Data processing method, device, storage medium and processor

By analyzing the address candidate word vectors using a semantic network model with an attention mechanism, the candidate word with the highest matching degree is determined as the center word. This solves the problem of address recognition difficulties caused by the diversity of synonyms, improves recognition accuracy, and reduces labor costs.

CN112307137BActive Publication Date: 2024-11-05ALIBABA GROUP HOLDING LTD
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN201910675745.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2019-07-24
Publication Date
2024-11-05
Estimated Expiration
2039-07-24

AI Technical Summary

Technical Problem

In existing technologies, the same POI can have a variety of names, making address identification difficult. Existing methods rely on heuristic rules to identify the POI's central words, which has low accuracy, lacks scalability, and has high labor costs.

Method used

A semantic network model based on attention mechanism is adopted. By converting multiple candidate words into word vectors, the matching degree between candidate words and target meaning is analyzed, and the candidate word corresponding to the largest dimension value is determined as the center word.

Benefits of technology

It improves the accuracy of POI central word recognition, reduces labor costs, has good scalability and generalization, and can handle complex situations such as aliases, abbreviations, and typos.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112307137B_ABST
    Figure CN112307137B_ABST
Patent Text Reader

Abstract

The present invention discloses a data processing method, device, storage medium and processor. The method includes: obtaining word vectors corresponding to a plurality of different candidate words, respectively, to obtain a word vector set, wherein the plurality of different candidate words are used to represent the same target meaning; based on a semantic network model including an attention mechanism, analyzing each word vector in the word vector set to obtain a target multidimensional vector, wherein the dimension value under each dimension in the target multidimensional vector is used to quantify the matching degree between the candidate word and the target meaning; determining the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among the plurality of different candidate words. The present invention solves the technical problem that when a central word is obtained from a plurality of synonyms to represent the target meaning, the central word obtained is difficult to accurately represent the target meaning.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a data processing method, device, storage medium and processor. Background Art

[0002] In the current address database, it is common that the same address POI (Point of Interest) has multiple names, such as "Taobao City" and "Alibaba Xixi Park", "Ant Z Space" and "Ali Center Hangzhou Z Space", "Yishu Flower Village" and "Yishu Flower Village Building". The following are the reasons why the same address POI has more messy and diverse POI names: the same POI has some informal names, such as aliases, abbreviations, etc.; POI has administrative division redundancy, such as "Shanghai Jinshan Seagull Building" relative to "Seagull Building"; some POIs have non-generic names, such as "China Anping Insurance Group Co., Ltd." relative to "China Anping"; in addition, POI names containing typos, express information, location information, redundant punctuation, etc. will also result in multiple POI names.

[0003] Since the messy and diverse POI names affect the recognition of addresses, in order to facilitate the management and maintenance of address databases, it is currently urgent to identify the POI core word for multiple synonym lists of the same POI, so as to accurately represent the POI address with the POI core word. The method of identifying the POI core word in the related art is mainly to select the POI core word based on heuristic rules, for example, taking the longest or most frequently used POI synonym as the core word, or comprehensively considering the length and frequency of use of the POI word to select the core word, but the method of selecting the POI core word based on heuristic rules can not only only cover part of the situation, lack of scalability and generalization, but also requires professionals to manually design some judgment rules to determine whether the synonym is the POI core word, such as the length of the POI synonym, the frequency of use, whether it contains punctuation and other judgment rules, resulting in high labor costs and low recognition accuracy.

[0004] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0005] The embodiments of the present invention provide a data processing method, device, storage medium and processor to at least solve the technical problem that when a central word is obtained from multiple synonyms to represent a target meaning, the obtained central word is difficult to accurately represent the target meaning.

[0006] According to one aspect of an embodiment of the present invention, there is provided a data processing method, including: obtaining word vectors corresponding to a plurality of different candidate words, respectively, to obtain a word vector set, wherein the plurality of different candidate words are used to represent the same target meaning; based on a semantic network model including an attention mechanism, analyzing each word vector in the word vector set to obtain a target multidimensional vector, wherein the dimension value under each dimension in the target multidimensional vector is used to quantify the degree of match between the candidate word and the target meaning; and determining the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among the plurality of different candidate words.

[0007] According to one aspect of an embodiment of the present invention, there is provided an address information processing method, comprising: obtaining a plurality of different candidate words input during a voice navigation process, wherein the plurality of different candidate words are used to represent target address information; obtaining word vectors corresponding to the plurality of different candidate words, respectively, to obtain a word vector set; based on a semantic network model including an attention mechanism, analyzing each word vector in the word vector set to obtain a target multidimensional vector, wherein the dimension value under each dimension in the target multidimensional vector is used to quantify the degree of match between the candidate word and the target address information; and determining the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among the plurality of different candidate words.

[0008] According to another aspect of an embodiment of the present invention, there is also provided a device for processing address information, including: an input layer, used to convert multiple different candidate words representing target address information into word vectors respectively; an attention layer, the attention layer is connected to the input layer, and the attention layer is used to extract attention weight information between each word vector, wherein the attention weight information is used to represent the weight of the contribution of the candidate word to identifying the target address information; a fully connected layer, the fully connected layer is connected to the attention layer, and the fully connected layer is used to extract and integrate multiple attention weight information in the attention layer; an output layer, the output layer is connected to the fully connected layer, and is used to convert the integrated multiple attention weight information into a target multi-dimensional vector, wherein the dimension value under each dimension in the target multi-dimensional vector is used to quantify the matching degree between the candidate word and the target address information.

[0009] According to another aspect of an embodiment of the present invention, a data processing device is also provided, including: an acquisition unit, used to acquire word vectors corresponding to multiple different candidate words, respectively, to obtain a word vector set, wherein the multiple different candidate words are used to represent the same target meaning; an analysis unit, used to analyze each word vector in the word vector set based on a semantic network model including an attention mechanism, to obtain a target multidimensional vector, wherein the dimension value under each dimension in the target multidimensional vector is used to quantify the matching degree between the candidate word and the target meaning; a determination unit, used to determine the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among the multiple different candidate words.

[0010] In an embodiment of the present invention, a data processing method is provided, which converts multiple candidate words of the target meaning into a word vector set, and then analyzes the word vector set through a semantic network model to obtain a target multidimensional vector for quantifying the matching degree between the candidate words and the target meaning, and determines the candidate word corresponding to the maximum dimensional value as the central word among multiple different candidate words, thereby achieving the purpose of obtaining a candidate word that accurately represents the target meaning from multiple synonyms, thereby achieving the technical effect that when the central word is obtained from multiple synonyms to represent the target meaning, the obtained central word can accurately represent the target meaning, and further solves the technical problem that when the central word is obtained from multiple synonyms to represent the target meaning, the obtained central word is difficult to accurately represent the target meaning. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0012] Figure 1 is a hardware structure block diagram of a computer terminal according to a data processing method according to an embodiment of the present invention;

[0013] Figure 2 is a flow chart of a data processing method according to an embodiment of the present invention;

[0014] Figure 3 is a schematic diagram of a semantic network model in a data processing method according to an embodiment of the present invention;

[0015] Figure 4 is a schematic diagram of obtaining a central word from a plurality of different candidate words in a data processing method according to an embodiment of the present invention;

[0016] Figure 5 is a flowchart of an address information processing method according to an embodiment of the present invention;

[0017] Figure 6 is a schematic diagram of a data processing device according to an embodiment of the present invention;

[0018] Figure 7 is a schematic diagram of an optional data processing device according to an embodiment of the present application Figure 1 ;

[0019] Figure 8 is a schematic diagram of an optional data processing device according to an embodiment of the present application Figure 2 ;

[0020] Fig. 9 is a schematic diagram of an optional data processing device according to an embodiment of the present application Figure 3;

[0021] Fig.10 is a schematic diagram of an optional data processing device according to an embodiment of the present application Figure 4 ;

[0022] Fig.11 is a schematic diagram of an optional data processing device according to an embodiment of the present application Figure 5 ;

[0023] Fig.12 is a schematic diagram of an optional data processing device according to an embodiment of the present application Figure 6 ;

[0024] Fig.13 is a schematic diagram of an address information processing device according to an embodiment of the present invention; and

[0025] Fig.14 It is a structural block diagram of a computer terminal according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0027] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0028] First, some nouns or terms that appear in the description of the embodiments of the present application are subject to the following explanations:

[0029] POI: Point of Interest, such as scenic spots, government agencies, companies, shopping malls, restaurants, etc. on electronic maps are all POIs.

[0030] Pre-trained word (character) vectors: Use large-scale Chinese corpus to train the neural network language model, and use the neural network language model to train a dense vector representation of each word (character). The word (character) vector contains semantic and syntactic information integrated into the context.

[0031] Attention mechanism: A mechanism used to extract contextual features in deep neural networks.

