An e-commerce commodity normalization method and system based on ontology words and vector retrieval

Abstract

The application discloses an e-commerce commodity normalization method and system based on ontology words and vector retrieval. The e-commerce commodity title is obtained, and the ontology words are extracted after hierarchical cleaning. The e-commerce commodity title and the ontology words are encoded to obtain a query vector, and approximate nearest neighbor retrieval is performed in a standard dictionary vector library according to the query vector to obtain an initial candidate set and a similarity score of each candidate item. Similarity distribution evaluation is performed according to the similarity score of each candidate item, and the initial candidate set is adaptively adjusted according to the evaluation result to obtain a final candidate set. The candidate set and the e-commerce commodity title are input into a local large language model to obtain an optimal candidate item and a confidence score. The standard product name, the category path and the candidate list of the normalized commodity are output according to the optimal candidate item and the confidence score. The e-commerce commodity normalization method and system based on ontology words and vector retrieval provided by the application realizes accurate normalization of e-commerce commodities.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, specifically to an e-commerce product normalization method and system based on ontology terms and vector retrieval. Background Technology

[0002] With the continuous development of the economy, e-commerce platforms have gradually become a major channel for consumers to shop. An e-commerce platform refers to an online marketplace that connects buyers and sellers through internet technology. It provides a virtual trading space where users can browse and purchase goods or services. E-commerce platforms typically include multiple functions such as product display, search, payment, logistics, and reviews, providing a one-stop service for consumers and merchants.

[0003] In e-commerce platforms, product titles are a key element in attracting consumer attention. However, with increasing competition on these platforms, merchants often use lengthy titles containing a variety of information to boost product exposure and clicks. These titles not only include a basic description of the product but also frequently incorporate promotional information, holiday-related modifiers, or regional labels. For example, "(Qixi Festival Exclusive) 2025 First Spring Unnamed King Ancient Tree Pu-erh Raw Tea Loose Material 70g / box" and "2025 First Spring Selected Material Manxi Longdai Ancient Tree Pu-erh Sun-dried Red Loose Material 50g / box"—although these two names differ significantly, their core term is "Pu-erh tea."

[0004] While lengthy titles containing multiple pieces of information can boost short-term sales, they can also lead to overly complex information, negatively impacting user experience and platform search efficiency. Therefore, how to normalize product titles to ensure concise and clear information while maintaining effective marketing has become a significant technical challenge for e-commerce platforms in optimizing product display and improving user experience. However, existing technologies for product normalization have the following problems:

[0005] 1. Limitations of the rule-based approach: It relies on manual rules or keyword matching, making it difficult to handle complex and ever-changing product titles;

[0006] 2. Large models are difficult to match directly: The standard dictionary is large (up to hundreds of thousands of entries), and directly inputting it into the model will exceed the maximum context length;

[0007] 3. Lack of tiered processing: Existing technologies are mostly "one-step matching" and lack phased screening, resulting in insufficient accuracy;

[0008] 4. Inadequate candidate set processing: The fixed Top-K candidate set cannot adapt to different query complexities, resulting in insufficient recall or excessive redundancy.

[0009] In summary, existing technologies struggle to accurately normalize e-commerce products due to issues such as redundant and complex product titles, excessively large dictionary sizes, difficulties in direct matching, and unreasonable candidate set processing. Summary of the Invention

[0010] To address this issue, this application provides a method and system for normalizing e-commerce products based on ontology terms and vector retrieval, in order to solve the problem that existing technologies struggle to accurately normalize e-commerce products.

[0011] To achieve the above objectives, this application provides the following technical solution:

[0012] Firstly, a normalization method for e-commerce products based on ontology terms and vector retrieval includes:

[0013] Step 1: Obtain the original e-commerce product titles and perform layered cleaning to obtain the core product titles;

[0014] Step 2: Predict the part-of-speech tag and domain label of the core title of the product based on the sequence labeling model to obtain a candidate word set;

[0015] Step 3: Align the candidate word set with the predefined ontology dictionary, and obtain standardized ontology words through similarity measurement and synonym cluster mapping mechanism;

[0016] Step 4: Encode the original e-commerce product title and the ontology terms to obtain the query vector;

