Attribute value adjusting method and device, computer equipment and storage medium

Abstract

The invention relates to an attribute value adjusting method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring a key value pair set of a to-be-processed attribute name; the key value pair set comprises key value pairs corresponding to the to-be-processed attribute name under a plurality of target entities; the key value pair comprises the to-be-processed attribute name and the attribute value; inputting the plurality of attribute values into a preset large language model to obtain an attribute value merging result of the plurality of attribute values; the attribute value merging result comprises a merged attribute value corresponding to each attribute value; and adjusting the attribute value of each target entity based on the attribute value merging result. By adopting the method, the product data analysis accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method, apparatus, computer device, and storage medium for adjusting attribute values. Background Art

[0002] With the development of the Internet, the variety of products on e-commerce platforms has become increasingly rich. In order to support product diversity, e-commerce platforms usually allow product providers to fill in customized attribute information when requiring them to fill in product attribute information. However, customized attribute information is prone to personalized descriptions and typos, resulting in similar products being divided into different attributes, which seriously affects the accuracy of subsequent product data analysis.

[0003] In traditional technology, product classification is achieved by comparing the product's attribute name with a pre-set attribute comparison table. However, this method relies on the accuracy of the attribute comparison table itself. As the number of product types increases, the accuracy and timeliness of the attribute comparison table itself cannot be guaranteed, and the accuracy of product data analysis is also low.

[0004] This shows that traditional technologies still have the problem of low accuracy in product data analysis. Summary of the Invention

[0005] Based on this, it is necessary to provide a method, device, computer equipment and storage medium for adjusting attribute values ​​that can improve the accuracy of product data analysis in response to the above technical problems.

[0006] In a first aspect, the present application provides a method for adjusting an attribute value, the method comprising:

[0007] Obtain a key-value pair set of attribute names to be processed; the key-value pair set includes key-value pairs corresponding to the attribute names to be processed under multiple target entities; the key-value pairs include the attribute names to be processed and attribute values;

[0008] Inputting the plurality of attribute values ​​into a preset large language model to obtain an attribute value merging result of the plurality of attribute values; the attribute value merging result includes a merged attribute value corresponding to each of the attribute values;

[0009] Based on the attribute value merging result, the attribute value of each target entity is adjusted.

[0010] In one embodiment, obtaining a set of key-value pairs to be processed includes:

[0011] Obtaining unstructured text of a plurality of target entities;

[0012] Inputting the unstructured text of each target entity into a pre-trained large language model to obtain an attribute key-value pair extraction result corresponding to each target entity;

[0013] Generate a key-value pair set of attribute names to be processed; the key-value pair set includes attribute key-value pair extraction results corresponding to the attribute names to be processed in multiple target entities.

[0014] In one embodiment, the unstructured text includes a description text, and inputting the unstructured text of each target entity into a pre-trained large language model to obtain an attribute key-value pair extraction result corresponding to each target entity includes:

[0015] Extracting key-value pairs from attribute information of multiple sample entities according to a preset attribute name set to obtain a first key-value pair set;

[0016] Obtain a second key-value pair set determined by voting of the plurality of sample entities based on the plurality of large language models;

[0017] Determining sample key-value pair training data based on the first key-value pair set and the second key-value pair set;

[0018] Based on the sample key-value pair training data, supervised fine-tuning is performed on the preset large language model to obtain the pre-trained large language model.

[0019] In one embodiment, obtaining a second key-value pair set determined by voting of the plurality of sample entities based on the plurality of large language models comprises:

[0020] Inputting the description text of the sample entity into a plurality of large language models respectively to obtain a key-value pair extraction result corresponding to each of the large language models; the key-value pair extraction result includes at least one second key-value pair;

[0021] Counting common attribute values ​​in the plurality of key-value pair extraction results;

[0022] Based on the second key-value pairs corresponding to the common attribute values, a second key-value pair set of the sample entities is determined.

[0023] In one embodiment, the preset attribute name set includes multiple preset attribute names and an attribute type corresponding to each of the preset attribute names; and the first key-value pair set obtained by extracting key-value pairs from attribute information of multiple sample entities according to the preset attribute name set includes:

[0024] The preset attribute name set and the attribute information of the multiple sample entities are input into the preset large language model to obtain the first key-value pair set; the first key-value pair set includes multiple first key-value pairs of each sample entity, and the attribute type corresponding to each first key-value pair.

[0025] In one embodiment, when the attribute type corresponding to the attribute name to be processed is a numeric type, and the multiple attribute values ​​include at least two value units, before inputting the multiple attribute values ​​into the preset large language model to obtain the attribute value merging result of the multiple attribute values, the step further includes:

[0026] For each pair of the value units of the attribute name to be processed, calling multiple large language models to respectively generate forward conversion coefficients and reverse conversion coefficients for each pair of the value units;

[0027] determining an inference confidence of each of the large language models according to a product result of the forward conversion coefficient and the reverse conversion coefficient of each pair of the value units;

[0028] Based on the inference confidence, a large language model that meets a preset confidence condition is selected, and unit mapping processing is performed on the plurality of attribute values.

[0029] In one embodiment, inputting the plurality of attribute values ​​into a preset large language model to obtain an attribute value merging result of the plurality of attribute values ​​includes:

[0030] According to a preset sampling rule, a plurality of attribute values ​​are sampled multiple times to obtain a plurality of random attribute value sets;

[0031] Inputting the plurality of random attribute value sets into the preset large language model respectively to obtain a sub-attribute value merging result corresponding to each of the random attribute value sets;

[0032] Frequency analysis and result merging are performed on the merging results of the plurality of sub-attribute values ​​to determine an attribute value merging result of the plurality of attribute values.

[0033] In a second aspect, the present application provides an attribute value processing device, the attribute value processing device comprising:

[0034] A key-value pair acquisition module is used to acquire a key-value pair set of attribute names to be processed; the key-value pair set includes key-value pairs corresponding to the attribute names to be processed under multiple target entities; the key-value pairs include the attribute names to be processed and attribute values;

[0035] an attribute value merging module, configured to input the plurality of attribute values ​​into a preset large language model to obtain an attribute value merging result of the plurality of attribute values; the attribute value merging result including a merged attribute value corresponding to each attribute value;

[0036] An attribute value processing module is used to adjust the attribute value of each target entity based on the attribute value merging result.

[0037] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the above-described method when executing the computer program.

[0038] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which implements the method described above when the computer program is executed by a processor.

