the structured comparison analysis report to the user.

[0007] This invention constructs a complete automated closed loop from demand input to comparison report output through the collaborative work of seven modules: demand analysis, product collection, knowledge graph construction, candidate recall, intelligent comparison, feedback optimization, and output. This reduces the information search cost and manual input for enterprise product selection. Furthermore, the feedback iterative optimization module enables adaptive and continuous improvement of the matching model, solving the technical problem that existing recommendation systems are static and cannot be continuously optimized with data accumulation.

[0008] The optimized requirement parsing module specifically includes: The domain dictionary construction unit is used to build an enterprise service domain dictionary and rule template library. The domain dictionary includes industry category terms, enterprise size terms, business scenario terms, and budget range terms. The entity recognition unit is used to segment and annotate the natural language requirement text using the domain dictionary, and to identify industry entities, scale entities, scene entities and budget entities in the requirement. Semantic encoding unit, used to semantically encode the segmented text using a pre-trained language model, to generate a demand question vector; A semantic network construction unit is used to construct a demand semantic network based on the entity recognition result and the demand question vector. The demand semantic network includes demand nodes, constraint nodes, and the associated edges between nodes.

[0009] The above solution achieves refined analysis of enterprise users' natural language needs through four units in the requirements analysis module: domain dictionary construction, entity recognition, semantic encoding, and semantic network construction. It can accurately identify semantic elements such as industry attributes, enterprise stage, and scenario constraints in complex requirements such as "lightweight CRM suitable for startup teams" and use a semantic network to represent them in a graph structure, thereby improving the accuracy and completeness of requirements analysis.

[0010] The optimized intelligent comparison engine includes: The semantic matching unit is used to obtain the product description text of each candidate product, perform semantic encoding on the product description text through a pre-trained language model to generate a product semantic vector, and calculate the cosine similarity or Euclidean distance between the demand question vector and the semantic vector of each product as the semantic matching degree. The feature similarity calculation unit is used to extract the demand structure features of enterprise users and the product structure features of candidate products. The demand structure features include at least industry category, enterprise size, business scenario and budget range. The product structure features include at least product category, applicable industry, applicable scale, service area and price range. The demand structure features and product structure features are encoded into feature vectors respectively. The weighted Euclidean distance or Manhattan distance between the demand structure feature vector and the product structure feature vector is calculated as the structure feature similarity. The knowledge reasoning unit is used to map the semantic representation of the demand to a query vector in the product knowledge graph, perform link prediction on the product knowledge graph based on the query vector, and calculate the reasoning association probability between each candidate product entity and the demand, which is used as the association degree based on the knowledge graph. The weighted fusion unit is used to perform a weighted summation of the semantic matching degree, structured feature similarity, and knowledge graph correlation degree according to the preset fusion weights to generate the comprehensive comparison score; The report generation unit is used to generate the structured comparison analysis report based on the comprehensive comparison score and the sub-item scores, combined with the adaptation relationship information in the product knowledge graph.

[0011] The above solution achieves a comprehensive comparison of candidate products from three progressive levels: semantics, structured features, and knowledge association, through the integrated setup of five units in the intelligent comparison engine: semantic matching, feature similarity calculation, knowledge reasoning, weighted fusion, and report generation. This avoids the one-sidedness problem caused by a single comparison dimension in existing technologies, making the comparison results more comprehensive, accurate, and reliable.

[0012] The optimized method for the report generation unit to generate the structured comparison analysis report includes: Extract the structured attribute information and product description text of the top-ranked candidate products from the product database; The structured attribute information is encoded into a structured feature vector, and the product description text is encoded into an unstructured feature vector through a pre-trained language model. The structured feature vector and the unstructured feature vector are then fused through a cross-modal attention mechanism to generate a comprehensive product representation vector for each candidate product. For each candidate product, its sub-scores on the three dimensions of semantic matching degree, structured feature similarity and knowledge graph relevance degree, as well as the comprehensive comparison score, are calculated based on the product comprehensive representation vector. A horizontal comparison matrix is ​​generated based on the sub-scores of each candidate product across various dimensions; Based on the comprehensive comparison score and sub-item score of each candidate product, and combined with the corresponding adaptation relationship information in the product knowledge graph, an adaptation evaluation result including selection recommendation level and adaptation risk warning is generated. The sub-scores, comprehensive comparison scores, horizontal comparison matrix, and suitability evaluation results are integrated according to a preset report template to generate the structured comparison analysis report.

[0013] The above solution uses a cross-modal attention mechanism in the report generation unit to fuse structured attribute information with unstructured descriptive information to generate a comprehensive product representation vector. It makes full use of multi-source heterogeneous data, and the generated comparative analysis report covers sub-item scores, horizontal comparison matrices and suitability assessments, providing users with fine-grained and interpretable decision-making basis. It solves the technical problem that existing recommendation systems only output a ranking list and users cannot intuitively understand the advantages and disadvantages of each product.

