Method and device for determining commodity category information

Abstract

The present application proposes a method and device for determining commodity category information, wherein the method for determining commodity category information includes: obtaining multiple first target product words corresponding to sample commodities, constructing a commodity category system based on the hierarchical relationship between the first target product words, and determining the category information of the commodity to be matched according to the commodity category system. The product word of a commodity is used to define the commodity and can reflect the category and essential attributes of the commodity. The method constructs a commodity category system based on the product word of the commodity. The constructed commodity category system contains the category information of all sample commodities. These categories are divided based on the definition of the commodity itself and can be better applied to other new e-commerce scenarios such as search and recommendation. The commodity to be matched can be matched with the category information of the sample commodity based on the commodity category system, thereby determining the category information of the commodity to be matched and reducing the cost of determining and updating the commodity category information.

Description

Technical Field

[0001] The present application relates to the field of data analysis technology, and in particular to a method and device for determining commodity category information. Background Art

[0002] Product category is one of the master data of the product management system. To meet the needs of product production, circulation and consumption activities, e-commerce platforms classify all products according to certain management purposes and gradually summarize them into several smaller sub-aggregates (categories) with more consistent characteristics, such as major categories, medium categories, minor categories, detailed categories, and even varieties and details, so that all products within the range can be clearly distinguished.

[0003] With the digitalization, e-commerce platforms are no longer faced with single product sales scenarios. There are also more interactive scenarios with users, such as recommendations and searches. However, compared with the original business operation and sales scenarios, these scenarios have inconsistent demands for product categories and category systems. They require category information based on the essential categories of the products to describe the essence of the products.

[0004] In the related technologies, the cost of determining and updating product category information is high, and product categories are mostly divided according to business sales scenarios, which are not very applicable to other new e-commerce scenarios such as search, recommendation and advertising. Summary of the invention

[0005] The present application aims to solve one of the technical problems in the related art at least to some extent.

[0006] To this end, the first objective of this application is to propose a method for determining product category information.

[0007] The second objective of the present application is to provide a device for determining product category information.

[0008] The third objective of the present application is to provide an electronic device.

[0009] A fourth object of the present application is to provide a non-transitory computer-readable storage medium.

[0010] A fifth object of the present application is to provide a computer program product.

[0011] To achieve the above-mentioned purpose, the first aspect embodiment of the present application proposes a method for determining product category information, including: obtaining multiple first target product words corresponding to sample products; constructing a product category system based on the hierarchical relationship between the first target product words; and determining the category information of the product to be matched based on the product category system.

[0012] This application obtains multiple first target product words corresponding to sample products; constructs a product category system based on the hierarchical relationship between the first target product words; and determines the category information of the product to be matched according to the product category system. The product word of a product is used to define the product and can reflect the category and essential attributes of the product. This method constructs a product category system based on the product word of the product. The constructed product category system contains the category information of all sample products. These categories are divided based on the definition of the product itself and can be better applied to other new e-commerce scenarios such as search, recommendation and advertising. The product to be matched can be matched based on the product category system and the category information of the sample product, thereby determining the category information of the product to be matched and reducing the cost of determining and updating the product category information.

[0013] To achieve the above-mentioned purpose, the second aspect embodiment of the present application proposes a device for determining product category information, including: an acquisition module, used to obtain multiple first target product words corresponding to sample products; a construction module, used to construct a product category system based on the hierarchical relationship between the first target product words; a determination module, used to determine the category information of the product to be matched according to the product category system.

[0014] This application obtains multiple first target product words corresponding to sample products; constructs a product category system based on the hierarchical relationship between the first target product words; and determines the category information of the product to be matched according to the product category system. The product word of a product is used to define the product and can reflect the category and basic attributes of the product. This method constructs a product category system based on the product word of the product. The constructed product category system contains the category information of all sample products. These categories are divided based on the definition of the product itself and can be better applied to other new e-commerce scenarios such as search, recommendation and advertising. The product to be matched can be matched based on the product category system and the category information of the sample product, thereby determining the category information of the product to be matched and reducing the cost of determining and updating the product category information.

[0015] To achieve the above-mentioned purpose, the third aspect embodiment of the present application proposes an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to implement the method for determining product category information as described in the first aspect embodiment of the present application.

[0016] To achieve the above-mentioned purpose, the fourth aspect embodiment of the present application proposes a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to implement the method for determining product category information as described in the first aspect embodiment of the present application.

[0017] To achieve the above-mentioned purpose, the fifth aspect embodiment of the present application proposes a computer program product, including a computer program, which, when executed by a processor, is used to implement the method for determining product category information as described in the first aspect embodiment of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A flowchart of a method for determining commodity category information provided in an embodiment of the present application;

[0019] Figure 2 A flowchart of a method for determining commodity category information provided in another embodiment of the present application;

[0020] Figure 3 A flowchart of a method for determining commodity category information provided in another embodiment of the present application;

[0021] Figure 4 A flowchart of a method for determining commodity category information provided in another embodiment of the present application;

[0022] Figure 5 A schematic diagram of a binary model for determining synonyms in a method for determining commodity category information provided in an embodiment of the present application;

[0023] Figure 6 A flowchart of a method for determining commodity category information provided in another embodiment of the present application;

[0024] Figure 7 A schematic diagram of a hierarchical relationship extraction model in a method for determining commodity category information provided in an embodiment of the present application;

[0025] Figure 8 A partial schematic diagram of a commodity category system of a method for determining commodity category information provided in an embodiment of the present application;

[0026] Fig. 9 A flowchart of a method for determining commodity category information provided in another embodiment of the present application;

[0027] Fig.10 A schematic diagram of the overall process of a method for determining commodity category information provided in an embodiment of the present application;

[0028] Fig.11 A block diagram of a device for determining commodity category information proposed in this application;

[0029] Fig.12 A block diagram of an electronic device provided for this application. DETAILED DESCRIPTION

[0030] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0031] Figure 1 This is a flow chart of a method for determining commodity category information provided in an embodiment of the present application. The method for determining commodity category information in the embodiment of the present application can be performed by a device for determining commodity category information provided in the embodiment of the present application. The device for determining commodity category information can be set in electronic devices such as terminals and servers. Figure 1 As shown, the method for determining the commodity category information in the embodiment of the present application includes the following steps:

[0032] S101, obtaining multiple first target product words corresponding to sample products.

[0033] The embodiment of the present application takes an e-commerce platform as an example. The sample products are the main products in the product library of the e-commerce platform or a large number of products accumulated by the platform. One product in the sample products usually corresponds to one or more product words. These product words belong to different granularities, such as jacket < sweatshirt, pants < skinny pants. These product words can reflect the essential attributes and categories of the products. This application uses these product words as the first target product words.

[0034] S102, constructing a product category system based on the hyponymous and hyponymous relationships between the first target product words.

