A method, device, system and storage medium for verifying the address of a real estate mortgage

Abstract

The present invention provides a method, apparatus, system and storage medium for verifying the address of real estate mortgage. The method includes: obtaining the information text of the real estate mortgage address; using a fine-grained address element parsing model to verify the address of the information text to obtain an address verification result; and confirming the missing key address elements in the information text according to the address verification result. After obtaining the information text in the real estate mortgage address, the present invention uses a pre-trained fine-grained address element parsing model to verify the address of the information text, so that the missing key address elements in the information text can be further confirmed according to the address verification result. The method provided by the present invention can accurately verify whether the key address elements are missing in the information text of the mortgaged real estate address filled in by the customer when handling the real estate mortgage loan of the bank, and can meet the requirements of business and supervision at the same time, providing convenience for the real estate mortgage business of the bank.

Description

Technical Field

[0001] The present invention relates to the financial field and the verification technology field of real estate mortgage address information. More specifically, it relates to a method, device, system and storage medium for verifying the real estate mortgage address. Background Art

[0002] In the business scenario where a customer applies for a bank real estate mortgage loan, the customer needs to fill in the specific address of the mortgaged real estate. And the Banking and Insurance Regulatory Commission requires that when banking business personnel handle real estate mortgage loans for customers, they need to inform the customers that the address of the mortgaged real estate they fill in should be accurate to the house number and key address elements cannot be missing.

[0003] However, in actual business, when banking business personnel handle real estate mortgage loans for customers, they often forget or neglect to remind the customers to accurately fill in the address information of the mortgaged real estate to the house number. Therefore, for many customers applying for bank real estate mortgage loans, the address information of the mortgaged real estate they fill in often lacks key address elements, such as building name, community name, building number, unit number, floor number or house number, etc.

[0004] In short, in the current business scenario where customers apply for bank real estate mortgage loans, it mainly relies on banking business personnel to remind customers to fill in the address information of their mortgaged real estate completely and accurately to the house number, and to verify it manually. However, this pure manual reminder and verification method has poor reliability and cannot guarantee that every customer applying for a bank real estate mortgage loan fills in the address information of their mortgaged real estate completely and accurately to the house number. Summary of the Invention

[0005] In view of this, the present invention provides a method for verifying a real estate mortgage address, including:

[0006] Obtain the information text of the real estate mortgage address;

[0007] Use a fine-grained address element parsing model to perform address verification on the information text to obtain an address verification result;

[0008] Confirm the missing key address elements in the information text according to the address verification result.

[0009] Preferably, the fine-grained address element parsing model includes a semantic fusion layer, an attention encoding calculation layer and a classification layer;

[0010] The step of using a fine-grained address element parsing model to perform address verification on the information text to obtain an address verification result includes:

[0011] Convert the information text into a semantic meaning encoding vector corresponding to the sub-characters in the information text through the semantic meaning fusion layer;

[0012] Input the semantic meaning encoding vector into the attention encoding calculation layer, perform attention encoding calculation between the semantic meaning encoding vectors, and obtain an attention encoding calculation result;

[0013] Use the classification layer to perform classification calculation on the attention encoding calculation result to obtain the address verification result.

[0014] Preferably, the step of converting the information text into a semantic meaning encoding vector corresponding to the sub-characters in the information text through the semantic meaning fusion layer includes:

[0015] Perform word segmentation on the information text using the Word-Bert Chinese semantic meaning pre-training model to obtain multiple sub-characters corresponding to the information text;

[0016] Input all the sub-characters into the semantic meaning fusion layer of the Word-Bert Chinese semantic meaning pre-training model to perform semantic meaning embedding encoding, and convert the sub-characters into semantic meaning encoding vectors.

[0017] Preferably, the semantic meaning encoding vector e i is calculated as follows:

[0018] Tokenization(t)∝(x 1 , x 2 ,..., x i ...x n ) 0≤i≤n;

[0019] e i =Word-Bert Embedding(x i );

[0020] where t represents the information text of the real estate mortgage address; x i represents all the sub-characters obtained after the information text is segmented; Word-Bert Embedding(x i ) represents performing semantic meaning embedding encoding on the sub-characters through the semantic meaning fusion layer of the Word-Bert Chinese semantic meaning pre-training model; e i is the semantic meaning encoding vector after the conversion of the sub-characters.

[0021] Preferably, the attention encoding calculation layer includes a hybrid attention mechanism; the hybrid attention mechanism includes a global attention mechanism and a cross-character dynamic convolution local attention mechanism;

[0022] Inputting the word meaning encoding vector into the attention encoding calculation layer to perform attention encoding calculation between the word meaning encoding vectors, and obtaining an attention encoding calculation result, including:

[0023] According to the global attention mechanism, performing global attention calculation on the word meaning encoding vector to obtain a global attention result; and according to the cross-character dynamic convolution local attention mechanism, performing local attention calculation on the word meaning encoding vector to obtain a cross-character dynamic convolution local attention result;

[0024] According to the hybrid attention mechanism, splicing the global attention result and the cross-character dynamic convolution local attention result to obtain a hybrid attention aggregation encoding vector, and using the hybrid attention aggregation encoding vector as the attention encoding calculation result.

