Address matching method and system, readable medium and equipment

Abstract

The invention relates to an address matching method and system, a readable medium and equipment, and relates to the technical field of natural language processing. According to the scheme, after a to-be-matched address text is obtained, vectorized semantic matching is carried out on the address text to obtain candidate matching addresses and corresponding semantic similarity, and rule matching is carried out on the address text according to an address matching rule corresponding to a scene type on the basis of scene recognition to obtain a rule hit score; and on the basis, performing weighted calculation on the semantic similarity and the rule hit score by using a dynamic weight corresponding to the scene type to obtain a target matching score, and determining a target matching address in the candidate matching addresses by using the target matching score. According to the scheme, semantic understanding and rule matching can be integrated, the scene type can dynamically influence rule selection and weight distribution, the address matching precision is effectively improved, mismatching is reduced, the decision tracing and reproduction difficulty is also reduced, recheck and verification are supported, and the rule is easy to dynamically expand and maintain and adapts to demand changes.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of natural language processing, and in particular, to an address matching method, an address matching system, a computer readable storage medium, and an electronic device. BACKGROUND

[0002] The matching of addresses is usually implemented by using regular expressions or keyword matching, and there are also schemes for text vectorization semantic matching. On this basis, a static rule hybrid model can be combined with semantic vectors and predefined rules to achieve the matching.

[0003] However, the foregoing schemes cannot exclude the influence of problems such as expression ambiguity, hierarchical missing, and business scene mismatch. For example, regular expressions or keyword matching have the problem of low precision. Text vectorization semantic matching is greatly affected by corpus quality, has no specific perception of characters related to business scenes in addresses, is not sensitive to administrative levels related to business in addresses, and has an opaque decision mechanism, which is difficult to review and check. The static rule hybrid model has the problem of business scene mismatch caused by rigid rules, and has high rule extension cost and is difficult to maintain dynamically.

[0004] It should be noted that the information disclosed in the above BACKGROUND section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that constitutes prior art known to those of ordinary skill in the art, and it can also include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present disclosure is to provide an address matching method, an address matching system, a computer readable storage medium, and an electronic device, thereby at least to some extent improving address matching precision, reducing the occurrence of business scene mismatch, and also reducing the difficulty of decision tracing and reproduction, supporting review and checking, and easily dynamically extending, maintaining, and adapting to changes in business requirements.

[0006] According to one aspect of the present disclosure, an address matching method is provided, which can include: obtaining address text to be matched; performing scene recognition on the address text, and performing rule matching on the address text by using address matching rules corresponding to the scene type to obtain a rule hit score corresponding to the address text; performing vectorization semantic matching on the address text to obtain at least one candidate matching address and a semantic similarity corresponding to each candidate matching address; performing weighted calculation on the semantic similarity and the rule hit score by using a dynamic weight corresponding to the scene type to obtain a target matching score of the candidate matching address; and determining a target matching address based on the target matching score in the candidate matching address.

[0007] In an optional exemplary embodiment of this disclosure, each address matching rule is configured with a rule weight corresponding to the scene type. The address text is matched with the address matching rule corresponding to the scene type to obtain the rule hit score corresponding to the address text. This includes: matching the address text with the address matching rule corresponding to the scene type to obtain the initial hit score corresponding to each address matching rule; and weighting the initial hit score with the rule weight to obtain the rule hit score.

[0008] In an optional exemplary embodiment of this disclosure, the number of address matching rules is greater than 1. The initial hit score is weighted and calculated using rule weights to obtain a rule hit score, including: calculating a weight decay parameter based on the priority of the address matching rules; the weight decay parameter decreases as the priority decreases; adjusting the rule weights of the address matching rules based on the weight decay parameter to obtain effective weights; and weighting the initial hit score using the effective weights to obtain a rule hit score.

[0009] In an optional exemplary embodiment of this disclosure, calculating the weight decay parameter based on the priority of the address matching rules includes: sorting each address matching rule according to its priority; calculating the weight decay parameter corresponding to each address matching rule in each position using a weight decay function; and defining the decay gradient for each increment of the weight decay parameter in each position.

[0010] In an optional exemplary embodiment of this disclosure, address matching rules are divided by rule type, and each rule type has a corresponding weight configuration range; under a scenario type, each address matching rule sets a rule weight within the weight configuration range according to its rule type.

[0011] In an optional exemplary embodiment of this disclosure, vectorized semantic matching is performed on address text to obtain at least one candidate matching address and the semantic similarity corresponding to each candidate matching address, including: performing semantic equivalence mapping and vectorization processing on the address text to obtain an initial character vector; using a pre-trained model to predict the label of the initial character vector to obtain a standard character vector; and performing vectorized semantic matching based on the standard character vector to obtain at least one candidate matching address and the semantic similarity corresponding to each candidate matching address.

[0012] In an optional exemplary embodiment of this disclosure, semantic equivalence mapping and vectorization are performed on address text to obtain an initial character vector, including: performing semantic equivalence mapping and embedding on address text to obtain a general character vector; extracting domain features from the general character vector based on a domain dictionary to obtain a domain feature vector; extracting self-attention features from the general character vector to obtain a context feature vector; and fusing the general character vector, the domain feature vector, and the context feature vector to obtain the initial character vector.

[0013] In an optional exemplary embodiment of this disclosure, determining a target matching address from candidate matching addresses based on a target matching score includes: determining a candidate matching address as a target matching address when the target matching score is greater than or equal to a first matching threshold; determining that the candidate matching address has failed to match when the target matching score is less than a second matching threshold; reviewing the candidate matching address and the target matching score when the target matching score is less than the first matching threshold and greater than or equal to the second matching threshold; and adjusting the rule weights based on the review results.

[0014] In an optional exemplary embodiment of this disclosure, the method further includes: performing a decision transparency analysis on the address matching process; the decision transparency analysis includes at least one of the following: constructing a decision graph in a decision space, wherein the dimensions of the decision space include at least rule contribution, semantic similarity, and target matching score; constructing a rule influence heatmap in a two-dimensional matrix, wherein the dimensions of the two-dimensional matrix include scene type and rule name, and the cell coloring of the two-dimensional matrix corresponds to the average influence score of the address matching rule; and monitoring the trend of address matching success rate changes.

[0015] According to one aspect of this disclosure, an address matching system is provided, which may include: an address input module for acquiring address text to be matched; a rule matching module for performing scene recognition on the address text and performing rule matching on the address text according to the address matching rules corresponding to the scene type to obtain the rule hit score corresponding to the address text; a vector matching module for performing vectorized semantic matching on the address text to obtain at least one candidate matching address and the semantic similarity corresponding to each candidate matching address; a dynamic weighting module for performing weighted calculation on the semantic similarity and rule hit score according to the dynamic weight corresponding to the scene type to obtain the target matching score of the candidate matching address; and an address matching module for determining the target matching address from the candidate matching addresses based on the target matching score.

