Intelligent keyword retrieval method for cross-border e-commerce English based on semantic understanding

Abstract

The application relates to the field of intelligent semantic processing, and particularly relates to a cross-border e-commerce English keyword intelligent retrieval method based on semantic understanding, which comprises semantic data construction, keyword normalization, multi-layer semantic modeling, retrieval intention analysis, semantic collaborative matching and scene constraint sorting. By constructing multi-layer semantic representation and introducing a pseudo missing data generation mechanism, multi-scale hierarchical semantic coding and a probability mapping mechanism are used to form keyword semantic representation results, so that the potential semantic distribution of standardized keywords in different commodity contexts can be accurately described, and the semantic representation capability of sparse or unevenly distributed keywords can be improved. By using a multi-view semantic consistency constraint and energy joint modeling method based on keyword semantic representation results and user retrieval intention semantic representation, a candidate keyword set is generated, and sorting is performed in combination with cross-border e-commerce scene constraints, so that the retrieval result relevance and sorting stability are improved, and the fine requirement in a complex cross-border retrieval scene is met.

Description

Technical Field

[0001] This invention relates to the field of intelligent semantic processing, specifically to an intelligent retrieval method for English keywords in cross-border e-commerce based on semantic understanding. Background Technology

[0002] With the rapid development of cross-border e-commerce platforms, the number and categories of goods are constantly expanding. Users typically search for goods using English keywords. However, existing cross-border e-commerce keyword search technologies mostly rely on methods based on word frequency statistics, exact keyword matching, or simple semantic vector representation. These methods struggle to fully capture the true semantic meaning of keywords in different product contexts and business scenarios. Specific problems include:

[0003] (1) Insufficient semantic modeling of keywords

[0004] Existing technologies typically rely solely on the explicit appearance of keywords in product titles, descriptions, or attribute information to construct training samples, ignoring the potential semantic distribution of keywords in different product contexts. Since keywords are sparse or unevenly distributed in cross-border e-commerce data, this method can lead to semantic representation bias of keywords, failing to accurately depict the semantic meaning of keywords in diverse product contexts, thus affecting the retrieval system's ability to identify potentially relevant keywords.

[0005] (2) Problem of insufficient accuracy in retrieval matching and sorting

[0006] Existing methods typically employ a single-scale similarity calculation approach when matching keywords with user search intent. They lack multi-perspective consistency constraints that comprehensively consider semantic hierarchy, structural relationships, and contextual information. Furthermore, these methods do not fully incorporate the specific business scenario constraints of cross-border e-commerce, resulting in insufficient relevance of search results and poor ranking stability, making it difficult to meet the refined needs of complex cross-border search scenarios. Summary of the Invention

[0007] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent keyword retrieval method for cross-border e-commerce based on semantic understanding. Addressing the problem of insufficient keyword semantic modeling, this invention constructs a multi-layered semantic representation and introduces a pseudo-missing data generation mechanism. It uses multi-scale hierarchical semantic encoding and probability mapping mechanisms to form keyword semantic representation results, thereby accurately depicting the potential semantic distribution of standardized keywords in different product contexts and improving the semantic representation capability for sparse or unevenly distributed keywords. Addressing the problem of insufficient retrieval matching and ranking accuracy, this invention generates a candidate keyword set through a multi-perspective semantic consistency constraint and energy joint modeling method based on keyword semantic representation results and user retrieval intent semantic representation. This set is then ranked in conjunction with cross-border e-commerce scenario constraints, improving the relevance and ranking stability of retrieval results and meeting the refined needs of complex cross-border retrieval scenarios.

[0008] The technical solution adopted by this invention is as follows: The intelligent retrieval method for English keywords in cross-border e-commerce based on semantic understanding provided by this invention includes the following steps:

[0009] Step S1: Semantic data construction. Collect multi-source English data from cross-border e-commerce, including product titles, product descriptions, attribute information, category structure data, and user behavior data. Perform semantic annotation and extract initial keywords. Based on product attributes, category structure, user behavior, and business rule information, construct cross-border e-commerce scenario constraint information and build a semantic basic dataset.