[0032] Position probability vector: The result of the softmax layer output of the neural network. The probability vector is an N-dimensional vector. The value of the i-th dimension of the vector indicates the probability that the i-th synonym is the POI center word among the N input POI synonyms.

[0033] Example 1

[0034] According to an embodiment of the present invention, an embodiment of a data processing method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0035] The method embodiment provided in the first embodiment of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 FIG. 1 shows a hardware structure block diagram of a computer terminal (or mobile device) for implementing a data processing method. Figure 1 As shown, the computer terminal 10 (or mobile device 10) may include one or more (102a, 102b, ..., 102n are used to illustrate) processors 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission device for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown.

[0036] The memory 104 can be used to store software programs and modules of application software, such as program instructions / data storage devices corresponding to the data processing method in the embodiment of the present invention. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, the data processing method of the above-mentioned application program is realized. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the computer terminal 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0037] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 (or mobile device).

[0038] Under the above operating environment, this application provides Figure 2 The data processing method shown. Figure 2 is a flow chart of a data processing method according to Embodiment 1 of the present invention.

[0039] S21, obtaining word vectors corresponding to a plurality of different candidate words respectively, and obtaining a word vector set, wherein the plurality of different candidate words are used to represent the same target meaning.

[0040] It should be noted that multiple different candidate words representing the same target meaning are obtained, and the multiple different candidate words are synonyms of each other. Each candidate word is converted into a word vector, and the multiple candidate word vectors constitute a word vector set.

[0041] For example, multiple candidate words for an address POI are obtained from the address database collected by the public security system. For example, POI candidate words located at the address of "No. 168, Shixin Road, Xintang Town, Zengcheng District, Guangzhou City, Guangdong Province" are obtained, including "Zhongsen Mingyuan Property Center", "Zhongsen Mingyuan Meiyijia", "Zhongsen Mingyuan Kindergarten", "Zhongsen Mingyuan Commercial Street", "Zhongsen Mingyuan Community", "Zhongsen Mingyuan Construction Site", "Zhongsen Mingyuan", "Zhongsen Mingyuan", "Zhongsen Mingyuan West Gate" and other 10 words. Each candidate word is converted into a vector of the language context in which the candidate word appears, that is, the word vector corresponding to the candidate word, and 10 word vectors are obtained. The 10 word vectors constitute a word vector set. It should be noted that this application does not limit the number of candidate words, and any N words can be selected as candidate words to obtain the central word of the target address.

[0042] According to the above embodiments of the present application, obtaining word vectors corresponding to multiple different candidate words includes: training word vectors corresponding to each character in the candidate words; and determining the word vector corresponding to the candidate word based on the respective word vectors of the characters in the candidate word.

[0043] Specifically, a neural network language model can be trained with a large-scale Chinese corpus to obtain a pre-trained word vector model, which is used to convert text into word vectors. For example, an open-source Bert model is pre-trained with a large-scale Chinese corpus. Each character in the "Zhongsen Mingyuan Community" is input into the open-source Bert model, and the word vector representations corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", and "Qu" are obtained respectively. Then, the word vector of each candidate word is obtained based on the respective word vector representations of the characters in the candidate word. For example, the word vector of the candidate word can be obtained by adding the word vectors.

[0044] There are various methods for obtaining the word vector of a candidate word based on word vectors. According to the above embodiments of the present application, determining the word vector corresponding to the candidate word based on the respective word vectors of the characters in the candidate word includes at least one of the following methods: adding the respective word vectors of the characters in the candidate word to obtain the word vector corresponding to the candidate word; or, using the respective word vectors of the characters in the candidate word as the input of a target model, and training the word vector corresponding to the candidate word through the target model, where the target model includes at least a recurrent neural network model.

[0045] Specifically, the word vector can be obtained by adding word vectors. For example, for the candidate word "Zhongsen Mingyuan Community", the word vectors corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", and "Qu" are respectively obtained by training with the word vector model. By adding the respective word vectors, the word vector corresponding to the candidate word "Zhongsen Mingyuan Community" can be accurately and quickly obtained.

[0046] For more complex word vectors, the word vector corresponding to the candidate word can be obtained by using a recurrent neural network model (Recurrent Neural Network). Specifically, for the candidate word "Zhongsen Mingyuan Community", the word vectors corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", and "Qu" are respectively obtained by training with the word vector model. The respective word vectors are input into the recurrent neural network model for training to obtain the word vector corresponding to "Zhongsen Mingyuan Community".

[0047] In addition, it should be noted that the method of obtaining word vectors by the recurrent neural network model is more complex than the method of adding word vectors, but the effect is more accurate. For relatively simple candidate words, the word vector of the candidate word can be obtained by adding word vectors. For relatively complex candidate words, this method can be used to replace the method of adding word vectors to obtain the word vector of the candidate word.

[0048] S22, based on a semantic network model including an attention mechanism, analyze each word vector in the word vector set to obtain a target multi-dimensional vector, wherein the dimension value under each dimension in the target multi-dimensional vector is used to quantify the matching degree between the candidate word and the target meaning.

[0049] Specifically, the semantic network model is also called the hierarchical network model of semantic memory. This model believes that the basic unit of semantic memory in long-term memory is the concept. The relevant concepts form a hierarchical network system according to the logical superior-subordinate relationship. The subordinate relationship between concepts and the relationship between concepts and features constitute a complex hierarchical network. Each level of the model stores the unique features of the concepts at that level, while the common features of the concepts at the same level are stored at the level of the previous concept.

[0050] It should be noted that the language for representing the target meaning is not limited to Chinese. The present application can also analyze synonyms in other languages ​​to obtain the central word representing the target meaning. Different semantic network models have different hierarchical structures and concept processing mechanisms. The embodiment of the present application analyzes each word vector in the word vector set based on the attention mechanism, extracts the features of the word vector, and determines the weight information of each word vector. The weight information is used to quantify the influence of the candidate word on the recognition of the central word. The multidimensional target vector is determined by the weight information of each word vector, and each dimension value of the multidimensional target vector is used to characterize the degree of match between each candidate word and the target meaning.

[0051] According to the above-mentioned embodiments of the present application, the semantic network model includes an input layer, an attention layer, a fully connected layer and an output layer. Based on the semantic network model including the attention mechanism, each word vector in the word vector set is analyzed to obtain a target multidimensional vector including: the word vectors in the word vector set are converted into multiple attention weight information through the attention mechanism in the attention layer, wherein the attention weight information is used to characterize the weight of the contribution of the candidate word to the recognition of the target meaning; the multiple attention weight information in the attention layer are extracted and integrated through the fully connected layer; and the integrated multiple attention weight information are converted into the target multidimensional vector through the output layer.

[0052] like Figure 3 As shown in the figure, the semantic network model includes an input layer, an attention layer, a fully connected layer and an output layer. The input of the model is a fixed number of candidate words. The input layer contains a character vector conversion model and a word vector generation mechanism, which are used to convert candidate words into word vectors. For example, if 10 POI synonyms are input, the input layer converts the 10 POI synonyms into 10 word vectors.

[0053] The attention layer is used to extract the features of 10 word vectors and calculate 10 attention weight information based on the 10 word vectors. The attention weight information is used to quantify the influence of candidate words on the recognition of the central word, that is, to quantify the contribution of candidate words to the recognition of the target meaning.

[0054] The fully connected layer is used to integrate the output of the attention layer and increase the complexity of the model. Specifically, each node in the fully connected layer is connected to all the nodes in the attention layer to integrate the model weight information extracted by the attention layer.

[0055] The output layer determines a multidimensional vector based on the attention weight information integrated by the attention layer, where each dimension value of the multidimensional vector is determined by an attention weight information to characterize the degree of match between the candidate word corresponding to the dimension and the target meaning.

[0056] This embodiment identifies the central word among multiple candidate words based on a deep semantic network model including an input layer, an attention layer, a fully connected layer and an output layer. It does not require manual design of recognition rules or features. The recognition method has good scalability and generalization, and can effectively handle various POI central word recognition problems such as aliases, abbreviations, typos, administrative division redundancy, etc., greatly improving the accuracy of POI central word recognition.

[0057] According to the above-mentioned embodiment of the present application, converting the integrated multiple attention weight information into a target multidimensional vector through the output layer includes: normalizing the integrated multiple attention weight information through the normalized exponential function in the output layer to obtain multiple probability values, wherein the sum of the multiple probability values ​​is 1; determining the target multidimensional vector based on the multiple probability values, wherein each dimension in the target multidimensional vector corresponds to a probability value.

