Automatic product feature extraction method and system based on large model semantic understanding

Abstract

The present invention belongs to the technical field of digital text information processing, and specifically relates to a method and system for automatically extracting product features based on large-scale model semantic understanding. The method utilizes existing product information to construct a training set and a verification set to achieve in-depth understanding and semantic extraction of product information. During feature proofreading, the feature vectors are divided into key vectors and non-key vectors, and the key and non-key parts of the product features are analyzed in detail. Weights are assigned according to their importance to further optimize the extraction results, and the changing trends of the feature vectors during the proofreading process are dynamically captured, thereby constructing an efficient small sample set to help screen out the optimal feature extraction model. Subsequently, when extracting new product features, multi-step extraction or single-step extraction can be selected based on the differences in the features, thereby improving the accuracy and efficiency of feature extraction. In practical applications, product features that meet the requirements can be extracted more quickly and accurately.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital text information processing, and in particular relates to a method and system for automatically extracting product features based on large-scale model semantic understanding. Background Art

[0002] With the rapid development of e-commerce platforms, the amount of product information has exploded. How to automatically extract representative feature information from massive product descriptions has become one of the key technologies to improve product retrieval efficiency, recommendation accuracy and user experience.

[0003] Traditional product feature extraction methods mainly rely on manual rules, keyword matching or shallow machine learning models. When faced with diverse, unstructured and semantically complex product description texts, these methods often suffer from low recognition accuracy, poor generalization ability and insufficient semantic understanding of long texts.

[0004] In recent years, with the advancement of natural language processing technology, pre-trained language models (such as BERT and GPT) have achieved significant breakthroughs in semantic understanding. These large models offer advantages such as strong contextual modeling capabilities and refined semantic representation, providing a new technical approach for extracting semantic features from product text. However, existing methods based on these large models are primarily used for general tasks such as text classification and named entity recognition. They lack specific design for feature extraction in the context of e-commerce products, particularly in handling mixed expressions of multiple product attributes, colloquial descriptions, and parsing implicit intent.

[0005] Furthermore, some current semantic extraction schemes fail to fully consider the impact of product category context on feature semantic weights, potentially misclassifying the same keyword as irrelevant attributes across different categories, impacting overall recognition accuracy. Therefore, there is an urgent need for an automated feature extraction method that combines the deep semantic understanding capabilities of large models with the context of product categories to achieve higher-quality, more robust product structured processing. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for automatically extracting product features based on large-scale model semantic understanding, which can improve the accuracy and efficiency of product feature extraction.

[0007] The technical solution adopted by the present invention is as follows: a method for automatically extracting product features based on large-scale model semantic understanding, comprising:

[0008] S1. Product information collection: obtaining product information of existing products; wherein the product information includes image information and text book information;

[0009] S2. Image Semantic Encoding Clustering: Input image information into a pre-trained large model to obtain deep features of the image information. These deep features are then clustered to obtain high-dimensional semantics that match each existing product.

[0010] S3. Model fine-tuning, proofreading, and optimization: Using high-dimensional semantics as input, the large model undergoes multiple rounds of fine-tuning. Proofreading is performed after each round of fine-tuning to obtain the target model for extracting product features.

[0011] S4. Target product semantic mapping feature matching: Input the image information of the product to be extracted into the target model to obtain the high-dimensional semantics of the product to be extracted. Then input the high-dimensional semantics into the target model to match the product features corresponding to the high-dimensional semantics and mark them as reference features of the product to be extracted.

[0012] S5. Feature confidence analysis, extraction and verification: The reference features are aggregated into a data set to be verified, and the text book information in the data set to be verified is input into the target model one by one to obtain the confidence score of each reference feature. The extraction method of each reference feature is determined based on the confidence score, and then verification extraction is performed to obtain the key features corresponding to each reference feature of the product to be extracted.

[0013] In the present invention, preferably, the step of clustering the deep features to obtain high-dimensional semantics matching each existing product includes:

[0014] Classify the deep features according to product information, and then perform offset operations on the deep feature vectors in the same category to obtain several baseline feature vectors;

[0015] Obtain a benchmark cluster radius, and using the benchmark eigenvector as a benchmark, calculate the Euclidean distance between each depth eigenvector and the benchmark eigenvector, and mark it as the first deviation;

[0016] Compare the first deviation with the benchmark cluster radius. If the first deviation is greater than or equal to the benchmark cluster radius, classify the corresponding depth feature vector into other categories. Then, match it to other categories based on the Euclidean distance between the first deviation and the benchmark feature vector in other categories.

[0017] If the first deviation is smaller than the reference cluster radius, it indicates that the corresponding depth feature vector and the reference feature vector belong to the same category.