[0010] Step S2: Keyword standardization, which involves uniformly mapping the initial keywords to generate a standardized keyword set;

[0011] Step S3: Multi-layer semantic modeling, constructing multi-layer semantic representation based on a standardized keyword set, and using multi-scale hierarchical semantic encoding and probability mapping mechanisms to form keyword semantic representation results;

[0012] Step S4: Search intent parsing, collect user search input and perform semantic parsing to generate a semantic representation of the user search intent;

[0013] Step S5: Semantic collaborative matching. Based on the semantic representation results of keywords and the semantic representation of search intent, collaborative matching is performed. Multi-view semantic consistency constraints and energy joint modeling methods are used to generate a set of candidate keywords.

[0014] Step S6: Sorting by scenario constraints. Based on the candidate keyword set, cross-border e-commerce scenario constraints are introduced for comprehensive sorting, and the search results are output.

[0015] Further, step S3 involves using multi-scale hierarchical semantic encoding and probabilistic mapping mechanisms to form keyword semantic representation results, specifically including the following steps:

[0016] Step S31: Data rasterization. Map the standardized keyword set to a rasterized matrix, where each row represents a standardized keyword and each column represents a data sample. Generate initial existing labels for each standardized keyword in the standardized keyword set, resulting in an initial label matrix, as shown below:

[0017] ;

[0018] in, Represents the initial label matrix. Indicates the first The standardized keyword in the first The labels appearing in each data sample This represents a standardized keyword index. Indicates the data sample index;

[0019] Step S32: Generating pseudo-missing data. Using the kernel density estimation method, elements with a value of 1 in the initial label matrix are taken as present data points, and the density value is obtained using a continuous density function. The formula used is as follows:

[0020] ;

[0021] in, Represents continuous density values. Indicates the current data point position. Indicates the first The standardized keyword in the first The location where data samples have already appeared. Indicates the first The number of existing keyword-sample combinations corresponding to each standardized keyword. For the search radius, The Gaussian kernel function;

[0022] The continuous density values ​​are normalized and mapped to a presence-absence label matrix, as shown below:

[0023] ;

[0024] in, Indicates sample Location, and For the preset threshold, This indicates the presence or absence of elements in the label matrix;

[0025] Step S33: Multi-scale feature construction. A hierarchical feature encoding network is used on the rasterized matrix and the presence-absence label matrix, which includes multi-level downsampling and semantic aggregation structures, to extract multi-scale features, as shown below:

[0026] ;

[0027] in: Representing multi-scale features, Represents a rasterized matrix. The resolution levels are 1 for the finest granularity and 4 for the coarsest granularity.

[0028] Step S34: Feature fusion, The data is concatenated and resampled to the original raster size, and then decoded using a feedforward multilayer network to generate the final semantic representation matrix. The formula used is as follows:

[0029] ;

[0030] in, Represents the semantic probability matrix of keywords. This indicates multi-scale feature splicing. Output the probability distribution values ​​for each category;

[0031] Step S35: Label correction. Use KDE kernel density thresholding to correct low-density misclassifications in the keyword semantic probability matrix to obtain the final keyword semantic representation matrix, as shown below:

[0032] ;

[0033] in, The average probability of the training set. Standard deviation This represents the final keyword semantic representation matrix;

[0034] Step S36: Output the keyword semantic representation result, and use the final keyword semantic representation matrix as the keyword semantic representation result.

[0035] Further, step S5, based on the keyword semantic representation results and the retrieval intent semantic representation, performs collaborative matching using a multi-perspective semantic consistency constraint and energy joint modeling method, specifically including the following steps:

[0036] Step S51: Multi-view semantic feature generation. Based on the keyword semantic representation matrix, multi-scale keyword semantic features are constructed from the semantic level, structural level and context level respectively. The semantic representation of the retrieval intent is mapped to the same scale to obtain a multi-view retrieval intent feature set, which includes feature vectors that correspond one-to-one with the standardized keyword features.