[0025] Preferably, in the process of performing global attention calculation on the word meaning encoding vector according to the global attention mechanism to obtain a global attention result, the calculation method of the global attention result is:

[0026]

[0027]

[0028] where Q, K, and V respectively represent the query vector, key vector, and value vector obtained after mapping transformation calculation for the word meaning encoding vector e i ; Self - Attn((K, V), q i ) represents the process of calculating the self - attention matrix using the query vector Q, the key vector K, and the value vector V, and calculating the global attention result of character i; is the global attention result of character i.

[0029] Preferably, in the process of performing local attention calculation on the word meaning encoding vector according to the cross - character dynamic convolution local attention mechanism to obtain a cross - character dynamic convolution local attention result, the calculation method of the cross - character dynamic convolution local attention result is:

[0030]

[0031]

[0032]

[0033]

[0034] where Q, V respectively represent the word meaning encoding vector ei The query vector and value vector obtained after mapping transformation calculation; K span The cross-character local context key-value vector representing character i; LConv(e, W, i) is the calculation process of the convolution kernel in the dynamic convolution attention; SDConv(Q, K span , V; W f , i) represents the process of calculating the local attention dependency matrix using the query vector Q, the value vector V, and the cross-character local context key-value vector K span and calculating the cross-character dynamic convolution local attention result of character i; is the cross-character dynamic convolution local attention result of character i.

[0035] Preferably, according to the hybrid attention mechanism, the global attention result and the cross-character dynamic convolution local attention result are concatenated to obtain a hybrid attention aggregation coding vector, and the hybrid attention aggregation coding vector is used as the attention coding calculation result. The calculation method of the hybrid attention aggregation coding vector is as follows:

[0036]

[0037] where Cat represents the concatenation operation; is the hybrid attention aggregation coding vector of character i obtained by concatenating the global attention result and the cross-character dynamic convolution local attention result.

[0038] Preferably, the classification layer is used to perform classification calculation on the attention coding calculation result to obtain the address verification result, including:

[0039] Using the long short-term memory bidirectional recurrent neural network model of the classification layer to perform forward recurrent coding calculation and backward recurrent coding calculation on the attention coding calculation result to obtain the forward array result and backward array result of multiple hidden layers for the attention coding calculation result;

[0040] Extracting the forward hidden state coding and backward hidden state coding of the sub-character sequence in the last hidden layer of the forward array result and the backward array result; and horizontally merging the forward hidden state coding and the backward hidden state coding to obtain a merged hidden state coding;

[0041] Inputting the merged hidden state coding into the classification sub-network of the classification layer for classification to obtain the address verification result.

[0042] Preferably, when using the classification layer to perform classification calculation on the attention encoding calculation result to obtain the address verification result, the calculation method of the address verification result is as follows:

[0043]

[0044]

[0045]

[0046]

[0047] Wherein, Bi-LSTM represents the long short-term memory bidirectional recurrent neural network model; represents the forward hidden state encoding of the sub-character sequence; represents the backward hidden state encoding of the sub-character sequence; is the hybrid attention aggregation encoding vector of the i-th sub-character in the information text After being calculated by the long short-term memory bidirectional recurrent neural network model Bi-LSTM, the merged hidden state encoding in the last hidden layer; Classification Head is the classification sub-network; O i represents the address verification result of the classification layer for the sub-address elements classification of all sub-character sequences in the information text.

[0048] In addition, to solve the above problems, the present invention also provides a verification device for real estate mortgage addresses, including:

[0049] An acquisition module, configured to acquire the information text of the real estate mortgage address;

[0050] A verification module, configured to perform address verification on the information text by using a fine-grained address element parsing model to obtain an address verification result;

[0051] A confirmation module, configured to confirm the missing key address elements in the information text according to the address verification result.

[0052] In addition, to solve the above problems, the present invention also provides a verification system for real estate mortgage addresses, including a memory and a processor, and the real estate mortgage address verification program. The processor runs the real estate mortgage address verification program so that the real estate mortgage address verification system executes the real estate mortgage address verification method as described above.

[0053] In addition, to solve the above problems, the present invention further provides a computer-readable storage medium, on which a verification program for the real estate mortgage address is stored. When the verification program for the real estate mortgage address is executed by a processor, the verification method for the real estate mortgage address as described above is implemented.