[0058] Specifically, the output layer in the semantic network model can include a normalization function. Through the normalized exponential function, that is, the softmax function, the multiple attention weight information integrated by the fully connected layer is normalized so that each dimension value in the obtained target multidimensional vector is distributed between 0 and 1. This can intuitively represent the probability that the candidate words corresponding to each dimension belong to the central word, such as Figure 4 As shown, the 10 candidate words "Zhongsen Mingyuan Property Center", "Zhongsen Mingyuan Meiyijia", "Zhongsen Mingyuan Kindergarten", "Zhongsen Mingyuan Commercial Street", "Zhongsen Mingyuan Garden", "Zhongsen Mingyuan Community", "Zhongsen Mingyuan Construction Site", "Zhongsen Mingyuan", "Zhongsen Mingyuan", and "Zhongsen Mingyuan West Gate" are input into the semantic network model, and the output dimension values ​​are 0.003, 0.044, 0.12, 0.045, 0.152, 0.335, 0.067, 0.138, 0.092, and 0.004, respectively.

[0059] In addition, it should be noted that in the process of testing the training results of the model, the probability corresponding to each candidate word in the standard set is set to 0 or 1. Specifically, the probability corresponding to a candidate word belonging to the central word in the standard set is 1, and the probability values ​​of the remaining candidate words that do not belong to the central word are all 0. In this embodiment, the training result of the model is a target multidimensional vector with dimension values ​​distributed between 0 and 1 through the output layer. When judging the accuracy of the training results, the variance can be taken between the training results and the test set, so as to conveniently calculate the accuracy of the semantic network model training results. When the accuracy of the training results is low, the semantic network model is further optimized.

[0060] Through this embodiment, the output results of the fully connected layer are normalized, so that the probability of each candidate word being determined as the central word can be intuitively known, and it is convenient to further optimize the model based on the dimensional value of the target multidimensional vector during the testing of the semantic network model.

[0061] According to the above-mentioned embodiment of the present application, the method also includes: before analyzing each word vector in the word vector set based on a semantic network model including an attention mechanism, determining whether the number of word vectors in the word vector set reaches a first preset number; when the number of word vectors in the word vector set is less than the first preset number, adding a second preset number of zero vectors to the word vector set to obtain an updated word vector set, and analyzing each word vector in the updated word vector set.

[0062] It should be noted that, in order to facilitate the setting of matrix parameters in the hierarchical matrix of the semantic model and the training of the target multidimensional vector, the semantic network model needs to take a candidate word list of fixed length N as input. If the candidate words of the target meaning obtained are less than N, they are padded with 0 vectors in the input layer. For example, the semantic network model needs to take a POI synonym list of length 10 as input. The POI synonyms obtained at the target address are 7, which is less than 10. First, the 7 candidate words obtained are converted into word vectors respectively, and then the 7 candidate word vectors and 3 0 vectors are input into the attention layer together, and the attention weights are extracted based on the 10 candidate word vectors.

[0063] Through this embodiment, the number of input vectors of the attention layer of the semantic network model is reduced, which facilitates the construction of each level of the model in advance and the training of the target multi-dimensional vector.

[0064] S23, determining the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among the multiple different candidate words.

[0065] It should be noted that since the dimension values ​​corresponding to each dimension in the target multidimensional vector are used to represent the weight of the candidate word's contribution to the identification of the target meaning, the candidate word corresponding to the maximum dimension value is the candidate word that best matches the target meaning. Therefore, the candidate word corresponding to the maximum dimension value is determined as the central word among multiple different candidate words, which can more accurately represent the target meaning.

[0066] According to the above embodiments of the present application, the method also includes: after determining the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among multiple different candidate words, judging whether the meaning expressed by the central word matches the target meaning; when the meaning expressed by the central word does not match the target meaning, adjusting the semantic network model, wherein the semantic network model is adjusted in at least one of the following ways: adjusting the hierarchical matrix in each level of the semantic network model, adding multiple attention layers to the semantic network model, and adding multiple fully connected layers to the semantic network model.

[0067] For example, the address POI is "No. 168, Shixin Road, Xintang Town, Zengcheng District, Guangzhou City, Guangdong Province", and "Zhongsen Mingyuan Community" can most accurately represent the address. The 10 candidate words obtained for the address POI include "Zhongsen Mingyuan Property Center", "Zhongsen Mingyuan Meiyijia", "Zhongsen Mingyuan Kindergarten", "Zhongsen Mingyuan Commercial Street", "Zhongsen Mingyuan Community", "Zhongsen Mingyuan Construction Site", "Zhongsen Mingyuan", "Zhongsen Mingyuan", and "Zhongsen Mingyuan West Gate". The candidate word corresponding to the maximum dimension value in the target multidimensional vector trained by the semantic network model is "Zhongsen Mingyuan West Gate". Since "Zhongsen Mingyuan West Gate" is quite different from the address POI, it means that the training result of the semantic network model does not match the target meaning, and the semantic network model needs to be adjusted.

[0068] There are many ways to adjust the semantic network model. For example, the semantic network model can be adjusted by adjusting the hierarchical matrix in each level of the semantic network model. It should be noted that the hierarchical matrix in each level is multiplied with the vector obtained from the previous level to obtain the vector output of the level. The matrix elements in the hierarchical matrix are adjusted to adjust the vector output of each level, thereby changing the output result of the semantic network model.

[0069] For another example, the semantic network model is adjusted by adding multiple attention layers and multiple fully connected layers to the semantic network model. Specifically, multiple attention layers are connected in series, the first attention layer extracts multiple attention weight information based on multiple word vectors obtained from the input layer, and the second attention layer further extracts the attention weight information obtained from the first attention layer. The multiple attention layers continuously extract the attention weight information, so that the attention weight information output by the last attention layer is more accurate. In addition, multiple fully connected layers can be set to perform multi-layer integration on the attention weight information output by the attention layer, so that the values ​​of each dimension of the target multi-dimensional vector can more accurately represent the degree of match between the candidate word and the target meaning, so that the candidate word can be accurately determined based on the target multi-dimensional vector.

[0070] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0071] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0072] Example 2

[0073] According to an embodiment of the present invention, a method for processing address information is also provided. Figure 5 As shown, including:

[0074] Step S51, obtaining a plurality of different candidate words input during the voice navigation process, wherein the plurality of different candidate words are used to represent target address information.

[0075] It should be noted that during the voice navigation process, the user will input the destination. Different users have different descriptions of the same address. The words input by the user can be saved and filtered to obtain multiple different candidate words, and the central word can be obtained from the multiple different candidate words input by the user, so as to more accurately describe the target address information.

[0076] Specifically, the target address that the user needs to reach is "No. 168, Shixin Road, Xintang Town, Zengcheng District, Guangzhou City, Guangdong Province", and the candidate words obtained from the user input include 10 words such as "Zhongsenmingyuan Property Center", "Zhongsenmingyuan Meiyijia", "Zhongsenmingyuan Kindergarten", "Zhongsenmingyuan Commercial Street", "Zhongsenmingyuan Community", "Zhongsenmingyuan Construction Site", "Zhongsenmingyuan", "Zhongsenmingyuan", "West Gate of Zhongsenmingyuan", etc.

[0077] Step S52: Obtain word vectors corresponding to multiple different candidate words respectively to obtain a word vector set.

[0078] For example, convert the 10 obtained candidate words "Zhongsenmingyuan Property Center", "Zhongsenmingyuan Meiyijia", "Zhongsenmingyuan Kindergarten", "Zhongsenmingyuan Commercial Street", "Zhongsenmingyuan Community", "Zhongsenmingyuan Construction Site", "Zhongsenmingyuan", "Zhongsenmingyuan", "West Gate of Zhongsenmingyuan" into corresponding word vectors respectively. Specifically, train the word vectors corresponding to each character in the candidate word, and then determine the word vector corresponding to the candidate word based on the word vectors of each character in the candidate word.

[0079] Specifically, a neural network language model can be trained with a large-scale Chinese corpus first to obtain a pre-trained word vector model, which is used to convert text into word vectors. For example, use the open-source Bert model pre-trained with a large-scale Chinese corpus, input each character in "Zhongsenmingyuan Community" into the open-source Bert model respectively, and obtain the word vector representations corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", "Qu" respectively. Then, based on the word vector representations of each character in the candidate word, obtain the word vector of each candidate word. For example, the word vector can be obtained by adding the word vectors. For example, for the candidate word "Zhongsenmingyuan Community", the word vectors corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", "Qu" are obtained by training with the word vector model respectively, and the word vectors are added together to accurately and quickly obtain the word vector corresponding to the candidate word "Zhongsenmingyuan Community". For more complex word vectors, the candidate word corresponding word vector can be obtained by using the Recurrent Neural Network model. Specifically, for the candidate word "Zhongsenmingyuan Community", the word vectors corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", "Qu" are obtained by training with the word vector model respectively, and each word vector is input into the recurrent neural network model for training to obtain the word vector corresponding to "Zhongsenmingyuan Community".