[0016] In an optional exemplary embodiment of this disclosure, each address matching rule is configured with a rule weight corresponding to the scene type. The rule matching module is specifically used to perform rule matching on the address text according to the address matching rule corresponding to the scene type to obtain the initial hit score corresponding to each address matching rule; and to perform weighted calculation on the initial hit score with the rule weight to obtain the rule hit score.

[0017] In an optional exemplary embodiment of this disclosure, the number of address matching rules is greater than 1. The rule matching module is specifically used to calculate a weight decay parameter based on the priority of the address matching rules; the weight decay parameter decreases as the priority decreases; the rule weight of the address matching rules is adjusted based on the weight decay parameter to obtain an effective weight; and the initial hit score is weighted and calculated using the effective weight to obtain a rule hit score.

[0018] In an optional exemplary embodiment of this disclosure, the rule matching module is specifically used to sort each address matching rule based on priority; calculate the weight decay parameter corresponding to each address matching rule in each bit order using a weight decay function; the weight decay function defines the decay gradient for each increment of the weight decay parameter in each bit order.

[0019] In an optional exemplary embodiment of this disclosure, address matching rules are divided by rule type, and each rule type has a corresponding weight configuration range; under a scenario type, each address matching rule sets a rule weight within the weight configuration range according to its rule type.

[0020] In an optional exemplary embodiment of this disclosure, the vector matching module is specifically used to perform semantic equivalence mapping and vectorization processing on the address text to obtain an initial character vector; to perform label prediction on the initial character vector using a pre-trained model to obtain a standard character vector; and to perform vectorized semantic matching based on the standard character vector to obtain at least one candidate matching address and the semantic similarity corresponding to each candidate matching address.

[0021] In an optional exemplary embodiment of this disclosure, the vector matching module is specifically used to perform semantic equivalence mapping and embedding processing on the address text to obtain a general character vector; to extract domain features from the general character vector based on the domain dictionary to obtain a domain feature vector; to extract self-attention features from the general character vector to obtain a context feature vector; and to fuse the general character vector, the domain feature vector, and the context feature vector to obtain an initial character vector.

[0022] In an optional exemplary embodiment of this disclosure, the address matching module is specifically configured to: determine a candidate matching address as a target matching address when the target matching score is greater than or equal to a first matching threshold; determine that the candidate matching address has failed to match when the target matching score is less than a second matching threshold; and review the candidate matching address and the target matching score when the target matching score is less than the first matching threshold and greater than or equal to the second matching threshold, and adjust the rule weights according to the review results.

[0023] In an optional exemplary embodiment of this disclosure, the system may further include a decision transparency enhancement module for performing decision transparency analysis on the address matching process; the decision transparency analysis includes at least one of the following: constructing a decision graph in a decision space, the dimensions of which include at least rule contribution, semantic similarity, and target matching score; constructing a rule influence heatmap in a two-dimensional matrix, the dimensions of which include scene type and rule name, and the cell coloring of the two-dimensional matrix corresponding to the average influence score of the address matching rule; and monitoring the trend of address matching success rate changes.

[0024] According to one aspect of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the address matching method described above.

[0025] According to one aspect of this disclosure, an electronic device is provided, comprising: The processor; and memory for storing the processor's executable instructions; The processor is configured to execute the address matching method described above by executing executable instructions.

[0026] This disclosure provides an address matching method, an address matching system, a computer-readable storage medium, and an electronic device. After obtaining the address text to be matched, the solution performs vectorized semantic matching on the address text to obtain candidate matching addresses and their corresponding semantic similarities. Based on scene recognition, it performs rule matching on the address text using address matching rules corresponding to the scene type to obtain a rule hit score. Furthermore, it uses dynamic weights corresponding to the scene type to weight the semantic similarity and rule hit score to obtain a target matching score, and uses this target matching score to determine the target matching address from the candidate matching addresses. This solution integrates semantic understanding and rule matching. The scene type dynamically influences the selection of address matching rules and the weight allocation of the weighted calculation, effectively improving address matching accuracy, reducing business scenario mismatches, lowering the difficulty of decision tracing and reproduction, supporting verification, and allowing for dynamic expansion, maintenance, and adaptation to changes in business needs.

[0027] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0028] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0029] Figure 1 An exemplary flowchart of the address matching method provided in this disclosure is shown.

[0030] Figure 2 An exemplary schematic diagram of the rule matching process provided in an embodiment of this disclosure is shown.

[0031] Figure 3 An exemplary schematic diagram of the vectorized semantic matching process provided in an embodiment of this disclosure is shown.

[0032] Figure 4 An exemplary schematic diagram of the address text vectorization process provided in an embodiment of this disclosure is shown.

[0033] Figure 5 This illustration shows a schematic diagram of the decision transparency analysis and address matching process optimization based on decision graphs provided in an embodiment of this disclosure.

[0034] Figure 6 An exemplary structural diagram of the address matching system provided in this disclosure is shown.

[0035] Figure 7 An electronic device for implementing an address matching method is illustrated by example according to an example embodiment of this disclosure. Detailed Implementation

[0036] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this disclosure more comprehensive and complete, and to fully convey the concept of the example embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more of the specific details omitted, or other methods, components, apparatus, steps, etc., can be employed. In other instances, well-known technical solutions are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0037] Furthermore, the accompanying drawings are merely illustrative of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0038] Conventional address matching schemes suffer from low matching accuracy due to a lack of semantic understanding, such as missing semantic equivalence understanding, misjudgment of letter case matching, and omission of administrative region level matching. Vectorized semantic matching, where the embedding representation of semantics is greatly affected by corpus quality, may also suffer from semantic equivalence misjudgment and insensitivity to missing or incorrect administrative region levels. Furthermore, the decision-making mechanism is opaque, making verification difficult. Static rule-based hybrid models suffer from rigid rules, high expansion costs, and maintenance difficulties. To address these issues, this application proposes an address matching method and an address matching system that implements this method. Based on scene recognition, it combines semantic understanding and rule matching according to different scene types, reducing the impact of corpus quality and improving the accuracy of semantic equivalence judgment. When combined with business scenarios, it can improve the perception of administrative region levels, facilitate convenient rule expansion and maintenance, and enable traceability and reproduction of the decision-making process, supporting verification. Specifically, as shown below: Figure 1 This is a flowchart illustrating the steps of the address matching method provided in an embodiment of this disclosure. Figure 1 As shown, the method may include steps 101 to 105 as follows.

[0039] Step 101: Obtain the address text to be matched.

[0040] In this embodiment of the disclosure, the source of the address text to be matched can be the input text string obtained directly, or the text string obtained through image character recognition, speech recognition, or other methods.

[0041] Step 102: Perform scene recognition on the address text, and perform rule matching on the address text according to the address matching rules corresponding to the scene type to obtain the rule hit score corresponding to the address text.

