Attribute information extraction model training method, attribute information extraction method and apparatus

By training the item title using a multi-head self-attention mechanism model, the problem of poor nested attribute extraction in existing models is solved, and more accurate attribute information extraction is achieved.

CN117744637BActive Publication Date: 2026-11-17BEIJING WODONG TIANJUN INFORMATION TECH CO LTD +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202211123944.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-09-15
Publication Date
2026-11-17
Estimated Expiration
2042-09-15

AI Technical Summary

Technical Problem

Existing natural language processing models are not effective at extracting attribute information from product titles, especially when nested attributes are present. Furthermore, the semantic and grammatical relationships in product titles are weak, resulting in inaccurate extraction.

Method used

A multi-head self-attention mechanism model is used to train the item title. Score sets are generated through character encoding and multi-head self-attention mechanism, and the model is trained in combination with labeled data to identify and extract nested and non-nested attribute information.

Benefits of technology

It improves the accuracy of extracting attribute information from item titles, and can more accurately identify and extract information from nested and non-nested attributes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117744637B_ABST
    Figure CN117744637B_ABST
Patent Text Reader

Abstract

This disclosure presents an embodiment of an attribute information extraction model training method, an attribute information extraction method, and an apparatus. One specific implementation of the method includes: acquiring pre-generated tag data for a target title; preprocessing the target title to obtain a character sequence; inputting each character in the character sequence into a character encoding model in an initial attribute information extraction model to generate a character vector sequence; inputting the character vector sequence into a multi-head self-attention mechanism model in the initial attribute information extraction model to generate a score set; and training the initial attribute information extraction model based on the score set and the tag data to obtain a trained attribute information extraction model. This implementation is related to artificial intelligence; using the attribute information extraction model, the various attribute information contained in the title can be extracted more accurately.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of computer technology, specifically to a method for training an attribute information extraction model, an attribute information extraction method, and an apparatus. Background Technology

[0002] An item title is a high-level summary of the item's characteristics, often containing rich information about its attributes. The common method for extracting this attribute information is to use natural language processing models (e.g., Long Short-Term Memory networks (LSTM)) to extract the title's attributes.

[0003] However, when using the above method to extract attribute information, the following technical problems often arise: First, item titles are often not fluent sentences and have weak semantic and grammatical relationships, making them difficult for general natural language processing models to extract effectively.

[0004] Second, when dealing with nested attributes in item titles, the general natural language processing model is not very effective at extracting attribute information from nested attributes.

[0005] The information disclosed in this background section is only intended to enhance the understanding of the background of the inventive concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this disclosure provide methods for training attribute information extraction models, methods for extracting attribute information, apparatus, electronic devices, and computer-readable media to address the technical problems mentioned in the background section above.

[0008] In a first aspect, some embodiments of this disclosure provide a method for training an attribute information extraction model, comprising: acquiring pre-generated tag data for a target title, wherein the target title is the title of a target item; performing data preprocessing on the target title to obtain a character sequence; inputting each character in the character sequence into a character encoding model in an initial attribute information extraction model to generate a character vector, thereby obtaining a character vector sequence; inputting the character vector sequence into a multi-head self-attention mechanism model in the initial attribute information extraction model to generate a score set, wherein the score set in the score set represents the probability that each word is word information under the corresponding tag category; and training the initial attribute information extraction model based on the score set and the tag data to obtain a trained attribute information extraction model.

[0009] Optionally, the aforementioned multi-head self-attention mechanism model includes: multiple fully connected models, wherein the fully connected models in the multiple fully connected models have a one-to-one correspondence with the label categories in the multiple label categories; and the aforementioned inputting the aforementioned character vector sequence into the aforementioned initial attribute information extraction model of the multi-head self-attention mechanism model to generate a score set includes: for each of the aforementioned multiple label categories, performing a score set generation step: inputting each character vector in the aforementioned character vector sequence into the fully connected model corresponding to the aforementioned label category to generate an output vector, thereby obtaining an output vector sequence; determining the vector similarity between every two output vectors in the aforementioned output vector sequence to generate a similarity score, thereby obtaining a score set.

[0010] Optionally, the above-mentioned training of the initial attribute information extraction model based on the score set and the label data to obtain the trained attribute information extraction model includes: for each score in the score set, performing a numerical determination step: determining two characters in the character sequence corresponding to the score as two target characters; determining a word corresponding to the two target characters based on the character sequence as a target word; determining the score set to which the score belongs as a target score set; determining the label category corresponding to the target score set as a target label category; determining a word set in the label data corresponding to the target label category; and generating a numerical value for the target label category based on the target word and the word set. Optionally, the above-mentioned training of the initial attribute information extraction model based on the above-mentioned score set and the above-mentioned label data to obtain the trained attribute information extraction model further includes: according to the obtained numerical set corresponding to the above-mentioned multiple label categories; adding the numerical values ​​included in the above-mentioned numerical set to obtain the summed value as the loss value; in response to determining that the above-mentioned loss value is greater than the target threshold, updating the model parameters in the initial attribute information extraction model to obtain the updated attribute information extraction model as the trained attribute information extraction model.

[0011] Optionally, the above-mentioned tag data is generated through the following steps: processing the target title according to multiple pre-set tag categories to obtain the above-mentioned tag data.