[0018] In the present invention, preferably, the step of performing multiple rounds of fine-tuning on the large model using high-dimensional semantics as input includes:

[0019] Construct training sets and validation sets based on each high-dimensional semantics;

[0020] Obtain the number of benchmark calibration rounds, input the training set within the benchmark calibration rounds into the large model, then input the validation set into the large model to verify the accuracy of the large model output results, and mark them as the first calibration parameters;

[0021] The first calibration parameter is compared with a preset calibration threshold, and if the first calibration parameter is greater than the calibration threshold within a reference calibration round number, the current large model is marked as a model to be calibrated.

[0022] In the present invention, preferably, after the step of marking the current large model as a model to be proofread when the first proofreading parameter is greater than the proofreading threshold within the benchmark proofreading rounds, the method further includes:

[0023] Obtain the models to be proofread whose first proofreading parameter is greater than or equal to the proofreading threshold within the benchmark proofreading round number, and sort the models to be proofread from largest to smallest according to the size of the first proofreading parameter;

[0024] The highest-ranked model to be proofread is marked as the model to be output, and the difference between the first proofreading parameter lower than the model to be output and the proofreading threshold within the benchmark proofreading rounds is marked as the parameter to be evaluated;

[0025] Obtain the model reference deviation interval and compare the upper limit of the model reference deviation interval with the parameter to be evaluated. If the parameter to be evaluated is within the model reference deviation interval, mark the corresponding large model as an output model after fine-tuning.

[0026] If the parameter to be evaluated is greater than the upper limit of the model reference deviation interval, the calibration operation will continue to be performed on the corresponding large model.

[0027] In the present invention, preferably, the step of performing a proofreading operation after the proofreading is completed includes:

[0028] Obtain the first calibration parameter that is higher than the corresponding parameter of the model to be output during the fine-tuning process, and mark it as the parameter to be calibrated;

[0029] Obtain the validation set wear corresponding to the parameter to be proofread, and mark the difference between the parameter to be proofread and the corresponding parameter of the output model as the amount to be proofread, and then divide the amount to be proofread into multiple intervals to be tested according to the validation set wear;

[0030] Compare the parameters corresponding to the model to be output with the interval to be checked with the highest order. If the interval to be checked is lower than the parameters corresponding to the model to be output, compare the parameters to be checked with the interval to be checked with the second highest order.

[0031] If the quantity to be calibrated is within the test interval of the corresponding parameter of the model to be output, then the corresponding verification set is input one by one into the large model corresponding to the quantity to be calibrated for verification;

[0032] If the performance parameters of the large model corresponding to the interval to be inspected with the second highest order are lower than the corresponding parameters of the large models of other intervals to be inspected during the verification and proofreading process, the verification will be stopped.

[0033] In the present invention, preferably, the step of determining the extraction method of each reference feature based on the confidence score includes:

[0034] Obtain confidence parameter values for each reference feature, and determine feature codes corresponding to the reference features based on each confidence parameter, and divide the feature codes into first-order codes and second-order codes;

[0035] Execute a first comparison mode, wherein the first comparison mode includes:

[0036] Obtaining the first-order codes of all feature codes, and comparing the feature code of the output result of the previous image information to be verified of the product to be extracted with the feature code of the output result of the current image information to be verified;

[0037] If the feature codes before and after extraction are consistent, the output second-order codes are consistent, that is, the output result of the previous image information to be verified of the extracted product is consistent with the output result of the current image information to be verified;

[0038] If the feature codes before and after extraction are inconsistent, the output second-order codes are inconsistent, that is, the output result of the previous image information to be verified of the extracted product is consistent with the output result of the current image information to be verified;

[0039] According to the output result of the second-order coding, the corresponding feature extraction method is executed. When the second-order coding is consistent, the output feature is verified according to the product level of the product to be extracted. When the second-order coding is inconsistent, the feature with a relatively small confidence parameter is extracted and output as a non-critical extracted feature, and the other extracted feature is output as a critical extracted feature.

[0040] In the present invention, preferably, if the second-order encodings among multiple features are consistent, the output results of the confidence parameters of each feature extraction are sorted, and the feature corresponding to the lowest-ranked confidence parameter is marked as the key extraction feature. In the case of the same rank, the corresponding key extraction feature is output according to the product level of the product to be extracted.

[0041] In the present invention, preferably, the target model supports extracting key features and non-key features from the commodities to be extracted, and the extracted commodities are commodities whose necessary key features are not supported by the target model;

[0042] The key extraction features include product name, product model, manufacturer, etc., and the non-key extraction features include product packaging, promotional information, etc.

[0043] In the present invention, preferably, the confidence parameters of the extraction method are divided into primary confidence parameters, secondary confidence parameters, and tertiary confidence parameters, with primary confidence parameters being prioritized, and the target model is optimized according to the confidence parameters of the selected features;

[0044] The tuning does not support the positioning requirements of necessary features, collects real image data for investigation, provides weights, and generates and outputs a custom target model accordingly.