[0042] In this embodiment of the disclosure, scene recognition can be used to identify the scene type of the business applied to the address text. During the recognition process, scene recognition can be performed based on business keywords in the address text. Keywords for medical businesses can include "hospital," "surgery," "internal medicine," "operation," "inpatient department," etc., while keywords for legal businesses can include "court," "lawyer," "consultation," etc. Scene recognition can also be performed through background information associated with the address text. Depending on the source of the address text, the background information can include scene description information obtained when the address text is input. This scene description information can indicate that the address text corresponds to medical business, legal business, real estate business, e-commerce delivery business, financial risk control business, etc. It can also include other information in images and voice, such as images categorized as e-commerce parcel slip images, financial risk control ID document images, real estate registration property address information, etc., and voice can correspond to other voice text or speaker identity information, etc.

[0043] In this embodiment, address matching rules can be configured to adapt to business needs under different scenario types. For example, for e-commerce delivery business requiring precise positioning, address matching rules for key building entities can be configured; for financial risk control business requiring legal validity, address matching rules for the integrity of administrative division hierarchy can be configured; similarly, for real estate business, address matching rules for the rationality of address logical chains can be configured; and for medical business, address matching rules for multiple condition combinations can be configured. Based on this, after scene recognition of the address text, rule matching can be performed on the address text according to the address matching rules corresponding to the determined scenario type, and the rule hit score corresponding to the address text can be obtained based on the hit result. The address matching rules corresponding to different scenario types can be different, partially the same, or identical, to adapt to the common needs of address matching in different scenario types; the scenario type determined by scene recognition for the address text can be one or several, thus allowing the address text to be matched with the address matching rules of different scenario types respectively.

[0044] Step 103: Perform vectorized semantic matching on the address text to obtain at least one candidate matching address and the semantic similarity corresponding to each candidate matching address.

[0045] In this embodiment, vectorized semantic matching refers to representing the characters in the address text semantically as vectors and then performing similarity retrieval and matching within the semantic vector of the standard address. Based on this, candidate matching addresses corresponding to the address text can be obtained through filtering based on semantic similarity. For example, N candidate matching addresses can be selected from largest to smallest semantic similarity, where N is a configurable integer; or, candidate matching addresses with semantic similarity greater than or equal to a preset semantic threshold can be selected. The semantic vectorization processing and similarity retrieval matching can be implemented using appropriate character embedding, character segmentation, and character retrieval algorithms based on data processing conditions and address matching requirements; this embodiment does not impose specific limitations on these methods.

[0046] Step 104: Calculate the semantic similarity and rule hit score by weighting the dynamic weights corresponding to the scene type to obtain the target matching score of the candidate matching address.

[0047] In this embodiment of the disclosure, the semantic similarity and rule hit score obtained above can be fused by weighted calculation to obtain a target matching score that comprehensively represents the degree of matching between the candidate matching address and the address text. Among them, the dynamic weight is used to represent the importance of semantic similarity and rule hit score to the target matching score. Under different scenario types, dynamic weight configuration can be adapted to business needs. For example, in e-commerce delivery business, the dynamic weight of semantic similarity can be reduced while the dynamic weight of rule hit score can be increased to enhance the accuracy of business rules; in financial risk control business, the dynamic weight of semantic similarity can be increased while the dynamic weight of rule hit score can be reduced, and the dynamic weights of semantic similarity and rule hit score can be made close to each other to balance the matching of semantics and rules; in address cleaning business, the dynamic weight of semantic similarity can be increased while the dynamic weight of rule hit score can be reduced, and the dynamic weights of semantic similarity and rule hit score can be made significantly different to focus on semantic generalization ability and avoid omissions. For example, the calculation of the target matching score can be expressed as the following formula (1):

[0048] Where α is the dynamic weight corresponding to semantic similarity, and β is the dynamic weight corresponding to rule hit score. For example, in the scenario of logistics and delivery business, α can be 0.4 and β can be 0.6; in the scenario of financial risk control business, α can be 0.6 and β can be 0.4; in the scenario of address cleaning business, α can be 0.7 and β can be 0.3.

[0049] Step 105: Determine the target matching address from the candidate matching addresses based on the target matching score.

[0050] In this embodiment of the disclosure, based on a comprehensive characterization of the matching degree between candidate matching addresses and address text using the target matching score, the target matching address can be determined by matching the candidate matching addresses with the target matching score. For example, the candidate matching address with the highest target matching score can be selected as the target matching address; or, candidate matching addresses with target matching scores greater than or equal to a preset matching threshold can be selected as the target matching address; or, M candidate matching addresses can be selected as the target matching address based on the target matching scores from largest to smallest, where M is a configurable integer, and this embodiment of the disclosure does not impose specific limitations on this.

[0051] In this embodiment, by performing parallel semantic vector matching and rule matching on address text, collaborative decision-making is achieved. Combining semantic understanding with adaptation to complex business logic, dual filtering and comprehensive evaluation are realized. This reduces invalid data by 90% through coarse screening, lowers computational load by 72%, and improves address matching accuracy. Furthermore, it can be tailored to the business needs of different scenario types, reducing expansion and maintenance costs.

[0052] In an optional method embodiment of this disclosure, each address matching rule is configured with a rule weight corresponding to the scenario type.

[0053] In this embodiment, the rule weight represents the importance of the address matching rule in the rule hit score calculation. The rule weight of the address matching rule can be configured based on the scenario type to adapt to the specific business needs of address matching. For example, for e-commerce delivery business requiring precise positioning, the rule weight of address matching rules that match key building entities can be increased; for financial risk control business requiring legal validity, the rule weight of address matching rules that ensure the integrity of administrative division can be increased; similarly, for real estate business, the rule weight of address matching rules that ensure the rationality of the address logical chain can be increased; for medical business, the rule weight of address matching rules with multiple condition combinations can be increased, and so on. Based on this, adjusting the rule weight of the address matching rule according to the scenario type can dynamically adapt to the business needs of different scenario types, with low maintenance costs and easy expansion.

[0054] Figure 2 A schematic diagram of the rule matching process provided in this disclosure embodiment is shown, such as... Figure 2 As shown, the aforementioned step 102 may include the following steps 201 to 202.

[0055] Step 201: Perform rule matching on the address text according to the address matching rules corresponding to the scene type, and obtain the initial hit score corresponding to each address matching rule.

[0056] Step 202: Calculate the initial hit score by weighting it according to the rule weights to obtain the rule hit score.

[0057] In this embodiment of the disclosure, based on the address matching rules configured with rule weights, the address text can first be matched using the address matching rule corresponding to the scene type to obtain its corresponding initial hit score. Then, the initial hit score is further weighted and calculated using the rule weights to obtain the rule hit score. Specifically, when there is only one address matching rule, the initial hit score can be directly weighted and calculated using the rule weights to obtain the rule hit score. When there are more than one address matching rule, the initial hit scores can be weighted and summed based on their respective rule weights to obtain the rule hit score.