[0035] In an embodiment of the present application, by identifying whether there is a hierarchical relationship between any two first target product words, and distinguishing between superordinate product words and subordinate product words for the two first target product words that have a hierarchical relationship, different category branches can be constructed based on the hierarchical relationship between the first target product words, thereby forming a commodity category system.

[0036] S103: Determine the category information of the product to be matched according to the product category system.

[0037] In an embodiment of the present application, for the goods to be matched, the category to which the goods to be matched belong can be determined based on the constructed product category system. The category information of the goods to be matched can also include the categories corresponding to the goods to be matched at different granularities, for example, the category corresponding to apples: fruit < fruit < apples.

[0038] The embodiment of the present application proposes a method for determining commodity category information, obtaining multiple first target product words corresponding to sample commodities, constructing a commodity category system based on the hierarchical relationship between the first target product words, and determining the category information of the commodity to be matched according to the commodity category system. The product word of a commodity is used to define the commodity and can reflect the category and essential attributes of the commodity. The method constructs a commodity category system based on the product word of the commodity. The constructed commodity category system contains the category information of all sample commodities. These categories are divided based on the definition of the commodity itself and can be better applied to other new e-commerce scenarios such as search and recommendation. The commodity to be matched can be matched with the category information of the sample commodity based on the commodity category system, thereby determining the category information of the commodity to be matched and reducing the cost of determining and updating the commodity category information.

[0039] Based on the above embodiments, Figure 2 As shown, the "obtaining multiple first target product words corresponding to the sample product" in the above step S101 may include the following steps:

[0040] S201, performing part-of-speech tagging on text information corresponding to the sample product based on preset part-of-speech tags.

[0041] In the embodiment of the present application, the text information corresponding to the sample product may be known text information such as the product title and search log of the product in the e-commerce platform. According to the characteristics of these text information, the part-of-speech tags or required part-of-speech tags that the word segments in the text information may belong to are sorted. For example, for e-commerce scenarios, the part-of-speech tags may be: brand words, model words, product words, applicable events, applicable time, applicable location, applicable population, other applicable objects, functional attributes, material attributes, style attributes, style attributes, origin attributes, color attributes, taste attributes and specification attributes, etc. These part-of-speech tags can be used as preset part-of-speech tags to perform part-of-speech tagging on the text information corresponding to the sample product.

[0042] In some embodiments, a part-of-speech tagging model can be used to perform part-of-speech tagging on the text information that has been uniformly converted to upper and lower case. For example, the text information can be segmented using word segmentation tools such as Jieba, Language Technology Platform (LTP), and then the CRF layer (BilSTM-CRF) in the bidirectional long short-term memory network or the Transformer Encoder for Name Entity Recognition (TENER) and other models can be used for part-of-speech tagging (named entity recognition), or the pre-trained language representation model (Bidirectional Encoder Representation from Transformers, BERT) can be pre-trained based on the existing text information, and then the BERT+CRF and other models can be used for part-of-speech tagging without word segmentation. As shown in Table 1, after the product title is segmented, each segmented word is tagged with a part-of-speech tag based on a preset part-of-speech tag.

[0043] Table 1 - Example table of product title part-of-speech tagging

[0044]

[0045] S202, screening out a plurality of first candidate product words from the text information after part-of-speech tagging.

[0046] From the text information after part-of-speech tagging, nouns whose part of speech is product words are selected and these nouns are used as the first candidate product words, such as sweatshirt and jacket in Table 1.

[0047] S203, performing synonymous merging on a plurality of first candidate product words to obtain a plurality of first target product words.

[0048] In the embodiment of the present application, some text information of the e-commerce platform is manually input by merchants or users, but different merchants or users have different writing habits, or intentionally or unintentionally input typos, such as synonyms such as shampoo, shampoo water, shampoo liquid and shampoo night, resulting in fragmentation and duplication of the first candidate product words obtained. Therefore, we need to synonymously merge the first candidate product words of different expressions of the same product to ensure the independence of each branch of the product category system.

[0049] In some embodiments, it is possible to determine whether any two first candidate product words are synonyms and merge the two first candidate product words that are synonyms into one noun, that is, to express them with one noun, thereby obtaining multiple first target product words with lower synonymy.

[0050] Based on the above embodiments, Figure 3 As shown, in the above step S203, "synonym-merging multiple first candidate product words to obtain multiple first target product words" may include the following steps:

[0051] S301, determining a plurality of similar product word pairs according to similarity scores between the first candidate product words.

[0052] In an embodiment of the present application, the similarity scores between the first candidate product words are calculated by vectorizing the first candidate product words. For example, based on the first representation vectors corresponding to the first candidate product words, the vector distance between the first representation vectors corresponding to any two first candidate product words is calculated to determine the similarity scores between any two first candidate product words, and any two first candidate product words with a similarity score greater than a first threshold are determined as similar product word pairs. The vector distance may include, but is not limited to, cosine distance and Euclidean distance.

[0053] In some embodiments, in order to more comprehensively characterize the first candidate product word, we obtain the representation vector (Embedding) of the first candidate product word from both semantic and behavioral aspects. The semantic representation vector (semantic Embedding) and the behavioral representation vector (behavioral Embedding) of the first candidate product word can be concatenated to generate the above-mentioned first representation vector.

[0054] Among them, semantic embedding can be obtained based on an unsupervised word vector model, where the unsupervised word vector model includes but is not limited to a word vector model (word2vec), text classification (fasttext) and BERT.

[0055] Behavior Embedding can be obtained based on search behavior, such as search (query). The specific steps are: obtain the text information such as the product title of the product clicked by the user under the search behavior corresponding to the first candidate product word; extract keywords from the product title to obtain keyword representation segmentation; obtain the semantic representation vector of the keyword representation segmentation; and use the weighted sum of the semantic representation vectors of the keyword representation segmentation as the behavior representation vector of the first candidate product word.

[0056] For example, based on a search formula containing a first candidate product word, a product search is performed on an e-commerce platform to obtain a sequence of products clicked by users under the search behavior, thereby obtaining the product titles of the products clicked by users, and then using an unsupervised keyword extraction method to obtain the key representation participles of the first candidate product word, and finally using the weighted sum of the semantic embeddings of the key representation participles to represent the behavioral embedding of the first candidate product word, wherein the unsupervised keyword extraction method may include but is not limited to keyword extraction based on term frequency-inverse document frequency (TF-IDF), keyword extraction and summarization algorithm (TextRank) and latent Dirichlet allocation model (LDA).

[0057] The semantic representation vector and the behavior representation vector of the first candidate product word are concatenated, for example, a semantic representation vector with a dimension of 25 and a behavior representation vector with a dimension of 25 are concatenated into a first representation vector with a dimension of 50.

