Knowledge graph construction method based on large model

Through large-scale models, the quality of product information on e-commerce platforms is solved, personalized recommendations and efficiency improvements are achieved, and different user needs are met.

CN119066209AInactive Publication Date: 2024-12-03JIANGSU RED NET TECH CO LTD
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411356165.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-27
Publication Date
2024-12-03
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The lack of detailed description of product information on e-commerce platforms and the different data quality leads to low efficiency in knowledge graph construction and recommendation systems. Especially in LBS scenarios, information quality is uneven, it is difficult to extract knowledge, and there are many data dimensions, which requires a large amount of processing and maintenance. Attributes that affect map construction and update.

Method used

The knowledge graph construction method based on large models is adopted to collect and integrate data, quality evaluation, integrity judgment and completion, feature information is extracted, entities and relationships are identified, knowledge graphs are constructed, and recommendations are made based on user data information.

Benefits of technology

It improves the accuracy and personalization of the recommendation system, ensures that recommendations are highly correlated with user interests, reduces the time for screening products, improves user experience and sales conversion rate, and optimizes inventory management and supply chain efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066209A_ABST
    Figure CN119066209A_ABST
Patent Text Reader

Abstract

The invention discloses a knowledge graph construction method based on a large model, and relates to the technical field of e-commerce platforms, and the method comprises the steps: carrying out the quality evaluation of collected data at regular intervals, carrying out the screening according to an evaluation result, and removing the data with insufficient accuracy, consistency and reliability; performing integrity judgment on the data passing the evaluation, and transmitting the data passing the judgment to the next step; according to the method, through accurate text similarity calculation, it is ensured that commodities provided by a recommendation system are highly related to interests and preferences of users, so that the satisfaction degree of the users is improved, recommendations received by the users better meet the actual requirements of the users, and the user experience is improved. The time for screening the commodities is shortened, the browsing and shopping experience of the user on the e-commerce platform is improved, the recommendation with high correlation is more likely to promote the user to purchase, and therefore the sales conversion rate of the e-commerce platform is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce platforms, and more specifically, to a knowledge graph construction method based on a large model. Background Art

[0002] Knowledge graphs can integrate knowledge from different sources and discover new knowledge and relationships through reasoning mechanisms, improving the accuracy and efficiency of information retrieval. In search engines, knowledge graphs can more accurately understand the user's query intent and context, and provide more relevant and rich search results. Knowledge graphs enable machines to understand and answer more complex questions, achieving more accurate information retrieval and knowledge discovery. By analyzing user behavior and preferences, knowledge graphs can help recommendation systems provide personalized and accurate recommendation results. Knowledge graphs are widely used in the field of artificial intelligence and are becoming increasingly important.

[0003] With the development of e-commerce platforms, in order to intelligently recommend suitable products to users, more and more e-commerce platforms will also build their own knowledge graphs. However, in actual use, since the product information on e-commerce platforms is often mainly titles and pictures, lacking detailed descriptive information, and the product data uploaded by merchants may be incomplete, especially in the LBS (location-based service) e-commerce scenario, the input of a large amount of product data is often accompanied by the problem of uneven information quality. In addition, although the product details page contains a lot of knowledge information, its quality and structure are different, making it difficult to extract knowledge from it;

[0004] The data dimensions involved in the commodity field are numerous, including general attributes (such as brand, specification, packaging, flavor, etc.) and category-specific attributes (such as fat content, whether it contains sugar, battery capacity, etc.). This leads to efficiency issues in data construction and requires processing and maintaining a large number of attribute dimensions, which is a challenge in the construction and update of knowledge graphs.

[0005] In daily life, users rely on a wealth of common sense knowledge to understand product information, such as inferring the true meaning of a product through a brief description. The construction of a knowledge graph needs to combine this common sense knowledge, which usually involves semantic understanding of a large amount of unstructured text. At the same time, in professional fields such as medicine and personal care, the construction of graphs also needs to rely on professional knowledge. The knowledge accuracy requirements in these fields are extremely high, and efficient graph construction requires a combination of experts and algorithms. Summary of the invention

[0006] To solve the above problems, the present invention provides a method for constructing a knowledge graph based on a large model.