[0014] The optimized feedback iterative optimization module specifically includes: The user behavior tracking unit is used to collect user interaction behavior data during the browsing of the structured comparison analysis report. The interaction behavior data includes at least one of the following: click operation, favorite operation, comparison dwell time and final selection result for each candidate product. A reward calculation unit is used to calculate a reward signal based on the interaction behavior data, wherein the reward signal is positively correlated with the user's preference for the product. The weight optimization unit is used to iteratively optimize the preset fusion weights using the preset fusion weights as the action space and the current demand features and product features as the state space, employing a policy gradient algorithm or a Q-learning algorithm with the goal of maximizing cumulative rewards.

[0015] The above scheme establishes an adaptive optimization mechanism based on reinforcement learning through three units in the feedback iterative optimization module: user behavior tracking, reward calculation, and weight optimization. By collecting behavioral data such as user clicks, favorites, dwell time, and final selection results, a reward signal is constructed, and the fusion weights are continuously optimized using policy gradient or Q-learning algorithms, thus achieving adaptive matching effect.

[0016] In this optimized version, the product data acquisition module collects product data from multiple data sources, cleans and standardizes the data, generates structured records in a unified format, and stores them in the product database. The product description text is preprocessed and then stored in a text database. This invention cleans and standardizes multi-source product data through the product data acquisition module, generating structured records in a unified format and storing them in the database. This eliminates the problems of inconsistent formats, data redundancy, and noise interference caused by multi-source data, providing a high-quality data foundation for product knowledge graph construction and subsequent comparison calculations.

[0017] In the optimized version, when the knowledge graph construction module constructs the product knowledge graph based on the product data, it extracts enterprise entities, product entities, and service provider entities from the product description text through named entity recognition, and identifies the relationship edges between entities through a relation extraction model. These relationship edges include at least enterprise-product compatibility relationships, product-service provider supply relationships, and enterprise-enterprise competitive / cooperative relationships. The extracted entities and relationship edges are stored in a graph database to form the product knowledge graph. This invention, through its knowledge graph construction module, uses named entity recognition to extract enterprise, product, and service provider entities, and identifies the relationship edges between entities through a relation extraction model, storing the data in a graph database. This enables the platform to uncover implicit associations and reasoning paths that cannot be discovered through keyword matching alone, providing a rich semantic foundation for candidate product recall and knowledge reasoning.

[0018] In the optimized version, when the candidate product recall module retrieves a matching set of candidate products from the product knowledge graph based on the demand semantic representation, it uses graph query language to retrieve candidate product entities in the product knowledge graph that match the industry tags, scenario tags, and constraints in the demand semantic representation. These are then initially screened and sorted according to their matching degree, and a preset number of candidate products are output to form the candidate product set. This solution uses graph query language in the candidate product recall module to quickly retrieve candidate product entities that match the user's demand semantics in the knowledge graph. While ensuring recall quality, this significantly reduces the computational scale of subsequent fine-grained comparisons, improving the platform's overall response speed and operational efficiency.

[0019] In this optimized approach, the requirement parsing module performs semantic vectorization on the natural language requirement text using a pre-trained language model fine-tuned for the enterprise service domain. This pre-trained language model is trained using a corpus of enterprise service domain data for domain-adaptive fine-tuning. This solution, by employing a pre-trained language model fine-tuned for the enterprise service domain in the requirement parsing module, enables the semantic encoding model to more accurately grasp the semantic features and industry expression habits of the enterprise service domain, thereby improving the accuracy and relevance of the requirement semantic representation.

[0020] The optimized platform also includes: The visualization module is used to present the structured comparison analysis report to the user in an interactive interface that combines charts and text. The interactive interface supports users in marking, annotating, and exporting candidate products. The user management module is used to manage the identity information and permissions of registered users and record the comparison history of each user.

[0021] This invention presents the comparison report in an interactive interface combining charts and text through a visualization module, supporting marking, annotation, and export operations, enabling users to intuitively understand complex comparison results and personalize them; the user management module manages user permissions and records comparison history, ensuring data access security and providing data support for personalized recommendations.

[0022] Compared with the prior art, the beneficial effects of the present invention are: 1. At the level of demand understanding and comparison strategy, this invention constructs a demand semantic parsing model that integrates knowledge from the enterprise service domain. This model achieves a refined understanding of users' natural language demands, accurately identifying semantic elements and their relationships in complex demand expressions, including industry attributes, enterprise stage, business scenarios, and budget constraints. This overcomes the shortcomings of existing technologies that only extract keywords or perform simple text classification, leading to inaccurate demand parsing. Based on this, a three-layer progressive comparison strategy combining semantic matching, structured feature similarity calculation, and knowledge graph reasoning is adopted to comprehensively score candidate products from multiple dimensions. Compared to existing recommendation schemes that only use a single matching method, this improves the comprehensiveness and accuracy of the comparison results. Particularly in the comparison report generation stage, this invention uses a cross-modal attention mechanism to fuse structured attribute information and unstructured descriptive information from multiple sources, automatically generating a structured comparison analysis report covering sub-item scores, horizontal comparison matrices, and suitability assessments. This provides fine-grained interpretability support for enterprise product selection decisions, solving the technical problem that existing recommendation systems only output ranked lists, making it difficult for users to intuitively understand the advantages, disadvantages, and suitability risks of each product.