[0045] The present invention also provides a system for automatically extracting product features based on large-scale model semantic understanding, which is used to implement the method for automatically extracting product features based on large-scale model semantic understanding, including:

[0046] A product information collection module is used to obtain product information of existing products, wherein the product information includes picture information and text book information;

[0047] The image semantic encoding and clustering module is used to input the image information into the pre-trained large model, extract the corresponding deep features, and perform clustering processing on the deep features to obtain high-dimensional semantics that match each existing product;

[0048] A model fine-tuning and proofreading optimization module, configured to use the high-dimensional semantics as input, perform multiple rounds of fine-tuning on the large model, and perform proofreading after each round of fine-tuning, thereby obtaining a target model for extracting product features;

[0049] The product semantic mapping and feature matching module is used to input the image information of the product to be extracted into the target model to obtain the corresponding high-dimensional semantics, and match the product features corresponding to the high-dimensional semantics through the target model and mark them as reference features;

[0050] A feature confidence analysis and verification module is used to aggregate the reference features into a dataset to be verified, input the text book information in the dataset into the target model one by one, calculate the confidence score of each reference feature; determine the feature extraction method based on the confidence score, and complete the final verification extraction output of the key features;

[0051] Among them, the key features include at least the product name, product model, and manufacturer, and the non-key features include at least the product packaging and promotional information.

[0052] As a preferred technical solution, the image semantic coding and clustering module specifically includes:

[0053] The feature classification unit is used to preliminarily classify the extracted deep features according to product information, and perform vector offset processing on each deep feature vector in each category to generate several baseline feature vectors;

[0054] A cluster determination unit is configured to obtain a reference cluster radius and, with the reference feature vector as the center, calculate the Euclidean distance between other depth feature vectors and the reference feature vector to obtain a first deviation;

[0055] A feature classification unit is used to compare the first deviation with the benchmark clustering radius. If the first deviation is greater than or equal to the clustering radius, the corresponding deep feature vector is classified into other categories and matched to the category according to the Euclidean distance; if the first deviation is less than the clustering radius, it is classified into the current category to complete the high-dimensional semantic clustering processing.

[0056] The present invention also provides a terminal for automatically extracting product features based on large-model semantic understanding, comprising:

[0057] one or more processors;

[0058] A storage device having one or more programs stored thereon, wherein the one or more programs are configured to be executable by one or more processors and executed by the one or more processors so that the terminal for automatically extracting product features based on large model semantic understanding executes the method for automatically extracting product features based on large model semantic understanding.

[0059] The beneficial effects of the present invention are as follows: the present invention uses the features of a large model and existing product information to construct a training set and a verification set, thereby achieving in-depth understanding and semantic extraction of product information. During feature proofreading, the feature vectors are divided into key vectors and non-key vectors, focusing on analyzing the key and non-key parts of the product features, and assigning weights based on their importance to further optimize the extraction results. By constructing a sliding window, the changing trend of the feature vectors during the proofreading process is dynamically captured, thereby constructing an efficient small sample set to help screen out the optimal feature extraction model. Subsequently, when extracting new product features, multi-step extraction or single-step extraction can be selected based on the differences in the features, thereby improving the accuracy and efficiency of feature extraction. In practical applications, product features that meet the requirements can be extracted more quickly and accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Schematic diagram of the method flow of the present invention;

[0061] Figure 2 Schematic diagram of sub-steps of step S2 of the present invention;

[0062] Figure 3 Schematic diagram of sub-steps of step S3 of the present invention;

[0063] Figure 4 Schematic diagram of sub-steps of step S4 of the present invention;

[0064] Figure 5 Schematic diagram of sub-steps of step S5 of the present invention;

[0065] Figure 6 This is a schematic diagram of the sub-steps of step S6 of the present invention. DETAILED DESCRIPTION

[0066] The present invention will be further described below with reference to specific examples. The following examples can enable those skilled in the art to more fully understand the present invention, but they are not intended to limit the present invention.

[0067] Example 1

[0068] This embodiment provides a method for automatically extracting product features based on semantic understanding of a large model, such as Figure 1 As shown, including:

[0069] Step S1: Obtain product information of an existing product, wherein the product information includes image information and text book information;

[0070] In step S1, the collection of product information can be captured or included through multiple online shopping websites or literature databases, wherein the image information can be understood as the appearance of the product, which can be color matching, style, background, etc. The text book information can include the product's identification text, logo or QR code. The product information covers all the characteristics and necessary characteristics of the product, wherein the conversion of image information into necessary information of the product is consistent with the upper limit of the number of image information of the product. The text book information of the corresponding product can be converted into multiple necessary characteristics of the corresponding product. In order to improve execution efficiency, all text book information is first parsed through image information, and then arranged by evaluating the important characteristics of the product from the user's perspective. The secondary confidence parameters require secondary judgment. For example: some key characteristics of the product are likely to be different. The most common ones are size, material, image ratio parameters, etc., which are selected as the first-level feature parameters.

