Brand association model training method, brand association model training system and brand information completion method
Through the combination of brand data enhancement and GAN discriminator, the problems of large manual workload and high computational complexity in brand association model training are solved, and efficient and accurate brand association model training is achieved.
Patent Information
- Application Number
- CN202411133879.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-16
- Publication Date
- 2025-02-21
AI Technical Summary
The prior art has problems such as large manual workload and high computational complexity in brand association model training, especially in the data set annotation and model training process.
The brand data enhancement steps are used to obtain positive and negative samples, and the pre-trained language model and GAN discriminator are used to discriminate and calculate the feature vectors and iteratively optimize the model parameters to obtain the trained brand association model.
It effectively reduces the workload and calculation complexity of manual labeling, improves the training efficiency and accuracy of the brand association model, and can more accurately judge the degree of brand association.
Smart Images
Figure CN119494370A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to computer information technology, and in particular to a brand association model training method and a training system thereof, and a brand information completion method. Background Art
[0002] Regarding brand mining, existing technical implementation methods are mainly divided into the following two schools:
[0003] (1) Rule-based method: The brand association of stores is mainly achieved by checking whether the keywords in brand information and store names overlap. For example, if “LELECHA Lelecha (Pudong Hang Seng Plaza Store)” contains the brand keyword “Lelecha”, the association between the two can be established.
[0004] (2) Model-based method: mainly uses machine learning or deep learning models to convert brand information and store names into feature vectors, uses a similarity measurement formula to calculate the similarity between the two, and selects the results with the highest similarity that exceeds the threshold to establish an association relationship between the two.
[0005] Regardless of whether it is a rule-based method or a model-based method, the shortcomings or problems of existing technologies are mainly concentrated on the manual construction of data sets, including:
[0006] (1) Rule-based method: It is necessary to collect keywords corresponding to all brands. Taking McDonald’s as an example, its brand keywords include but are not limited to {“McDonald’s”, “Golden Arches”, “McDonald”}. The workload of manually sorting data is too large.
[0007] (2) Model-based methods: A large amount of associations between brand information and store names need to be labeled as training sets, and the training model is used to extract text feature vectors. In addition, the time complexity of directly comparing the similarity between brand information and store names is too high (it is necessary to compare with all brands in the brand library), and it is often necessary to combine other methods for model pruning. Summary of the invention
[0008] In order to solve the above problems in the prior art, it is intended to provide a brand association model training method and a training system thereof that can avoid excessive manual workload and reduce computational complexity.
[0009] A brand association model training method according to one aspect of the present invention comprises:
[0010] A brand data augmentation step to obtain positive and negative samples from the collection of company names and trademark names respectively; and
[0011] The brand association model training step includes inputting the positive samples and negative samples into a specified pre-trained language model to obtain a feature vector, using a discriminator in the GAN to discriminate the feature vector and calculate the loss of the positive and negative samples, and iteratively optimizing the parameters of the pre-trained language model and the GAN based on the loss to obtain a trained brand association model.
[0012] Optionally, obtaining a positive sample includes:
[0013] A data pair consisting of a company name and a trademark name whose similarity is higher than a preset first threshold is obtained as a positive sample from a company-trademark relationship table containing the correspondence between company names and trademark names using a prescribed similarity calculation model.
[0014] Optionally, obtaining a positive sample further includes:
[0015] A specified language model is used to obtain data pairs consisting of enterprise names and enterprise abbreviations from an enterprise total-branch relationship table containing the correspondence between enterprise names and enterprise branch names as positive samples.
[0016] Optionally, obtaining a positive sample includes:
[0017] The data pairs consisting of the company names and corresponding brand names obtained through manual annotation are taken as positive samples.
[0018] Optionally, obtaining negative samples from the enterprise trademark relationship table includes:
[0019] Data pairs consisting of enterprise names and trademark names are randomly selected from the enterprise-trademark relationship table as negative samples.
[0020] Optionally, obtaining negative samples from the enterprise trademark relationship table further includes:
[0021] A prescribed similarity calculation model is used to obtain data pairs consisting of enterprise names and trademark names whose similarity is lower than a preset second threshold from the enterprise-trademark relationship table as negative samples.