[0039] The above-mentioned attribute value adjustment method, device, computer equipment and storage medium obtain a key-value pair set of attribute names to be processed; the key-value pair set includes key-value pairs corresponding to the attribute names to be processed under multiple target entities; the key-value pairs include the attribute names to be processed and attribute values; the multiple attribute values ​​are input into a preset large language model to obtain attribute value merging results of the multiple attribute values; the attribute value merging results include merged attribute values ​​corresponding to each attribute value; based on the attribute value merging results, the attribute value of each target entity is adjusted, so that the large language model can provide a dynamically adaptable attribute value merging result according to the overall situation of the key-value pair set, avoiding the errors that may be caused by the traditional reliance on attribute comparison tables, and achieving the technical effect of reducing manual intervention, improving the accuracy and efficiency of attribute value classification, and improving the quality of product data analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A diagram illustrating an application environment of a method for adjusting attribute values ​​in one embodiment;

[0041] Figure 2 Schematic diagram of a flow chart of a method for adjusting attribute values ​​in one embodiment;

[0042] Figure 3 is a structural block diagram of an attribute value adjustment device in one embodiment;

[0043] Figure 4 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0045] The attribute value adjustment method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. The terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. The terminal 102 obtains a key-value pair set of attribute names to be processed by communicating with the server 104; the key-value pair set includes key-value pairs corresponding to the attribute names to be processed under multiple target entities; the key-value pairs include the attribute names to be processed and attribute values; the multiple attribute values ​​are input into the preset large language model to obtain the attribute value merging result of the multiple attribute values; the attribute value merging result includes the merged attribute value corresponding to each attribute value; based on the attribute value merging result, the attribute value of each target entity is adjusted. The terminal 102 can be, but is not limited to, various personal computers, laptops, smart phones and tablet computers. The server 104 can be implemented as an independent server or a server cluster consisting of multiple servers.

[0046] In one embodiment, Figure 2 As shown, a property value adjustment method is provided, which is applied to Figure 1 Taking the terminal 102 in FIG. 1 as an example, the method includes the following steps:

[0047] Step S100: Obtain a key-value pair set of attribute names to be processed.

[0048] The key-value pair set includes key-value pairs corresponding to the attribute name to be processed under multiple target entities, and the key-value pairs include the attribute name to be processed and the attribute value.

[0049] A key-value pair set can be a collection of associated data used to describe the characteristics of a target entity, where each key-value pair includes a correspondence between an attribute name and an attribute value. This set can be obtained by extracting relevant information from multiple target entities. For example, a key-value pair set can include, but is not limited to, product attributes such as color and size and their corresponding attribute values. For example, "color: red" is a form of a key-value pair.

[0050] The target entity may be an object that needs to be classified by attributes and can be distinguished by a unique identifier. The target entity may include but is not limited to one or more of commodities, virtual items or physical objects in the e-commerce platform.

[0051] The attribute name to be processed can be a specific attribute name that needs to be classified or adjusted, such as "color", "size", etc. The attribute value can be a specific description of the target entity under the specific attribute name, which can be expressed in different units, such as "180cm", "pure cotton", etc.

[0052] To obtain a set of key-value pairs for the attribute name to be processed, data can be collected from a database or e-commerce platform to obtain the set of attribute values ​​corresponding to the same attribute name under all target entities. Furthermore, the key-value pair set can be cleansed (for example, removing duplicate values) and filtered (for example, excluding invalid data) to ensure that the data source is complete and covers all possible attribute value variations.

[0053] In step S200 , the plurality of attribute values ​​are input into a preset large language model to obtain an attribute value merging result of the plurality of attribute values.

[0054] The attribute value merging result includes a merged attribute value corresponding to each attribute value.

[0055] The pre-set large language model can be a language processing model trained on a large amount of text. It can be implemented using a Transformer architecture, such as GPT, and is used to perform semantic analysis and classification tasks. These models can understand natural language and generate semantically relevant output. The attribute value merging result can refer to the merged result obtained by the model based on the overall judgment of the input attribute values. The merged attribute value can be a standard value determined by the large language model to unify similar attribute values. It has a certain degree of semantic consistency, thereby automatically identifying similar or equivalent attribute values ​​and reducing the need for human intervention.

[0056] For example, the pre-set large language model can utilize its semantic analysis capabilities (e.g., identifying synonyms, correcting spelling errors), contextual understanding, and other capabilities to generate a merged attribute value corresponding to each attribute value. For example, the model may merge "pure cotton" and "100% cotton" into "pure cotton," thereby reducing reliance on static comparison tables and improving classification flexibility and accuracy.

[0057] Step S300: adjusting the attribute value of each target entity based on the attribute value merging result.

[0058] Adjusting the attribute value can be replacing the original attribute value of the target entity with the merged and standardized attribute value. For example, if the "size" attribute value of a product is "L" and the merged attribute value determined by the model is "Large", it will be replaced with "Large".

[0059] This embodiment provides a method for adjusting attribute values, which obtains a key-value pair set of attribute names to be processed; the key-value pair set includes key-value pairs corresponding to the attribute names to be processed under multiple target entities; the key-value pairs include the attribute names to be processed and attribute values; the multiple attribute values ​​are input into a preset large language model to obtain an attribute value merging result of the multiple attribute values; the attribute value merging result includes a merged attribute value corresponding to each attribute value; based on the attribute value merging result, the attribute value of each target entity is adjusted, so that the large language model can provide a dynamically adaptable attribute value merging result according to the overall situation of the key-value pair set, avoiding the errors that may be caused by the traditional method of relying on attribute comparison tables, thereby achieving the technical effect of reducing manual intervention, improving the accuracy and efficiency of attribute value classification, and improving the quality of product data analysis.

[0060] In one embodiment, obtaining a set of key-value pairs to be processed includes:

[0061] Obtaining unstructured text of a plurality of target entities;

[0062] Inputting the unstructured text of each target entity into a pre-trained large language model to obtain an attribute key-value pair extraction result corresponding to each target entity;

[0063] Generate a set of key-value pairs of attribute names to be processed;

[0064] The key-value pair set includes attribute key-value pair extraction results corresponding to the attribute names to be processed in multiple target entities.

[0065] The attribute key-value pairs of the target entity can be obtained by parsing the unstructured text associated with it to extract structured information. Unstructured text can be free text data that has not been formatted. For example, unstructured text can include but is not limited to entity descriptions, free text fields that can be filled in by oneself, or natural language descriptions of product details pages. It can also include evaluation information of the target entity. Unstructured text information can be obtained through e-commerce platform interfaces, database queries, or external data sources. For example, descriptions such as "the material is pure cotton, the color is red, and the sizes are S / M / L" can be extracted from the product details page. By obtaining unstructured text from multiple sources, the attribute expressions contained in the data source can be made more comprehensive and rich, thereby improving the accuracy and comprehensiveness of the key-value pair set.