[0017] Step 5: Perform approximate nearest neighbor retrieval in a pre-built standard dictionary vector library based on the query vector to obtain an initial candidate set, and return the similarity score of each candidate item in the initial candidate set;

[0018] Step 6: Evaluate the similarity distribution based on the similarity score of each candidate item in the initial candidate set, and adaptively adjust the initial candidate set according to the evaluation results to obtain the final candidate set;

[0019] Step 7: Input the final candidate set and the original e-commerce product titles into the local large language model to obtain the optimal candidate items and confidence scores;

[0020] Step 8: Output the normalized standard product name, category path, confidence score, and candidate list based on the best candidate item and confidence score.

[0021] Optionally, in step 1, the layered cleaning includes: removing holiday modifiers, marketing terms, and packaging modification information from the original e-commerce product title based on regularization rules and a stop word list, and extracting brand entities and capacity specification entities using a named entity recognition model to obtain the core product title.

[0022] Optionally, in step 5, each entry in the standard dictionary vector library is a multimodal vector formed by extracting semantic vectors through a text encoder and concatenating other attribute fields in the entry into vectors.

[0023] Optionally, in step 5, when performing approximate nearest neighbor retrieval in a pre-built standard dictionary vector library based on the query vector, an IVF-PQ index structure is used.

[0024] Optionally, step 6 specifically includes:

[0025] Statistical analysis is performed on the similarity scores of each candidate item in the initial candidate set to determine the degree of concentration or dispersion of the distribution. The elbow rule or threshold judgment is used to determine the optimal cutoff point. When the distribution is relatively dispersed and the score difference is large, the number of initial candidates is increased. When the distribution is relatively concentrated and the scores are close, the initial candidate set is shrunk.

[0026] Optionally, in step 7, the large language model needs to determine in stages whether the candidate entries are consistent with the ontology words, whether the category path of the candidate entries is consistent with the context constraints, and whether the key attributes conflict.

[0027] Optionally, in step 7, if the confidence score is lower than the threshold, it is marked as requiring manual review.

[0028] Secondly, an e-commerce product normalization system based on ontology terms and vector retrieval includes:

[0029] The layered cleaning module is used to obtain the original e-commerce product titles and perform layered cleaning to obtain the core product titles;

[0030] The tag prediction module is used to predict the part-of-speech and domain tags of the core title of the product based on the sequence labeling model, and obtain a set of candidate words;

[0031] The ontology word extraction module is used to align the candidate word set with a predefined ontology dictionary and obtain standardized ontology words through similarity measurement and synonym cluster mapping mechanism;

[0032] The encoding module is used to encode the original e-commerce product title and the ontology words to obtain a query vector;

[0033] The vector retrieval module is used to perform approximate nearest neighbor retrieval in a pre-built standard dictionary vector library based on the query vector, obtain an initial candidate set, and return the similarity score of each candidate item in the initial candidate set;

[0034] The candidate set adjustment module is used to evaluate the similarity distribution based on the similarity score of each candidate item in the initial candidate set, and adaptively adjust the initial candidate set according to the evaluation results to obtain the final candidate set.

[0035] The precision judgment module is used to input the final candidate set and the original e-commerce product title into the local large language model, and obtain the optimal candidate items and confidence scores;

[0036] The output module is used to output the normalized standard product name, category path, confidence score, and candidate list based on the best candidate item and confidence score.

[0037] Thirdly, a computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the e-commerce product normalization method based on ontology and vector retrieval.

[0038] Fourthly, a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the e-commerce product normalization method based on ontology and vector retrieval.