[0037] Step S52: Feature consistency constraint calculation. The cross-perspective consistency constraint score is calculated based on the multi-perspective keyword semantic features and the multi-perspective retrieval intent feature set, as shown below:

[0038] ;

[0039] in, Standardized keywords In the Feature vectors at a semantic scale This represents the feature vector at the corresponding scale representing the search intent. Indicates the number of semantic scales. Multi-scale semantic inconsistency between standardized keywords and search intent;

[0040] Step S53: Collaborative matching energy modeling. The consistency constraint results and standardized keyword semantic probabilities are jointly modeled to construct a collaborative matching energy function, as shown below:

[0041] ;

[0042] in, Standardized keywords The semantic existence probability, For weight adjustment parameters, This represents the energy value of the collaborative matching between standardized keywords and search intent;

[0043] Step S54: Candidate keyword generation. A preset keyword threshold is established. All standardized keywords are ranked based on their collaborative matching energy values, and keywords with energy values ​​less than the preset threshold are selected. The keywords constitute a set of candidate keywords.

[0044] The beneficial effects achieved by adopting the above solution are as follows:

[0045] (1) To address the problem of insufficient semantic modeling of keywords, this invention constructs a multi-layer semantic representation and introduces a pseudo-missing data generation mechanism. It uses multi-scale hierarchical semantic encoding and probability mapping mechanism to form the semantic representation result of keywords, thereby accurately depicting the potential semantic distribution of standardized keywords in different product contexts and improving the semantic representation ability of sparse or unevenly distributed keywords.

[0046] (2) To address the problem of insufficient accuracy in retrieval matching and ranking, this invention generates a set of candidate keywords by using a multi-perspective semantic consistency constraint and energy joint modeling method based on keyword semantic representation results and user retrieval intent semantic representation. The results are then combined with cross-border e-commerce scenario constraints for ranking, thereby improving the relevance and ranking stability of retrieval results and meeting the refined requirements of complex cross-border retrieval scenarios. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the intelligent retrieval method for English keywords in cross-border e-commerce based on semantic understanding provided by the present invention.

[0048] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0049] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0050] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0051] Example 1, see Figure 1 The present invention provides a semantic understanding-based intelligent retrieval method for English keywords in cross-border e-commerce. This embodiment uses the sale of electronic products on a cross-border e-commerce platform as an example. When users search for products on the platform, they use English keywords to perform the retrieval. The method includes the following steps:

[0052] Step S1: Semantic data construction. The platform collects the titles, descriptions, and attribute information of all products, including brand, model, color, category structure, and user history behavior. Semantic annotation and keyword extraction are performed on this data. "noise-cancelling headphones" is labeled as the initial keyword.

[0053] Step S2: Keyword standardization, unifying the mapping of different forms of initial keywords, and unifying "noisereduction" as "noise-cancelling";

[0054] Step S3: Multi-layer semantic modeling, mapping "noise-cancelling" to the product sample matrix, and generating keyword semantic representation results through multi-scale semantic encoding and probability mapping mechanisms;

[0055] Step S4: Search Intent Parsing. When a user enters "wireless noise-cancellingheadphones" in the search box, the system parses the user's search intent, including keyword combinations, intent preferences, and product attribute preferences, and converts them into semantic vector representations.

[0056] Step S5: Semantic collaborative matching. Based on the semantic representation results of keywords and the user's search intent, a set of candidate keywords is generated by combining multi-perspective semantic consistency constraints and collaborative matching rules. "wirelessheadphones" and "bluetooth earbuds" are highly related at the semantic level, and the system will include them in the candidate keywords.

