E-commerce commodity normalization method and system based on ontology word and vector retrieval

Abstract

The invention discloses an e-commerce commodity normalization method and system based on ontology words and vector retrieval, and the method comprises the steps: obtaining an e-commerce commodity title, carrying out the layering cleaning, and extracting ontology words; encoding an e-commerce commodity title and an ontology word to obtain a query vector, and performing approximate nearest neighbor retrieval 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; performing similarity distribution evaluation according to the similarity score of each candidate item, and adaptively adjusting the initial candidate set according to an evaluation result to obtain a final candidate set; inputting the candidate set and the e-commerce commodity title into a local large language model to obtain an optimal candidate entry and a confidence score; and outputting a normalized standard product name, a category path and a candidate list of the commodity according to the optimal candidate item and the confidence score. According to the e-commerce commodity normalization method and system based on ontology word and vector retrieval provided by the invention, accurate normalization of e-commerce commodities is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, in particular to an e-commerce commodity normalization method and system based on ontology words and vector retrieval. BACKGROUND

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

[0003] In e-commerce platforms, product titles are one of the key elements that attract consumers' attention. However, as competition on the platform intensifies, merchants often use lengthy titles that contain multiple information to increase the exposure and click-through rate of their products. These titles not only include basic descriptions of the product, but also often incorporate promotional information, holiday modifiers, or regional annotations. For example, “(Qixi Special) 2025 Head Spring Unknown King Ancient Tree Pu'er Raw Tea Loose Material 70g / Box” and “2025 Head Spring Selected Material Manxilong Dai Ancient Tree Pu'er Sun-dried Red Loose Material 50g / Box”. Although the two names differ greatly, their ontology words are both “Pu'er Tea”.

[0004] Using lengthy titles that contain multiple information may help short-term sales, but it can also make the information too complex, affecting user experience and search efficiency on the platform. Therefore, how to normalize product titles to ensure concise and clear information while maintaining effective marketing has become an important technical challenge for e-commerce platforms to optimize product display and improve user experience. However, existing technologies have the following problems in product normalization:

[0005] 1. Limitations of rule-based methods: relying on manual rules or keyword matching, making it difficult to deal with complex and variable product titles;

[0006] 2. Difficulty in direct matching with large models: the size of the standard dictionary is large (up to hundreds of thousands of entries), and direct input into the model will exceed the maximum context length;

[0007] 3. Lack of hierarchical processing: existing technologies are mostly “one-step matching”, lacking phased screening, resulting in insufficient accuracy;

[0008] 4. Unreasonable handling of candidate sets: Top-K candidate sets are fixed and cannot adapt to different query complexities, resulting in insufficient recall or excessive redundancy.

[0009] In summary, the prior art causes the e-commerce goods to be difficult to accurately realize normalization due to redundant and complex commodity titles, excessively large dictionary size, direct matching difficulty, and unreasonable candidate set processing. SUMMARY

[0010] To this end, the present application provides an e-commerce goods normalization method and system based on ontology words and vector retrieval to solve the problem that the e-commerce goods are difficult to accurately realize normalization in the prior art.

[0011] To achieve the above object, the present application provides the following technical solutions:

[0012] In a first aspect, an e-commerce goods normalization method based on ontology words and vector retrieval includes:

[0013] Step 1: Obtain an original e-commerce goods title and perform hierarchical cleaning to obtain a goods core title;

[0014] Step 2: Predict the part of speech and domain label of the goods core title according to a sequence labeling model to obtain a candidate word set;

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

[0016] Step 4: Encode the original e-commerce goods title and the ontology words and obtain a query vector;

[0017] Step 5: Perform approximate nearest neighbor retrieval in a pre-constructed standard dictionary vector library according to 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: Perform similarity distribution evaluation according to the similarity score of each candidate item in the initial candidate set, and adaptively adjust the initial candidate set according to the evaluation result to obtain a final candidate set;

[0019] Step 7: Input the final candidate set and the original e-commerce goods title into a local large language model to obtain an optimal candidate item and a confidence score;

[0020] Step 8: Output the normalized standard product name, category path, confidence, and candidate list of the goods according to the optimal candidate item and the confidence score.

