Residential electricity utilization address word segmentation analysis method and device

Abstract

The invention relates to the technical field of word segmentation of a power address text, in particular to a word segmentation analysis method and device for a residential electricity utilization address text, and the method specifically comprises the following steps: (1) based on a residential electricity utilization address corpus of a certain region, carrying out the statistical calculation of connection feature information between all sememes in the corpus; (2) carrying out combination and vectorization representation on connection feature information between address meaning elements; (3) adopting a pre-training language model, fusing the connection feature information between the vectorized address semantic elements into the language model, carrying out model training and fine tuning, and forming a power utilization address word segmentation model conforming to the linguistic data features of the residential power utilization addresses in a certain region; and (4) based on the trained and fine-tuned address word segmentation model, performing word segmentation analysis on the residential electricity utilization address in a certain region. The problems that in the prior art, the training cost is high, the requirement for hardware is large, and a large amount of labeled data is needed are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of word segmentation of electric power address texts, and in particular to a word segmentation and parsing method and device for residential electricity address texts. Background Art

[0002] In the power industry, low-voltage residential customer electricity address data on the marketing side is crucial. However, there are many problems with the existing low-voltage residential electricity address data.

[0003] On the one hand, the completeness of address elements is poor, and administrative divisions, roads, detailed information, etc. are often missing;

[0004] On the other hand, the accuracy of address elements is not high, such as wrong road house numbers and community names. These problems lead to business application problems such as wrong work order dispatch, delayed repairs, and inaccurate power outage and restoration notifications, which seriously affect the normal development of power marketing business. At the same time, the potential value of basic marketing address data elements has not been effectively applied. It is urgent to perform word segmentation and analysis on the low-voltage residential customer electricity address data of marketing business in various regions, and divide the manually filled address strings into valid standard address elements.

[0005] In the field of address segmentation in the power industry, traditional methods mainly rely on power address dictionaries and rule-based segmentation methods. The accuracy of segmentation for methods that rely on dictionaries depends on the dictionary itself. When encountering unlisted words or new address expressions, it is difficult to accurately segment words. The rule-based method has shortcomings in model construction. The labeled training address samples often lack the address expression language preferences of the power industry and specific regions, which makes it difficult to verify the model parameters and cannot meet the requirements of high precision and intelligence. Summary of the invention

[0006] The purpose of the embodiment of the present invention is to provide a method and device for word segmentation and parsing of residential electricity addresses, which can fully understand the language preference characteristics of the address expression of the power industry in a specific area, and realize efficient parsing and word segmentation of residential electricity addresses in a specific area. It aims to solve the technical problems of high training cost, high hardware requirements and large amount of labeled data in the prior art, and has low hardware requirements.

[0007] The objective of the present invention is achieved through the following technical solutions:

[0008] In a first aspect, the present invention provides a residential electricity address segmentation and parsing method, which specifically includes the following steps:

[0009] (1) Based on a corpus of residential electricity addresses in a certain area, the connection feature information between each semantic element in the corpus is statistically calculated;

[0010] (2) Combine and vectorize the connection feature information between address semantic elements;

[0011] (3) Using a pre-trained language model, the connection feature information between vectorized address semantic elements is integrated into the language model, and the model is trained and fine-tuned to form an electricity address segmentation model that conforms to the characteristics of the electricity address corpus of residents in a certain area;

[0012] (4) Based on the trained and fine-tuned address segmentation model, the electricity addresses of residents in a certain area are segmented and parsed.

[0013] Preferably, in step (1), the calculation of the connection feature information between the semantic elements in the address corpus specifically includes:

[0014] 1) Using regular expressions, the original data of residential electricity address corpus in a certain area is preprocessed to obtain a regular residential electricity address corpus dataset containing only Chinese characters, English letters, numbers and common symbols in addresses;

[0015] 2) From the regularized residential electricity address corpus data set, count the occurrence frequencies of 1-N semantic elements and filter out the semantic elements with absolutely low frequency;

[0016] 3) Traverse all 1-N semantic elements and use formula (1) to calculate the minimum value of the point-to-point mutual information within each semantic element as the point-to-point mutual information of the semantic element.

[0017]

[0018] Assuming that the current semantic element is composed of m characters, p(·) represents the probability of the semantic element expression “·” appearing in the address corpus, all sub-semesemes of the current semantic element are enumerated to calculate the minimum mutual information as the overall point-to-point mutual information PMI(g), which is used to estimate the connection between the sub-semesemes within the current semantic element. The higher the PMI(g), the more likely the current semantic element is to be an overall address unit.