[0057] Step S6: Sorting by scenario constraints. The candidate keywords are sorted by scenario constraints of cross-border e-commerce, including category priority, brand preference and price range restrictions, and the final search results list is output.

[0058] Example 2, based on the above examples, step S3, uses a multi-scale hierarchical semantic encoding and probability mapping mechanism to form the keyword semantic representation result, specifically including the following steps:

[0059] Step S31: Data rasterization. Map the standardized keyword set to a rasterized matrix, where each row represents a standardized keyword and each column represents a data sample. Generate initial existing labels for each standardized keyword in the standardized keyword set, resulting in an initial label matrix, as shown below:

[0060] ;

[0061] in, Represents the initial label matrix. Indicates the first The standardized keyword in the first The labels appearing in each data sample This represents a standardized keyword index. Indicates the data sample index;

[0062] Step S32: Generating pseudo-missing data. Using the kernel density estimation method, elements with a value of 1 in the initial label matrix are taken as present data points, and the density value is obtained using a continuous density function. The formula used is as follows:

[0063] ;

[0064] in, Represents continuous density values. Indicates the current data point position. Indicates the first The standardized keyword in the first The location where data samples have already appeared. Indicates the first The number of existing keyword-sample combinations corresponding to each standardized keyword. For the search radius, The Gaussian kernel function;

[0065] The continuous density values ​​are normalized and mapped to a presence-absence label matrix, as shown below:

[0066] ;

[0067] in, Indicates sample Location, and For the preset threshold, This indicates the presence or absence of elements in the label matrix;

[0068] Step S33: Multi-scale feature construction. A hierarchical feature encoding network is used on the rasterized matrix and the presence-absence label matrix, which includes multi-level downsampling and semantic aggregation structures, to extract multi-scale features, as shown below:

[0069] ;

[0070] in: Representing multi-scale features, Represents a rasterized matrix. The resolution levels are 1 for the finest granularity and 4 for the coarsest granularity.

[0071] Step S34: Feature fusion, The data is concatenated and resampled to the original raster size, and then decoded using a feedforward multilayer network to generate the final semantic representation matrix. The formula used is as follows:

[0072] ;

[0073] in, Represents the semantic probability matrix of keywords. This indicates multi-scale feature splicing. Output the probability distribution values ​​for each category;

[0074] Step S35: Label correction. Use KDE kernel density thresholding to correct low-density misclassifications in the keyword semantic probability matrix to obtain the final keyword semantic representation matrix, as shown below:

[0075] ;

[0076] in, The average probability of the training set. Standard deviation This represents the final keyword semantic representation matrix;

[0077] Step S36: Output the keyword semantic representation result, and use the final keyword semantic representation matrix as the keyword semantic representation result.

[0078] In this embodiment, the code used is as follows:

[0079] import numpy as np

[0080] from scipy.stats import gaussian_kde

[0081] from sklearn.preprocessing import normalize

[0082] from sklearn.neural_network import MLPClassifier

[0083] # Assuming a standardized set of keywords and data samples

[0084] keywords = ["wireless", "headphones", "bluetooth", "noise-cancelling"]

[0085] samples = ["sample1", "sample2", "sample3", "sample4", "sample5"]

[0086] # -----------------------------

[0087] # Step S31: Data Rasterization

[0088] # -----------------------------

[0089] # Construct the initial label matrix Y_init

[0090] #1 indicates that the keyword appeared in the sample, and 0 indicates that it did not appear.