[0021] Optionally, in step 1, the hierarchical cleaning includes removing holiday modifiers, marketing words, and packaging modification information in the original e-commerce goods 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 goods core title.

[0022] Optionally, in step 5, each entry in the standard dictionary vector library is a multi-modal vector obtained by extracting semantic vectors through a text encoder and splicing other attribute fields in the entry into vectors.

[0023] Optionally, in step 5, a VF-PQ index structure is used when the query vector is subjected to approximate nearest neighbor search in the pre-constructed standard dictionary vector library.

[0024] Optionally, step 6 specifically includes:

[0025] Statistical analysis is performed according to the similarity scores of each candidate entry in the initial candidate set, and the concentration or dispersion degree of the distribution is determined, and the elbow rule or threshold is used to determine the optimal cutoff point; when the distribution is relatively dispersed and the scores differ greatly, the initial candidate number is expanded; when the distribution is relatively concentrated and the scores are close, the initial candidate set is contracted.

[0026] Optionally, in step 7, the large language model needs to determine the consistency of the candidate entry with the ontology word, whether the category path of the candidate entry is consistent with the context constraint, and whether the key attributes conflict in stages.

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

[0028] In a second aspect, an e-commerce commodity normalization system based on ontology words and vector retrieval includes:

[0029] A hierarchical cleaning module is configured to obtain an original e-commerce commodity title and perform hierarchical cleaning to obtain a commodity core title.

[0030] A label prediction module is configured to predict the part-of-speech and domain label of the commodity core title according to a sequence labeling model to obtain a candidate word set.

[0031] An ontology word extraction module is configured 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] An encoding module is configured to encode the original e-commerce commodity title and the ontology words to obtain a query vector.

[0033] A vector retrieval module is configured to perform approximate nearest neighbor search according to the query vector in a pre-constructed standard dictionary vector library to obtain an initial candidate set and return the similarity score of each candidate entry in the initial candidate set.

[0034] a candidate set adjustment module configured to perform a similarity distribution evaluation according to the similarity scores of each candidate item in the initial candidate set, and to adaptively adjust the initial candidate set according to the evaluation result to obtain a final candidate set;

[0035] a fine judgment module configured to input the final candidate set and the original e-commerce commodity title into a local large language model to obtain an optimal candidate item and a confidence score;

[0036] an output module configured to output a normalized standard product name, a category path, a confidence score and a candidate list of the commodity according to the optimal candidate item and the confidence score.

[0037] In a third aspect, a computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the e-commerce commodity normalization method based on ontology words and vector retrieval when executing the computer program.

[0038] In a fourth aspect, a computer readable storage medium stores a computer program, and the computer program implements the steps of the e-commerce commodity normalization method based on ontology words and vector retrieval when executed by a processor.

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

[0040] The present application provides an e-commerce commodity normalization method based on ontology words and vector retrieval. The original e-commerce commodity title is extracted after hierarchical cleaning, and the original e-commerce commodity title and the ontology words are encoded to obtain a query vector. The query vector is subjected to approximate nearest neighbor search in a pre-constructed standard dictionary vector library to obtain an initial candidate set and a similarity score of each candidate item. The similarity distribution of each candidate item in the initial candidate set is evaluated according to the similarity score, and the initial candidate set is adaptively adjusted according to the evaluation result to obtain a final candidate set. The final candidate set and the original e-commerce commodity title are input into a local large language model to obtain an optimal candidate item and a confidence score. The normalized standard product name, category path, confidence score and candidate list of the commodity are output according to the optimal candidate item and the confidence score. The e-commerce commodity normalization method based on ontology words and vector retrieval provided by the present application realizes accurate normalization of e-commerce commodities, and solves the problem that e-commerce commodities are difficult to accurately normalize due to the reasons such as redundant and complex commodity titles, excessively large dictionary size and direct matching difficulty in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more intuitively illustrate the prior art and the present application, exemplary drawings are given below. It should be understood that the specific shapes, structures shown in the drawings should not be considered as limiting conditions in the implementation of the present application; for example, based on the technical concepts disclosed in the present application and the exemplary drawings, those skilled in the art can easily make routine adjustments or further optimizations to the increase / decrease / assignment of certain units (components), specific shapes, positional relationships, connection methods, size ratio relationships, etc.