[0066] The pre-trained large language model can be a natural language processing model trained based on a large-scale corpus. For example, it can be a trained model such as GPT. The trained large language model can have the ability to identify implicit attribute associations in the text.

[0067] The attribute key-value pair extraction result can be the structured data output by the model. The correspondence between the attribute name and the attribute value can be located through the text understanding ability of the model. For example, the key-value pair of "color: red" can be extracted from the string "color is red". This embodiment can reduce the dependence on manual labeling or rule definition in an automated way. At the same time, it can also reduce the risk of attribute omission due to differences in text formats.

[0068] The key-value pair set can be a structured data set for a specific attribute name, and can be formed by filtering the attribute key-value pair extraction results of all target entities. For example, when the target attribute name is "color", all key-value pairs containing this attribute name can be extracted from the attribute key-value pair extraction results of each target entity and aggregated into a unified set.

[0069] Furthermore, for different attribute names to be processed, multiple key-value pair sets corresponding to different attribute names can be generated, and the attribute value information of the corresponding attribute name of each target entity can be adjusted in turn using the attribute value adjustment method of this embodiment.

[0070] This embodiment provides a method for adjusting attribute values, which expands the data source by acquiring unstructured text of multiple target entities, automatically parses the attribute association relationships in the text using a pre-trained model to form a structured result, and then filters and aggregates the key-value pair set based on the target attribute name. Through comprehensive parsing of the unstructured text, the coverage of the data source can be improved, the need for manual intervention can be reduced, and the risk of missing attribute values ​​due to format differences or redundant information can be reduced, thereby achieving the effect of improving the accuracy of data analysis.

[0071] In one embodiment, the unstructured text includes a description text, and inputting the unstructured text of each target entity into a pre-trained large language model to obtain an attribute key-value pair extraction result corresponding to each target entity includes:

[0072] Extracting key-value pairs from attribute information of multiple sample entities according to a preset attribute name set to obtain a first key-value pair set;

[0073] Obtain a second key-value pair set determined by voting of the plurality of sample entities based on the plurality of large language models;

[0074] Determining sample key-value pair training data based on the first key-value pair set and the second key-value pair set;

[0075] Based on the sample key-value pair training data, supervised fine-tuning is performed on the preset large language model to obtain the pre-trained large language model.

[0076] The preset attribute name set may be a predefined set containing standard attribute names, used to guide the attribute extraction process for sample entities. For example, the preset attribute name set may be obtained through manual annotation, domain knowledge, or prior knowledge. For example, the preset attribute name set may include attribute names such as "color," "size," and "material."

[0077] Sample entities can be example target entities used to train or verify models. Their attribute information can be manually constructed or obtained from existing e-commerce platforms and data sets. For example, it can be product information containing structured fields.

[0078] Key-value pair extraction can be performed by extracting the association between attribute names and attribute values ​​from the structured or semi-structured data of the sample entity. In this embodiment, the first key-value pair set can be obtained by extracting the attribute information. Furthermore, the first key-value pair set can be obtained by matching the attribute information of the sample entity item by item using a preset attribute name set.

[0079] In this embodiment, multiple large language models may be used to respectively identify key-value pairs for unstructured texts of the same sample entity, thereby obtaining a second key-value pair set corresponding to each unstructured text.

[0080] Voting by multiple large language models can be a mechanism for determining the final result by performing a majority vote or other aggregation strategy on the results generated by multiple large language models. In this embodiment, the unstructured text of the sample entity can be input into multiple large language models, the output results of each model can be collected, and the key-value pairs with statistical consistency among the majority of models can be screened, thereby reducing the prediction bias of a single model and improving the robustness of key-value pair extraction.

[0081] The sample key-value pair training data can be a training data set formed by fusing the first key-value pair data and the second key-value pair data, and can be achieved by merging overlapping or consistent items in the first key-value pair set and the second key-value pair set to form a more comprehensive and reliable training data set.

[0082] The preset large language model may be a general language model that has not been initially trained for a specific task. For example, an open source large language model may be used for further supervised fine-tuning.

[0083] Supervised fine-tuning involves adjusting model parameters using training data to optimize the model's performance for a specific task. For example, the model is fed with input text and corresponding key-value pairs, and the parameters are updated by calculating the difference between the predicted result and the label and backpropagating the result. For example, if the model predicts "Material: Cotton" and the actual attribute value is "Pure Cotton," the loss function drives the adjustment of model parameters to increase the output probability of "Pure Cotton," thereby reducing reliance on general corpus and improving the accuracy of attribute extraction.

[0084] This embodiment provides an attribute value adjustment method, which can significantly improve the quality of training data and reduce annotation deviation by combining a preset attribute name set with rule extraction of sample entities to generate benchmark data, using multi-model voting to enhance the reliability of key-value pair results, fusing two first key-value pair sets and a second key-value pair set to construct training data, and optimizing model parameters based on supervised fine-tuning. Through supervised fine-tuning, the model can be made more adaptable to the unstructured text features of specific business scenarios, improve the accuracy of attribute extraction, and reduce the reliance on manual repeated annotation through automated voting and iterative optimization processes, thereby achieving the effect of improving the accuracy of attribute value merging.

[0085] In one embodiment, obtaining a second key-value pair set determined by voting of the plurality of sample entities based on the plurality of large language models comprises:

[0086] Inputting the description text of the sample entity into a plurality of large language models respectively to obtain a key-value pair extraction result corresponding to each of the large language models; the key-value pair extraction result includes at least one second key-value pair;

[0087] Counting common attribute values ​​in the plurality of key-value pair extraction results;

[0088] Based on the second key-value pairs corresponding to the common attribute values, a second key-value pair set of the sample entities is determined.

[0089] Among them, the multiple large language models can be general large language models based on different architectures or parameter configurations, such as GPT and llama architectures. For example, open source large language models such as deepseekv3 and qwen2.5 can be called locally or online, or closed source large language models such as chatgpt can be called online to achieve complementarity between models through parallel deployment.

[0090] The key-value pair extraction result may be a list of attribute key-value pairs output by each model for the input text, and may include, for example, structured expressions such as "material: pure cotton" and "color: red".

[0091] In this embodiment, by inputting the same descriptive text into multiple large language models in parallel, each model can be used to perform attribute recognition based on its own training data and logic, thereby capturing the potential correlation between the expressions obtained by models with different capabilities.

