Method, device, system and storage medium for resolving address text

Abstract

The present disclosure relates to a method, device, system and computer medium for parsing address text. The method comprises processing, by a first bidirectional recurrent neural network, an encoding sequence corresponding to each character in the address text to obtain a feature matrix indicating semantic information of the address text and an output matrix of the first bidirectional recurrent neural network; processing, by a second bidirectional recurrent neural network, the output matrix to determine a category to which each character in the address text belongs; processing, by a multi-layer fully connected neural network, the feature matrix to identify administrative division address information at each level corresponding to the address text; and determining administrative division address information at each level and non-administrative division address information of the address text according to the category to which each character in the address text belongs and the identified administrative division address information at each level. According to the above scheme, the address information at each level of the address text can be parsed more accurately and completely.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of data processing, and more particularly, to a method, device, system and computer readable storage medium for parsing address text in the field of data processing. BACKGROUND

[0002] In today's rapidly developing information, data automation processing is getting more and more attention. For example, for any address text, people hope to identify as accurately as possible the administrative division address information such as province, city, district / county and the non-administrative division address information such as street, community name, unit number, door number and the like contained in the address text, so as to construct the complete structured address information of the address text.

[0003] Currently, the traditional way of parsing Chinese address text mainly includes two categories. One is to process the address text as a normal text to get a word segmentation sequence, and then to identify the syntax structure of different words in the text by constructing a syntax structure tree. Although this method can get different words, it is difficult to distinguish the relationship between different words in the address hierarchy, so it is difficult to identify the address text hierarchically. In addition, if the word segmentation processing is not proper, the errors introduced by word segmentation will accumulate and affect the judgment of the syntax structure, and the use of syntax structure tree requires the format of the input address text to have specific rules, so it cannot parse any address text. The other is to set matching rules in advance, and then match according to the rules to find the corresponding hierarchical information in the address. Although this method can identify the address information with key words in the address text, it is difficult to accurately identify the address text hierarchically if the address text is incomplete or lacks key words, and if the matching is wrong, it will further lead to inaccurate address information.

[0004] Since the current way of parsing Chinese address text is not accurate, manual review is often needed after parsing to correct errors. However, since the auditors can only review the parsed address information by eye, under the heavy workload, errors or omissions are likely to occur, making it difficult to improve the accuracy of address parsing.

[0005] In addition, the existing parsing method not only has low accuracy, but also is difficult to complete if the address text lacks one or more administrative division address information, so it is difficult to get hierarchical address information such as province, city, district / county, street, community and the like. SUMMARY

[0006] The present disclosure provides a method, device, system and computer readable storage medium for parsing address text, which can more accurately and completely identify the address information of the address text.

[0007] According to an aspect of the present disclosure, a method for parsing address text is provided. The method comprises: processing, by a first bidirectional recurrent neural network, an encoding sequence corresponding to each word in the address text to obtain a feature matrix indicating semantic information of the address text and an output matrix of the first bidirectional recurrent neural network; processing, by a second bidirectional recurrent neural network, the output matrix to determine a category to which each word in the address text belongs; processing, by a multi-layer fully connected neural network, the feature matrix to identify administrative division address information at each level corresponding to the address text; and determining administrative division address information at each level and non-administrative division address information of the address text according to the category to which each word in the address text belongs and the identified administrative division address information at each level.

[0008] According to another aspect of the present disclosure, an apparatus for parsing address text is provided. The apparatus comprises means for performing the steps of the above method.

[0009] According to still another aspect of the present disclosure, an apparatus for parsing address text is provided. The apparatus comprises a memory storing computer executable instructions, and a processor coupled to the memory, wherein the computer executable instructions, when executed by the processor, cause the processor to perform the above method.

[0010] According to yet another aspect of the present disclosure, a computer readable storage medium is provided. The medium stores computer executable instructions, wherein the computer executable instructions, when executed by a processor, cause the processor to perform the above method.

[0011] According to the above technical solution, by using the first and second bidirectional recurrent neural networks, the address information at the corresponding level contained in the address text itself can be identified on a word-by-word basis. By using the multi-layer fully connected neural network, according to the address text feature information extracted by the first bidirectional recurrent neural network, the administrative division address information at each level corresponding to the address text can be comprehensively identified, so that the administrative division address information that may be missing in the address text can be completed. By combining the two types of address information identified above, the administrative division address information and the non-administrative division address information of the address text can be more accurately and completely parsed. In addition, since the above technical solution does not have special requirements for the format of the address text, it can parse address texts of any format, thereby having stronger universality.