[0007] The present invention provides a method for constructing a knowledge graph based on a large model, comprising the following steps:

[0008] Collect the internal and external data of the platform's products and integrate the collected data into a unified data warehouse. The specific steps are as follows:

[0009] Regularly conduct quality assessments on the collected data and screen them based on the assessment results to remove data that lacks accuracy, consistency, and reliability;

[0010] The integrity of the data that has passed the evaluation is further judged, and the data that has passed the judgment is transmitted to the next step;

[0011] The data that has been judged for integrity is completed, and then the completed data is transmitted to the next step;

[0012] Preprocess the commodity data in the data warehouse and extract the required feature information;

[0013] The extracted feature information is input into the large model to identify entities in the text and the relationships between entities;

[0014] Determine the classification system and the relationship between entities in the text ontology, store the classification system, entities and the relationship between entities in the database to build a knowledge graph;

[0015] Obtain user data information and make recommendations to users based on the user data information and the similarity of products in the knowledge graph.

[0016] Preferably, the collected data are regularly evaluated for quality, and the data are screened according to the evaluation results to remove data with insufficient accuracy, consistency and reliability in the following specific manner:

[0017] The texts in the collected data are randomly checked to obtain the number of random check errors, and the number of text random check errors is divided by the total number of random checks to obtain the accuracy of the data;

[0018] Check whether each text in the collected data follows the predefined format, and obtain the consistency of the data by dividing the number of texts that meet the consistency requirements by the total number of texts in the data;

[0019] Obtain the number of stable texts in the data, divide the number of stable texts by the total number of texts in the data, and obtain the reliability of the data;

[0020] The accuracy, consistency and reliability of the acquired data are compared with the corresponding historical data averages. If any one of the accuracy, consistency and reliability of the data is lower than the corresponding historical data average, the data is judged to be not in compliance with the requirements. Otherwise, the data is judged to be in compliance with the requirements and is transferred to the next step.

[0021] Preferably, the specific process of performing integrity judgment on the data that has passed the evaluation is as follows:

[0022] For the texts in the evaluation data, the number of information points N that each text should contain is set in advance, and the number of information points n actually contained is obtained;

[0023] According to the formula T , calculate and obtain the coverage T of each text;

[0024] The average coverage of all texts in the data is calculated to obtain the average information point coverage of the entire data;

[0025] If the average information point coverage of the entire data is lower than the preset standard range, the data is marked as incomplete data, otherwise, the data is marked as complete data and transmitted to the next step.

[0026] Preferably, the specific process of completing the data after integrity judgment is as follows:

[0027] Determine whether the text data in the incomplete data is numerical data, and if so, determine whether the incomplete data is normally distributed or skewed distributed;

[0028] If the data are normally distributed or approximately normally distributed, the missing values ​​of the incomplete data are filled using mean interpolation;

[0029] If the distribution was skewed, median imputation was used to fill in the incomplete data;

[0030] If the incomplete data have an obvious mode, use mode imputation to fill in the incomplete data.

[0031] Preferably, if the text data in the incomplete data is product text data, it is determined whether the incomplete data is missing the length of the product description or the data sparsity;

[0032] If the missing data is the length of the product description, the median length is used as an interpolation;

[0033] If the missing data is due to data sparsity, determine whether the incomplete data has common descriptive phrases. If so, use the majority interpolation. If not, use the large model to generate the missing description.

[0034] Preferably, the specific process of preprocessing the commodity data in the data warehouse is as follows:

[0035] First, perform data cleaning to remove duplicate records;

[0036] The data was then converted into a format suitable for analysis, specifically by unifying the date format and standardizing the text encoding;

[0037] Scale the data in your data to a specific range or distribution.

[0038] Preferably, the identification of entities in the text and the identification of the relationship between entities are specifically as follows: firstly, entity recognition is performed, specifically: the text is divided into separate words or tokens, and the part of speech is marked for each word;

[0039] Identify specific types of entities using predefined patterns or regular expressions;

[0040] Use pre-trained models for entity recognition;

[0041] Then identify the relationship between entities, specifically:

[0042] Dependency parsing, analyzing the dependency relationship between words in a sentence and determining the grammatical structure between them;

[0043] Path analysis, which finds paths between entity pairs in the dependency syntax tree to determine the relationship between them;

[0044] Relation classification, using machine learning or deep learning methods to classify the relationships between pairs of entities.