[0039] Compared with the prior art, this application has at least the following beneficial effects:

[0040] This application provides an e-commerce product normalization method based on ontology and vector retrieval. The method involves: extracting ontology terms from the obtained original e-commerce product titles after hierarchical cleaning; encoding the original e-commerce product titles and ontology terms to obtain query vectors; performing approximate nearest neighbor retrieval in a pre-built standard dictionary vector library based on the query vectors to obtain an initial candidate set and a similarity score for each candidate item; evaluating the similarity distribution based on the similarity scores of each candidate item in the initial candidate set, and adaptively adjusting the initial candidate set based on the evaluation results to obtain the final candidate set; inputting the final candidate set and the original e-commerce product titles into a local large language model to obtain the optimal candidate items and confidence scores; and outputting the normalized standard product name, category path, confidence score, and candidate list based on the optimal candidate items and confidence scores. This e-commerce product normalization method based on ontology and vector retrieval achieves accurate normalization of e-commerce products, solving the problem of inaccurate normalization of e-commerce products caused by redundant and complex product titles, excessively large dictionary sizes, and difficulties in direct matching in existing technologies. Attached Figure Description

[0041] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0042] Figure 1 A flowchart of an e-commerce product normalization method based on ontology terms and vector retrieval provided in Embodiment 1 of this application;

[0043] Figure 2 A schematic diagram of the structure of an e-commerce product normalization method based on ontology and vector retrieval provided in Embodiment 1 of this application;

[0044] Figure 3 This is a flowchart of the ontology extraction process provided in Embodiment 1 of this application;

[0045] Figure 4 This is a flowchart of the vector recall process provided in Embodiment 1 of this application;

[0046] Figure 5 The flowchart for adaptive candidate set adjustment provided in Embodiment 1 of this application;

[0047] Figure 6 This is a schematic diagram of an e-commerce product normalization device based on ontology and vector retrieval, provided in Embodiment 2 of this application. Detailed Implementation

[0048] The present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] In the description of this application: unless otherwise stated, "a plurality of" means two or more. The terms "first," "second," "third," etc., in this application are intended to distinguish the objects referred to and do not have any special meaning in terms of technical connotation (e.g., they should not be construed as an emphasis on importance or order). Expressions such as "including," "comprising," and "having" also mean "not limited to" (certain units, components, materials, steps, etc.).

[0050] The terms used in this application, such as "upper," "lower," "left," "right," and "middle," are generally used to indicate the general relative positional relationship for the purpose of intuitive understanding by referring to the accompanying drawings, and are not absolute limitations on the positional relationship in the actual product.

[0051] Example 1

[0052] Please see Figure 1 and Figure 2 This embodiment provides an e-commerce product normalization method based on ontology terms and vector retrieval, applicable to product name standardization, category mapping, and catalog normalization on e-commerce platforms. The method includes:

[0053] S1: Obtain the original e-commerce product titles and perform layered cleaning to obtain the core product titles;

[0054] In this step, the layered cleaning of the original e-commerce product titles specifically includes: removing holiday modifiers, marketing terms, and packaging modification information from the original e-commerce product titles based on regularization rules and stop word lists, and using the Named Entity Recognition (NER) model to extract brand entities and capacity specification entities, retaining core noun phrases to obtain the core product title.

[0055] S2: Predict the part-of-speech and domain tags of the core product title based on the sequence labeling model to obtain a set of candidate words;

[0056] Specifically, this step uses a sequence labeling model (e.g., Transformer-CRF) to predict the part-of-speech and domain tags of the core product title, thereby obtaining a set of candidate words.

[0057] S3: Align the candidate word set with the predefined ontology dictionary, and obtain standardized ontology words through similarity measurement and synonym cluster mapping mechanism;

[0058] Specifically, this step aligns the obtained candidate word set with a predefined ontology dictionary and obtains standardized ontology words O through similarity measurement and synonym cluster mapping mechanisms, such as "Pu'er tea", "basketball shoes", and "laundry detergent", etc. Figure 3 As shown, steps S1 to S3 above achieve the extraction of ontology words.

[0059] Unlike existing technologies that rely solely on keyword matching, this embodiment employs a dual-channel extraction mechanism combining rule templates (i.e., an ontology dictionary) and neural sequence annotation, and introduces confidence gating and synonym cluster mapping to achieve robust extraction of ontology terms from complex titles. This combined approach enables accurate extraction of core ontology information even in scenarios with severe marketing interference or disordered title word order, significantly improving the robustness of the normalization process.

[0060] S4: Encode the original e-commerce product titles and ontology terms to obtain the query vector;

[0061] For details, please refer to Figure 4 This step encodes the original e-commerce product title T and the extracted ontology term O to obtain the query vector E. T E O .