[0012] Other features and advantages of the present disclosure will become apparent from the following detailed description of exemplary embodiments thereof, taken in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0013] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0014] Figure 1 is a flowchart of a method for parsing address text according to an embodiment of the present disclosure.

[0015] Figure 2 is a structural schematic diagram of a bidirectional recurrent neural network according to an embodiment of the present disclosure.

[0016] Figure 3 is a functional block diagram of a parsing model for parsing address text according to an embodiment of the present disclosure.

[0017] Figure 4 is a flowchart of a method for training a parsing model according to an embodiment of the present disclosure.

[0018] Figure 5 is a flowchart of another method for training a parsing model according to an embodiment of the present disclosure.

[0019] Figure 6 is a structural block diagram of an apparatus for parsing address text according to an embodiment of the present disclosure.

[0020] Figure 7 is a structural block diagram of another apparatus for parsing address text according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0021] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement, numerical expressions, and values of components and steps set forth in these embodiments do not limit the scope of the present disclosure unless otherwise specifically stated.

[0022] At the same time, it should be understood that the size of each part shown in the drawings is not drawn in accordance with the actual proportional relationship for the convenience of description. The following description of at least one exemplary embodiment is actually only illustrative, but by no means as any limitation on the present disclosure and its application or use. The technology, methods and equipment known to those skilled in the related art can not be discussed in detail, but should be considered as part of the specification under appropriate circumstances. In all examples shown and discussed here, any specific value should be interpreted as merely exemplary, not as a limitation. Therefore, other examples of exemplary embodiments can have different values. It should be noted that similar reference numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0023] First, with reference toFigure 1 A flow chart of a method 100 for parsing address text according to an embodiment of the present disclosure is described.

[0024] In S110, the encoding sequence corresponding to each word in the address text is processed by a bidirectional recurrent neural network to obtain a feature matrix indicating semantic information of the address text and an output matrix of the bidirectional recurrent neural network.

[0025] Specifically, the address text to be parsed is split by word to obtain a sequence of text characters. For each character in the sequence of text characters, it is encoded to be converted into an encoding sequence. For example, the characters representing each word can be mapped by a vectorization model, so that each word is mapped to an encoding sequence of 1*N. For an address text containing M words, a vectorization matrix of M*N size can be obtained by mapping of the vectorization model, where each row can correspond to the encoding sequence of a word. In this way, the vectorization of the text is realized, and the original address text is finally converted into a floating-point vectorization matrix, for example. The vectorization model used for vectorizing the text can use a model known to those skilled in the art for vectorizing text.

[0026] After determining the encoding sequence of each word in the address text, the encoding sequence of each word can be sequentially input into the bidirectional recurrent neural network. Figure 2 A schematic diagram of the bidirectional recurrent neural network is shown in represents the input of the bidirectional recurrent neural network, h i represents the output of the bidirectional recurrent neural network, c1 and c2 represent the state matrix of the bidirectional recurrent neural network, the first row composed of LSTM (Long Short Term Memory) is called the backward path of the bidirectional recurrent neural network, and the second row composed of LSTM is called the forward path of the bidirectional recurrent neural network. It should be noted that matrices and vectors can be used interchangeably in this document, and those skilled in the art can fully understand their meanings.

[0027] The encoding sequence of each word is sequentially input from After the encoding sequence of the last word of the address text is input into the bidirectional recurrent neural network for processing, at this time, the state matrix c1 in the backward path and the state matrix c2 in the forward path are concatenated to form a feature matrix C = [c1, c2] that can indicate the semantic information of the address text, and an output matrix H = [h0, h1, h2, …] is obtained.

[0028] In S120, the above output matrix is processed by another bidirectional recurrent neural network to determine the category to which each word in the address text belongs.