[0042] Figure 1 A flow chart of an e-commerce commodity normalization method based on ontology words and vector retrieval is provided for Embodiment One of the present application.

[0043] Figure 2 A structural schematic diagram of an e-commerce commodity normalization method based on ontology words and vector retrieval is provided for Embodiment One of the present application.

[0044] Figure 3 A flow chart of ontology word extraction is provided for Embodiment One of the present application.

[0045] Figure 4 A flow chart of vector recall is provided for Embodiment One of the present application.

[0046] Figure 5 A flow chart of self-adaptive candidate set adjustment is provided for Embodiment One of the present application.

[0047] Figure 6 A structural schematic diagram of an e-commerce commodity normalization device based on ontology words and vector retrieval is provided for Embodiment Two of the present application. DETAILED DESCRIPTION

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

[0049] In the description of the present application: unless otherwise specified, the meaning of "multiple" is two or more. The terms "first", "second", "third", etc. in the present application are intended to distinguish the objects referred to, and do not have special technical connotations (for example, should not be understood as emphasizing importance or order, etc.). The expressions "include", "contain", "have", etc. also mean "not limited to" (certain units, components, materials, steps, etc.).

[0050] The terms such as "upper", "lower", "left", "right", "middle", etc. cited in the present application are generally indications of the relative positional relationship for the purpose of intuitive understanding with reference to the drawings, and are not absolute limitations on the positional relationship in the actual product.

[0051] Embodiment One

[0052] Please refer to Figure 1 andFigure 2 The embodiment provides an e-commerce commodity normalization method based on ontology words and vector retrieval, which is suitable for commodity name standardization, category mapping and catalog normalization of an e-commerce platform. The method comprises the following steps:

[0053] S1: Obtain an original e-commerce commodity title, and perform hierarchical cleaning to obtain a commodity core title;

[0054] In this step, when the original e-commerce commodity title is subjected to hierarchical cleaning, holiday modification words, marketing words and packaging modification information in the original e-commerce commodity title are removed based on a regularization rule and a stop word table, and a named entity recognition model (NER) is used to extract brand entities and capacity specification entities, and a core noun phrase is reserved to obtain the commodity core title.

[0055] S2: According to a sequence labeling model, the part of speech and the domain label of the commodity core title are predicted to obtain a candidate word set;

[0056] Specifically, a sequence labeling model (for example, a Transformer-CRF) is used to predict the part of speech and the domain label of the commodity core title to obtain the candidate word set.

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

[0058] Specifically, the obtained candidate word set is aligned with the predefined ontology dictionary, and a standardized ontology word O is obtained through a similarity measurement and a synonym cluster mapping mechanism, for example, “Pu'er tea”, “basketball shoes” and “laundry detergent”, as shown in Figure 3 The steps S1 to S3 realize extraction of the ontology word.

[0059] Unlike the method of the prior art which only relies on keyword matching, the embodiment realizes robust extraction of the ontology word in a complex title through a rule template (that is, an ontology dictionary) + neural sequence labeling double-channel extraction mechanism, and introduces a confidence gating and a synonym cluster mapping. The combination scheme can still accurately extract core ontology information in the scene where marketing interference is serious or the title sequence is disordered, and significantly improves the robustness of the normalization process.