[0080] Step S53: Based on a semantic network model including an attention mechanism, analyze each word vector in the word vector set to obtain a target multi-dimensional vector, where the dimension value of each dimension in the target multi-dimensional vector is used to quantify the matching degree between the candidate word and the target address information.

[0081] Specifically, the semantic network model is also called the hierarchical network model of semantic memory. This model believes that the basic unit of semantic memory in long-term memory is the concept. The relevant concepts form a hierarchical network system according to the logical superior-subordinate relationship. The subordinate relationship between concepts and the relationship between concepts and features constitute a complex hierarchical network. Each level of the model stores the unique features of the concepts at that level, while the common features of the concepts at the same level are stored at the level of the previous concept.

[0082] For example, the semantic network model including the attention mechanism analyzes the word vectors of 10 candidate words "Zhongsen Mingyuan Property Center", "Zhongsen Mingyuan Meiyijia", "Zhongsen Mingyuan Kindergarten", "Zhongsen Mingyuan Commercial Street", "Zhongsen Mingyuan Garden", "Zhongsen Mingyuan Community", "Zhongsen Mingyuan Construction Site", "Zhongsen Mingyuan", "Zhongsen Mingyuan", and "Zhongsen Mingyuan West Gate", and outputs target multi-dimensional vectors with dimension values ​​of 0.003, 0.044, 0.12, 0.045, 0.152, 0.335, 0.067, 0.138, 0.092, and 0.004 respectively.

[0083] It should be noted that different semantic network models have different hierarchical structures and concept processing mechanisms. The embodiment of the present application analyzes each word vector in the word vector set based on the attention mechanism, extracts the characteristics of the word vector, and determines the weight information of each word vector. The weight information is used to quantify the influence of the candidate word on the recognition of the central word. The multidimensional target vector is determined by the weight information of each word vector, and each dimension value of the multidimensional target vector is used to characterize the matching degree between each candidate word and the target address information.

[0084] Step S54: determine the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among the multiple different candidate words.

[0085] It should be noted that since the dimension values ​​corresponding to each dimension in the target multidimensional vector are used to represent the weight of the candidate word's contribution to identifying the target meaning, the candidate word corresponding to the maximum dimension value is the candidate word that best matches the target meaning. Therefore, the candidate word corresponding to the maximum dimension value is determined as the central word among multiple different candidate words, which can more accurately represent the target address information.

[0086] Example 3

[0087] According to an embodiment of the present invention, there is also provided a device for processing address information, comprising: an input layer, an attention layer, a fully connected layer and an output layer.

[0088] Specifically, the input layer is used to convert multiple different candidate words representing the target address information into word vectors respectively.

[0089] For example, multiple different candidate words representing target address information are multiple candidate words for the same address POI. Multiple candidate words for an address POI are obtained from the address database collected from the public security system. For example, candidate words for the POI located at the address "No. 168, Shixin Road, Xintang Town, Zengcheng District, Guangzhou City, Guangdong Province" are obtained, including "Zhongsenmingyuan Property Management Center", "Zhongsenmingyuan Meiyijia", "Zhongsenmingyuan Kindergarten", "Zhongsenmingyuan Commercial Street", "Zhongsenmingyuan Community", "Zhongsenmingyuan Construction Site", "Zhongsenmingyuan", "Zhongsenmingyuan", "West Gate of Zhongsenmingyuan", etc. A total of 10 candidate words are obtained and input into the input layer. The input layer converts the multiple candidate words into corresponding word vectors respectively. Specifically, the word vectors corresponding to each character in the candidate words are trained, and then the word vector corresponding to the candidate word is determined based on the word vectors of each character in the candidate word.

[0090] Specifically, a neural network language model can be trained first using a large-scale Chinese corpus to obtain a pre-trained word vector model, which is used to convert text into word vectors. For example, an open-source Bert model is pre-trained using a large-scale Chinese corpus. Each character in "Zhongsenmingyuan Community" is input into the open-source Bert model respectively to obtain the word vector representations corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", and "Qu". Then, the word vector of each candidate word is obtained based on the word vector representations of each character in the candidate word. For example, the word vector can be obtained by adding the word vectors. For example, for the candidate word "Zhongsenmingyuan Community", the word vectors corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", and "Qu" are obtained by training with the word vector model respectively, and the word vectors are added together to accurately and quickly obtain the word vector corresponding to the candidate word "Zhongsenmingyuan Community". For more complex word vectors, the word vector corresponding to the candidate word can be obtained by using a recurrent neural network model (Recurrent Neural Network). Specifically, for the candidate word "Zhongsenmingyuan Community", the word vectors corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", and "Qu" are obtained by training with the word vector model respectively, and each word vector is input into the recurrent neural network model for training to obtain the word vector corresponding to "Zhongsenmingyuan Community".

[0091] Attention layer. The attention layer is connected to the input layer and is used to extract the attention weight information between each word vector. Among them, the attention weight information is used to represent the weight of the contribution of the candidate word to the recognition of the target address information.

[0092] For example, if 10 POI synonyms are input, the input layer converts the 10 POI synonyms into 10 word vectors. The attention layer is used to extract the features of the 10 word vectors and calculate 10 attention weights based on the 10 word vectors. The attention weights are used to quantify the influence of the candidate words on the recognition of the central word, that is, to quantify the contribution of the candidate words to the recognition of the target meaning.

[0093] The fully connected layer is connected to the attention layer. The fully connected layer is used to extract and integrate multiple attention weight information in the attention layer.

[0094] It should be noted that the fully connected layer is used to integrate the output of the attention layer and increase the complexity of the model. Specifically, each node in the fully connected layer is connected to all the nodes of the attention layer to integrate the model weight information extracted by the attention layer.

[0095] The output layer is connected to the fully connected layer and is used to convert the integrated multiple attention weight information into a target multi-dimensional vector, wherein the dimension value of each dimension in the target multi-dimensional vector is used to quantify the matching degree between the candidate word and the target address information.

[0096] For example, the output layer may include a normalization function, which normalizes the multiple attention weight information integrated by the fully connected layer through a normalized exponential function, that is, a softmax function, so that each dimension value in the obtained target multidimensional vector is distributed between 0 and 1, thereby intuitively representing the probability that the candidate words corresponding to each dimension belong to the central word, such as Figure 4 As shown, the 10 candidate words "Zhongsen Mingyuan Property Center", "Zhongsen Mingyuan Meiyijia", "Zhongsen Mingyuan Kindergarten", "Zhongsen Mingyuan Commercial Street", "Zhongsen Mingyuan Garden", "Zhongsen Mingyuan Community", "Zhongsen Mingyuan Construction Site", "Zhongsen Mingyuan", "Zhongsen Mingyuan", and "Zhongsen Mingyuan West Gate" are input into the semantic network model, and the output dimension values ​​are 0.003, 0.044, 0.12, 0.045, 0.152, 0.335, 0.067, 0.138, 0.092, and 0.004, respectively.

[0097] Through this embodiment, multiple different candidate words used to represent the target address information are converted into a multidimensional vector through the input layer, attention layer, fully connected layer and output layer. Each dimension value of the multidimensional vector is determined by an attention weight information to represent the degree of match between the candidate word corresponding to the dimension and the target meaning, so that the probability of each candidate word being determined as the central word can be intuitively known.

[0098] Example 4

[0099] According to an embodiment of the present invention, a device for implementing the above data processing is also provided. Figure 6 As shown, the device includes: an acquisition unit 61, an analysis unit 62 and a determination unit 63.

[0100] Specifically, the acquisition unit 61 is used to acquire word vectors corresponding to a plurality of different candidate words, respectively, to obtain a word vector set, wherein the plurality of different candidate words are used to represent the same target meaning.

[0101] It should be noted that multiple different candidate words representing the same target meaning are obtained, and the multiple different candidate words are synonyms of each other. Each candidate word is converted into a word vector, and the multiple candidate word vectors constitute a word vector set.

[0102] For example, multiple candidate words for an address POI are obtained from the address database collected by the public security system. For example, candidate words for the POI located at the address of "No. 168, Shixin Road, Xintang Town, Zengcheng District, Guangzhou City, Guangdong Province" are obtained, including "Zhongsen Mingyuan Property Center", "Zhongsen Mingyuan Meiyijia", "Zhongsen Mingyuan Kindergarten", "Zhongsen Mingyuan Commercial Street", "Zhongsen Mingyuan Community", "Zhongsen Mingyuan Construction Site", "Zhongsen Mingyuan", "Zhongsen Mingyuan", "Zhongsen Mingyuan West Gate" and other 10 words. Each candidate word is converted into a vector of the language context in which the candidate word appears, that is, the word vector corresponding to the candidate word, and 10 word vectors are obtained. The 10 word vectors constitute a word vector set.