[0071] Step S2: Input the image information into the pre-trained large model to obtain the deep features of the image information, and cluster the deep features to obtain high-dimensional semantics that match each existing product;

[0072] In step S2, after the image information is input into the pre-trained large model, the deep features outputted are used as the understanding of the existing products. Specifically, the degree of understanding of the products can be equated to the first observation of a person. Through cluster analysis, the parts of the specific image that may be consistent with other products in the production process are extracted, such as small parts. These small parts can be sold separately and can also be used as parts of large parts. After adding text book information, they can be distinguished. Small parts are just an example. Correspondingly, the clustering here can be fine-tuned by selecting a small sample set to output the mimicking large model.

[0073] Step S3: Using high-dimensional semantics as input, the large model is fine-tuned multiple times, and a proofreading operation is performed after each round of fine-tuning to obtain a target model for extracting product features.

[0074] In step S3, fine-tuning is a processing method for adjusting the large model. Its purpose is to simulate the deep extraction of image information and then append it to the large model to shorten the large model's extraction speed of these high-frequency features of daily commodities and improve the user experience. Specifically, the fine-tuning process will construct a training set and a validation set. Both can be obtained by crawling or collecting in the first step. During the proofreading process, the accuracy of the large model's understanding is reflected, that is, the output is compared with the text book information.

[0075] Step S4: Input the image information of the product to be extracted into the target model to obtain the high-dimensional semantics of the product to be extracted, then input the high-dimensional semantics into the target model, match the product features corresponding to the high-dimensional semantics, and mark them as reference features of the product to be extracted;

[0076] In step S4, during the proofreading process, the initial recognition is to match the product features corresponding to the input image. The construction of high-dimensional semantics is also achieved through a similar accumulation method, that is, by fine-tuning the model to train a non-scalable model (custom target model) that covers all information, decomposing these necessary features into small samples, and then converting them into important parameters, such as: for subsequent entry features, image segmentation, and quality assessment output. Obviously, the subsequent fine-tuning model will improve the accuracy based on the input features, and will inevitably output a more efficient target custom model.

[0077] Step S5: Summarize the reference features into a data set to be verified, and then input the text book information in the data set to be verified into the target model one by one to obtain the confidence score of each reference feature, and determine the extraction method of each reference feature based on the confidence score, and then perform verification extraction to obtain the key features corresponding to each reference feature of the product to be extracted.

[0078] In step S5, the text book information in the data set to be verified is input into the large model one by one, the user perspective is truly simulated, and the corresponding confidence parameters are derived. The importance and accuracy of the product features are clarified based on the confidence score. The higher the confidence score, the more accurate the corresponding product reference features are, that is, the more key features there are, and vice versa, the more non-key features. The difference in the extraction method lies in the feature parameters, and the ones with higher confidence scores will be retained as the output results. The same emphasis is placed on fine-tuning each aspect of the product as an independent target, and then adjusting the combination according to the corresponding product situation and the corresponding purpose. For example: prioritize identifying the corresponding trademark and other main body identifications, and output comparison parameters. This is combined with the target custom model, and each product can be specified as the most matching result.

[0079] Example 2

[0080] like Figure 2 As shown, in this embodiment, the steps of clustering the deep features to obtain high-dimensional semantics matching each existing product include:

[0081] Step S201: Classify the depth features according to product information, and then perform an offset operation on each depth feature vector in the same category to obtain a number of reference feature vectors;

[0082] In step S201, the depth features are classified according to their corresponding product information. The depth features of the same category (such as image information and description information may match the information that has been collected) are offset by utilizing the correlation between the depth features to achieve leveling of the features. It can be understood that extraction is essentially a way to realize the front-end display of image information. Some images that are different but have high similarity are grouped together. The operation here will be reduced to a set, so that it will not cause a large model load and will not lead to frequent adjustment of large model parameters to generate more high-frequency and low-precision extraction schemes.

[0083] Step S202: obtaining a reference cluster radius, and using the reference feature vector as a reference, calculating the Euclidean distance between each depth feature vector and the reference feature vector, and marking it as a first deviation;

[0084] In step S202, the benchmark clustering radius is a value range calculated by the TD&R model. This is also to reserve an adjustment margin for the benchmark feature vector. It refers to the maximum difference between different vectors when performing deep feature clustering. Here, its value range is ensured to be a defined interval or a specific value, and the radius size is affected by the selected parameters. If the benchmark feature vector is not unique, multiple benchmark feature vectors are required. Specifically, it is suitable for all situations where the coordination between images and information is relatively high, that is, the image cannot be quickly identified to its precise category and needs to be used in combination with other conditions.

[0085] Step S203: Compare the first deviation with the reference cluster radius. If the first deviation is greater than or equal to the reference cluster radius, classify the corresponding depth feature vector into other categories. Then, match the depth feature vector to the other categories based on the Euclidean distance between the first deviation and the reference feature vector in the other categories.