[0058] For example, when a user enters "mobile phone" to search, many products related to mobile phones will appear. The titles of the products clicked by the user are obtained for keyword extraction, and the top 30% of the keywords that appear in these product titles are selected. These keywords are used as key representation word segmentations in the search behavior of "mobile phone", and semantic representation is performed on these key representation word segmentations to obtain the semantic representation vectors of the key representation word segmentations.

[0059] In some embodiments, the similarity between the two first candidate product words is characterized by calculating the cosine value between the first representation vectors of any two first candidate product words, and the cosine value is used as the similarity score, wherein the calculation formula of the similarity score is as follows:

[0060]

[0061]

[0062] Among them, product word 1 and product word 2 are any two first candidate product words, Y is the first representation vector of any first candidate product word, θ is the angle between the two first representation vectors, and X is the semantic representation vector obtained based on the unsupervised word vector model (X 产品词1 is the semantic representation vector of product word 1, X 产品词1的关键表征分词i is the semantic representation vector of the i-th key representation word corresponding to product word 1), n ​​is the number of key representation words of product word 1 based on search behavior, and n can be selected as 30, 50, etc. based on experience.

[0063] As an example of the results, assuming that the vector dimension of the semantic representation vector and the behavior representation vector is 25, the vector dimension of the first representation vector is 50, and the vectorized representation of the first candidate product words "shampoo", "shampoo", and "shower gel" are shown in Table 2 below. The calculated cosine distance of the first candidate product word pair [shampoo, shampoo] is 0.9512, and the cosine distance of the first candidate product word pair [shampoo, shower gel] is 0.8730. The vector distance shows that the similarity of the first candidate product word pair [shampoo, shampoo] is significantly higher than that of the first candidate product word pair [shampoo, shower gel].

[0064] Table 2 - Example of vector representation results of the first candidate product word

[0065]

[0066]

[0067] S302: Based on the synonym discrimination model, multiple synonymous product word pairs are selected from multiple similar product word pairs.

[0068] In some embodiments, similar product word pairs may be similar words or synonyms, and the obtained multiple similar product word pairs need to be screened to obtain synonymous product word pairs.

[0069] In some embodiments, the model can be iteratively trained based on the active learning method to obtain a synonym discrimination model, such as a linear model and a deep model (Wide & Deep model). It is not difficult to understand that the active learning method refers to first using a small number of labeled data for training, then scoring the training, sorting the samples in the middle segment according to the scores, then annotating and updating the data set, and then training again, so as to iterate until the model indicators meet the requirements.

[0070] S303: normalize each synonymous product word pair into a second candidate product word.

[0071] In an embodiment of the present application, synonymous product words are normalized, and two synonymous product words in a synonymous product word pair are represented by one product word, which is used as a second candidate product word, thereby obtaining multiple second candidate product words with lower synonymy.

[0072] S304: Select multiple first target product words from multiple second candidate product words according to the commodity coverage of the second candidate product words and the frequency of occurrence of the second candidate product words in the search behavior.

[0073] In practice, the number of product words on e-commerce platforms can reach hundreds of thousands, and a category system of this magnitude would be too large and complex in actual application. In addition, the coverage of product words on commodities is a long-tail distribution, that is, the top 20,000 product words cover 98% of the commodities on the platform. This ranking can be the frequency ranking of product words in all commodity titles on the platform. Therefore, the main product words of the e-commerce platform are obtained by screening the obtained second candidate product words to build a commodity category system.

[0074] For example, according to the distribution of product words on the e-commerce platform, select the product words that cover the top 98% of the number of products ranked from high to low from multiple second candidate product words, and the union of the product words that appear in the top 98% of the frequency under search behavior (the frequency of occurrence under search behavior can be understood as: the user inputs the search formula "xx mobile phone", and the identified product word is mobile phone, then the frequency corresponding to mobile phone is increased by 1), and each product word in the union set (i.e., the second candidate product word) is used as the first target product word. This takes into account the coverage of the products sold on the e-commerce platform on the one hand, and the coverage rate of the products that users pay attention to on the other hand. If the e-commerce platform already has an original category system, the item "product words that cover the top 98% of the number of products ranked from high to low" can be refined into the union of product words that cover the top 95% of the number of products ranked from high to low under a certain granularity category. "The top 98% of product terms in terms of the number of covered products ranked from high to low" is counted based on the total number of products on the platform. If the e-commerce platform focuses on home appliances or clothing, then the proportion of home appliances or clothing will be higher, and some products with a smaller proportion will be difficult to cover. Counting the number of covered products under a certain granularity category can increase the comprehensiveness of coverage. For example, select the top 95% of product terms in each category of a certain granularity, such as 95% for fresh food, 95% for clothing, 95% for beauty products, and so on.

[0075] Based on the above embodiments, Figure 4 As shown, the above step S302 "selecting multiple synonymous product word pairs from multiple similar product word pairs based on the synonym discrimination model" may include the following steps:

[0076] S401, obtaining a first frequency of two product words in a similar product word pair appearing in a preset scenario respectively, and a second frequency of the similar product word pair appearing simultaneously in the preset scenario.

[0077] In an embodiment of the present application, the preset scenario includes at least one of the following: a product title library, a search click behavior, and a recommendation click behavior.

[0078] If the preset scenario is a product title library, the first frequency and the second frequency are obtained based on all product titles on the e-commerce platform: the frequency of the two product words in the similar product word pairs appearing in all product titles on the e-commerce platform is obtained as the first frequency; based on all product titles on the e-commerce platform, the frequency of similar product word pairs appearing simultaneously in one product title is obtained as the second frequency.

[0079] If the preset scenario is search and click behavior, based on the search log, the frequency of similar product word pairs appearing simultaneously in the product click list corresponding to each search behavior of the user is obtained as the second frequency, and the frequency of any product word in the similar product words appearing in the product click list is obtained as the first frequency.

[0080] If the preset scenario is a recommendation click behavior, the product click sequence corresponding to the recommendation click behavior can be understood as a product sequence clicked based on the platform's recommendation (such as homepage product recommendation), which can include homepage products clicked by users and products clicked in the drop-down page triggered by homepage products. The frequency of similar product word pairs appearing simultaneously in the product click sequence corresponding to the recommendation click behavior is obtained as the second frequency, and the frequency of the two product words in the similar product word pairs appearing separately is obtained as the first frequency.

[0081] It should be noted that the first frequency is the frequency of a product word appearing in a preset scenario, and only considers whether the product word appears, that is, the first frequency includes the number of times the product word appears alone and the number of times it appears at the same time as another product word in a similar product word pair.

[0082] S402: Calculate the relevance score of the similar product word pair in a preset scenario according to the first frequency and the second frequency.