[0045] Preferably, after identifying entities in the text and identifying the relationships between entities, it is also necessary to evaluate the performance of entity recognition and relationship extraction tasks, specifically in the following manner:

[0046] The accuracy score of the recognized entities in the text is obtained by dividing the number of correctly recognized entities by the number of correctly recognized entities and instances that are incorrectly recognized as entities;

[0047] The recall in identifying entities in text is obtained by dividing the number of correctly recognized entities by the sum of the number of instances where the entity actually exists but is incorrectly not recognized as an entity by the model and the number of correctly recognized entities.

[0048] Preferably, the specific working method of recommending products to users based on the user's data information and the similarity of products in the knowledge graph is as follows:

[0049] Obtain the vocabulary set G of the text in the user data, and obtain the vocabulary set H of each product text;

[0050] Obtain the intersection of vocabulary set G and vocabulary set H, and obtain the union of the intersection of vocabulary set G and vocabulary set H;

[0051] The similarity K between the user and the product is obtained by dividing the intersection of the vocabulary set G and the vocabulary set H by the union of the vocabulary set G and the vocabulary set H;

[0052] A minimum value of similarity K is set in advance, and all products with similarity K greater than the minimum value are recommended to the user's e-commerce platform.

[0053] Preferably, the specific working method of recommending products to users based on the user's data information and the similarity of products in the knowledge graph may also be:

[0054] Before segmenting the user's text, preprocess it to remove stop words and punctuation marks, perform stemming or word form restoration, etc., to improve the quality of the vocabulary set;

[0055] Assign a weight to each word in the vocabulary set;

[0056] Even if the similarity k is lower than the minimum value, the product will be pushed to the user's mobile terminal at intervals.

[0057] Beneficial effects: Through precise text similarity calculation, the products provided by the recommendation system are ensured to be highly relevant to the user's interests and preferences, thereby improving user satisfaction. The recommendations received by users are more in line with their actual needs, reducing the time for screening products and improving the user's browsing and shopping experience on the e-commerce platform. Recommendations with strong relevance are more likely to prompt users to make purchases, thereby improving the sales conversion rate of the e-commerce platform. By analyzing user preferences and purchasing behaviors, it can help e-commerce platforms more accurately predict hot-selling products, optimize inventory management and supply chain efficiency. The knowledge graph enables the recommendation system to provide more personalized services to meet the specific needs of different users. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0059] Application scenarios: In actual applications, when insurance companies or agents recommend pet insurance plans, they often make recommendations based on the specific circumstances of the pet and the needs of the customer. However, customers may have their own expectations of the insurance plan, and companies usually recommend it based on actual needs. This may lead to a lack of a mechanism to evaluate the rationality of customer expectations, thereby affecting customer experience and possibly leading to customer churn.

[0060] like Figure 1As shown: A method for constructing a knowledge graph based on a large model, comprising the following steps: It should be noted that the knowledge graph can integrate knowledge from different sources and discover new knowledge and relationships through reasoning mechanisms, thereby improving the accuracy and efficiency of information retrieval. In search engines, the knowledge graph can more accurately understand the user's query intent and context, and provide more relevant and rich search results. The knowledge graph enables machines to understand and answer more complex questions, and achieve more accurate information retrieval and knowledge discovery. By analyzing user behavior and preferences, the knowledge graph can help the recommendation system provide personalized and accurate recommendation results. The application of knowledge graphs in the field of artificial intelligence is extensive and increasingly important.

