Commodity publishing method, system and equipment based on small model and large model fusion
Through the product release method that integrates small models and large models, product classification is automated, and the problems of inefficiency and accuracy of traditional manual classification are solved, efficient and accurate product release and market change tracking are achieved, and user experience is improved.
Patent Information
- Application Number
- CN202410983724.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-19
AI Technical Summary
Traditional manual product release and classification are inefficient, prone to errors, and difficult to keep up with market changes, resulting in merchants missing business opportunities and poor user experience.
Using the product release method based on the integration of small models and large models, the product category labels and confidence are automatically determined by pre-trained small models and large models, and the classification is improved by combining advanced classification judgment rules. Accuracy and efficiency.
There is no need for manual classification, which improves the efficiency and accuracy of product releases, keeps up with market changes, avoids merchants missing out on business opportunities, and improves user experience.
Smart Images

Figure CN118982404A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of natural language processing technology, and in particular to a method, system and device for publishing goods based on the fusion of a small model and a large model. Background Art
[0002] With the booming development of the e-commerce industry and the rapid advancement of technology, not only has the speed of product updates increased exponentially, but consumers' shopping habits, demand preferences and market trends are also changing rapidly, which has directly led to the amount of data accumulated in the e-commerce industry reaching an unprecedented scale. Multi-dimensional data such as product descriptions, pictures, videos, user reviews, price change history, etc., also makes product information more and more complex.
[0003] At present, the traditional manual method of product release and classification is becoming increasingly inadequate. This method is not only inefficient and prone to errors, but also difficult to keep up with the pace of market changes, causing merchants to miss business opportunities and affecting user experience. Summary of the invention
[0004] To solve the above problems, the embodiments of the present application provide a product publishing method, system and device based on the fusion of small model and large model.
[0005] On the one hand, an embodiment of the present application provides a product publishing method based on the fusion of a small model and a large model, the method comprising:
[0006] Obtaining initial product information from a user terminal;
[0007] Determine, according to the initial product information and a pre-trained small product classification model, a first product category label and a first classification confidence level corresponding to the initial product information;
[0008] Based on the first commodity category label, the first classification confidence, and the advanced classification judgment rule, determine whether the first commodity category label belongs to the advanced classification label; wherein the advanced classification judgment rule is used to determine whether the first commodity category label and / or the first classification confidence meets the advanced classification standard; the judgment attributes of the advanced classification standard at least include long-tail commodities and confidence threshold;
[0009] If yes, the initial product information is sent to the large model server, so as to obtain a second product category label corresponding to the initial product information based on the product classification large model in the large model server;
[0010] Based on the first commodity category label or the second commodity category label, a third commodity category label corresponding to the initial commodity information is determined, so as to send the third commodity category label to the user terminal.
[0011] In one implementation of the present application, before determining the first product category label and the first classification confidence corresponding to the initial product information according to the initial product information and the pre-trained product classification model, the method further includes:
[0012] Based on the preset joint training strategy and the preset training samples, determine the first key classification data corresponding to the small commodity classification model and the second key classification data corresponding to the large commodity classification model; wherein the first key classification data is the classification data of the small commodity classification model; the second key classification data is the classification data of which the classification accuracy of the large commodity classification model after training and fine-tuning is less than the preset classification threshold;
[0013] Taking the first key classification data as a first sample and taking the second key classification data as a second sample;
[0014] The first sample is input into the large commodity classification model, and the second sample is input into the small commodity classification model, so as to jointly train the small commodity classification model and the large commodity classification model.
[0015] In one implementation of the present application, determining the first product category label and the first classification confidence corresponding to the initial product information according to the initial product information and the pre-trained product classification model specifically includes:
[0016] Determine the product introduction text corresponding to the initial product information;
[0017] The product introduction text is input into the pre-trained product classification model to determine one or more first product category labels corresponding to the product introduction text and their corresponding first classification confidences; the first classification confidence is the Softmax function value obtained by the product classification model based on the first product category label.
[0018] In one implementation of the present application, determining whether the first product category label belongs to an advanced classification label based on the first product category label, the first classification confidence, and an advanced classification judgment rule specifically includes:
[0019] In a case where the initial product information corresponds to one of the first product category labels, determining a long-tail product curve corresponding to the first product category label, and determining whether the first product category label belongs to an advanced classification label based on a determination result of whether the initial product information belongs to a long-tail product and the first classification confidence;
[0020] In the case that the initial product information corresponds to multiple first product category labels, the long-tail product curves corresponding to each of the first product category labels are determined respectively, so as to determine whether all of the first product category labels belong to the advanced classification label based on the judgment result that the initial product information belongs to the long-tail product under each of the first product category labels and the confidence of each of the first classifications.