[0019] 5) Traverse all 1-N semantic elements and use formulas (2) and (3) to calculate the left and right information entropy between each semantic element and its adjacent semantic elements, LE(g) and RE(g).

[0020]

[0021] The left information entropy is calculated by calculating the n adjacent to the left of the current semantic element in the address corpus. l The conditional probability of the semantic elements is obtained. Similarly, the right information entropy is obtained by calculating the n adjacent elements to the right of the current semantic element. r The conditional probability of a semantic element is obtained, and the two are used to measure the closeness of the connection between the current semantic element and the left and right adjacent semantic elements respectively;

[0022] 7) Based on the above calculation results, a statistical dictionary of the connection tightness of 1-N semantic elements is constructed to facilitate subsequent model expression, calculation and training, and micro-survey. The statistical dictionary includes semantic elements, semantic element frequencies, mutual information values ​​between semantic element points, and information entropy values ​​of the left and right semantic elements;

[0023] 8) The determination of the value of N in the 1-N semantic elements needs to comprehensively consider the size of the address corpus and the actual computer performance. Generally, for the statistics of the address corpus of a certain region, the value of N ranges from 3 to 5;

[0024] Preferably, in step (2), the connection feature information between address semantic elements is combined and vectorized, specifically including:

[0025] 1) Given a text address a containing n characters = {c1, c2, ... c n}, traverse each character c in the text address i ;

[0026] 2) Based on the address corpus 1-N semantic connection closeness statistical dictionary constructed in step (2), query a containing character c i The statistical information of 1-N semantic elements, namely the mutual information value between points and the left and right information entropy value;

[0027] 3) Sequentially use c i The left information entropy values ​​of the 1-N semantic elements at the end of the character are arranged in order from the smallest to the largest number of characters to form a vector containing N elements.

[0028] 4) Sequentially include the character c i The mutual information value between the 1-N semantic elements is from small to large in terms of the number of characters. i Arrange the elements from left to right in the sub-element to form a vector containing N*(N+1) / 2 elements.

[0029] 5) Sequentially use c i The right information entropy values ​​of the 1-N semantic elements at the beginning of the character are arranged in order from the smallest to the largest number of characters to form a vector containing N elements.

[0030] 6) According to formula (4), the above constructed vectors are connected and combined to construct a vector e containing N*(N+5) / 2 elements. i ;

[0031] 7) Repeat the above steps 1-6 to obtain the vectorized representation E of the semantic feature information of the text address a a ={e1,e2…e n}.

[0032] 8) Traverse all addresses in the address corpus data set, and repeat the above steps 1-7 to obtain the complete vectorized expression of the connection feature information between the semantic elements of each text address of the current address corpus.

[0033] Preferably, in step (3), the connection feature information between semantic elements of the vectorized address is integrated into the language model, which specifically includes:

[0034] 1) Given an address a processed by step (2), the connection feature information matrix between its internal address semantic elements is E a ={e1,e2…e n};

[0035] 2) According to formula (5), the features in the pre-trained language model are projected into the space of semantic connection feature information, so that the address semantic connection feature information is integrated into the pre-trained language model.

[0036]

[0037] Among them, L BA is the fusion training objective function, MSE() is the mean square error loss, W B is a trainable matrix used to project the pre-trained language model into the semantic feature information space. Represents the hidden layer features of the lth layer in the pre-trained language model.

[0038] Preferably, in step (3), the model is trained, specifically including:

[0039] 1) Select a pre-trained language model that is friendly to Chinese and has superior performance, align it for initialization, and randomly initialize related parameters;

[0040] 2) Traversing all addresses in the address corpus data set, and further training the pre-trained language model according to the training objective function between the vectorized address semantic element connection feature information and the pre-trained language model defined in claim 4, so that the pre-trained language model can master the connection features between the semantic elements of each address in the address corpus data set of a certain region.

[0041] Preferably, in step (3), the model is fine-tuned to establish a local place name dictionary containing address elements of the four-level administrative divisions of a region, namely, province, city, county and township. Based on the dictionary, the language model that has integrated the connection feature information between semantic elements of the address corpus of a region is fine-tuned.