[0023] 2. Regarding model adaptive capability, this invention introduces a feedback iterative optimization mechanism based on reinforcement learning. By collecting user behavioral data such as clicks, favorites, comparison dwell time, and final selection results of the structured comparison analysis report, a reward signal is constructed. Using preset fusion weights as the action space and current demand features and product features as the state space, the fusion weights of each comparison dimension are continuously optimized using a policy gradient algorithm or Q-learning algorithm. This achieves the technical effect of dynamically and adaptively adjusting the comparison model according to user preferences, overcoming the static architecture defects of existing technologies where matching model parameters are fixed and cannot be continuously improved based on user feedback. This allows the platform's comparison accuracy and user satisfaction to gradually improve with the accumulation of usage data, enabling long-term service capabilities.

[0024] 3. At the overall system architecture level, this invention constructs a complete platform architecture in which the demand analysis module, product data acquisition module, knowledge graph construction module, candidate product recall module, intelligent comparison engine, feedback iteration optimization module, and output module work together. This forms a fully automated closed loop from demand input to comparison report output, reducing the information search cost and manual research input in the enterprise product selection process. At the same time, the product data acquisition module cleans and standardizes multi-source heterogeneous data, and the knowledge graph construction module uses named entity recognition and relation extraction technology to construct a product knowledge graph containing ternary relationships of enterprises, products, and service providers, as well as multiple types of relation edges. This provides a high-quality data foundation and rich semantic association support for efficient graph query recall and knowledge reasoning of candidate products, enabling the platform to maintain high response speed and operating efficiency when processing large-scale enterprise-level product data. Attached Figure Description

[0025] Figure 1 This is a system architecture diagram of an AI product comparison platform for enterprise needs according to the present invention; Figure 2 This is a flowchart of the semantic parsing process in this invention; Figure 3 This is a flowchart of the three-layer progressive comparison strategy in this invention; Figure 4 This is a flowchart of the structured comparison analysis report generation process in this invention; Figure 5 This is a schematic diagram of the feedback iterative optimization mechanism in this invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] like Figures 1-5 As shown, this invention introduces an AI product comparison platform for enterprise needs, comprising: The requirement parsing module is used to receive natural language requirement text input by enterprise users, perform word segmentation, entity recognition and semantic vectorization processing on the natural language requirement text, and generate a requirement semantic representation containing industry tags, scenario tags and constraints. The product data acquisition module is used to collect product data from multiple data sources, including structured attribute information and unstructured description information of the product. The knowledge graph construction module is used to construct a product knowledge graph based on the product data, which includes enterprise entities, product entities, service provider entities, and relationship edges between entities. The candidate product recall module is used to retrieve a set of matching candidate products from the product knowledge graph based on the demand semantic representation. The intelligent comparison engine is used to calculate the semantic matching degree, structured feature similarity, and knowledge graph-based correlation degree between the semantic representation of the demand and each candidate product. The engine weights and fuses the above three scores according to the preset fusion weight to generate a comprehensive comparison score for each candidate product. The candidate products are ranked according to the comprehensive comparison score. For the top-ranked candidate products, the engine automatically extracts the structured attribute information and unstructured description information of each product to generate a structured comparison analysis report containing multi-dimensional matching scores, horizontal comparison matrix, and suitability assessment. The feedback iterative optimization module is used to collect user interaction data on the structured comparison analysis report and iteratively update the preset fusion weights based on a reinforcement learning algorithm. The output module is used to output the structured comparison analysis report to the user.

[0028] In this invention, the requirement parsing module specifically includes a domain dictionary construction unit, an entity recognition unit, a semantic encoding unit, and a semantic network construction unit. The domain dictionary construction unit pre-builds an enterprise service domain dictionary and a rule template library. The domain dictionary covers multiple core vocabulary categories for enterprise services, such as industry category terms, enterprise size terms, business scenario terms, and budget range terms. The rule template library contains parsing rules for various requirement expression patterns, providing a domain knowledge foundation for subsequent entity recognition and semantic parsing. The entity recognition unit, upon receiving natural language requirement text input by an enterprise user, calls the domain dictionary to perform word segmentation on the natural language requirement text and annotates the segmentation results based on the rule template library, thereby accurately identifying key requirement elements such as industry entities, size entities, scenario entities, and budget entities from the requirement text. The semantic encoding unit uses a pre-trained language model to process the segmented text. After word and entity annotation processing, the text undergoes semantic encoding, transforming the user's textual requirements into a dense requirement question vector. This vector incorporates the contextual semantic information of the requirement text, providing a numerical representation basis for subsequent semantic matching calculations. The semantic network construction unit is used to construct a structured requirement semantic network based on the entity recognition results output by the entity recognition unit and the requirement question vector generated by the semantic encoding unit. This network uses the identified entities as requirement nodes, the corresponding constraints of each entity as constraint nodes, and the semantic relationships and constraint associations between entities as association edges. The requirement semantic network fully represents the various elements in the user's requirements and their inherent relationships in a graph structure, providing a clear and complete requirement input for subsequent candidate product recall and multi-dimensional comparison.