[0021] In one implementation of the present application, based on the determination result of whether the initial product information belongs to a long-tail product and the first classification confidence, determining whether the first product category label belongs to an advanced classification label specifically includes:
[0022] When the initial product information belongs to the long-tail product or the initial product information does not belong to the long-tail product and the first classification confidence is less than a first preset threshold, determining that the first product category label belongs to the advanced classification label;
[0023] Otherwise, the first product category label does not belong to the advanced classification label;
[0024] Based on the determination result that the initial product information belongs to the long-tail product under each of the first product category labels and each of the first classification confidences, determining whether all of the first product category labels belong to the advanced classification label specifically includes:
[0025] When the initial product information, under the corresponding first product category labels, all belong to the long-tail products, or the first classification confidences are all less than the first preset threshold, or the first classification confidences corresponding to non-long-tail products are less than the first preset threshold, it is determined that all the first product category labels belong to the advanced classification label;
[0026] Otherwise, it is determined that the first product category label corresponding to the non-long-tail product and the first classification confidence is greater than or equal to the first preset threshold does not belong to the advanced classification label.
[0027] In one implementation of the present application, the method further includes:
[0028] In the case where it is determined that the first product category label does not belong to the advanced category label, determining the number of non-advanced category labels corresponding to the first product category label;
[0029] When the number of the non-advanced classification tags is greater than 1, determining the sales volume curves of the products corresponding to the initial product information in the first product category tags;
[0030] Determine the ranking of the product corresponding to the initial product information in each of the sales volume curves;
[0031] According to the ranking, each of the first commodity category labels is sequentially added to the sales label update sequence.
[0032] In one implementation of the present application, determining a third product category label corresponding to the initial product information based on the first product category label or the second product category label specifically includes:
[0033] In the case where the first product category label does not belong to the advanced classification label, using one of the first product category labels as the third product category label;
[0034] In the case that the first product category label belongs to the advanced classification label, the second product category label is used as the third product category label.
[0035] In one implementation of the present application, the method further includes:
[0036] In the case where the sales tag update sequence exists or there are multiple second commodity category tags, the continuous life cycle corresponding to each commodity category tag is determined based on a preset sales prediction model; wherein the continuous life cycle includes a preset time period in which the sales volume is greater than a second preset threshold within a preset time;
[0037] According to each of the continuous life cycles and each of the product category tags, in chronological order, each of the product category tags is used as the third product category tag, and a product category tag update sequence is generated and sent to the user terminal, so as to replace the tags at the end of the continuous life cycle corresponding to each of the product category tags according to the product category tag update sequence.
[0038] On the other hand, the embodiment of the present application further provides a product release system based on the fusion of a small model and a large model, the system comprising:
[0039] An acquisition module, used to acquire initial commodity information from a user terminal;
[0040] A first determination module, configured to determine a first commodity category label and a first classification confidence corresponding to the initial commodity information according to the initial commodity information and a pre-trained commodity classification small model;
[0041] A second determination module is used to determine whether the first commodity category label belongs to an advanced classification label based on the first commodity category label, the first classification confidence, and an advanced classification judgment rule; wherein the advanced classification judgment rule is used to determine whether the first commodity category label and / or the first classification confidence meets an advanced classification standard; and the judgment attributes of the advanced classification standard at least include long-tail commodities and a confidence threshold;
[0042] a sending module, configured to send the initial commodity information to a large model server, so as to obtain a second commodity category label corresponding to the initial commodity information based on a commodity classification large model in the large model server;
[0043] The third determining module is used to determine a third commodity category label corresponding to the initial commodity information based on the first commodity category label or the second commodity category label, so as to send the third commodity category label to the user terminal.
[0044] On the other hand, the embodiment of the present application further provides a product release device based on the fusion of a small model and a large model, the device comprising:
[0045] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned method for releasing goods based on the fusion of small models and large models.