[0022] Optionally, obtaining negative samples in the brand data enhancement step further includes: generating two random feature vectors by using a GAN generator to generate random noise,
[0023] The brand association model training step includes:
[0024] Inputting the negative sample into a specified pre-trained language model to obtain a feature vector, and concatenating the feature vectors to obtain a first negative sample set, concatenating the two random feature vectors to obtain a second negative sample set, and merging the first negative sample set and the second negative sample set to obtain a negative sample set;
[0025] Inputting the positive samples into the specified pre-trained language model to obtain feature vectors, and concatenating the feature vectors to obtain a positive sample set;
[0026] Using the discriminator in the GAN to discriminate positive and negative labels for the positive sample set and the negative sample set, and calculating the loss of the positive and negative samples; and
[0027] Parameters of the pre-trained language model and the GAN are iteratively optimized based on the loss to obtain a trained brand association model.
[0028] Optionally, the brand association model training step includes:
[0029] Inputting the positive samples and negative samples into a specified pre-trained language model to obtain feature vectors, and concatenating the feature vectors to obtain a positive and negative sample set;
[0030] Using the discriminator in GAN to discriminate positive and negative labels for the positive and negative sample sets, and calculating the loss of positive and negative samples; and
[0031] Parameters of the pre-trained language model and the GAN are iteratively optimized based on the loss to obtain a trained brand association model.
[0032] Optionally, the pre-trained language model is any one of BERT, RoBERTa, and XLNet.
[0033] A method for completing brand information according to one aspect of the present invention comprises:
[0034] Get the company names that have no corresponding brands;
[0035] According to the enterprise name without a corresponding brand, a candidate brand is selected from an enterprise trademark relationship table including a correspondence between enterprise names and trademark names and an enterprise total-branch relationship table including a correspondence between enterprise names and branch names based on character overlap of the enterprise names;
[0036] Calculating the similarity between the company name without a corresponding brand and the candidate brand using the brand association model trained by the brand association model training method according to any one of claims 1 to 9; and
[0037] A brand association result associated with the enterprise name without a corresponding brand is obtained based on the calculation result of the similarity.
[0038] A brand association model training system according to one aspect of the present invention comprises:
[0039] A brand data enhancement module, used to obtain positive samples and negative samples from a corporate trademark relationship table containing the correspondence between corporate names and trademark names and a corporate total score relationship table containing corporate names and branch names; and
[0040] The brand association model training module inputs the positive samples and negative samples into a specified pre-trained language model to obtain a feature vector, uses a discriminator in the GAN to discriminate the feature vector and calculates the loss of the positive and negative samples, and iteratively optimizes the parameters of the pre-trained language model and the GAN based on the loss to obtain a trained brand association model.
[0041] Optionally, the brand data enhancement module includes:
[0042] A positive sample acquisition submodule for acquiring positive samples; and
[0043] Negative sample submodule for obtaining negative samples.
[0044] Optionally, the positive sample acquisition submodule uses a specified similarity calculation model to acquire data pairs consisting of enterprise names and trademark names whose similarity is higher than a preset first threshold from the enterprise-trademark relationship table as positive samples.
[0045] Optionally, the positive sample acquisition submodule uses a specified language model to acquire a data pair consisting of an enterprise name and an enterprise abbreviation from a data containing an enterprise name and an enterprise branch name as a positive sample.
[0046] Optionally, the positive sample acquisition submodule uses data pairs consisting of company names and corresponding brand names obtained through manual annotation as positive samples.
[0047] Optionally, the negative sample submodule randomly selects data pairs consisting of enterprise names and trademark names from the enterprise-trademark relationship table as negative samples.
[0048] Optionally, the negative sample submodule uses a specified similarity calculation model to obtain data pairs consisting of enterprise names and trademark names whose similarity is lower than a preset second threshold from the enterprise-trademark relationship table as negative samples.
[0049] Optionally, the negative sample submodule uses a GAN generator to generate random noise to generate two random feature vectors as negative samples.
[0050] Optionally, the brand association model training module is used to input the positive samples and negative samples into a specified pre-trained language model and obtain feature vectors, concatenate the feature vectors to obtain positive and negative sample sets, use the discriminator in the GAN to discriminate positive and negative labels for the positive and negative sample sets, and calculate the loss of positive and negative samples, and iteratively optimize the parameters of the pre-trained language model and the GAN based on the loss to obtain a trained brand association model.
[0051] Optionally, the pre-trained language model is any one of BERT, RoBERTa, and XLNet.
[0052] A computer-readable medium according to one aspect of the present invention stores a computer program, which, when executed by a processor, implements the brand association model training method according to any one of claims 1 to 9.