[0092] The common attribute value may be an attribute value that appears repeatedly in multiple key-value pair extraction results. For example, "pure cotton" is recognized by at least two models in the three models as the attribute value corresponding to the "material" attribute name. The counting of common attribute values ​​in the multiple key-value pair extraction results can be achieved, for example, by calculating the frequency of occurrence of each attribute value in different models and setting a threshold for screening. For example, attribute values ​​that appear more than a preset frequency or appear more than a preset number of times can be regarded as common attribute values. In a specific embodiment, if two of the three models output "material: pure cotton" and one outputs "material: cotton", then "pure cotton" can be taken as a common attribute value and included in the common attribute value set.

[0093] In the case where there are multiple common attribute values, there may also be multiple second key-value pairs corresponding to the common attribute values. The second key-value pair set of the sample entity is determined based on all the second key-value pairs. The key-value pairs that meet the occurrence threshold are aggregated to obtain the second key-value pair set.

[0094] This embodiment provides an attribute value adjustment method, which inputs the descriptive text of a sample entity into multiple large language models to obtain key-value pair extraction results, counts the common attribute values ​​in the output of each model, and then forms a second key-value pair set based on the consensus result. The method reduces the prediction deviation of a single model due to differences in training data or architecture through multi-model parallel processing, filters out abnormal results through screening of common attribute values, thereby reducing noise interference, thereby improving data reliability while achieving the effect of improving the accuracy of the key-value pair set while improving the efficiency and quality of automated labeling.

[0095] In one embodiment, extracting key-value pairs from attribute information of multiple sample entities according to a preset attribute name set to obtain a first key-value pair set includes:

[0096] The preset attribute name set and the attribute information of the plurality of sample entities are input into a preset large language model to obtain the first key-value pair set.

[0097] The preset attribute name set includes a plurality of preset attribute names and an attribute type corresponding to each of the preset attribute names. The attribute type may be a classification identifier used to limit the semantic scope or format requirements of the attribute name itself. Exemplarily, the attribute type may include a numerical type, which indicates that there may be multiple value units and attribute values ​​with different numerical values ​​under the attribute name. The attribute type may also include a label type, or a text type, which indicates that the attribute value under the attribute name is generally described by a relatively fixed label or text.

[0098] In the process of obtaining the first key-value pair set, the large language model uses the type constraint mechanism in the preset attribute name set to identify the corresponding attribute name while determining the attribute category of the attribute name itself, thereby effectively reducing the interference of irrelevant information and improving the accuracy of key-value pair extraction.

[0099] The first key-value pair set includes multiple first key-value pairs for each sample entity, and the attribute type corresponding to each first key-value pair. Each key-value pair in the first key-value pair set includes not only an attribute name and an attribute value, but also an attribute type tag, thereby facilitating targeted merging based on different attribute types during attribute value merging.

[0100] This embodiment provides a method for adjusting attribute values. By constructing a preset attribute name set containing attribute type constraints, the model's semantic understanding of attribute values ​​is enhanced using type information, thereby reducing the ambiguity of attributes with the same name. By refining data annotations through type labels, a more comprehensive reference can be provided for attribute value merging, thereby improving the semantic consistency and standardization of the attribute extraction and merging process summary.

[0101] It is understandable that a large language model may have hallucination characteristics and numerical insensitivity. Therefore, when comparing two numbers, for example, 9.11 and 9.8, the model may give an erroneous result that 9.11 is greater than 9.8. Therefore, in one embodiment, when the attribute type corresponding to the attribute name to be processed is a numeric type, and the multiple attribute values ​​include at least two value units, the inputting the multiple attribute values ​​into the preset large language model to obtain the attribute value merging result of the multiple attribute values ​​further includes:

[0102] For each pair of the value units of the attribute name to be processed, calling multiple large language models to respectively generate forward conversion coefficients and reverse conversion coefficients for each pair of the value units;

[0103] determining an inference confidence of each of the large language models according to a product result of the forward conversion coefficient and the reverse conversion coefficient of each pair of the value units;

[0104] Based on the inference confidence, a large language model that meets a preset confidence condition is selected, and unit mapping processing is performed on the plurality of attribute values.

[0105] Among them, when the attribute type corresponding to the attribute name to be processed is a numerical type, and multiple attribute values ​​include at least two value units, then before merging the attribute values, the value units can be unified through a large language model to avoid the impact of different value units on the merging results.

[0106] The value unit can be a measurement unit of a numerical attribute value, such as "centimeter", "inch", "milliliter", "liter", etc., which is used to quantify the physical meaning of the attribute value. For example, it can be obtained by constructing a unit recognition algorithm.

[0107] The forward conversion coefficient can be the mathematical ratio that converts a first unit to a second unit. For example, in "1 inch = 2.54 centimeters," the forward coefficient from centimeters to inches is 0.3937 (1 / 2.54), which can be obtained through model reasoning. The reverse conversion coefficient can be the mathematical ratio that converts the second unit back to the first unit. For example, the reverse coefficient from inches to centimeters is 2.54, and its calculation method is the opposite of the forward coefficient.

[0108] In one specific embodiment, when attribute values ​​include "Size: 180cm" and "Size: 70 inches," multiple large language models can be invoked to process the unit pair "cm and inches," respectively, requiring the models to output the forward and reverse conversion coefficients for each. By using model inference rather than manually defining unit conversion rules, it can adapt to conversion needs between emerging or non-standard units, avoiding the limitations of traditional manually maintained comparison tables.

[0109] Inference confidence can be used as an indicator to measure the reliability of model output results. Inference confidence can be obtained by multiplying the forward conversion coefficient and the reverse conversion coefficient. It can be understood that if the product is close to 1 (for example, between 0.99 and 1.01), it indicates that the model's derivation of the conversion relationship is self-consistent. For example, if a model outputs a forward coefficient of 0.3937 and a reverse coefficient of 2.54 for "cm to inches", the product of these coefficients is 1.0001, indicating that the model's inference results are reliable.

[0110] By calculating the inference confidence of the large language model through the forward conversion coefficient and the reverse conversion coefficient, we can verify the self-consistency of the model reasoning and screen out models with a more accurate understanding of the unit conversion logic, thereby reducing the risk of conversion bias caused by model reasoning errors.

[0111] Based on the inference confidence, a large language model that meets a preset confidence condition is selected, and unit mapping processing is performed on the plurality of attribute values.

[0112] Unit mapping can be the process of converting attribute values ​​with different units under the same attribute name into a preset unit, for example, converting all "inches" to "centimeter." Numerical conversion can be performed by selecting a large language model with a confidence level above a threshold and using its output forward conversion coefficient. The preset unit can be pre-set based on prior knowledge or determined based on the most frequently occurring unit among multiple attribute values.