[0046] Compared with the prior art, the present invention has the following significant effects:
[0047] Through the above solution, this application can obtain the first product category label or the second product category label of the product by fusing the small model with the large model, and then provide the third product category label to the store to realize product release. There is no need for manual product classification, which improves the efficiency and accuracy of product release, and enables product classification to keep up with market changes, avoiding merchants from missing business opportunities and improving user experience. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:
[0049] Figure 1 A flowchart of a method for publishing products based on the fusion of a small model and a large model in an embodiment of the present application;
[0050] Figure 2 This is a structural schematic diagram of a commodity publishing device based on the fusion of a small model and a large model in an embodiment of the present application;
[0051] Figure 3 This is a structural diagram of a product publishing system based on the fusion of a small model and a large model in an embodiment of the present application;
[0052] Figure 4This is a structural diagram of a product publishing device based on the fusion of a small model and a large model in an embodiment of the present application. DETAILED DESCRIPTION
[0053] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.
[0054] The embodiments of the present application provide a product publishing method, system and device based on the fusion of small models and large models, which are used to solve the problems that the current traditional manual product publishing and classification methods are inefficient, have high error rates, are difficult to adapt to market changes, and affect user experience.
[0055] The following describes in detail various embodiments of the present application in conjunction with the accompanying drawings.
[0056] The embodiment of the present application provides a product publishing method based on the fusion of small model and large model, such as Figure 1 As shown, the method may include steps S101-S105:
[0057] S101, the server obtains initial product information from a user terminal.
[0058] It should be noted that the server, as the executor of the product publishing method based on the fusion of small models and large models, is only an example. The executor is not limited to the server, and this application does not make any specific limitations on this.
[0059] The user terminal can be understood as the merchant's mobile phone, computer and other devices for publishing goods, and this application does not make specific restrictions on this. The initial product information can be information about the product uploaded by the user terminal, such as the product name, product specifications, etc., or the product appearance image, product description text, etc., and this application does not make specific restrictions on this.
[0060] S102: The server determines a first product category label and a first classification confidence level corresponding to the initial product information based on the initial product information and a pre-trained small product classification model.
[0061] The small product classification model can be a pre-trained FastTest classification model or a MobileBERT model. This application does not make any specific limitations on this. It can perform preliminary screening and classification based on the initial product information to obtain the first product category label.
[0062] In the embodiment of the present application, before determining the first commodity category label and the first classification confidence corresponding to the initial commodity information according to the initial commodity information and the pre-trained commodity classification model, the method further includes:
[0063] Based on the preset joint training strategy and the preset training samples, determine the first key classification data corresponding to the small commodity classification model and the second key classification data corresponding to the large commodity classification model. The first key classification data is the classification data of the small commodity classification model. The second key classification data is the classification data whose classification accuracy is less than the preset classification threshold after the large commodity classification model is trained and fine-tuned. The first key classification data is used as the first sample, and the second key classification data is used as the second sample. The first sample is input into the large commodity classification model, and the second sample is input into the small commodity classification model, so as to jointly train the small commodity classification model and the large commodity classification model.
[0064] In other words, the present application can jointly train and fine-tune the small model for commodity classification and the large model for commodity classification through joint training. The present application can add the first key classification data that is misclassified by the small model for commodity classification to the training of the large model for commodity classification, so as to enhance the analytical reasoning ability of the large model for such data. The large model for commodity classification and the second key classification data with poor fine-tuning effect are put into the small model for commodity classification for repeated training and parameter adjustment, and the weight of these classification data is increased. The training of the large model and the small model on each other's misclassified data is combined to improve the accuracy of the model as a whole. The above-mentioned preset classification threshold can be set by the user according to the actual usage scenario, and the present application does not make specific limitations on this.
[0065] The above technical solution can optimize the collaboration between the two models, and adopt a joint training strategy to enable the small model and the large model to learn from each other during the training process, especially to perform collaborative optimization at the feature level. The model fusion step can use weighted averaging, voting mechanism or ensemble learning method to combine the outputs of the large model and the small model.
[0066] This application also reduces the storage and computing requirements of large models through technologies such as model quantization and sparsification, and uses edge computing and containerization technologies to achieve flexible and efficient deployment strategies to ensure that the system can run efficiently on various hardware platforms. Modeling: Convert the weights in the model from high precision to low precision to reduce the model size and computing resource requirements. Sparsification: Remove unimportant weights in the model through pruning technology to make the model sparse, thereby reducing the amount of calculation and storage requirements. Structured sparsification or unstructured sparsification can be used, the former is more conducive to hardware acceleration. Edge computing: Decentralize computing tasks from the cloud to the device side or the edge of the network to reduce data transmission delays and bandwidth requirements. This requires the model to be lightweight enough to run on resource-constrained devices. Containerization technology: Use container technologies such as Docker to package the model and its operating environment to ensure that the model can run seamlessly on different hardware platforms and improve deployment efficiency and flexibility. By comprehensively applying these technologies and strategies, the resource consumption of large models can be effectively reduced, and their deployment efficiency and operating performance on various hardware platforms can be improved.