[0083] According to the first frequency of two product words in the similar product word pair appearing in the preset scenario and the second frequency of simultaneous appearance, the relevance score of the similar product word pair is calculated. The relevance calculation formula is as follows:

[0084]

[0085] Among them, F(product word a, product word b) is the frequency of product word a and product word b appearing at the same time in a preset scenario (i.e., the second frequency), F(product word b) is the total frequency of product word b appearing in the scenario (i.e., the first frequency corresponding to product word b), and F(product word a) is the total frequency of product word a appearing in the scenario (i.e., the first frequency corresponding to product word a), where product word a and product word b are two product words in any similar product word pair.

[0086] The larger the correlation score is, the more correlated product word a and product word b are in this scenario, and the more similar they are.

[0087] S403: Input the relevance score and the first representation vectors respectively corresponding to the two product words in the similar product word pair into a synonym discrimination model to discriminate whether the similar product word pair is a synonym product word pair.

[0088] In an embodiment of the present application, the association score of each similar product word pair and the first representation vectors corresponding to the two product words in the similar product word pair are used as inputs to a synonym discrimination model, and synonymous product word pairs are determined through the model, wherein the first representation vectors corresponding to the two product words in the similar product word pair are similar to the first representation vector of the first candidate product word mentioned above, which will not be repeated here.

[0089] In addition, if the original category system exists before the new product category system is constructed in this application, the quantity distribution characteristics of the product words in the original category system can also be added to the input of the model.

[0090] like Figure 5 As shown, the present disclosure is based on the Wide&Deep model as a synonym discrimination model, and the model includes a Deep side and a Wide side. On the Deep side of the model, the first representation vector corresponding to product word a and product word b, and the quantity distribution characteristics of product word a and product word b in the original category system are used as inputs, and a fully connected layer is used to reduce the dimension and extract the quantity distribution characteristics of product word a and product word b in the original category system; on the Wide side of the model, the correlation of similar product word pairs in a preset scenario is used as input, such as the statistical characteristics of the correlation of similar product word pairs under the commodity title library, search click behavior and recommendation click behavior. Among them, the quantity distribution in the original category system can be understood as: the distribution of product words under a category of a certain granularity, for example, the third-level category may include three thousand third-level categories, and based on the number of product words in these three thousand third-level categories, a distribution vector is formed, and each third-level category corresponds to a dimension, thus forming a three thousand-dimensional vector, and then the dimension is reduced through the fully connected layer.

[0091] Based on the above embodiments, Figure 6 As shown, in the above step S102, "building a commodity category system based on the hyponymous relationship between the first target product words" may include the following steps:

[0092] S601, determining any two first target product words that have a hyponymy relationship as a first hyponymy product word pair.

[0093] In an embodiment of the present application, for any two first target product words, it is determined whether there is a hierarchical relationship between the two first target product words, and the two first target product words with a hierarchical relationship are determined as the first hierarchical product word pair, such as "zongzi" and "rouzongzi".

[0094] In some embodiments, product word pairs that may have a hierarchical relationship can be recalled by two methods: vector recall and rule recall. Among them, the vector recall method is similar to the calculation process of the similarity score described above. The cosine value of the first characterization vector between product words is calculated two by two, and the product word pairs whose cosine value is greater than the threshold are recalled, that is, they are determined to be the first hierarchical product word pairs. The rule recall method is mainly based on the pattern method (pattern is the expression mode after compilation of the regular expression, which can be understood as a rule in this application such as: "xx pants" is "pants") to recall, and recall product word pairs that may have a hierarchical relationship according to the Chinese text structure. For example, the Chinese text structure can be a series of patterns such as "xx table" is a hyponym of "table" manually sorted, or a series of patterns such as "xx, containing aa, bb" can be set to indicate that xx is a hypernym of aa, bb, and then a fuzzy match is performed based on Baidu Encyclopedia or Wikipedia to obtain specific hierarchical product word pairs as the first hierarchical product word pairs.

[0095] S602: Perform relevance filtering on the plurality of first hyponymy and hyponymy product word pairs to determine second hyponymy and hyponymy product word pairs.

[0096] In the embodiment of the present application, relevance filtering can be performed based on at least one of the search log, the recommendation log, and the original category system, that is, the first hyponym and hyponym product word pairs are screened and filtered by judging at least one of the search behavior relevance, the recommendation behavior relevance, and the original category system relevance:

[0097] Based on the search log, the first hyponymy product word pair whose third frequency is greater than the second threshold is determined as the second hyponymy product word pair, and the third frequency is the frequency at which the search expressions corresponding to the two product words in the first hyponymy product word pair have intersections.

[0098] For example, it is checked based on the search log whether the frequency of intersection of all search expressions input by the user corresponding to the product word is greater than a preset threshold (ie, the third threshold).

[0099] Based on the recommendation log, the first hyponym product word pair whose fourth frequency is greater than the third threshold is determined as the second hyponym product word pair, and the fourth frequency is the frequency of the first hyponym product word pair appearing simultaneously in the product click sequence triggered by the recommended product. The recommended product may be a homepage product.

[0100] The third threshold and the fourth threshold can be set as needed, and this application does not limit them.

[0101] In response to the existence of an original category system before constructing the product category system of the present application, the first hierarchical product word pair in which the first product set and the second product set have an intersection is determined as the second hierarchical product word pair, wherein the first product set is a set of products covered by any product word in the first hierarchical product word pair in the original category system, and the second product set is a set of products covered by product words in the first hierarchical product word pair except any product word in the original category system.

[0102] S603: For each second hyponymy and hyponymy product word pair, determine the hypernymy product word and the hyponymy product word in the second hyponymy and hyponymy product word pair based on the relationship extraction model.

[0103] In the embodiment of the present application, the hypernymy relationship of each second hypernymy product word pair is confirmed based on the relationship extraction model, and the hypernymy product word and the hypernymy product word are distinguished. The relationship extraction model includes but is not limited to the knowledge graph embedding conversion model: TransR and TransD, and BERT and other models.

[0104] Take the TransR model as an example of a relation extraction model. Figure 7 As shown, the upper and lower relationships are expressed as<h,r,t> Triplet, where the triangle and circle in the figure represent product words in the second hyponymous product word pair, circle h is product word A, circle t is product word B, r is the relationship between product word A and product word B, product word A and product word B are two product words in a second hyponymous product word pair, Mr is the projection matrix of whether h is the hypernym of t, hr and tr are the representations of h and t in the hypernymous relationship space after being mapped by the projection matrix, respectively, where the projection matrix can be obtained through active learning model training.

[0105] Among them, the product word can be projected to the superordinate relationship space based on the second representation vector corresponding to the product word in the physical space and the projection matrix, so that different product words can be dispersed in order to better determine the superordinate and subordinate relationships between product words, wherein the second representation vector corresponding to the product word can be composed of the above-mentioned first representation vector (the concatenation vector of the semantic representation vector and the behavior representation vector), the quantity distribution vector of the product word in the original category system, and the frequency distribution vector of the product word and the other product word in the second superordinate and subordinate relationship product word pair appearing simultaneously in the original category system, wherein the quantity distribution vector and the frequency distribution vector are output to the same dimension as the first representation vector through a fully connected layer, and the second representation vector corresponding to the product word is obtained by concatenating the three vectors.