[0062] S5: Perform approximate nearest neighbor retrieval in a pre-built standard dictionary vector library based on the query vector to obtain an initial candidate set, and return the similarity score of each candidate item in the initial candidate set;

[0063] Specifically, this step performs an approximate nearest neighbor search (ANN, using IVF-PQ index) in the standard dictionary vector library to obtain an initial candidate set C0, and returns the similarity score S for each candidate item in the initial candidate set. i .

[0064] In this step, the standard dictionary vector library is pre-built. When building the standard dictionary vector library, the standard dictionary entries D={d1,d2,…,d…} are... n The data is stored in a vectorized manner, and each entry simultaneously generates a multimodal representation, including: the semantic vector E output by the text encoder. d The semantic vector of each entry is constructed by embedding the semantic vector and other attribute fields (category path, specifications, etc.). In other words, in the standard dictionary vector library, each entry is a multimodal vector formed by extracting semantic vectors through a text encoder and then concatenating the other attribute fields of the entry into vectors.

[0065] In this embodiment, steps S4 and S5 achieve vector recall. Unlike existing methods that rely solely on a single perspective (e.g., the complete title), this embodiment employs a multi-perspective recall mechanism combining title vectors and ontology word vectors. This reduces noise interference while enhancing coverage of long-tail categories and diverse expressions. This "multi-vector joint retrieval" strategy reduces the probability of erroneous candidates entering subsequent discrimination during the recall stage, thereby improving overall accuracy.

[0066] S6: Analyze the similarity distribution based on the similarity score of each candidate item in the initial candidate set, and adaptively adjust the initial candidate set according to the evaluation results to obtain the final candidate set;

[0067] Specifically, this step performs statistical analysis on the similarity score distribution of each candidate item in the initial candidate set. It uses indicators such as variance and range to determine the degree of centralization or dispersion of the distribution, and employs the elbow rule or threshold to determine the optimal cutoff point. When the distribution is sparse (discrete) and the scores differ significantly, the size of the initial candidate set is appropriately increased to ensure recall; when the distribution is concentrated and the scores are similar, the range of the initial candidate set is narrowed to improve the purity of the results. Finally, the optimal candidate set is adaptively determined based on the distribution characteristics, achieving a balance between recall and precision. Please refer to [link to relevant documentation]. Figure 5 .

[0068] In this embodiment, step S6 achieves adaptive candidate set adjustment. Its dynamic K-adjustment mechanism based on similarity distribution features overcomes the limitation of using a fixed Top-K in existing technologies. It can adaptively adjust the candidate size according to retrieval uncertainty, thus balancing recall and precision in both high-confidence and fuzzy scenarios. Especially in scenarios involving long-tail products and semantically similar categories (such as "Pu'er tea" and "black tea"), this method can effectively improve the accuracy of subsequent precise judgment.

[0069] S7: Input the final candidate set and the original e-commerce product titles into the local large language model to obtain the optimal candidate items and confidence scores;

[0070] Specifically, in this step, the large language model needs to determine the consistency between candidate entries and ontology words, whether the category path of candidate entries is consistent with the context constraints, and whether key attributes conflict in stages, and finally output the optimal candidate entry. With confidence scores .

[0071] In this embodiment, step S7 above realizes the precision judgment normalization. This step adopts a combination of "candidate set constraint + large model reasoning" in the normalization judgment: first, the candidate size is controlled by vector recall, and then the semantic reasoning of the large model is used to complete the final precision judgment.

[0072] S8: Output the normalized standard product name, category path, confidence score, and candidate list based on the best candidate item and confidence score.

[0073] Specifically, this step ultimately outputs a standardized product name. and the corresponding category path It also provides a confidence score for the normalized results. And candidate list comparison information (Top-K items and similarity / rejection reasons). If If the result is below the threshold θ, it is marked as "requires manual review" to support manual result writing back and continuous model optimization.

[0074] This embodiment not only outputs the final normalized result but also provides confidence quantification and candidate control interpretation, ensuring the traceability and auditability of the results. Unlike existing methods that only output a single label, this embodiment features a dual mode of "automatic adoption + manual confirmation," guaranteeing the reliable implementation of normalization in large-scale e-commerce environments.

