CN113256379B

CN113256379B - A method for associating shopping needs with commodities

Info

Publication number: CN113256379B
Application number: CN202110564384.XA
Authority: CN
Inventors: 华钰; 陈帅; 彭力
Original assignee: Beijing Xiaomi Mobile Software Co Ltd; Beijing Xiaomi Pinecone Electronic Co Ltd
Current assignee: Beijing Xiaomi Mobile Software Co Ltd; Beijing Xiaomi Pinecone Electronic Co Ltd
Priority date: 2021-05-24
Filing date: 2021-05-24
Publication date: 2024-12-20
Anticipated expiration: 2041-05-24
Also published as: CN113256379A

Abstract

One or more embodiments of the present specification provide a method of associating shopping requirements for an item. The commodity title sample can be obtained, the trade names contained in the commodity title sample are named entities, shopping requirements met by the commodity names in the commodity title sample are named entity types, the NER algorithm is identified based on the named entity types, model training is conducted, and a shopping requirement identification model is obtained. Therefore, the shopping demand recognition model can be utilized to correlate the commodities of the electronic commerce platform to the shopping demands, and the correlation can be utilized to recommend the commodities correlated with the shopping demands for the user on the premise of knowing the shopping demands of the user.

Description

Method for associating shopping requirements for commodities

Technical Field

One or more embodiments of the present disclosure relate to the field of information technology, and more particularly, to a method for associating shopping needs for merchandise.

Background

In practical applications, users often do not use e-commerce platforms with the purpose of explicitly purchasing a particular item. The user will typically enter search terms into the e-commerce platform, such as "how to expel insects", "what needs to be used for kitchen cleaning", etc., which generally only indicate certain shopping needs of the user, such as expelling insects, kitchen cleaning, etc., but not specifically directed to a certain type of merchandise.

Assuming that the e-commerce platform can determine the shopping requirement of the user according to the search statement input by the user, how to recommend the commodity capable of meeting the shopping requirement to the user is a technical problem to be solved.

Disclosure of Invention

To overcome the problems in the related art, the present disclosure provides a method for associating shopping requirements for goods, a method for recommending goods for users, and apparatuses, electronic devices, and storage media.

According to a first aspect of embodiments of the present disclosure, there is provided a method of associating shopping requirements for a commodity, the method comprising:

acquiring a commodity title corresponding to a commodity of the commodity platform for commodities to be associated with shopping requirements;

inputting the acquired commodity titles into a shopping demand identification model, and identifying shopping demands met by commodity names in the acquired commodity titles;

establishing an association between the identified shopping need and the merchandise;

The training method of the shopping demand identification model comprises the steps of obtaining a commodity title sample, taking a commodity name contained in the commodity title sample as a named entity, taking shopping demands met by commodity names in the commodity title sample as named entity types, and carrying out model training based on a named entity type identification NER algorithm.

According to a second aspect of the present disclosure, a method for recommending commodities to a user is disclosed, and the method is applied to an e-commerce platform, and comprises:

acquiring a search statement input by a user, and determining a matched shopping requirement according to the search statement;

recommending at least part of commodities associated with the determined shopping requirements to the user, wherein the association between the shopping requirements and the commodities is determined by the method of the 1 st aspect.

According to a third aspect of the present disclosure, an apparatus for associating shopping needs for merchandise is disclosed, comprising:

the acquisition module is used for acquiring commodity titles corresponding to commodities to be associated with shopping demands of the electronic commerce platform;

the identification module inputs the acquired commodity titles into a shopping demand identification model, and identifies shopping demands satisfied by commodity names in the acquired commodity titles;

an association module that establishes an association between the identified shopping need and the commodity;

According to a fourth aspect of the present disclosure, an apparatus for recommending goods for a user is disclosed, applied to an e-commerce platform, the apparatus comprising:

the acquisition module acquires a search statement input by a user and determines a matched shopping requirement according to the search statement;

And the recommending module is used for recommending at least part of commodities associated with the determined shopping demands to the user, wherein the association between the shopping demands and the commodities is determined by the method of the first aspect.

According to a fifth aspect of the present disclosure, an electronic device is disclosed, comprising a processor, a memory for storing processor executable instructions, wherein the processor is configured to execute the instructions to implement the methods of the first and second aspects of the claims.

According to a sixth aspect of the present disclosure, a non-transitory computer readable storage medium is disclosed, having stored thereon a computer program which, when executed by a processor, implements the methods of the first and second aspects.

In the above technical solution, the problem of "how to recommend goods for users knowing the shopping demands of users" is converted into the problem of "related shopping demands for goods". The problem of "how to associate shopping demands for commodities" is further converted into a problem of "how to identify a named entity from a commodity title (commodity names in commodity titles are defined as named entities)".

In this way, a plurality of commodity title samples can be obtained, the trade names contained in each commodity title sample are named entities, shopping requirements met by the commodity names in each commodity title sample are named entity types, NER algorithm is identified based on the named entity types, and model training is carried out to obtain a shopping requirement identification model. Therefore, the shopping demand recognition model can be utilized to correlate the commodities of the electronic commerce platform to various shopping demands, and the correlation can be utilized to recommend commodities with higher matching degree of the shopping demands for the user on the premise of knowing the shopping demands of the user.

Drawings

Fig. 1 is a schematic diagram of a user searching using an e-commerce platform provided in the present specification.

FIG. 2 is a flow chart of a method for associating shopping requirements for merchandise according to an embodiment of the present disclosure.

FIG. 3 is a flow chart of a method for determining shopping needs of a user according to an embodiment of the present disclosure.

Fig. 4 is a schematic structural diagram of a demand keyword recognition model provided in the present specification.

Fig. 5 is a schematic diagram of an architecture of an integrated model according to an embodiment of the present disclosure.