[0067] In the embodiment of the present application, according to the initial product information and the pre-trained product classification model, determining the first product category label and the first classification confidence corresponding to the initial product information specifically includes:
[0068] Determine the product introduction text corresponding to the initial product information. Input the product introduction text into the pre-trained product classification model to determine one or more first product category labels corresponding to the product introduction text and their corresponding first classification confidences. The first classification confidence is the Softmax function value obtained by the product classification model based on the first product category label.
[0069] In other words, the server can extract the product introduction text in the initial product information. For example, if the initial product information is a picture, the server can obtain the product introduction text therein through image recognition technology. The product introduction text can be a product name or a product code, etc. The server inputs the product introduction text into a trained small product classification model for processing. The small product classification model can be trained by using a number of product introduction text samples marked with product category labels. When the small product classification model outputs the product category label, it can also output its corresponding classification confidence, thereby obtaining a first product category label and a first classification confidence.
[0070] S103: The server determines whether the first product category label is an advanced category label based on the first product category label, the first category confidence, and the advanced category judgment rule.
[0071] The advanced classification judgment rule is used to judge whether the first product category label and / or the first classification confidence level meet the advanced classification standard. The judgment attributes of the advanced classification standard at least include long-tail products and confidence threshold.
[0072] In the embodiment of the present application, based on the first product category label, the first classification confidence, and the advanced classification judgment rule, determining whether the first product category label belongs to the advanced classification label specifically includes:
[0073] In the case where the initial product information corresponds to one first product category label, the long-tail product curve corresponding to the first product category label is determined, and based on the determination result of whether the initial product information belongs to a long-tail product and the first classification confidence, it is determined whether the first product category label belongs to an advanced classification label. In the case where the initial product information corresponds to multiple first product category labels, each long-tail product curve corresponding to each first product category label is determined respectively, so as to determine whether all first product category labels belong to advanced classification labels based on the determination result of whether the initial product information belongs to a long-tail product under each first product category label and each first classification confidence.
[0074] That is to say, the initial product information may have one first product category label, or may have multiple first product category labels. If the initial product information corresponds to one first product category label, the server can obtain the long-tail product curve where the first product category label is located from the historical product sales data, and determine whether the product corresponding to the initial product information belongs to a long-tail product, and further determine whether the first product category label belongs to an advanced classification label based on the determination result and the first classification confidence. For example, the initial product information corresponds to a water cup, and the first product category label is household items. The server can determine the sales data of each product corresponding to the household items, generate a long-tail product curve, and determine the position of the water cup in the long-tail product curve to determine whether the water cup is a long-tail product in the curve. Long-tail products refer to those products with low sales but a wide variety.
[0075] At the same time, the initial product information output by the model may correspond to multiple first product category labels. The server will also determine whether each first product category label belongs to a long-tail product, and further determine the advanced classification label based on the first classification confidence.
[0076] In the embodiment of the present application, based on the determination result of whether the initial product information belongs to a long-tail product and the first classification confidence, determining whether the first product category label belongs to an advanced classification label specifically includes:
[0077] When the initial product information belongs to a long-tail product or the initial product information does not belong to a long-tail product and the first classification confidence is less than a first preset threshold, it is determined that the first product category label belongs to an advanced classification label. Otherwise, the first product category label does not belong to an advanced classification label.
[0078] In other words, if the product is a long-tail product or the first classification confidence is less than the first preset threshold, the server will determine that the product category label belongs to the advanced classification label. If not, the first product category label does not belong to the advanced classification label.
[0079] Based on the determination result that the initial product information belongs to long-tail products under each first product category label and the confidence of each first classification, determining whether all first product category labels belong to advanced classification labels specifically includes:
[0080] If the initial product information, under each corresponding first product category label, all belong to long-tail products or each first classification confidence is less than the first preset threshold or the first classification confidence corresponding to the non-long-tail product is less than the first preset threshold, it is determined that each first product category label belongs to the advanced classification label. Otherwise, it is determined that the first product category label corresponding to the non-long-tail product and the first classification confidence is greater than or equal to the first preset threshold does not belong to the advanced classification label.