[0058] In an optional embodiment of the method disclosed herein, if the number of address matching rules is greater than 1, then the aforementioned step 202 may include the following steps A1 to A3.

[0059] Step A1: Calculate the weight decay parameter based on the priority of the address matching rule; the weight decay parameter decreases as the priority decreases.

[0060] Step A2: Adjust the rule weights of the address matching rules based on the weight decay parameter to obtain effective weights.

[0061] Step A3: Calculate the initial hit score using effective weights to obtain the rule hit score.

[0062] In this embodiment of the disclosure, when the number of address matching rules is greater than one, there may be a situation where multiple address matching rules accumulate, leading to an excessively high score for the hit rule. To balance the proportion of semantic similarity, corresponding priorities can be set for different address matching rules. This priority can characterize the relative importance of an address matching rule among different address matching rules. Based on this, a weight decay parameter is set to decrease as the priority decreases. Thus, the rule weights corresponding to address matching rules at different priorities can be adjusted based on the weight decay parameter. The resulting effective weights remain unchanged or change little relative to the rule weights at higher priorities, while the relative rule weights are appropriately reduced at lower priorities. This can generally control the proportion of the hit rule score in the target matching score, maintaining the expected contribution of semantic similarity to the target matching score.

[0063] In this embodiment, the weight decay parameter can be flexibly set according to data processing conditions and address matching requirements. As the priority decreases, the weight decay parameter can decrease at the same gradient, or decrease at a different gradient, or the magnitude and trend of the decrease can be freely set. This embodiment does not impose specific restrictions on this. Based on adjusting the rule weights with the weight decay parameter corresponding to the address matching rule to obtain the effective weights, the target matching score can be calculated as shown in the following formula (2):

[0064] The effective weight is obtained by adjusting the rule weights using a weight decay parameter.

[0065] In an optional embodiment of the method disclosed herein, the aforementioned step A1 may include the following steps A11 to A12.

[0066] Step A11: Sort the address matching rules according to their priority.

[0067] Step A12: Calculate the weight decay parameter corresponding to the address matching rule of each bit order using the weight decay function; the weight decay function defines the decay gradient for each increment of the bit order weight decay parameter.

[0068] In this embodiment, the address matching rules can be sorted based on priority, so that the address matching rules are arranged in priority order. As the position increases in the sequence, the priority decreases, and the weight decay parameter decreases accordingly, resulting in a greater reduction in the rule weight. The weight decay parameter can be calculated using a weight decay function, which defines the decay gradient of the weight decay parameter for each increment in position. The decay gradient can be a linear descent gradient. The effective weight is obtained by subtracting the weight decay parameter from the rule weight of each address matching rule. The weight decay function can then be expressed as the following formula (3):

[0069] Assuming the decay gradient is 0.1, the weight decay parameter is 0 when the position is 1, and the regular weight is equal to the effective weight; the weight decay parameter is 0.1 when the position is 2, and the effective weight is the regular weight minus 0.1; the weight decay parameter is 0.2 when the position is 3, and the effective weight is the regular weight minus 0.2, and so on.

[0070] The decay gradient can also be an exponential descent gradient. The effective weight is obtained by multiplying the rule weight of each address matching rule by the weight decay coefficient. The weight decay function can be expressed as the following formula (4).

[0071]

[0072] Assuming the decay gradient is 0.85, the weight decay parameter is 1 when the position is 1, and the regular weight is equal to the effective weight; the weight decay parameter is 0.85 when the position is 2, and the regular weight is multiplied by 0.85 to obtain the effective weight; the weight decay parameter is 0.7225 when the position is 3, and the regular weight is multiplied by 0.7225 to obtain the effective weight, and so on.

[0073] The aforementioned weight decay function and decay gradient are merely examples. Those skilled in the art can select, adjust, and extend the weight decay function and decay gradient according to actual needs. This disclosure does not impose specific limitations on these aspects.

[0074] For example, taking the weight decay function as shown in the aforementioned formula (4), the address matching rules corresponding to the scenario types of financial risk control business are shown in Table 1 below: Table 1

[0075] In an optional method embodiment of this disclosure, address matching rules are divided by rule type, and each rule type has a corresponding weight configuration range; under scenario type, each address matching rule sets a rule weight within the weight configuration range according to its rule type.

[0076] In this embodiment, rule types can be categorized based on the function of address matching rules. These can include keyword rules for matching key entities, such as verifying the existence of keywords like "seat" or "building"; hierarchical rules for verifying the integrity of administrative levels, such as whether a three-tiered structure of province, city, and district is included; and structural rules for verifying the rationality of address logical chains, such as whether they are logically structured using [road][address][building]. Different rule types can have corresponding weight configuration ranges. The weight configuration range can be the recommended rule weight setting range for the address matching rule; it is recommended that the rule weight set for the address matching rule fall within the weight configuration range. Based on this, the rule weight can be set according to the rule type to which the address matching rule belongs, within the weight configuration range according to the scenario type. For example, the rule types and their weight configuration ranges can be shown in Table 2 below: Table 2

[0077] Based on this, taking e-commerce logistics as an example, keyword rules include floor matching rules, building matching rules, and temperature control point matching rules. Specifically, within the range of 0.6 to 0.9, the rule weight for floor matching rules can be 0.8, the rule weight for building matching rules can be 0.9, and the rule weight for temperature control point matching rules can be 0.7, all set at a relatively high level to achieve accurate positioning. Similarly, financial risk control can set the rule weight of hierarchical rules at a relatively high level between 0.7 and 1.0, such as setting the rule weight of administrative division verification rules to 0.80; real estate registration can set the rule weight of structural rules at a relatively high level between 0.5 and 0.8, such as setting the rule weight of address chain logic verification rules to 0.75; and medical emergency rescue can set the rule weight of composite rules at a relatively high level between 0.8 and 1.0, such as setting the rule weight of key rules to 0.95. The above weight configuration ranges and rule weights are only examples. Those skilled in the art can adjust the weight configuration ranges and the values ​​of rule weights within the range according to actual needs. This disclosure does not impose specific limitations on these aspects.

[0078] In this embodiment of the disclosure, different scene types can be configured with corresponding rule bases, so that the rule base of the corresponding scene type can be directly called for rule matching based on scene recognition.

[0079] In an optional method embodiment of this disclosure, Figure 3 A schematic diagram of the vectorized semantic matching process provided in an embodiment of this disclosure is shown, such as... Figure 3 As shown, the aforementioned step 103 may include the following steps 301 to 303.

[0080] Step 301: Perform semantic equivalence mapping and vectorization processing on the address text to obtain an initial character vector.