[0053] A computer device according to one aspect of the present invention comprises a storage module, a processor, and a computer program stored in the storage module and executable on the processor, wherein the processor implements the brand association model training method when executing the computer program.
[0054] A computer program product according to one aspect of the present invention comprises a computer program, which implements the brand association model training method when executed by a processor. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 It is a schematic diagram of a brand association model training method according to an embodiment of the present invention.
[0056] Figure 2 This is an example of a table showing the total and branch relationship of a company.
[0057] Figure 3 This is an example of a table representing corporate trademark relationships.
[0058] Figure 4 It is a structural block diagram of a brand association model training system according to one embodiment of the present invention.
[0059] The above and other objects and advantages of the present application will become more fully apparent from the following detailed description in conjunction with the accompanying drawings, in which the same or similar elements are represented by the same reference numerals. DETAILED DESCRIPTION
[0060] The following describes some of the various embodiments of the present invention, which are intended to provide a basic understanding of the present invention, but are not intended to identify the key or decisive elements of the present invention or to limit the scope of protection.
[0061] For the purpose of brevity and illustration, the principles of the present invention are described herein mainly with reference to its exemplary embodiments. However, those skilled in the art will readily recognize that the same principles are equally applicable to all types of brand association model training methods and training systems thereof, and these same principles can be implemented therein, and any such changes do not depart from the true spirit and scope of the present patent application.
[0062] Moreover, in the following description, reference is made to the accompanying drawings, which illustrate specific exemplary embodiments. Electrical, mechanical, logical and structural changes may be made to these embodiments without departing from the spirit and scope of the present invention. In addition, although the features of the present invention are disclosed in conjunction with only one of several implementations / embodiments, this feature may be combined with one or more other features of other implementations / embodiments as may be desirable and / or advantageous for any given or identifiable function. Therefore, the following description should not be regarded as limiting, and the scope of the present invention is defined by the appended claims and their equivalents.
[0063] Terms such as “having” and “including” indicate that in addition to the units (modules) and steps directly and explicitly stated in the specification and claims, the technical solution of the present invention does not exclude the situation where it has other units (modules) and steps that are not directly or explicitly stated.
[0064] Before describing the brand association model training method and brand association model training system of the present invention, several related technical terms are briefly described.
[0065] (1)BERT (Bidirectional Encoder Representation fromTransformers)
[0066] The Transformer-based bidirectional encoder representation mainly uses its feature extraction function to convert text into feature vectors in digital form.
[0067] (2)GAN (Generative Adversarial Networks)
[0068] Generative Adversarial Networks,
[0069] (3) Generator and Discriminator
[0070] The main use of his sample generation and discrimination functions is to generate more diverse negative samples and expand the size of the training set. On the other hand, his discriminator is used to calculate the similarity between company names and brands.
[0071] (4) Adam Optimizer (Adaptive Moment Estimation)
[0072] It is an optimization algorithm widely used in deep learning. It was proposed by Diederik P. Kingma and Jimmy Ba in 2014. It combines the advantages of the momentum method and RMSProp optimization algorithms, and aims to accelerate the training process of the neural network by adaptively adjusting the learning rate of each parameter.
[0073] Figure 1 It is a schematic diagram of a brand association model training method according to an embodiment of the present invention.
[0074] like Figure 1 As shown, the brand association model training method according to one embodiment of the present invention mainly includes: a brand data enhancement step S100 and a brand association model training step S200, and the merchant brand information completion step S300 is described as an application example of the brand association model trained by the brand association model training method. These steps are described in detail below.
[0075] First, the brand data enhancement step S100 includes:
[0076] (1) Filter and obtain the enterprise total-branch relationship table and enterprise trademark relationship table from the enterprise map. The filtering rules can be selected by region, industry, keyword, etc.
[0077] For ease of understanding, an example of a company's total-score relationship table and a company's trademark relationship table is given for illustration. Figure 2 This is an example of a table showing the total-to-score relationship of a company. Figure 3 is an example of a table representing the relationship between corporate trademarks. Figure 2 In the enterprise total branch relationship table, the vertical columns include enterprise name, credit code, branch name, and branch credit code. Figure 3 The vertical columns in the enterprise trademark relationship table include, for example, enterprise identifier, branch name, trademark code, and trademark name.