FIG. 6 is a schematic diagram of an apparatus for associating shopping needs for merchandise provided herein.

Fig. 7 is a schematic structural diagram of an apparatus for recommending goods to a user according to the present disclosure.

FIG. 8 is a schematic diagram of an apparatus for determining shopping needs of a user provided herein.

FIG. 9 is a schematic diagram of an apparatus for determining a demand characterization word for shopping demand provided in the present specification.

Fig. 10 is a schematic structural diagram of an apparatus for determining similarity between demand keywords provided in the present description.

Fig. 11 is a schematic diagram of an electronic device 1500, according to an example embodiment.

Fig. 12 is a schematic diagram of another electronic device 1600, according to an example embodiment.

Detailed Description

Reference will now be made in detail to exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numbers in different drawings refer to the same or similar elements, unless otherwise indicated. The implementations described in the following exemplary examples are not representative of all implementations consistent with the present disclosure. Rather, they are merely examples of apparatus and methods consistent with some aspects of the present disclosure as detailed in the accompanying claims.

The terminology used in the present disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. As used in this disclosure and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used herein refers to and encompasses any or all possible combinations of one or more of the associated listed items.

It should be understood that although the terms first, second, third, etc. may be used in this disclosure to describe various information, these information should not be limited to these terms. These terms are only used to distinguish one type of information from another. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of the present disclosure. The term "if" as used herein may be interpreted as "at..once" or "when..once" or "in response to a determination", depending on the context.

For the e-commerce platform, firstly, the shopping demand of a user is determined according to a search statement input by the user, and then, a plurality of commodities capable of meeting the shopping demand are recommended to the user according to the shopping demand of the user. It will be appreciated that the e-commerce platform often desires to determine that the shopping needs based on the search statement are accurate, and also desires to match the user with the shopping needs of the user as well as being as comprehensive as possible (i.e., covering as many categories of merchandise as possible) based on the merchandise recommended to the user.

The disclosure proposes a technical scheme for recommending commodities to a user on the premise of knowing shopping requirements, and specifically:

the problem of knowing the shopping requirements of the user and recommending goods for the user is converted into the problem of associating the shopping requirements for the goods. The problem of "how to associate shopping demands for commodities" is further converted into a problem of "how to identify a named entity from a commodity title (commodity names in commodity titles are defined as named entities)".

In this way, a plurality of commodity title samples can be obtained, the trade names contained in each commodity title sample are named entities, shopping requirements met by the commodity names in each commodity title sample are named entity types, NER algorithm is identified based on the named entity types, and model training is carried out to obtain a shopping requirement identification model. Therefore, the shopping demand recognition model can be utilized to correlate the commodities of the electronic commerce platform to various shopping demands, the correlation can be utilized, the commodities correlated with the shopping demands are recommended to the user on the premise of knowing the shopping demands of the user, and the shopping demands of the user can be met as comprehensively as possible by the commodities recommended to the user.

Fig. 1 is a schematic diagram of a user searching using an e-commerce platform provided in the present specification. As shown in FIG. 1, a user inputs a search statement of "household cleaning" into a search box in a page, and an e-commerce platform presents and recommends a plurality of commodities meeting shopping requirements expressed by the search statement for the user.

FIG. 2 is a flow chart of a method for associating shopping requirements for merchandise according to an embodiment of the present disclosure, comprising the following steps:

S200, acquiring commodity titles corresponding to commodities to be associated with shopping demands of the electronic commerce platform.

In the e-commerce field, merchants typically write a commodity name in the title of a commodity purchase page (i.e., commodity title). Sometimes, in order to raise the probability of searching for the commodity, the merchant lists commodity names in a plurality of different expression modes in the commodity title. For example, names of electric kettles, electric kettles and the like correspond to one commodity.

S202, inputting the acquired commodity titles into a shopping demand identification model, and identifying shopping demands met by each commodity name in the acquired commodity titles.

When a commodity needs to be associated with a matching shopping demand, the title of the commodity is input to a shopping demand identification model.

The training method of the shopping demand identification model can comprise the steps of obtaining a commodity title sample, taking a commodity name contained in the commodity title sample as a named entity, taking shopping demands met by commodity names in the commodity title sample as named entity types, and carrying out model training based on a named entity type identification NER algorithm. In practical application, there are usually a plurality of obtained commodity title samples, and each commodity name contained in each commodity title sample is called a named entity, and shopping requirements satisfied by each commodity name in each commodity title sample are used as named entity types.

In other words, the model may be input for each commodity title sample, and the model may be output for a named entity tag corresponding to each commodity name in the commodity title sample.

There are a variety of NER algorithms. In some embodiments, the BERT algorithm, biLSTM algorithm, and CRF algorithm may be combined to implement the NER algorithm, and a model including the BERT algorithm layer, biLSTM algorithm layer, and CRF algorithm layer connected in sequence may be constructed for training.

In some embodiments, a model including a word2vec algorithm layer, a BiLSTM algorithm layer and a CRF algorithm layer connected in sequence may also be constructed to train by combining the word2vec algorithm, biLSTM algorithm and CRF algorithm.

It should be noted that, in the two models described above, the first layer of the model is generally used for vectorizing (or coding) each chinese character in the inputted commodity heading sample, so that the commodity heading sample represented by the vector enters BiLSTM algorithm layer for training or recognition. The BiLSTM algorithm layer can learn which information and which information is forgotten in the training process, so that the dependency relationship with a longer distance can be better captured, and the BiLSTM algorithm layer can learn semantic information in the front and back directions of sentences.

Because the output of BiLSTM algorithm layers in the model application stage is the probability that each word in a sentence is a named entity of each type, the probabilities corresponding to each word respectively lack a connection (the probability that a single word is considered as a granularity and belongs to a part of the named entity) and the probability that a word is considered as a granularity and belongs to the named entity. This easily results in the named entity, not the desired commodity name, being identified from the output of the BiLSTM algorithm layer alone for each word component that is part of the named entity.