[0029] In this invention, the intelligent comparison engine is the core decision-making component of the platform. It is used to comprehensively evaluate and rank the candidate products from three dimensions: semantics, structured features, and knowledge association, based on the candidate product set output by the candidate product recall module, and automatically generate a fine-grained comparison analysis report. Specifically, it includes a semantic matching unit, a feature similarity calculation unit, a knowledge reasoning unit, a weighted fusion unit, and a report generation unit. The semantic matching unit is used to acquire product description text of each candidate product, semantically encode the product description text using a pre-trained language model to generate product semantic vectors, and then calculate the cosine similarity or Euclidean distance between the demand question vector and each product semantic vector. The obtained similarity value is used as the semantic matching degree reflecting the degree of relevance between the demand and the product at the text semantic level. The feature similarity calculation unit is used to extract the structural features of enterprise users' demands and the structural features of candidate products. The demand structural features include at least industry category, enterprise size, business scenario, and budget range. The product structural features include at least product category, applicable industry, applicable scale, service area, and price range. The demand structural features and product structural features are encoded into structured feature vectors respectively. By calculating the weighted Euclidean distance or Manhattan distance between the demand structured feature vector and the product structured feature vector, the structured feature similarity reflecting the degree of matching between the demand and the product in each explicit parameter dimension is obtained. The knowledge reasoning unit is used to map the demand semantic representation to the query vector in the product knowledge graph. The query vector is used to predict links in the product knowledge graph, and the inference association probability between each candidate product entity and the demand is calculated. This inference association probability is used as the association degree based on the knowledge graph to measure the strength of the association between the demand and the product through the implicit relationship path in the knowledge graph. The weighted fusion unit is used to perform weighted summation of the semantic matching degree, structured feature similarity and knowledge graph association degree according to preset fusion weights, thereby generating a comprehensive comparison score for each candidate product, and sorting the candidate products in descending order according to the comprehensive comparison score. The report generation unit is used to generate a structured comparison analysis report containing the sub-scores of each candidate product in each dimension, the horizontal comparison matrix between each candidate product and the corresponding adaptation relationship information extracted from the product knowledge graph, based on the comprehensive comparison score of each candidate product and its sub-scores in the three dimensions of semantic matching degree, structured feature similarity and knowledge graph association degree. This analysis report presents the comprehensive performance and relative advantages and disadvantages of each candidate product in a structured form, providing fine-grained and interpretable comparison basis for the final selection decision of enterprise users.

[0030] In this invention, the report generation unit generates the structured comparison analysis report in the following manner: First, it extracts the structured attribute information and product description text of each candidate product ranked at the top after being sorted by the weighted fusion unit from the product database. The structured attribute information includes fields such as product category, applicable industry, applicable scale, service area, price range, and service rating. The product description text includes unstructured content such as product function introduction, applicable scenario description, and user reviews. Then, it performs numerical encoding and normalization on each field in the structured attribute information to generate a structured feature vector. At the same time, it uses a pre-trained language to generate the product description text. The model performs semantic encoding to generate unstructured feature vectors, and then fuses these structured and unstructured feature vectors using a cross-modal attention mechanism. This mechanism calculates the attention weight distribution of structured features relative to unstructured features, giving higher contribution weights to fields in the structured data most relevant to the current need, while effectively preserving semantically complementary descriptive content in the unstructured text. This generates a comprehensive product representation vector that integrates both types of heterogeneous information. Based on this, for each candidate product, its semantic matching degree and structured features are calculated using the comprehensive product representation vector. The system calculates a comprehensive comparison score by weighting and fusing sub-scores across three dimensions: similarity, knowledge graph relevance, and overall performance. A horizontal comparison matrix is ​​generated based on these sub-scores, presenting a table format that clearly shows the relative performance of all candidate products across each dimension, allowing users to quickly identify the strengths and weaknesses of each product. Furthermore, based on the comprehensive comparison score and sub-scores of each candidate product, and combined with enterprise-product compatibility information and product-service provider supply relationship information extracted from the product knowledge graph, a compatibility assessment result is generated, including a selection recommendation level and compatibility risk warnings. The selection recommendation level is determined by the comprehensive comparison score and sub-scores. The ranking interval mapping of the comparison scores is determined, and the adaptation risk warning is triggered by the identification of dimensions where the confidence of the associated path in the knowledge graph is lower than a preset threshold. Finally, the sub-scores, comprehensive comparison scores, horizontal comparison matrices, and adaptation evaluation results of each candidate product are structured and integrated according to a preset report template to generate a complete, intuitive, and directly applicable structured comparison analysis report for enterprise procurement decisions. This analysis report presents the comprehensive performance, horizontal differences, and selection suggestions of each candidate product in each evaluation dimension in a standardized format, solving the problem that existing recommendation systems only output product ranking lists and users cannot obtain fine-grained decision-making basis.