[0078] (2) Obtain positive samples from the enterprise total-score relationship table and the enterprise trademark relationship table. The following three methods can be used to obtain positive samples:
[0079] Get similar companies and similar trademarks as positive samples
[0080] Use rule similarity calculation methods such as BLEU, Rouge, or general model methods such as LSTM, BERT, GPT, LLM large language model, etc. Figure 3In the table, data pairs consisting of company names and trademark names whose similarity is significantly higher than a threshold (as an example, the threshold is set to 0.8) are selected as positive samples of the brand association model, for example: Shandong Luneng Commercial Management Co., Ltd., Guihe InterContinental Hotel, Guihe.
[0081] Get the company name and abbreviation as positive samples
[0082] Use the abbreviation extraction model trained by large language models such as LSTM, BERT, GPT, and LLM to generate positive samples, for example: Shandong Luneng Commercial Management Co., Ltd. Guihe InterContinental Hotel, Luneng.
[0083] Manually annotate company names and brand names as positive samples
[0084] The data pairs consisting of manually labeled company names and corresponding brand names are used as positive samples, for example: Shandong Luneng Commercial Management Co., Ltd. Guihe InterContinental Hotel, Guihe InterContinental.
[0085] (3) Obtaining negative samples
[0086] There are three ways to obtain negative samples:
[0087] ① Select random companies and random trademarks as negative samples
[0088] from Figure 3 Randomly select data pairs consisting of company names and trademark names from the table as negative samples of the brand association model, for example, Anhui Youruijia Technology Co., Ltd. Bengbu Branch, Guihe.
[0089] ② Obtain companies and trademarks with low similarity as negative samples
[0090] Use rule similarity calculation methods such as BLEU, Rouge, or general model methods such as LSTM, BERT, GPT, LLM large language model, etc. Figure 3 Data pairs consisting of company names and trademark names whose similarity is significantly lower than a threshold (as an example, the threshold is set to 0.4) are selected as negative samples of the brand association model, such as Shanghai Tobacco Group Hongkou Tobacco, Sugar and Wine Co., Ltd. Famous Tobacco and Famous Wine Wan'an Store, Baiguoyuan.
[0091] ③Generate random noise as negative samples
[0092] Since the subsequent brand association model training step S200 will convert positive and negative samples into feature vectors for similarity determination, the negative sample acquisition process can be skipped by generating random noise and converting the random noise into random feature vector 1 and random feature vector 2 through the GAN generator as negative samples of the brand association model.
[0093] noise=torch.normal(0,1,size=(-1,128))
[0094] Generator=torch.nn.Sequential(
[0095] torch.nn.Linear(in_features=128, out_features=256),
[0096] torch.nn.Tanh(),
[0097] torch.nn.Linear(in_features=256, out_features=512),
[0098] torch.nn.ReLU(),
[0099] torch.nn.Linear(in_features=512, out_features=768),
[0100] torch.nn.Tanh() )
[0102] fake_feature=Generator(noise)
[0103] Here, the torch.normal function is used to generate a random noise vector (noise) based on the standard normal distribution (mean 0, standard deviation 1). The dimension of this vector is 128, and the number and dimension of the generated noise vector can be changed by adjusting the size parameter. Next, torch.nn.Sequential is used to define a simple neural network as a generator (Generator). This network contains three fully connected layers, using Tanh and ReLU as activation functions respectively. The input feature of the network is 128 dimensions, and the final output feature is 768 dimensions. The random noise vector (noise) is passed as input to the generator (Generator) to obtain the output random feature vector (fake_feature). The dimension of this feature vector is 768, which can be used as a negative sample for the brand association model. In other words, the 128-dimensional noise vector noise randomly generated based on the (0, 1) uniform distribution can be converted into a 768-dimensional feature vector (fake_feature).
[0104] The benefit of this method is that it provides a way to quickly generate a large number of negative samples without collecting or labeling them from actual data. This is particularly useful for application scenarios where negative samples are difficult to obtain or expensive to label. At the same time, since the generated negative samples are random, it also helps to improve the generalization ability of the model.
[0105] Next, the brand association model training step S200 is described.
[0106] The brand association model training step S200 includes:
[0107] (1) Generate feature vectors corresponding to positive and negative samples
[0108] Use a feature encoder (for example, download a deep learning model from huggingface.co) to map multi-source brand data to the same vector space and align it with the feature vector generated by the generator. Here, as an example in this implementation, a pre-trained BERT model (for example, bert-base-uncased) is used as a feature encoder.