Therefore, the output of BiLSTM algorithm layer is further input to CRF algorithm layer, the CRF algorithm can consider the connection between single words, the probability of various possible combinations belonging to named entities can be analyzed, the word combination (i.e. word) with the highest probability of each type of named entity can be determined according to the output of CRF algorithm layer, and the named entity type corresponding to the highest probability in the probabilities represents the shopping requirement satisfied by commodity name.

In addition, when training the shopping demand recognition model, usually, B, I, O labels are adopted to name and label the commodity names in the search statement sample. The first word of the requirement keyword is marked by B, the other words of the requirement keyword are marked by I, and the part irrelevant to the requirement keyword in the search sentence is marked by O. A search term may contain more than one demand keyword. When the labeling is actually performed, the label is followed by a named entity type, such as B-S (S represents the named entity type) and I-S. In the method flow shown in fig. 3, there are a plurality of named entity types (corresponding to a plurality of shopping requirements respectively).

In addition, in the process of training a shopping demand recognition model, an actual commodity title can be obtained, at least one other commodity name meeting the shopping demand is determined according to the shopping demand met by the commodity names in the actual commodity title, and the commodity title sample is obtained by adding the at least one other commodity name into the actual commodity title.

In practical applications, a plurality of actual product titles may be obtained from a database of an e-commerce platform. For each actual commodity title, determining at least one other commodity name meeting the shopping requirement according to the shopping requirement met by each commodity name in the actual commodity title, and adding the at least one other commodity name into the actual commodity title to obtain a commodity title sample. Since there may be a plurality of actual commodity titles acquired, one commodity title sample can be obtained for each actual commodity title, and thus a plurality of commodity title samples can be obtained.

Therefore, the data volume can be expanded under the condition of limited labeling manpower, and the model can learn shopping requirements expected to be identified more specifically.

And S204, establishing an association between the identified shopping requirements and the commodity.

In this way, in the application stage, the search statement input by the user can be obtained, and the matched shopping requirement is determined according to the search statement. At least a portion of the merchandise associated with the determined shopping need is then recommended to the user.

In addition, in some embodiments, the matching shopping requirement is determined according to the search statement, specifically, a requirement keyword is extracted from the search statement, and the matching shopping requirement is determined according to the extracted requirement keyword.

In some embodiments, a corresponding demand characterization word may be pre-assigned to each shopping demand. In this way, if the requirement characterization word which is the same as the extracted requirement keyword exists, the shopping requirement corresponding to the requirement characterization word is determined to be the matched shopping requirement. If the demand characterization words which are the same as the extracted demand keywords do not exist, the demand characterization words which have the closest meaning with the extracted demand keywords are determined, and shopping demands corresponding to the demand characterization words are determined to be matched shopping demands.

When determining the demand characterization word closest to the meaning of the extracted demand keyword, actually, performing similarity calculation on the extracted demand keyword and each demand testimony one by one, and taking the demand testimony corresponding to the calculated maximum similarity as the demand characterization word closest to the meaning of the extracted demand keyword. The method for calculating the similarity in meaning between two words used herein may be a method well known to those skilled in the art, or may be a similarity calculation method as given herein.

In some embodiments, after determining the requirement characterization word corresponding to the search statement, if the requirement characterization word is found to be inconsistent with the requirement keyword in the search statement input by the user, the e-commerce platform may prompt the requirement characterization word to the user, and after clicking the requirement characterization word, the e-commerce platform recommends the commodity associated with the requirement characterization word to the user for purchase.

It should be noted that, the method for extracting the requirement keywords may be implemented by using a requirement keyword recognition model described later, or may be implemented by using other methods. For example, the following method may be used to extract the demand keywords:

And segmenting the search sentence, and inputting each segmented word into a word classification model. And determining matched shopping requirements according to the word segmentation classified as the requirement keywords.

The training method of the word classification model comprises the steps of obtaining a plurality of search sentence samples, and extracting a requirement keyword from each search sentence sample. And taking each extracted keyword as a white sample, and performing model training by adopting a classification algorithm. The black samples may be specified according to the actual situation.

If the required keywords are extracted from the search sentences through the word classification model, the search sentences input by the user are required to be segmented, each segmented word is input into the word classification model one by one, and whether each segmented word is the required keywords is judged.

In addition, the word classification model may include a BERT algorithm layer, a BiLSTM algorithm layer, and a full connection layer connected in sequence. The BERT algorithm layer may be replaced by another algorithm layer for vectorizing the chinese characters in the words, such as word2vec algorithm layer.

In the embodiment where the word classification model includes the BERT algorithm layer, the text classification output by BERT may be used to represent the vector representation corresponding to CLS, so that the semantic information of each word in the word may be fused more "fairly". The vector representation corresponding to the CLS is also input to the fully connected layer, so that the fully connected layer synthesizes the text representation of the word to be recognized and the output of BiLSTM on the whole, and determines the classification.

The above scheme for extracting the requirement keywords from the search sentences based on the word classification model is to actually convert the task of extracting the requirement keywords from the search sentences into a problem of judging whether each word in the search sentences belongs to the requirement keywords one by one. And inputting each word in the search sentence into the word classification model, and judging whether each word in the search sentence belongs to the required keyword or not, so that the aim of identifying the required keyword from the search sentence is fulfilled. And then, according to the identified demand keywords, determining shopping demands of the users, and recommending commodities meeting the shopping demands for the users.