[0029] Specifically, the output matrix H of the bidirectional recurrent neural network in S110 is input into another bidirectional recurrent neural network in S120. The bidirectional recurrent neural network also has the structure shown in the above formula (1) and (2). After processing by the other bidirectional recurrent neural network, a feature matrix that reflects more semantic information of the address text than the feature matrix extracted by the bidirectional recurrent neural network in S110 can be extracted, so that the contextual semantic relationship of the address text is further represented by the new feature matrix. Figure 2

[0030] According to the new feature matrix for representing the contextual semantic relationship of the address text, it can be more accurately identified which words in the address text belong to the same semantic, so as to be identified as the same category. For example, each element in the output matrix [h0, h1, h2, …] in S110 can be sequentially input into the hidden state matrix of the bidirectional recurrent neural network in S120. After processing the last element in the output matrix H = [h0, h1, h2, …], at this time, the state matrix c1' in the backward path and the state matrix c2' in the forward path of the bidirectional recurrent neural network in S120 are spliced to form another feature matrix C' = [c1', c2']. Taking C' as the input of the conditional random field algorithm, the category to which each word in the address text belongs is determined by the conditional random field algorithm, and the words belonging to the same category are marked with the same number. The conditional random field algorithm is a technology known to those skilled in the art, and how it operates is well known, which will not be described here.

[0031] For example, for the address text to be parsed "Anhui Province Hefei Qili Station Wanfu Garden 5 402", after processing by S110 and S120, the category information of each word can be obtained [1, 1, 1, 2, 2, 4, 4, 4, 5, 5, 5, 5, 0, 0, 0, 0, 0]. In the predefined rules, category 1 can represent province name, category 2 can represent city name, category 3 can represent district / county name, category 4 can represent street name, category 5 can represent community name, and category 0 can represent house number, room number, etc. From this, according to the category to which each word belongs, the corresponding hierarchical address information represented by the corresponding character can be determined.

[0032] In S130, the feature matrix C in S110 is processed by a multi-layer fully connected neural network to identify the hierarchical administrative division address information corresponding to the address text.

[0033] The multi-layer fully connected neural network here can also be referred to as a classifier in the art, which is a concept well known to those skilled in the art, and it can be trained to classify input data.

[0034] ​Here, the feature matrix C in S110 is multiplied by a standard address feature matrix E determined in advance according to a plurality of external standard administrative division address texts, to obtain a correlation matrix K = C*E T . Then, the correlation matrix K is input into the multi-layer fully connected neural network for processing, and an output vector is obtained at the output end of the multi-layer fully connected neural network. Each element in the output vector respectively represents the administrative division address information containing provinces, cities and districts / counties specified in advance in the training stage. The maximum value element in the current output vector is taken, and the pre-specified administrative division address information corresponding to the maximum value element is determined as the administrative division address information of each level corresponding to the address text.

[0035] For example, after the correlation matrix K between the feature matrix C of the address text to be parsed "Anhui Province Hefei Qilizhan Wanfu Garden 5 402" and the standard address feature matrix E is input into the multi-layer fully connected neural network, the maximum value element in the output vector corresponds to the pre-specified administrative division address information "Anhui Province, Hefei City, Yaohai District", so it can be determined that the administrative division address information of each level of the address text to be parsed is "Anhui Province, Hefei City, Yaohai District". As can be seen, although the address information at the district / counties level is not included in the original address text, the missing level information can be completed after classification by the multi-layer fully connected neural network, so that more complete address information can be obtained.

[0036] The standard address feature matrix E can be obtained in a similar manner to the manner of obtaining the feature matrix C of the address text to be parsed. Specifically, for each external standard administrative division address text, the same vectorization model as in S110 is used to map each character in the external standard administrative division address text to a corresponding code sequence, and the code sequence corresponding to each character is input into the bidirectional recurrent neural network in S110 in order. After all the characters are input into the bidirectional recurrent neural network in S110 for processing, the state matrix in the backward path and the state matrix in the forward path of the bidirectional recurrent neural network together constitute a feature matrix for representing the semantic information of the external standard administrative division address text. The feature matrices of each external standard administrative division address text are concatenated to form the standard address feature matrix E.

[0037] In S140, according to the category to which each character in the address text belongs determined in S120 and the administrative division address information of each level identified in S130, the administrative division address information of each level and the non-administrative division address information of the address text are determined.