[0113] This embodiment provides a method for adjusting attribute values. By calling multiple large language models to generate unit conversion coefficients, verifying the reliability of the model based on the product of forward and reverse coefficients, and selecting a high-confidence model to perform unit standardization, this method can solve the problem in traditional technologies that manually maintained comparison tables are difficult to cover non-standard units. By utilizing the self-consistency verification mechanism of the forward and reverse coefficients, the reliability of the model output can be ensured. By standardizing the values ​​after unifying the units, the accuracy of attribute value merging can be improved, thereby achieving the effect of improving data processing efficiency and comparability of results.

[0114] In some other exemplary embodiments, before inputting the plurality of attribute values ​​into a preset large language model to obtain an attribute value merging result of the plurality of attribute values, the process may further include:

[0115] For each pair of the value units of the attribute name to be processed, calling the current large language model to generate a forward conversion coefficient and a reverse conversion coefficient for each pair of the value units;

[0116] Determining the inference confidence of the current large language model according to a product result of the forward conversion coefficient and the reverse conversion coefficient of each pair of the value units;

[0117] If the inference confidence of the current large language model meets the preset confidence condition, the unit mapping processing is performed on the multiple attribute values ​​through the large language model; if not, the large language model is replaced based on the preset large language model list.

[0118] In these embodiments, the reasoning ability of the current large language model can be judged by the reasoning confidence. If the reasoning confidence does not meet the requirements of the preset confidence conditions, a large language model with better reasoning ability is selected from the preset large language model list to reprocess the steps of this embodiment.

[0119] In one embodiment, inputting the plurality of attribute values ​​into a preset large language model to obtain an attribute value merging result of the plurality of attribute values ​​includes:

[0120] According to a preset sampling rule, a plurality of attribute values ​​are sampled multiple times to obtain a plurality of random attribute value sets;

[0121] Inputting the plurality of random attribute value sets into the preset large language model respectively to obtain a sub-attribute value merging result corresponding to each of the random attribute value sets;

[0122] Frequency analysis and result merging are performed on the merging results of the plurality of sub-attribute values ​​to determine an attribute value merging result of the plurality of attribute values.

[0123] The preset sampling rule may be a random sampling rule for data adoption. For example, the preset sampling rule may set a fixed ratio or a fixed number. The random attribute value set may be a subset of the original attribute values ​​generated by the sampling rule, and may be obtained by randomly selecting different combinations from the original attribute value set.

[0124] Multiple random attribute value sets are input into the preset large language model to obtain a merged result of the sub-attribute values ​​corresponding to each random attribute value set. The large language model can be invoked to perform semantic analysis on each subset and generate a merged result. The number of merged attribute values ​​in the merged result can be dynamically determined by the model based on the conditions of all attribute values, or can be pre-set. For example, the number of merged attribute values ​​can be limited to no more than 30 in the prompt word.

[0125] For example, if a set of random attribute values ​​includes "large", "L", and "XL", the model may output three merged results of "L", "XL", and "other", or it may output two merged results of "L" and "XL", or it may output one merged result of "large"; if another subset includes "extra large" and "XXL", it may output two merged results of "XXL" and "other", or it may output a merged result of "XXL".

[0126] Frequency analysis and result merging of the merged results of the multiple sub-attribute values ​​can be performed by counting the frequency of occurrence of each merge suggestion and combining it with a threshold or weighting rule to determine the final result. Frequency analysis can be the process of counting the number of occurrences of each merged attribute value, and result merging can be the step of integrating sub-results based on frequency or other rules. Exemplary, result merging can include strategies such as selecting the highest frequency value or a weighted average. Reducing random fluctuations in model output through statistical aggregation can improve the stability and reliability of the merged results.

[0127] This embodiment provides an attribute value adjustment method that generates multiple sets of random attribute values ​​by introducing preset sampling rules. This can effectively avoid situations where the model cannot handle the context due to excessive length. A large language model is used to perform multiple rounds of inference on different subsets to reduce the impact of the sampling behavior itself on the merging results. By eliminating noise and determining the final merging result through frequency analysis and result merging, the technical effect of reducing model output deviation and improving the coverage capability of the merging results can be achieved.

[0128] In order to more clearly illustrate the technical solution of this application, this application also provides a detailed embodiment.

[0129] In one embodiment, a method for adjusting attribute values ​​is provided, taking a target entity as an e-commerce platform product as an example, including:

[0130] 1. Training Data Acquisition and Cleaning

[0131] Analyze the attributes that already exist on the Amazon page and perform data acquisition, collection, cleaning, warehousing and other processes.

[0132] Attribute name type classification: Attribute names are classified according to commonly used attribute names, mainly into numerical type and label type. Numerical type is attributes with units such as kilograms, kilos, and grams, and label type mainly includes enumeration values, such as color, material, etc. By maintaining a mapping relationship between commonly used attribute names and label types, the large language model is instructed to extract. If the real-time maintenance of the mapping relationship is low, that is, the attribute information contains attribute names that are not provided by the mapping relationship, a preset number of high-frequency values ​​under the attribute name can be extracted (to avoid interference from abnormal values), and the attribute type corresponding to the attribute name can be determined by calling the large language model. For example, commonly used numerical attribute name keywords include: range, weight, width, size, count, dimensions, number, etc.

[0133] 2. Fine-grained attribute extraction

[0134] It is understandable that in the relevant information of the product, there may be attributes that are not listed in the attribute information, and they can be extracted from the product description, five-point description, and comment text.

[0135] (1) Constructing attribute extraction training set

[0136] In this embodiment, in order to improve the information extraction capability of the large language model for unstructured text, this is achieved by performing supervised fine-tuning (SFT) on the large language model.

[0137] Construct an attribute extraction training set, which includes two aspects: recall rate evaluation and precision rate.

[0138] 1. Build an extraction test benchmark based on the relationship between structural tags and product information hits. Specifically, when a structured attribute value is included in the description information, it is considered a Y-value label. This yields a subset that verifies the extraction recall rate of the large model. This type of attribute value is a clear result, so the large language model needs to emphasize it.

[0139] 2. Use multiple large language models to parse the description and extract valuable key-value pairs. For example, when multiple models vote for a common attribute value, this attribute value becomes the true attribute.

[0140] 3. If there are still omissions in the recognition results of unstructured text, the oov data is marked or keyword granularity is achieved, and the information extraction results are supplemented through ngram (n = 1-3) + llm recognition.

[0141] (2) In the training data set, the format of each training data can be: {description:xxxx,

[0142] [{attr_name:color,attr_value:red},...]}. Here, description can be unstructured text, attr_name can be the attribute name, and attr_value can be the attribute value.