[0106] S604, determine whether the superordinate product word in the second superordinate product word pair is a cross-layer superordinate product word. If so, the second superordinate product word pair is discarded, where the cross-layer superordinate product word is a superordinate product word of multiple subordinate product words, and there is a superordinate relationship between the multiple subordinate product words. If not, the second superordinate product word pair is determined as the third superordinate product word pair.

[0107] Since a product word may have multiple superordinate product words, in order to ensure the tree structure of the category system, it is necessary to select the most appropriate superordinate product word for the product word. In the embodiment of the present application, the cross-layer superordinate product words that are both superordinate words of product word A and superordinate words of a superordinate word of product word A are removed to ensure a clear hierarchy. For example, C is the superordinate word of A, B is the superordinate word of A, and C is the superordinate word of B, then in the second superordinate and subordinate product word pair composed of C and A, C is the cross-layer superordinate word.

[0108] S605, calculating the relevance score of the third hyponymous and hyponymous product word pair.

[0109] In an embodiment of the present application, the correlation score can be calculated based on at least one of the frequency of simultaneous appearance of the third hyponymous and hyponymous product word pairs in the product title library, the association score in the product title library scenario, the association score in the search click behavior scenario, and the association score in the recommendation click behavior scenario.

[0110] For example, if the four indicators of the frequency of simultaneous appearance of the third hierarchical product word pair in the product title library, the correlation score in the product title library scenario, the correlation score in the search click behavior scenario, and the correlation score in the recommendation click behavior scenario are fused to obtain a comprehensive correlation score for the third hierarchical product word pair, the multi-indicator fusion method includes but is not limited to weighted fusion, hierarchical analysis method, entropy method, etc.

[0111] S606, selecting the third hyponymy product word pair with the highest correlation score from the third hyponymy product word pairs containing the same hyponymy product word, and determining it as the target hyponymy product word pair.

[0112] In the embodiment of the present application, the same subordinate product word may exist in multiple third and lower-level product word pairs, corresponding to different upper-level product words. For each subordinate product word, the third and lower-level product word pair with the highest correlation score is selected from the multiple third and lower-level product word pairs containing the subordinate product word, and the third and lower-level product word pair is determined as the target upper-level product word pair, that is, the upper-level product word in the target product word pair is the only upper-level word corresponding to the subordinate product word.

[0113] S607, constructing a product category system based on the target hyponymous and hyponymous product word pairs.

[0114] In the embodiment of the present application, based on the obtained multiple target hyponym and hyponym product word pairs, a hierarchical construction of commodity categories is performed to form a commodity category system.

[0115] Figure 8 A partial schematic diagram of a commodity category system of a method for determining commodity category information provided in an embodiment of the present application, such as Figure 8 As shown, fitness equipment is used as a category level, and the hyponyms of fitness equipment: supine board, push-up chair and inversion chair are used as a category level. The products can be attached to the corresponding categories based on the product words of the products. For example, the product with the product word "supine board" is attached to the category of fitness equipment < supine board, and the product corresponding to the sit-up board, which has a high synonymy with the supine board, can also be attached to the category of fitness equipment < supine board.

[0116] Based on the above embodiments, Fig. 9 As shown, in the above step S103, "determining the category information of the to-be-matched commodity according to the commodity category system" may include the following steps:

[0117] S901, obtaining a second target product word of the commodity to be matched.

[0118] In the embodiment of the present application, the to-be-matched product may be a newly-listed product to be registered by a platform user, or a product waiting to obtain category information. According to the text information corresponding to the to-be-matched product (such as the product title of the to-be-matched product, etc.), the second candidate product word and attribute information of the to-be-matched product are obtained; based on the conditional probability of the second candidate product word existing under the attribute information, the score of the second candidate product word is determined; and the second candidate product word with the highest score is determined as the second target product word.

[0119] Among them, the second candidate product word and attribute information can be obtained through part-of-speech tagging. The attribute information can be information such as the brand, color, style, function and specification of the product. The conditional probability of each second candidate product word under various attribute information is calculated based on the platform products, and then each second candidate product word is scored, and the second candidate product word with the highest score is used as the unique identification product word (second target product word) of the product to be matched. These conditional probabilities can be further stored as a probability dictionary for easy call.

[0120] The score of the second candidate product word can be determined based on the conditional probability by the following formula:

[0121]

[0122]

[0123] Among them, pt i is the i-th second candidate product word, attribute kis the kth attribute identified based on part-of-speech tagging, P(pt i ) is the probability of the i-th second candidate product word appearing on the e-commerce platform, that is, the number of products covered by the i-th second candidate product word divided by the total number of platform products, P (attribute j |pt i ) is the attribute j and pt i The number of simultaneous occurrences divided by pt i The number of occurrences is based on the naive Bayesian assumption that each attribute is independently distributed. In addition, for different second candidate product words of the same matching product, their P (attribute 1, attribute 2, ..., attribute k ) are the same, so the formula can be simplified and the numerators can be calculated and compared.

[0124] As a feasible implementation, image information may be added on the basis of the conditional probability to determine the score of the second candidate product word.

[0125] Obtain an image representation vector of the second candidate product word and an image representation vector of the product to be matched; calculate a cosine value between the image representation vector of the second candidate product word and the image representation vector of the product to be matched; and perform multi-index fusion of the cosine value and the conditional probability to obtain a score for the second candidate product word.

[0126] Each product word may correspond to multiple different commodities, and an image representation vector of the product word may be obtained based on the images of these commodities. The image representation vector of the second candidate product word may be obtained by the following process: obtaining multiple commodity images corresponding to the second candidate product word; obtaining a representation vector of the commodity image; and taking a weighted sum of the multiple representation vectors as the image representation vector of the second candidate product word.

[0127] S902, searching for product words in the commodity category system whose synonymy score with the second target product word is greater than a fourth threshold.

[0128] As a feasible implementation method, a product word having a synonymy score greater than a fourth threshold with the second target product word is searched from the commodity category system based on a synonym relationship classifier or the above-mentioned synonym discrimination model.

[0129] S903: Determine the category information corresponding to the product word as the category information of the commodity to be matched.

[0130] In some embodiments, the constructed product category system may contain a product word that is the same as the second target product word. If so, the category information corresponding to the product word can be used as the category information of the product to be matched, or the product to be matched can be directly attached to the category corresponding to the product word.

[0131] If it does not exist, the synonymy score between the second target product word and the product word in the commodity category system can be calculated. If the score is greater than the preset first threshold, the category information corresponding to the product word can be used as the category information of the commodity to be matched, or the commodity to be matched can be indirectly linked to the category corresponding to the product word. If the score does not reach the preset fourth threshold, the second target product word is regarded as a new word, and a new branch of the category system can be constructed for the word based on the above embodiment to determine the category information of the commodity to be matched.