[0038] Specifically, the recognition results in S120 and S130 are combined to obtain the hierarchical address resolution result of the address text. For example, for the address text to be resolved "Anhui Province Hefei Qilizhan Wanfu Garden 5 402", the category recognition result in S120 is [1, 1, 1, 2, 2, 4, 4, 4, 5, 5, 5, 5, 0, 0, 0, 0, 0], and the classification result in S130 is "Anhui Province, Hefei City, Yaohai District". By combining the two, the hierarchical address information of the address text can be determined as "Province: Anhui Province, City: Hefei City, District / County: Yaohai District, Street: Qilizhan, Community: Wanfu Garden, House Information: 5 402", thereby accurately and completely resolving the information of each level of the address text.

[0039] As can be understood by those skilled in the art, if the recognition results of each word category in S120 and the classification results in S130 are inconsistent in terms of administrative division address information, conflict resolution specifications (for example, taking the results in S130 as the standard) can be set to more accurately obtain the administrative division address information.

[0040] By the method provided by the embodiments of the present disclosure, the state matrix-based feature matrix that more fully represents the context semantic information of the address text can be obtained by using the bidirectional recurrent neural network in the cascaded S110 and S120, thereby identifying the address information contained in the address text on a word-by-word basis. With the help of the feature matrix representing semantic information provided by the bidirectional recurrent neural network in S110, through the classification of the multi-layer fully connected neural network in S130, the administrative division address information corresponding to the address text at each level can be identified, thereby the administrative division address information that may be missing in the address text can be completed. By combining the recognition results in S120 and S130, the hierarchical address information of the address text can be more accurately and completely resolved.

[0041] Figure 3 A functional block diagram of a resolution model 300 for resolving an address text according to an embodiment of the present disclosure is shown. The resolution model 300 includes a text analysis module 320, an element extraction module 330, an intelligent inference module 340, and a result determination module 350. Each of these functional modules can be implemented by program code written in various programming languages, can be implemented by hardware or firmware capable of implementing the corresponding functions, can be implemented by a development software package (such as tensorflow) specially used for writing and training neural networks, or can be implemented by any combination of the above. In addition, one or more of these functional modules can be combined into a larger module, or part of the operation of a certain functional module can be implemented by another functional module.

[0042] The address text 310 to be parsed is input into a vectorization model 322 in a text analysis module 320. Each character in the address text 310 is mapped into a sequence of codes by the vectorization model 322, so as to realize vectorization of the address text. Then, the sequence of codes corresponding to each character is input into a trained bidirectional recurrent neural network 324. The bidirectional recurrent neural network 324 processes the input data to extract relevant features of the address text 310, so as to obtain a feature matrix C composed of state matrices in a backward path and state matrices in a forward path of the bidirectional recurrent neural network 324, and also obtain an output matrix H.

[0043] The output matrix H is input into a trained bidirectional recurrent neural network 332 in an element extraction module 330. After processing by the bidirectional recurrent neural network 332, a feature matrix C’ composed of state matrices in a backward path and state matrices in a forward path of the bidirectional recurrent neural network 332 is obtained, which can represent semantic information of the address text 310 more than the feature matrix C, so as to more accurately reflect the semantic relationship between characters in the address text 310. The feature matrix C’ is input into a conditional random field algorithm module 334. Through processing by the existing conditional random field algorithm, a class 336 to which each character belongs can be identified and labeled with a class number, for example, a character labeled with class 1 corresponds to a province name, a character labeled with class 2 corresponds to a city name, and so on. In this way, through processing by the element extraction module 330, administrative division address information, community information, house information, and the like contained in the address text 310 itself can be more accurately identified character by character.

[0044] The feature matrix C obtained in the bidirectional recurrent neural network 324 in the text analysis module 320 is input into an association matrix calculation module 342 in an intelligent inference module 340. Another input of the calculation module 342 is a standard address feature matrix E as described above. By calculating an association matrix K = C * E T , the association matrix K is input into a trained multi-layer fully connected neural network 344 as a classifier, so as to determine the maximum element in the output vector of the multi-layer fully connected neural network 344 as the predetermined province, city, and district / county information corresponding to the address text 310 as the administrative division address information 346 of each level.

