Address unit analysis method and device based on regulation and control threshold iterative learning

Abstract

The invention relates to the technical field of address unit analysis, and provides an address unit analysis method and device based on regulation and control threshold iterative learning, and the address unit analysis method based on regulation and control threshold iterative learning comprises the following steps: dividing an address recognition data set; training an address unit labeling model; iteratively optimizing the address unit labeling model based on the labeling result and the regulation and control threshold value; labeling a test set based on a voting method; and extracting the address units in the test set. The device comprises an address unit data set division component, an address unit labeling model training component, an address unit standard model iterative optimization component, an address unit labeling component and an address unit extraction component. According to the method, the credibility of a sample amplification result can be improved, the amplification speed can be ensured, the quality of a training set is further improved, a slow-amplification iteration process is avoided, and the address unit analysis effect can be improved from the aspects of precision and speed.

Description

Technical Field

[0001] The present invention relates to the technical field of address unit parsing, and in particular to an address unit parsing method and device based on iterative learning of a control threshold. Background Art

[0002] Address information is widely present in various texts such as insurance policies, express delivery orders, work orders, etc. Address unit parsing can extract various spatial location elements from the above texts, including: province, city, county, township, village, road, building, house number, direction, distance, etc., which is of great significance and value in daily life, commercial activities, government office, security control and other fields.

[0003] Commonly used address unit parsing methods include two categories: matching-based methods and probability-statistics-based methods. The former has a strong ability to identify existing address units, but lacks learning ability and cannot identify unknown address units; the latter uses labeled data to train models with learning ability and can identify unknown address units. However, the effect of probability-statistics-based methods is based on a large amount of labeled data, and the cost of manually labeling data is high; in actual applications, there are relatively few labeled training samples, and there are a large number of test samples without labeled information. The data distribution characteristics of the two may not be exactly the same, and the trained model is difficult to accurately identify each unlabeled test sample, which affects the effect of address unit parsing.

[0004] To address this problem, Vapink proposed a direct learning method, which is essentially semi-supervised learning. This method assumes that unlabeled test samples have the same independent and identically distributed characteristics of data sources as labeled training samples, and its own implicit information helps to improve the prediction effect of the model. This method uses an existing model to machine-label test samples, regards the labeling results with better quality as credible samples, mixes them with the original labeled data set to form a new training set, and trains again to obtain a new model. Repeatedly iterating the above process can continuously increase the size of the training set. Among them, "good quality" is the standard for screening machine labeling results. Usually, multiple models are used to predict at the same time, and test samples with the same prediction results are regarded as credible samples and added to the training set. Therefore, in address unit parsing, the test set is processed by direct learning, which realizes the automatic labeling of address units and reduces the burden of manual labeling; adding "good quality" test samples to the training set can increase the size of the training set, thereby improving the prediction quality of the address unit labeling model; this method is suitable for scenarios with a small number of training sets and a large amount of data to be tested in address unit parsing, and can effectively alleviate the overfitting problem caused by the small size of the training set and the inability to cover the distribution space of the test set.

[0005] This direct learning method generally uses multiple address unit annotation models to make predictions at the same time, and adds test samples with the same prediction results to the training set; however, it does not consider the quality of the model. If the quality of most models is not good, then the prediction results are not reliable, and the quality of the training set will be reduced after amplification. In addition, this method requires iterative execution. As the number of iterations increases, the number of reliable samples amplified in each round gradually decreases, but the computational cost consumed in each round hardly changes. The benefits brought by sample amplification gradually decrease, and the cost-effectiveness also decreases, resulting in a waste of resources and violating the principle of reducing costs and increasing efficiency. Summary of the invention

[0006] Based on this, in order to solve the problem that the quality of the training set may be reduced after the amplification of the address unit parsing method based on direct induction learning because the quality of the model is not considered, the present invention provides an address unit parsing method and device based on iterative learning of control thresholds, and its specific technical solution is as follows:

[0007] An address unit parsing method based on iterative learning of control thresholds comprises the following steps:

[0008] S1: partition address recognition dataset;

[0009] S2: training address unit annotation model;

[0010] S3: Iteratively optimize the address unit annotation model based on the annotation results and control thresholds;

[0011] S4: Annotate the test set based on voting method;

[0012] S5: Extract the address units in the test set.