[0132] In some embodiments, if the category system is updated on a weekly or monthly basis, the number of new words that cannot be directly or indirectly linked will be in the single or tens digits, so the category branches can also be added manually.

[0133] This application can perform full daily updates or incremental real-time acquisition of product categories as needed. Product category affiliation can be automated and real-time, and new product terms can be better mined and affiliating categories can be achieved.

[0134] This application provides a method for constructing a commodity product word category system from scratch at low cost for the construction of the existing commodity category system of e-commerce, as well as a method for determining the category information of commodities based on the commodity category system. The commodity product word category system is based on the definition of the commodity itself. Except for some one-time labeling data that relies on manual labor, its construction and operation are autonomous. Compared with the existing category systems that are more determined based on the scope of business operations, it is more objective and universal, and has higher applicability to scenarios such as search, recommendation and advertising. The product words, synonyms, and hierarchical relationship data accumulated in the process of building the category system are also an important data asset that can be provided to the corresponding demand parties.

[0135] Fig.10 This is an overall schematic diagram of a method for determining commodity category information provided by an embodiment of the present application. Fig.10 As shown, the method for determining the commodity category information in the embodiment of the present application may specifically include the following steps:

[0136] S1001, performing part-of-speech tagging on text information corresponding to the sample product based on preset part-of-speech tags.

[0137] S1002, screening out a plurality of first candidate product words from the text information after part-of-speech tagging.

[0138] S1003, obtaining a semantic representation vector and a behavior representation vector of the first candidate product word.

[0139] S1004: Concatenate the semantic representation vector and the behavior representation vector to generate a first representation vector.

[0140] S1005: Determine a similarity score between any two first candidate product words based on the first representation vector, and determine any two first candidate product words with a similarity score greater than a first threshold as a similar product word pair.

[0141] S1006, based on the synonym discrimination model, multiple synonymous product word pairs are selected from multiple similar product word pairs.

[0142] S1007, normalize each synonymous product word pair into a second candidate product word.

[0143] S1008: Select multiple first target product words from multiple second candidate product words based on the commodity coverage of the second candidate product word and the frequency of occurrence of the second candidate product word in the search behavior.

[0144] S1009: Determine any two first target product words that have a hyponymy relationship as a first hyponymy product word pair.

[0145] S1010, performing relevance filtering on a plurality of first hyponymy and hyponymy product word pairs to determine second hyponymy and hyponymy product word pairs.

[0146] S1011, for each second hyponymy and hyponymy product word pair, determine the hypernymy product word and the hyponymy product word in the second hyponymy and hyponymy product word pair based on the relationship extraction model.

[0147] S1012, determine whether the superordinate product word in the second superordinate product word pair is a cross-layer superordinate product word. If so, the second superordinate product word pair is discarded, where the cross-layer superordinate product word is a superordinate product word of multiple subordinate product words, and there is a superordinate relationship between the multiple subordinate product words. If not, the second superordinate product word pair is determined as the third superordinate product word pair.

[0148] S1013, calculate the correlation score according to at least one of the frequency of simultaneous appearance of the third hyponymous product word pair in the product title library, the association score in the product title library scenario, the association score in the search click behavior scenario, and the association score in the recommendation click behavior scenario.

[0149] S1014, selecting the third hyponymy product word pair with the highest correlation score from the third hyponymy product word pairs containing the same hyponymy product word, and determining it as the target hyponymy product word pair.

[0150] S1015, construct a product category system based on target hyponymous and hyponymous product word pairs.

[0151] S1016, obtaining the second target product word of the commodity to be matched.

[0152] S1017, searching for product words in the commodity category system whose synonymy score with the second target product word is greater than a fourth threshold.

[0153] S1018: Determine the category information corresponding to the product word as the category information of the commodity to be matched.

[0154] Fig.11 This is a block diagram of a device for determining commodity category information proposed in this application, such as Fig.11 As shown, the device 1100 for determining the commodity category information includes: an acquisition module 1101, a construction module 1102 and a determination module 1103, wherein:

[0155] The acquisition module 1101 is used to acquire multiple first target product words corresponding to the sample products.

[0156] The construction module 1102 is used to construct a commodity category system based on the hyponymous and hyponymous relationships between the first target product words.

[0157] The determination module 1103 is used to determine the category information of the commodity to be matched according to the commodity category system.

[0158] According to one embodiment of the present application, the acquisition module 1011 can be further used to: perform part-of-speech tagging on the text information corresponding to the sample product based on preset part-of-speech tags; filter out multiple first candidate product words from the text information after part-of-speech tagging; and perform synonymous merging on the multiple first candidate product words to obtain multiple first target product words.

[0159] According to one embodiment of the present application, the acquisition module 1011 can be further used to: determine multiple similar product word pairs based on the similarity scores between the first candidate product words; screen out multiple synonymous product word pairs from the multiple similar product word pairs based on the synonym discrimination model; normalize each synonymous product word pair into a second candidate product word; and select multiple first target product words from the multiple second candidate product words based on the product coverage of the second candidate product word and the frequency of occurrence of the second candidate product word in the search behavior.

[0160] According to one embodiment of the present application, the acquisition module 1011 is further used to: obtain a first representation vector of the first candidate product word; determine a similarity score between any two first candidate product words based on the first representation vector; and determine any two first candidate product words whose similarity scores are greater than a first threshold as a similar product word pair.

[0161] According to one embodiment of the present application, the acquisition module 1011 can be further used to: acquire a semantic representation vector of the first candidate product word; acquire a behavioral representation vector of the first candidate product word; and concatenate the semantic representation vector and the behavioral representation vector to generate a first representation vector.

[0162] According to one embodiment of the present application, the acquisition module 1011 is further used to: obtain the product title of the product clicked by the user under the search behavior corresponding to the first candidate product word; extract keywords from the product title to obtain keyword representation segmentation; obtain the semantic representation vector of the keyword representation segmentation; and use the weighted sum of the semantic representation vectors of the keyword representation segmentation as the behavior representation vector of the first candidate product word.

[0163] According to one embodiment of the present application, the acquisition module 1011 can be further used to: obtain the first frequency of two product words in the similar product word pair appearing in a preset scenario respectively, and the second frequency of the similar product word pair appearing at the same time in the preset scenario; calculate the association score of the similar product word pair based on the first frequency and the second frequency; input the association score and the first representation vectors corresponding to the two product words in the two similar product word pairs into the synonym discrimination model to determine whether the similar product word pairs are synonymous product word pairs.

[0164] According to one embodiment of the present application, the preset scenario includes at least one of the following: a product title library, a search click behavior, and a recommendation click behavior.