[0045] The class 336 to which each character recognized by the element extraction module 330 belongs and the administrative division address information 346 of each level recognized by the intelligent inference module 340 are input into a result determination module 350 to combine these information to obtain an address parsing result, so as to determine the province, city, and district / county address information corresponding to the address text 310 and possible street information, community information, house information, and the like.

[0046] The bidirectional recurrent neural network 324, the bidirectional recurrent neural network 332 and the multi-layer fully connected neural network 344 are all neural network models that are trained on a large amount of training texts in advance. Figure 4 A flowchart of a method 400 for jointly training them is specifically described.

[0047] In S410, a large number of training texts, i.e., a large number of known address texts, are prepared, and each of them is annotated, not only the category to which each word in each known address text belongs is annotated, but also the administrative division address information corresponding to each known address text, including province, city, district / county information, is annotated.

[0048] For any address text in a large number of training texts, S420 , S430 and S440 are executed.

[0049] In S420, the encoding sequence corresponding to each word in the text is input into the first bidirectional recurrent neural network for processing to obtain a feature matrix indicating the semantic information of the text and an output matrix of the first bidirectional recurrent neural network. The first bidirectional recurrent neural network is set with random initialization parameters when S420 is first executed. The first bidirectional recurrent neural network and the method for obtaining the feature matrix and output matrix here are similar to those in S110 and will not be repeated here.

[0050] In S430, the output matrix in S420 is input into the second bidirectional recurrent neural network for processing to determine the category to which each word in the text belongs using the conditional random field algorithm, and calculate the conditional random field loss (crf_loss) between the determined category and the marked category. The second bidirectional recurrent neural network is set with random initialization parameters when S430 is executed for the first time. The second bidirectional recurrent neural network here and the method of obtaining the category are similar to those in S120 and will not be repeated here. In addition, the conditional random field loss (crf_loss) is a concept well known to those skilled in the art, and can be implemented, for example, by the tensorflow software development kit.

[0051] In S440, the feature matrix in S420 is input into the multi-layer fully connected neural network for processing to identify the administrative division address information at each level corresponding to the text, and to calculate the multi-class cross-entropy loss (softmax_loss) between the identified administrative division address information at each level and the labeled administrative division address information at each level. The multi-layer fully connected neural network is provided with random initial parameters when it is first executed in S440. The multi-layer fully connected neural network and the way of obtaining the administrative division address information at each level are similar to those in S130, and will not be described here again. In addition, the multi-class cross-entropy loss (softmax_loss) is a concept known to those skilled in the art, which can be implemented by, for example, the tensorflow software development kit.

[0052] In S450, the parameters in the first bidirectional recurrent neural network, the second bidirectional recurrent neural network and the multi-layer fully connected neural network are adjusted, and S420, S430 and S440 are continued to be executed for another text in the large number of training texts, so that the sum of the conditional random field loss and the multi-class cross-entropy loss calculated again is smaller.

[0053] Then, in order to minimize the sum of the conditional random field loss and the multi-class cross-entropy loss, S450 is continued to be executed, so that the parameters in the neural networks can be further optimized, and thus the parsing model 300 containing the neural networks which can more accurately and completely parse address texts is trained.

[0054] In addition, Figure 3 The vectorization model 322 in the text vectorization in S420 can also be a neural network model, and can also be trained together with the neural networks 324, 332 and 344. In Figure 5 In S510, the training starts, and the large number of training address texts are labeled in the manner in S410. In addition, the neural networks are respectively initialized with random parameters.

[0055] In S510, the training starts, and the large number of training address texts are labeled in the manner in S410. In addition, the neural networks are respectively initialized with random parameters.

[0056] In S520, a certain training address text is input into the text analysis module 320 shown in S420, so as to obtain the feature matrix C and the output matrix H of the training address text by the vectorization module 322 and the bidirectional recurrent neural network 324. Figure 3 In S520, a certain training address text is input into the text analysis module 320 shown in S420, so as to obtain the feature matrix C and the output matrix H of the training address text by the vectorization module 322 and the bidirectional recurrent neural network 324. Figure 3The text analysis module 320 shown in FIG. 2 obtains a respective feature matrix for each external standard administrative division address text, and splices the feature matrices together to form a standard address feature matrix E.