[0103] The analysis unit 62 is used to analyze each word vector in the word vector set based on a semantic network model including an attention mechanism to obtain a target multi-dimensional vector, wherein the dimension value under each dimension in the target multi-dimensional vector is used to quantify the matching degree between the candidate word and the target meaning.

[0104] Specifically, the semantic network model is also called the hierarchical network model of semantic memory. This model believes that the basic unit of semantic memory in long-term memory is the concept. The relevant concepts form a hierarchical network system according to the logical superior-subordinate relationship. The subordinate relationship between concepts and the relationship between concepts and features constitute a complex hierarchical network. Each level of the model stores the unique features of the concepts at that level, while the common features of the concepts at the same level are stored at the level of the previous concept.

[0105] It should be noted that different semantic network models have different hierarchical structures and concept processing mechanisms. The embodiment of the present application analyzes each word vector in the word vector set based on the attention mechanism, extracts the characteristics of the word vector, and determines the weight information of each word vector. The weight information is used to quantify the influence of the candidate word on the recognition of the central word. The multidimensional target vector is determined by the weight information of each word vector, and each dimension value of the multidimensional target vector is used to characterize the degree of match between each candidate word and the target meaning.

[0106] A determining unit 63, configured to determine, from the plurality of different candidate words, a central word as the candidate word corresponding to the maximum dimension value in the target multi-dimensional vector.

[0107] It should be noted that, since the dimension values corresponding to the respective dimensions in the target multi-dimensional vector are used to represent the weights of the contributions of the candidate words to the recognition of the target meaning, the candidate word corresponding to the maximum dimension value is the candidate word that best matches the target meaning. Therefore, determining the candidate word corresponding to the maximum dimension value as the central word among the plurality of different candidate words can more accurately represent the target meaning.

[0108] It should be noted here that the above-mentioned obtaining unit 61, analyzing unit 62, and determining unit 63 correspond to steps S21 to S23 in Embodiment 1. The examples and application scenarios implemented by the three modules and the corresponding steps are the same, but are not limited to the content disclosed in the above-mentioned Embodiment 1. It should be noted that the above-mentioned modules, as part of the device, can run in the computer terminal 10 provided in Embodiment 1.

[0109] According to the above embodiments of the present application, as Figure 7 shown, the obtaining unit 61 includes: a training module 611 and a determining module 612.

[0110] The training module 611 is configured to train a word vector corresponding to each character in the candidate words.

[0111] The determining module 612 is configured to determine a word vector corresponding to the candidate word based on the respective word vectors in the candidate words.

[0112] Specifically, a neural network language model can be trained with a large-scale Chinese corpus to obtain a pre-trained word vector model, which is used to convert text into word vectors. For example, an open-source Bert model is pre-trained with a large-scale Chinese corpus. Each character in "Zhongsen Mingyuan Community" is input into the open-source Bert model respectively to obtain the word vector representations corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", and "Qu", and then the word vector of each candidate word is obtained based on the respective word vector representations in the candidate words. For example, the word vector of the candidate word can be obtained by adding the word vectors.

[0113] It should be noted here that the above-mentioned training module 611 and determining module 612 correspond to the steps in Embodiment 1. The examples and application scenarios implemented by the two modules and the corresponding steps are the same, but are not limited to the content disclosed in the above-mentioned Embodiment 1. It should be noted that the above-mentioned modules, as part of the device, can run in the computer terminal 10 provided in Embodiment 1.

[0114] According to the above embodiments of the present application, as Figure 8As shown, the determination module 612 includes at least one of the following sub-modules: an addition sub-module 6121 or a training sub-module 6122.

[0115] The addition sub-module 6121 is configured to add the word vectors of each character in the candidate word to obtain the word vector corresponding to the candidate word; or,

[0116] The training sub-module 6122 is configured to use the word vectors of each character in the candidate word as the input of the target model, and train the word vector corresponding to the candidate word through the target model, where the target model includes at least a recurrent neural network model.

[0117] Specifically, the word vector can be obtained by adding word vectors. For example, for the candidate word "Zhongsen Mingyuan Community", the word vectors corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", and "Qu" are respectively trained by the word vector model, and the word vectors are added to accurately and quickly obtain the word vector corresponding to the candidate word "Zhongsen Mingyuan Community".

[0118] For more complex word vectors, the word vector corresponding to the candidate word can be obtained by using a recurrent neural network model (Recurrent Neural Network). Specifically, for the candidate word "Zhongsen Mingyuan Community", the word vectors corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", and "Qu" are respectively trained by the word vector model, and the word vectors are input into the recurrent neural network model for training to obtain the word vector corresponding to "Zhongsen Mingyuan Community".

[0119] In addition, it should be noted that the method of obtaining word vectors by the recurrent neural network model is more complex than the method of adding word vectors, but the effect is more accurate. For relatively simple candidate words, the word vector of the candidate word can be obtained by adding word vectors. For relatively complex candidate words, this method can be used to replace the method of adding word vectors to obtain the word vector of the candidate word.

[0120] It should be noted here that the above addition sub-module 6121 or training sub-module 6122 corresponds to the steps in Embodiment 1. The instances and application scenarios implemented by the two modules and the corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1 above. It should be noted that the above modules, as part of the device, can run in the computer terminal 10 provided in Embodiment 1.

[0121] According to the above embodiments of the present application, as Fig. 9 shown, the device further includes: a first determination unit 91 and an update unit 92.

[0122] The first judgment unit 91 is used to judge whether the number of word vectors in the word vector set reaches a first preset number before analyzing each word vector in the word vector set based on a semantic network model including an attention mechanism.

[0123] An updating unit 92 is used to add a second preset number of zero vectors to the word vector set when the number of word vectors in the word vector set is less than the first preset number to obtain an updated word vector set, and analyze each word vector in the updated word vector set.

[0124] It should be noted that, in order to facilitate the setting of matrix parameters in the hierarchical matrix of the semantic model and the training of the target multidimensional vector, the semantic network model needs to take a candidate word list of fixed length N as input. If the candidate words of the target meaning obtained are less than N, they are padded with 0 vectors in the input layer. For example, the semantic network model needs to take a POI synonym list of length 10 as input. The POI synonyms obtained at the target address are 7, which is less than 10. First, the 7 candidate words obtained are converted into word vectors respectively, and then the 7 candidate word vectors and 3 0 vectors are input into the attention layer together, and the attention weights are extracted based on the 10 candidate word vectors.

[0125] Through this embodiment, the number of input vectors of the attention layer of the semantic network model is reduced, which facilitates the construction of each level of the model in advance and the training of the target multi-dimensional vector.

[0126] It should be noted that the first judgment unit 91 and the updating unit 92 correspond to the steps in Example 1, and the examples and application scenarios implemented by the two modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above-mentioned Example 1. It should be noted that the above-mentioned modules as part of the device can be run in the computer terminal 10 provided in Example 1.

[0127] According to the above embodiments of the present application, Fig.10 As shown, the semantic network model includes an input layer, an attention layer, a fully connected layer and an output layer, and the analysis unit 62 includes: a first conversion module 621, an extraction module 622 and a second conversion module 623.

[0128] The first conversion module 621 is used to convert the word vectors in the word vector set into multiple attention weight information through the attention mechanism in the attention layer, wherein the attention weight information is used to represent the weight of the contribution of the candidate word to the recognition of the target meaning.

[0129] The extraction module 622 is used to extract and integrate multiple attention weight information in the attention layer through the fully connected layer.

[0130] The second conversion module 623 is used to convert the integrated multiple attention weight information into a target multi-dimensional vector through the output layer.

[0131] like Figure 3 As shown in the figure, the semantic network model includes an input layer, an attention layer, a fully connected layer and an output layer. The input of the model is a fixed number of candidate words. The input layer contains a character vector conversion model and a word vector generation mechanism, which are used to convert candidate words into word vectors. For example, if 10 POI synonyms are input, the input layer converts the 10 POI synonyms into 10 word vectors.

[0132] The attention layer is used to extract the features of 10 word vectors and calculate 10 attention weight information based on the 10 word vectors. The attention weight information is used to quantify the influence of candidate words on the recognition of the central word, that is, to quantify the contribution of candidate words to the recognition of the target meaning.

[0133] The fully connected layer is used to integrate the output of the attention layer and increase the complexity of the model. Specifically, each node in the fully connected layer is connected to all the nodes in the attention layer to integrate the model weight information extracted by the attention layer.

[0134] The output layer determines a multidimensional vector based on the attention weight information integrated by the attention layer, where each dimension value of the multidimensional vector is determined by an attention weight information to characterize the degree of match between the candidate word corresponding to the dimension and the target meaning.