[0086] If the first deviation is smaller than the reference cluster radius, it indicates that the corresponding depth feature vector and the reference feature vector belong to the same category;

[0087] In step S203, the Euclidean distance is used to measure the offset between the depth feature vector and the baseline feature vector to extract whether there is a difference. Correspondingly, the large model tends to be combined with the depth feature vector and then adjusted. Here, the deviation is the key parameter for measuring the accuracy of feature vector classification. The classification effect of the baseline feature vector can be adjusted by evaluating the size of the deviation.

[0088] Example 3

[0089] like Figure 3 As shown, in this embodiment, the steps of performing multiple rounds of fine-tuning on the large model using high-dimensional semantics as input include:

[0090] Step S301: constructing a training set and a validation set based on each high-dimensional semantics;

[0091] In step S301, a training set is constructed that includes a portion of corresponding high-dimensional semantics. This is the same as the execution method of the traditional training set. Cross-validation is performed, specifically including: deep feature classification, baseline feature offset and judgment classification and other operations. The verification set includes input image information, which can be one or more images, reflecting the large model's understanding of image information during fine-tuning. This execution method is commonly used in the image modeling process, that is, when the sample size is insufficient, the current output content will be used as subsequent verification data, so that the execution feedback for better multi-experience can be improved.

[0092] Step S302: Obtain a benchmark number of proofreading rounds, input the training set within the benchmark number of proofreading rounds into the large model, then input the validation set into the large model, verify the accuracy of the large model output result, and mark it as the first proofreading parameter;

[0093] In step S302, the training set and validation set within the benchmark proofreading rounds are input into the large model one by one (multiple times) to improve the large model's ability to understand image information and correspondingly reduce the probability of errors in its application process. The proofreading parameters marked here will be used to participate in the corresponding multiple cycles to reduce the erroneous results in the validation set, thereby assisting it in performing further adjustments to the large model, such as the number of cycles and the use of erroneous data as a training set again.

[0094] Step S303: Compare the first proofreading parameter with a preset proofreading threshold. If the first proofreading parameter is greater than the proofreading threshold within the benchmark proofreading round number, mark the current large model as a model to be proofread.

[0095] In step S303, the threshold is determined according to the calculation ability of the large model, which can be understood here as the maximum loss that the large model can accept in a short period of time. When it exceeds this value, it is necessary to perform a round of parameter adjustment large cycle and coordination operations between feature vectors. Specifically, it can be a comprehensive execution status display, such as: user feedback, product failure rate, etc. The parameters here will select their selectable range according to the situation of the large model. At the same time, for the large model that has reached the critical value and implemented corresponding fine-tuning, the corresponding end time will be recorded, and the high-frequency target model will be directly generated for the small sample that has reached the critical value.

[0096] Example 4

[0097] like Figure 4 As shown, in this embodiment, when the first calibration parameter is greater than the calibration threshold within the benchmark calibration round number, after the step of marking the current large model as the model to be calibrated, the following steps are further included:

[0098] Step S401: Obtain models to be proofread whose first proofreading parameter is greater than or equal to the proofreading threshold within a benchmark proofreading round number, and sort the models to be proofread from largest to smallest according to the magnitude of the first proofreading parameter;

[0099] Step S402: Mark the model to be proofread with the highest position as the model to be output, and then mark the difference between the first proofreading parameter of the model to be output and the proofreading threshold within the benchmark proofreading rounds as the parameter to be evaluated;

[0100] Step S403: Obtain a model reference deviation interval and compare the upper limit of the model reference deviation interval with the parameter to be evaluated. If the parameter to be evaluated is within the model reference deviation interval, mark the corresponding large model as an output model after fine-tuning is completed.

[0101] If the parameter to be evaluated is greater than the upper limit of the model reference deviation interval, the corresponding large model will continue to be calibrated;

[0102] In step S403, fine-tuning and calibration operations are performed to determine the ranking of the model to be calibrated. First, a reference deviation interval is determined, which contains a set of values related to the parameter to be evaluated. The position of the parameter to be evaluated is determined and its deviation from the upper limit of the model reference deviation interval is calculated.

[0103] In steps S401 to S403, the first proofreading parameter is compared with the preset proofreading threshold value through the model to be proofreaded with the highest position, and is ranked at the top among the models to be proofreaded. We determine the output of the proofreading model by marking the model that is higher than or equal to the preset proofreading threshold value as the model to be output.

[0104] Example 5

[0105] like Figure 5 As shown, in this embodiment, the step of performing the proofreading operation after the proofreading is completed includes:

[0106] Step S501: obtaining a first calibration parameter that is higher than the corresponding parameter of the model to be output during the fine-tuning process, and marking it as a parameter to be calibrated;

[0107] Step S502: Obtain the validation set wear corresponding to the parameter to be verified, and mark the difference between the parameter to be verified and the corresponding parameter of the model to be output as the amount to be verified, and then divide the amount to be verified into multiple intervals to be tested according to the validation set wear;

[0108] In steps S501 to S502, the wear degree is determined, and the position of the quantity to be proofread in the wear degree to be tested interval of the validation set is judged based on the difference between the parameter to be proofread and the corresponding parameter of the model to be output, combined with the determination of the wear degree of the validation set associated with the parameter to be proofread.