In addition, the present disclosure also proposes a technical solution for how to determine shopping requirements according to search sentences, specifically:

The problem of how to determine shopping demands according to the search sentences is converted into the problem of how to extract demand keywords from the search sentences, and the demand keywords can accurately reflect the shopping demands. The problem of "how to extract a demand keyword from a search term" is converted into a problem of "how to identify a named entity from a search term (the demand keyword in the search term is defined as a named entity)".

Therefore, a plurality of search statement samples can be obtained, the requirement keywords contained in each search statement sample are used as named entities, a NER algorithm is identified based on the named entities, model training is conducted, and a requirement keyword identification model is obtained. Therefore, the demand keyword recognition model can be utilized to recognize the demand keyword from the search statement input by the user, and shopping demands can be accurately matched based on the recognized demand keyword.

FIG. 3 is a flowchart of a method for determining shopping needs of a user according to an embodiment of the present disclosure, including the following steps:

s300, acquiring a search statement input by a user.

The method shown in fig. 3 is applied to a system of an e-commerce platform. The method flow shown in fig. 3 is for any search term entered by any user.

The term search statement refers to a statement input in a page provided to an e-commerce platform when a user uses the e-commerce platform to make shopping, and the e-commerce platform searches for goods based on the term search statement and recommends the goods to the user.

S302, inputting the search sentence into a demand keyword recognition model, and recognizing a demand keyword.

The training method of the demand keyword recognition model comprises the steps of obtaining a plurality of search statement samples, taking demand keywords contained in each search statement sample as named entities, and carrying out model training based on a named entity recognition NER algorithm.

In other words, the model is input as each search statement sample, and the named entity label of the required keyword in the statement sample is output.

There are a variety of NER algorithms. In some embodiments, a model comprising a BERT algorithm layer, a BiLSTM algorithm layer, and a CRF algorithm layer connected in sequence may be constructed for training in combination with the BERT algorithm, biLSTM algorithm, and CRF algorithm.

It should be noted that, in the above two models, the first layer of the model is generally used for vectorizing (or coding) each chinese character in the input search sentence, so that the search sentence represented by the vector enters BiLSTM algorithm layer for training or recognition. The BiLSTM algorithm layer can learn which information and which information is forgotten in the training process, so that the dependency relationship with a longer distance can be better captured, and the BiLSTM algorithm layer can learn semantic information in the front and back directions of sentences.

In the model application stage, the output of BiLSTM algorithm layers is the probability of whether each word in a sentence is a named entity or not, and the probabilities corresponding to each word respectively lack a relation (the probability of belonging to a part of the named entity is considered by taking a single word as granularity), and the probability of belonging to the named entity is not considered by taking a word as granularity. This easily results in the recognition of named entities consisting of each word that is part of the named entity based solely on the BiLSTM algorithm layer output, and not the desired keywords to be obtained.

Therefore, the output of BiLSTM algorithm layer is further input to CRF algorithm layer, the CRF algorithm can consider the connection between single words, can analyze the probability of various possible combinations belonging to named entities, and can determine the word combination (i.e. word) with the highest probability of belonging to named entities according to the output of CRF algorithm layer, and the word is the required keyword.

In addition, when the requirement keyword model is trained, three kinds of labels B, I, O are usually adopted to name and label the requirement keywords in the search statement sample. The first word of the requirement keyword is marked by B, the other words of the requirement keyword are marked by I, and the part irrelevant to the requirement keyword in the search sentence is marked by O. A search term may contain more than one demand keyword. When the labeling is actually performed, the label is followed by a named entity type, such as B-S (S represents the named entity type) and I-S. In the method flow shown in fig. 1, only 1 named entity type is actually needed keywords.

Fig. 4 is a schematic structural diagram of a demand keyword recognition model provided in the present specification. As shown in fig. 4, the input search sentence (Query) is "what is required for outdoor riding", the search sentence is input to the BERT layer first, then the code of the search sentence output by the BERT layer is input to the BiLSTM (bi-directional LSTM) neural network layer, the output recognition result is input to the CRF layer again for correction, and the CRF layer finally outputs the corrected recognition result, because the CRF layer is a training stage, the output of the CRF layer is good, the "outdoor riding" is designated as a requirement keyword, and the model is required to be recognized as a named entity through learning.

S304, determining matched shopping requirements according to the identified requirement keywords.

It should be noted here that how to find the matched shopping demand according to the demand keyword may be implemented by "find demand witness words having meaning similar to the demand keyword", where the demand witness words are used to characterize the shopping demand.

In some embodiments, the identified demand keywords may be defined as a demand witness for the shopping demand. A shopping demand may have multiple demand meter witness words.

Further, in some embodiments, a corresponding one of the demand tokens is pre-assigned to each shopping demand. If the requirement characterization words which are the same as the identified requirement keywords exist, the shopping requirements corresponding to the requirement characterization words are determined to be matched shopping requirements. If the demand characterization words which are the same as the identified demand keywords do not exist, the demand characterization words closest to the meanings of the extracted demand keywords are determined, and shopping demands corresponding to the demand characterization words are determined to be matched shopping demands.

In addition, the disclosure also provides a method for determining the requirement characterization words for the shopping requirements. Specifically, a plurality of history search sentences may be acquired. For each history search statement, the history search statement is input into a demand keyword recognition model, and the demand keywords are recognized. And adding each identified requirement keyword into the requirement characterization word alternative set. And dividing the set into a plurality of subsets according to the similarity among the requirement keywords in the set, wherein the similarity among the requirement keywords in the same subset is larger than a specified threshold. Thus, the requirement keywords in the same subset satisfy the same shopping requirement, and different subsets correspond to different shopping requirements. Then, based on different subsets, different requirement characterization words are determined. The requirement characterization words corresponding to each subset are the notary words used for characterizing the requirements corresponding to the subset.