[0060] S4: Encode the original e-commerce commodity title and the ontology word, and obtain a query vector;

[0061] Specifically, please refer to Figure 4 In this step, the original e-commerce commodity title T and the extracted ontology word O are encoded to obtain a query vector E T , E O .

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

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

[0064] In this step, the standard dictionary vector library is pre-constructed. When constructing the standard dictionary vector library, the standard dictionary items D = {d1, d2, …, d n} are stored in vector form, and each item generates a multi-modal representation, including: the semantic vector E d output by the text encoder and the attribute field (category path, specifications, etc.) after embedding and splicing. That is, in the standard dictionary vector library, each item is a multi-modal vector obtained by extracting the semantic vector through the text encoder and splicing the other attribute fields in the item into vectors.

[0065] In this embodiment, steps S4 and S5 realize vector recall. Unlike existing methods that only rely on single perspective (e.g., complete title) recall, this embodiment uses a multi-perspective recall mechanism of title vector + ontology word vector to reduce noise interference while enhancing coverage of long-tail categories and variable expressions. This "multi-vector joint retrieval" strategy reduces the probability of incorrect candidates entering the subsequent discrimination phase, improving overall precision.

[0066] S6: Perform similarity distribution evaluation according to the similarity score of each candidate item in the initial candidate set, and adaptively adjust the initial candidate set according to the evaluation result 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, determines the concentration or dispersion degree of the distribution through variance, range, etc., and determines the optimal cutoff point using the elbow rule or threshold judgment. When the distribution is sparse (dispersed) and the score difference is large, the initial candidate set size is appropriately expanded to ensure recall. When the distribution is concentrated and the scores are close, the initial candidate set range is contracted to improve the purity of the results. Finally, the optimal candidate set is adaptively determined according to the distribution characteristics to balance recall rate and accuracy. Please refer to Figure 5 .

[0068] In the embodiment, the step S6 realizes the adaptive candidate set adjustment, which provides a dynamic K adjustment mechanism based on the similarity distribution characteristics, breaks through the limitation of using fixed Top-K in the prior art, and can adaptively adjust the candidate size according to the retrieval uncertainty, so that the recall rate and the precision can be considered in both high confidence and fuzzy scenarios. In particular, in the long-tail commodity and the similar category (such as “Pu’er tea” and “black tea”) scenarios, the method can effectively improve the accuracy of subsequent accurate judgment.

[0069] S7: inputting the final candidate set and the original e-commerce commodity title into a local large language model to obtain an optimal candidate item and a confidence score;

[0070] Specifically, in this step, the large language model needs to judge the consistency of the candidate item and the ontology word, whether the category path of the candidate item is consistent with the context constraint, and whether the key attributes conflict in stages, and finally output the optimal candidate item with the confidence score .

[0071] In the embodiment, the step S7 realizes the accurate judgment normalization. In the normalization judgment, the combination of “candidate set constraint + large model reasoning” is used: the candidate size is controlled by vector recall, and the final accurate judgment is completed by the semantic reasoning of the large model.

[0072] S8: outputting the normalized standard product name, category path, confidence, and candidate list of the commodity according to the optimal candidate item and the confidence score.

[0073] Specifically, the step finally outputs the standardized product name and the corresponding category path , and provides the confidence score of the normalization result and the candidate list comparison information (Top-K item and similarity / rejection reason). If is lower than the threshold θ, it is marked as “manual review is required”, supporting the manual result write-back and model continuous optimization.

[0074] The embodiment not only can output the final normalization result, but also can provide confidence quantification and candidate comparison explanation, realize the traceability and auditability of the result. Unlike the prior art which only outputs a single label, the embodiment has a “automatic adoption + manual confirmation” dual mode, which provides a guarantee for the reliable landing of normalization in a large-scale e-commerce environment.