[0057] In S530, the output matrix H is input to Figure 3 The element extraction module 330 shown in FIG. 3 determines the category to which each character in the text belongs by using a bidirectional recurrent neural network 332 and a conditional random field algorithm, thereby obtaining address information contained in the address text itself.

[0058] In S540, the feature matrix C is input to Figure 3 The intelligent inference module 340 shown in FIG. 4 is also input the standard address feature matrix E. By calculating the correlation matrix K of C and E, K can be input to a multi-layer fully connected neural network 344 acting as a classifier for processing to determine the administrative division address information at each level of the address text.

[0059] In S550, the address information determined in S530 and the administrative division address information at each level determined in S540 are merged to obtain address information at each level related to the address text used for training.

[0060] In S560, the loss between the predicted address information and the address information labeled in S510 is calculated. That is, for the final determined administrative division address information at each level and the administrative division address information labeled in S510, a multi-class cross-entropy loss is calculated, and for the final determined category to which each character belongs and the category to which each character belongs labeled in S510, a conditional random field loss is calculated, and the sum of the two types of losses is taken as the total loss of the address resolution result obtained in S550.

[0061] In S570, by means of a back propagation mechanism, the total loss of the final result is conducted back to the parameters preset in the vectorization model 322, the bidirectional recurrent neural network 324, the bidirectional recurrent neural network 332 and the multi-layer fully connected neural network 344, and each parameter is optimized by gradient descent to minimize the sum of the multi-class cross-entropy loss and the conditional random field loss.

[0062] In S580, S520-S570 are continuously executed until the training of the resolution model converges, for example, a preset number of training rounds is reached, or a total loss threshold is lower than a preset value, or an expected address resolution accuracy is reached.

[0063] Through the above training process, the resolution model 300 shown in FIG. 5 can be obtained, so that the resolution model 300 is used to implement hierarchical resolution of any Chinese address text by executing the method 100. Figure 3

[0064] ​According to the address resolution method provided by the embodiment of the present disclosure, since the resolution model comprises multiple trained neural networks, the address text can be accurately and completely resolved in a hierarchical manner, and even if the administrative division address information is missing, it can be completed. Moreover, since the external standard administrative division address text is used in training the resolution model, the identification of the administrative division address information at each level can be more accurate, avoiding the interference caused by the extra information in the address text. In addition, since the resolution model divides the address text by word and extracts features by using a bidirectional recurrent neural network, the address text to be resolved does not need to rely on any specific format, thereby being more versatile.

[0065] The method for resolving the address text according to the embodiment of the present disclosure is described above, and the structure block diagram of the device for resolving the address text according to the embodiment of the present disclosure is described below. Figure 6 to 7 The structure block diagram of the device for resolving the address text according to the embodiment of the present disclosure is described above, and the structure block diagram of the device for resolving the address text according to the embodiment of the present disclosure is described below.

[0066] Figure 6 The device 600 for resolving the address text shown in the embodiment of the present disclosure comprises a first processing component 610, a second processing component 620, a third processing component 630 and a result determining component 640. Each of these components can be a processor or a part thereof, or a program code stored in a memory or other storage space, any of them can be realized in the form of hardware, software, firmware or any combination thereof.

[0067] According to the embodiment of the present disclosure, the first processing component 610 can be configured to process the encoding sequence corresponding to each word in the address text by the first bidirectional recurrent neural network to obtain a feature matrix for indicating semantic information of the address text and an output matrix of the first bidirectional recurrent neural network. The second processing component 620 can be configured to process the output matrix by the second bidirectional recurrent neural network to determine the category to which each word in the address text belongs. The third processing component 630 can be configured to process the feature matrix by the multi-layer fully connected neural network to identify the administrative division address information at each level corresponding to the address text. The result determining component 640 can be configured to determine the administrative division address information at each level and the non-administrative division address information of the address text according to the category to which each word in the address text belongs and the identified administrative division address information at each level.

[0068] In addition, the first processing component 610, the second processing component 620, the third processing component 630 and the result determining component 640 can be further configured to perform the related steps in the above address resolution method, and for the sake of brevity, the description is not repeated.

[0069] The above and other operations and / or functions of the first processing component 610, the second processing component 620, the third processing component 630, and the result determining component 640 can refer to the relevant descriptions in Figures 1 to 5 , which will not be repeated here.