[0054] The present invention provides a method, device, system and storage medium for verifying a real estate mortgage address. The method includes: obtaining an information text of the real estate mortgage address; using a fine-grained address element parsing model to verify the address of the information text to obtain an address verification result; and confirming the missing key address elements in the information text according to the address verification result. After obtaining the information text in the real estate mortgage address, the present invention uses a pre-trained fine-grained address element parsing model to verify the address of the information text, so as to further confirm the missing key address elements in the information text according to the address verification result. The method provided by the present invention can accurately verify whether the key address elements are missing in the information text of the mortgaged real estate address filled in by the customer when handling the real estate mortgage loan of the bank, and can meet the requirements of business and supervision, providing convenience for the real estate mortgage business of the bank. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic structural diagram of the hardware operating environment involved in the embodiment of the method for verifying a real estate mortgage address of the present invention;

[0056] Figure 2 It is a schematic flowchart of the first embodiment of the method for verifying a real estate mortgage address of the present invention;

[0057] Figure 3 It is a schematic flowchart of the refinement of step S200 in the first embodiment of the method for verifying a real estate mortgage address of the present invention;

[0058] Figure 4 It is a schematic flowchart of adding a judgment step to the confirmation step in the first embodiment of the method for verifying a real estate mortgage address of the present invention;

[0059] Figure 5 It is a schematic flowchart of the refinement of step S210 in the second embodiment of the method for verifying a real estate mortgage address of the present invention;

[0060] Figure 6 It is a schematic flowchart of the refinement of step S220 in the third embodiment of the method for verifying a real estate mortgage address of the present invention;

[0061] Figure 7 It is a schematic flowchart of the refinement of step S230 in the fourth embodiment of the method for verifying a real estate mortgage address of the present invention;

[0062] Figure 8It is a schematic diagram of module connection of the verification device for the real estate mortgage address of the present invention.

[0063] The realization of the object, functional features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with embodiments. Specific embodiments

[0064] The embodiments of the present invention will be described in detail below, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions throughout.

[0065] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality" means two or more unless otherwise specifically defined.

[0066] In the present invention, unless otherwise clearly defined and limited, the terms "installation", "connection", "connection", "fixation" and other terms should be understood in a broad sense. For example, it may be a fixed connection, a detachable connection or integrated; it may be a mechanical connection or an electrical connection; it may be directly connected or indirectly connected through an intermediate medium, and it may be the internal communication of two elements or the interaction relationship between two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0067] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0068] As Figure 1 shown, it is a schematic diagram of the structure of the hardware operating environment of the terminal involved in the embodiment of the present invention.

[0069] The verification system for the real estate mortgage address in the embodiments of the present invention can be a PC, or a mobile terminal device such as a smart phone, a tablet computer, or a portable computer, etc. The verification system for the real estate mortgage address can include: a processor 1001, such as a CPU, a network interface 1004, a user interface 1003, a memory 1005, and a communication bus 1002. Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 can include a display screen, an input unit such as a keyboard, a remote control. Optionally, the user interface 1003 can also include a standard wired interface and a wireless interface. The network interface 1004 can optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 can be a high-speed RAM memory, or a stable memory, such as a disk memory. Optionally, the memory 1005 can also be a storage device independent of the aforementioned processor 1001. Optionally, the verification system for the real estate mortgage address can also include an RF (Radio Frequency) circuit, an audio circuit, a WiFi module, etc. In addition, the verification system for the real estate mortgage address can also be configured with other sensors such as a gyroscope, a barometer, a hygrometer, a thermometer, an infrared sensor, etc., which will not be elaborated here.

[0070] Those skilled in the art can understand that Figure 1 the verification system for the real estate mortgage address shown in Figure 1 does not constitute a limitation thereto, and may include more or fewer components than shown in the figure, or combine some components, or arrange different components. As

[0071] Embodiment 1:

[0072] Referring to Figure 2 a verification method for a real estate mortgage address provided by the first embodiment of the present invention includes:

[0073] Step S100, obtaining the information text of the real estate mortgage address.

[0074] As described above, the information text is an information string containing an address input by the user, which may include, but is not limited to, the names of administrative divisions at all levels, and information detailed to the location number that can represent its uniqueness. For example, from "country, province, city, county, township, village, brigade" at the top to "area, community, building, group, committee, floor, house number" at the bottom.

[0075] When users input information, since it is a bank real estate mortgage business, it is not excluded that some users input incomplete information, or due to misunderstandings of detailed addresses, some key information is missed when inputting information. In this case, bank business personnel need to give further manual reminders and verifications. Since manual reminders and verifications are somewhat subjective, it may lead to situations such as incomplete filling of business information, inability to continue with the next node of business processing, and illegal handling of business.

[0076] Therefore, in this embodiment, the information text filled in by the user is automatically verified to solve this problem. First, the information text is obtained. The obtaining step can be passive reception for automatic verification or active acquisition for automatic verification.

[0077] Step S200: Use the fine-grained address element parsing model to perform address verification on the information text to obtain an address verification result.

[0078] Furthermore, the fine-grained address element parsing model includes a semantic fusion layer, an attention encoding calculation layer, and a classification layer;

[0079] As described above, different functional layers represent different functional models, that is, corresponding calculations are performed using the corresponding models in this functional layer to achieve the corresponding functions.