[0013] Preferably, the specific method of dividing the address recognition data set includes the following steps:

[0014] The address recognition dataset is divided into a training set ADS train and test set ADS test ;

[0015] Among them, the training set ADS train The test set ADS consists of text containing address unit annotation information. test Contains text only, no address unit labeling information.

[0016] Preferably, the specific method of training the address unit labeling model includes the following steps:

[0017] Using the training set ADS train Train M address unit annotation models i ;

[0018] Among them, 1≤i≤M.

[0019] Preferably, the specific method for iteratively optimizing the address unit annotation model based on the annotation results and the control threshold comprises the following steps:

[0020] Controlling the training set ADS according to the regulation threshold train Iterative amplification process and optimize the address unit labeling model;

[0021] Among them, the regulation threshold includes a score threshold, an amplification threshold and a round number threshold.

[0022] Preferably, the training set ADS is controlled according to the regulation threshold train The specific method of iterative expansion process and optimizing the address unit labeling model includes the following steps:

[0023] S31: Initialize I=0, read the score threshold δ S , amplification threshold δ N , round number threshold δ I ;

[0024] S32: Initialize N=0;

[0025] S33: Read test set ADS test The first text to be processed Str;

[0026] S34: Use address units to label the model i Process Str, 1≤i≤M, and obtain M address types and scores for each address unit;

[0027] S35: Select M addresses with the same type and scores exceeding δ S The address unit;

[0028] S36: Count the number n of selected address units;

[0029] S37: If n≥1, go to S38, otherwise go to S3D;

[0030] S38: considering the selected n address units as trusted address units;

[0031] S39: Combine the trusted address unit and Str into a trusted sample;

[0032] S3A: Adding credible samples to the training set ADS train ;

[0033] S3B: From the test set ADS test Delete Str from

[0034] S3C: Update the number of amplification samples in the current round N=N+1;

[0035] S3D: If ADS test If there is text to be processed, then go to S3E, otherwise go to S3F;

[0036] S3E: read the next text Str to be processed, and go to S34;

[0037] S3F: Using ADS train Training address unit annotation model i , 1≤i≤M;

[0038] S3G: update iteration number I = I + 1;

[0039] S3H: If N ≥ δ N And I≤δ I , then go to S32, otherwise go to S3I;

[0040] S3I: Return to ADS train 、ADS test and Model i , 1≤i≤M;

[0041] Among them, Str represents the text in the test set, N represents the number of amplified samples in the current round, and I represents the number of iterations.

[0042] Preferably, the specific method for labeling the test set based on the voting method comprises the following steps:

[0043] S41: Read test set ADS test The first text to be processed Str;

[0044] S42: Use address units to label the model i Process Str, 1≤i≤M, and obtain M temporary address types for each address unit;

[0045] S43: using a voting method to fuse the processing results of different address unit annotation models to obtain a final address type for each address unit;

[0046] S44: If ADS test If there is text to be processed, go to S45, otherwise go to S46;

[0047] S45: read the next text Str to be processed, and go to S42;

[0048] S46: Return to ADS test The address unit marking results.

[0049] Preferably, the specific method for extracting address units in the test set includes the following steps:

[0050] According to the test set ADS test The address unit annotation results are extracted from the text to extract the address unit entities and their types.

[0051] An address unit parsing device based on iterative learning of control thresholds, comprising:

[0052] An address unit data set division component M1, used for dividing a data set containing address units into a training set and a test set;

[0053] An address unit annotation model training component M2, used for training multiple address unit annotation models;

[0054] The address unit annotation model iterative optimization component M3 is used to iteratively expand the training set according to the annotation results and the control threshold, and continuously optimize the address unit annotation model;

[0055] The address unit marking component M4 is used to obtain the marking results of the address units in the test sample by using a voting method;

[0056] The address unit extraction component M5 is used to extract the address unit entity and its type from the address unit marking result.

[0057] Preferably, the address unit annotation model iterative optimization component M3 specifically includes:

[0058] Iterative optimization parameter initialization component M31, used to set initial parameters when iteratively optimizing the address unit annotation model based on the annotation results and the control threshold;

[0059] Iterative optimization process control component M32, used to control the normal execution and end of the iterative optimization address unit annotation model process based on the annotation results and the regulation threshold;

[0060] A first-category address unit annotation result generating component M33, used to generate address types and scores of address units in a test sample using multiple address unit annotation models;

[0061] A trusted sample generating component M34, used to generate a trusted sample according to the quality of the address unit annotation result;

[0062] The training set augmentation component M35 is used to transfer the credible samples in the test set to the training set.