[0165] According to one embodiment of the present application, construction module 1102 can be further used to: determine any two first target product words that have a hierarchical relationship as a first hierarchical product word pair; perform relevance filtering on multiple first hierarchical product word pairs to determine a second hierarchical product word pair; for each second hierarchical product word pair, determine the superordinate product word and the subordinate product word in the second hierarchical product word pair based on the relationship extraction model; determine whether the superordinate product word in the second hierarchical product word pair is a cross-layer superordinate product word; if so, discard the second hierarchical product word pair, wherein the cross-layer superordinate product word is a superordinate product word of multiple subordinate product words, and there is a hierarchical relationship between the multiple subordinate product words; if not, determine the second hierarchical product word pair as a third hierarchical product word pair; calculate the relevance score of the third hierarchical product word pair; select the third hierarchical product word pair with the highest relevance score from the third hierarchical product word pairs containing the same subordinate product word, and determine it as the target hierarchical product word pair; and construct a product category system based on the target hierarchical product word pairs.

[0166] According to one embodiment of the present application, the construction module 1102 can be further used to: determine the first hierarchical product word pair whose third frequency is greater than the second threshold as the second hierarchical product word pair based on the search log, the third frequency being the frequency at which the search expressions corresponding to the two product words in the first hierarchical product word pair respectively have an intersection; determine the first hierarchical product word pair whose fourth frequency is greater than the third threshold as the second hierarchical product word pair based on the recommendation log, the fourth frequency being the frequency at which the first hierarchical product word pair appears simultaneously in the product click sequence triggered by the recommended product; and in response to the existence of the original category system before the construction of the product category system, determine the first hierarchical product word pair whose first product set and the second product set have an intersection as the second hierarchical product word pair, the first product set being the set of products covered by any product word in the first hierarchical product word pair in the original category system, and the second product set being the set of products covered by product words in the first hierarchical product word pair except any product word in the original category system.

[0167] According to one embodiment of the present application, construction module 1102 is further used to calculate the correlation score based on at least one of the frequency of simultaneous appearance of the third hyponymous and hyponymous product word pairs in the product title library, the association score in the product title library scenario, the association score in the search click behavior scenario, and the association score in the recommendation click behavior scenario.

[0168] According to one embodiment of the present application, the determination module 1103 is further used to: obtain a second target product word for the commodity to be matched; search for a product word in the commodity category system whose synonymy score with the second target product word is greater than a fourth threshold; and determine the category information corresponding to the product word as the category information of the commodity to be matched.

[0169] According to one embodiment of the present application, the determination module 1103 can be further used to: obtain the second candidate product word and attribute information of the to-be-matched product based on the text information corresponding to the to-be-matched product; determine the score of the second candidate product word based on the conditional probability of the second candidate product word existing under the attribute information; and determine the second candidate product word with the highest score as the second target product word.

[0170] According to one embodiment of the present application, the determination module 1103 can be further used to: obtain the image representation vector of the second candidate product word and the image representation vector of the product to be matched; calculate the cosine value between the image representation vector of the second candidate product word and the image representation vector of the product to be matched; and perform multi-indicator fusion of the cosine value and the conditional probability to obtain the score of the second candidate product word.

[0171] It should be noted that the above explanation of the embodiment of the method for determining product category information is also applicable to the device for determining product category information of this embodiment, and the specific process will not be repeated here.

[0172] The embodiment of the present application proposes a device for determining commodity category information, which obtains multiple first target product words corresponding to sample commodities, constructs a commodity category system based on the hierarchical relationship between the first target product words, and determines the category information of the commodity to be matched according to the commodity category system. The product word of a commodity is used to define the commodity, and can reflect the category and essential attributes of the commodity. The method constructs a commodity category system based on the product word of the commodity. The constructed commodity category system contains the category information of all sample commodities. These categories are divided based on the definition of the commodity itself, and can be better applied to other new e-commerce scenarios such as search and recommendation. The commodity to be matched can be matched with the category information of the sample commodity based on the commodity category system, so as to determine the category information of the commodity to be matched, and reduce the cost of determining and updating the commodity category information.

[0173] In order to implement the above embodiment, the present application embodiment also proposes an electronic device 1200, such as Fig.12 As shown, the electronic device 1200 includes: a processor 1201 and a memory 1202 communicatively connected to the processor, the memory 1202 stores instructions executable by at least one processor, and the instructions are executed by at least one processor 1201 to implement the method for determining product category information as shown in the above embodiment.

[0174] In order to implement the above embodiment, the embodiment of the present application also proposes a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable the computer to implement the method for determining the commodity category information as shown in the above embodiment.

[0175] In order to implement the above embodiments, the embodiments of the present application further propose a computer program product, including a computer program, which implements the method for determining the commodity category information as shown in the above embodiments when executed by a processor.

[0176] In the description of the present application, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0177] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0178] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0179] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A method for determining commodity category information, characterized in that: include: Obtain multiple first target product words corresponding to the sample products; Building a product category system based on the hyponymous and hyponymous relationships between the first target product words; Determining category information of the product to be matched according to the product category system; The step of constructing a commodity category system based on the hyponymous and hyponymous relationships between the first target product words includes: Determine any two of the first target product words that have a hyponymy relationship as a first hyponymy product word pair; Performing relevance filtering on the plurality of first hyponymous and hyponymous product word pairs to determine second hyponymous and hyponymous product word pairs; For each of the second hyponymous and hyponymous product word pairs, determining a hypernymous product word and a hyponymous product word in the second hyponymous and hyponymous product word pair based on a relationship extraction model; Determine whether the superordinate product word in the second superordinate product word pair is a cross-layer superordinate product word. If so, discard the second superordinate product word pair, wherein the cross-layer superordinate product word is a superordinate product word of multiple subordinate product words, and there is a superordinate relationship between the multiple subordinate product words. If not, determine the second superordinate product word pair as the third superordinate product word pair; Calculating the correlation scores of the third hyponymy and hyponymy product word pairs, and selecting the third hyponymy and hyponymy product word pair with the highest correlation score from the third hyponymy and hyponymy product word pairs containing the same hyponymy product word, and determining it as the target hyponymy and hyponymy product word pair; The commodity category system is constructed based on the target hyponymous and hyponymous product word pairs.

2. The determination method according to claim 1, characterized in that: The obtaining of a plurality of first target product words corresponding to the sample product includes: Performing part-of-speech tagging on the text information corresponding to the sample product based on preset part-of-speech tags; Filtering out a plurality of first candidate product words from the text information after part-of-speech tagging; Synonymically merge the plurality of the first candidate product words to obtain a plurality of the first target product words.