[0109] Bert_model=transformers.AutoModel.from_pretrained
[0110] (PATH_OF_MODEL)
[0111] The company name, company abbreviation, trademark name, and manually labeled brand name in the positive and negative samples are input into the BERT model to generate a 768-dimensional feature vector name_feature (the feature dimension pre-defined by the bert-base-uncased model), such as:
[0112] inputs=tokenizer(name, return_tensors="pt")
[0113] outputs=Bert_model(**inputs)
[0114] name_feature=outputs["last_hidden_state"]
[0115] Optionally, BERT can be replaced with other text feature encoders such as RoBERTa and XLNet.
[0116] (2) Use the GAN discriminator to distinguish the labels corresponding to positive and negative samples
[0117] According to the above content, the feature vectors in all samples are concatenated to form positive and negative sample sets, which are as follows:
[0118] Positive sample set: pos_feature∈{(company name, trademark), (company name, company abbreviation), (company name, brand name)}
[0119] Negative sample set: neg_feature∈{(random enterprise, random trademark), (random feature vector 1, random feature vector 2)}
[0120] Use the discriminator in GAN to input the above positive and negative sample feature vectors into the brand association model to calculate the loss, as follows:
[0121] Discriminator=torch.nn.Sequential(
[0122] torch.nn.Linear(in_features=768*2, out_features=512),
[0123] torch.nn.Tanh(),
[0124] torch.nn.Linear(in_features=512, out_features=256),
[0125] torch.nn.Tanh(),
[0126] torch.nn.Linear(in_features=256, out_features=128),
[0127] torch.nn.Tanh(),
[0128] torch.nn.Linear(in_features=128, out_features=1),
[0129] torch.nn.Sigmoid() )
[0131] pos_score=Discriminator(pos_feature.detach())
[0132] neg_score=Discriminator(neg_feature.detach())
[0133] loss_fn = torch.nn.BCELoss()
[0134] pos_loss=loss_fn(pos_score, torch.ones_like(pos_score))
[0135] neg_loss=loss_fn(neg_score, torch.zeros_like(neg_score))
[0136] loss = pos_loss + neg_loss
[0137] In the above content, a multi-layer fully connected neural network is used as the discriminator, the dimension of the input feature vector is 768*2, and the output is a score between 0 and 1, indicating the probability that the sample is a positive sample. Then, the loss is calculated for positive and negative samples separately, and the binary cross entropy loss function (BCELoss) is used.
[0138] (3) Iteratively optimize the parameters of BERT and GAN and train the brand association model as follows:
[0139] paras=Bert_model.paramaters()+Generator.paramaters()+Discriminator.paramaters()
[0140] optim=torch.optim.Adam(paras,lr=1e-4)
[0141] optim.zero_grad()
[0142] loss.backward()
[0143] optim.step()
[0144] In the above content, the parameters of the BERT model, generator, and discriminator are combined for optimization. The Adam optimizer is used, the learning rate is set to 1e-4, and backpropagation and parameter updates are also performed.
[0145] As described above, a brand association model can be trained by illustrating the brand association model training step S200, and the brand association model can determine the degree of association between a given brand name and a company name, a trademark name, etc.
[0146] According to the brand association model training method of the present invention, by using the association relationship between the enterprise total-score relationship table and the enterprise trademark relationship table in the enterprise map, a data set of annotations of the association relationship between the enterprise name and the trademark name as the brand information and the store name is mined, and a brand association model is trained on the basis of this part of the data set, which can be used to complete the brand information of existing or incremental merchants later.
[0147] As an example that can be used to subsequently complete the brand information of existing or incremental merchants, in this embodiment, the merchant brand information completion step S300 is understood for explanation.
[0148] The merchant brand information completion step S300 is an example of an application scenario of the brand association model of the present invention. The application of the brand association model of the present invention is not limited to the application in merchant brand information completion, but can also be applied to other fields such as brand management and trademark infringement detection. Here, product brand information completion is used as an example for explanation.
[0149] The merchant brand information completion step S300 includes:
[0150] (1) Define the candidate brands based on the names of unbranded companies
[0151] As an example, for example: the brand company name to be completed name = "Yonghui Supermarket Co., Ltd. Fujian Fuzhou Hongxingyuan Supermarket".
[0152] Optional, if the merchant has a unified social credit code "9135****6536", search Figure 2 The enterprise total-branch relationship table is used to obtain the corresponding parent enterprise information of the merchant, "Yonghui Superstores Co., Ltd.", and continue searching Figure 3 The enterprise trademark relationship table is used to obtain the trademark set corresponding to the parent enterprise, candidates = {YH, Yonghui, Yonghui Youxuan, ...}, as the candidate brand range.