[0075] In order to more clearly illustrate the method provided in the embodiment, the method provided in the embodiment will be further introduced below by taking the normalization of Pu’er tea commodities as an example:

[0076] S1: Input the product title: "(Qixi Special) 2025 Spring Head Unknown King Ancient Tree Pu'er Raw Tea Loose Tea 70g / Box";

[0077] S2: Ontology word extraction: "Pu'er tea" is obtained;

[0078] S3: Vector recall: search in the dictionary library, 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 judgment normalization: the large model determines the result as "Pu'er tea" in the candidate set, and outputs the confidence 0.92;

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

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

[0083] 1. Two-stage matching architecture: combining "vector recall + model fine judgment", the problem of direct matching of large-scale dictionaries being infeasible is solved.

[0084] 2. Ontology word system: through ontology word extraction, the matching accuracy is improved and redundant interference information is reduced.

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

[0086] 4. Fine judgment strategy: introduce a local large model for secondary fine judgment to significantly improve the reliability of the normalization result.

[0087] 5. Strong scalability: support dynamic dictionary update to adapt to large-scale e-commerce platform management.

[0088] In summary, the e-commerce product normalization method based on ontology words and vector retrieval provided in this embodiment realizes efficient, accurate and scalable product normalization, and solves the problem that the existing technology causes e-commerce products to be difficult to accurately normalize due to the redundancy and complexity of product titles, the large size of dictionaries, and the difficulty of direct matching.

[0089] Embodiment Two

[0090] Please refer to Figure 6 , the embodiment provides an e-commerce product normalization system based on ontology words and vector retrieval, comprising:

[0091] A hierarchical cleaning module is configured to obtain an original e-commerce commodity title and perform hierarchical cleaning to obtain a commodity core title.

[0092] A label prediction module is configured to predict the part-of-speech and domain label of the commodity core title according to a sequence labeling model to obtain a candidate word set.

[0093] An ontology word extraction module is configured to align the candidate word set with a predefined ontology dictionary and obtain standardized ontology words through a similarity measure and a synonym cluster mapping mechanism.

[0094] An encoding module is configured to encode the original e-commerce commodity title and the ontology words to obtain a query vector.

[0095] A vector retrieval module is configured to perform approximate nearest neighbor retrieval in a pre-constructed standard dictionary vector library according to the query vector to obtain an initial candidate set and return a similarity score of each candidate item in the initial candidate set.

[0096] A candidate set adjustment module is configured to perform similarity distribution evaluation according to the similarity score of each candidate item in the initial candidate set and adaptively adjust the initial candidate set according to the evaluation result to obtain a final candidate set.

[0097] A fine judgment module is configured to input the final candidate set and the original e-commerce commodity title into a local large language model to obtain an optimal candidate item and a confidence score.

[0098] An output module is configured to output a normalized standard product name, a category path, a confidence score, and a candidate list of the commodity according to the optimal candidate item and the confidence score.

[0099] The specific implementation of each module in the e-commerce commodity normalization system based on ontology words and vector retrieval can be referred to the limitation of the e-commerce commodity normalization method based on ontology words and vector retrieval in the foregoing, and will not be described here.

[0100] Embodiment Three

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

[0102] Embodiment Four

[0103] The embodiment provides a computer readable storage medium, which stores a computer program, and the computer program implements the steps of the e-commerce commodity normalization method based on ontology words and vector retrieval when executed by a processor.

[0104] Any technical features in the above embodiments can be combined (as long as the combination of the technical features does not cause contradiction), and for the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; the embodiments not explicitly written in the above description should also be considered as the scope of the present disclosure.