[0070] According to the apparatus for address resolution provided in the embodiments of the present disclosure, by using the first and second bidirectional recurrent neural networks, the corresponding hierarchical address information contained in the address text itself can be identified on a character-by-character basis. By using the multi-layer fully connected neural network, according to the address text feature information extracted by the first bidirectional recurrent neural network, the hierarchical administrative division address information corresponding to the address text can be comprehensively identified, so that the administrative division address information that may be missing in the address text can be completed. By combining the two types of address information identified above, the administrative division address information and the non-administrative division address information of the address text can be more accurately and completely resolved. In addition, since there is no special requirement for the format of the address text, it has stronger universality.

[0071] Figure 7 The apparatus 700 shown includes a memory 710 and a processor 720. The memory 710 can be a read-only memory, an optical disc, a hard disk, a magnetic disk, a flash memory, or any other non-volatile storage medium. The memory can store computer executable instructions for implementing one or more steps of at least one of the methods 100, 400, and 500 and the methods involved in the resolution model 300 described above.

[0072] The processor 720 can be coupled to the memory 710, for example, through a bus, and can be implemented as one or more integrated circuits, such as a microprocessor or a microcontroller. The processor 720 is configured to execute computer executable instructions stored in the memory 710 for implementing one or more steps of the methods described above. Through execution of the computer executable instructions, the hierarchical address information of the address text can be more accurately and completely resolved, and it has stronger universality since there is no special requirement for the format of the address text.

[0073] As in existing computer apparatuses, the apparatus 700 can be connected to external storage devices through a read-write interface to invoke external data, and can be connected to a network or other computer apparatuses through a network interface, which will not be described in detail here.

[0074] According to the embodiments of the present disclosure, computer executable instructions for performing one or more steps of the methods described above can be stored on a computer readable medium, which when executed by a processor, can cause the processor to perform the corresponding steps, thereby more accurately and completely resolving the hierarchical address information of the address text.

[0075] The apparatuses 600 and 700 described above can be connected in a network system as networked devices to communicate with other devices. The network system can be a local area network, a metropolitan area network, a wide area network, or a wireless cellular network, etc. In such a network system, at least one apparatus is capable of implementing the method and the analytical model described above.

[0076] Those skilled in the art will understand that the embodiments of the present disclosure can be provided as methods, apparatuses, or computer program products. Therefore, the present disclosure can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present disclosure can take the form of a computer program product implemented on one or more computer-usable non-transitory storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) having computer usable program code embodied therein.

[0077] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, apparatuses (systems) and computer program products according to the embodiments of the present disclosure. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as a combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus generate a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that performs the functions specified in the flowcharts and / or block diagrams.

[0078] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatuses that implement the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that performs the functions specified in the flowcharts and / or block diagrams.

[0079] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that performs the functions specified in the flowcharts and / or block diagrams.

[0080] So far, the present disclosure has been described in detail. In order to avoid obscuring the concept of the present disclosure, some details known in the art are not described. Those skilled in the art can fully understand how to implement the technical solutions disclosed herein according to the above description.

[0081] The method and system of the present disclosure can be implemented in many ways. For example, the method and system of the present disclosure can be implemented by software, hardware, firmware, or any combination of software, hardware, firmware. The above order of steps for the method is only for illustration, and the steps of the method of the present disclosure are not limited to the above specific description, unless otherwise specifically described. In addition, in some embodiments, the present disclosure can also be implemented as programs recorded in recording media, which include machine-readable instructions for implementing the method according to the present disclosure. Therefore, the present disclosure also covers the recording media storing the programs for executing the method according to the present disclosure.

[0082] Although some specific embodiments of the present disclosure have been described in detail through examples, those skilled in the art should understand that the above examples are only for illustration, not for limiting the scope of the present disclosure. Those skilled in the art should understand that the above embodiments can be modified without departing from the scope and spirit of the present disclosure. The scope of the present disclosure is defined by the appended claims.