[0153] Optionally, brands with overlapping characters are selected from the brand library as candidate brand range candidates = {Yonghui, ParknShop Yonghui, Red Star Macalline...}.
[0154] (2) Calculate the similarity between the unbranded company name and the candidate brand
[0155] The company name to be completed and the candidate brand range are converted into feature vectors through the BERT model, and then the GAN discriminator calculates the similarity between the two. As follows:
[0156] name_input=tokenizer(name, return_tensors="pt")
[0157] name_output=Bert_model(**name_input)
[0158] name_feature=name_output["last_hidden_state"]
[0159] cand_inputs=tokenizer(candidates, return_tensors="pt")
[0160] cand_outputs=Bert_model(**cand_inputs)
[0161] cand_features=cand_outputs["last_hidden_state"]
[0162] similarity=Discriminator([name_feature,cand_features].detach())
[0163] (3) Select the result with the greatest similarity, determine its relationship with the similarity threshold, and obtain the brand association result.
[0164] For example, taking name = "Yonghui Supermarket Co., Ltd., Hongxingyuan Supermarket, Fuzhou, Fujian" as an example, after calculation, sim(name, "Yonghui")>sim(name, "Yonghui Select")>sim(name, "ParknShop Yonghui")>sim(name, "Red Star Macalline"), and finally "Yonghui" is taken as the brand association result of "Yonghui Supermarket Co., Ltd., Hongxingyuan Supermarket, Fuzhou, Fujian".
[0165] As described above, according to the brand association model training method of the present invention, by making use of the relationship between the total and divided enterprises, and the relationship between enterprises and trademarks in the enterprise graph, a data set of annotations of the relationship between enterprise names and trademark names as the relationship between brand information and store names is mined, and a brand association model is trained on the basis of this part of the data set. Subsequently, the brand association model can be used to complete the brand information of existing / incremental merchants.
[0166] Furthermore, according to the brand association model training method of the present invention, by using a variety of data enhancement technical means such as GAN generator, similarity calculation, abbreviated sequence labeling, etc., it is possible to expand the brand association model training data set and solve the problems of lack of real brand data and excessive workload of manual labeling.
[0167] Figure 4 It is a structural block diagram of a brand association model training system according to one embodiment of the present invention.
[0168] like Figure 4 As shown, a brand association model training system 100 according to an embodiment of the present invention includes:
[0169] A brand data enhancement module 110, for respectively obtaining positive samples and negative samples from a corporate trademark relationship table containing the correspondence between corporate names and trademark names and a corporate total-score relationship table containing corporate names and branch names; and
[0170] The brand association model training module 120 inputs the positive samples and negative samples into a specified pre-trained language model to obtain a feature vector, uses a discriminator in the GAN to discriminate the feature vector and calculates the loss of the positive and negative samples, and iteratively optimizes the parameters of the pre-trained language model and the GAN based on the loss to obtain a trained brand association model.
[0171] The brand data enhancement module 110 includes:
[0172] A positive sample acquisition submodule 111 for acquiring positive samples; and
[0173] A negative sample submodule 112 is used to obtain negative samples.
[0174] The positive sample acquisition submodule 111 uses a specified similarity calculation model to acquire enterprise names and trademark names whose similarity is higher than a preset first threshold from the enterprise-trademark relationship table as positive samples.
[0175] Optionally, the positive sample acquisition submodule 111 uses a specified language model to acquire the enterprise name and the enterprise abbreviation as positive samples from a database containing the enterprise name and the enterprise branch name.
[0176] Optionally, the positive sample acquisition submodule 111 uses the company names and corresponding brand names obtained through manual annotation as positive samples.
[0177] Optionally, the negative sample submodule 112 randomly selects enterprise names and trademark names from the enterprise-trademark relationship table as negative samples.
[0178] Optionally, the negative sample submodule 112 uses a specified similarity calculation model to obtain enterprise names and trademark names whose similarity is lower than a preset second threshold from the enterprise-trademark relationship table as negative samples.
[0179] Optionally, the negative sample submodule 112 uses a GAN generator to generate random noise to generate a random feature vector as a negative sample.
[0180] Among them, the brand association model training module 200 is used to input the positive samples and negative samples into a specified pre-trained language model and obtain feature vectors, concatenate the feature vectors to obtain positive and negative sample sets, use the discriminator in the GAN to discriminate positive and negative labels for the positive and negative sample sets, and calculate the loss of positive and negative samples, and iteratively optimize the parameters of the pre-trained language model and the GAN based on the loss to obtain a trained brand association model.