[0081] In the embodiments of the present disclosure, during the process of vectorized semantic matching, the obtained address text can be vectorized first to represent the address text through the initial character vector, which is convenient for subsequent model processing, analysis, similarity calculation, and retrieval matching. Among them, the vectorization processing can be implemented by an algorithm and model for embedding text characters. Those skilled in the art can select according to data processing conditions and address matching requirements, and the embodiments of the present disclosure do not limit this.

[0082] In the embodiments of the present disclosure, before the vectorization processing, semantic equivalence mapping can also be performed on the characters in the address text to convert characters with the same or similar semantics into mapped standard words, such as performing semantic equivalence mapping on characters such as "No. = Lane = Building". On this basis, the foregoing vectorization processing is performed to obtain an initial character vector.

[0083] Step 302: Use a pre-trained model to perform label prediction on the initial character vector to obtain a standard character vector.

[0084] In the embodiments of the present disclosure, the pre-trained model can perform further inference analysis such as encoding, word segmentation, and label prediction on the initial character vector to improve the accuracy of semantic similarity calculation and matching. Exemplarily, the pre-trained model can adopt a BiLSTM-CRF (Bidirectional Long Short-Term Memory Network - Conditional Random Field) word segmentation model, and can also be adjusted for address matching, configure an address dictionary containing 12.7 address entities such as roads, buildings, and administrative regions, and provide an ambiguity resolution mechanism to solve the word segmentation ambiguity between "Courtyard No. 12 on Road B" and "No. 12 on Road B / Courtyard", so that the word segmentation accuracy rate reaches 98.3%.

[0085] Step 303: Perform vectorized semantic matching based on the standard character vector to obtain at least one candidate matching address and the semantic similarity corresponding to each candidate matching address.

[0086] In the embodiments of the present disclosure, based on the obtained standard character vector, vectorized semantic matching can be performed to calculate the similarity between the standard character vector and the feature vector of the preset standard address, so as to determine at least one candidate matching address whose semantic similarity meets the candidate conditions in the preset standard address, and the semantic similarity corresponding to the candidate matching address. Among them, the semantic similarity can be realized by calculating the cosine similarity, or other types of similarities can be calculated. The embodiments of the present disclosure do not make specific limitations on this.

[0087] Exemplarily, the HNSW algorithm (Hierarchical Navigable Small World) can be used for vectorized semantic matching. With a preset semantic threshold of 0.65, the time-consuming of address retrieval can be effectively improved through graph-based approximate nearest neighbor search. Based on this, candidate matching addresses with a cosine similarity greater than or equal to 0.65 are screened. Those skilled in the art can also select other retrieval matching algorithms according to data processing conditions and address matching requirements, and the embodiments of the present disclosure do not make specific limitations thereto.

[0088] In an optional method embodiment of the present disclosure, during the process of vectorized representation of address text, feature enhancement can also be performed to enrich the dimensions and content of character representation in the address text. Figure 4 The flowchart of vectorizing the address text provided by the embodiment of the present disclosure is shown, as Figure 4 shown, the foregoing step 301 may include the following steps 401 to step 404.

[0089] Step 401: Perform semantic equivalence mapping and embedding processing on the address text to obtain a general character vector.

[0090] In the embodiment of the present disclosure, the semantic equivalence mapping can be correspondingly referred to the relevant description in the foregoing step 301. To avoid repetition, it will not be elaborated here. The embedding processing can be realized by mapping and transforming the address text through an embedding layer, and a general character vector corresponding to each character in the address text is obtained. This embedding layer can be implemented by using pre-trained models such as BERT (Bidirectional Encoder Representation from Transformers) and BGE (BAAI General Embedding). For example, the BGE model obtained by using 6.8 million Chinese addresses covering 283 city variants as training data is used to perform embedding processing on the address text. Usually, the dimension of the general character vector can be 256 dimensions or 768 dimensions.

[0091] Exemplarily, taking the dimension of the general character vector as 4 dimensions, the character "Su" in the address text, through vectorization processing, the obtained general character vector can be expressed as V_general = [0.1, 0.5, -0.3, 0.8].

[0092] Step 402: Extract domain features from the general character vector based on a domain dictionary to obtain a domain feature vector.

[0093] In the embodiments of the present disclosure, the domain dictionary contains address entries specifically configured in the address matching domain. By matching the address entries for annotation, the domain feature embedding of the general character vector is achieved, thereby supplementing relevant domain knowledge. Based on the obtained general character vector, domain feature enhancement can be performed on the general character vector based on the domain dictionary to obtain the corresponding domain feature vector.

[0094] Exemplarily, the domain dictionary includes the address entry "Suzhou Street". By matching the character "Su", it is determined that "Su" is the beginning part of the address entry "Suzhou Street", and then "Su" is annotated with the label B-Road, indicating the start of the road name. On this basis, the vector representation of the label B-Road is used as the domain feature vector. Assuming it is 4-dimensional, the domain feature vector of the label B-Road can be represented as V_dict = [0.9, -0.1, 0.2, 0.0].

[0095] Step 403: Perform self-attention feature extraction on the general character vector to obtain the context feature vector.

[0096] In the embodiments of the present disclosure, self-attention feature extraction can also be performed. The calculation of self-attention features can dynamically adjust the degree of attention to characters, consider the dependency relationships between characters at different positions in the address text, and represent the context semantics of the general character vector to obtain the context feature vector.

[0097] Exemplarily, the general character vector V_general of the character "Su" can be input into the lightweight self-attention layer. In the self-attention layer, self-attention feature extraction is performed to analyze the character sequence "Suzhou Street" of the address text, capture the importance of the characters "Zhou" and "Jie" for understanding the semantics of the character "Su", and calculate the context feature vector of the character "Su" based on the fact that they jointly form a complete place name. Assuming it is 4-dimensional, the context feature vector can be represented as V_attn = [-0.2, 0.6, 0.1, 0.5]. This context feature vector represents the semantics of the character "Su" in the context of the place name "Suzhou Street".

[0098] Step 404: Fuse the general character vector, the domain feature vector, and the context feature vector to obtain the initial character vector.

[0099] In the embodiments of the present disclosure, based on obtaining the domain feature vector and the context feature vector, they can be fused into the general character vector. By supplementing domain knowledge and context knowledge, the character representation can be feature-enhanced to obtain an initial character vector that more comprehensively and fully represents the character features. Among them, the fusion can be vector concatenation, weighted fusion, or other fusion methods; in addition, those skilled in the art can also expand other feature enhancement schemes by referring to the foregoing disclosed domain feature vectors, context feature vectors, etc., and the embodiments of the present disclosure do not specifically limit this.