The method for determining the requirement characterization words effectively unifies a plurality of requirement keywords with similar meanings into one shopping requirement, and summarizes the plurality of requirement keywords with similar meanings into one requirement keyword. In some embodiments, a demand keyword may be selected from the subset as a demand witness word representing such shopping demand, or a new word may be created as a demand witness word based on each demand keyword in the subset.

In some embodiments, before calculating the similarity between the requirement keywords in the set, a plurality of search statement samples for training the requirement keyword recognition model may be further obtained, and the requirement keywords in each obtained search statement sample are added to the set. In this way, as many demand keywords as possible can be taken into account.

For how to determine the similarity between the demand keywords in the collection, various algorithms known to calculate meaning similarity between two words can be used.

The disclosure provides a technical scheme for integrating multiple similarity determination methods, which comprises the following steps:

And aiming at any two requirement keywords in the set, adopting at least two different similarity determination methods to respectively determine the similarity between the two requirement keywords. And carrying out weighted calculation according to the obtained at least two similarities, and re-using the calculation result as the similarity between the two requirement keywords.

The above-mentioned weight calculation means that different weights are assigned to different similarity determination methods, and the results of the various similarity determination methods are multiplied by the weights and added. The method can integrate similarity determination methods of various different principles, and analyze the similarity between two requirement keywords from multiple angles.

One similarity determination method used in the present disclosure is to calculate the Jaro distance for two demand keywords. The Jaro distance focuses on the literal similarity of two demand keywords, and on which of the same words the two demand keywords contain. The closer the literal meaning is, the smaller the Jaro distance is. The Jaro distance calculation formula is as follows:

Wherein s1 and s2 respectively represent a demand keyword, m represents the number of repeated characters contained in the two demand keywords, and t represents the number of characters to be modified in the process of converting one demand keyword into another demand keyword.

Further, the calculation can be performed by adopting a Jaro-Winkler distance method, and the formula is as follows:

The Jaro-Winkler distance is further obtained after the Jaro distance is calculated, and it also considers the number p of characters corresponding to the same prefix of the two demand keywords, that is, the two demand keywords are identical with p consecutive characters starting from the 1 st character.

Another similarity determination method used in the present disclosure is to calculate the cosine distance of two demand keywords. Firstly, training a word2vec (a word vector mapping algorithm) model based on commodity titles on an e-commerce platform, then adopting the trained word2vec model to map two required keywords with similarity to word vectors respectively, and comparing cosine distances between the two word vectors to be used as similarity characterization.

Before training the word2vec model, the word2vec model may be set as follows:

Parameter name	Meaning of
		Sg	0, Sg=0 denotes CBOW model, sg=1 denotes Skip-gram model
Size	128, Set the dimension of the word vector
		window	10, Setting the maximum window length of the current word and the predicted word in the text
alpha	0.03, Set learning rate of model
		min_count	1, Discarding if the word frequency is less than min_count, wherein the default value is 5
iter	5, Representing the iteration number of the model
		negative	5, Setting the number of negative samples to indicate how many noise words are contained

In addition, before the cosine distance between the requirement keywords is calculated, the editing distance between the requirement keywords can be calculated first, that is, the minimum editing operation number required for converting from one requirement keyword to another requirement keyword is calculated. A plurality of requirement keywords with editing distances smaller than the specified distance are divided into a group, and then cosine distances among the requirement keywords are calculated in each group, so that the calculated amount of the cosine distances can be effectively reduced.

In addition, the disclosure also provides a method for determining the similarity between the requirement keywords. It should be noted that, the method for determining the similarity between the requirement keywords may be applied to the foregoing embodiment to determine the similarity between two requirement keywords, or may not be applied to the foregoing embodiment, in other words, the requirement keywords described in the method for determining the similarity between the requirement keywords may be requirement keywords identified from search sentences in the foregoing embodiment, or may be requirement keywords identified from search sentences, or may be requirement keywords related to shopping requirements of users acquired from other paths (for example, from a questionnaire survey of a batch of users), or may be requirement keywords generated based on a certain requirement keyword generation rule.

The method for determining the similarity between the requirement keywords can be as follows:

Based on a first demand keyword, carrying out commodity retrieval on an electronic commerce platform, and splicing the commodity name of each commodity in at least part of retrieved commodities into a first commodity name text corresponding to the first demand keyword;

based on the second demand keywords, carrying out commodity retrieval on an electronic commerce platform, and splicing the commodity names of each commodity in at least part of retrieved commodities into a second commodity name text corresponding to the second demand keywords;

Based on a text coding model, acquiring a first text matrix corresponding to the first commodity name text and a second text matrix corresponding to the second commodity name text, and inputting a combined matrix spliced by the first text matrix and the second text matrix into a pre-trained text classification model to obtain probability of positive classification; when training the text classification model, a combination matrix spliced by text matrixes of commodity name texts corresponding to two similar requirement keywords is used as input corresponding to positive classification; the combination matrix spliced by text matrixes of commodity name texts respectively corresponding to two dissimilar demand keywords is input corresponding to negative classification;

And determining the obtained probability as the similarity between the first requirement key and the second requirement key.

The method converts the similarity calculation problem between two requirement keywords into a text classification problem, and indirectly solves the similarity calculation problem by using a text classification method.

On one hand, the method utilizes the demand keywords to search on the E-commerce platform, and the commodity names in the search results are formed into name texts which can be used as the characteristics of the demand keywords. On the other hand, two different name texts are spliced into one text, and a definition is classified into a positive classification (the label may be 1) and a negative classification (the label may be 0), the positive classification indicating that two parts of the text are similar, and the negative classification indicating that two parts of the text are dissimilar.

In the model training stage, the text classification model learns the similarity rule between two parts contained in the spliced texts, so that in the model application stage, after the text spliced by two name texts needing similarity calculation is input into the text classification model, the output probability belonging to positive classification can represent the similarity of the two name texts, namely the similarity between the two name texts corresponding to the required keywords respectively.