[0080] As described above, the fine-grained address element parsing model is the "Semantic Fusion Embed-ConvBert-LSTM Deep Model", which can accurately verify whether a certain address element is missing in the mortgaged real estate address information text of the customer, such as building name, community name, building number, unit number, floor number, or house number, etc.

[0081] The Semantic Fusion Embed-ConvBert-LSTM Deep Model mainly includes three sub-network models: "Semantic Fusion Layer Embedding", "Attention Encoding Calculation Layer ConvBert", and "Classification Layer LSTM". Correspondingly, in the fine-grained address element parsing model, there are a semantic fusion layer, an attention encoding calculation layer, and a classification layer.

[0082] Reference Figure 3 As described above, in step S200, using the fine-grained address element parsing model to perform address verification on the information text to obtain an address verification result includes:

[0083] Step S210: Convert the information text into a semantic encoding vector corresponding to the sub-characters in the information text through the semantic fusion layer;

[0084] "Lexical Meaning Fusion Layer Embedding" is mainly used to fuse the lexical meaning information in the information text. "Lexical Meaning Fusion Layer Embedding" is actually a "pre-trained lexical meaning embedding encoding layer", which is taken from the lexical meaning embedding encoding layer in the Word-Bert Chinese lexical meaning pre-training model after lexical meaning pre-training on a large number of Chinese texts.

[0085] At this time, the "pre-trained lexical meaning embedding encoding layer" (the trained lexical meaning fusion layer) contains the lexical meaning encodings learned during the lexical meaning pre-training of a large number of Chinese texts. It can convert the Chinese characters in the information text into lexical meaning encoding vectors, and the converted lexical meaning encoding vectors can then be input into the "Attention Encoding Calculation Layer ConvBert" for attention encoding calculation between the "sub-character lexical meaning encoding vectors".

[0086] Step S220: Input the lexical meaning encoding vector into the attention encoding calculation layer to perform attention encoding calculation between the lexical meaning encoding vectors, and obtain the attention encoding calculation result.

[0087] After the "Lexical Meaning Fusion Layer Embedding" converts the mortgaged real estate address text into individual sub-character lexical meaning encoding vectors, all the lexical meaning encoding vectors are input into the "Attention Encoding Calculation Layer ConvBert" model for attention encoding calculation between the "sub-character lexical meaning encoding vectors", thereby obtaining the corresponding attention encoding calculation result.

[0088] Step S230: Use the classification layer to perform classification calculation on the attention encoding calculation result to obtain the address verification result.

[0089] After the "Attention Encoding Calculation Layer ConvBert" calculates the attention encoding calculation results of all sub-characters i in the mortgaged real estate address text, it is input into the "Classification Layer LSTM" model for fine-grained sub-address element classification calculation. Here, the "Classification Layer LSTM" model will classify and parse each sub-character i in the address text into one of the 17 different granularity sub-address element categories described above.

[0090] As mentioned above, the sub-address element categories may include, but are not limited to, the 17 categories in Table 1 provided in this embodiment:

[0091] Table 1. 17 Kinds of Granularity Sub-Address Elements (Arranged from Largest to Smallest Granularity Level)

[0092]

[0093]

[0094] Step S300: Confirm the missing key address elements in the information text according to the address verification result.

[0095] In addition, after step S300, it may further include judging whether there are missing key address elements in the information text according to the address verification result;

[0096] If so, return to execute the "obtain the information text of the real estate mortgage address" and generate an alarm message to remind the business personnel to perform operations;

[0097] If not, it is determined that the information text passes the verification.

[0098] Here, a certain amount of information can be preset as key address elements first, and then in the judgment, if the missing information is a key address element, an alarm message is generated, and the process returns to execute the operation of obtaining the information text of the real estate mortgage address until there are no missing key address elements.

[0099] In the business scenario of a customer applying for a bank real estate mortgage loan, a "fine-grained address element parsing model" is required. When the customer fills in the address information of the mortgaged real estate, this model can automatically verify the address information filled in by the customer and judge whether there are missing key address elements, such as building name, community name, building number, unit number, floor number or house number, etc. The "fine-grained address element parsing model" is a pre-trained neural network model, which can not only reduce the workload of bank business personnel, but also meet the compliance supervision requirements of the China Banking and Insurance Regulatory Commission.

[0100] The overall process reference Figure 2 and the brief process Figure 4 In this embodiment, after obtaining the information text in the real estate mortgage address, the pre-trained fine-grained address element parsing model is used to verify the address of the information text, so that the missing key address elements in the information text can be further confirmed according to the address verification result. The method provided in this embodiment can accurately verify whether there are missing key address elements in the information text of the mortgaged real estate address filled in by the customer when applying for a bank real estate mortgage loan, and can meet the requirements of business and supervision at the same time, providing convenience for the bank's real estate mortgage business.