[0143] 3. General Merging Solution Based on Large Language Model

[0144] It is understandable that manually filled attribute values ​​may have data inconsistencies, so similar attributes need to be normalized and merged.

[0145] (1) LLM-based attribute merging scheme

[0146] 1. Construct a prompt word template to guide the large language model to analyze attribute names. In a specific embodiment, the requirements for the large language model's output results may include: merging attribute values ​​with similar content; normalizing numerical values ​​according to the most frequent unit, dividing the interval range according to equal frequency, and dividing the corresponding attributes into corresponding intervals; for dimensions, dividing the intervals according to the longest side; for scientific notation, restoring numerical values ​​to decimal; classifying attribute values ​​that should not appear under the attribute name, invalid attribute values, or attribute values ​​that cannot be classified as "other"; limiting the number of merged attribute values ​​to 30 or less; and returning the processing results in JSON format, including the original attribute value information and the merged attribute value information.

[0147] Exemplarily, the output format of each merged attribute value may be: {"origin_attribute":"xxx","combined_attribute":"ooo"}.

[0148] (2) Merging when there are many attribute values

[0149] When there are many units or values ​​under an attribute name, the context length of the large language model is limited. As the number of item attribute values ​​increases with the number of items, sometimes exceeding the model's context, a batch extraction approach can be used, which is then fed into the large language model in sequence. However, this approach may result in some attribute values ​​that should have been assigned to the same combined attribute value being assigned to different batches. If the batch contains very little data with the same label, the attribute value may be classified as "Other," causing errors.

[0150] To this end, this embodiment adopts a random sampling method with multiple samplings to randomly sample all attribute values ​​multiple times, with the number of samplings being N / batch times, to obtain multiple sets of random key-value pairs, and then input each random key-value pair set into the large language model in turn. This can effectively avoid the situation where normalized attributes are not divided into an integrated merge label due to window issues, thereby improving the accuracy of the merge result.

[0151] 4. Special treatment of numerical attributes

[0152] (1) Merging of single-dimensional attributes

[0153] It is understandable that if the same merging scheme is used for numerical attributes as for label attributes, the large language model may have hallucination problems when judging the size of values ​​between, for example, 9.9 and 9.18.

[0154] For example, when the first column is the original attribute value and the second column is the attribute label after normalization by the large language model, where 16ounces = 1pounds, the values ​​less than 16ounces should be in the range of 0-1pounds, but are incorrectly assigned to 0-0.1pounds.

[0155] In this embodiment, the unit unification problem is solved by converting the value unit into the value unit that appears most frequently. The conversion is achieved by constructing a mapping relationship from the original unit to the target unit.

[0156] In this embodiment, the knowledge of the large model is used for cold start to solve the problem of unit standardization.

[0157] Taking length units as an example, the units of numerical attribute values ​​include 10cm, 0.1m, and 10inch. By cold-starting and building a unit mapping based on the large language model, the corresponding unit enumeration values ​​are collected according to the attributes, and then the large language model is called to generate unidirectional unit conversion coefficients for each pair, including the forward conversion coefficient oef{a→b} and the reverse conversion coefficient coef{b→a}. If the coefficients cannot be generated, None is output. Then, the llm is called again to generate the reverse coefficient. If the two coefficients multiplied together are equal to 1, that is, coef{a→

[0158] b}*coef{b→a}=1, it is considered that there is bidirectional transformation consistency, and this bidirectional transformation is considered to be credible.

[0159] The reasoning capabilities of large models are ranked according to this credibility, and then inconsistent situations are handled in this order.

[0160] If the bidirectional conversion consistency of the output results of the current large language model is weak, you can call other large language models with stronger specialized capabilities from the topN large model sequence list. For example, you can upgrade from a general large model to an inference large model, such as DeepSeekV3 to DeepSeekR1. If the number of upgrades exceeds the preset limit, for example, more than three times, and there are still inconsistencies, the conversion is considered unreliable and will not be adopted.

[0161] In a specific embodiment, the data format of the unit mapping relationship is as follows: [{unit1:m, unit2:inch, coefficient:39.3701}, {unit1:m, unit2:inch, coefficient:39.3701}].

[0162] (2) Merging of multi-dimensional attributes

[0163] Under some attribute names, there are more complex attribute values ​​containing multiple dimensions, such as volume, area, etc. These attribute values ​​generally contain numerical dimensions of multiple dimensions. This embodiment decomposes the multi-dimensional attributes and processes each dimension separately according to the merging method of single-dimensional attributes.

[0164] 5. Attribute Distribution Evaluation after Merger

[0165] (1) Evaluation of the proportion of “other” attribute values

[0166] Ideally the number of "other" classes should be in the bottom 10% of the combined attribute values.

[0167] If the proportion of "Other" after merging ranks among the top three merged attribute values, there may be a problem with the distribution. In this case, check whether the result is caused by a huge difference in data distribution or a problem with the merge.

[0168] (2) Overall category distribution evaluation

[0169] The labeling results from a large language model may exhibit overall bias. In this embodiment, a T-distribution or normal distribution hypothesis test is performed on the data. If multiple distribution patterns exhibit significant deviations, the model is adjusted and the predictions are re-evaluated to ensure the reliability of the results. Furthermore, when the sample data volume is less than a preset data volume threshold, a T-distribution hypothesis test can be used. When the sample data volume meets the preset data volume threshold, a normal distribution hypothesis test is used.