[0042] In a second aspect, the present invention provides a computer-readable storage medium having a program stored thereon, wherein when the program is executed by a processor, the method of the first aspect is implemented.

[0043] Compared with the prior art, the beneficial effects are:

[0044] 1) Large language model, low training cost, high hardware requirements and small demand for large amounts of labeled data. Based on the pre-trained language model and combined with the language preference characteristics of the address expression of the power industry in a specific region, a dedicated model for the segmentation and parsing tasks of residential electricity addresses in a specific region is constructed. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 is a flow chart of a method provided by an embodiment of the present application;

[0047] Figure 2 It is a relationship diagram between a semantic library and a dictionary provided in an embodiment of the present application;

[0048] Figure 3 This is a schematic diagram of an address search process provided by an embodiment of the present application;

[0049] Figure 4 This is a software processing architecture diagram provided in one embodiment of the present application. DETAILED DESCRIPTION

[0050] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0051] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0052] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0053] like Figures 1 to 3As shown, this embodiment provides a residential electricity address word segmentation and parsing method, which specifically includes the following:

[0054] (1) Based on a corpus of residential electricity addresses in a certain area, the connection feature information between each semantic element in the corpus is statistically calculated;

[0055] (2) Combine and vectorize the connection feature information between address semantic elements;

[0056] (3) Using a pre-trained language model, the connection feature information between vectorized address semantic elements is integrated into the language model, and the model is trained and fine-tuned to form an electricity address segmentation model that conforms to the characteristics of the electricity address corpus of residents in a certain area;

[0057] (4) Based on the trained and fine-tuned address segmentation model, the electricity addresses of residents in a certain area are segmented and parsed.

[0058] The calculation of the connection feature information between the semantic elements in the address corpus specifically includes:

[0059] 1) Using regular expressions, the original data of residential electricity address corpus in a certain area is preprocessed to obtain a regular residential electricity address corpus dataset containing only Chinese characters, English letters, numbers and common symbols in addresses;

[0060] 2) From the regularized residential electricity address corpus data set, count the occurrence frequencies of 1-N semantic elements and filter out the semantic elements with absolutely low frequency;

[0061] 3) Traverse all 1-N semantic elements and use formula (1) to calculate the minimum value of the point-to-point mutual information within each semantic element as the point-to-point mutual information of the semantic element.

[0062]

[0063] Assuming that the current semantic element is composed of m characters, p(·) represents the probability of the semantic element expression “·” appearing in the address corpus, all sub-semesemes of the current semantic element are enumerated to calculate the minimum mutual information as the overall point-to-point mutual information PMI(g), which is used to estimate the connection between the sub-semesemes within the current semantic element. The higher the PMI(g), the more likely the current semantic element is to be an overall address unit.

[0064] 4) Traverse all 1-N semantic elements and use formulas (2) and (3) to calculate the left and right information entropy between each semantic element and its adjacent semantic elements, LE(g) and RE(g).

[0065]

[0066] The left information entropy is calculated by calculating the n adjacent to the left of the current semantic element in the address corpus. lThe conditional probability of the semantic elements is obtained. Similarly, the right information entropy is obtained by calculating the n adjacent elements to the right of the current semantic element. r The conditional probability of a semantic element is obtained, and the two are used to measure the closeness of the connection between the current semantic element and the left and right adjacent semantic elements respectively;

[0067] 5) Based on the above calculation results, a statistical dictionary of the connection tightness of 1-N semantic elements is constructed to facilitate subsequent model expression, calculation and training, and micro-survey. The statistical dictionary includes semantic elements, semantic element frequencies, mutual information values ​​between semantic element points, and information entropy values ​​of the left and right semantic elements;

[0068] 6) The determination of the value of N in the 1-N semantic elements needs to comprehensively consider the address corpus size and the actual computer performance. Generally, for the statistics of the address corpus of a certain region, the value of N ranges from 3 to 5; combining and vectorizing the connection feature information between address semantic elements, specifically including:

[0069] 1) Given a text address a containing n characters = {c1, c2, ... c n}, traverse each character c in the text address i ;

[0070] 2) Based on the address corpus 1-N semantic connection closeness statistical dictionary constructed in step (2), query a containing character c i The statistical information of 1-N semantic elements, namely the mutual information value between points and the left and right information entropy value;

[0071] 3) Sequentially use c i The left information entropy values ​​of the 1-N semantic elements at the end of the character are arranged in order from the smallest to the largest number of characters to form a vector containing N elements.