[0101] Embodiment 2:

[0102] Refer to Figure 5 Based on the above Embodiment 1, the step S210 of converting the information text into a semantic encoding vector corresponding to the sub-characters in the information text through the semantic fusion layer in the second embodiment of the present invention includes:

[0103] Step S211, segment and split the information text using the Word-Bert Chinese word sense pre-training model to obtain multiple sub-characters corresponding to the information text;

[0104] As described above, the Word-Bert Chinese word sense pre-training model is a pre-trained recognition and splitting model. Using this model, the character strings in the information text are segmented and split into multiple sub-characters corresponding to the original information text.

[0105] Step S212, input all the sub-characters into the word sense fusion layer of the Word-Bert Chinese word sense pre-training model for word sense embedding encoding, and convert the sub-characters into word sense encoding vectors.

[0106] As described above, all the sub-characters obtained after segmentation will be input into the Word-Bert Chinese word sense pre-training model in the "Word Sense Fusion Layer Embedding" for word sense embedding encoding, and then all the sub-characters will be converted into word sense encoding vectors.

[0107] For the above steps S211 and S212, the word sense encoding vector e i The calculation method, that is, the calculation process, can be as follows:

[0108] Tokenization(t) ∝ (x 1 , x 2 ,..., x i ...x n ) 0 ≤ i ≤ n; (1)

[0109] e i = Word-Bert Embedding(x i ); (2)

[0110] Wherein, t represents the information text of the real estate mortgage address for fine-grained address element parsing and splitting; x i represents all the sub-characters obtained after the information text is segmented and split; Word-Bert Embedding(x i ) represents performing word sense embedding encoding on the sub-characters through the word sense fusion layer of the Word-Bert Chinese word sense pre-training model; e i is the word sense encoding vector after the conversion of the sub-characters.

[0111] In this embodiment, by inputting the information text into the Word-Bert Chinese word sense pre-training model, for the text string of the mortgage real estate address for fine-grained address element parsing, it can convert Chinese characters into word sense encoding vectors, so as to perform further attention encoding calculation layers.

[0112] Example 3:

[0113] Referring to Figure 6 , the third embodiment of the present invention provides a method for verifying the address of a real estate mortgage. Based on the above-mentioned Embodiment 1, the attention encoding calculation layer includes a hybrid attention mechanism; the hybrid attention mechanism includes a global attention mechanism and a cross-character dynamic convolution local attention mechanism;

[0114] The "attention encoding calculation layer ConvBert" model is essentially a pre-trained model based on the cross-character dynamic convolution attention mechanism. However, this model actually contains both the "global self-attention calculation mechanism self-attention" and the "cross-character dynamic convolution local attention mechanism span-based dynamic convolution attention", and the two can together form a hybrid attention mechanism (mixed attention mechanism).

[0115] In step S220, inputting the semantic encoding vector into the attention encoding calculation layer to perform attention encoding calculation between the semantic encoding vectors to obtain an attention encoding calculation result, including:

[0116] Step S221, according to the global attention mechanism, perform global attention calculation on the semantic encoding vector to obtain a global attention result; and, according to the cross-character dynamic convolution local attention mechanism, perform local attention calculation on the semantic encoding vector to obtain a cross-character dynamic convolution local attention result;

[0117] When all the semantic encoding vectors e of the sub-characters in the real estate mortgage address text i are input into the ConvBert model, the "global self-attention" will be calculated first for the semantic encoding vector to obtain a global attention result.

[0118] In step S221, according to the global attention mechanism, perform global attention calculation on the semantic encoding vector to obtain a global attention result. The calculation method of the global attention result is:

[0119]

[0120]

[0121] where Q, K, and V respectively represent the semantic encoding vector e iThe query vector, key vector, and value vector obtained after mapping transformation calculation; Self - Attn((K, V), q i ) represents the process of calculating the self - attention matrix using the query vector Q, the key vector K, and the value vector V, and calculating the global attention result for character i; is the global attention result for character i.

[0122] As described above, based on the semantic encoding vectors e of all sub - characters i , the "cross - character dynamic convolution local attention span - based dynamic convolution attention" will also be calculated simultaneously, which is the cross - character dynamic convolution local attention result.

[0123] In step S221, according to the cross - character dynamic convolution local attention mechanism, for the local attention calculation of the semantic encoding vector, the calculation method of the cross - character dynamic convolution local attention result is as follows:

[0124]

[0125]

[0126]

[0127]

[0128] Among them, Q and V respectively represent the query vector and value vector obtained after mapping transformation calculation of the semantic encoding vector e i ; K span represents the cross - character local context key - value vector of character i; LConv(e, W, i) is the calculation process of the convolution kernel in the dynamic convolution attention; SDConv(Q, K span , V; W f , i) represents the process of calculating the local attention dependence matrix using the query vector Q, the value vector V, and the cross - character local context key - value vector K span and calculating the cross - character dynamic convolution local attention result for character i; is the cross - character dynamic convolution local attention result for character i.