[0081] That is to say, when each first product category label meets the above conditions, the server will use the first product category label as an advanced classification label. If among multiple first product category labels, the first product category label corresponding to a non-long-tail product has a first classification confidence greater than or equal to a first preset threshold, the corresponding first product category label will be used as a non-advanced classification label.
[0082] Through the above scheme, the advanced classification label can be determined simply and quickly.
[0083] In an embodiment of the present application, when it is determined that the first commodity category tag does not belong to an advanced classification tag, the number of non-advanced classification tags of the corresponding first commodity category tag is determined. When the number of non-advanced classification tags is greater than 1, the sales volume curve corresponding to each first commodity category tag of the commodity corresponding to the initial commodity information is determined. The ranking of the commodity corresponding to the initial commodity information in each sales volume curve is determined. According to the ranking, each first commodity category tag is added to the sales tag update sequence in sequence.
[0084] That is, the server can count the number of non-advanced classification tags. If the server determines that there are multiple non-advanced classification tags, it further determines the sales volume curve of the first commodity category tag and determines the ranking of the corresponding commodity in the sales volume curve. The sales volume curve can be understood as the sales volume ranking of each commodity in the first commodity category tag. The server can rank the first commodity category tag according to the ranking noun, thereby generating a sales tag update sequence.
[0085] For example, in the classification of e-commerce products, based on the product name, the small model will first be used to quickly predict a classification for the product and give the confidence of this classification. If the confidence of the small model is less than 70%, it is considered that the small model cannot give a credible answer to the predicted classification of the product, that is, the label is a non-advanced classification label; at this time, the large model is involved and the large model is used to continue to predict the classification of the product. Confidence: The small model will predict the category to which the product name belongs, and use the softmax function to give a predicted probability value for each classification. The one with the largest probability value is considered to be the classification of the product, and the probability value of the classification is the confidence.
[0086] S104, when the server determines that the first product category label is an advanced classification label, the server sends the initial product information to the large model server to obtain a second product category label corresponding to the initial product information based on the product classification large model in the large model server.
[0087] That is to say, the commodity classification large model in this application can be set on the large model server to facilitate calling, and further classify and process the commodities that are not accurately classified by the small model to obtain the second commodity category label. The commodity classification large model can be a large model trained by the chatGLM base model.
[0088] S105: The server determines a third product category tag corresponding to the initial product information based on the first product category tag or the second product category tag, and sends the third product category tag to the user terminal.
[0089] In the embodiment of the present application, determining the third product category label corresponding to the initial product information based on the first product category label or the second product category label specifically includes:
[0090] In the case where the first product category tag does not belong to the advanced classification tag, a first product category tag is used as the third product category tag. In the case where the first product category tag belongs to the advanced classification tag, the second product category tag is used as the third product category tag.
[0091] In other words, when the number of first product category tags that are not advanced category tags is 1, the tag can be used as the third product category tag; if all first product category tags are advanced category tags, the server uses the second product category tag as the third product category tag.
[0092] In addition, when there is a sales tag update sequence or multiple second commodity category tags, the continuous life cycle corresponding to each commodity category tag is determined based on the preset sales prediction model. The continuous life cycle includes a preset time period in which the sales volume is greater than the second preset threshold within the preset time. According to each continuous life cycle and each commodity category tag, in chronological order, each commodity category tag is used as a third commodity category tag, and a commodity category tag update sequence is generated and sent to the user terminal, so that the tag is replaced at the end of the continuous life cycle corresponding to each commodity category tag according to the commodity category tag update sequence.
[0093] That is to say, the server can train a neural network model through a number of commodity sales data samples and the continuous life cycle of the annotations to obtain a preset sales prediction model. The continuous life cycle corresponding to each commodity category label is determined by the preset sales prediction model, and the server arranges the continuous life cycles corresponding to different commodity category labels to generate a commodity category label update sequence of the third commodity category label within a preset time. The commodity category label update sequence is composed of different commodity category labels or multiple second commodity category labels in the sales label update sequence. In different periods within the preset time, the commodity has different commodity category labels, thereby realizing label replacement.
[0094] Through the above solution, this application can obtain the first product category label or the second product category label of the product by fusing the small model with the large model, and then provide the third product category label to the store to realize product release. There is no need for manual product classification, which improves the efficiency and accuracy of product release, and enables product classification to keep up with market changes, avoiding merchants from missing business opportunities and improving user experience.