[0135] This embodiment identifies the central word among multiple candidate words based on a deep semantic network model including an input layer, an attention layer, a fully connected layer and an output layer. It does not require manual design of recognition rules or features. The recognition method has good scalability and generalization, and can effectively handle various POI central word recognition problems such as aliases, abbreviations, typos, administrative division redundancy, etc., greatly improving the accuracy of POI central word recognition.

[0136] It should be noted that the first conversion module 621, the extraction module 622 and the second conversion module 623 correspond to the steps in Example 1, and the three modules and the corresponding steps implement the same examples and application scenarios, but are not limited to the contents disclosed in the above-mentioned Example 1. It should be noted that the above-mentioned modules, as part of the device, can be run in the computer terminal 10 provided in Example 1.

[0137] According to the above embodiments of the present application, Fig.11 As shown, the second conversion module 623 includes: a normalization submodule 6231 and a determination submodule 6232 .

[0138] The normalization submodule 6231 is used to normalize the integrated multiple attention weight information through the normalized exponential function in the output layer to obtain multiple probability values, wherein the sum of the multiple probability values ​​is 1.

[0139] The determination submodule 6232 is used to determine the target multi-dimensional vector based on multiple probability values, wherein each dimension in the target multi-dimensional vector corresponds to a probability value.

[0140] Specifically, the output layer in the semantic network model can include a normalization function. Through the normalized exponential function, that is, the softmax function, the multiple attention weight information integrated by the fully connected layer is normalized so that each dimension value in the obtained target multidimensional vector is distributed between 0 and 1. This can intuitively represent the probability that the candidate words corresponding to each dimension belong to the central word, such as Figure 4 As shown, the 10 candidate words "Zhongsen Mingyuan Property Center", "Zhongsen Mingyuan Meiyijia", "Zhongsen Mingyuan Kindergarten", "Zhongsen Mingyuan Commercial Street", "Zhongsen Mingyuan Garden", "Zhongsen Mingyuan Community", "Zhongsen Mingyuan Construction Site", "Zhongsen Mingyuan", "Zhongsen Mingyuan", and "Zhongsen Mingyuan West Gate" are input into the semantic network model, and the output dimension values ​​are 0.003, 0.044, 0.12, 0.045, 0.162, 0.335, 0.067, 0.138, 0.092, and 0.004, respectively.

[0141] In addition, it should be noted that in the process of testing the training results of the model, the probability corresponding to each candidate word in the standard set is set to 0 or 1. Specifically, the probability corresponding to a candidate word belonging to the central word in the standard set is 1, and the probability values ​​of the remaining candidate words that do not belong to the central word are all 0. In this embodiment, the training result of the model is a target multidimensional vector with dimension values ​​distributed between 0 and 1 through the output layer. When judging the accuracy of the training results, the variance can be taken between the training results and the test set, so as to conveniently calculate the accuracy of the semantic network model training results. When the accuracy of the training results is low, the semantic network model is further optimized.

[0142] Through this embodiment, the output results of the fully connected layer are normalized, so that the probability of each candidate word being determined as the central word can be intuitively known, and it is convenient to further optimize the model based on the dimensional value of the target multidimensional vector during the testing of the semantic network model.

[0143] It should be noted that the above normalization submodule 6231 and determination submodule 6232 correspond to the steps in Example 1, and the examples and application scenarios implemented by the two modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above Example 1. It should be noted that the above modules, as part of the device, can be run in the computer terminal 10 provided in Example 1.

[0144] According to the above embodiments of the present application, Fig.12 As shown, the device further includes: a second judgment unit 121 and an adjustment unit 122.

[0145] The second judgment unit 121 is used to judge whether the meaning expressed by the central word matches the target meaning after determining the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among the multiple different candidate words;

[0146] The adjustment unit 122 is used to adjust the semantic network model when the meaning expressed by the central word does not match the target meaning, wherein the semantic network model is adjusted in at least one of the following ways: adjusting the hierarchical matrix in each level of the semantic network model, adding multiple attention layers to the semantic network model, and adding multiple fully connected layers to the semantic network model.

[0147] For example, the address POI is "No. 168, Shixin Road, Xintang Town, Zengcheng District, Guangzhou City, Guangdong Province", and "Zhongsen Mingyuan Community" can most accurately represent the address. The 10 candidate words obtained for the address POI include "Zhongsen Mingyuan Property Center", "Zhongsen Mingyuan Meiyijia", "Zhongsen Mingyuan Kindergarten", "Zhongsen Mingyuan Commercial Street", "Zhongsen Mingyuan Community", "Zhongsen Mingyuan Construction Site", "Zhongsen Mingyuan", "Zhongsen Mingyuan", and "Zhongsen Mingyuan West Gate". The candidate word corresponding to the maximum dimension value in the target multidimensional vector trained by the semantic network model is "Zhongsen Mingyuan West Gate". Since "Zhongsen Mingyuan West Gate" is quite different from the address POI, it means that the training result of the semantic network model does not match the target meaning, and the semantic network model needs to be adjusted.

[0148] There are many ways to adjust the semantic network model. For example, the semantic network model can be adjusted by adjusting the hierarchical matrix in each level of the semantic network model. It should be noted that the hierarchical matrix in each level is multiplied with the vector obtained from the previous level to obtain the vector output of the level. The matrix elements in the hierarchical matrix are adjusted to adjust the vector output of each level, thereby changing the output result of the semantic network model.

[0149] For another example, the semantic network model is adjusted by adding multiple attention layers and multiple fully connected layers to the semantic network model. Specifically, multiple attention layers are connected in series, the first attention layer extracts multiple attention weight information based on multiple word vectors obtained from the input layer, and the second attention layer further extracts the attention weight information obtained from the first attention layer. The multiple attention layers continuously extract the attention weight information, so that the attention weight information output by the last attention layer is more accurate. In addition, multiple fully connected layers can be set to perform multi-layer integration on the attention weight information output by the attention layer, so that the values ​​of each dimension of the target multi-dimensional vector can more accurately represent the degree of match between the candidate word and the target meaning, so that the candidate word can be accurately determined based on the target multi-dimensional vector.

[0150] It should be noted that the second judgment unit 121 and the adjustment unit 122 correspond to the steps in Example 1, and the examples and application scenarios implemented by the two modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above-mentioned Example 1. It should be noted that the above-mentioned modules as part of the device can be run in the computer terminal 10 provided in Example 1.

[0151] Example 5

[0152] According to an embodiment of the present invention, there is also provided an address information processing device, such as Fig.13 As shown, including:

[0153] The target first acquisition unit 131 is used to acquire a plurality of different candidate words input during the voice navigation process, wherein the plurality of different candidate words are used to represent the target address information.

[0154] It should be noted that during the voice navigation process, the user will input the destination. Different users have different descriptions of the same address. The words input by the user can be saved and filtered to obtain multiple different candidate words, and the central word can be obtained from the multiple different candidate words input by the user, so as to more accurately describe the target address information.

[0155] Specifically, the target address the user needs to reach is "No. 168, Shixin Road, Xintang Town, Zengcheng District, Guangzhou City, Guangdong Province". The candidate words entered by the user include "Zhongsen Mingyuan Property Center", "Zhongsen Mingyuan Meiyajia", "Zhongsen Mingyuan Kindergarten", "Zhongsen Mingyuan Commercial Street", "Zhongsen Mingyuan Community", "Zhongsen Mingyuan Construction Site", "Zhongsen Mingyuan", "Zhongsen Mingyuan", "Zhongsen Mingyuan West Gate" and other 10 words.

[0156] The target second acquisition unit 132 is used to obtain word vectors corresponding to multiple different candidate words, and obtain a word vector set.

[0157] For example, convert the 10 candidate words obtained, namely "Zhongsen Mingyuan Property Management Center", "Zhongsen Mingyuan Meiyijia", "Zhongsen Mingyuan Kindergarten", "Zhongsen Mingyuan Commercial Street", "Zhongsen Mingyuan Community", "Zhongsen Mingyuan Construction Site", "Zhongsen Mingyuan", "Zhongsen Mingyuan", and "West Gate of Zhongsen Mingyuan" into corresponding word vectors. Specifically, train the word vectors corresponding to each character in the candidate words, and then determine the word vector corresponding to the candidate word based on the word vectors of each character in the candidate word.