[0012] Secondly, some embodiments of this disclosure provide an attribute information extraction model training apparatus, comprising: a first acquisition unit configured to acquire pre-generated tag data for a target title, wherein the target title is the title of a target item; a first data preprocessing unit configured to preprocess the target title to obtain a character sequence; a first input unit configured to input each character in the character sequence into a character encoding model in an initial attribute information extraction model to generate a character vector, thereby obtaining a character vector sequence; a second input unit configured to input the character vector sequence into a multi-head self-attention mechanism model in the initial attribute information extraction model to generate a score set, wherein the score set in the score set represents the probability that each word is word information under the corresponding tag category; and a training unit configured to train the initial attribute information extraction model based on the score set and the tag data to obtain a trained attribute information extraction model.

[0013] Optionally, the multi-head self-attention mechanism model includes: multiple fully connected models, wherein the fully connected models in the multiple fully connected models have a one-to-one correspondence with the label categories in the multiple label categories; and the second input unit can be further configured to: for each of the multiple label categories, perform a score set generation step: input each character vector in the character vector sequence into the fully connected model corresponding to the label category to generate an output vector, thereby obtaining an output vector sequence; determine the vector similarity between every two output vectors in the output vector sequence to generate a similarity score, thereby obtaining a score set.

[0014] Optionally, the training unit can be further configured to: for each score in the above score set, perform a numerical determination step: determine two characters in the above character sequence that correspond to the above score as two target characters; determine the word corresponding to the above two target characters according to the above character sequence as a target word; determine the score set to which the above score belongs as a target score set; determine the label category corresponding to the above target score set as a target label category; determine the word set in the above label data that corresponds to the above target label category; and generate a numerical value for the above target label category according to the above target word and the above word set.

[0015] Optionally, the training unit can be further configured to: based on the obtained numerical set corresponding to the above multiple label categories; add the numerical values ​​included in the numerical set to obtain the summed value as the loss value; in response to determining that the loss value is greater than the target threshold, update the model parameters in the above initial attribute information extraction model to obtain the updated attribute information extraction model as the above trained attribute information extraction model.

[0016] Optionally, the above-mentioned tag data is generated through the following steps: processing the target title according to multiple pre-set tag categories to obtain the above-mentioned tag data.

[0017] Thirdly, some embodiments of this disclosure provide an attribute information extraction method, including: obtaining a title to be extracted; performing data preprocessing on the title to be extracted to obtain a character sequence; inputting the character sequence into a pre-trained attribute information extraction model to obtain a set of attribute information corresponding to multiple label categories, wherein the attribute information extraction model is generated by a method as described in any implementation of the first aspect.

[0018] Optionally, the above-mentioned inputting the character sequence into a pre-trained attribute information extraction model to obtain attribute information sets corresponding to multiple label categories includes: inputting each character in the character sequence into the character encoding model in the attribute information extraction model to generate a character vector, thereby obtaining a character vector set; inputting the character vector set into the multi-head self-attention mechanism model in the attribute information extraction model to obtain a score set corresponding to multiple label categories; for each score set corresponding to each label category among the multiple label categories, performing an attribute information set determination step: filtering out scores greater than a predetermined threshold from the score set as filtering scores, thereby obtaining a filtering score set; determining the character corresponding to each filtering score in the filtering score set to generate a character group, thereby obtaining a character group set; determining the word corresponding to each character group in the character group set to obtain a word set corresponding to the label category, thereby obtaining an attribute information set.

[0019] Fourthly, some embodiments of this disclosure provide an attribute information extraction apparatus, comprising: a second acquisition unit configured to acquire a title to be extracted; a second data preprocessing unit configured to preprocess the title to be extracted to obtain a character sequence; and a third input unit configured to input the character sequence into a pre-trained attribute information extraction model to obtain a set of attribute information corresponding to multiple tag categories, wherein the attribute information extraction model is generated by a method as described in any implementation of the first aspect.

[0020] Optionally, the third input unit can be further configured to: input each character in the above character sequence into the character encoding model in the above attribute information extraction model to generate a character vector, thereby obtaining a character vector set; input the above character vector set into the multi-head self-attention mechanism model in the above attribute information extraction model to obtain a set of scores corresponding to multiple label categories; for each set of scores corresponding to each of the above multiple label categories, perform the attribute information set determination step: filter out scores greater than a predetermined threshold from the above score set as filter scores, thereby obtaining a set of filter scores; determine the character corresponding to each filter score in the above set of filter scores to generate a character group, thereby obtaining a character group set; determine the word corresponding to each character group in the above set of character groups to obtain a word set corresponding to the above label category, thereby obtaining an attribute information set.

[0021] Fifthly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any implementation of the first or third aspect.

[0022] Sixthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method as described in any of the implementations of the first or third aspect.

[0023] The above embodiments of this disclosure have the following beneficial effects: By using the attribute information extraction model training method of some embodiments of this disclosure, the attribute information extraction model can more accurately extract the various attribute information contained in the title. Specifically, the reasons for the insufficient accuracy of attribute information extraction are: First, item titles are often not fluent sentences, and the semantic and grammatical relationships are weak, so general natural language processing models often have poor extraction results. Second, for the case of nested attributes in item titles, the general natural language processing model has poor results in extracting the attribute information of nested attributes. Based on this, the attribute information extraction model training method of some embodiments of this disclosure first obtains pre-generated label data for the target title as labels for the initial attribute information extraction model to be trained subsequently. Then, the target title is preprocessed to obtain a character sequence. Here, by dividing the target title into character granularities, the subsequent multi-head self-attention mechanism model can more accurately identify the attribute information of nested attributes and the attribute information of non-nested attributes. Next, each character in the aforementioned character sequence is input into the character encoding model of the initial attribute information extraction model, converting each character in the character sequence into a model input vector, facilitating the subsequent training of the initial attribute information extraction model. Then, the aforementioned character vector sequence is input into the multi-head self-attention mechanism model of the initial attribute information extraction model to generate score sets. Here, the multi-head self-attention mechanism model can accurately determine the score information of the two words corresponding to every two character vectors, thereby accurately identifying the attribute information of nested attributes and non-nested attributes. Finally, based on the aforementioned score sets and the aforementioned label data, the initial attribute information extraction model is trained to obtain the trained attribute information extraction model, which serves as the attribute information extraction model. By using the attribute information extraction model, the attribute information of nested attributes and non-nested attributes contained in the title can be extracted more accurately. Attached Figure Description