[0063] Preferably, the address unit marking component M4 specifically includes:

[0064] A second type address unit labeling result generating component M41, used to generate a temporary address type of an address unit in a test sample by using multiple address unit labeling models;

[0065] The address unit labeling result fusion component M42 is used to determine the final address type of the address unit in the test sample by using a voting method.

[0066] The present invention provides an address unit parsing method and device based on iterative learning of control thresholds, and the beneficial effects thereof are as follows:

[0067] 1. Establish a reliable sample generation method based on score threshold: According to the positive correlation between the prediction score and the credibility of the prediction result, the present invention only selects the address units whose prediction score exceeds the threshold and discards the address units whose quality does not meet the requirements. This can improve the credibility of the amplification results, and then improve the quality of the training set, which is helpful to obtain a high-quality address unit labeling model.

[0068] 2. Establish an iterative management and control method based on the amplification threshold and the round number threshold: The present invention counts the number of credible samples generated in each iteration process. When its value is less than the amplification threshold, it is considered that the cost of iterative optimization is less than the benefit brought by sample amplification, and the iterative process is terminated in advance; when the number of iteration rounds exceeds the round number threshold, it is considered that the iterative optimization occupies resources for too long, and the iterative process needs to be terminated in advance to release resources for use by other processes; through this management and control method, the goal of optimizing resources, reducing costs and increasing efficiency can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The present invention can be further understood from the following description in conjunction with the accompanying drawings. The components in the figures are not necessarily drawn to scale, but the emphasis is placed on illustrating the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0070] Figure 1 It is a schematic diagram of the overall process of an address unit parsing method based on iterative learning of a control threshold according to the present invention;

[0071] Figure 2 It is a schematic diagram of a process of iteratively optimizing an address unit labeling model based on labeling results and a control threshold in the present invention;

[0072] Figure 3 It is a schematic diagram of the process of labeling a test set based on a voting method in the present invention;

[0073] Figure 4 It is a schematic diagram of the overall structure of an address unit parsing device based on iterative learning of a control threshold value according to the present invention. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with its embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the protection scope of the present invention.

[0075] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only and do not represent the only implementation method.

[0076] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0077] The “first” and “second” mentioned in the present invention do not represent specific quantities and orders, but are merely used to distinguish names.

[0078] Address information is widely present in various texts such as insurance policies, express delivery orders, work orders, tax bills, invoices, contracts, etc. Address unit parsing can extract various spatial location elements from the above texts, including: province, city, county, township, village, road, building, house number, direction, distance, etc., which is of great significance and value in daily life, commercial activities, government office, security control and other fields. The sentence "The two parties agreed to meet at the 2nd basement floor of the South Building at No. 5 Dadi Road, Ouhai High-tech Park, Ouhai District, Wenzhou City, Zhejiang Province." is parsed by address unit, and the result is "Prov = Zhejiang Province", "City = Wenzhou City", "Dist = Ouhai District", "City = Wenzhou", "DevArea = Ouhai High-tech Park", "Road = Dadi Road", "RoadNo = No. 5", "Poi = South Building", "SPoi = 2nd basement floor"; among which Prov, City, Dist, DevArea, Road, RoadNo, Poi, and SPoi represent the provincial-level administrative region, municipal-level administrative region, district-level administrative region, economic development zone, road, building number of the road, point of interest, and sub-point of interest, respectively.

[0079] Transductive learning is a semi-supervised learning method that can amplify samples with good quality in the test set into the training set, which helps improve the prediction effect of the model. It is suitable for scenarios such as address unit parsing where there are few labeled training samples and many unlabeled test samples. The address unit parsing method based on transductive learning generally uses multiple address unit annotation models to make predictions at the same time, and adds test samples with the same prediction results to the training set; however, it does not consider the quality of the model. If the quality of most models is not good, the prediction results are not reliable, and the quality of the training set will be reduced after amplification.