[0061] With the development of e-commerce platforms, in order to intelligently recommend suitable products to users, more and more e-commerce platforms will also build their own knowledge graphs. However, in actual use, since the product information on e-commerce platforms is often mainly titles and pictures, lacking detailed descriptive information, and the product data uploaded by merchants may be incomplete, especially in the LBS (location-based service) e-commerce scenario, the input of a large amount of product data is often accompanied by the problem of uneven information quality. In addition, although the product details page contains a lot of knowledge information, its quality and structure are different, making it difficult to extract knowledge from it;

[0062] The data dimensions involved in the commodity field are numerous, including general attributes (such as brand, specification, packaging, flavor, etc.) and category-specific attributes (such as fat content, whether it contains sugar, battery capacity, etc.). This leads to efficiency issues in data construction and requires processing and maintaining a large number of attribute dimensions, which is a challenge in the construction and update of knowledge graphs.

[0063] In daily life, users rely on a wealth of common sense knowledge to understand product information, such as inferring the true meaning of a product through a brief description. The construction of a knowledge graph needs to combine this common sense knowledge, which usually involves the semantic understanding of a large amount of unstructured text. At the same time, in professional fields such as medicine and personal care, the construction of graphs also needs to rely on professional knowledge. The accuracy of knowledge in these fields is extremely high, and efficient graph construction requires a combination of experts and algorithms.

[0064] Collect the internal and external data of the products on the platform and integrate the collected data into a unified data warehouse. The specific steps are as follows: It should be noted that in this embodiment, the internal data refers to the transaction records, user behavior logs and product information extracted from the e-commerce platform, and the external data refers to the data obtained from channels such as social media, review websites and third-party data suppliers;

[0065] It should also be noted that, in this embodiment, the collected data is text data that integrates internal data and external data, which is a plurality of data sets consisting of text, each of which represents transaction records, user behavior logs and product information or data obtained from channels such as social media, review websites and third-party data suppliers;

[0066] Regularly conduct quality assessments on the collected data, and screen based on the assessment results to remove data that lacks accuracy, consistency, and reliability. It should be noted that the regularity can be one day, one week, or one month. By conducting quality assessments on the data sets consisting of the collected text within one day, one week, or one month, it is possible to determine which of the transaction records, user behavior logs, and product information, or data obtained from channels such as social media, review sites, and third-party data suppliers meet the needs.

[0067] The integrity of the data that has passed the evaluation is further judged, and the data that has passed the judgment is transmitted to the next step;

[0068] The data that has been judged for integrity is completed, and then the completed data is transmitted to the next step;

[0069] Preprocess the commodity data in the data warehouse and extract the required feature information;

[0070] The extracted feature information is input into the large model to identify entities in the text and the relationships between entities;

[0071] Determine the classification system and the relationship between entities in the text ontology, store the classification system, entities and the relationship between entities in the database to build a knowledge graph; it should be noted that the nodes in the database represent entities and the edges represent relationships;

[0072] Obtain user data information and make recommendations to users based on the user data information and the similarity of products in the knowledge graph.

[0073] As an optional embodiment: the specific method of regularly evaluating the quality of the collected data and screening according to the evaluation results to remove data with insufficient accuracy, consistency and reliability is as follows:

[0074] The texts in the collected data are randomly checked to obtain the number of random check errors, and the number of text random check errors is divided by the total number of random checks to obtain the accuracy of the data;

[0075] Check whether each text in the collected data follows the predefined format, and obtain the consistency of the data by dividing the number of texts that meet the consistency requirements by the total number of texts in the data;

[0076] Obtain the number of stable texts in the data, divide the number of stable texts by the total number of texts in the data, and obtain the reliability of the data;

[0077] The accuracy, consistency and reliability of the acquired data are compared with the corresponding historical data average. If any of the data accuracy, consistency and reliability is lower than the corresponding historical data average, the data is judged to be unqualified. Otherwise, the data is judged to be qualified and transferred to the next step. It should be noted that by comprehensively considering the accuracy, consistency and reliability of the data, the data that does not meet the above requirements is screened to ensure the accuracy of the data input into the large model, which is conducive to the construction of the subsequent knowledge graph.