[0100] Exemplarily, based on the obtained general character vector V_general = [0.1, 0.5, -0.3, 0.8], which represents the general semantics of the character "Su"; the domain feature vector V_dict = [0.9, -0.1, 0.2, 0.0], which represents that the character "Su" is the start of the road name "Suzhou Street"; the context feature vector V_attn = [-0.2, 0.6, 0.1, 0.5], which represents that the character "Su" is the first character in the road name 'Suzhou Street'. Through the concatenation fusion operation, the obtained 12-dimensional initial character vector is represented as V_enhanced = concatenate(V_general, V_dict, V_attn) = [0.1, 0.5, -0.3, 0.8, 0.9, -0.1, 0.2, 0.0, -0.2, 0.6, 0.1, 0.5].

[0101] On this basis, this 12-dimensional initial character vector can be input into the BiLSTM-CRF for further encoding and label prediction. Thus, when receiving the character "Su", the BiLSTM-CRF can fully understand the domain and context information of the character "Su", so as to be able to more efficiently and accurately determine that the label to which the character "Su" belongs can be B-Road, rather than other labels, for efficient encoding.

[0102] In an optional method embodiment of the present disclosure, the foregoing step 105 may include the following steps B1 to B3.

[0103] Step B1: When the target matching score is greater than or equal to the first matching threshold, determine the candidate matching address as the target matching address.

[0104] Step B2: When the target matching score is less than the second matching threshold, determine that the candidate matching address fails to match.

[0105] Step B3: When the target matching score is less than the first matching threshold and greater than or equal to the second matching threshold, review the candidate matching address and the target matching score, and adjust the rule weight according to the review result.

[0106] In this embodiment, based on the obtained target matching score, a threshold judgment can be made on the target matching score to determine whether to output the target matching address. Based on this, processing schemes corresponding to different threshold levels can be set to achieve hierarchical decision-making. For example, three processing schemes can be divided by setting a first matching threshold and a second matching threshold from large to small. When the target matching score is greater than or equal to the first matching threshold, the matching degree between the candidate matching address and the address text is considered to meet expectations, thus determining the candidate matching address as the target matching address. When the target matching score is less than the second matching threshold, the matching degree between the candidate matching address and the address text is determined to be unsatisfactory, indicating a matching failure, thus rejecting the output of the candidate matching address. When the target matching score is less than the first matching threshold and greater than or equal to the second matching threshold, the matching degree between the candidate matching address and the address text has a probability of meeting expectations, so a review can be performed, and the rule weights of the aforementioned address matching rules can be adjusted based on the review results to improve the business adaptability of the address matching rules. The review can be based on static rules and model inference to evaluate whether the candidate matching address text and the target matching score representation are accurate, or it can be performed manually; this embodiment does not impose specific limitations on this.

[0107] The address matching method disclosed herein can significantly improve the matching accuracy to 68% to 76% compared to conventional solutions. It can also flexibly adapt to changes in scenario types and effectively reduce the cost of rule expansion. Furthermore, hierarchical decision-making can improve review efficiency and optimize the address matching process.

[0108] In an optional embodiment of the method disclosed herein, the method may further include the following step C.

[0109] Step C: Conduct a transparent decision-making analysis of the address matching process.

[0110] In this embodiment of the disclosure, for the address matching process, decision transparency analysis can be performed on any step or the entire process, thereby increasing the amount of information provided for address matching, supporting higher compliance assurance, and also improving the risk detection rate.

[0111] In an optional embodiment of the method disclosed herein, decision transparency analysis includes constructing a decision graph in a decision space, wherein the dimensions of the decision space include at least rule contribution, semantic similarity, and target matching score.

[0112] In this embodiment, decision transparency analysis may include constructing a decision graph, which can visualize the complex decision-making process in address matching as an intuitive spatial model. The decision graph can be constructed within a decision space, which may include at least different dimensions such as rule contribution, semantic similarity, and target matching score. For example, rule contribution can be the X-axis, representing the aggregated value of address matching rule contributions, such as a weighted normalized rule hit score; semantic similarity can be the Y-axis, representing the similarity setting obtained from vectorized semantic matching calculations, ranging from [0, 1]; and target matching score can be the Z-axis, representing the final comprehensive score. Based on the coordinate points of the aforementioned three-dimensional representation, other scene types, timestamps, address text, and other address matching metadata may also be included. On this basis, a decision space is first constructed, and visualization mapping relationships are defined, such as the mapping relationship between the result status of the address matching process and color representation, and the mapping relationship between different weights and graph size. A reference plane and region division are further loaded into the decision space. Thus, the rule contribution, semantic similarity, and target matching score in the address matching process can be plotted onto the decision space to provide a decision graph.

[0113] Figure 5 This illustration shows a flowchart of decision transparency analysis based on decision graphs and address matching process optimization provided in an embodiment of this disclosure. Figure 5 As shown, the address matching process may include the following steps 501 to 506.

[0114] Step 501: Obtain metadata for the address matching process.

[0115] Step 502: Extract dimensional data from the metadata, including rule contribution, semantic similarity, and target matching score, as well as additional data including scene type, timestamp, and address text.

[0116] Step 503: Generate a three-dimensional decision point based on rule contribution, semantic similarity, and target matching score, and generate attribute information for the three-dimensional decision point with additional data.

[0117] Step 504: Load the reference plane and region division of the decision space, draw the three-dimensional decision point in the decision space based on the visualization mapping relationship, and display the corresponding decision map in response to the view filtering conditions indicated by the user.

[0118] Step 505: In response to the user's selection of a 3D decision point, display the attribute information and dimensional data of that 3D decision point. If manual review is triggered, proceed to step 506.

[0119] Step 506: Perform a manual review of the address matching process corresponding to the three-dimensional decision point, and adjust the address matching process based on the review results.

[0120] In an optional embodiment of the method disclosed herein, decision transparency analysis includes constructing a rule influence heatmap in a two-dimensional matrix, wherein the dimensions of the two-dimensional matrix include scene type and rule name, and the cell coloring of the two-dimensional matrix corresponds to the average influence score of the address matching rule.

[0121] In this embodiment of the disclosure, the rule influence heatmap is used to identify the effect of address matching rules, distinguish between efficient rules that play a role in address matching decisions, and invalid rules that actually contribute little or no to the decision, or conflicting rules that actually have a counterproductive effect, so as to support the optimization and maintenance of address matching rules. The rule influence heatmap can be constructed on a two-dimensional matrix. For example, the X-axis can be configured as the scenario type, such as ["logistics delivery", "financial risk control", "medical emergency", "government census", ...], and the Y-axis can be configured as the name of the address matching rule, such as ["administrative district planning", "house number rule", "building rule", "synonym rule", ...]. In the two-dimensional matrix, the average influence score of a specific address matching rule can be represented by the grid coloring. Among them, the average influence score Avg_Influence can be represented by aggregated statistical values ​​and calculated as follows (5):

[0122] The number of matches refers to the total number of times the address matching rule participates in the matching process.

[0123] For example, the grid color can be dark red corresponding to a higher average influence score, indicating that the address matching rule is more effective in this scenario type; light yellow or white corresponding to a lower average influence score, indicating that the address matching rule is less effective in this scenario type, or rarely triggered.