The text encoding model is used for mapping Cheng Wenben the commodity name text to a matrix, wherein each row in the text matrix represents a vector corresponding to one word in the text, or each column represents a vector corresponding to one word in the text.

In the method for determining the similarity between the required keywords, the classification algorithms adopted by the text classification model can be various, and the coding algorithms adopted by the text coding model can be various.

In some embodiments, the text classification model may include one fully connected layer, or at least two connected fully connected layers, where each fully connected layer is configured to perform vector compression (vector fully connected for each row in the text matrix, then compressed), and the last fully connected layer contains an activation function configured to map the vector output by the corresponding fully connected layer to a probability of belonging to a positive classification with a probability of belonging to a negative classification.

In some embodiments, the text encoding model may include a Bert algorithm layer, biLSTM algorithm layer, self-attention mechanism self-attention algorithm layer connected in sequence. The text encoding model may further include a max pooling layer connected after the self-attention algorithm layer.

The self-attention algorithm layer is introduced because, on one hand, the importance of different words in the trade name text in characterizing shopping needs is divided into high and low, and the self-attention algorithm can give different attention weights to different words in the text. On the other hand, the self-attention algorithm can capture long-distance dependence among words in the text, so that the coding effect of the text coding model is improved.

The effect of the max pooling layer is to reduce the amount of encoded data and to prevent overfitting.

In some embodiments, an integrated model may be constructed, and training of the text encoding model and the text classification model is accomplished through a process of training the integrated model.

Specifically, the integration model may include a text classification model, a matrix stitching layer, two text encoding models. The text classification model takes a combination matrix output by the matrix splicing layer as input, the matrix splicing layer takes a text matrix output by each text coding model as input, and the two text coding models respectively take different commodity name texts as input.

Fig. 5 is a schematic diagram of an architecture of an integrated model according to an embodiment of the present disclosure. As shown in fig. 5, the integrated model includes a text classification model (i.e., a fully connected layer), a matrix product level, a text coding model 1, and a text coding model 2. The text coding model further comprises a BERT algorithm layer, a BiLSTM algorithm layer, a self-attention algorithm layer and a maximum pooling layer.

In the training phase of the combined model, a plurality of positive samples and negative samples are acquired. Each positive sample can be commodity name text respectively corresponding to two requirement keywords marked as similar, and each negative sample can be commodity name text respectively corresponding to two requirement keywords marked as dissimilar. It will be appreciated that the probability that a positive sample belongs to a positive class is 1 and the probability that a negative sample belongs to a positive class is 0.

After the combination model is trained, in an application stage of the combination model, for a first demand keyword and a second demand keyword which need to be subjected to similarity calculation, a first commodity name text corresponding to the first demand keyword is obtained, and a second commodity name text corresponding to the second demand keyword is obtained. And inputting the first commodity name text and the second commodity name text into the trained integration model, and outputting to obtain the probability of positive classification, wherein the probability can be used as similarity characterization between the first demand keywords and the second demand keywords.

In the case where at least one fully connected layer (for example, 2 layers) is used as the text classification model, the process of training the text classification model is actually a process of iteratively adjusting parameters in the fully connected layer. The goal of the iterative adjustment may be defined as an objective function:

L(sim,g)=glog(sim)+(1-g)log(1-sim);

Where sim represents the probability of belonging to a positive class, g represents whether the sample belongs to a positive class (1) or to a negative class (0). It can be understood that the text classification model with better effect can be trained by iteratively adjusting the parameters in the full connection layer with the aim of maximizing the value of the objective function.

For the foregoing method embodiments, for simplicity of explanation, the methodologies are shown as a series of acts, but one of ordinary skill in the art will appreciate that the present disclosure is not limited by the order of acts described, as some steps may occur in other orders or concurrently in accordance with the disclosure.

Further, those skilled in the art will also appreciate that the embodiments described in the specification are all alternative embodiments, and that the acts and modules referred to are not necessarily required by the present disclosure.

Corresponding to the embodiment of the application function implementation method, the disclosure also provides an embodiment of the application function implementation device and a corresponding terminal.

FIG. 6 is a schematic structural diagram of an apparatus for associating shopping needs for merchandise provided in the present specification, comprising:

The acquiring module 601 acquires a commodity title corresponding to a commodity of the commodity platform for the commodity to be associated with the shopping demand;

The identification module 602 inputs the acquired commodity title to a shopping demand identification model, and identifies shopping demands satisfied by commodity names in the acquired commodity title;

An association module 603 that establishes an association between the identified shopping need and the commodity;

Fig. 7 is a schematic structural diagram of an apparatus for recommending commodities to a user, provided in the present disclosure, and applied to an e-commerce platform, the apparatus includes:

The acquisition module 701 acquires a search statement input by a user, and determines a matched shopping requirement according to the search statement;

And a recommending module 702, which recommends at least part of commodities associated with the determined shopping requirements to the user.

FIG. 8 is a schematic structural diagram of an apparatus for determining shopping needs of a user, applied to an e-commerce platform, provided in the present specification, the apparatus includes:

an acquisition module 801 for acquiring a search sentence input by a user;

the recognition module 802 inputs the search sentence into a demand keyword recognition model to recognize a demand keyword, wherein the training method of the demand keyword recognition model comprises the steps of obtaining a plurality of search sentence samples, taking the demand keyword contained in each search sentence sample as a named entity, and carrying out model training based on a named entity recognition NER algorithm;

a determining module 803 determines a matching shopping requirement based on the identified requirement keywords.