[0078] As an optional embodiment: the specific process of further judging the integrity of the data that has passed the evaluation is as follows:

[0079] For the texts in the evaluation data, the number of information points N that each text should contain is set in advance, and the number of information points n actually contained is obtained;

[0080] According to the formula T , calculate and obtain the coverage T of each text;

[0081] The average coverage of all texts in the data is calculated to obtain the average information point coverage of the entire data;

[0082] If the average information point coverage of the entire data is lower than the preset standard range, the data is marked as incomplete data. Otherwise, the data is marked as complete data and transmitted to the next step. It should be noted that, assuming that the product description text of an e-commerce platform needs to include the following information points: product name, brand, price, color, size, material, for the product description text in the data, if it contains brand, price and color, but lacks product name, size and material, the coverage of the text is 3 / 6 = 50%. If this ratio is lower than a preset standard (for example, 80%), the text is considered incomplete;

[0083] Through this method, the integrity of text data can be quantitatively evaluated and corresponding measures can be taken to improve data quality.

[0084] As an optional embodiment: the specific process of completing the data after the integrity judgment is as follows:

[0085] Determine whether the text data in the incomplete data is numerical data, and if so, determine whether the incomplete data is normally distributed or skewed distributed;

[0086] If the data is normally distributed or approximately normally distributed, mean interpolation is used to fill in the missing values ​​of incomplete data. It should be noted that when the missing fields are random and the field values ​​are relatively independent, mean interpolation is more appropriate. It should also be noted that mean interpolation is to fill in the values ​​using the mean of all numerical data.

[0087] If the distribution is skewed, use median interpolation to fill in the incomplete data; it should be noted that when there are many outliers in the data set or the data is unevenly distributed, median interpolation is a better choice; it should also be noted that median interpolation uses the median of all numerical data for filling;

[0088] If the incomplete data has an obvious mode, use mode interpolation to fill in the incomplete data. It should be noted that mode interpolation is appropriate when a value appears most frequently in the field and the missing data is random; it should also be noted that mode interpolation uses the mode of all numerical data for filling.

[0089] As an optional embodiment: if the text data in the incomplete data is commodity text data, it is determined whether the incomplete data is missing the length of the commodity description or the data sparsity;

[0090] If the length of the product description is missing, the median length is used as interpolation. It should be noted that the median length of the product category is used based on the product category. For missing values, the GPT generation model can be used to generate missing text based on non-missing text data.

[0091] If the missing data is data sparsity, determine whether the incomplete data has common descriptive phrases. If so, use the majority interpolation. If not, use the large model to generate the missing description. It should be noted that the GPT generation model can be used to generate the missing description using the large model.

[0092] As an optional embodiment: the specific process of preprocessing the commodity data in the data warehouse is as follows:

[0093] First, data cleaning is performed to remove duplicate records; it should be noted that data redundancy can be reduced and data accuracy and consistency can be improved.

[0094] The data is then converted into a format suitable for analysis, specifically by unifying the date format and standardizing the text encoding; it should be noted that ensuring the data format is consistent is convenient for subsequent processing and analysis;

[0095] Scaling the data in the data to a specific range or distribution, such as 0 to 1 or with a standard normal distribution, eliminates the influence of different dimensions and improves the convergence speed and performance of the algorithm.

[0096] As an optional embodiment: the identification of entities in the text and the identification of the relationship between entities are as follows: first, entity recognition is performed, specifically: the text is divided into separate words or tags, and the part of speech is marked for each word; it should be noted that the specific part of speech is noun, verb, etc.;

[0097] Use predefined patterns or regular expressions to identify specific types of entities; for example, dates and phone numbers.

[0098] Use pre-trained models for entity recognition; it should be noted that these models, such as the BERT model, can understand contextual information and improve recognition accuracy;

[0099] Then identify the relationship between entities, specifically:

[0100] Dependency syntactic analysis analyzes the dependency relationship between words in a sentence and determines the grammatical structure between them. It should be noted that dependency syntactic analysis is a text analysis method used to determine the dependency relationship between words in a sentence. Each word has a dependency relationship with another word in the sentence (usually the subject, predicate or object). This relationship is represented by a dependency syntactic tree. The main steps are: tokenization: decompose the sentence into separate words or tokens, determine the dependency relationship of each word, and usually use a tree structure to represent these relationships. Analyze the dependency tree to understand the structure of the sentence, for example, which word is the central word and which word is a modifier.