[0095] The device corresponding to the present application can be composed of Figure 2 The system consists of a text preprocessing module, a small model preclassification module, a large model fine classification module, a fusion decision module, and an optimization deployment module. The preprocessing module is responsible for text cleaning and standardization; the small model preclassification module quickly filters text; the large model fine classification module accurately classifies the preclassification results; the fusion decision module decides whether to call the large model based on the confidence of the small model output; and the optimization deployment module ensures the efficient operation of the entire system.
[0096] Figure 3 A structural diagram of a product publishing system 300 based on the fusion of a small model and a large model provided in an embodiment of the present application, such as Figure 3 As shown, the system includes:
[0097] The acquisition module 301 is used to obtain the initial product information from the user terminal. The first determination module 302 is used to determine the first product category label and its first classification confidence corresponding to the initial product information based on the initial product information and the pre-trained product classification small model. The second determination module 303 is used to determine whether the first product category label belongs to the advanced classification label based on the first product category label, the first classification confidence, and the advanced classification judgment rule. Among them, the advanced classification judgment rule is used to determine whether the first product category label and / or the first classification confidence meets the advanced classification standard. The judgment attributes of the advanced classification standard include at least long-tail products and confidence thresholds. The sending module 304 is used to send the initial product information to the large model server if so, so as to obtain the second product category label corresponding to the initial product information based on the product classification large model in the large model server. The third determination module 305 is used to determine the third product category label corresponding to the initial product information based on the first product category label or the second product category label, so as to send the third product category label to the user terminal.
[0098] In the embodiment of the present application, the system can also:
[0099] Based on the preset joint training strategy and the preset training samples, determine the first key classification data corresponding to the small commodity classification model and the second key classification data corresponding to the large commodity classification model. The first key classification data is the classification data of the small commodity classification model. The second key classification data is the classification data whose classification accuracy is less than the preset classification threshold after the large commodity classification model is trained and fine-tuned. The first key classification data is used as the first sample, and the second key classification data is used as the second sample. The first sample is input into the large commodity classification model, and the second sample is input into the small commodity classification model, so as to jointly train the small commodity classification model and the large commodity classification model.
[0100] In the embodiment of the present application, the first determining module 302 is specifically used for:
[0101] Determine the product introduction text corresponding to the initial product information. Input the product introduction text into the pre-trained product classification model to determine one or more first product category labels corresponding to the product introduction text and their corresponding first classification confidences. The first classification confidence is the Softmax function value obtained by the product classification model based on the first product category label.
[0102] In the embodiment of the present application, the second determining module 303 is specifically used to:
[0103] In the case where the initial product information corresponds to one first product category label, the long-tail product curve corresponding to the first product category label is determined, and based on the determination result of whether the initial product information belongs to a long-tail product and the first classification confidence, it is determined whether the first product category label belongs to an advanced classification label. In the case where the initial product information corresponds to multiple first product category labels, each long-tail product curve corresponding to each first product category label is determined respectively, so as to determine whether all first product category labels belong to advanced classification labels based on the determination result of whether the initial product information belongs to a long-tail product under each first product category label and each first classification confidence.
[0104] In the embodiment of the present application, the second determining module 303 is further configured to:
[0105] When the initial product information belongs to a long-tail product or the initial product information does not belong to a long-tail product and the first classification confidence is less than a first preset threshold, it is determined that the first product category label belongs to an advanced classification label. Otherwise, the first product category label does not belong to an advanced classification label.
[0106] In the embodiment of the present application, the second determining module 303 is further configured to:
[0107] If the initial product information, under each corresponding first product category label, all belong to long-tail products or each first classification confidence is less than the first preset threshold or the first classification confidence corresponding to the non-long-tail product is less than the first preset threshold, it is determined that each first product category label belongs to the advanced classification label. Otherwise, it is determined that the first product category label corresponding to the non-long-tail product and the first classification confidence is greater than or equal to the first preset threshold does not belong to the advanced classification label.
[0108] In the embodiment of the present application, the system can also:
[0109] When it is determined that the first product category tag does not belong to the advanced classification tag, the number of non-advanced classification tags of the corresponding first product category tag is determined. When the number of non-advanced classification tags is greater than 1, the sales volume curves corresponding to the products corresponding to the initial product information in each first product category tag are determined. The ranking of the products corresponding to the initial product information in each sales volume curve is determined. According to the ranking, each first product category tag is added to the sales tag update sequence in sequence.