[0080] In addition, this address unit parsing method requires iterative execution. As the number of iterations increases, the number of trusted samples amplified in each round gradually decreases, but the computational cost consumed in each round remains almost unchanged. The benefits brought by sample amplification gradually decrease, and the cost-effectiveness also decreases, resulting in a waste of resources and violating the principle of reducing costs and increasing efficiency.

[0081] In order to solve the problems existing in the address unit parsing method based on direct induction learning, such as Figure 1 As shown, one embodiment of the present invention provides an address unit parsing method based on iterative learning of a control threshold, comprising the following steps:

[0082] S1: Partition the address recognition dataset.

[0083] Preferably, the specific method of dividing the address recognition data set includes the following steps:

[0084] The address recognition dataset is divided into a training set ADS train and test set ADS test ;

[0085] Among them, the training set ADS train The test set ADS consists of text containing address unit annotation information. test Contains text only, no address unit labeling information.

[0086] S2, train the address unit labeling model.

[0087] This step uses the training set ADS train Train M address unit annotation models i , 1≤i≤M.

[0088] S3, iteratively optimize the address unit labeling model based on the labeling results and control thresholds.

[0089] This step controls the training set ADS according to the regulation threshold trainThe iterative amplification process is carried out, and the address unit annotation model is optimized; the control threshold includes three categories: score threshold, amplification threshold and round number threshold; in each round of iteration, the generated M address unit annotation model Model is first used i Processing test set ADS test The text in , where 1≤i≤M.

[0090] For each address unit in any text, we get M address types and corresponding scores, and then we add M addresses with exactly the same type and scores exceeding the score threshold δ. S The address unit is regarded as a credible address unit, and the credible address unit and the text in it are combined into a credible sample and added to the training set ADS train and from the test set ADS test , and then use the expanded training set ADS train Retrain M address unit annotation models, if the number of amplified samples N in the current round is not less than the amplification threshold δ N And the number of iterations I does not exceed the round number threshold δ I , then continue to iterate.

[0091] In this way, by establishing an iterative control method based on the amplification threshold and the round number threshold, the number of credible samples generated in each iteration process is counted. When the number of credible samples is less than the amplification threshold, it is considered that the cost of iterative optimization is less than the benefit brought by sample amplification, and the iteration process is terminated early; when the number of iteration rounds exceeds the round number threshold, it is considered that the iterative optimization occupies resources for too long, and the iteration process needs to be terminated early to release resources for other processes to use; through this control method, the goal of optimizing resources, reducing costs and increasing efficiency can be achieved.

[0092] Further, if Figure 2 As shown, the training set ADS is controlled according to the regulation threshold train The specific method of iterative expansion process and optimizing the address unit labeling model includes the following steps:

[0093] S31: Initialize I=0, read the score threshold δ S , amplification threshold δ N , round number threshold δ I ;

[0094] S32: Initialize N=0;

[0095] S33: Read test set ADS test The first text to be processed Str:

[0096] S34: Use address units to label the model iProcess Str, 1≤i≤M; this step uses M address unit annotation models to process Str, and for each address unit in Str, M address types and corresponding scores are obtained;

[0097] S35: Select addresses with the same type and scores exceeding δ S This step selects address units from Str that meet the following requirements: M address types are exactly the same and the scores are all greater than δ S .

[0098] S36: Count the number n of selected address units;

[0099] S37: If n≥1, go to S38, otherwise go to S3D;

[0100] S38: considering the selected n address units as trusted address units;

[0101] S39: Combine the trusted address unit and Str into a trusted sample;

[0102] S3A: Adding credible samples to the training set ADS train ;

[0103] S3B: From the test set ADS test Delete Str from

[0104] S3C: Update the number of amplification samples in the current round N=N+1;

[0105] S3D: If ADS test If there is text to be processed, then go to S3E, otherwise go to S3F;

[0106] S3E: read the next text Str to be processed, and go to S34;

[0107] S3F: Using ADS train Training address unit annotation model i , 1≤i≤M;

[0108] S3G: update iteration number I = I + 1;

[0109] S3H: If N ≥ δ N And I≤δ I , then go to S32, otherwise go to S3I;

[0110] S3I: Return to ADS train 、ADS test and Model i , 1≤i≤M;

[0111] Among them, Str represents the text in the test set, N represents the number of amplified samples in the current round, and I represents the number of iterations.