[0091] Y_init = np.array([

[0092] [1, 0, 1, 0, 0], # wireless

[0093] [0, 1, 1, 0, 0], # headphones

[0094] [0, 0, 1, 1, 0], # bluetooth

[0095] [0, 0, 0, 1, 1], # noise-cancelling ])

[0097] # -----------------------------

[0098] # Step S32: Generating Pseudo-Missing Data

[0099] # -----------------------------

[0100] Y_kde = np.zeros_like(Y_init, dtype=float)

[0101] threshold_high = 0.7

[0102] threshold_low = 0.3

[0103] for i in range(Y_init.shape[0]): # Iterate through keywords

[0104] # Get the index of the samples that have appeared

[0105] appeared_idx = np.where(Y_init[i] == 1)[0]

[0106] if len(appeared_idx) == 0:

[0107] continue

[0108] # Using kernel density estimation

[0109] kde = gaussian_kde(appeared_idx, bw_method=0.5) # h ~ 0.5

[0110] for j in range(Y_init.shape[1]):

[0111] density = kde(j)

[0112] # Mapping to Existence-Non-existence labels

[0113] if density >= threshold_high:

[0114] Y_kde[i, j] = 1 # Exists

[0115] elif density <= threshold_low:

[0116] Y_kde[i, j] = 0 # Does not exist

[0117] else:

[0118] Y_kde[i, j] = 0.5 # Potentially exists

[0119] # -----------------------------

[0120] # Step S33: Multi-scale feature construction

[0121] # -----------------------------

[0122] # Simulate multi-scale encoding, generating random features at each scale as an example.

[0123] F = []

[0124] for l in range(4): # 1-finest granularity, 4-coarsest granularity

[0125] # Generate 10-dimensional features for each keyword and each sample.

[0126] features = np.random.rand(Y_init.shape[0], Y_init.shape[1], 10) *(l+1)

[0127] F.append(features)

[0128] # -----------------------------

[0129] # Step S34: Feature Fusion

[0130] # -----------------------------

[0131] # Simple concatenation and resampling

[0132] F_concat = np.concatenate([f.reshape(Y_init.shape[0], -1) for f inF], axis=1)

[0133] # Using feedforward MLP decoding to generate probability matrices

[0134] mlp = MLPClassifier(hidden_layer_sizes=(32,), max_iter=500)

[0135] # Simulated training: Use Y_kde > 0.5 as positive sample labels

[0136] y_train = (Y_kde > 0.5).astype(int).flatten()

[0137] mlp.fit(F_concat, y_train)

[0138] S_prob = mlp.predict_proba(F_concat)[:, 1].reshape(Y_init.shape[0],Y_init.shape[1])

[0139] # -----------------------------

[0140] # Step S35: Label Correction

[0141] # -----------------------------

[0142] mu = np.mean(S_prob)

[0143] sigma = np.std(S_prob)

[0144] S_final = np.zeros_like(S_prob)

[0145] for i in range(S_prob.shape[0]):

[0146] for j in range(S_prob.shape[1]):

[0147] if S_prob[i, j] > mu + 3*sigma:

[0148] S_final[i, j] = 1

[0149] elif S_prob[i, j] < mu:

[0150] S_final[i, j] = 0

[0151] else:

[0152] S_final[i, j] = S_prob[i, j]

[0153] # -----------------------------

[0154] # Step S36: Output keyword semantic representation results

[0155] # -----------------------------

[0156] keyword_semantic_matrix = S_final

[0157] print("Final keyword semantic representation matrix:\n", keyword_semantic_matrix).

[0158] Example 3, based on the above examples, takes the user's input search intent "highquality bluetooth headphones" as an example. Semantic collaborative matching is performed on the standardized keyword set based on step S5. The standardized keyword set is ["wireless", "headphones", "bluetooth", "noise-cancelling", "portable"], corresponding to the final keyword semantic representation matrix generated in step S3. The semantic existence probabilities in are respectively Step S5 specifically includes the following steps:

[0159] Step S51: Multi-view semantic feature generation. Multi-scale keyword features are extracted from the semantic, structural, and contextual levels. The feature vector for "bluetooth" is as follows at the four levels:

[0160] ;

[0161] The user's search intent "high quality Bluetooth headphones" is also mapped to four levels of feature vectors;

[0162] ;

[0163] Step S52: Feature consistency constraint calculation. Based on vector cosine similarity, calculate the semantic inconsistency between keywords and search intent at each level.