[0075] To more clearly illustrate the method provided in this embodiment, the following example of Pu'er tea commodity normalization will be used to further introduce the method provided in this embodiment:

[0076] S1: Enter the product title: "(Qixi Festival Exclusive) 2025 First Spring Unnamed King Ancient Tree Pu'er Raw Tea Loose Material 70g / box";

[0077] S2: Ontology extraction: Resulting in "Pu'er tea";

[0078] S3: Vector Recall: Search the dictionary database, and the Top-K candidate results include "Pu'er tea", "black tea", and "green tea";

[0079] S4: Adaptive candidate adjustment: Because the similarity between "Pu'er tea" and "black tea" in the candidate set is close, K is automatically expanded to 5;

[0080] S5: Fine-grained normalization: The large model determines the result as "Pu'er tea" in the candidate set and outputs a confidence score of 0.92;

[0081] S6: Final output: The product is normalized to the standard name "Pu'er Tea", and the category path is "Beverages > Tea > Pu'er Tea".

[0082] The e-commerce product normalization method based on ontology and vector retrieval provided in this embodiment has the following advantages:

[0083] 1. Two-stage matching architecture: Combining "vector recall + model precision judgment" solves the problem that direct matching of large-scale dictionaries is not feasible.

[0084] 2. Ontology system: Improve matching accuracy and reduce redundant interference information by extracting ontology terms.

[0085] 3. Adaptive candidate set: Dynamically adjust the size of the Top-K candidate set to improve recall efficiency and accuracy.

[0086] 4. Fine-grained judgment strategy: Introduce a local large model for secondary fine-grained judgment, which significantly improves the reliability of the normalization results.

[0087] 5. High scalability: Supports dynamic dictionary updates, adapting to large-scale product management on e-commerce platforms.

[0088] In summary, the e-commerce product normalization method based on ontology terms and vector retrieval provided in this embodiment achieves efficient, accurate, and scalable product normalization, solving the problem of difficulty in accurately normalizing e-commerce products caused by redundant and complex product titles, excessively large dictionary size, and difficulty in direct matching in existing technologies.

[0089] Example 2

[0090] Please see Figure 6 This embodiment provides an e-commerce product normalization system based on ontology terms and vector retrieval, including:

[0091] The layered cleaning module is used to obtain the original e-commerce product titles and perform layered cleaning to obtain the core product titles;

[0092] The tag prediction module is used to predict the part-of-speech and domain tags of the core title of the product based on the sequence labeling model, and obtain a set of candidate words;

[0093] The ontology word extraction module is used to align the candidate word set with a predefined ontology dictionary and obtain standardized ontology words through similarity measurement and synonym cluster mapping mechanism;

[0094] The encoding module is used to encode the original e-commerce product title and the ontology words to obtain a query vector;

[0095] The vector retrieval module is used to perform approximate nearest neighbor retrieval in a pre-built standard dictionary vector library based on the query vector, obtain an initial candidate set, and return the similarity score of each candidate item in the initial candidate set;

[0096] The candidate set adjustment module is used to evaluate the similarity distribution based on the similarity score of each candidate item in the initial candidate set, and adaptively adjust the initial candidate set according to the evaluation results to obtain the final candidate set.

[0097] The precision judgment module is used to input the final candidate set and the original e-commerce product title into the local large language model, and obtain the optimal candidate items and confidence scores;

[0098] The output module is used to output the normalized standard product name, category path, confidence score, and candidate list based on the best candidate item and confidence score.

[0099] For details on the implementation of each module in an e-commerce product normalization system based on ontology terms and vector retrieval, please refer to the above description of the limitations of an e-commerce product normalization method based on ontology terms and vector retrieval, which will not be repeated here.

[0100] Example 3

[0101] This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of an e-commerce product normalization method based on ontology and vector retrieval.

[0102] Example 4

[0103] This embodiment provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps of an e-commerce product normalization method based on ontology and vector retrieval.

[0104] The technical features of the above embodiments can be combined in any way (as long as there is no contradiction in the combination of these technical features). For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; these embodiments not explicitly written should also be considered to be within the scope of this specification.