[0031] In this invention, the feedback iterative optimization module is a key component driving the platform to achieve adaptive continuous improvement. Specifically, it includes a user behavior tracking unit, a reward calculation unit, and a weight optimization unit. The user behavior tracking unit collects various interactive behavior data from the user during the entire process of browsing the structured comparison analysis report after it is output to the user. This interactive behavior data includes, but is not limited to, user clicks on the details pages of each candidate product, user favorites of specific candidate products, the duration of page dwell when comparing candidate products, and the final selection result submitted by the user after completing the comparison. This interactive behavior data reflects the user's implicit preferences and actual selection tendencies for candidate products from different dimensions. The reward calculation unit receives the interactive behavior data collected by the user behavior tracking unit, calculates the corresponding reward value for each interactive behavior event based on a preset reward function, and accumulates all reward values ​​generated by the same user in the same comparison session to generate a comprehensive reward signal. The reward function is set based on the principle that the reward signal is positively correlated with the user's preference for the product; that is, candidate products that are clicked, favorited, have a longer dwell time, or are selected as the final result receive higher rewards. The reward value is calculated based on the user's current state and the reward value. Unnoticed candidate products receive lower or even zero reward values. The weight optimization unit uses the preset fusion weights as the agent's action space, the feature vector resulting from concatenating the current user's needs and the candidate products' features as the state space, and the comprehensive reward signal output by the reward calculation unit as the immediate feedback. It iteratively optimizes the preset fusion weights using a policy gradient algorithm or a Q-learning algorithm, aiming to maximize the cumulative reward in multiple comparison sessions. Specifically, in each comparison session, the agent outputs a set of fusion weights based on its current state. The weighted fusion unit generates the comparison results according to these weights and presents them to the user. After the user's interaction behavior is fed back by the reward calculation unit with a comprehensive reward signal, the weight optimization unit adjusts the parameters of the weight policy network based on this reward signal through policy gradient updates or Q-value updates. This allows the platform to gradually learn the optimal weight configuration that adapts to the current user group's preference distribution, ultimately achieving an adaptive optimization effect where the comparison strategy continuously evolves with the accumulation of user usage data.

[0032] In this invention, the product data acquisition module serves as the data entry point for the platform. It collects raw product data from multiple heterogeneous data sources, including at least the product's official website, third-party enterprise service evaluation platforms, publicly available data from app stores, and standardized product information forms provided by suppliers. After collecting the raw product data, the product data acquisition module first performs data cleaning on the raw product data. This data cleaning process includes removing duplicate records, correcting incorrectly formatted fields, completing missing key field values, and identifying and removing abnormal data records. Subsequently, the product data acquisition module performs standardization processing on the cleaned product data. This standardization process maps the heterogeneous product data from different data sources into a standardized field structure according to a preset unified data model. The unified data model includes at least the following fields: product identifier, product name, product category, applicable industry, applicable scale, service region, price range, service rating, and product description text. The value format and range of each field are uniformly converted according to preset specifications to generate consistent and well-structured product data records, which are then stored in the product database. Simultaneously, the product data acquisition module performs separate text preprocessing on the product description text. This preprocessing includes, but is not limited to, removing HTML tags and special characters, performing word segmentation, removing stop words, and normalizing the text content. After this preprocessing, standardized product description text is obtained, and this standardized product description text is associated with its corresponding product identifier and stored in the text database. This provides a high-quality, uniformly formatted product data input foundation for the subsequent knowledge graph construction module, semantic matching unit, and report generation unit.

[0033] In this invention, the knowledge graph construction module is the core foundational component of the platform for realizing the structured organization of product knowledge and implicit relational reasoning. It is used to construct the product knowledge graph based on the product data output by the product data acquisition module. The specific construction process is as follows: First, the knowledge graph construction module reads the structured attribute information of each product from the product database and the corresponding standardized product description text from the text database, and extracts entities from the product description text through a named entity recognition model. The named entity recognition model employs a deep learning-based sequence labeling architecture. After fine-tuning and training with enterprise service domain corpora, it can accurately identify and extract enterprise entities, product entities, and service provider entities from unstructured product description text, assigning each extracted entity a globally unique identifier and entity type label. Then, the knowledge graph construction module uses a relation extraction model to identify and classify the semantic relationships between the extracted entities. This relation extraction model is built based on a pre-trained language model, taking entity pairs and their context text as input and outputting the classification probability of the relationship type between entity pairs. It identifies at least several types of relation edges, including enterprise-product fit relationships, product-service provider supply relationships, and enterprise-enterprise competitive / cooperative relationships. The enterprise-product fit relationship is used to characterize the applicability between a specific enterprise entity and a specific product entity. The product-service provider supply relationship is used to characterize that a specific product entity is provided by a specific service provider entity, and the enterprise-enterprise competition and cooperation relationship is used to characterize the direct competition or cooperation between different enterprise entities in a specific product domain. Finally, the knowledge graph construction module organizes the extracted entities as nodes and the identified various relationship edges as the connection edges between nodes according to the graph structure data model, and stores them in a graph database to form a product knowledge graph containing multiple types of entity nodes and multiple types of relationship edges. The product knowledge graph systematically stores the three core entities of enterprises, products and service providers in the product domain and their multi-dimensional relationship information in the form of a graph structure, providing a clear and semantically rich knowledge support foundation for the subsequent candidate product recall module to perform graph query retrieval and for the knowledge reasoning unit to perform link prediction and correlation calculation.