[0170] This embodiment provides a method for adjusting attribute values, which obtains a key-value pair set of attribute names to be processed; the key-value pair set includes key-value pairs corresponding to the attribute names to be processed under multiple target entities; the key-value pairs include the attribute names to be processed and attribute values; the multiple attribute values ​​are input into a preset large language model to obtain an attribute value merging result of the multiple attribute values; the attribute value merging result includes a merged attribute value corresponding to each attribute value; based on the attribute value merging result, the attribute value of each target entity is adjusted, so that the large language model can provide a dynamically adaptable attribute value merging result according to the overall situation of the key-value pair set, thereby avoiding the problem of transmission error. The method can reduce the errors that may be caused by the traditional dependency attribute comparison table method, thereby reducing manual intervention, improving the accuracy and efficiency of attribute value classification, and improving the technical effect of product data analysis quality; by obtaining a key-value pair set of attribute names to be processed; the key-value pair set includes key-value pairs corresponding to the attribute names to be processed under multiple target entities; the key-value pairs include the attribute names to be processed and attribute values; the multiple attribute values ​​are input into a preset large language model to obtain the attribute value merging result of the multiple attribute values; the attribute value merging result includes the merged attribute value corresponding to each attribute value; based on the attribute value merging result, the attribute value of each target entity is adjusted, which can achieve The large language model gives a dynamically adaptable attribute value merging result based on the overall situation of the key-value pair set, avoiding the errors that may be caused by the traditional reliance on attribute comparison tables, achieving the technical effect of reducing manual intervention, improving the accuracy and efficiency of attribute value classification, and improving the quality of product data analysis; by obtaining unstructured texts of multiple target entities to expand the data source, using pre-trained models to automatically parse the attribute association relationship in the text to form a structured result, and then filtering and aggregating the key-value pair set based on the target attribute name. Through comprehensive analysis of unstructured texts, the coverage of the data source can be improved, the need for manual intervention can be reduced, and the attribute errors caused by format differences or redundant information can be reduced. The risk of missing values ​​is reduced, thereby improving the accuracy of data analysis. By combining a preset attribute name set with sample entity rule extraction to generate benchmark data, using multi-model voting to enhance the reliability of key-value pair results, fusing two first key-value pair sets and a second key-value pair set to construct training data, and optimizing model parameters based on supervised fine-tuning, the quality of training data can be significantly improved, annotation bias can be reduced, and supervised fine-tuning can make the model more adaptable to the unstructured text features of specific business scenarios, improving the accuracy of attribute extraction, and through automated voting and iterative optimization processes, it can reduce the reliance on manual repeated annotation, thereby achieving the effect of improving the accuracy of attribute value merging.By inputting the description text of the sample entity into multiple large language models to obtain the key-value pair extraction results, counting the common attribute values ​​in the output of each model, and then forming a second key-value pair set based on the consensus results, the prediction deviation of a single model due to differences in training data or architecture is reduced through multi-model parallel processing. By screening common attribute values, abnormal results are filtered out, thereby reducing noise interference, thereby improving data reliability while improving the efficiency and quality of automated annotation and improving the accuracy of the key-value pair set; by constructing a preset attribute name set containing attribute type constraints, using type information to strengthen the model's semantic understanding of attribute values, reducing the ambiguity of attributes with the same name, and refining data annotations through type labels, a more comprehensive reference can be provided for attribute value merging, thereby improving the semantic consistency and standardization of attribute extraction and merging process summary; by calling multiple By using a large language model to generate unit conversion coefficients, verifying model reliability based on the product of forward and reverse coefficients, and selecting high-confidence models for unit standardization, this approach addresses the difficulty of manually maintained comparison tables in traditional technologies in covering non-standard units. The self-consistency verification mechanism of forward and reverse coefficients ensures the reliability of model output. Standardizing numerical values ​​after unit unification improves the accuracy of attribute value merging, thereby enhancing data processing efficiency and comparability of results. By introducing preset sampling rules to generate multiple sets of random attribute values, this effectively avoids situations where the model cannot process data due to excessive context length. Using a large language model to perform multiple rounds of inference on different subsets reduces the impact of sampling on the merging results. Frequency analysis and result merging eliminate noise and determine the final merging result, reducing model output bias and improving the coverage of merging results.

[0171] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0172] Based on the same inventive concept, embodiments of the present application also provide a property value adjustment device for implementing the aforementioned property value adjustment method. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more property value adjustment device embodiments provided below can be found in the aforementioned limitations of the property value adjustment method and will not be further elaborated here.

[0173] In one embodiment, Figure 3 As shown, a property value processing device is provided, and the property value processing device includes:

[0174] The key-value pair acquisition module 100 is used to acquire a key-value pair set of a to-be-processed attribute name. The key-value pair set includes key-value pairs corresponding to the to-be-processed attribute name under multiple target entities. The key-value pairs include the to-be-processed attribute name and attribute value.

[0175] The attribute value merging module 200 is configured to input the plurality of attribute values ​​into a preset large language model to obtain an attribute value merging result of the plurality of attribute values, wherein the attribute value merging result includes a merged attribute value corresponding to each attribute value.

[0176] The attribute value processing module 300 is configured to adjust the attribute value of each target entity based on the attribute value merging result.

[0177] In one embodiment, the key-value pair acquisition module 100 is further configured to:

[0178] Obtaining unstructured text of a plurality of target entities;

[0179] Inputting the unstructured text of each target entity into a pre-trained large language model to obtain an attribute key-value pair extraction result corresponding to each target entity;

[0180] Generate a key-value pair set of attribute names to be processed; the key-value pair set includes attribute key-value pair extraction results corresponding to the attribute names to be processed in multiple target entities.

[0181] In one embodiment, the unstructured text includes description text, and the attribute value merging module 200 is further configured to:

[0182] Extracting key-value pairs from attribute information of multiple sample entities according to a preset attribute name set to obtain a first key-value pair set;

[0183] Obtain a second key-value pair set determined by voting of the plurality of sample entities based on the plurality of large language models;

[0184] Determining sample key-value pair training data based on the first key-value pair set and the second key-value pair set;

[0185] Based on the sample key-value pair training data, supervised fine-tuning is performed on the preset large language model to obtain the pre-trained large language model.

[0186] In one embodiment, the attribute value merging module 200 is further configured to:

[0187] Inputting the description text of the sample entity into a plurality of large language models respectively to obtain a key-value pair extraction result corresponding to each of the large language models; the key-value pair extraction result includes at least one second key-value pair;

[0188] Counting common attribute values ​​in the plurality of key-value pair extraction results;

[0189] Based on the second key-value pairs corresponding to the common attribute values, a second key-value pair set of the sample entities is determined.

[0190] In one embodiment, the preset attribute name set includes multiple preset attribute names and an attribute type corresponding to each preset attribute name; the attribute value merging module 200 is further configured to:

[0191] The preset attribute name set and the attribute information of the multiple sample entities are input into the preset large language model to obtain the first key-value pair set; the first key-value pair set includes multiple first key-value pairs of each sample entity, and the attribute type corresponding to each first key-value pair.

[0192] In one embodiment, when the attribute type corresponding to the attribute name to be processed is a numeric type, and the multiple attribute values ​​include at least two value units, the attribute value merging module 200 is further configured to:

[0193] For each pair of the value units of the attribute name to be processed, calling multiple large language models to respectively generate forward conversion coefficients and reverse conversion coefficients for each pair of the value units;

[0194] determining an inference confidence of each of the large language models according to a product result of the forward conversion coefficient and the reverse conversion coefficient of each pair of the value units;

[0195] Based on the inference confidence, a large language model that meets a preset confidence condition is selected, and unit mapping processing is performed on the plurality of attribute values.