Claims

1. A method for normalizing e-commerce products based on ontology terms and vector retrieval, characterized in that, include: Step 1: Obtain the original e-commerce product titles and perform layered cleaning to obtain the core product titles; Step 2: Predict the part-of-speech tag and domain label of the core title of the product based on the sequence labeling model to obtain a candidate word set; Step 3: Align the candidate word set with the predefined ontology dictionary, and obtain standardized ontology words through similarity measurement and synonym cluster mapping mechanism; Step 4: Encode the original e-commerce product title and the ontology terms to obtain the query vector; Step 5: Perform approximate nearest neighbor retrieval in a pre-built standard dictionary vector library based on the query vector to obtain an initial candidate set, and return the similarity score of each candidate entry in the initial candidate set; the IVF-PQ index structure is used when performing approximate nearest neighbor retrieval in the pre-built standard dictionary vector library based on the query vector; in the standard dictionary vector library, each entry is a multimodal vector formed by extracting semantic vectors through a text encoder and concatenating other attribute fields in the entry into vectors; Step 6: Evaluate the similarity distribution based on the similarity score of each candidate item in the initial candidate set, and adaptively adjust the initial candidate set according to the evaluation results to obtain the final candidate set; Step 7: Input the final candidate set and the original e-commerce product titles into the local large language model to obtain the optimal candidate items and confidence scores; Step 8: Output the normalized standard product name, category path, confidence score, and candidate list based on the best candidate item and confidence score.

2. The e-commerce product normalization method based on ontology and vector retrieval according to claim 1, characterized in that, In step 1, the layered cleaning includes: removing holiday modifiers, marketing terms and packaging modification information from the original e-commerce product title based on regularization rules and a stop word list, and extracting brand entity and capacity specification entity using a named entity recognition model to obtain the core title of the product.

3. The e-commerce product normalization method based on ontology and vector retrieval according to claim 1, characterized in that, Step 6 specifically includes: Statistical analysis is performed on the similarity scores of each candidate item in the initial candidate set to determine the degree of centralization or dispersion of the distribution. The elbow rule or threshold judgment is used to determine the optimal cutoff point. When the distribution is discrete and the scores differ greatly, the number of initial candidates is increased. When the distribution is concentrated and the scores are close, the initial candidate set is shrunk.

4. The e-commerce product normalization method based on ontology and vector retrieval according to claim 1, characterized in that, In step 7, the large language model needs to determine in stages whether the candidate entries are consistent with the ontology words, whether the category path of the candidate entries is consistent with the context constraints, and whether the key attributes conflict.

5. The e-commerce product normalization method based on ontology and vector retrieval according to claim 1, characterized in that, In step 7, if the confidence score is lower than the threshold, it is marked as requiring manual review.

6. An e-commerce product normalization system based on ontology and vector retrieval, characterized in that, include: The layered cleaning module is used to obtain the original e-commerce product titles and perform layered cleaning to obtain the core product titles; The tag prediction module is used to predict the part-of-speech and domain tags of the core title of the product based on the sequence labeling model, and obtain a set of candidate words; The ontology word extraction module is used to align the candidate word set with a predefined ontology dictionary and obtain standardized ontology words through similarity measurement and synonym cluster mapping mechanism; The encoding module is used to encode the original e-commerce product title and the ontology words to obtain a query vector; The vector retrieval module is used to perform approximate nearest neighbor retrieval in a pre-built standard dictionary vector library based on the query vector, obtain an initial candidate set, and return the similarity score of each candidate entry in the initial candidate set; the IVF-PQ index structure is used when performing approximate nearest neighbor retrieval in the pre-built standard dictionary vector library based on the query vector; in the standard dictionary vector library, each entry is a multimodal vector formed by extracting semantic vectors through a text encoder and concatenating other attribute fields in the entry into vectors; The candidate set adjustment module is used to evaluate the similarity distribution based on the similarity score of each candidate item in the initial candidate set, and adaptively adjust the initial candidate set according to the evaluation results to obtain the final candidate set. The precision judgment module is used to input the final candidate set and the original e-commerce product title into the local large language model, and obtain the optimal candidate items and confidence scores; The output module is used to output the normalized standard product name, category path, confidence score, and candidate list based on the best candidate item and confidence score.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images