[0164] ;

[0165] Taking "bluetooth" as an example, the cosine inconsistencies of the four levels are 0.011, 0.018, 0.013, and 0.017, respectively. The average is calculated as follows: ;

[0166] Step S53: Collaborative matching energy modeling, combining consistency constraints with semantic existence probability. Joint modeling is performed to construct a cooperative matching energy function, as follows:

[0167] ;

[0168] Assuming weight parameters Then the cooperative matching energy value of "bluetooth" is:

[0169] ;

[0170] The energy values ​​of the other keywords are as follows:

[0171] wireless: ;

[0172] headphones: ;

[0173] noise cancellation: ;

[0174] portable: ;

[0175] Step S54: Candidate keyword generation, setting threshold Filter keywords whose collaborative matching energy value is less than the threshold to obtain the candidate keyword set as: ["bluetooth", "wireless", "headphones"].

[0176] Example 4: This example is based on the above examples. In this example, the user searches for "wireless earbuds" on a cross-border e-commerce platform, hoping to find Bluetooth earphone products.

[0177] The platform collected 50,000 product data entries, including titles, descriptions, and attributes, and extracted standardized keywords, including "wireless", "earbuds", "bluetooth", and "noise-canceling". In the grid matrix, if a product title contains "wireless earbuds", the corresponding position is marked as 1; otherwise, it is marked as 0.

[0178] Many product descriptions only mention "bluetooth earbuds" without explicitly stating "wireless earbuds";

[0179] Using kernel density estimation, the potential existence probability is generated based on the distribution of "wireless earbuds" products that have already appeared, and the potential existence probability is marked as 0.5.

[0180] Multi-scale feature construction and feature fusion:

[0181] F1: Product-level features;

[0182] F2: Category-level features;

[0183] F3: User behavior characteristics;

[0184] F4: Platform global distribution characteristics;

[0185] Generate keyword semantic probability matrix through MLP fusion;

[0186] Collaborative matching and candidate keyword generation:

[0187] The user's search intent was "wireless earbuds";

[0188] Calculating multi-view semantic consistency reveals that "bluetooth earbuds" is highly consistent with user intent across F1 to F4 layers of features;

[0189] After the energy function is calculated, the keyword is selected into the candidate keyword set, and the final sorted output is "wirelessearbuds", "bluetooth earbuds", and "noise-canceling earbuds".

[0190] When users enter "wireless earbuds", relevant products can be retrieved even if the term is not explicitly mentioned in some products, thus improving the hit rate and recommendation relevance.

[0191] Example 5, based on the above examples, in this example, when a user searches for "summer beachhat", the system needs to identify potential related keywords;

[0192] The platform collected 30,000 pieces of data on summer clothing and accessories, extracted keywords such as "summer hat", "beach hat", "sun protection", and "wide brim", and generated an initial tag matrix using a grid matrix to mark whether the keywords appeared or not.

[0193] Some product descriptions only state "sun hat" without specifying "beach hat";

[0194] The KDE method generates a potential existence probability, which is marked as 0.5 to improve semantic coverage;

[0195] Multi-scale feature construction and fusion:

[0196] F1: Product title and attribute text features;

[0197] F2: Category hierarchy characteristics;

[0198] F3: User behavior characteristics;

[0199] F4: Popularity distribution characteristics across all platforms;

[0200] After concatenating multi-scale features, a keyword semantic probability matrix is ​​generated using MLP.

[0201] Collaborative matching and candidate keyword generation:

[0202] When a user inputs "summer beach hat", the system calculates the semantic consistency of the keywords.

[0203] "sun hat" and "wide brim hat" have a high degree of matching with user intent and low collaborative energy, and are selected into the candidate set;

[0204] The final set of candidate keywords is "summer beach hat", "sun hat", and "wide brim hat".