[0112] S4: Label the test set based on voting method.

[0113] This step uses the address unit to label the model Model i Processing test set ADS test For each address unit in any text, 1≤i≤M, the corresponding address type is determined according to the voting method.

[0114] Preferably, the specific method for labeling the test set based on the voting method comprises the following steps:

[0115] S41: Read test set ADS test The first text to be processed Str;

[0116] S42: Use address units to label the model i Process Str, 1≤i≤M; this step uses M address unit annotation models to process Str, and for each address unit in Str, M temporary address types are obtained;

[0117] S43: Using a voting method to fuse the processing results of different address unit annotation models; this step selects the address type with the largest number of occurrences among the M prediction results as the final address type of each address unit;

[0118] S44: If ADS test If there is text to be processed, go to S45, otherwise go to S46;

[0119] S45: read the next text Str to be processed, and go to S42;

[0120] S46: Return to ADS test The address unit marking result;

[0121] S5: Extract the address units in the test set. Preferably, this step is based on the test set ADS test The address unit annotation results are extracted from the text to extract the address unit entities and their types.

[0122] For M address units, the labeling model Model i, 1≤i≤M, when predicting each address unit in any text, M address types and corresponding scores can be obtained. The score (i.e., prediction score) refers to the probability value when the address unit annotation model is used to predict the address unit. The probability value range is 0 to 1. The larger the value, the more accurate the model prediction and the higher the credibility. Otherwise, it is inaccurate. In other words, there is a positive correlation between the prediction score and the credibility of the prediction result. By selecting address units with prediction scores exceeding the score threshold and discarding address units that do not meet the quality requirements, the credibility of the amplification results can be improved, thereby improving the quality of the training set and helping to obtain a high-quality address unit annotation model.

[0123] For example, for the text "We came to Guangzhou for a trip today", there is one city-level address unit "Guangzhou". When the four-address unit annotation model is used for prediction, the model prediction results (address type and score) are: provincial administrative region / 0.4, city-level administrative region / 0.9, city-level administrative region / 0.85, district-level administrative region / 0.6. According to the principle of maximum score, the model believes that "Guangzhou" is the most credible city-level administrative region.

[0124] In one embodiment, if Figure 4 As shown, the present invention also provides an address unit parsing device based on threshold control iterative learning, which includes an address unit data set division component M1, an address unit annotation model training component M2, an address unit annotation model iterative optimization component M3, an address unit annotation component M4 and an address unit extraction component M5.

[0125] The address unit data set division component M1 is used to divide the data set containing address units into a training set and a test set; the address unit annotation model training component M2 is used to train multiple address unit annotation models; the address unit annotation model iterative optimization component M3 is used to iteratively expand the training set according to the annotation results and the control threshold, and continuously optimize the address unit annotation model; the address unit annotation component M4 is used to obtain the annotation results of the address units in the test sample by using the voting method; the address unit extraction component M5 is used to extract the address unit entities and their types from the address unit annotation results.

[0126] The address unit labeling model iterative optimization component M3 specifically includes an iterative optimization parameter initialization component M31, an iterative optimization process control component M32, a first type of address unit labeling result generation component M33, a trusted sample generation component M34 and a training set expansion component M35.

[0127] The iterative optimization parameter initialization component M31 is used to set the initial parameters when iteratively optimizing the address unit annotation model based on the annotation results and the control threshold; the iterative optimization process control component M32 is used to control the normal execution and termination of the iterative optimization address unit annotation model process based on the annotation results and the control threshold; the first-category address unit annotation result generation component M33 is used to use multiple address unit annotation models to generate the address type and score of the address unit in the test sample; the trusted sample generation component M34 is used to generate trusted samples according to the quality of the address unit annotation results; the training set expansion component M35 is used to transfer the trusted samples in the test set to the training set.

[0128] The address unit labeling component M4 specifically includes a second type address unit labeling result generating component M41 and an address unit labeling result fusing component M42.

[0129] The second type of address unit labeling result generating component M41 is used to generate temporary address types of address units in the test sample by using multiple address unit labeling models; the address unit labeling result fusion component M42 is used to determine the final address type of the address unit in the test sample by using a voting method.