[0181] Optionally, the pre-trained language model is any one of BERT, RoBERTa, and XLNet.
[0182] As described above, the brand association model training method and brand association model training system according to the present invention can achieve the following technical effects:
[0183] (1) Efficient feature representation
[0184] Unified vector space: Through the feature encoder, brand data from different sources and in various formats (such as company names, abbreviations, trademark names, etc.) are converted into feature vectors of fixed dimensions, so that these originally heterogeneous data can be compared and calculated in the same vector space, which greatly simplifies the subsequent processing flow and improves the efficiency and accuracy of data processing.
[0185] (2) Rich semantic information
[0186] Pre-trained deep learning models (such as BERT) have learned a wealth of language knowledge and semantic information during the training process, and can therefore generate feature vectors containing rich semantics. These feature vectors can more accurately reflect the inherent connections and differences between brand data.
[0187] (3) Strong discrimination ability
[0188] The discriminator of GAN is used. Since the discriminator in GAN is designed to distinguish between real samples and generated samples, it has strong discriminative ability. Here, the discriminator is used to distinguish between positive and negative samples (i.e., brand data pairs with associated relationships and random data pairs without associated relationships), so as to train a model that can accurately judge the degree of brand association. In addition, adversarial training is carried out. Specifically, the adversarial training mechanism of GAN is used to enable the discriminator and generator to continuously improve their respective capabilities in the mutual game. In this process, the discriminator gradually learns to distinguish positive and negative samples more finely, improving the accuracy and robustness of the brand association model.
[0189] (4) Flexibility and scalability
[0190] Multi-source data support: Since the feature encoder can map multi-source brand data into the same vector space, this method is very flexible and can be easily extended to more data sources and more complex scenarios.
[0191] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Those skilled in the art can think of other feasible changes or substitutions based on the technical scope disclosed in the present application, and such changes or substitutions are all included in the protection scope of the present application. In the absence of conflict, the implementation modes of the present application and the features in the implementation modes can also be combined with each other. The protection scope of the present application shall be subject to the description of the claims.
Claims
1. A brand association model training method, characterized in that: The method includes: A brand data augmentation step to obtain positive and negative samples from the collection of company names and trademark names respectively; and The brand association model training step includes inputting the positive samples and negative samples into a specified pre-trained language model to obtain a feature vector, using a discriminator in the GAN to discriminate the feature vector and calculate the loss of the positive and negative samples, and iteratively optimizing the parameters of the pre-trained language model and the GAN based on the loss to obtain a trained brand association model.
2. The brand association model training method according to claim 1, characterized in that: The obtaining of positive samples comprises: A data pair consisting of a company name and a trademark name whose similarity is higher than a preset first threshold is obtained as a positive sample from a company-trademark relationship table containing the correspondence between company names and trademark names using a prescribed similarity calculation model.
3. The brand association model training method according to claim 2, characterized in that: The obtaining of positive samples further comprises: A specified language model is used to obtain data pairs consisting of enterprise names and enterprise abbreviations from an enterprise total-branch relationship table containing the correspondence between enterprise names and enterprise branch names as positive samples.
4. The brand association model training method according to claim 3, characterized in that: Obtaining positive samples includes: The data pairs consisting of the company names and corresponding brand names obtained through manual annotation are taken as positive samples.
5. The brand association model training method according to claim 4, characterized in that: Obtaining negative samples from the enterprise trademark relationship table includes: Data pairs consisting of enterprise names and trademark names are randomly selected from the enterprise-trademark relationship table as negative samples.
6. The brand association model training method according to claim 5, characterized in that: Acquiring negative samples from the enterprise trademark relationship table further includes: A prescribed similarity calculation model is used to obtain data pairs consisting of enterprise names and trademark names whose similarity is lower than a preset second threshold from the enterprise-trademark relationship table as negative samples.
7. The brand association model training method according to claim 6, characterized in that: Acquiring negative samples in the brand data enhancement step further includes: generating two random feature vectors by using a GAN generator to generate random noise, The brand association model training step includes: Inputting the negative sample into a specified pre-trained language model to obtain a feature vector, and concatenating the feature vectors to obtain a first negative sample set, concatenating the two random feature vectors to obtain a second negative sample set, and merging the first negative sample set and the second negative sample set to obtain a negative sample set; Inputting the positive samples into the specified pre-trained language model to obtain feature vectors, and concatenating the feature vectors to obtain a positive sample set; Using the discriminator in the GAN to discriminate positive and negative labels for the positive sample set and the negative sample set, and calculating the loss of the positive and negative samples; and Parameters of the pre-trained language model and the GAN are iteratively optimized based on the loss to obtain a trained brand association model.