[0101] Path analysis, finding the path between entity pairs in the dependency syntax tree to determine the relationship between them; it should be noted that in the dependency syntax tree, the path between entity pairs can reveal the relationship between them. Path analysis involves identifying the dependency paths from one entity to another and inferring semantic relationships based on these paths. First, the dependency syntax tree of the sentence is constructed. In the dependency tree, the path from one entity to another may include multiple intermediate nodes, which represent modifiers or other grammatical elements. Specific vocabulary or patterns on the path may indicate specific relationships between entities.

[0102] Relation classification, using machine learning or deep learning methods to classify the relationships between entity pairs. It should be noted that relationship classification is the process of classifying the relationships between entity pairs into predefined categories. This usually involves machine learning or deep learning methods, where the model is trained to recognize different relationship types, extracting features from dependency paths and sentence structures, and using labeled relational data to train classification models such as support vector machines, random forests, or neural networks. The model is applied to entity pairs in new sentences to predict the relationship category between them;

[0103] It should be noted that through entity recognition and relationship extraction, the accuracy and consistency of the data set can be improved. Accurate entity and relationship information can improve the performance of machine learning models. Entity and relationship information can support more complex data analysis and queries, and help discover new knowledge patterns and connections from large amounts of text.

[0104] As an optional embodiment: After identifying entities in the text and identifying the relationships between entities, it is also necessary to evaluate the performance of entity recognition and relationship extraction tasks, specifically in the following manner:

[0105] The accuracy score of the entity in the recognized text is obtained by dividing the number of correctly recognized entities by the number of correctly recognized entities and the number of instances that are incorrectly recognized as entities. It should be noted that the accuracy score measures the proportion of positive categories predicted by the model that are actually positive, that is, the proportion of correctly predicted positive category samples to all predicted positive category samples;

[0106] The recall rate of entities in recognized text is obtained by dividing the number of correctly recognized entities by the sum of the number of instances of entities that actually exist but are not recognized as entities by the model and the number of correctly recognized entities. It should be noted that the recall rate measures the proportion of positive class samples correctly recognized by the model to all actual positive class samples, that is, the proportion of actual positive class samples captured. It is very important in scenarios where all positive class samples need to be discovered as much as possible, such as credit card fraud detection. A low recall rate may mean that many fraudulent behaviors are not detected.

[0107] The F1 score is obtained by multiplying 2 by the product of the accuracy score and the recall rate and then dividing it by the accuracy score and the recall rate. It should be noted that the F1 score is the harmonic mean of the accuracy and the recall rate, providing a single indicator to comprehensively consider the accuracy and the recall rate. The F1 score strikes a balance between the two, especially in the case of imbalanced categories. When you care about both accuracy and recall, the F1 score is a useful indicator. It is a very general performance metric that is suitable for most scenarios that need to evaluate the accuracy and completeness of the model;

[0108] Precision score, recall rate and F1 score are standard indicators for evaluating model performance, which can help you understand the possible performance of the model in practical applications. If the precision rate, recall rate and F1 score are lower than the standard value, the process of identifying entities in the text and identifying the relationship between entities should be optimized and improved.

[0109] As an optional embodiment: the specific working method of recommending users based on the user's data information and the similarity of commodities in the knowledge graph is as follows:

[0110] Obtain a vocabulary set G of the text in the user data, and obtain a vocabulary set H of each product text; it should be noted that each text is divided into separate words, and these words are combined into a set;

[0111] Obtain the intersection of vocabulary set G and vocabulary set H, and obtain the union of the intersection of vocabulary set G and vocabulary set H;

[0112] The similarity K between the user and the product is obtained by dividing the intersection of the vocabulary set G and the vocabulary set H by the union of the vocabulary set G and the vocabulary set H. It should be noted that the words that appear in both text vocabulary sets are found to form an intersection, and all different words that appear in the two text vocabulary sets, including repeated words, are found to form a union.