[0205] Even though the user did not explicitly enter "wide brim hat", the system identified potentially related products through semantic modeling, improving search accuracy and coverage.

[0206] Example 6: This example is based on the above examples. In this example, Ms. Zhang is purchasing a baby food processor for her baby through a cross-border e-commerce platform in Singapore. She enters "baby food processor" in the search bar. The system collects real baby product data from the platform, including titles, product descriptions, brand information, specifications, and user reviews. The initial keywords are extracted as "baby food processor", "blender", "infant puree maker", "BPA-free", and "electric baby blender". The product category and brand constraints are marked as "Kitchen Appliances > Baby Products" and the brands are "Philips Avent" and "Beaba".

[0207] Map 2000 standardized keywords into a 2000-row × 10000-column matrix;

[0208] Ms. Zhang searched for "baby food processor". In the matrix, if a product title contains this word, it is marked as 1; otherwise, it is marked as 0. The product title "Beaba Babycook Deluxe" only contains "Babycook", so the matrix is ​​initially set to 0.

[0209] The system analyzed the distribution of products containing "baby food processor" and found that "Babycook Deluxe" and "Baby Food Maker" are highly related in terms of function and brand, but their initial tags were 0.

[0210] Using kernel density estimation, these potentially relevant goods are labeled with a value of 0.5;

[0211] F1: Product-level text features;

[0212] F2: Category hierarchy characteristics;

[0213] F3: User behavior characteristics;

[0214] F4: Popularity characteristics across all platforms;

[0215] Multi-scale features are fused using MLP to generate a keyword semantic probability matrix;

[0216] The user intent semantically represents "baby food processor";

[0217] Calculating the multi-perspective consistency between each product keyword and user intent, "Babycook Deluxe" shows a high match with "baby food processor" in terms of text, category, and user behavior, but with low synergy energy. "Philips Avent BabyBlender" also shows a high match. Candidate keyword sets are then filtered based on synergy matching energy values.

[0218] "baby food processor";

[0219] “Babycook Deluxe”;

[0220] "Philips Avent Baby Blender";

[0221] The system takes into account the cross-border scenario constraints of the user being located in Singapore, the goods being available for direct shipping, and the current discount promotion. It sorts the candidate products and outputs the search results. Even if Ms. Zhang only entered "baby food processor", she can see the most matching brand products and see the information "available for direct shipping to Singapore" and "10% discount".

[0222] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0223] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0224] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A semantic understanding-based intelligent keyword retrieval method for cross-border e-commerce, characterized by: The present invention provides an intelligent retrieval method for English keywords in cross-border e-commerce based on semantic understanding. The method includes the following steps: Step S1: Semantic data construction. Collect multi-source English data from cross-border e-commerce, including product titles, product descriptions, attribute information, category structure data, and user behavior data. Perform semantic annotation and extract initial keywords. Based on product attributes, category structure, user behavior, and business rule information, construct cross-border e-commerce scenario constraint information and build a semantic basic dataset. Step S2: Keyword standardization, which involves uniformly mapping the initial keywords to generate a standardized keyword set; Step S3: Multi-layer semantic modeling, constructing multi-layer semantic representation based on a standardized keyword set, and using multi-scale hierarchical semantic encoding and probability mapping mechanisms to form keyword semantic representation results; Step S4: Search intent parsing, collect user search input and perform semantic parsing to generate a semantic representation of the user search intent; Step S5: Semantic collaborative matching. Based on the semantic representation results of keywords and the semantic representation of search intent, collaborative matching is performed. Multi-view semantic consistency constraints and energy joint modeling methods are used to generate a set of candidate keywords. Step S6: Sorting by scenario constraints. Based on the candidate keyword set, cross-border e-commerce scenario constraints are introduced for comprehensive sorting, and the search results are output.