Claims

1. An e-commerce commodity normalization method based on ontology word and vector retrieval, characterized in that, The method comprises the following steps: Step 1: Obtain the original e-commerce commodity title and perform hierarchical cleaning to obtain a commodity core title; Step 2: According to the part-of-speech and domain label of the commodity core title predicted by the sequence labeling model, a candidate word set is obtained; Step 3: Align the candidate word set with the predefined ontology dictionary, and obtain the standardized ontology word through the similarity measurement and synonym cluster mapping mechanism; Step 4: Encode the original e-commerce commodity title and the ontology word to obtain a query vector; Step 5: According to the query vector, approximate nearest neighbor retrieval is performed in the pre-constructed standard dictionary vector library to obtain an initial candidate set, and the similarity score of each candidate item in the initial candidate set is returned; Step 6: According to the similarity score of each candidate item in the initial candidate set, the similarity distribution is evaluated, and the initial candidate set is adaptively adjusted according to the evaluation result to obtain a final candidate set; Step 7: The final candidate set and the original e-commerce commodity title are input into a local large language model to obtain an optimal candidate item and a confidence score; Step 8: According to the optimal candidate item and the confidence score, the normalized standard product name, category path, confidence and candidate list of the commodity are output. 2.The e-commerce product normalization method based on ontology word and vector retrieval according to claim 1, wherein, In step 1, the hierarchical cleaning comprises: removing holiday modifiers, marketing words and packaging modification information in the original e-commerce commodity title based on regular rules and a stop word table, and extracting brand entities and capacity specification entities by using a named entity recognition model to obtain a commodity core title. 3.The e-commerce product normalization method based on ontology word and vector retrieval according to claim 1, characterized in that, In step 5, each item in the standard dictionary vector library is a multi-modal vector obtained by extracting semantic vectors through a text encoder and concatenating other attribute fields in the item into vectors. 4.The e-commerce product normalization method based on ontology word and vector retrieval according to claim 1, wherein, In step 5, when performing approximate nearest neighbor retrieval of the query vector in the pre-constructed standard dictionary vector library, a VF-PQ index structure is adopted. 5.The e-commerce product normalization method based on ontology word and vector retrieval according to claim 1, wherein, Step 6 specifically comprises: Statistical analysis is performed according to the similarity score of each candidate item in the initial candidate set, and the concentration or dispersion degree of the distribution is judged, and the elbow rule or threshold value is used to determine the optimal cut-off point; when the distribution is relatively dispersed and the score difference is large, the initial candidate number is expanded; when the distribution is relatively concentrated and the score is close, the initial candidate set is contracted.

6. The ontology word and vector retrieval based e-commerce product normalization method of claim 1, wherein, In step 7, the large language model needs to judge the consistency of the candidate item and the ontology word, whether the category path of the candidate item is consistent with the context constraint, and whether the key attributes conflict in stages.

7. The ontology word and vector retrieval based e-commerce product normalization method of claim 1, wherein, In step 7, if the confidence score is lower than the threshold value, it is marked for manual review.

8. An e-commerce product normalization system based on ontology word and vector retrieval, characterized in that, The method comprises the following steps: A hierarchical cleaning module is configured to obtain an original e-commerce commodity title and perform hierarchical cleaning to obtain a commodity core title; A label prediction module is configured to predict the part-of-speech and domain label of the commodity core title according to a sequence labeling model to obtain a candidate word set; An ontology word extraction module is configured to align the candidate word set with a predefined ontology dictionary, and obtain a standardized ontology word through similarity measurement and synonym cluster mapping mechanism; An encoding module is configured to encode the original e-commerce commodity title and the ontology word to obtain a query vector; a vector retrieval module configured to perform approximate nearest neighbor search in a pre-constructed standard dictionary vector library according to the query vector to obtain an initial candidate set and return a similarity score of each candidate item in the initial candidate set; a candidate set adjustment module configured to perform similarity distribution evaluation according to the similarity score of each candidate item in the initial candidate set and adaptively adjust the initial candidate set according to the evaluation result to obtain a final candidate set; a fine judgment module configured to input the final candidate set and the original e-commerce commodity title into a local large language model to obtain an optimal candidate item and a confidence score; an output module configured to output a normalized standard product name, a category path, a confidence, and a candidate list of the commodity according to the optimal candidate item and the confidence score. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the steps of the method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method in any one of claims 1 to 7.

Images