[0034] In this invention, the candidate product recall module is a crucial bridge connecting demand analysis and fine-grained comparison within the platform. Based on the demand semantic representation generated by the demand analysis module, it quickly retrieves and filters a set of candidate products that initially match the user's needs from the product knowledge graph, thereby narrowing the scope of subsequent fine-grained comparison. Specifically, the candidate product recall module first analyzes the demand semantic representation, extracting industry tags, scenario tags, and various constraints as search keywords. The industry tags identify the user's industry sector, the scenario tags identify the user's specific business application scenario, and the constraints include at least enterprise size constraints and budget range constraints. Then, based on the extracted search keywords, the candidate product recall module performs a multi-condition combined search in the product knowledge graph using graph query language. The specific search strategy is as follows: using the industry tags and scenario tags as the primary matching conditions, it retrieves candidate product entity nodes and their associated enterprise entity nodes and service provider entity nodes that match the industry tags and scenario tags in the product knowledge graph; then, it uses the enterprise size constraints and budget range constraints as filtering criteria. The system further filters candidate product entities in the search results, eliminating those that do not meet the constraints, thereby obtaining a list of candidate product entities that match the semantics of the user's needs. Subsequently, the candidate product recall module performs an initial screening and ranking based on the degree of matching between each candidate product entity and the search keywords. The degree of matching is calculated based on the number of matching industry tags, scene tags, and the number of constraints met for each candidate product entity; the more matching items, the higher the ranking. Finally, according to the initial screening and ranking results, the candidate product recall module selects a preset number of candidate product entities from the top-ranked list, extracts the identification information and corresponding structured attribute information of each candidate product entity, forms the candidate product set, and outputs it to the intelligent comparison engine as the input data range for subsequent three-layer progressive fine comparison. Through the above-mentioned multi-condition combination retrieval and initial screening and ranking mechanism based on graph query language, the candidate product recall module can quickly locate candidate products related to the semantics of user needs in a large-scale product knowledge graph, reducing the retrieval scale of subsequent fine comparisons while ensuring recall quality, thereby improving the overall response speed and operating efficiency of the platform.

[0035] In this invention, the pre-trained language model used by the requirement parsing module when semantically vectorizing the natural language requirement text is not a general basic language model, but a dedicated semantic encoding model that has been adaptively fine-tuned for the enterprise service domain. Specifically, this dedicated semantic encoding model is obtained as follows: First, a large-scale original corpus of enterprise service domain data is collected from multiple channels, including internet enterprise service websites, official documentation of enterprise-level software products, third-party evaluation reports, enterprise user selection survey questionnaires, and product service agreement texts. Second, the collected original corpus undergoes data cleaning, deduplication, text normalization, and domain tagging to construct a dedicated corpus of enterprise service domain data covering multiple industries and enterprise scale scenarios. This dedicated corpus of enterprise service domain data at least covers professional terminology, product function descriptions, industry scenario descriptions, and enterprise requirement expression patterns from multiple enterprise service sub-sectors such as enterprise resource planning systems, customer relationship management systems, supply chain management systems, human resource management platforms, collaborative office platforms, and cloud infrastructure services. Then, based on the dedicated corpus of enterprise service domain data, a masked language model pre-training task is used to further refine the general basic pre-trained language model. Domain-adaptive pre-training continues, driving model parameters to iteratively update on the domain corpus through a masked language model loss function. This allows the model to learn and master the vocabulary distribution characteristics, sentence expression habits, and semantic association patterns of the enterprise service domain. Finally, after convergence of the domain-adaptive pre-training, the finely tuned pre-trained language model for the enterprise service domain is obtained. After the encoder part of this pre-trained language model is fully trained on the domain corpus, the semantic vectors generated by it can more accurately reflect the semantic distance between enterprise service domain-specific terms and industry expressions in the vector space. That is, synonymous or near-synonymous demand expressions have a closer distance in the vector space, and the distinction from unrelated expressions is more obvious. This improves the domain-specificity and representation accuracy of the demand parsing module in semantically vectorizing user natural language demand text, providing a more accurate numerical input basis for subsequent semantic matching calculations and comparison analyses.