[0158] Specifically, a neural network language model can be trained with a large-scale Chinese corpus to obtain a pre-trained word vector model, which is used to convert text into word vectors. For example, use a large-scale Chinese corpus to pre-train the open-source Bert model. Input each character in "Zhongsen Mingyuan Community" into the open-source Bert model respectively to obtain the word vector representations corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", and "Qu". Then, based on the word vector representations of each character in the candidate word, obtain the word vector of each candidate word. For example, the word vector can be obtained by adding the word vectors. For example, for the candidate word "Zhongsen Mingyuan Community", the word vectors corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", and "Qu" are obtained by training with the word vector model respectively, and the word vectors are added together to accurately and quickly obtain the word vector corresponding to the candidate word "Zhongsen Mingyuan Community". For more complex word vectors, the candidate word corresponding word vector can be obtained by using a Recurrent Neural Network model. Specifically, for the candidate word "Zhongsen Mingyuan Community", the word vectors corresponding to "Zhong", "Sen", "Ming", "Yuan", "Xiao", and "Qu" are obtained by training with the word vector model respectively, and each word vector is input into the Recurrent Neural Network model for training to obtain the word vector corresponding to "Zhongsen Mingyuan Community".

[0159] The target analysis unit 133 is used to analyze each word vector in the word vector set based on a semantic network model including an attention mechanism to obtain a target multi-dimensional vector. Among them, the dimension value under each dimension in the target multi-dimensional vector is used to quantify the matching degree between the candidate word and the target address information.

[0160] Specifically, the semantic network model is also called the hierarchical network model of semantic memory. This model believes that the basic unit of semantic memory in long-term memory is the concept. Relevant concepts form a hierarchical network system according to the logical superior-subordinate relationship. A complex hierarchical network is formed by the subordinate relationship between concepts and the relationship between concepts and features. Each level of the model stores the unique features of the concepts at that level, while the common features shared by the concepts at the same level are stored at the upper-level concept level.

[0161] For example, the semantic network model including the attention mechanism analyzes the word vectors of 10 candidate words "Zhongsen Mingyuan Property Center", "Zhongsen Mingyuan Meiyijia", "Zhongsen Mingyuan Kindergarten", "Zhongsen Mingyuan Commercial Street", "Zhongsen Mingyuan Garden", "Zhongsen Mingyuan Community", "Zhongsen Mingyuan Construction Site", "Zhongsen Mingyuan", "Zhongsen Mingyuan", and "Zhongsen Mingyuan West Gate", and outputs target multi-dimensional vectors with dimension values ​​of 0.003, 0.044, 0.12, 0.045, 0.152, 0.335, 0.067, 0.138, 0.092, and 0.004 respectively.

[0162] It should be noted that different semantic network models have different hierarchical structures and concept processing mechanisms. The embodiment of the present application analyzes each word vector in the word vector set based on the attention mechanism, extracts the characteristics of the word vector, and determines the weight information of each word vector. The weight information is used to quantify the influence of the candidate word on the recognition of the central word. The multidimensional target vector is determined by the weight information of each word vector, and each dimension value of the multidimensional target vector is used to characterize the matching degree between each candidate word and the target address information.

[0163] The target determination unit 134 is used to determine the candidate word corresponding to the maximum dimension value in the target multi-dimensional vector as the central word among the multiple different candidate words.

[0164] It should be noted that since the dimension values ​​corresponding to each dimension in the target multidimensional vector are used to represent the weight of the candidate word's contribution to identifying the target meaning, the candidate word corresponding to the maximum dimension value is the candidate word that best matches the target meaning. Therefore, the candidate word corresponding to the maximum dimension value is determined as the central word among multiple different candidate words, which can more accurately represent the target address information.

[0165] Example 6

[0166] The embodiment of the present invention can provide a computer terminal, which can be any computer terminal device in a computer terminal group. Optionally, in this embodiment, the computer terminal can also be replaced by a terminal device such as a mobile terminal.

[0167] Optionally, in this embodiment, the computer terminal may be located in at least one network device among a plurality of network devices of the computer network.

[0168] In this embodiment, the above-mentioned computer terminal can execute the program code of the following steps in the data processing method of the application: obtaining word vectors corresponding to multiple different candidate words respectively, and obtaining a word vector set, wherein the multiple different candidate words are used to represent the same target meaning; based on a semantic network model including an attention mechanism, analyzing each word vector in the word vector set to obtain a target multidimensional vector, wherein the dimension value under each dimension in the target multidimensional vector is used to quantify the matching degree between the candidate word and the target meaning; determining the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among multiple different candidate words.

[0169] Optionally, Fig.14 is a structural block diagram of a computer terminal according to an embodiment of the present invention. Fig.14 As shown, the computer terminal A may include: one or more (only one is shown in the figure) processors and a memory.

[0170] Among them, the memory can be used to store software programs and modules, such as program instructions / modules corresponding to the data processing method and device in the embodiment of the present invention. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, that is, realizing the above-mentioned data processing method. The memory may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the terminal A via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0171] The processor can call the information and application programs stored in the memory through the transmission device to perform the following steps: obtain word vectors corresponding to multiple different candidate words respectively, and obtain a word vector set, wherein the multiple different candidate words are used to represent the same target meaning; based on a semantic network model including an attention mechanism, analyze each word vector in the word vector set to obtain a target multi-dimensional vector, wherein the dimension value under each dimension in the target multi-dimensional vector is used to quantify the matching degree between the candidate word and the target meaning; determine the candidate word corresponding to the maximum dimension value in the target multi-dimensional vector as the central word among the multiple different candidate words.

[0172] Optionally, the processor may also execute program code of the following steps: obtaining word vectors corresponding to a plurality of different candidate words respectively includes: training word vectors corresponding to each word in the candidate words; and determining the word vector corresponding to the candidate word based on each word vector in the candidate words.

[0173] Optionally, the processor may also execute the following steps of program code: determining the word vector corresponding to the candidate word based on the individual character vectors in the candidate word includes at least one of the following methods: adding the individual character vectors in the candidate word to obtain the word vector corresponding to the candidate word; or, using the individual character vectors in the candidate word as input to a target model, and training the word vector corresponding to the candidate word through the target model, wherein the target model includes at least a recurrent neural network model.

[0174] Optionally, the processor may also execute the program code of the following steps: before analyzing each word vector in the word vector set based on a semantic network model including an attention mechanism, determine whether the number of word vectors in the word vector set reaches a first preset number; when the number of word vectors in the word vector set is less than the first preset number, add a second preset number of zero vectors to the word vector set to obtain an updated word vector set, and analyze each word vector in the updated word vector set.

[0175] Optionally, the processor may also execute program code of the following steps: the semantic network model includes an input layer, an attention layer, a fully connected layer and an output layer. Based on the semantic network model including the attention mechanism, each word vector in the word vector set is analyzed to obtain a target multidimensional vector including: converting the word vectors in the word vector set into multiple attention weight information through the attention mechanism in the attention layer, wherein the attention weight information is used to characterize the weight of the contribution of the candidate word to the recognition of the target meaning; extracting and integrating multiple attention weight information in the attention layer through the fully connected layer; and converting the integrated multiple attention weight information into a target multidimensional vector through the output layer.

[0176] Optionally, the processor may also execute the program code of the following steps: converting the integrated multiple attention weight information into a target multidimensional vector through the output layer, including: normalizing the integrated multiple attention weight information through the normalized exponential function in the output layer to obtain multiple probability values, wherein the sum of the multiple probability values ​​is 1; determining the target multidimensional vector based on the multiple probability values, wherein each dimension in the target multidimensional vector corresponds to a probability value.

[0177] Optionally, the processor may also execute the program code of the following steps: after determining the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among multiple different candidate words, judging whether the meaning expressed by the central word matches the target meaning; if the meaning expressed by the central word does not match the target meaning, adjusting the semantic network model, wherein the semantic network model is adjusted in at least one of the following ways: adjusting the hierarchical matrix in each level of the semantic network model, adding multiple attention layers to the semantic network model, and adding multiple fully connected layers to the semantic network model.

[0178] By adopting an embodiment of the present invention, a data processing method is provided, which converts multiple candidate words of the target meaning into a word vector set, and then analyzes the word vector set through a semantic network model to obtain a target multidimensional vector for quantifying the matching degree between the candidate words and the target meaning, and determines the candidate word corresponding to the maximum dimensional value as the central word among multiple different candidate words, thereby achieving the purpose of obtaining a candidate word that accurately represents the target meaning from multiple synonyms, and further solving the technical problem that when a central word is obtained from multiple synonyms to represent the target meaning, the obtained central word is difficult to accurately represent the target meaning.

[0179] It can be understood by those skilled in the art that Fig.14 The structure shown is for illustration only, and the computer terminal may also be a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a handheld computer, a mobile Internet device (Mobile Internet Devices, MID), a PAD, or other terminal devices. Fig.14 The structure of the electronic device is not limited. For example, the computer terminal 10 may also include Fig.14 More or fewer components (such as network interfaces, display devices, etc.) shown in, or having Fig.14 Different configurations are shown.