[0113] A minimum value of similarity K is set in advance, and all products with similarity K greater than the minimum value are recommended to the user's e-commerce platform. It should be noted that, assuming that the user's text is "I like product A and product B";

[0114] The text of the goods is Goods B and Goods C;

[0115] Segment the text into a vocabulary set G (product A, product B);

[0116] Vocabulary set H (product B, product C);

[0117] The intersection of vocabulary set G and vocabulary set H is (product B), and the union of vocabulary set G and vocabulary set H is (product A, product B, product C), which means that the similarity K is about 33.33%;

[0118] As an optional embodiment: the specific working method of recommending to users based on the user's data information and the similarity of commodities in the knowledge graph can also be:

[0119] Before segmenting the user's text, preprocess it to remove stop words and punctuation marks, perform stemming or word form restoration, etc., to improve the quality of the vocabulary set;

[0120] Assign a weight to each word in the vocabulary set; it should be noted that in actual use, the importance of each text word is different, so it is necessary to consider assigning a weight to each word in the vocabulary set instead of simply counting;

[0121] It should also be noted that the context of words in the text is also important, and more advanced NLP techniques may be needed to understand the semantics of words, not just whether they appear;

[0122] Even if the similarity k is lower than the minimum value, the product is pushed to the user's mobile terminal at intervals. It should be noted that even if the similarity is not high, it may be necessary to recommend some products that are slightly different from the user's interests to increase the diversity of recommendations.

[0123] How it works

[0124] Through precise text similarity calculation, we ensure that the products provided by the recommendation system are highly relevant to the user's interests and preferences, thereby improving user satisfaction. The recommendations received by users are more in line with their actual needs, reducing the time for screening products and improving the user's browsing and shopping experience on the e-commerce platform. Recommendations with strong relevance are more likely to prompt users to purchase, thereby increasing the sales conversion rate of the e-commerce platform. By analyzing user preferences and purchasing behaviors, it can help e-commerce platforms more accurately predict hot-selling products, optimize inventory management and supply chain efficiency. The knowledge graph enables the recommendation system to provide more personalized services to meet the specific needs of different users.

[0125] The above are only preferred implementations of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical staff in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of this template.

Claims

1. A method for constructing a knowledge graph based on a large model, characterized in that: The following steps are involved: Collect the internal and external data of the platform's products and integrate the collected data into a unified data warehouse. The specific steps are as follows: Regularly conduct quality assessments on the collected data and screen them based on the assessment results to remove data that lacks accuracy, consistency, and reliability; The integrity of the data that has passed the evaluation is further judged, and the data that has passed the judgment is transmitted to the next step; The data that has been judged for integrity is completed, and then the completed data is transmitted to the next step; Preprocess the commodity data in the data warehouse and extract the required feature information; The extracted feature information is input into the large model to identify entities in the text and the relationships between entities; Determine the classification system and the relationship between entities in the text ontology, store the classification system, entities and the relationship between entities in the database to build a knowledge graph; Obtain user data information and make recommendations to users based on the user data information and the similarity of products in the knowledge graph.

2. According to the method for constructing a knowledge graph based on a large model according to claim 1, it is characterized in that: The specific method of regularly conducting quality assessment on the collected data and screening based on the assessment results to remove data with insufficient accuracy, consistency and reliability is as follows: The texts in the collected data are randomly checked to obtain the number of random check errors, and the number of text random check errors is divided by the total number of random checks to obtain the accuracy of the data; Check whether each text in the collected data follows the predefined format, and obtain the consistency of the data by dividing the number of texts that meet the consistency requirements by the total number of texts in the data; Obtain the number of stable texts in the data, divide the number of stable texts by the total number of texts in the data, and obtain the reliability of the data; The accuracy, consistency and reliability of the acquired data are compared with the corresponding historical data averages. If any one of the accuracy, consistency and reliability of the data is lower than the corresponding historical data average, the data is judged to be not in compliance with the requirements. Otherwise, the data is judged to be in compliance with the requirements and is transferred to the next step.

3. According to the method for constructing a knowledge graph based on a large model according to claim 1, it is characterized in that: The specific process of conducting integrity judgment on the data that has passed the evaluation is as follows: For the texts in the evaluation data, the number of information points N that each text should contain is set in advance, and the number of information points n actually contained is obtained; According to the formula T , calculate and obtain the coverage T of each text; The average coverage of all texts in the data is calculated to obtain the average information point coverage of the entire data; If the average information point coverage of the entire data is lower than the preset standard range, the data is marked as incomplete data, otherwise, the data is marked as complete data and transmitted to the next step.