[0109] Step S503: Compare the parameters corresponding to the model to be output with the interval to be checked with the highest order. If the interval to be checked is lower than the parameters corresponding to the model to be output, compare the parameters to be checked with the interval to be checked with the next highest order.

[0110] If the quantity to be calibrated is within the test interval of the corresponding parameter of the model to be output, then the corresponding verification set is input one by one into the large model corresponding to the quantity to be calibrated for verification;

[0111] If the performance parameters of the large model corresponding to the interval to be tested with the second highest order are lower than those of the large models corresponding to other intervals to be tested during the verification and proofreading process, the verification will be stopped.

[0112] In step S503, through the proofreading operation, the parameter to be proofread is compared with the highest and second highest ranked segments to be inspected, and then the parameter to be proofread is determined. The main basis is to evaluate whether the amount to be proofread is in the second highest position in the verification set. In order to save unnecessary resources, we stop performing verification and comparison between it and the highest proofreading parameter.

[0113] like Figure 6 As shown, the specific steps are to determine the extraction method of each reference feature based on the confidence score, including:

[0114] Step S601: Obtain confidence parameter values of each reference feature, and determine feature codes corresponding to the reference features based on each confidence parameter, and divide the feature codes into first-order codes and second-order codes;

[0115] In step S601, after the confidence score and feature parameters are preliminarily determined, products with the same style but different colors may appear during feature extraction. A vector can be assigned to represent the corresponding product. For example, during the production process, a corresponding feature code is assigned to this situation. This code includes the confidence parameters and extraction parameters, making it easier to retrieve and use it later. The vector label is not unique, which is the basis for subsequent comparison and is not restricted by product type. This can be done by adjusting the corresponding vector in the subsequent update process.

[0116] Step S602: Execute a first comparison mode, wherein the first comparison mode includes:

[0117] Obtaining the first-order codes of all feature codes, and comparing the feature code of the output result of the previous image information to be verified of the product to be extracted with the feature code of the output result of the current image information to be verified;

[0118] If the feature codes before and after extraction are consistent, the output second-order codes are consistent, that is, the output result of the previous image information to be verified of the extracted product is consistent with the output result of the current image information to be verified;

[0119] If the feature codes before and after extraction are inconsistent, the output second-order codes are inconsistent, that is, the output result of the previous image information to be verified of the extracted product is consistent with the output result of the current image information to be verified;

[0120] Step S603: Based on the output result of the second-order coding, execute the corresponding feature extraction method, wherein when the second-order coding is consistent, the output feature is verified according to the product level of the product to be extracted. When the second-order coding is inconsistent, the feature with a relatively small confidence parameter is extracted and output as a non-critical extracted feature, and the other extracted feature is output as a critical extracted feature.

[0121] In steps S602 to S603, when determining the comparison of characteristic parameters, this part of the comparison results is determined as a second-order code, retaining the commodity features that are more urgent for specific commodities for output, and the rest can be output as optional display content, mainly based on some bulk commodities (commodities subject to container volume restrictions) to define their key parameters.

[0122] By using standard data of key extracted features and then accurately comparing the extracted product features with the key features, various parameters can be determined to evaluate the tuning direction of the target large model.

[0123] The target model supports extracting key features and non-key features from the products to be extracted. The extracted products are products whose necessary key features are not supported by the target model.

[0124] Key extracted features include product name, product model, manufacturer, etc., and non-key extracted features include product packaging, promotional information, etc.

[0125] In this embodiment, through the characteristics of the large model, the existing product information is used to construct a training set and a verification set to achieve in-depth understanding and semantic extraction of the product information. During feature proofreading, the feature vector is divided into key vectors and non-key vectors, and the key and non-key parts of the product features are analyzed in detail. Weights are assigned according to their importance to further optimize the extraction results. By constructing a sliding window, the changing trend of the feature vector during the proofreading process is dynamically captured, thereby constructing an efficient small sample set to help screen out the optimal feature extraction model. When extracting new product features subsequently, multi-step extraction or single-step extraction can be selected based on the differences in the features, thereby improving the accuracy and efficiency of feature extraction. In practical applications, product features that meet the requirements can be extracted more quickly and accurately.

[0126] Example 6

[0127] This embodiment provides a system for automatically extracting product features based on large-scale semantic understanding. This system, built on the Python programming framework and a deep learning platform (such as TensorFlow or PyTorch), combines multimodal semantic modeling with feature confidence analysis mechanisms to automatically extract key features from product images and text. Specifically, it includes the following modules:

[0128] Product Information Collection Module: This module accesses the product management database API to batch capture product images and corresponding text descriptions (including product name, brand description, packaging instructions, etc.). The system supports image formats such as JPEG and PNG, and standardizes text information, encoding it into a UTF-8 text stream.