[0196] In one embodiment, the attribute value merging module 200 is further configured to:

[0197] According to a preset sampling rule, a plurality of attribute values ​​are sampled multiple times to obtain a plurality of random attribute value sets;

[0198] Inputting the plurality of random attribute value sets into the preset large language model respectively to obtain a sub-attribute value merging result corresponding to each of the random attribute value sets;

[0199] Frequency analysis and result merging are performed on the merging results of the plurality of sub-attribute values ​​to determine an attribute value merging result of the plurality of attribute values.

[0200] Each module in the aforementioned property value adjustment device may be implemented in whole or in part through software, hardware, or a combination thereof. Each module may be embedded in or independent of a processor in a computer device in the form of hardware, or may be stored in a memory in the computer device in the form of software, so that the processor can call and execute the corresponding operations of each module.

[0201] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a method for adjusting a property value is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.

[0202] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0203] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the attribute value adjustment method of any of the above embodiments is implemented:

[0204] Obtain a key-value pair set of attribute names to be processed; the key-value pair set includes key-value pairs corresponding to the attribute names to be processed under multiple target entities; the key-value pairs include the attribute names to be processed and attribute values;

[0205] Inputting the plurality of attribute values ​​into a preset large language model to obtain an attribute value merging result of the plurality of attribute values; the attribute value merging result includes a merged attribute value corresponding to each of the attribute values;

[0206] Based on the attribute value merging result, the attribute value of each target entity is adjusted.

[0207] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the attribute value adjustment method of any of the above embodiments is implemented:

[0208] Obtain a key-value pair set of attribute names to be processed; the key-value pair set includes key-value pairs corresponding to the attribute names to be processed under multiple target entities; the key-value pairs include the attribute names to be processed and attribute values;

[0209] Inputting the plurality of attribute values ​​into a preset large language model to obtain an attribute value merging result of the plurality of attribute values; the attribute value merging result includes a merged attribute value corresponding to each of the attribute values;

[0210] Based on the attribute value merging result, the attribute value of each target entity is adjusted.

[0211] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0212] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processor involved in the various embodiments provided herein may be, but are not limited to, a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic unit, a data processing logic unit based on quantum computing, and the like.

[0213] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0214] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A method for adjusting an attribute value, characterized in that: The attribute value adjustment method includes: Obtain a key-value pair set of attribute names to be processed; the key-value pair set includes key-value pairs corresponding to the attribute names to be processed under multiple target entities; the key-value pairs include the attribute names to be processed and attribute values; Inputting the plurality of attribute values ​​into a preset large language model to obtain an attribute value merging result of the plurality of attribute values; the attribute value merging result includes a merged attribute value corresponding to each of the attribute values; Based on the attribute value merging result, the attribute value of each target entity is adjusted.

2. The attribute value adjustment method according to claim 1, characterized in that: Obtaining a set of key-value pairs to be processed includes: Obtaining unstructured text of a plurality of target entities; Inputting the unstructured text of each target entity into a pre-trained large language model to obtain an attribute key-value pair extraction result corresponding to each target entity; Generate a key-value pair set of attribute names to be processed; the key-value pair set includes attribute key-value pair extraction results corresponding to the attribute names to be processed in multiple target entities.

3. The attribute value adjustment method according to claim 2, characterized in that: The unstructured text includes a description text. Inputting the unstructured text of each target entity into a pre-trained large language model to obtain an attribute key-value pair extraction result corresponding to each target entity includes: Extracting key-value pairs from attribute information of multiple sample entities according to a preset attribute name set to obtain a first key-value pair set; Obtain a second key-value pair set determined by voting of the plurality of sample entities based on the plurality of large language models; Determining sample key-value pair training data based on the first key-value pair set and the second key-value pair set; Based on the sample key-value pair training data, supervised fine-tuning is performed on the preset large language model to obtain the pre-trained large language model.

4. The attribute value adjustment method according to claim 3, characterized in that: The obtaining of a second key-value pair set determined by voting of the plurality of sample entities based on the plurality of large language models comprises: Inputting the description text of the sample entity into a plurality of large language models respectively to obtain a key-value pair extraction result corresponding to each of the large language models; the key-value pair extraction result includes at least one second key-value pair; Counting common attribute values ​​in the plurality of key-value pair extraction results; Based on the second key-value pairs corresponding to the common attribute values, a second key-value pair set of the sample entities is determined.

5. The attribute value adjustment method according to claim 3, characterized in that: The preset attribute name set includes a plurality of preset attribute names and an attribute type corresponding to each of the preset attribute names; The key-value pair extraction is performed on the attribute information of the plurality of sample entities according to the preset attribute name set to obtain a first key-value pair set including: Inputting the preset attribute name set and the attribute information of the plurality of sample entities into a preset large language model to obtain the first key-value pair set; The first key-value pair set includes a plurality of first key-value pairs of each sample entity and an attribute type corresponding to each first key-value pair.

6. The attribute value adjustment method according to claim 5, characterized in that: When the attribute type corresponding to the attribute name to be processed is a numeric type, and the multiple attribute values ​​include at least two value units, before inputting the multiple attribute values ​​into the preset large language model to obtain the attribute value merging result of the multiple attribute values, the method further includes: For each pair of the value units of the attribute name to be processed, calling multiple large language models to respectively generate forward conversion coefficients and reverse conversion coefficients for each pair of the value units; determining an inference confidence of each of the large language models according to a product result of the forward conversion coefficient and the reverse conversion coefficient of each pair of the value units; Based on the inference confidence, a large language model that meets a preset confidence condition is selected, and unit mapping processing is performed on the plurality of attribute values.

7. The attribute value adjustment method according to claim 1, characterized in that: Inputting the plurality of attribute values ​​into a preset large language model to obtain an attribute value merging result of the plurality of attribute values ​​includes: According to a preset sampling rule, a plurality of attribute values ​​are sampled multiple times to obtain a plurality of random attribute value sets; Inputting the plurality of random attribute value sets into the preset large language model respectively to obtain a sub-attribute value merging result corresponding to each of the random attribute value sets; Frequency analysis and result merging are performed on the merging results of the plurality of sub-attribute values ​​to determine an attribute value merging result of the plurality of attribute values.

8. An attribute value processing device, characterized in that: The attribute value processing device includes: A key-value pair acquisition module is used to acquire a key-value pair set of attribute names to be processed; the key-value pair set includes key-value pairs corresponding to the attribute names to be processed under multiple target entities; the key-value pairs include the attribute names to be processed and attribute values; an attribute value merging module, configured to input the plurality of attribute values ​​into a preset large language model to obtain an attribute value merging result of the plurality of attribute values; the attribute value merging result including a merged attribute value corresponding to each attribute value; An attribute value processing module is used to adjust the attribute value of each target entity based on the attribute value merging result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images