[0024] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0025] Figure 1-3 This is a schematic diagram of an application scenario of the attribute information extraction model training method according to some embodiments of this disclosure; Figure 4 This is a flowchart of some embodiments of the attribute information extraction model training method according to this disclosure; Figure 5These are flowcharts of some other embodiments of the attribute information extraction model training method based on this disclosure; Figure 6 This is a schematic diagram illustrating an application scenario of the attribute information extraction method according to some embodiments of this disclosure; Figure 7 This is a flowchart of some embodiments of the attribute information extraction method according to this disclosure; Figure 8 These are schematic diagrams illustrating the structure of some embodiments of the attribute information extraction model training apparatus according to this disclosure; Figure 9 These are schematic diagrams illustrating the structure of some embodiments of the attribute information extraction apparatus according to this disclosure; Figure 10 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0026] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0027] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0028] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0029] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0030] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0031] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0032] Figures 1-3This is a schematic diagram of an application scenario of the attribute information extraction model training method according to some embodiments of this disclosure.

[0033] exist Figures 1-3In the application scenario, first, the electronic device 101 may acquire pre-generated label data 104 for the target title 103. The above target title 103 is the title of the target item 102. In this application scenario, the target item 102 may be "tea". The target title 103 may be "L***n / Li** black tea bag, carefully selected black tea 400g, 200 sachets of bagged tea, for guest rooms, offices and catering". The label data 104 may be "{“Main commodity brand term”: {“Lipton”: [[0,5]], “Li**”: [[7,8]]}, “Main commodity product term”: {“Black tea bag”: [[9,11]], “Black tea”: [[9,10], [16,17]], “Bagged tea”: [[22,24]]}, “Main commodity series term”: {“Yellow label”: [[12,13]]}, “Main commodity style attribute”: {“Carefully selected”: [[14,15]]}, “Main commodity specification attribute”: {“400g”: [[18,21]], “200 sachets”: [[25,28]]}, “Main commodity applicable location”: {“Guest room”: [[29,30]]}, “Main commodity applicable event”: {“Office”: [[31,32]]}, “Main commodity style attribute”: {“Catering pack”: [[33,35]]}}". Then, the electronic device 101 may perform data preprocessing on the above target title 103 to obtain a character sequence 105. In this application scenario, the character sequence 105 may be ""L", "*", "*", "*", "n", "Li", "*", "*", "Red", "Tea", "Bag", "*", "*", "Care", "Selected", "Red", "Tea", "4", "0", "0", "g", "Bag", "Steep", "Tea", "2", "0", "0", "Sachet", "Guest", "Room", "Of", "fice", "Ca", "tering", "Pack"". Next, the electronic device 101 may input each character in the above character sequence 105 into the character encoding model 106 in the initial attribute information extraction model 110 to generate character vectors, so as to obtain a character vector sequence 107.In this application scenario, the character vector sequence 107 can be "[2, 5, 4], [1, 5, 4], [4, 3, 4], [2, 2, 1], [9, 4, 4], [0, 1, 4], [2, 3, 5], [6, 3, 1], [2, 6, 1], [8, 1, 4], [0, 1, 2], [0, 2, 4], [3, 3, 1], [5, 3, 5], [3, 3, 3], [2, 3, 2], [8, 1, 4]". The sequence of characters 107 is given by the electronic device 101. The electronic device 101 can then input the character vector sequence 107 into the multi-head self-attention mechanism model 108 in the initial attribute information extraction model 110 to generate a score set 109. The score set in the score set 109 represents the probability that each word belongs to the corresponding label category. Finally, the electronic device 101 can train the initial attribute information extraction model 110 based on the above-mentioned fraction set 109 and the above-mentioned label data 104 to obtain the trained attribute information extraction model 111. In this application scenario, the fraction set 109 can be "12, 23, -23, 12, 34, -89, 0, 90, ..., 39, 0".

[0034] It should be noted that the aforementioned electronic device 101 can be either hardware or software. When the electronic device is hardware, it can be implemented as a distributed cluster consisting of multiple servers or terminal devices, or as a single server or a single terminal device. When the electronic device is software, it can be installed in the hardware devices listed above. It can be implemented as, for example, multiple software programs or software modules used to provide distributed services, or as a single software program or software module. No specific limitations are made here.

[0035] It should be understood that Figures 1-3 The number of electronic devices shown is merely illustrative. Any number of electronic devices can be used depending on the implementation requirements.

[0036] Continue to refer to Figure 4 The diagram illustrates a flow 400 of some embodiments of an attribute information extraction model training method according to the present disclosure. This attribute information extraction model training method includes the following steps: Step 401: Obtain pre-generated tag data for the target title.

[0037] In some embodiments, the execution entity of the above-described attribute information extraction model training method (e.g.) Figure 1 The electronic device 101 shown can acquire pre-generated tag data for the target title via a wired or wireless connection. The target title is the title of the target item. The target item is the item used for training the subsequent attribute information extraction model. The tag data can be tags used for training the subsequent initial attribute information extraction model. The data format of the tag data can be pre-defined, which will not be elaborated further here.