[0036] In this invention, the platform further includes a visualization module and a user management module, which respectively enhance the overall functional architecture of the platform from the dimensions of user interaction experience and system security management. The visualization module presents the structured comparison analysis report generated by the intelligent comparison engine to the user in an interactive interface that organically combines charts and text. Specifically, the interactive interface includes at least one radar chart display area, at least one horizontal bar chart comparison area, and at least one detailed data table display area. The radar chart display area visually presents the comprehensive performance profile of each candidate product in three dimensions—semantic matching degree, structured feature similarity, and knowledge graph relevance—in the form of a multi-dimensional radar chart. The horizontal bar chart comparison area visually compares the scores of each candidate product in each dimension using horizontal bar charts. The detailed data table area lists in detail the structured attribute information, sub-scores for each dimension, comprehensive comparison scores, and selection recommendations from the suitability evaluation results for each candidate product in tabular form. The interactive interface includes risk warnings and adaptation features. Furthermore, it allows users to mark candidate products of interest while browsing the comparison report, including adding custom tags or importance star ratings, custom notes, recording user evaluations or questions regarding specific product dimensions, and exporting the current comparison report and all its charts and text content in PDF or Excel format. This enables users to save the comparison analysis results locally or share and discuss them within the team, significantly improving decision-making and collaboration efficiency. The user management module is used to uniformly manage the identity information and operation permissions of registered users using the platform. Specifically, the user management module includes at least a user registration and login management unit, a user role and permission allocation unit, and a user operation log recording unit. The user registration and login management unit reviews user registration applications and assigns unique user identifiers and initial access credentials to users who pass the review. The user role and permission allocation unit assigns different levels of access permissions to different users based on their organization and job title. The user operation log recording unit records every operation behavior of each user on the platform, including at least the user's login time, the input requirements of each comparison session, the comparison reports viewed, and the final selection result. It also constructs a comparison history database for each user. This comparison history database can be used to support the user behavior tracking unit in collecting user preference data and calculating reward signals, and can also serve as the basic data source for the feedback iterative optimization module to perform weight optimization. Furthermore, it can provide historical reference for different users in the same organization when selecting similar products in the future. Thus, while ensuring the security of platform data access and the traceability of user operations, it further supports the platform's personalized recommendation and adaptive optimization capabilities.

[0037] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Although the applicant has described the present invention in detail with reference to preferred embodiments, those skilled in the art should understand that any modifications or equivalent substitutions made to the technical solutions of the present invention cannot depart from the spirit and scope of the present invention and should be covered within the scope of the claims of the present invention.

Claims

1. An AI product comparison platform for enterprise needs, characterized in that, include: The requirement parsing module is used to receive natural language requirement text input by enterprise users, perform word segmentation, entity recognition and semantic vectorization processing on the natural language requirement text, and generate a requirement semantic representation containing industry tags, scenario tags and constraints. The product data acquisition module is used to collect product data from multiple data sources, including structured attribute information and unstructured description information of the product. The knowledge graph construction module is used to construct a product knowledge graph based on the product data, which includes enterprise entities, product entities, service provider entities, and relationship edges between entities. The candidate product recall module is used to retrieve a set of matching candidate products from the product knowledge graph based on the demand semantic representation. The intelligent comparison engine is used to calculate the semantic matching degree, structured feature similarity, and knowledge graph-based correlation degree between the semantic representation of the demand and each candidate product. The engine weights and fuses the above three scores according to the preset fusion weight to generate a comprehensive comparison score for each candidate product. The candidate products are ranked according to the comprehensive comparison score. For the top-ranked candidate products, the engine automatically extracts the structured attribute information and unstructured description information of each product to generate a structured comparison analysis report containing multi-dimensional matching scores, horizontal comparison matrix, and suitability assessment. The feedback iterative optimization module is used to collect user interaction data on the structured comparison analysis report and iteratively update the preset fusion weights based on a reinforcement learning algorithm. The output module is used to output the structured comparison analysis report to the user.

2. The AI ​​product comparison platform for enterprise needs as described in claim 1, characterized in that, The requirement analysis module specifically includes: The domain dictionary construction unit is used to build an enterprise service domain dictionary and rule template library. The domain dictionary includes industry category terms, enterprise size terms, business scenario terms, and budget range terms. The entity recognition unit is used to segment and annotate the natural language requirement text using the domain dictionary, and to identify industry entities, scale entities, scene entities and budget entities in the requirement. Semantic encoding unit, used to semantically encode the segmented text using a pre-trained language model, to generate a demand question vector; A semantic network construction unit is used to construct a demand semantic network based on the entity recognition result and the demand question vector. The demand semantic network includes demand nodes, constraint nodes, and the associated edges between nodes.