[0130] In one embodiment, the number of annotated and unannotated address texts in the address text data set is 5,000 and 50,000 respectively, and the amount of annotated data is significantly lower than that of unannotated data. In order to improve the effect of address unit annotation, it is necessary to use an annotated data training model to perform address unit parsing on the unannotated data, and add the results that meet the parsing quality requirements to the annotated address text data set, so as to improve the effect of address unit annotation by expanding the training set scale. This expansion process needs to be iteratively executed, and the parameters in the iterative process are as follows: score threshold δ S =0.7, amplification threshold δ N =80, round number threshold δ I =100.

[0131] Execute S1: Set the 5000 labeled address text dataset as the training set ADS train , set the 50,000 unlabeled address text dataset as the test set ADS test .

[0132] Execute S2: In the training set ADS trainIn the above, CRF (conditional random field), BiLSTM (Bi-directional Long Short-Term Memory), BiGRU+CRF (Bidirectional Gated Recurrent Unit+conditional random field), Bert (Bidirectional Encoder Representations from Transformers), Bert+CRF (Bidirectional Encoder Representations from Transformers+conditional random field), and Bert+BiGRU+CRF (Bidirectional Encoder Representationsfrom Transformers+Bidirectional Gated Recurrent Unit+conditional randomfield) are used to train six address unit labeling models Model1, Model2, Model3, Model4, Model5, and Model6 respectively.

[0133] Specifically, CRF, BiLSTM, BiGRU+CRF, Bert, Bert+CRF, and Bert+BiGRU+CRF are six entity recognition models used for sequence labeling tasks to identify entities in sentences. For example, address units include: province, city, county, township, village, etc., and each level is a type 1 entity.

[0134] Execute S31: Initialize the number of iterations I = 0, read the score threshold δ S =0.7, amplification threshold δ N =80, round number threshold δ I =100.

[0135] Execute S32: initialize the number of amplification samples in the current round N=0.

[0136] Execute S33: Read test set ADS test The first piece of text to be processed Str.

[0137] Execute S34: Use address units to label the model Model i Processing Str, 1≤i≤6, for each address unit in Str, 6 address types and corresponding scores are obtained.

[0138] Execute S35: select address units that meet the following requirements from Str: the 6 address types are completely the same and the scores are all greater than 0.7.

[0139] Execute S36: count the number n of address units selected from Str.

[0140] Execute S37: If n≥1, then go to S38; otherwise go to S3D.

[0141] Execute S38: regard the selected n address units as trusted address units;

[0142] Execute S39: combine the trusted address unit and Str into a trusted sample.

[0143] Execute S3A: Add trusted samples to the training set ADS train .

[0144] Execute S3B: From the test set ADS test Delete Str.

[0145] Execute S3C: update the number of amplification samples in the current round N=N+1.

[0146] Execute S3D: If ADS test If there is text to be processed, go to S3E, otherwise go to S3F.

[0147] Execute S3E: read the next text Str to be processed and go to S34.

[0148] Repeat S34 to S3E several times for the test set ADS test All texts in the address unit are annotated and the results that meet the quality requirements are added to the training set ADS train In this way, the scale of the training set can be expanded.

[0149] Executing S3F: Using ADS train Training address unit annotation model i , 1≤i≤6.

[0150] Execute S3G: update the number of iterations I=I+1.

[0151] Execute S3H: If N ≥ 80 and I ≤ 100, then go to S32; otherwise, go to S3I.

[0152] Repeat S32 to S3H several times to achieve the training set ADS train Scale expansion and optimize and update the address unit annotation model Model i , 1≤i≤6, improve the labeling effect.

[0153] Execute S3I: Return training set ADS train , test set ADS test and address unit annotation model i , 1≤i≤6.

[0154] Execute S41: Read test set ADS test The first piece of text to be processed Str.