[0180] A person of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing the hardware related to the terminal device through a program, and the program can be stored in a computer-readable storage medium, and the storage medium may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0181] Example 7

[0182] The embodiment of the present invention further provides a storage medium. Optionally, in this embodiment, the storage medium can be used to store the program code executed by the data processing method provided in the first embodiment.

[0183] Optionally, in this embodiment, the above storage medium may be located in any computer terminal in a computer terminal group in a computer network, or in any mobile terminal in a mobile terminal group.

[0184] Optionally, in this embodiment, the storage medium is configured to store program code for executing the following steps: obtaining word vectors corresponding to multiple different candidate words, respectively, to obtain a word vector set, wherein the multiple different candidate words are used to represent the same target meaning; based on a semantic network model including an attention mechanism, analyzing each word vector in the word vector set to obtain a target multidimensional vector, wherein the dimension value under each dimension in the target multidimensional vector is used to quantify the degree of match between the candidate word and the target meaning; determining the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among the multiple different candidate words.

[0185] Obtaining word vectors corresponding to a plurality of different candidate words respectively includes: training a word vector corresponding to each word in the candidate word; and determining a word vector corresponding to the candidate word based on each word vector in the candidate word.

[0186] Determining the word vector corresponding to the candidate word based on the individual word vectors in the candidate word includes at least one of the following methods: adding the individual word vectors in the candidate word to obtain the word vector corresponding to the candidate word; or, using the individual word vectors in the candidate word as input to a target model, and training the word vector corresponding to the candidate word through the target model, wherein the target model includes at least a recurrent neural network model.

[0187] Before analyzing each word vector in a word vector set based on a semantic network model including an attention mechanism, determine whether the number of word vectors in the word vector set reaches a first preset number; when the number of word vectors in the word vector set is less than the first preset number, add a second preset number of zero vectors to the word vector set to obtain an updated word vector set, and analyze each word vector in the updated word vector set.

[0188] The semantic network model includes an input layer, an attention layer, a fully connected layer and an output layer. Based on the semantic network model including the attention mechanism, each word vector in the word vector set is analyzed to obtain a target multidimensional vector, including: converting the word vectors in the word vector set into multiple attention weight information through the attention mechanism in the attention layer, wherein the attention weight information is used to characterize the weight of the contribution of the candidate word to the recognition of the target meaning; extracting and integrating the multiple attention weight information in the attention layer through the fully connected layer; and converting the integrated multiple attention weight information into the target multidimensional vector through the output layer.

[0189] Converting the integrated multiple attention weight information into a target multidimensional vector through the output layer includes: normalizing the integrated multiple attention weight information through a normalized exponential function in the output layer to obtain multiple probability values, wherein the sum of the multiple probability values ​​is 1; determining the target multidimensional vector based on the multiple probability values, wherein each dimension in the target multidimensional vector corresponds to a probability value.

[0190] After determining the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among multiple different candidate words, determine whether the meaning expressed by the central word matches the target meaning; when the meaning expressed by the central word does not match the target meaning, adjust the semantic network model, wherein the semantic network model is adjusted in at least one of the following ways: adjusting the hierarchical matrix in each level of the semantic network model, adding multiple attention layers to the semantic network model, and adding multiple fully connected layers to the semantic network model.

[0191] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0192] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0193] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0194] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0195] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0196] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0197] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A data processing method, wherein: Obtain word vectors corresponding to a plurality of different candidate words respectively to obtain a word vector set, wherein the plurality of different candidate words are used to represent the same target meaning; Based on a semantic network model including an attention mechanism, each word vector in the word vector set is analyzed to obtain a target multidimensional vector, wherein the dimension value under each dimension in the target multidimensional vector is used to quantify the matching degree between the candidate word and the target meaning; The candidate word corresponding to the maximum dimension value in the target multidimensional vector is determined as the central word among the multiple different candidate words.

2. The data processing method according to claim 1, wherein: Obtaining word vectors corresponding to multiple different candidate words includes: Train the word vector corresponding to each word in the candidate word; The word vector corresponding to the candidate word is determined based on each character vector in the candidate word.

3. The data processing method according to claim 2, wherein: Determining the word vector corresponding to the candidate word based on each word vector in the candidate word includes at least one of the following methods: Add up the word vectors of each candidate word to get the word vector corresponding to the candidate word; or, Each character vector in the candidate word is used as an input of a target model, and a word vector corresponding to the candidate word is trained through the target model, wherein the target model at least includes a recurrent neural network model.

4. The data processing method according to claim 1, wherein: The method further comprises: Before analyzing each word vector in the word vector set based on the semantic network model including the attention mechanism, determining whether the number of word vectors in the word vector set reaches a first preset number; When the number of word vectors in the word vector set is less than the first preset number, a second preset number of zero vectors are added to the word vector set to obtain an updated word vector set, and each word vector in the updated word vector set is analyzed.

5. The data processing method according to claim 1, wherein: The semantic network model includes an input layer, an attention layer, a fully connected layer and an output layer. Based on the semantic network model including the attention mechanism, each word vector in the word vector set is analyzed to obtain a target multidimensional vector including: Converting the word vectors in the word vector set into a plurality of attention weight information through the attention mechanism in the attention layer, wherein the attention weight information is used to represent the weight of the contribution of the candidate word to the recognition of the target meaning; Extracting and integrating the multiple attention weight information in the attention layer through the fully connected layer; The integrated multiple attention weight information is converted into the target multi-dimensional vector through the output layer.

6. The data processing method according to claim 5, wherein: Converting the integrated multiple attention weight information into the target multi-dimensional vector through the output layer includes: Normalizing the integrated multiple attention weight information by a normalized exponential function in the output layer to obtain multiple probability values, wherein the sum of the multiple probability values ​​is 1; The target multi-dimensional vector is determined based on the multiple probability values, wherein each dimension in the target multi-dimensional vector corresponds to a probability value.

7. The data processing method according to claim 5, wherein: The method further comprises: After determining the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among the multiple different candidate words, determining whether the meaning expressed by the central word matches the target meaning; When the meaning expressed by the central word does not match the target meaning, the semantic network model is adjusted, wherein the semantic network model is adjusted in at least one of the following ways: adjusting the hierarchical matrix in each level of the semantic network model, adding multiple attention layers to the semantic network model, and adding multiple fully connected layers to the semantic network model.

8. A method for processing address information, wherein: Acquire multiple different candidate words input during the voice navigation process, wherein the multiple different candidate words are used to represent target address information, and the multiple different candidate words are used to represent the same target meaning; Obtain word vectors corresponding to the multiple different candidate words respectively to obtain a word vector set; Based on a semantic network model including an attention mechanism, each word vector in the word vector set is analyzed to obtain a target multi-dimensional vector, wherein the dimension value under each dimension in the target multi-dimensional vector is used to quantify the matching degree between the candidate word and the target address information; The candidate word corresponding to the maximum dimension value in the target multidimensional vector is determined as the central word among the multiple different candidate words.

9. The address information processing method according to claim 8, wherein the semantic network model comprises: The input layer is used to convert multiple candidate words representing the target address information into word vectors respectively; An attention layer, the attention layer is connected to the input layer, and the attention layer is used to extract attention weight information between each word vector, wherein the attention weight information is used to represent the weight of the contribution of the candidate word to the identification of the target address information; A fully connected layer, the fully connected layer is connected to the attention layer, and the fully connected layer is used to extract and integrate multiple attention weight information in the attention layer; An output layer, the output layer is connected to the fully connected layer, and is used to convert the integrated multiple attention weight information into a target multi-dimensional vector, wherein the dimension value under each dimension in the target multi-dimensional vector is used to quantify the matching degree between the candidate word and the target address information.

10. A data processing device, wherein: An acquisition unit, used to acquire word vectors corresponding to a plurality of different candidate words, respectively, to obtain a word vector set, wherein the plurality of different candidate words are used to represent the same target meaning; An analysis unit, configured to analyze each word vector in the word vector set based on a semantic network model including an attention mechanism to obtain a target multidimensional vector, wherein a dimension value under each dimension in the target multidimensional vector is used to quantify a matching degree between a candidate word and the target meaning; A determination unit is used to determine the candidate word corresponding to the maximum dimension value in the target multidimensional vector as the central word among the multiple different candidate words.

11. A storage medium, wherein: The storage medium includes a stored program, wherein when the program is executed, the device where the storage medium is located is controlled to execute the data processing method according to any one of claims 1 to 7.

Citation Information

Patent Citations

  • Individualized knowledge active pushing method based on task decomposition

    CN103995858A

  • Enhanced question answering interaction system and method for universal mapping knowledge domain on the basis of deep learning

    CN108509519A

  • A single-semantic supervised word vector training method and device

    CN109271633A

  • Method and system for automatically extracting relations between concepts included in text

    US20080275694A1

  • PORTALS and DOORS for the Semantic Web and Grid

    US20080313229A1