3. The determination method according to claim 2, characterized in that: The synonymous merging of the plurality of the first candidate product words to obtain the plurality of the first target product words includes: Determining a plurality of similar product word pairs according to the similarity scores between the first candidate product words; Based on the synonym discrimination model, a plurality of synonymous product word pairs are screened out from the plurality of similar product word pairs; normalizing each of the synonymous product word pairs into a second candidate product word; According to the commodity coverage of the second candidate product word and the frequency of occurrence of the second candidate product word in the search behavior, multiple first target product words are selected from multiple second candidate product words.

4. The determination method according to claim 3, characterized in that: The determining a plurality of similar product word pairs according to the similarity scores between the first candidate product words includes: Obtaining a first representation vector of the first candidate product word; Determine the similarity score between any two of the first candidate product words based on the first representation vector; Any two of the first candidate product words whose similarity scores are greater than a first threshold are determined as the similar product word pair.

5. The determination method according to claim 4, characterized in that: The obtaining of a first representation vector of the first candidate product word includes: Obtaining a semantic representation vector of the first candidate product word; Obtaining a behavior representation vector of the first candidate product word; The semantic representation vector and the behavior representation vector are concatenated to generate the first representation vector.

6. The determination method according to claim 5, characterized in that: The step of obtaining the behavior representation vector of the first candidate product word includes: Obtaining the product title of the product clicked by the user in the search behavior corresponding to the first candidate product word; Extract keywords from the product title to obtain keyword representation segmentation; Obtaining a semantic representation vector of the keyword representation segmentation; The weighted sum of the semantic representation vectors of the keyword representation segmentation is used as the behavior representation vector of the first candidate product word.

7. The determination method according to claim 3, characterized in that: The method of selecting a plurality of synonymous product word pairs from the plurality of similar product word pairs based on the synonym discrimination model includes: Obtaining first frequencies of two product words in the similar product word pair appearing in preset scenarios respectively, and second frequencies of the similar product word pair appearing simultaneously in the preset scenarios; Calculating, according to the first frequency and the second frequency, a relevance score of the similar product word pair in the preset scenario; The relevance score and the first representation vectors respectively corresponding to the two product words in the similar product word pair are input into the synonym discrimination model to discriminate whether the similar product word pair is the synonymous product word pair.

8. The determination method according to claim 7, characterized in that: The preset scenario includes at least one of the following: a product title library, a search click behavior, and a recommendation click behavior.

9. The determination method according to claim 1, characterized in that: The step of performing correlation filtering on the plurality of the first hyponymy and hyponymy product word pairs to determine the second hyponymy and hyponymy product word pairs includes at least one of the following: Based on the search log, the first hyponymy product word pair whose third frequency is greater than the second threshold is determined as the second hyponymy product word pair, wherein the third frequency is the frequency at which the search expressions corresponding to the two product words in the first hyponymy product word pair respectively have an intersection; Based on the recommendation log, the first hyponymy product word pair whose fourth frequency is greater than a third threshold is determined as the second hyponymy product word pair, wherein the fourth frequency is the frequency at which the first hyponymy product word pair appears simultaneously in a product click sequence triggered by the recommended product; as well as In response to the existence of an original category system before the product category system is constructed, the first hierarchical product word pairs in which the first product set and the second product set have an intersection are determined as the second hierarchical product word pairs. The first product set is a set of products covered by any product word in the first hierarchical product word pairs in the original category system, and the second product set is a set of products covered by product words in the first hierarchical product word pairs except any product word in the original category system.

10. The determination method according to claim 1, characterized in that: The calculating the relevance score of the third hyponymous and hyponymous product word pair includes: The correlation score is calculated based on at least one of the frequency of simultaneous appearance of the third hyponymous and hyponymous product word pairs in the product title library, the association score in the product title library scenario, the association score in the search click behavior scenario, and the association score in the recommendation click behavior scenario.

11. The determination method according to claim 1, characterized in that: Determining the category information of the to-be-matched commodity according to the commodity category system includes: Obtaining a second target product word of the commodity to be matched; Searching for a product word in the commodity category system whose synonymy score with the second target product word is greater than a fourth threshold; The category information corresponding to the product word is determined as the category information of the commodity to be matched.

12. The determination method according to claim 11, characterized in that: The step of obtaining the second target product word of the commodity to be matched includes: Acquire a second candidate product word and attribute information of the to-be-matched product according to the text information corresponding to the to-be-matched product; determining a score of the second candidate product word based on a conditional probability that the second candidate product word exists under the attribute information; The second candidate product word with the highest score is determined as the second target product word.

13. The determination method according to claim 12, characterized in that: The determining the score of the second candidate product word based on the conditional probability that the second candidate product word exists under the attribute information includes: Obtaining an image representation vector of the second candidate product word and an image representation vector of the commodity to be matched; Calculating a cosine value between the image representation vector of the second candidate product word and the image representation vector of the commodity to be matched; The cosine value and the conditional probability are integrated into a multi-index manner to obtain a score for the second candidate product word.

14. The determination method according to claim 13, characterized in that: The step of obtaining the image representation vector of the second candidate product word includes: Acquire multiple product images corresponding to the second candidate product word; Obtaining a representation vector of the product image; The weighted sum of the multiple representation vectors is used as the image representation vector of the second candidate product word.

15. A device for storing commodity category information, characterized in that: include: An acquisition module, used to acquire multiple first target product words corresponding to the sample product; A construction module, used to construct a commodity category system based on the hyponymous and hyponymous relationships between the first target product words; A determination module, used to determine the category information of the commodity to be matched according to the commodity category system; The building block is further used for: Determine any two of the first target product words that have a hyponymy relationship as a first hyponymy product word pair; Performing relevance filtering on the plurality of first hyponymous and hyponymous product word pairs to determine second hyponymous and hyponymous product word pairs; For each of the second hyponymous and hyponymous product word pairs, determining a hypernymous product word and a hyponymous product word in the second hyponymous and hyponymous product word pair based on a relationship extraction model; Determine whether the superordinate product word in the second superordinate product word pair is a cross-layer superordinate product word. If so, discard the second superordinate product word pair, wherein the cross-layer superordinate product word is a superordinate product word of multiple subordinate product words, and there is a superordinate relationship between the multiple subordinate product words. If not, determine the second superordinate product word pair as the third superordinate product word pair; Calculating the correlation scores of the third hyponymy and hyponymy product word pairs, and selecting the third hyponymy and hyponymy product word pair with the highest correlation score from the third hyponymy and hyponymy product word pairs containing the same hyponymy product word, and determining it as the target hyponymy and hyponymy product word pair; The commodity category system is constructed based on the target hyponymous and hyponymous product word pairs.

16. The device according to claim 15, characterized in that The determining module is further used for: Obtaining a second target product word of the commodity to be matched; Searching for a product word in the commodity category system whose synonymy score with the second target product word is greater than a fourth threshold; The category information corresponding to the product word is determined as the category information of the commodity to be matched.

17. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 14.

18. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-14.

19. A computer program product comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 14.

Images