[0072] 4) Sequentially include the character c i The mutual information value between the 1-N semantic elements is from small to large in terms of the number of characters. i Arrange the elements from left to right in the sub-element to form a vector containing N*(N+1) / 2 elements.

[0073] 5) Sequentially use c i The right information entropy values ​​of the 1-N semantic elements at the beginning of the character are arranged in order from the smallest to the largest number of characters to form a vector containing N elements.

[0074] 6) According to formula (4), the above constructed vectors are connected and combined to construct a vector e containing N*(N+5) / 2 elements. i ;

[0075] 7) Repeat the above steps 1-6 to obtain the vectorized representation E of the semantic feature information of the text address a a ={e1,e2…e n}.

[0076] 8) Traverse all addresses in the address corpus data set, and repeat the above steps 1-7 to obtain the complete vectorized expression of the connection feature information between the semantic elements of each text address of the current address corpus.

[0077] The connection feature information between vectorized address semantics is integrated into the language model, including:

[0078] 1) Given an address a processed by step (2), the connection feature information matrix between its internal address semantic elements is E a ={e1,e2…e n};

[0079] 2) According to formula (5), the features in the pre-trained language model are projected into the space of semantic connection feature information, so that the address semantic connection feature information is integrated into the pre-trained language model.

[0080]

[0081] Among them, L BA is the fusion training objective function, MSE() is the mean square error loss, W B is a trainable matrix used to project the pre-trained language model into the semantic feature information space. Represents the hidden layer features of the lth layer in the pre-trained language model.

[0082] In step (3), the model is trained, which specifically includes:

[0083] 1) Select a pre-trained language model that is friendly to Chinese and has superior performance, align it for initialization, and randomly initialize related parameters;

[0084] 2) Traversing all addresses in the address corpus data set, and further training the pre-trained language model according to the training objective function between the vectorized address semantic element connection feature information and the pre-trained language model defined in claim 4, so that the pre-trained language model can master the connection features between the semantic elements of each address in the address corpus data set of a certain region.

[0085] Fine-tune the model and establish a local place name dictionary that contains the address elements of the four-level administrative divisions of a certain region: province, city, county and township. Based on this dictionary, fine-tune the language model that has integrated the connection feature information between the semantic elements of the address corpus of a certain region.

[0086] like Figure 4As shown, another embodiment provides a residential electricity address segmentation and parsing device, including a memory and one or more processors, wherein the memory stores executable code, and when the processor executes the executable code, the address semantic element statistical dictionary defined. The internal semantic element connection information vector database of the full address corpus defined implements the above method.

[0087] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0088] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

Claims

1. A residential electricity address word segmentation and parsing method, characterized in that: The details include: (1) Based on a corpus of residential electricity addresses in a certain area, the connection feature information between each semantic element in the corpus is statistically calculated; (2) Combine and vectorize the connection feature information between address semantic elements; (3) Using a pre-trained language model, the connection feature information between vectorized address semantic elements is integrated into the language model, and the model is trained and fine-tuned to form an electricity address segmentation model that conforms to the characteristics of the electricity address corpus of residents in a certain area; (4) Based on the trained and fine-tuned address segmentation model, the electricity addresses of residents in a certain area are segmented and parsed.

2. A residential electricity address word segmentation and parsing method according to claim 1, characterized in that: In step (1), the calculation of the connection feature information between the semantic elements in the address corpus specifically includes: 1) Using regular expressions, the original data of residential electricity address corpus in a certain area is preprocessed to obtain a regular residential electricity address corpus dataset containing only Chinese characters, English letters, numbers and common symbols in addresses; 2) From the regularized residential electricity address corpus data set, count the occurrence frequencies of 1-N semantic elements and filter out the semantic elements with absolutely low frequency; 3) Traverse all 1-N semantic elements and use formula (1) to calculate the minimum value of the point-to-point mutual information within each semantic element as the point-to-point mutual information of the semantic element. Assuming that the current semantic element is composed of m characters, p(·) represents the probability of the semantic element expression "·" appearing in the address corpus, all sub-semesemes of the current semantic element are enumerated to calculate the minimum mutual information as the overall point-to-point mutual information PMI(g), which is used to estimate the connection density between the sub-semesemes within the current semantic element. The higher the PMI(g), the more likely the current semantic element is to be an overall address unit. 4) Traverse all 1-N semantic elements and use formulas (2) and (3) to calculate the left and right information entropy between each semantic element and its adjacent semantic elements, LE(g) and RE(g). The left information entropy is calculated by calculating the n adjacent to the left of the current semantic element in the address corpus. l The conditional probability of the semantic elements is obtained. Similarly, the right information entropy is obtained by calculating the n adjacent elements to the right of the current semantic element. r The conditional probability of a semantic element is obtained, and the two are used to measure the closeness of the connection between the current semantic element and the left and right adjacent semantic elements respectively; 5) Based on the above calculation results, a statistical dictionary of the connection tightness of 1-N semantic elements is constructed to facilitate subsequent model expression, calculation and training, and micro-survey. The statistical dictionary includes semantic elements, semantic element frequencies, mutual information values ​​between semantic element points, and information entropy values ​​of the left and right semantic elements; 6) The determination of the value of N in the 1-N semantic elements needs to comprehensively consider the size of the address corpus and the actual computer performance. Generally, for the statistics of the address corpus of a certain region, the value of N ranges from 3 to 5.