[0124] In an optional embodiment of the method disclosed herein, decision transparency analysis includes monitoring the trend of changes in the success rate of address matching.

[0125] In this embodiment, the success rate of address matching can also be monitored to promptly detect abnormal situations such as a large increase in the number of failed address matching cases or a sudden drop in the success rate. At this time, an alarm can be issued, a review process can be automated, or address matching can be actively interrupted. For example, an alarm threshold can be configured to calculate the address matching failure rate based on the number of failed address matching cases and the number of address matching requests per unit time. An alarm can be triggered after the address matching failure rate exceeds the alarm threshold for a sustained period, enabling timely problem localization, root cause analysis, and operational maintenance. For example, the alarm content could be "The matching failure rate has reached 35.2% in the last 5 minutes."

[0126] This disclosure provides an address matching method. After obtaining the address text to be matched, the method performs vectorized semantic matching on the address text to obtain candidate matching addresses and their corresponding semantic similarities. Based on scene recognition, it performs rule matching on the address text according to the address matching rules corresponding to the scene type to obtain a rule hit score. Furthermore, it uses dynamic weights corresponding to the scene type to weight the semantic similarity and rule hit score to obtain a target matching score, and uses this target matching score to determine the target matching address from the candidate matching addresses. This method integrates semantic understanding and rule matching. The scene type dynamically influences the selection of address matching rules and the weight allocation of the weighted calculation, effectively improving address matching accuracy, reducing business scenario mismatches, reducing the difficulty of decision tracing and reproduction, supporting verification, and allowing for easy dynamic expansion, maintenance, and adaptation to changes in business needs.

[0127] The following are system embodiments of this disclosure, which can be used to execute the method embodiments of this disclosure. For details not disclosed in the device embodiments of this disclosure, please refer to the method embodiments of this disclosure.

[0128] This disclosure also provides an example structural diagram of an address matching system 600. Specifically, refer to... Figure 6 As shown, the system 600 may include: an address input module 601, used to acquire the address text to be matched; a rule matching module 602, used to perform scene recognition on the address text and perform rule matching on the address text according to the address matching rules corresponding to the scene type, to obtain the rule hit score corresponding to the address text; a vector matching module 603, used to perform vectorized semantic matching on the address text, to obtain at least one candidate matching address and the semantic similarity corresponding to each candidate matching address; a dynamic weighting module 604, used to perform weighted calculation on the semantic similarity and rule hit score according to the dynamic weight corresponding to the scene type, to obtain the target matching score of the candidate matching address; and an address matching module 605, used to determine the target matching address among the candidate matching addresses based on the target matching score.

[0129] In an optional system embodiment of this disclosure, each address matching rule is configured with a rule weight corresponding to the scene type. The rule matching module 602 is specifically used to perform rule matching on the address text according to the address matching rule corresponding to the scene type to obtain the initial hit score corresponding to each address matching rule; and to perform weighted calculation on the initial hit score according to the rule weight to obtain the rule hit score.

[0130] In an optional system embodiment of this disclosure, the number of address matching rules is greater than 1. The rule matching module 602 is specifically used to calculate a weight decay parameter based on the priority of the address matching rules; the weight decay parameter decreases as the priority decreases; the rule weight of the address matching rules is adjusted based on the weight decay parameter to obtain an effective weight; and the initial hit score is weighted and calculated using the effective weight to obtain a rule hit score.

[0131] In an optional system embodiment of this disclosure, the rule matching module 602 is specifically used to sort each address matching rule based on priority; calculate the weight decay parameter corresponding to each address matching rule in each bit order using a weight decay function; the weight decay function defines the decay gradient for each increment of the weight decay parameter in each bit order.

[0132] In an optional system embodiment of this disclosure, address matching rules are divided by rule type, and each rule type has a corresponding weight configuration range; under scenario type, each address matching rule sets a rule weight within the weight configuration range according to its rule type.

[0133] In an optional system embodiment of this disclosure, the vector matching module 603 is specifically used to perform semantic equivalence mapping and vectorization processing on the address text to obtain an initial character vector; to use a pre-trained model to predict the label of the initial character vector to obtain a standard character vector; and to perform vectorized semantic matching based on the standard character vector to obtain at least one candidate matching address and the semantic similarity corresponding to each candidate matching address.

[0134] In an optional system embodiment of this disclosure, the vector matching module 603 is specifically used to perform semantic equivalence mapping and embedding processing on the address text to obtain a general character vector; to extract domain features from the general character vector based on the domain dictionary to obtain a domain feature vector; to extract self-attention features from the general character vector to obtain a context feature vector; and to fuse the general character vector, the domain feature vector, and the context feature vector to obtain an initial character vector.

[0135] In an optional system embodiment of this disclosure, the address matching module 605 is specifically used to determine a candidate matching address as a target matching address when the target matching score is greater than or equal to a first matching threshold; to determine that the candidate matching address has failed to match when the target matching score is less than a second matching threshold; and to review the candidate matching address and the target matching score when the target matching score is less than the first matching threshold and greater than or equal to the second matching threshold, and to adjust the rule weights according to the review results.

[0136] In an optional system embodiment of this disclosure, the system may further include a decision transparency enhancement module for performing decision transparency analysis on the address matching process; the decision transparency analysis includes at least one of the following: constructing a decision graph in a decision space, the dimensions of which include at least rule contribution, semantic similarity, and target matching score; constructing a rule influence heatmap in a two-dimensional matrix, the dimensions of which include scene type and rule name, and the color of the cells in the two-dimensional matrix corresponding to the average influence score of the address matching rule; and monitoring the trend of address matching success rate changes.

[0137] This disclosure provides an address matching system. After obtaining the address text to be matched, the system performs vectorized semantic matching on the address text to obtain candidate matching addresses and their corresponding semantic similarities. Based on scene recognition, it performs rule matching on the address text using address matching rules corresponding to the scene type to obtain a rule hit score. Furthermore, it uses dynamic weights corresponding to the scene type to weight the semantic similarity and rule hit score to obtain a target matching score, and uses this target matching score to determine the target matching address from the candidate matching addresses. This system integrates semantic understanding and rule matching. The scene type dynamically influences the selection of address matching rules and the weight allocation of the weighted calculation, effectively improving address matching accuracy, reducing business scenario mismatches, lowering the difficulty of decision tracing and reproduction, supporting verification, and allowing for dynamic expansion, maintenance, and adaptation to changes in business needs.

[0138] The specific details of each module in the above address matching system have been described in detail in the corresponding address matching methods, so they will not be repeated here.

[0139] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0140] Furthermore, although the steps of the method in this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or a step may be broken down into multiple steps.

[0141] In exemplary embodiments of this disclosure, an electronic device capable of implementing the above-described methods is also provided. Those skilled in the art will understand that various aspects of this disclosure can be implemented as systems, methods, or program products. Therefore, various aspects of this disclosure can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or an implementation combining hardware and software aspects, collectively referred to herein as a "circuit," "module," or "system."