[0129] Image Semantic Encoding and Clustering Module: This module feeds collected product images into a large model based on the CLIP (Contrastive Language–Image Pre-training) architecture. After extracting the visual depth features of each image, the system initially groups the products by category. It then performs vector offset processing on the depth feature vectors within each group and calculates the Euclidean distance between the image vectors. Cluster radii are constructed based on the category center vectors. If the deviation distance of an image vector is less than the cluster radius, it is included in the current cluster. Otherwise, it is reclassified by comparing it with other cluster centers, ultimately outputting a high-dimensional semantic label for each product image.

[0130] Model Fine-tuning and Proofreading Optimization Module: This module receives the high-dimensional semantics of the clustered images and, combined with the annotated features of each product category, constructs training and validation sets for fine-tuning the parameters of the initial large model. After each round of fine-tuning, the validation set results are compared with the true labels, the accuracy is calculated, and the initial proofreading parameters are generated and compared with the preset proofreading threshold. After multiple rounds of proofreading, if the model output accuracy meets the proofreading criteria, it is marked as the target product feature extraction model.

[0131] Product Semantic Mapping and Feature Matching Module: This module inputs the target product image to be extracted into a calibrated target model and outputs its corresponding high-dimensional semantic label. The system then compares this label with existing semantic clustering information to match the historical product semantic set most relevant to the target product and extracts the annotated graphic and text features as reference features for the product.

[0132] Feature Confidence Analysis and Verification Module: This module evaluates all reference features item by item and obtains their confidence parameter values. The system maps the confidence values into encoding formats, divided into first-order and second-order encodings, and verifies the encoding consistency between the previous and next images. If the second-order encodings of the corresponding features in the two images are consistent, the feature is marked as stable and its output as a key feature is determined based on the product grade. If they are inconsistent, the confidence parameter values are compared, and the feature with the higher confidence value is selected as the key feature, while the other feature is marked as a non-key feature.

[0133] Output result module: The system outputs the key features and non-key features finally extracted to the product management system in a JSON structure. The key feature fields include at least the product name, product model, and manufacturer. Non-key features include packaging style, promotional statements, etc.

[0134] The above embodiments are only preferred embodiments of the present invention and are not limitations on the technical solutions of the present invention. All equivalent transformations and improvements made herein based on the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. The method for automatically extracting product features based on semantic understanding of a large model is characterized by: include: S1. Product information collection: obtaining product information of existing products; wherein the product information includes image information and text book information; S2. Image Semantic Encoding Clustering: Input image information into a pre-trained large model to obtain deep features of the image information. These deep features are then clustered to obtain high-dimensional semantics that match each existing product. This includes: Classify the deep features according to product information, and then perform offset operations on the deep feature vectors in the same category to obtain several baseline feature vectors; Obtain a benchmark cluster radius, and using the benchmark eigenvector as a benchmark, calculate the Euclidean distance between each depth eigenvector and the benchmark eigenvector, and mark it as the first deviation; Compare the first deviation with the benchmark cluster radius. If the first deviation is greater than or equal to the benchmark cluster radius, classify the corresponding depth feature vector into other categories. Then, match it to other categories based on the Euclidean distance between the first deviation and the benchmark feature vector in other categories. If the first deviation is smaller than the reference cluster radius, it indicates that the corresponding depth feature vector and the reference feature vector belong to the same category; S3. Model fine-tuning, proofreading, and optimization: Using high-dimensional semantics as input, the large model undergoes multiple rounds of fine-tuning. Proofreading is performed after each round of fine-tuning to obtain the target model for extracting product features. S4. Target product semantic mapping feature matching: Input the image information of the product to be extracted into the target model to obtain the high-dimensional semantics of the product to be extracted. Then input the high-dimensional semantics into the target model to match the product features corresponding to the high-dimensional semantics and mark them as reference features of the product to be extracted. S5. Feature Confidence Analysis, Extraction, and Verification: Aggregate the reference features into a dataset to be verified, then input the text information in the dataset into the target model one by one to obtain the confidence score of each reference feature. Based on the confidence score, determine the extraction method for each reference feature, and then perform verification extraction to obtain the key features corresponding to each reference feature of the extracted product; including: Obtain confidence parameter values for each reference feature, and determine feature codes corresponding to the reference features based on each confidence parameter, and divide the feature codes into first-order codes and second-order codes; Execute a first comparison mode, wherein the first comparison mode includes: Obtaining the first-order codes of all feature codes, and comparing the feature code of the output result of the previous image information to be verified of the product to be extracted with the feature code of the output result of the current image information to be verified; If the feature codes before and after extraction are consistent, the output second-order codes are consistent, that is, the output result of the previous image information to be verified of the extracted product is consistent with the output result of the current image information to be verified; If the feature codes before and after extraction are inconsistent, the output second-order codes are inconsistent, that is, the output result of the previous image information to be verified of the extracted product is consistent with the output result of the current image information to be verified; According to the output result of the second-order coding, the corresponding feature extraction method is executed. When the second-order coding is consistent, the output feature is verified according to the product level of the product to be extracted. When the second-order coding is inconsistent, the feature with a relatively small confidence parameter is extracted and output as a non-critical extracted feature, and the other extracted feature is output as a critical extracted feature.