3. A residential electricity address word segmentation and parsing method according to claim 1, characterized in that: In the step (2), the connection feature information between address semantic elements is combined and vectorized, specifically including: 1) Given a text address a containing n characters = {c1, c2, ... c n }, traverse each character c in the text address i ; 2) Based on the address corpus 1-N semantic connection closeness statistical dictionary constructed in step (2), query a containing character c i The statistical information of 1-N semantic elements, namely the mutual information value between points and the left and right information entropy value; 3) Sequentially use c i The left information entropy values ​​of the 1-N semantic elements at the end of the character are arranged in order from the smallest to the largest number of characters to form a vector containing N elements. 4) Sequentially include the character c i The mutual information value between the 1-N semantic elements is from small to large in terms of the number of characters. i Arrange the elements from left to right in the sub-element to form a vector containing N*(N+1) / 2 elements. 5) Sequentially use c i The right information entropy values ​​of the 1-N semantic elements at the beginning of the character are arranged in order from the smallest to the largest number of characters to form a vector containing N elements. 6) According to formula (4), the above constructed vectors are connected and combined to construct a vector e containing N*(N+5) / 2 elements. i ; 7) Repeat the above steps 1-6 to obtain the vectorized representation E of the semantic feature information of the text address a a ={e1,e2…e n }. 8) Traverse all addresses in the address corpus data set, and repeat the above steps 1-7 to obtain the complete vectorized expression of the connection feature information between the semantic elements of each text address of the current address corpus.

4. A residential electricity address word segmentation and parsing method according to claim 1, characterized in that: In step (3), the connection feature information between the vectorized address semantic elements is integrated into the language model, which specifically includes: 1) Given an address a processed by step (2), the connection feature information matrix between its internal address semantic elements is E a ={e1,e2…e n }; 2) According to formula (5), the features in the pre-trained language model are projected into the space of semantic connection feature information, so that the address semantic connection feature information is integrated into the pre-trained language model. Among them, L BA is the fusion training objective function, MSE() is the mean square error loss, W B is a trainable matrix used to project the pre-trained language model into the semantic feature information space. Represents the hidden layer features of the lth layer in the pre-trained language model.

5. A residential electricity address word segmentation and parsing method according to claim 1, characterized in that: In step (3), the model is trained, which specifically includes: 1) Select a pre-trained language model that is friendly to Chinese and has superior performance, align it for initialization, and randomly initialize related parameters; 2) Traversing all addresses in the address corpus data set, and further training the pre-trained language model according to the training objective function between the vectorized address semantic element connection feature information and the pre-trained language model defined in claim 4, so that the pre-trained language model can master the connection features between the semantic elements of each address in the address corpus data set of a certain region.

6. A residential electricity address word segmentation and parsing method according to claim 1, characterized in that: In the step (3), the model is fine-tuned to establish a local place name dictionary containing address elements of the four-level administrative divisions of a certain region, namely, province, city, county and township. Based on the dictionary, the language model that has been integrated with the connection feature information between semantic elements of the address corpus of a certain region is fine-tuned.

7. A residential electricity address word segmentation and parsing device, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the processor executes the executable code, the address semantic component statistical dictionary defined in claim 2, or the internal semantic component connection information vector database of the full address corpus defined in claim 3, implements the method as described in any one of claims 1-6.

Images