[0129] Step S222, according to the hybrid attention mechanism, splice the global attention result and the cross - character dynamic convolution local attention result to obtain a hybrid attention aggregation encoding vector, and use the hybrid attention aggregation encoding vector as the attention encoding calculation result.

[0130] As described above, under the hybrid attention mechanism, the calculation result of the "global self-attention calculation mechanism" (global attention result) and the calculation result of the "cross-character dynamic convolution local attention mechanism" (cross-character dynamic convolution local attention result) are concatenated to calculate the hybrid attention aggregation coding vector, which is the attention coding calculation result.

[0131] In step S222, according to the hybrid attention mechanism, the global attention result and the cross-character dynamic convolution local attention result are concatenated to obtain a hybrid attention aggregation coding vector, and the hybrid attention aggregation coding vector is used as the attention coding calculation result. The hybrid attention aggregation coding vector The calculation method is as follows:

[0132]

[0133] where Cat represents the concatenation operation; is the hybrid attention aggregation coding vector of character i obtained after concatenating the global attention result and the cross-character dynamic convolution local attention result.

[0134] In this embodiment, through the attention coding calculation of the ConvBert model in the attention coding calculation layer, the model can further fuse the semantic coding vectors e of all sub-characters in the mortgaged real estate address i At the same time, the model is enabled to have a stronger ability to understand, represent, and transform the sub-character semantic coding vectors, and finally the result of the address element parsing and splitting of the mortgaged real estate address text can be made more accurate.

[0135] Embodiment 4:

[0136] Referring to Figure 7 In the third embodiment of the present invention, a method for verifying a real estate mortgage address is provided. Based on the above Embodiment 1, in step S230, the classification layer is used to perform classification calculation on the attention coding calculation result to obtain the address verification result, including:

[0137] Step S231, using the long short-term memory bidirectional recurrent neural network model of the classification layer to perform forward recurrent coding calculation and backward recurrent coding calculation on the attention coding calculation result to obtain a forward array result and a backward array result of multiple hidden layers for the attention coding calculation result;

[0138] As described above, the "classification layer LSTM" model is also called the "long short-term memory bidirectional recurrent neural network".

[0139] As described above, first, using the long short-term memory bidirectional recurrent neural network of the "classification layer LSTM" model, the "hybrid attention aggregation encoding vectors" of all sub-characters i in the information text are subjected to forward and reverse recurrent encoding calculations to obtain a forward array result and a reverse array result respectively.

[0140] Among them, the forward array result and the reverse array result are two array matrices respectively.

[0141] Step S232: Extract the forward hidden state encoding and the reverse hidden state encoding of the sub-character sequence in the last hidden layer of the forward array result and the reverse array result respectively; and horizontally merge the forward hidden state encoding and the reverse hidden state encoding to obtain a merged hidden state encoding.

[0142] Horizontally merge the forward hidden state encoding and the reverse hidden state encoding of the sub-character sequence in the last hidden layer.

[0143] As described above, in the classification layer LSTM, the two array matrices of the forward array result and the reverse array result respectively represent matrices composed of multiple data hidden layers, and the forward hidden state encoding and the reverse hidden state encoding in the self-character sequence of the last hidden layer are taken for horizontal merging.

[0144] As described above, extract the forward hidden state encoding of the sub-character sequence in the last hidden layer of the forward array result respectively; and the reverse hidden state encoding of the sub-character sequence in the last hidden layer of the reverse array result.

[0145] Step S233: Input the merged hidden state encoding into the classification sub-network of the classification layer for classification to obtain the address verification result.

[0146] Finally, input the hidden state encoding obtained by horizontally merging the forward and reverse of the sub-character sequence into the classification sub-network in the "classification layer LSTM" model, so as to facilitate the classification of 17 different granularity sub-address elements as shown in Table 1.

[0147] Furthermore, in step S230, using the classification layer to perform classification calculation on the attention encoding calculation result to obtain the address verification result, the calculation method of the address verification result is:

[0148]

[0149]

[0150]

[0151]

[0152] Among them, Bi-LSTM represents the long short-term memory bidirectional recurrent neural network model; represents the forward hidden state encoding of the sub-character sequence; represents the backward hidden state encoding of the sub-character sequence; is the hybrid attention aggregation encoding vector of the sub-character i in the information text After being calculated by the long short-term memory bidirectional recurrent neural network model Bi-LSTM, the merged hidden state encoding in the last hidden layer; ClassificationHead is the classification sub-network; O i represents the address verification result of the classification layer for the sub-address elements classification of all sub-character sequences in the information text.

[0153] In summary, in this embodiment, after obtaining the information text in the real estate mortgage address, the pre-trained fine-grained address element parsing model is used to perform address verification on the information text, so that the missing key address elements in the information text can be further confirmed according to the address verification result. The method provided in this embodiment can accurately verify whether the key address elements are missing in the information text of the mortgage real estate address filled in by the customer when handling the bank real estate mortgage loan, and at the same time can meet the requirements of business and supervision, providing convenience for the bank real estate mortgage business.