[0038] For example, the target item is "black tea". The target title could be "L***n / Li** black tea bags, **selected black tea 400g, tea bags 200 bags, hotel / office / catering packaging". The tag data could be {"main product brand name": {"Lipton": [[0,5]], "Li**": [[7,8]]}, "main product product name": {"black tea bags": [[9,11]], "black tea": [[9,10], [[16,17]], "tea bags": [[22,24]]}, "main product series name": {"yellow card": [[12,13]]}, "main product style name": {"yellow card": [[12,13]]}, {" ... "Sex": {"Selected": [[14, 15]}, "Main Product Specification Attributes": {"400g": [[18, 21]], "200 packs": [[25, 28]]}, "Main Product Applicable Location": {"Guest Room": [[29, 30]]}, "Main Product Applicable Event": {"Office": [[31, 32]]}, "Main Product Style Attributes": {"Catering Packaging": [[33, 35]]}}}.

[0039] In some optional implementations of certain embodiments, the above-mentioned tag data is generated through the following steps: Based on multiple pre-defined tag categories, the target title is processed to obtain the tag data. The tag categories can be attribute categories of the item in the title. For example, attribute categories can include, but are not limited to, at least one of the following: item material, item style attribute, item pattern attribute, whether the item is the main product in the title, whether the item is an accessory item, and whether the item is a suitable product. The main product can be the item currently for sale corresponding to the title. Accessory items are salable accessories that appear in the title. Suitable products are items applicable to the main product. The number of title categories included in the multiple title categories can be up to 54. For example, the main product is a battery. The suitable product is a flashlight. Batteries can be used in flashlights. That is, a flashlight is a suitable product for batteries.

[0040] Step 402: Perform data preprocessing on the target title to obtain a character sequence.

[0041] In some embodiments, the execution entity may perform data preprocessing on the target title to obtain a character sequence.

[0042] As an example, firstly, the aforementioned execution entity can remove symbols from the target title to obtain the de-symbolized text. Then, the aforementioned execution entity can perform character segmentation on the de-symbolized text to obtain a character sequence.

[0043] Step 403: Input each character in the above character sequence into the character encoding model in the initial attribute information extraction model to generate a character vector and obtain a character vector sequence.

[0044] In some embodiments, the execution entity can input each character in the character sequence into the character encoding model in the initial attribute information extraction model to generate a character vector, resulting in a character vector sequence. The initial attribute information extraction model can be an attribute information extraction model after model parameter initialization. The attribute information extraction model can be a model used to extract attribute information of each item in the title. In practice, item attributes can include, but are not limited to, at least one of the following: item name, item color, and item weight. The character encoding model can be a model that performs vector conversion on characters. For example, the character encoding model can be, but is not limited to, one of the following: BERT encoding model or Transformer encoding model.

[0045] Step 404: Input the above character vector sequence into the multi-head self-attention mechanism model in the above initial attribute information extraction model to generate score set groups.

[0046] In some embodiments, the execution entity can input the character vector sequence into the multi-head self-attention mechanism model in the initial attribute information extraction model to generate a score set. The score set in the score set represents the probability that each word belongs to the corresponding label category. The score set in the score set corresponds one-to-one with the label categories in the multiple label categories. The multi-head self-attention mechanism model is used to determine the degree of association between every two characters in the character sequence. The score set in the score set can be in matrix form. Each word is a word composed of consecutive characters in the character sequence.

[0047] In some optional implementations of certain embodiments, training the initial attribute information extraction model based on the score set and the label data to obtain the trained attribute information extraction model may include the following steps: For each fraction set in the above fraction set group, perform the numerical determination step: In the first sub-step, for each score in the above-mentioned score set, the above-mentioned execution subject may determine two characters corresponding to the above-mentioned score in the above-mentioned character sequence as a target character group.

[0048] As an example, first, for each score in the above-mentioned score set, the above-mentioned execution subject may determine two character vectors corresponding to the above-mentioned score. Then, the above-mentioned execution subject may determine two characters corresponding to the above two character vectors as a target character group.

[0049] For example, the character sequence may be {"child", "child", "blue", "color", "sleep", "clothes"}. The target character group may be {"blue", "clothes"}.

[0050] In the second sub-step, the above-mentioned execution subject may determine, according to the above-mentioned character sequence, a word corresponding to each target character group in the obtained target character group set as a target word, to obtain a target word set.

[0051] As an example, first, the above-mentioned execution subject may determine the sequence position of the target character group in the character sequence. Then, character concatenation is performed on each character between the two sequence positions corresponding to the target character group to obtain a target word.

[0052] For example, the character sequence may be {"child", "child", "blue", "color", "sleep", "clothes"}. The target character group may be "blue" and "clothes" respectively. The two sequence positions are the third position and the sixth position respectively. Each character between the two sequence positions may include: "blue", "color", "sleep" and "clothes". Then the target word may be "blue pajama".

[0053] In the third sub-step, the above-mentioned execution subject may determine the label category corresponding to the above-mentioned score set as a target label category. There is a one-to-one correspondence between the score sets in the score set group and the label categories in the plurality of label categories. Each score in the score set is a score corresponding to each character group in the character sequence.

[0054] In the fourth sub-step, the above-mentioned execution subject may determine the word set corresponding to the above-mentioned target label category in the above-mentioned label data.

[0055] Wherein, the word set corresponding to the target label category is each word in the target title whose corresponding attribute category is the target label category.

[0056] In the fifth sub-step, the above-mentioned execution subject may generate a value for the above-mentioned target label category according to the above-mentioned target word set and the above-mentioned word set.