3. The AI ​​product comparison platform for enterprise needs as described in claim 1, characterized in that, The intelligent comparison engine includes: The semantic matching unit is used to obtain the product description text of each candidate product, perform semantic encoding on the product description text through a pre-trained language model to generate a product semantic vector, and calculate the cosine similarity or Euclidean distance between the demand question vector and the semantic vector of each product as the semantic matching degree. The feature similarity calculation unit is used to extract the demand structure features of enterprise users and the product structure features of candidate products. The demand structure features include at least industry category, enterprise size, business scenario and budget range. The product structure features include at least product category, applicable industry, applicable scale, service area and price range. The demand structure features and product structure features are encoded into feature vectors respectively. The weighted Euclidean distance or Manhattan distance between the demand structure feature vector and the product structure feature vector is calculated as the structure feature similarity. The knowledge reasoning unit is used to map the semantic representation of the demand to a query vector in the product knowledge graph, perform link prediction on the product knowledge graph based on the query vector, and calculate the reasoning association probability between each candidate product entity and the demand, which is used as the association degree based on the knowledge graph. The weighted fusion unit is used to perform a weighted summation of the semantic matching degree, structured feature similarity, and knowledge graph correlation degree according to the preset fusion weights to generate the comprehensive comparison score; The report generation unit is used to generate the structured comparison analysis report based on the comprehensive comparison score and the sub-item scores, combined with the adaptation relationship information in the product knowledge graph.

4. The AI ​​product comparison platform for enterprise needs as described in claim 3, characterized in that, The specific methods by which the report generation unit generates the structured comparison analysis report include: Extract the structured attribute information and product description text of the top-ranked candidate products from the product database; The structured attribute information is encoded into a structured feature vector, and the product description text is encoded into an unstructured feature vector through a pre-trained language model. The structured feature vector and the unstructured feature vector are then fused through a cross-modal attention mechanism to generate a comprehensive product representation vector for each candidate product. For each candidate product, its sub-scores on the three dimensions of semantic matching degree, structured feature similarity and knowledge graph relevance degree, as well as the comprehensive comparison score, are calculated based on the product comprehensive representation vector. A horizontal comparison matrix is ​​generated based on the sub-scores of each candidate product across various dimensions; Based on the comprehensive comparison score and sub-item score of each candidate product, and combined with the corresponding adaptation relationship information in the product knowledge graph, an adaptation evaluation result including selection recommendation level and adaptation risk warning is generated. The sub-scores, comprehensive comparison scores, horizontal comparison matrix, and suitability evaluation results are integrated according to a preset report template to generate the structured comparison analysis report.

5. The AI ​​product comparison platform for enterprise needs as described in claim 1, characterized in that, The feedback iterative optimization module specifically includes: The user behavior tracking unit is used to collect user interaction behavior data during the browsing of the structured comparison analysis report. The interaction behavior data includes at least one of the following: click operation, favorite operation, comparison dwell time and final selection result for each candidate product. A reward calculation unit is used to calculate a reward signal based on the interaction behavior data, wherein the reward signal is positively correlated with the user's preference for the product. The weight optimization unit is used to iteratively optimize the preset fusion weights using the preset fusion weights as the action space and the current demand features and product features as the state space, employing a policy gradient algorithm or a Q-learning algorithm with the goal of maximizing cumulative rewards.

6. The AI ​​product comparison platform for enterprise needs as described in claim 1, characterized in that, After collecting product data from multiple data sources, the product data acquisition module performs data cleaning and standardization processing on the product data, generates a unified format of structured product data records and stores them in the product database, and performs text preprocessing on the product description text before storing it in the text database.

7. The AI ​​product comparison platform for enterprise needs as described in claim 1, characterized in that, When the knowledge graph construction module constructs the product knowledge graph based on the product data, it extracts enterprise entities, product entities, and service provider entities from the product description text through named entity recognition, and identifies the relationship edges between entities through a relationship extraction model. The relationship edges include at least enterprise-product adaptation relationships, product-service provider supply relationships, and enterprise-enterprise competitive cooperation relationships. The extracted entities and relationship edges are stored in the form of a graph database to form the product knowledge graph.

8. The AI ​​product comparison platform for enterprise needs as described in claim 1, characterized in that, When the candidate product recall module retrieves a matching set of candidate products from the product knowledge graph based on the demand semantic representation, it uses graph query language to retrieve candidate product entities in the product knowledge graph that match the industry tags, scenario tags, and constraints in the demand semantic representation, and performs initial screening and sorting according to the degree of matching, outputting a preset number of candidate products to form the candidate product set.

9. The AI ​​product comparison platform for enterprise needs as described in claim 1, characterized in that, When the requirement parsing module performs semantic vectorization processing on the natural language requirement text, it adopts a pre-trained language model that is fine-tuned for the enterprise service domain. The pre-trained language model is used as training data by a corpus of the enterprise service domain for domain adaptive fine-tuning.

10. The AI ​​product comparison platform for enterprise needs according to any one of claims 1 to 9, characterized in that, The platform also includes: The visualization module is used to present the structured comparison analysis report to the user in an interactive interface that combines charts and text. The interactive interface supports users in marking, annotating, and exporting candidate products. The user management module is used to manage the identity information and permissions of registered users and record the comparison history of each user.

Citation Information

Patent Citations

  • Intelligent matching method and system for intelligent supply chain platform

    CN120353921A

  • Enterprise service-oriented big data intelligent recommendation and matching platform

    CN121685066A

  • Product recommendation method and related equipment

    CN122288817A