[0155] Execute S42: Use address units to label the model Model i Process Str, 1≤i≤6, and for each address unit in Str, 6 temporary address types are obtained. Take "Xiao Ming came to Dongwu Village, Mushan Town, Yuyao City, Ningbo City, Zhejiang Province" as an example, using Result i Representation Model i The processing results of the 6 address unit annotation models are as follows:

[0156] Result1: Xiao Ming / O came to / O Zhejiang Province / Prov Ningbo City / City Yuyao City / DevZone Mushan Town / Town Dongwu Village / Village

[0157] Result2: Xiao Ming / O came to / O Zhejiang Province / Prov Ningbo City / City Yuyao City / City Mushan Town / Road Dongwu Village / Community

[0158] Result3: Xiao Ming / O came to / O Zhejiang Province / Prov Ningbo City / City Yuyao City / City Mushan Town / Town Dongwu Village / Community

[0159] Result4: Xiao Ming / O came to / O Zhejiang Province / Prov Ningbo City / City Yuyao City / District Mushan Town / DevZone Dongwu Village / Community

[0160] Result5: Xiao Ming / O came to / O Zhejiang Province / Prov Ningbo City / City Yuyao City / District Mushan Town / Town Dongwu Village / Village

[0161] Result6: Xiao Ming / O came to / O Zhejiang Province / Prov Ningbo City / City Yuyao City / District Mushan Town / Town Dongwu Village / Community

[0162] Among them, Prov, City, DevZone, Town, Village, Road, Community, and District represent provincial administrative districts, municipal administrative districts, economic development zones, township administrative districts, village brigade communes, roads, village administrative districts, and county administrative districts respectively.

[0163] Execute S43: select the address type that appears most frequently among the six prediction results, and regard it as the final address type of each address unit, thereby fusing the processing results of different address unit labeling models.

[0164] Taking the annotation results of S4 as an example, the fused results are as follows:

[0165] Result merg :Xiao Ming / O came to / O Zhejiang Province / Prov Ningbo City / City Yuyao City / District Mushan Town / Town Dongwu Village / Community

[0166] Execute S44: If the test set ADS test If there is text to be processed, go to S45; otherwise, go to S46.

[0167] Execute S45: Read test set ADS test The next text to be processed Str, go to S42.

[0168] Repeat S42 to S45 several times to annotate the test set ADS test All address units present in the text.

[0169] Execute S46: Return to test set ADS test The address unit marking results.

[0170] Execute S5: According to the test set ADS test The address unit annotation results are extracted from the text to extract the address unit entity and its type. Taking "Xiao Ming / O came to / O Zhejiang Province / Prov Ningbo City / City Yuyao City / District Mushan Town / Town Dongwu Village / Community" as an example, the extracted results are as follows:

[0171] Prov=Zhejiang Province, City=Ningbo City, District=Yuyao City, Town=Mushan Town, Community=Dongwu Village.

[0172] In summary, the present invention establishes a credible sample generation method based on a score threshold. According to the positive correlation between the prediction score and the credibility of the prediction result, only the address units whose prediction score exceeds the threshold are selected, and the address units whose quality does not meet the requirements are discarded. This can improve the credibility of the amplification results, and then improve the quality of the training set, which is helpful to obtain a high-quality address unit labeling model.

[0173] In addition, the present invention establishes an iterative control method based on an amplification threshold and a round number threshold, and counts the number of credible samples generated in each iteration process. When the value is less than the amplification threshold, it is considered that the cost of iterative optimization is less than the benefit brought by sample amplification, and the iterative process is terminated early; when the number of iteration rounds exceeds the round number threshold, it is considered that the iterative optimization occupies resources for too long, and the iterative process needs to be terminated early to release resources for use by other processes; through this control method, the goal of optimizing resources, reducing costs and increasing efficiency can be achieved.

[0174] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0175] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. An address unit parsing method based on iterative learning of control thresholds, characterized in that: The address unit parsing method based on iterative learning of control thresholds comprises the following steps: S1: partition address recognition dataset; S2: training address unit annotation model; S3: Iteratively optimize the address unit annotation model based on the annotation results and control thresholds; S4: Annotate the test set based on voting method; S5: Extract the address units in the test set.

2. The address unit parsing method based on iterative learning of control threshold according to claim 1, characterized in that: The specific method of dividing the address recognition data set includes the following steps: The address recognition dataset is divided into a training set ADS train and test set ADS test ; Among them, the training set ADS train The test set ADS consists of text containing address unit annotation information. test Contains text only, without address unit labeling information.

3. The address unit parsing method based on iterative learning of control threshold according to claim 1, characterized in that: The specific method of training the address unit labeling model includes the following steps: Using the training set ADS train Train M address unit annotation models i ; Among them, 1≤i≤M.