[0057] As an example, firstly, the executing entity can determine at least one word belonging to the target word set. Then, it can input at least one score corresponding to that word into a first function to obtain a first sub-value. Next, it can input the score set corresponding to the remaining word set in the target word set (i.e., the word set after removing at least one word) into a second function to obtain a second sub-value. Finally, it can add the first and second sub-values ​​to obtain a value for the target label category.

[0058] In practice, the first function is as follows: , in, This is the first sub-value. For tag categories. It can be a tag category The corresponding vocabulary set. It can be the first in the target title The character and the first Words between characters. It can be in the label category In the case of the first The character and the first The score corresponding to the word between each character.

[0059] In practice, the second function is as follows: , in, This is the second sub-value. For each word in the target title, the corresponding tag category does not belong to the tag type. A collection of words.

[0060] Optionally, the above-mentioned training of the initial attribute information extraction model based on the above-mentioned score set and the above-mentioned label data to obtain the trained attribute information extraction model further includes the following steps: In the first sub-step, the aforementioned executing entity can add the values ​​included in the value set to obtain a summed value, which serves as the loss value. Here, the aforementioned value set is the obtained value set corresponding to the multiple label categories mentioned above.

[0061] In the second sub-step, in response to determining that the loss value is greater than the target threshold, the execution entity can update the model parameters in the initial attribute information extraction model to obtain the updated attribute information extraction model, which serves as the trained attribute information extraction model.

[0062] The aforementioned target threshold can be preset.

[0063] Optionally, in response to determining that the loss value is less than or equal to the target threshold, the executing entity may determine the initial attribute information extraction model as the trained attribute information extraction model.

[0064] Step 405: Based on the above score set and the above label data, train the above initial attribute information extraction model to obtain the trained attribute information extraction model.

[0065] In some embodiments, the execution entity may train the initial attribute information extraction model based on the score set and the label data to obtain a trained attribute information extraction model. The trained attribute information extraction model may be the model after the parameter update is complete.

[0066] The above embodiments of this disclosure have the following beneficial effects: By using the attribute information extraction model training method of some embodiments of this disclosure, the attribute information extraction model can more accurately extract the various attribute information contained in the title. Specifically, the reasons for the insufficient accuracy of attribute information extraction are: First, item titles are often not fluent sentences, and the semantic and grammatical relationships are weak, so general natural language processing models often have poor extraction results. Second, for the case of nested attributes in item titles, the general natural language processing model has poor results in extracting the attribute information of nested attributes. Based on this, the attribute information extraction model training method of some embodiments of this disclosure first obtains pre-generated label data for the target title as labels for the initial attribute information extraction model to be trained subsequently. Then, the target title is preprocessed to obtain a character sequence. Here, by dividing the target title into character granularities, the subsequent multi-head self-attention mechanism model can more accurately identify the attribute information of nested attributes and the attribute information of non-nested attributes. Next, each character in the aforementioned character sequence is input into the character encoding model of the initial attribute information extraction model, converting each character in the character sequence into a model input vector, facilitating the subsequent training of the initial attribute information extraction model. Then, the aforementioned character vector sequence is input into the multi-head self-attention mechanism model of the initial attribute information extraction model to generate score sets. Here, the multi-head self-attention mechanism model can accurately determine the score information of the two words corresponding to every two character vectors, thereby accurately identifying the attribute information of nested attributes and non-nested attributes. Finally, based on the aforementioned score sets and the aforementioned label data, the initial attribute information extraction model is trained to obtain the trained attribute information extraction model, which serves as the attribute information extraction model. By using the attribute information extraction model, the attribute information of nested attributes and non-nested attributes contained in the title can be extracted more accurately.

[0067] Further reference Figure 5 The flowchart 500 illustrates some other embodiments of the attribute information extraction model training method according to the present disclosure. This attribute information extraction model training method includes the following steps: Step 501: Obtain pre-generated tag data for the target title.

[0068] Step 502: Perform data preprocessing on the target title to obtain a character sequence.

[0069] Step 503: Input each character in the above character sequence into the character encoding model in the initial attribute information extraction model to generate a character vector and obtain a character vector sequence.

[0070] Step 504: For each of the above multiple label categories, perform the score set generation step: Step 5041: Input each character vector in the above character vector sequence into the fully connected model corresponding to the above label category to generate an output vector and obtain an output vector sequence.

[0071] In some embodiments, the executing entity (e.g. Figure 1 The electronic device 101 shown can input each character vector in the above character vector sequence into the fully connected model corresponding to the above label category to generate an output vector, thus obtaining an output vector sequence. The fully connected model may include at least one fully connected layer.

[0072] Step 5042: Determine the vector similarity between every two output vectors in the above output vector sequence to generate a similarity score, thus obtaining a score set.

[0073] In some embodiments, the execution entity may determine the vector similarity between every two output vectors in the output vector sequence to generate a similarity score, thereby obtaining a score set.

[0074] As an example, the aforementioned execution entity can determine the cosine distance between every two output vectors in the aforementioned output vector sequence, and obtain a set of fractions as a fraction.

[0075] Step 505: Based on the above score set and the above label data, train the above initial attribute information extraction model to obtain the trained attribute information extraction model.

[0076] In some embodiments, the specific implementation of steps 501-503 and 505 and their resulting technical effects can be found in [reference needed]. Figure 4 Steps 401-403 and 405 in the corresponding embodiments will not be repeated here.

[0077] from Figure 5 It can be seen that, compared with Figure 4 the descriptions of some corresponding embodiments, Figure 5 the process 500 of the attribute information extraction model training method in some corresponding embodiments further highlights the specific steps of generating the score group through a plurality of fully-connected models. Accordingly, for each label category, the solutions described in these embodiments can accurately generate the score between every two characters in the character sequence through the fully-connected model corresponding to each label category.