FIG. 9 is a schematic structural diagram of an apparatus for determining a demand token for shopping demand provided in the present specification, including:

An acquisition module 901 for acquiring a plurality of history search sentences;

The recognition module 902 inputs the history search sentences into a requirement keyword recognition model for each history search sentence, and recognizes the requirement keywords;

the adding module 903 adds each identified requirement keyword to the alternative set of requirement characterization words;

the dividing module 904 divides the set into a plurality of subsets according to the similarity between the requirement keywords in the set, wherein the similarity between the requirement keywords in the same subset is greater than a specified threshold;

The determining module 905 determines different requirement tokens based on different subsets.

Fig. 10 is a device for determining similarity between requirement keywords provided in the present description, where the device includes:

The first obtaining module 101 performs commodity retrieval on the electronic commerce platform based on the first demand keyword, and splices the commodity name of each commodity in at least part of the retrieved commodities into a first commodity name text corresponding to the first demand keyword;

The second obtaining module 1002 performs commodity searching on the e-commerce platform based on the second requirement keyword, and splices the commodity name of each commodity in at least part of the searched commodities into a second commodity name text corresponding to the second requirement keyword;

The calculation module 1003 obtains a first text matrix corresponding to the first commodity name text and a second text matrix corresponding to the second commodity name text based on a text coding model, inputs a combined matrix spliced by the first text matrix and the second text matrix into a pre-trained text classification model to obtain probability of positive classification, wherein when the text classification model is trained, the combined matrix spliced by the text matrices of two different commodity name texts corresponding to the same demand keyword is input corresponding to positive classification, the combined matrix spliced by the text matrices of commodity name texts corresponding to two different demand keywords is input corresponding to negative classification, and the obtained probability is determined to be similarity between the first demand keyword and the second demand keyword.

For the device embodiments, reference is made to the description of the method embodiments for the relevant points, since they essentially correspond to the method embodiments. The apparatus embodiments described above are merely illustrative, wherein the elements described above as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the objectives of the disclosed solution. Those of ordinary skill in the art will understand and implement the present invention without undue burden.

Fig. 11 is a schematic diagram of an electronic device 1500, according to an example embodiment. For example, device 1500 may be a user device, which may be embodied as a mobile phone, computer, digital broadcast terminal, messaging device, game console, tablet device, medical device, fitness device, personal digital assistant, wearable device such as a smart watch, smart glasses, smart bracelet, smart running shoe, and the like.

Referring to FIG. 11, device 1500 can include one or more of a processing component 1502, a memory 1504, a power component 1506, a multimedia component 1508, an audio component 1510, an input/output (I/O) interface 1512, a sensor component 1514, and a communication component 1516.

The processing component 1502 generally controls overall operation of the device 1500, such as operations associated with display, telephone calls, data communications, camera operations, and recording operations. The processing component 1502 may include one or more processors 1520 to execute instructions to perform all or part of the steps of the methods described above. Further, the processing component 1502 may include one or more modules that facilitate interactions between the processing component 1502 and other components. For example, the processing component 1502 may include a multimedia module to facilitate interaction between the multimedia component 1508 and the processing component 1502.

The memory 1504 is configured to store various types of data to support operations at the device 1500. Examples of such data include instructions for any application or method operating on device 1500, contact data, phonebook data, messages, pictures, video, and the like. The memory 1504 may be implemented by any type or combination of volatile or nonvolatile memory devices such as Static Random Access Memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic or optical disk.

The power supply assembly 1506 provides power to the various components of the device 1500. The power supply component 1506 can include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the device 1500.

The multimedia component 1508 comprises a screen between the device 1500 and the user that provides an output interface. In some embodiments, the screen may include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensor may sense not only a boundary of a touch or a sliding action but also a duration and a pressure related to the touch or the sliding operation. In some embodiments, multimedia assembly 1508 includes a front camera and/or a rear camera. The front camera and/or the rear camera may receive external multimedia data when the device 1500 is in an operational mode, such as a shooting mode or a video mode. Each front camera and rear camera may be a fixed optical lens system or have focal length and optical zoom capabilities.

The audio component 1510 is configured to output and/or input audio signals. For example, the audio component 1510 includes a Microphone (MIC) configured to receive external audio signals when the device 1500 is in an operational mode, such as a call mode, a recording mode, and a voice recognition mode. The received audio signals may be further stored in the memory 1504 or transmitted via the communication component 1516. In some embodiments, the audio component 1510 further comprises a speaker for outputting audio signals.

The I/O interface 1512 provides an interface between the processing component 1502 and peripheral interface modules, which can be keyboards, click wheels, buttons, and the like. These buttons may include, but are not limited to, a home button, a volume button, an activate button, and a lock button.

The sensor assembly 1514 includes one or more sensors for providing status assessment of various aspects of the device 1500. For example, the sensor assembly 1514 may detect an on/off state of the device 1500, a relative positioning of the assemblies such as a display and keypad of the device 1500, the sensor assembly 1514 may also detect a change in position of the device 1500 or one of the assemblies of the device 1500, the presence or absence of a user's contact with the device 1500, an orientation or acceleration/deceleration of the device 1500, and a change in temperature of the device 1500. The sensor assembly 1514 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. The sensor assembly 1514 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 1514 may also include an acceleration sensor, a gyroscopic sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

The communication component 1516 is configured to facilitate communication between the device 1500 and other devices, either wired or wireless. The device 1500 may access a wireless network based on a communication standard, such as WiFi,2G or 3G,4G LTE, 5G NR, or a combination thereof. In one exemplary embodiment, the communication component 1516 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, the communication component 1516 described above further includes a Near Field Communication (NFC) module to facilitate short range communications. For example, the NFC module may be implemented based on Radio Frequency Identification (RFID) technology, infrared data association (IrDA) technology, ultra Wideband (UWB) technology, bluetooth (BT) technology, and other technologies.