[0110] In the embodiment of the present application, the third determination module 305 is specifically used to:
[0111] In the case where the first product category tag does not belong to the advanced classification tag, a first product category tag is used as the third product category tag. In the case where the first product category tag belongs to the advanced classification tag, the second product category tag is used as the third product category tag.
[0112] In the embodiment of the present application, the system can also:
[0113] In the case where there is a sales tag update sequence or multiple second commodity category tags, the continuous life cycle corresponding to each commodity category tag is determined based on the preset sales prediction model. The continuous life cycle includes a preset time period in which the sales volume is greater than the second preset threshold within the preset time. According to each continuous life cycle and each commodity category tag, in chronological order, each commodity category tag is used as a third commodity category tag, and a commodity category tag update sequence is generated and sent to the user terminal, so that the tag is replaced at the end of the continuous life cycle corresponding to each commodity category tag according to the commodity category tag update sequence.
[0114] Figure 4 A structural diagram of a commodity publishing device based on the fusion of a small model and a large model provided in an embodiment of the present application, such as Figure 4 As shown, the device includes:
[0115] At least one processor; and a memory in communication with the at least one processor. The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to:
[0116] Obtain initial product information from the user terminal. Determine the first product category label and its first classification confidence corresponding to the initial product information based on the initial product information and the pre-trained product classification small model. Determine whether the first product category label belongs to the advanced classification label based on the first product category label, the first classification confidence, and the advanced classification judgment rule. Among them, the advanced classification judgment rule is used to determine whether the first product category label and / or the first classification confidence meets the advanced classification standard. The judgment attributes of the advanced classification standard include at least long-tail products and confidence thresholds. If so, send the initial product information to the large model server to obtain the second product category label corresponding to the initial product information based on the product classification large model in the large model server. Based on the first product category label or the second product category label, determine the third product category label corresponding to the initial product information, and send the third product category label to the user terminal.
[0117] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system and device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.
[0118] The system and equipment provided in the embodiments of the present application correspond one-to-one to the method. Therefore, the system and equipment also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the system and equipment will not be repeated here.
[0119] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.
[0120] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.
Claims
1. A product release method based on the fusion of small model and large model, characterized in that: The method comprises: Obtaining initial product information from a user terminal; Determine, according to the initial product information and a pre-trained small product classification model, a first product category label and a first classification confidence level corresponding to the initial product information; Based on the first commodity category label, the first classification confidence, and the advanced classification judgment rule, determine whether the first commodity category label belongs to the advanced classification label; wherein the advanced classification judgment rule is used to determine whether the first commodity category label and / or the first classification confidence meets the advanced classification standard; the judgment attributes of the advanced classification standard at least include long-tail commodities and confidence threshold; If yes, the initial product information is sent to the large model server, so as to obtain a second product category label corresponding to the initial product information based on the product classification large model in the large model server; Based on the first commodity category label or the second commodity category label, a third commodity category label corresponding to the initial commodity information is determined, so as to send the third commodity category label to the user terminal.
2. According to claim 1, a method for releasing goods based on the fusion of small model and large model is characterized in that: Before determining the first commodity category label and the first classification confidence corresponding to the initial commodity information according to the initial commodity information and the pre-trained commodity classification model, the method further includes: Based on the preset joint training strategy and the preset training samples, determine the first key classification data corresponding to the small commodity classification model and the second key classification data corresponding to the large commodity classification model; wherein the first key classification data is the classification data of the small commodity classification model; the second key classification data is the classification data of which the classification accuracy of the large commodity classification model after training and fine-tuning is less than the preset classification threshold; Taking the first key classification data as a first sample and taking the second key classification data as a second sample; The first sample is input into the large commodity classification model, and the second sample is input into the small commodity classification model, so as to jointly train the small commodity classification model and the large commodity classification model.
3. According to claim 1, a method for releasing goods based on the fusion of small model and large model is characterized in that: Determining a first product category label and a first classification confidence level corresponding to the initial product information according to the initial product information and a pre-trained product classification model, specifically includes: Determine the product introduction text corresponding to the initial product information; The product introduction text is input into the pre-trained product classification model to determine one or more first product category labels corresponding to the product introduction text and their corresponding first classification confidences; the first classification confidence is the Softmax function value obtained by the product classification model based on the first product category label.