[0078] Figure 6 is a schematic diagram of an application scenario of the attribute information extraction method according to some embodiments of the present disclosure.

[0079] In Figure 6 the application scenario, first, the electronic device 601 may acquire a to-be-attribute-extracted title 603. In this application scenario, the to-be-attribute-extracted title 603 may be "Xiang**Hao Tea Set, Green Bamboo Health Product". The item 602 corresponding to the to-be-attribute-extracted title 603 is "teapot".

[0080] Then, the electronic device 601 may perform data preprocessing on the above target title 603 to obtain a character sequence 604. In this application scenario, the character sequence 604 may be ""Xiang", "*", "*", "Hao", "Cha", "Ju", "Lv", "Zhu", "Jian", "Kang", "Pin"".

[0081] Finally, the electronic device 601 may input the above character sequence 604 into the pre-trained attribute information extraction model 605 to obtain the attribute information set group 606 corresponding to a plurality of label categories. The above attribute information extraction model 605 is generated by the attribute information extraction model training method according to some embodiments of the present disclosure. In this application scenario, the attribute information set group 606 may be "{"Main Product Brand Word": {"Xiang*": [[0,1]]}, "Main Product Series Word": {"Health": [[9,10]]}, "Main Product Word": {"Green Bamboo": [[8,9]]}}".

[0082] It should be noted that the above electronic device 601 may be hardware or software. When the electronic device is hardware, it may be implemented as a distributed cluster composed of a plurality of servers or terminal devices, or may be implemented as a single server or a single terminal device. When the electronic device is embodied as software, it may be installed in the hardware devices listed above. It may be implemented as, for example, a plurality of software or software modules for providing distributed services, or may be implemented as a single software or software module. No specific limitation is made here.

[0083] It should be understood that Figure 6The number of electronic devices shown is merely illustrative. Any number of electronic devices can be used depending on the implementation requirements.

[0084] Continue to refer to Figure 7 The diagram illustrates a flow 700 of some embodiments of the attribute information extraction method according to the present disclosure. This attribute information extraction method includes the following steps: Step 701: Obtain the title of the attribute to be extracted.

[0085] In some embodiments, the execution entity of the above-described attribute information extraction model training method (e.g.) Figure 1 The electronic device 601 shown can obtain the title of the attribute to be extracted via wired or wireless means.

[0086] Step 702: Perform data preprocessing on the title to be attribute extracted to obtain a character sequence.

[0087] In some embodiments, the execution entity may perform data preprocessing on the target title to obtain a character sequence.

[0088] As an example, firstly, the aforementioned execution entity can remove symbols from the target title to obtain the de-symbolized text. Then, the aforementioned execution entity can perform character segmentation on the de-symbolized text to obtain a character sequence.

[0089] Step 703: Input the above character sequence into the pre-trained attribute information extraction model to obtain attribute information sets corresponding to multiple label categories.

[0090] In some embodiments, the execution entity can input the character sequence into a pre-trained attribute information extraction model to obtain attribute information sets corresponding to multiple label categories. The attribute information extraction model is generated using the attribute information extraction model training method of some embodiments of this disclosure.

[0091] In some optional implementations of certain embodiments, the above-mentioned input of the character sequence into a pre-trained attribute information extraction model to obtain attribute information sets corresponding to multiple label categories may include the following steps: The first step is for the execution entity to input each character in the character sequence into the character encoding model in the attribute information extraction model to generate a character vector and obtain a character vector set.

[0092] In the second step, the aforementioned execution entity can input the aforementioned character vector set into the multi-head self-attention mechanism model in the aforementioned attribute information extraction model to obtain multiple score sets corresponding to label categories.

[0093] The third step is to determine the attribute information set for the score set corresponding to each of the multiple tag categories mentioned above: In the first sub-step, the executing entity can filter out scores greater than a predetermined threshold from the score set to obtain a set of filtered scores. For example, the predetermined threshold can be "0".

[0094] In the second sub-step, the aforementioned executing entity can determine the character corresponding to each filtering score in the aforementioned filtering score set to generate a character group and obtain a character group set.

[0095] Fourth, the aforementioned executing entity can determine the word corresponding to each character group in the aforementioned character group set, and obtain a word set corresponding to the aforementioned tag category, which serves as the attribute information set.

[0096] The above embodiments of this disclosure have the following beneficial effects: by using the attribute information extraction model through the attribute information extraction method of some embodiments of this disclosure, the attribute information of nested attributes and non-nested attributes contained in the title can be extracted more accurately.

[0097] Further reference Figure 8 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of an attribute information extraction model training device, which are similar to... Figure 4 Corresponding to the method embodiments shown, the device can be specifically applied to various electronic devices.

[0098] like Figure 8 As shown, an attribute information extraction model training device 800 includes: a first acquisition unit 801, a first data preprocessing unit 802, a first input unit 803, a second input unit 804, and a training unit 805. The first acquisition unit 801 is configured to acquire pre-generated tag data for a target title, wherein the target title is the title of a target item; the first data preprocessing unit 802 is configured to preprocess the target title to obtain a character sequence; the first input unit 803 is configured to input each character in the character sequence into a character encoding model in an initial attribute information extraction model to generate a character vector, resulting in a character vector sequence; the second input unit 804 is configured to input the character vector sequence into a multi-head self-attention mechanism model in the initial attribute information extraction model to generate a score set, wherein the score set in the score set represents the probability that each word is word information under the corresponding tag category; the training unit 805 is configured to train the initial attribute information extraction model based on the score set and the tag data to obtain a trained attribute information extraction model.