In an exemplary embodiment, the apparatus 1500 may be implemented by one or more Application Specific Integrated Circuits (ASICs), digital Signal Processors (DSPs), digital Signal Processing Devices (DSPDs), programmable Logic Devices (PLDs), field Programmable Gate Arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic elements for executing the methods described above.

In an exemplary embodiment, a non-transitory computer readable storage medium is also provided, such as a memory 1504 including instructions that, when executed by the processor 1520 of the apparatus 1500, enable the apparatus 1500 to perform the methods of the embodiments of the present specification.

The non-transitory computer readable storage medium may be a ROM, random-access memory (RAM), CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.

As shown in fig. 12, fig. 12 is a schematic diagram of another electronic device 1600, according to an example embodiment. For example, device 1600 may be provided as an application server. Referring to fig. 7, device 1600 includes a processing component 1622 that further includes one or more processors and memory resources, represented by memory 1616, for storing instructions, such as application programs, executable by processing component 1622. An application program stored in memory 1616 may include one or more modules each corresponding to a set of instructions. Further, processing component 1622 is configured to execute instructions to perform the methods described above.

The device 1600 may also include a power component 1626 configured to perform power management of the device 1600, a wired or wireless network interface 1650 configured to connect the device 1600 to a network, and an input output (I/O) interface 1658. The device 1600 may operate based on an operating system stored in memory 1616, such as Android, iOS, windows server (tm), mac OS XTM, unixTM, linuxTM, freeBSDTM, or the like.

In an exemplary embodiment, a non-transitory computer-readable storage medium is also provided, such as a memory 1616, that includes instructions executable by the processing component 1622 of the device 1600 to perform the above-described methods. For example, the non-transitory computer readable storage medium may be ROM, random Access Memory (RAM), CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.

Wherein the instructions in the memory 1616, when executed by the processing component 1622, enable the device 1600 to perform the methods of the various embodiments of the present description.

Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure disclosed herein. This disclosure is intended to cover any adaptations, uses, or adaptations of the disclosure following the general principles of the disclosure and including such departures from the present disclosure as come within known or customary practice within the art to which the disclosure pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.

It is to be understood that the present disclosure is not limited to the precise arrangements and instrumentalities shown in the drawings, and that various modifications and changes may be effected without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method of associating shopping requirements for merchandise, comprising:

2. The method of claim 1, wherein the shopping need identification model comprises a BERT algorithm layer, a BiLSTM algorithm layer, and a CRF algorithm layer connected in sequence.

3. The method of claim 1, wherein obtaining a commodity title sample comprises:

Acquiring an actual commodity title;

and determining at least one other commodity name meeting the shopping requirement according to the shopping requirement met by the commodity names in the actual commodity title, and adding the at least one other commodity name into the actual commodity title to obtain the commodity title sample.

4. The method of claim 1, wherein the shopping requirement satisfied by the commodity name in the commodity title sample is a named entity type, comprising:

and aiming at the commodity title sample, if shopping requirements respectively met by any two commodity names in the commodity title sample are different, determining different named entity types for the two commodity names.

5. A method for recommending goods for a user, characterized by being applied to an e-commerce platform, the method comprising:

Recommending at least part of the goods associated with the determined shopping demand to the user, wherein the association between the shopping demand and the goods is determined by the method of any one of claims 1-4.

6. The method of claim 5, wherein determining matching shopping needs based on the search statement comprises:

and extracting a demand keyword from the search statement, and determining the matched shopping demand according to the extracted demand keyword.

7. The method of claim 6, wherein,

Determining matching shopping requirements according to the extracted requirement keywords, including:

a demand characterization word is allocated for shopping demands in advance;

If the requirement characterization words which are the same as the extracted requirement keywords exist, determining shopping requirements corresponding to the requirement characterization words as the matched shopping requirements;

if the demand representation words which are the same as the extracted demand keywords do not exist, determining the demand representation words closest to the meanings of the extracted demand keywords from the demand meter witness words, and determining shopping demands corresponding to the closest demand representation words as the matched shopping demands.

8. The method of claim 6, wherein determining matching shopping needs based on the search statement comprises:

and inputting the search statement into a demand keyword recognition model, and determining the matched shopping demand according to the recognized demand keyword.

9. The method of claim 8, wherein the training method of the demand keyword recognition model comprises:

And obtaining a search statement sample, taking a demand keyword contained in the search statement sample as a named entity, and performing model training based on an NER algorithm.

10. The method of claim 9, wherein the demand keyword recognition model comprises a BERT algorithm layer, a BiLSTM algorithm layer, and a CRF algorithm layer connected in sequence.

11. The method of claim 6, wherein determining matching shopping needs based on the search statement comprises:

Performing word segmentation on the search sentence, and inputting the word segmentation into a word classification model;

and determining the matched shopping demands according to the word segmentation classified as the demand keywords.

12. The method of claim 11, wherein the training method of the word classification model comprises:

obtaining a search statement sample, and extracting a requirement keyword from the search statement sample;

and taking the extracted demand keywords as white samples, and performing model training by adopting a classification algorithm.

13. The method of claim 12, wherein the word classification model comprises a BERT algorithm layer, a BiLSTM algorithm layer, and a full connection layer connected in sequence.

14. An apparatus for correlating shopping needs for merchandise, comprising:

15. An apparatus for recommending items for a user, the apparatus being adapted for use with an e-commerce platform, the apparatus comprising:

A recommending module for recommending at least part of commodities associated with the determined shopping demands to the user, wherein the association between the shopping demands and the commodities is determined by the method of any one of claims 1-4.

16. An electronic device comprising a processor, a memory for storing processor-executable instructions, wherein the processor is configured to execute the instructions to implement the method of any of claims 1-13.

17. A non-transitory computer readable storage medium having stored thereon a computer program, characterized in that the program, when executed by a processor, implements the method of any of claims 1-13.