2. The method for automatically extracting product features based on large model semantic understanding according to claim 1 is characterized in that: In step S3, the steps of performing multiple rounds of fine-tuning on the large model using high-dimensional semantics as input include: Construct training sets and validation sets based on each high-dimensional semantics; Obtain the number of benchmark calibration rounds, input the training set within the benchmark calibration rounds into the large model, then input the validation set into the large model to verify the accuracy of the large model output results, and mark them as the first calibration parameters; The first calibration parameter is compared with a preset calibration threshold, and if the first calibration parameter is greater than the calibration threshold within a reference calibration round number, the current large model is marked as a model to be calibrated.

3. The method for automatically extracting product features based on large model semantic understanding according to claim 2 is characterized in that: After the step of marking the current large model as a model to be proofread if the first proofreading parameter is greater than the proofreading threshold within the benchmark proofreading rounds, the method further includes: Obtain the models to be proofread whose first proofreading parameter is greater than or equal to the proofreading threshold within the benchmark proofreading round number, and sort the models to be proofread from largest to smallest according to the size of the first proofreading parameter; The highest-ranked model to be proofread is marked as the model to be output, and the difference between the first proofreading parameter lower than the model to be output and the proofreading threshold within the benchmark proofreading rounds is marked as the parameter to be evaluated; Obtain the model reference deviation interval and compare the upper limit of the model reference deviation interval with the parameter to be evaluated. If the parameter to be evaluated is within the model reference deviation interval, mark the corresponding large model as an output model after fine-tuning is completed. If the parameter to be evaluated is greater than the upper limit of the model reference deviation interval, the calibration operation will continue to be performed on the corresponding large model.

4. The method for automatically extracting product features based on large model semantic understanding according to claim 3 is characterized in that: If the second-order encodings of multiple features are consistent, they are sorted based on the output results of the confidence parameters of each feature extraction, and the feature corresponding to the lowest confidence parameter is marked as the key extraction feature. In the case of the same ranking, the corresponding key extraction feature is output according to the product level of the product to be extracted.

5. The method for automatically extracting product features based on large model semantic understanding according to claim 1 is characterized in that: The key extraction features include at least product name, product model, and manufacturer, and the non-key extraction features include at least product packaging and promotional information.

6. A system for automatically extracting product features based on large-scale semantic understanding, for implementing the method for automatically extracting product features based on large-scale semantic understanding according to any one of claims 1 to 5, characterized in that: include: A product information collection module is used to obtain product information of existing products, wherein the product information includes picture information and text book information; The image semantic encoding and clustering module is used to input the image information into the pre-trained large model, extract the corresponding deep features, and perform clustering processing on the deep features to obtain high-dimensional semantics that match each existing product; A model fine-tuning and proofreading optimization module, configured to use the high-dimensional semantics as input, perform multiple rounds of fine-tuning on the large model, and perform proofreading after each round of fine-tuning, thereby obtaining a target model for extracting product features; The product semantic mapping and feature matching module is used to input the image information of the product to be extracted into the target model to obtain the corresponding high-dimensional semantics, and match the product features corresponding to the high-dimensional semantics through the target model and mark them as reference features; A feature confidence analysis and verification module is used to aggregate the reference features into a dataset to be verified, input the text book information in the dataset into the target model one by one, calculate the confidence score of each reference feature; determine the feature extraction method based on the confidence score, and complete the final verification extraction output of the key features; Among them, the key features include at least the product name, product model, and manufacturer, and the non-key features include at least the product packaging and promotional information.

7. The automatic product feature extraction system based on large model semantic understanding according to claim 6 is characterized in that: The image semantic coding and clustering module specifically includes: The feature classification unit is used to preliminarily classify the extracted deep features according to product information, and perform vector offset processing on each deep feature vector in each category to generate several baseline feature vectors; A cluster determination unit is configured to obtain a reference cluster radius and, with the reference feature vector as the center, calculate the Euclidean distance between other depth feature vectors and the reference feature vector to obtain a first deviation; A feature classification unit is used to compare the first deviation with the benchmark clustering radius. If the first deviation is greater than or equal to the clustering radius, the corresponding deep feature vector is classified into other categories and matched to the category according to the Euclidean distance; if the first deviation is less than the clustering radius, it is classified into the current category to complete the high-dimensional semantic clustering processing.

8. A terminal for automatically extracting product features based on large-scale model semantic understanding, characterized in that: include: at least one processor; and a memory communicatively connected to the processor; Wherein, executable instructions are stored in the memory, and when the executable instructions are executed by the processor, the processor is enabled to execute the method for automatic extraction of product features based on large model semantic understanding as described in any one of claims 1 to 5.

Images