[0099] In some optional implementations of certain embodiments, the multi-head self-attention mechanism model includes: multiple fully connected models, wherein the fully connected models in the multiple fully connected models have a one-to-one correspondence with the label categories in the multiple label categories; and the second input unit 804 can be further configured to: for each of the multiple label categories, perform a score set generation step: input each character vector in the character vector sequence into the fully connected model corresponding to the label category to generate an output vector, thereby obtaining an output vector sequence; determine the vector similarity between every two output vectors in the output vector sequence to generate a similarity score, thereby obtaining a score set.

[0100] In some optional implementations of certain embodiments, the training unit 805 may be further configured to: for each score in the aforementioned score set, perform a numerical determination step: determine two characters in the aforementioned character sequence corresponding to the aforementioned score as two target characters; determine a word corresponding to the aforementioned two target characters based on the aforementioned character sequence as a target word; determine the score set to which the aforementioned score belongs as a target score set; determine the label category corresponding to the aforementioned target score set as a target label category; determine the word set in the aforementioned label data corresponding to the aforementioned target label category; and generate a numerical value for the aforementioned target label category based on the aforementioned target word and the aforementioned word set. In some optional implementations of some embodiments, the training unit 805 may be further configured to: add the values ​​included in the value set to obtain a summed value as a loss value, wherein the value set is the obtained value set corresponding to the plurality of label categories; in response to determining that the loss value is greater than a target threshold, update the model parameters in the initial attribute information extraction model to obtain an updated attribute information extraction model as the trained attribute information extraction model.

[0101] In some optional implementations of certain embodiments, the above-mentioned tag data is generated by the following steps: processing the target title according to a plurality of pre-set tag categories to obtain the above-mentioned tag data.

[0102] It is understandable that the units described in the device 800 are related to the reference. Figure 4 The steps in the described method correspond to each other. Therefore, the operations, features, and beneficial effects described above for the method also apply to the device 800 and the units contained therein, and will not be repeated here.

[0103] Further reference Figure 9 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of an attribute information extraction device, which are similar to... Figure 6Corresponding to the method embodiments shown, the device can be specifically applied to various electronic devices.

[0104] like Figure 9 As shown, an attribute information extraction device 900 includes: a second acquisition unit 901, a second data preprocessing unit 902, and a third input unit 903. The second acquisition unit 901 is configured to acquire a title to be extracted; the second data preprocessing unit 902 is configured to preprocess the title to obtain a character sequence; and the third input unit 903 is configured to input the character sequence into a pre-trained attribute information extraction model to obtain multiple attribute information sets corresponding to various tag categories. The attribute information extraction model is generated using the attribute information extraction model training method of some embodiments of this disclosure.

[0105] In some optional implementations of certain embodiments, the third input unit 903 may be further configured to: input each character in the character sequence into the character encoding model in the attribute information extraction model to generate a character vector, thereby obtaining a character vector set; input the character vector set into the multi-head self-attention mechanism model in the attribute information extraction model to obtain a set of scores corresponding to multiple label categories; for each set of scores corresponding to each of the multiple label categories, perform the attribute information set determination step: filter out scores greater than a predetermined threshold from the set of scores as filtering scores, thereby obtaining a set of filtering scores; determine the character corresponding to each filtering score in the set of filtering scores to generate a character group, thereby obtaining a character group set; determine the word corresponding to each character group in the set of character groups to obtain a word set corresponding to the label category, thereby obtaining an attribute information set.

[0106] It is understandable that the units described in the device 900 are related to the reference. Figure 6 The steps in the described method correspond accordingly. Therefore, the operations, features, and beneficial effects described above for the method also apply to the device 900 and the units contained therein, and will not be repeated here.

[0107] The following is for reference. Figure 10 It illustrates electronic devices suitable for implementing some embodiments of this disclosure (e.g., Figure 1 A schematic diagram of the structure of electronic device 101)1000. Figure 10 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0108] like Figure 10As shown, the electronic device 1000 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1008 into a random access memory (RAM) 1003. The RAM 1003 also stores various programs and data required for the operation of the electronic device 1000. The processing unit 1001, ROM 1002, and RAM 1003 are interconnected via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0109] Typically, the following devices can be connected to the I / O interface 1005: input devices 1006 including, for example, a touchscreen, touchpad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; output devices 1007 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1008 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows electronic device 1000 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 10 An electronic device 1000 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 10 Each box shown can represent a device or multiple devices as needed.

[0110] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 1009, or installed from storage device 1008, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of some embodiments of this disclosure.

[0111] It should be noted that, in some embodiments of this disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0112] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0113] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to: acquire pre-generated tag data for a target title, wherein the target title is the title of a target item; preprocess the target title to obtain a character sequence; input each character in the character sequence into a character encoding model in an initial attribute information extraction model to generate a character vector, resulting in a character vector sequence; input the character vector sequence into a multi-head self-attention mechanism model in the initial attribute information extraction model to generate a score set, wherein the score set in the score set represents the probability that each word is word information under the corresponding tag category; and train the initial attribute information extraction model based on the score set and the tag data to obtain a trained attribute information extraction model. Obtain the title to be extracted; perform data preprocessing on the title to be extracted to obtain a character sequence; input the character sequence into a pre-trained attribute information extraction model to obtain attribute information sets corresponding to multiple label categories, wherein the attribute information extraction model is generated by the attribute information extraction model training method of some embodiments of this disclosure.

[0114] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0115] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0116] The units described in some embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including a first acquisition unit, a first data preprocessing unit, a first input unit, a second input unit, and a training unit. The names of these units do not necessarily limit the specific unit; for example, the first acquisition unit may also be described as "a unit that acquires pre-generated tag data for a target title."