4. The address unit parsing method based on iterative learning of control threshold according to claim 3, characterized in that: The specific method for iteratively optimizing the address unit labeling model based on the labeling results and the regulation threshold comprises the following steps: Controlling the training set ADS according to the regulation threshold train Iterative amplification process and optimize the address unit labeling model; Among them, the regulation threshold includes a score threshold, an amplification threshold and a round number threshold.

5. The address unit parsing method based on iterative learning of control threshold according to claim 4, characterized in that: The training set ADS is controlled according to the regulation threshold train The specific method of iterative expansion process and optimizing the address unit labeling model includes the following steps: S31: Initialize I=0, read the score threshold δ S , amplification threshold δ N , round number threshold δ I ; S32: Initialize N=0; S33: Read test set ADS test The first text to be processed Str; S34: Use address units to label the model i Process Str, 1≤i≤M, and obtain M address types and scores for each address unit; S35: Select M addresses with the same type and scores exceeding δ S The address unit; S36: Count the number n of selected address units; S37: If n ≥ 1, go to S38, otherwise go to S3D; S38: considering the selected n address units as trusted address units; S39: Combine the trusted address unit and Str into a trusted sample; S3A: Adding credible samples to the training set ADS train ; S3B: From the test set ADS test Delete Str from S3C: Update the number of amplification samples in the current round N=N+1; S3D: If ADS test If there is text to be processed, then go to S3E, otherwise go to S3F; S3E: read the next text Str to be processed, and go to S34; S3F: Using ADS train Training address unit annotation model i , 1≤i≤M; S3G: update iteration number I = I + 1; S3H: If N ≥ δ N And I≤δ I , then go to S32, otherwise go to S3I; S3I: Return to ADS train 、ADS test and Model i , 1≤i≤M; Among them, Str represents the text in the test set, N represents the number of amplified samples in the current round, and I represents the number of iterations.

6. The address unit parsing method based on iterative learning of control threshold according to claim 5, characterized in that: The specific method of labeling the test set based on the voting method includes the following steps: S41: Read test set ADS test The first text to be processed Str; S42: Use address units to label the model i Process Str, 1≤i≤M, and obtain M temporary address types for each address unit; S43: using a voting method to fuse the processing results of different address unit annotation models to obtain a final address type for each address unit; S44: If ADS test If there is text to be processed, go to S45, otherwise go to S46; S45: read the next text Str to be processed, and go to S42; S46: Return to ADS test The address unit marking results.

7. The address unit parsing method based on iterative learning of control threshold according to claim 1, characterized in that: The specific method for extracting address units in the test set includes the following steps: According to the test set ADS test The address unit annotation results are extracted from the text to extract the address unit entities and their types.

8. An address unit parsing device based on iterative learning of control threshold, characterized in that: The address unit parsing device based on iterative learning of control thresholds includes: An address unit data set division component M1, used for dividing a data set containing address units into a training set and a test set; An address unit annotation model training component M2, used for training multiple address unit annotation models; The address unit annotation model iterative optimization component M3 is used to iteratively expand the training set according to the annotation results and the control threshold, and continuously optimize the address unit annotation model; The address unit marking component M4 is used to obtain the marking results of the address units in the test sample by using a voting method; The address unit extraction component M5 is used to extract the address unit entity and its type from the address unit marking result.

9. The address unit parsing device based on iterative learning of control threshold according to claim 8, characterized in that: The address unit annotation model iterative optimization component M3 specifically includes: Iterative optimization parameter initialization component M31, used to set initial parameters when iteratively optimizing the address unit annotation model based on the annotation results and the control threshold; Iterative optimization process control component M32, used to control the normal execution and end of the iterative optimization address unit annotation model process based on the annotation results and the regulation threshold; A first-category address unit annotation result generating component M33, used to generate address types and scores of address units in a test sample using multiple address unit annotation models; A trusted sample generating component M34, used to generate a trusted sample according to the quality of the address unit annotation result; The training set augmentation component M35 is used to transfer the credible samples in the test set to the training set.

10. The address unit parsing method based on iterative learning of control threshold according to claim 7, characterized in that: The address unit marking component M4 specifically includes: A second type address unit labeling result generating component M41, used to generate a temporary address type of an address unit in a test sample by using multiple address unit labeling models; The address unit labeling result fusion component M42 is used to determine the final address type of the address unit in the test sample by using a voting method.

Images