4. A method for releasing goods based on the fusion of small model and large model according to claim 3, characterized in that: Determining whether the first product category label is an advanced category label based on the first product category label, the first category confidence, and an advanced category judgment rule specifically includes: In a case where the initial product information corresponds to one of the first product category labels, determining a long-tail product curve corresponding to the first product category label, and determining whether the first product category label belongs to an advanced classification label based on a determination result of whether the initial product information belongs to a long-tail product and the first classification confidence; In the case that the initial product information corresponds to multiple first product category labels, the long-tail product curves corresponding to each of the first product category labels are determined respectively, so as to determine whether all of the first product category labels belong to the advanced classification label based on the judgment result that the initial product information belongs to the long-tail product under each of the first product category labels and the confidence of each of the first classifications.
5. According to claim 4, a method for releasing goods based on the fusion of small model and large model is characterized in that: Based on the determination result of whether the initial product information belongs to a long-tail product and the first classification confidence, determining whether the first product category label belongs to an advanced classification label specifically includes: When the initial product information belongs to the long-tail product or the initial product information does not belong to the long-tail product and the first classification confidence is less than a first preset threshold, determining that the first product category label belongs to the advanced classification label; Otherwise, the first product category label does not belong to the advanced classification label; Based on the determination result that the initial product information belongs to the long-tail product under each of the first product category labels and each of the first classification confidences, determining whether all of the first product category labels belong to the advanced classification label specifically includes: When the initial product information, under the corresponding first product category labels, all belong to the long-tail products, or the first classification confidences are all less than the first preset threshold, or the first classification confidences corresponding to non-long-tail products are less than the first preset threshold, it is determined that all the first product category labels belong to the advanced classification label; Otherwise, it is determined that the first product category label corresponding to the non-long-tail product and the first classification confidence is greater than or equal to the first preset threshold does not belong to the advanced classification label.
6. A method for releasing goods based on the fusion of small model and large model according to claim 5, characterized in that: The method further comprises: In the case where it is determined that the first product category label does not belong to the advanced category label, determining the number of non-advanced category labels corresponding to the first product category label; When the number of the non-advanced classification tags is greater than 1, determining the sales volume curves of the products corresponding to the initial product information in the first product category tags; Determine the ranking of the product corresponding to the initial product information in each of the sales volume curves; According to the ranking, each of the first commodity category labels is sequentially added to the sales label update sequence.
7. A method for releasing goods based on the fusion of small model and large model according to claim 6, characterized in that: Determining a third product category label corresponding to the initial product information based on the first product category label or the second product category label specifically includes: In the case where the first product category label does not belong to the advanced classification label, using one of the first product category labels as the third product category label; In the case that the first product category label belongs to the advanced classification label, the second product category label is used as the third product category label.
8. A method for releasing goods based on the fusion of small model and large model according to claim 7, characterized in that: The method further comprises: In the case where the sales tag update sequence exists or there are multiple second commodity category tags, the continuous life cycle corresponding to each commodity category tag is determined based on a preset sales prediction model; wherein the continuous life cycle includes a preset time period in which the sales volume is greater than a second preset threshold within a preset time; According to each of the continuous life cycles and each of the product category tags, in chronological order, each of the product category tags is used as the third product category tag, and a product category tag update sequence is generated and sent to the user terminal, so as to replace the tags at the end of the continuous life cycle corresponding to each of the product category tags according to the product category tag update sequence.
9. A product release system based on the fusion of small model and large model, characterized in that: The system comprises: An acquisition module, used to acquire initial commodity information from a user terminal; A first determination module, configured to determine a first commodity category label and a first classification confidence corresponding to the initial commodity information according to the initial commodity information and a pre-trained commodity classification small model; A second determination module is used to determine whether the first commodity category label belongs to an advanced classification label based on the first commodity category label, the first classification confidence, and an advanced classification judgment rule; wherein the advanced classification judgment rule is used to determine whether the first commodity category label and / or the first classification confidence meets an advanced classification standard; and the judgment attributes of the advanced classification standard at least include long-tail commodities and a confidence threshold; a sending module, configured to send the initial commodity information to a large model server, so as to obtain a second commodity category label corresponding to the initial commodity information based on a commodity classification large model in the large model server; The third determining module is used to determine a third commodity category label corresponding to the initial commodity information based on the first commodity category label or the second commodity category label, so as to send the third commodity category label to the user terminal.
10. A commodity publishing device based on the fusion of small model and large model, characterized in that: The device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a product release method based on the fusion of a small model and a large model as described in any one of claims 1 to 8 above.