[0154] The method provided in this embodiment "can parse and split 17 fine-grained sub-address elements such as those in Table 1 for the information text of the mortgage real estate address filled in by the customer when handling the bank real estate mortgage loan, and it can accurately judge whether the key address elements are missing in the information text of the customer's mortgage real estate address, such as building name, community name, building number, unit number, floor number or house number, etc.

[0155] Moreover, the method provided in this embodiment can automatically verify whether the key address elements in the information text of the mortgage real estate address filled in by the customer are complete when handling the bank real estate mortgage loan business. And it can reduce the workload of bank business personnel, and at the same time can meet the compliance supervision requirements of the China Banking and Insurance Regulatory Commission.

[0156] In addition, referring to Figure 8 , the present invention also provides a verification device for a real estate mortgage address, including:

[0157] An acquisition module 10, configured to acquire the information text of the real estate mortgage address;

[0158] A verification module 20, configured to perform address verification on the information text by using a fine-grained address element parsing model to obtain an address verification result;

[0159] A confirmation module 30 is configured to confirm the missing key address elements in the information text according to the address verification result.

[0160] In addition, the present invention further provides a verification system for real estate mortgage addresses, including a memory and a processor, and a verification program for real estate mortgage addresses. The processor runs the verification program for real estate mortgage addresses so that the verification system for real estate mortgage addresses executes the verification method for real estate mortgage addresses as described above.

[0161] In addition, the present invention further provides a computer-readable storage medium, on which a verification program for real estate mortgage addresses is stored. When the verification program for real estate mortgage addresses is executed by a processor, it implements the verification method for real estate mortgage addresses as described above.

[0162] 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.

[0163] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-described embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they 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, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium as described above (such as ROM / RAM, magnetic disk, optical disc), and includes several instructions for causing a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present invention. The above is only the preferred embodiment of the present invention, and does not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A method for verifying the address of real estate mortgage, characterized in that, it includes: Obtaining the information text of the real estate mortgage address; Using a fine-grained address element parsing model to verify the address of the information text to obtain an address verification result; The fine-grained address element parsing model includes a semantic fusion layer, an attention encoding calculation layer, and a classification layer; The using the fine-grained address element parsing model to verify the address of the information text to obtain an address verification result includes: converting the information text into a semantic encoding vector corresponding to the sub-characters in the information text through the semantic fusion layer; inputting the semantic encoding vector into the attention encoding calculation layer to perform attention encoding calculation between the semantic encoding vectors to obtain an attention encoding calculation result; using the classification layer to perform classification calculation on the attention encoding calculation result to obtain the address verification result; The attention encoding calculation layer includes a hybrid attention mechanism; the hybrid attention mechanism includes a global attention mechanism and a cross-character dynamic convolution local attention mechanism; The inputting the semantic encoding vector into the attention encoding calculation layer to perform attention encoding calculation between the semantic encoding vectors to obtain an attention encoding calculation result includes: According to the global attention mechanism, performing global attention calculation on the semantic encoding vector to obtain a global attention result; and, according to the cross-character dynamic convolution local attention mechanism, performing local attention calculation on the semantic encoding vector to obtain a cross-character dynamic convolution local attention result; According to the hybrid attention mechanism, splicing the global attention result and the cross-character dynamic convolution local attention result to obtain a hybrid attention aggregation encoding vector, and using the hybrid attention aggregation encoding vector as the attention encoding calculation result; The calculation method of the mixed attention aggregation coding vector is as follows: Among them, represents a splicing operation; is the character obtained after splicing the global attention result and the cross-character dynamic convolution local attention result of the hybrid attention aggregation coding vector; Confirming the missing key address elements in the information text according to the address verification result.

2. The method for verifying the address of real estate mortgage according to claim 1, characterized in that, The converting the information text into a semantic encoding vector corresponding to the sub-characters in the information text through the semantic fusion layer includes: Performing word segmentation on the information text using the Word-Bert Chinese semantic pre-training model to obtain multiple sub-characters corresponding to the information text; Inputting all the sub-characters into the semantic fusion layer of the Word-Bert Chinese semantic pre-training model to perform semantic embedding encoding, and converting the sub-characters into semantic encoding vectors.

3. The method for verifying the address of real estate mortgage according to claim 2, characterized in that, The semantic meaning encoding vector is calculated as follows: ； ； Among them, The information text representing the real estate mortgage address; All sub-characters obtained by segmenting and splitting the information text; Indicates that the word meaning embedding coding is performed on the sub-characters through the word meaning fusion layer of the Word-Bert Chinese word meaning pre-training model; Is the word meaning coding vector after conversion of the sub-characters.

4. The method for verifying the address of real estate mortgage according to claim 1, characterized in that, In the performing global attention calculation on the semantic encoding vector according to the global attention mechanism to obtain a global attention result, the calculation method of the global attention result is: ； ； Among them, , and respectively represent the query vector, key vector, and value vector obtained after mapping transformation calculation of the semantic coding vector ; represents the process of calculating the self-attention matrix using the query vector , the key vector and the value vector , and calculating the global attention result of the character ; is the global attention result of the character .