Claims

1. A method for parsing an address text, comprising: Processing a coding sequence corresponding to each word in the address text by a first bidirectional recurrent neural network to obtain a feature matrix indicating semantic information of the address text and an output matrix of the first bidirectional recurrent neural network; Processing the output matrix through a second bidirectional recurrent neural network to determine the category to which each word in the address text belongs; Processing the feature matrix through a multi-layer fully connected neural network to identify address information of each level of administrative division corresponding to the address text, including: calculating an association matrix between the feature matrix and a standard address feature matrix pre-determined based on a plurality of external standard administrative division address texts, and inputting the association matrix into the multi-layer fully connected neural network for processing to identify predetermined administrative division address information corresponding to a maximum value element in an output of the multi-layer fully connected neural network as the address information of each level of administrative division corresponding to the address text; and Determine the administrative division address information and non-administrative division address information of the address text according to the category of each word in the address text and the identified administrative division address information of each level corresponding to the address text, The standard address feature matrix is ​​obtained by the following operations: for each external standard administrative division address text, the coding sequence corresponding to each word in the external standard administrative division address text is input into the first bidirectional recurrent neural network for processing, wherein, after each word in the external standard administrative division address text is input into the first bidirectional recurrent neural network for processing, the state matrix in the backward path and the state matrix in the forward path of the first bidirectional recurrent neural network together constitute a feature matrix for representing the semantic information of the external standard administrative division address text; the feature matrix of each external standard administrative division address text is combined into a standard address feature matrix.

2. The method according to claim 1, wherein The feature matrix is ​​obtained by the following operations: Inputting the encoding sequence corresponding to each word in the address text into the first bidirectional recurrent neural network for processing; and After all the words in the address text are input into the first bidirectional recurrent neural network for processing, the state matrix in the backward path and the state matrix in the forward path of the first bidirectional recurrent neural network together constitute the feature matrix.

3. The method according to claim 1, wherein The processing of the output matrix by a second bidirectional recurrent neural network to determine the category to which each word in the address text belongs includes: Processing the output matrix through a second bidirectional recurrent neural network to obtain a state matrix of the second bidirectional recurrent neural network; and According to the state matrix of the second bidirectional recurrent neural network, a conditional random field algorithm is used to determine the category to which each word in the address text belongs.

4. The method according to claim 3, wherein: Determining the category to which each word in the address text belongs by using a conditional random field algorithm according to the state matrix of the second bidirectional recurrent neural network includes: After all output matrices of the first bidirectional recurrent neural network are input into the second bidirectional recurrent neural network, the state matrix in the backward path and the state matrix in the forward path of the second bidirectional recurrent neural network together constitute another feature matrix; and According to the another feature matrix, a conditional random field algorithm is used to determine the category to which each word in the address text belongs.

5. The method according to claim 1, wherein The first bidirectional recurrent neural network, the second bidirectional recurrent neural network, and the multi-layer fully connected neural network are obtained through the following training: For each text in a large number of known address texts, mark the category to which each word in the text belongs and the address information of each level of administrative division corresponding to the text; For any text in the large number of known address texts, the following determining step, first loss calculation step and second loss calculation step are performed, a determining step for inputting a coding sequence corresponding to each word in the text into a first bidirectional recurrent neural network for processing to obtain a feature matrix indicating semantic information of the text and an output matrix of the first bidirectional recurrent neural network, The first loss calculation step is used to process the output matrix through a second bidirectional recurrent neural network to determine the category to which each word in the text belongs using a conditional random field algorithm, and calculate the conditional random field loss between the determined category and the labeled category, A second loss calculation step is used to process the feature matrix through a multi-layer fully connected neural network to identify the administrative division address information at each level corresponding to the text, and calculate the multi-classification cross entropy loss between the identified administrative division address information at each level and the labeled administrative division address information at each level; as well as Adjust parameters in the first bidirectional recurrent neural network, the second bidirectional recurrent neural network, and the multi-layer fully connected neural network, and continue to perform the determining step, the first loss calculation step, and the second loss calculation step for other arbitrary texts in the large amount of known address texts to minimize the sum of the conditional random field loss and the multi-classification cross entropy loss.

6. An apparatus for parsing address text, comprising components for executing the steps of the method according to any one of claims 1 to 5.

7. A device for parsing address text, comprising: a memory storing computer-executable instructions; as well as A processor is coupled to the memory, and the computer executable instructions, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 5.

8. A computer-readable storage medium storing computer-executable instructions, wherein when the computer-executable instructions are executed by a processor, the processor is caused to perform the method according to any one of claims 1 to 5.

Images