[0142] The following reference Figure 7 To describe an electronic device 700 according to such an embodiment of the present disclosure. Figure 7 The electronic device 700 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0143] like Figure 7 As shown, the electronic device 700 is manifested in the form of a general-purpose computing device. The components of the electronic device 700 may include, but are not limited to: at least one processing unit 710, at least one storage unit 720, a bus 730 connecting different system components (including storage unit 720 and processing unit 710), and a display unit 740.

[0144] The storage unit stores program code that can be executed by the processing unit 710, causing the processing unit 710 to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this disclosure. For example, the processing unit 710 can perform actions such as... Figures 1 to 5 Any of the address matching methods shown in the table.

[0145] Storage unit 720 may include readable media in the form of volatile storage units, such as random access memory (RAM) 7201 and / or cache memory 7202, and may further include read-only memory (ROM) 7203.

[0146] The storage unit 720 may also include a program / utility 7204 having a set (at least one) program module 7205, such program module 7205 including but not limited to: an operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0147] Bus 730 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0148] Electronic device 700 can also communicate with one or more external devices 800 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 700, and / or with any device that enables electronic device 700 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 750. Furthermore, electronic device 700 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 760. As shown, network adapter 760 communicates with other modules of electronic device 700 via bus 730. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 700, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0149] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal device, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0150] In exemplary embodiments of this disclosure, a computer-readable storage medium is also provided, on which a program product capable of implementing the methods described above is stored. In some possible implementations, various aspects of this disclosure may also be implemented as a program product including program code that, when the program product is run on a terminal device, causes the terminal device to perform the steps of the various exemplary embodiments of this disclosure described in the "Exemplary Methods" section above.

[0151] The program product for implementing the above-described method according to embodiments of the present disclosure may employ a portable compact disc read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto. In this document, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0152] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0153] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting programs for use by or in conjunction with an instruction execution system, apparatus, or device.

[0154] The program code contained on the readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0155] Program code for performing the operations of this disclosure can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0156] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of this disclosure and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0157] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention described herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not invented by this disclosure. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

Claims

1. An address matching method, characterized in that, The method includes: Get the address text to be matched; The address text is subjected to scene recognition, and the address text is matched with the rule corresponding to the scene type to obtain the rule hit score corresponding to the address text; The address text is subjected to vectorized semantic matching to obtain at least one candidate matching address and the semantic similarity corresponding to each candidate matching address; The semantic similarity and the rule hit score are weighted and calculated using the dynamic weights corresponding to the scene type to obtain the target matching score of the candidate matching address. The target matching address is determined from the candidate matching addresses based on the target matching score.

2. The method according to claim 1, characterized in that, Each address matching rule is configured with a corresponding rule weight. The step of matching the address text with the address matching rule corresponding to the scene type to obtain the rule hit score corresponding to the address text includes: The address text is matched according to the address matching rules corresponding to the scene type to obtain the initial hit score corresponding to each address matching rule; The initial hit score is weighted by the rule weights to obtain the rule hit score.

3. The method according to claim 2, characterized in that, The number of address matching rules is greater than 1. The step of weighting the initial hit score with the rule weights to obtain the rule hit score includes: The weight decay parameter is calculated based on the priority of the address matching rule; the weight decay parameter decreases as the priority decreases. The weights of the address matching rules are adjusted based on the weight decay parameter to obtain effective weights; The initial hit score is weighted using the effective weights to obtain the rule hit score.

4. The method according to claim 3, characterized in that, The step of calculating the weight decay parameter based on the priority of the address matching rule includes: The address matching rules are sorted based on their priority. The weight decay parameter corresponding to the address matching rule for each bit sequence is calculated using a weight decay function; the weight decay function defines the decay gradient of the weight decay parameter for each increment of the bit sequence.

5. The method according to claim 2, characterized in that, The address matching rules are divided into rule types, and each rule type has a corresponding weight configuration range. Under the aforementioned scenario type, each address matching rule sets its weight within the weight configuration range according to its corresponding rule type.

6. The method according to claim 1, characterized in that, The step of performing vectorized semantic matching on the address text to obtain at least one candidate matching address and the semantic similarity corresponding to each candidate matching address includes: The address text is semantically equivalent mapped and vectorized to obtain an initial character vector; A pre-trained model is used to predict the labels of the initial character vectors to obtain standard character vectors. Vectorized semantic matching is performed based on the standard character vector to obtain at least one candidate matching address and the semantic similarity corresponding to each candidate matching address.

7. The method according to claim 6, characterized in that, The step of performing semantic equivalence mapping and vectorization on the address text to obtain an initial character vector includes: The address text is subjected to semantic equivalence mapping and embedding processing to obtain a general character vector; Based on the domain dictionary, the general character vector is subjected to domain feature extraction to obtain a domain feature vector; Self-attention feature extraction is performed on the general character vector to obtain the context feature vector; The initial character vector is obtained by fusing the general character vector, the domain feature vector, and the context feature vector.

8. The method according to claim 2, characterized in that, Determining the target matching address from the candidate matching addresses based on the target matching score includes: If the target matching score is greater than or equal to the first matching threshold, the candidate matching address is determined as the target matching address; If the target matching score is less than the second matching threshold, it is determined that the candidate matching address has failed to match. If the target matching score is less than the first matching threshold and greater than or equal to the second matching threshold, the candidate matching address and the target matching score are reviewed, and the rule weights are adjusted according to the review results.

9. The method according to claim 1, characterized in that, The method further includes: Conduct a transparent analysis of the address matching process; The decision transparency analysis includes at least one of the following: Construct a decision graph in a decision space, wherein the dimensions of the decision space include at least rule contribution, semantic similarity, and target matching score; A rule influence heatmap is constructed in a two-dimensional matrix, wherein the dimensions of the two-dimensional matrix include scene type and rule name, and the cell color of the two-dimensional matrix corresponds to the average influence score of the address matching rule; Trends in the success rate of monitoring address matching.

10. An address matching system, characterized in that, The system includes: The address input module is used to obtain the address text to be matched; The rule matching module is used to perform scene recognition on the address text, and to match the address text with the address matching rules corresponding to the scene type to obtain the rule hit score corresponding to the address text; The vector matching module is used to perform vectorized semantic matching on the address text to obtain at least one candidate matching address and the semantic similarity corresponding to each candidate matching address. The dynamic weighting module is used to calculate the semantic similarity and the rule hit score by weighting the scene type with dynamic weights to obtain the target matching score of the candidate matching address. The address matching module is used to determine the target matching address from the candidate matching addresses based on the target matching score.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the address matching method according to any one of claims 1-9.

12. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the address matching method of any one of claims 1-9 by executing the executable instructions.

Images