[0117] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0118] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for training an attribute information extraction model, comprising: Obtain pre-generated tag data for the target title, wherein the target title is the title of the target item; The target title is preprocessed to obtain a character sequence; Each character in the character sequence is input into the character encoding model in the initial attribute information extraction model to generate a character vector, thus obtaining a character vector sequence; The character vector sequence is input into the multi-head self-attention mechanism model in the initial attribute information extraction model to generate a score set group, wherein the score set in the score set group represents the probability that each word is word information under the corresponding label category; Based on the score set and the label data, the initial attribute information extraction model is trained to obtain the trained attribute information extraction model. The multi-head self-attention mechanism model includes: multiple fully connected models, wherein each fully connected model has a one-to-one correspondence with a label category among multiple label categories; and The step of inputting the character vector sequence into the multi-head self-attention mechanism model in the initial attribute information extraction model to generate a score set includes: For each of the multiple label categories, perform the score set generation step: Each character vector in the character vector sequence is input into the fully connected model corresponding to the label category to generate an output vector, thus obtaining an output vector sequence; Determine the similarity between every two output vectors in the output vector sequence to generate a similarity score, thus obtaining a score set.

2. The method of claim 1, wherein, The step of training the initial attribute information extraction model based on the score set and the label data to obtain the trained attribute information extraction model includes: For each fraction set in the aforementioned fraction set group, perform the numerical determination step: For each score in the score set, determine two characters in the character sequence that correspond to the score, and use them as the target character group; Based on the character sequence, determine the word corresponding to each target character group in the obtained target character group set, and use it as the target word to obtain the target word set; Determine the label category corresponding to the score set as the target label category; Determine the set of words in the tag data that corresponds to the target tag category; Based on the target word set and the word set, generate a numerical value for the target tag category.

3. The method according to claim 2, wherein, The step of training the initial attribute information extraction model based on the score set and the label data to obtain the trained attribute information extraction model further includes: The numerical values ​​included in the numerical set are added together to obtain a summed value, which is used as the loss value. The numerical set is the obtained numerical set corresponding to the multiple label categories. In response to determining that the loss value is greater than the target threshold, the model parameters in the initial attribute information extraction model are updated to obtain the updated attribute information extraction model, which is used as the trained attribute information extraction model.

4. The method according to claim 1, wherein, The tag data is generated through the following steps: The target title is processed according to multiple pre-set tag categories to obtain the tag data.

5. A method for extracting attribute information, comprising: Get the title of the attribute to be extracted; The title to be attributed is preprocessed to obtain a character sequence; The character sequence is input into a pre-trained attribute information extraction model to obtain attribute information sets corresponding to multiple label categories, wherein the attribute information extraction model is generated based on the method described in any one of claims 1-4.

6. The method according to claim 5, wherein, The step of inputting the character sequence into a pre-trained attribute information extraction model to obtain multiple attribute information sets corresponding to label categories includes: Each character in the character sequence is input into the character encoding model in the attribute information extraction model to generate a character vector, thus obtaining a character vector set; The character vector set is input into the multi-head self-attention mechanism model in the attribute information extraction model to obtain multiple score sets corresponding to label categories; For the score set corresponding to each of the multiple label categories, perform the attribute information set determination step: Scores greater than a predetermined threshold are selected from the score set and used as the selected scores to obtain the selected score set; Determine the character corresponding to each filter score in the filter score set to generate a character group, thus obtaining a character group set; The word corresponding to each character group in the character group set is determined to obtain the word set corresponding to the tag category, which serves as the attribute information set.

7. A training device for an attribute information extraction model, comprising: The first acquisition unit is configured to acquire pre-generated tag data for a target title, wherein the target title is the title of the target item; The first data preprocessing unit is configured to preprocess the target title to obtain a character sequence; The first input unit is configured to input each character in the character sequence into the character encoding model in the initial attribute information extraction model to generate a character vector and obtain a character vector sequence; The second input unit is configured to input the character vector sequence into the multi-head self-attention mechanism model in the initial attribute information extraction model to generate a score set group, wherein the score set in the score set group represents the probability that each word is word information under the corresponding label category; The training unit is configured to train the initial attribute information extraction model based on the score set and the label data to obtain the trained attribute information extraction model. The multi-head self-attention mechanism model includes: multiple fully connected models, wherein there is a one-to-one correspondence between the fully connected models and the label categories in the multiple label categories; The second input unit is further configured to: For each of the multiple label categories, perform the score set generation step: Each character vector in the character vector sequence is input into the fully connected model corresponding to the label category to generate an output vector, thus obtaining an output vector sequence; Determine the similarity between every two output vectors in the output vector sequence to generate a similarity score, thus obtaining a score set.

8. An attribute information extraction device, comprising: The second acquisition unit is configured to acquire the title to be extracted from the attributes; The second data preprocessing unit is configured to preprocess the title to be attributed and obtained a character sequence. The third input unit is configured to input the character sequence into a pre-trained attribute information extraction model to obtain attribute information sets corresponding to multiple label categories, wherein the attribute information extraction model is generated based on the method described in any one of claims 1-4.

9. An electronic device, comprising: One or more processors; Storage device, on which one or more programs are stored, When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-6.

10. A computer-readable medium having a computer program stored thereon, wherein, When the program is executed by the processor, it implements the method as described in any one of claims 1-6.

Citation Information

Patent Citations

  • A method of extracting topics and key information from news

    CN109508459A

  • Commodity description text generation method and device and storage medium

    CN113157910A