2. The intelligent retrieval method for English keywords in cross-border e-commerce based on semantic understanding according to claim 1, characterized in that: Step S3 involves using multi-scale hierarchical semantic encoding and probability mapping mechanisms to form keyword semantic representation results, specifically including the following steps: Step S31: Data rasterization. Map the standardized keyword set to a rasterized matrix, where each row represents a standardized keyword and each column represents a data sample. Generate initial existing labels for each standardized keyword in the standardized keyword set, resulting in an initial label matrix, as shown below: ； in, Represents the initial label matrix. Indicates the first The standardized keyword in the first The labels appearing in each data sample This represents a standardized keyword index. Indicates the data sample index; Step S32: Generating pseudo-missing data. Using the kernel density estimation method, elements with a value of 1 in the initial label matrix are taken as present data points, and the density value is obtained using a continuous density function. The formula used is as follows: ； in, Represents continuous density values. Indicates the current data point position. Indicates the first The standardized keyword in the first The location where data samples have already appeared. Indicates the first The number of existing keyword-sample combinations corresponding to each standardized keyword. For the search radius, The Gaussian kernel function; The continuous density values ​​are normalized and mapped to a presence-absence label matrix, as shown below: ； in, Indicates sample Location, and For the preset threshold, This indicates the presence or absence of elements in the label matrix; Step S33: Multi-scale feature construction. A hierarchical feature encoding network is used on the rasterized matrix and the presence-absence label matrix, which includes multi-level downsampling and semantic aggregation structures, to extract multi-scale features, as shown below: ； in: Representing multi-scale features, Represents a rasterized matrix. The resolution levels are 1 for the finest granularity and 4 for the coarsest granularity. Step S34: Feature fusion, The data is concatenated and resampled to the original raster size, and then decoded using a feedforward multilayer network to generate the final semantic representation matrix. The formula used is as follows: ； in, Represents the semantic probability matrix of keywords. This indicates multi-scale feature splicing. Output the probability distribution values ​​for each category; Step S35: Label correction. Use KDE kernel density thresholding to correct low-density misclassifications in the keyword semantic probability matrix to obtain the final keyword semantic representation matrix, as shown below: ； in, The average probability of the training set. Standard deviation This represents the final keyword semantic representation matrix; Step S36: Output the keyword semantic representation result, and use the final keyword semantic representation matrix as the keyword semantic representation result.

3. The intelligent retrieval method for English keywords in cross-border e-commerce based on semantic understanding according to claim 1, characterized in that: Step S5 involves performing collaborative matching based on the keyword semantic representation results and the retrieval intent semantic representation, using a multi-perspective semantic consistency constraint and energy joint modeling method. This includes the following steps: Step S51: Multi-view semantic feature generation. Based on the keyword semantic representation matrix, multi-scale keyword semantic features are constructed from the semantic level, structural level and context level respectively. The semantic representation of the retrieval intent is mapped to the same scale to obtain a multi-view retrieval intent feature set, which includes feature vectors that correspond one-to-one with the standardized keyword features. Step S52: Feature consistency constraint calculation. The cross-perspective consistency constraint score is calculated based on the multi-perspective keyword semantic features and the multi-perspective retrieval intent feature set, as shown below: ； in, Standardized keywords In the Feature vectors at a semantic scale This represents the feature vector at the corresponding scale representing the search intent. Indicates the number of semantic scales. Multi-scale semantic inconsistency between standardized keywords and search intent; Step S53: Collaborative matching energy modeling. The consistency constraint results and standardized keyword semantic probabilities are jointly modeled to construct a collaborative matching energy function, as shown below: ； in, Standardized keywords The semantic existence probability, For weight adjustment parameters, This represents the energy value of the collaborative matching between standardized keywords and search intent; Step S54: Candidate keyword generation. A preset keyword threshold is established. All standardized keywords are ranked based on their collaborative matching energy values, and keywords with energy values ​​less than the preset threshold are selected. The keywords constitute a set of candidate keywords.

Images