8. The brand association model training method according to claim 6, characterized in that: The brand association model training step includes: Inputting the positive samples and negative samples into a specified pre-trained language model to obtain feature vectors, and concatenating the feature vectors to obtain a positive and negative sample set; Using the discriminator in GAN to discriminate positive and negative labels for the positive and negative sample sets, and calculating the loss of positive and negative samples; and Parameters of the pre-trained language model and the GAN are iteratively optimized based on the loss to obtain a trained brand association model.
9. The brand association model training method according to claim 8, characterized in that: The pre-trained language model is any one of BERT, RoBERTa, and XLNet.
10. A method for completing brand information, characterized in that: The method includes: Get the company names that have no corresponding brands; According to the enterprise name without a corresponding brand, selecting a candidate brand based on character overlap of the enterprise name from an enterprise-trademark relationship table including the correspondence between enterprise names and trademark names and an enterprise-to-branch relationship table including the correspondence between enterprise names and branch names; Calculating the similarity between the company name without a corresponding brand and the candidate brand using the brand association model trained by the brand association model training method according to any one of claims 1 to 9; and A brand association result associated with the enterprise name without a corresponding brand is obtained based on the calculation result of the similarity.
11. A brand association model training system, characterized in that: include: A brand data enhancement module, used to obtain positive samples and negative samples from an enterprise trademark relationship table containing the correspondence between enterprise names and trademark names and an enterprise total score relationship table containing enterprise names and branch names; as well as The brand association model training module inputs the positive samples and negative samples into a specified pre-trained language model to obtain a feature vector, uses a discriminator in the GAN to discriminate the feature vector and calculates the loss of the positive and negative samples, and iteratively optimizes the parameters of the pre-trained language model and the GAN based on the loss to obtain a trained brand association model.
12. The brand association model training system according to claim 11, characterized in that: The brand data enhancement module includes a positive sample acquisition submodule for acquiring positive samples and a negative sample submodule for acquiring negative samples.
13. The brand association model training system according to claim 12, characterized in that: The positive sample acquisition submodule uses a specified similarity calculation model to acquire data pairs consisting of enterprise names and trademark names with a similarity higher than a preset first threshold from the enterprise-trademark relationship table as positive samples.
14. The brand association model training system according to claim 13, characterized in that: The positive sample acquisition submodule uses a specified language model to acquire a data pair consisting of an enterprise name and an enterprise abbreviation from a data containing an enterprise name and an enterprise branch name as a positive sample.
15. The brand association model training system according to claim 14, characterized in that: The positive sample acquisition submodule uses the data pairs consisting of the company names and corresponding brand names obtained through manual annotation as positive samples.
16. The brand association model training system according to claim 15, characterized in that: The negative sample submodule randomly selects data pairs consisting of enterprise names and trademark names from the enterprise-trademark relationship table as negative samples.
17. The brand association model training system according to claim 16, characterized in that: The negative sample submodule uses a specified similarity calculation model to obtain data pairs consisting of enterprise names and trademark names whose similarity is lower than a preset second threshold from the enterprise-trademark relationship table as negative samples.
18. The brand association model training system according to claim 17, characterized in that: The negative sample submodule uses the GAN generator to generate random noise to generate two random feature vectors as negative samples.
19. The brand association model training system according to claim 18, characterized in that: The brand association model training module is used to input the positive samples and negative samples into a specified pre-trained language model and obtain feature vectors, concatenate the feature vectors to obtain positive and negative sample sets, use the discriminator in the GAN to discriminate positive and negative labels for the positive and negative sample sets, and calculate the loss of positive and negative samples, and iteratively optimize the parameters of the pre-trained language model and the GAN based on the loss to obtain a trained brand association model.
20. The brand association model training system according to claim 19, characterized in that: The pre-trained language model is any one of BERT, RoBERTa, and XLNet.
21. A computer readable medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the brand association model training method described in any one of claims 1 to 9 is implemented.
22. A computer device comprising a storage module, a processor, and a computer program stored in the storage module and executable on the processor, characterized in that: When the processor executes the computer program, the brand association model training method described in any one of claims 1 to 9 is implemented.
23. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the brand association model training method described in any one of claims 1 to 9 is implemented.