4. A method for constructing a knowledge graph based on a large model according to claim 1, characterized in that: The specific process of completing the data after integrity judgment is as follows: Determine whether the text data in the incomplete data is numerical data, and if so, determine whether the incomplete data is normally distributed or skewed distributed; If the data are normally distributed or approximately normally distributed, the missing values ​​of the incomplete data are filled using mean interpolation; If the distribution was skewed, median imputation was used to fill in the incomplete data; If the incomplete data have an obvious mode, use mode imputation to fill in the incomplete data.

5. A method for constructing a knowledge graph based on a large model according to claim 4, characterized in that: If the text data in the incomplete data is product text data, determine whether the missing part of the incomplete data is the length of the product description or data sparsity; If the missing data is the length of the product description, the median length is used as an interpolation; If the missing data is due to data sparsity, determine whether the incomplete data has common descriptive phrases. If so, use the majority interpolation. If not, use the large model to generate the missing description.

6. A method for constructing a knowledge graph based on a large model according to claim 1, characterized in that: The specific process of preprocessing the commodity data in the data warehouse is as follows: First, perform data cleaning to remove duplicate records; The data was then converted into a format suitable for analysis, specifically by unifying the date format and standardizing the text encoding; Scale the data in your data to a specific range or distribution.

7. A method for constructing a knowledge graph based on a large model according to claim 6, characterized in that: The specific steps of identifying entities in the text and identifying the relationship between entities are as follows: first, entity recognition is performed, specifically: the text is divided into separate words or tokens, and the part of speech is marked for each word; Identify specific types of entities using predefined patterns or regular expressions; Use pre-trained models for entity recognition; Then identify the relationship between entities, specifically: Dependency parsing, analyzing the dependency relationship between words in a sentence and determining the grammatical structure between them; Path analysis, which finds paths between entity pairs in the dependency syntax tree to determine the relationship between them; Relation classification, using machine learning or deep learning methods to classify the relationships between pairs of entities.

8. A method for constructing a knowledge graph based on a large model according to claim 7, characterized in that: After identifying entities in the text and the relationships between entities, it is necessary to evaluate the performance of entity recognition and relationship extraction tasks. The specific method is as follows: The accuracy score of the recognized entities in the text is obtained by dividing the number of correctly recognized entities by the number of correctly recognized entities and instances that are incorrectly recognized as entities; The recall in identifying entities in text is obtained by dividing the number of correctly recognized entities by the sum of the number of instances where the entity actually exists but is incorrectly not recognized as an entity by the model and the number of correctly recognized entities.

9. The method for constructing a knowledge graph based on a large model according to claim 1, characterized in that: The specific working method of recommending users based on the user's data information and the similarity of products in the knowledge graph is as follows: Obtain the vocabulary set G of the text in the user data, and obtain the vocabulary set H of each product text; Obtain the intersection of vocabulary set G and vocabulary set H, and obtain the union of the intersection of vocabulary set G and vocabulary set H; The similarity K between the user and the product is obtained by dividing the intersection of the vocabulary set G and the vocabulary set H by the union of the vocabulary set G and the vocabulary set H; A minimum value of similarity K is set in advance, and all products with similarity K greater than the minimum value are recommended to the user's e-commerce platform.

10. A method for constructing a knowledge graph based on a large model according to claim 9, characterized in that: The specific working method of recommending users based on the user's data information and the similarity of commodities in the knowledge graph may also be: Before segmenting the user's text, preprocess it to remove stop words and punctuation marks, perform stemming or word form restoration, etc., to improve the quality of the vocabulary set; Assign a weight to each word in the vocabulary set; Even if the similarity k is lower than the minimum value, the product will be pushed to the user's mobile terminal at intervals.

Citation Information

Patent Citations

  • Recommendation method and system based on commodity knowledge graph feature learning

    CN111369318A

  • Method for constructing knowledge graph in medical field

    CN116541472A

  • Knowledge graph analysis method based on non-privacy data association

    CN116955648A

  • Knowledge graph-based unstructured data retrieval method and system

    CN117807191A

  • AI-based rating process digital management method

    CN117875706A