5. The method for verifying the address of real estate mortgage according to claim 1, characterized in that, According to the cross-character dynamic convolutional local attention mechanism, for the local attention calculation of the semantic encoding vector to obtain the cross-character dynamic convolutional local attention result, the calculation method of the cross-character dynamic convolutional local attention result is as follows: ； ； ； ； Among them, and respectively represent the query vector and the value vector obtained after the mapping transformation calculation of the semantic encoding vector ; represents the cross-character local context key-value vector of the character ; is the calculation process of the convolution kernel in the dynamic convolution attention; represents the process of calculating the local attention dependence matrix using the query vector , the value vector and the cross-character local context key-value vector , and calculating the cross-character dynamic convolution local attention result of the character ; is the cross-character dynamic convolution local attention result of the character .

6. The method for verifying the real estate mortgage address according to claim 1, characterized in that the use of the classification layer to perform classification calculation on the attention encoding calculation result to obtain the address verification result includes: using the long short-term memory bidirectional recurrent neural network model of the classification layer to perform forward recurrent encoding calculation and backward recurrent encoding calculation on the attention encoding calculation result to obtain a forward array result and a backward array result of multiple hidden layers for the attention encoding calculation result; extracting the forward hidden state encoding and the backward hidden state encoding of the sub-character sequence in each last hidden layer of the forward array result and the backward array result; and horizontally merging the forward hidden state encoding and the backward hidden state encoding to obtain a merged hidden state encoding; inputting the merged hidden state encoding into the classification sub-network of the classification layer for classification to obtain the address verification result.

7. The method for verifying the real estate mortgage address according to claim 6, characterized in that in the use of the classification layer to perform classification calculation on the attention encoding calculation result to obtain the address verification result, the calculation method of the address verification result is as follows: ； ； ； ； Among them, represents the long short-term memory bidirectional recurrent neural network model; represents the forward hidden state encoding of the sub-character sequence; represents the backward hidden state encoding of the sub-character sequence; is the hybrid attention aggregation encoding vector of the sub-character i in the information text after passing through the long short-term memory bidirectional recurrent neural network model the combined hidden state encoding in the last hidden layer after calculation; is the classification sub-network; represents the address verification result of the classification layer for the sub-address element classification of all sub-character sequences in the information text.

8. A device for verifying a real estate mortgage address, characterized in that it includes: an acquisition module for acquiring the information text of the real estate mortgage address; a verification module for using the fine-grained address element parsing model to perform address verification on the information text to obtain an address verification result; the fine-grained address element parsing model includes a semantic fusion layer, an attention encoding calculation layer, and a classification layer; the use of the fine-grained address element parsing model to perform address verification on the information text to obtain an address verification result includes: converting the information text into a semantic encoding vector corresponding to the sub-characters in the information text through the semantic fusion layer; inputting the semantic encoding vector into the attention encoding calculation layer to perform attention encoding calculation between the semantic encoding vectors to obtain an attention encoding calculation result; using the classification layer to perform classification calculation on the attention encoding calculation result to obtain the address verification result; the attention encoding calculation layer includes a hybrid attention mechanism; the hybrid attention mechanism includes a global attention mechanism and a cross-character dynamic convolutional local attention mechanism; the inputting of the semantic encoding vector into the attention encoding calculation layer to perform attention encoding calculation between the semantic encoding vectors to obtain an attention encoding calculation result includes: performing global attention calculation on the semantic encoding vector according to the global attention mechanism to obtain a global attention result; and performing local attention calculation on the semantic encoding vector according to the cross-character dynamic convolutional local attention mechanism to obtain a cross-character dynamic convolutional local attention result; According to the hybrid attention mechanism, the global attention result and the cross-character dynamic convolution local attention result are concatenated to obtain a hybrid attention aggregation coding vector, and the hybrid attention aggregation coding vector is used as the attention coding calculation result; The calculation method of the mixed attention aggregation encoding vector is as follows: ； Among them, represents a splicing operation; is the character obtained after splicing the global attention result and the cross-character dynamic convolution local attention result of the hybrid attention aggregation encoding vector of the character; A confirmation module for confirming the missing key address elements in the information text according to the address verification result.

9. A verification system for real estate mortgage addresses Characterized in that It includes a memory and a processor, and the verification program for the real estate mortgage address. The processor runs the verification program for the real estate mortgage address so that the verification system for the real estate mortgage address executes the verification method for the real estate mortgage address according to any one of claims 1-7.

10. A computer-readable storage medium Characterized in that The computer-readable storage medium stores a verification program for a real estate mortgage address, and when the verification program for the real estate mortgage address is executed by a processor, it implements the verification method for the real estate mortgage address according to any one of claims 1-7.

Images