CN112329467B

CN112329467B - Address recognition method and device, electronic equipment and storage medium

Info

Publication number: CN112329467B
Application number: CN202011211412.1A
Authority: CN
Inventors: 张龙
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2020-11-03
Filing date: 2020-11-03
Publication date: 2022-09-30
Anticipated expiration: 2040-11-03
Also published as: CN112329467A

Abstract

The embodiment of the application discloses an address identification method and device, electronic equipment and a storage medium, and the method and device can be applied to the fields of artificial intelligence, big data, maps and the like. The method comprises the following steps: acquiring an address text to be identified; acquiring identification guide information of an address text to be identified, wherein the identification guide information comprises at least one item of basic information of words contained in the address text to be identified, identification information of target address words or characteristic information of words, and the identification information of the target address words represents identification results of the target address words; and obtaining an address role recognition result of the address text to be recognized according to the address text to be recognized and the recognition guidance information. By adopting the method and the device, the address character recognition result of the address text to be recognized can be obtained through the recognition guidance information of the address text to be recognized and the recognition guidance information of the address text to be recognized, and the accuracy is high.

Description

Address recognition method and device, electronic equipment and storage medium

Technical Field

The present application relates to the field of artificial intelligence, big data processing, and map technologies, and in particular, to an address identification method and apparatus, an electronic device, and a storage medium.

Background

With the continuous development of random Artificial Intelligence (AI) and big data, text recognition is one of the most important technologies. In the address role identification scenario, the address role of the address text line to be identified needs to be identified.

In the existing scheme, because the comprehension capability of the address text to be recognized is insufficient, the address recognition error is easily caused. For example, the address "Handan City Yi Yuan Ji No. 4-7", wherein the first is the first district, the first is the building, and in the almost same structure address "Haizhan Yongfeng Lu xi shan Yi building No. 5", the Xishan Yi Ji as a whole represents a district name. The existing scheme can not effectively distinguish 'one number yard' in two addresses with different sources during address identification. That is, in the existing address recognition scheme, the generalization capability is weak for different source addresses.

Therefore, how to improve the accuracy of address identification becomes an urgent problem to be solved.

Disclosure of Invention

The embodiment of the application provides an address identification method, an address identification device, an electronic device and a storage medium, wherein an address character identification result of an address text to be identified can be obtained through identification guide information of the address text to be identified and identification guide information of the address text to be identified, and the identification accuracy can be effectively improved.

In a first aspect, an embodiment of the present application provides an address identification method, including:

acquiring an address text to be identified;

acquiring identification guidance information of the address text to be identified, wherein the identification guidance information comprises at least one item of basic information of words contained in the address text to be identified, identification information of target address words or character characteristic information of words, and the identification information of the target address words represents an identification result of the target address words;

and obtaining an address role recognition result of the address text to be recognized according to the address text to be recognized and the recognition guidance information.

In a second aspect, an embodiment of the present application provides an address recognition apparatus, including:

the text to be recognized acquisition module is used for acquiring an address text to be recognized;

a guidance information obtaining module, configured to obtain recognition guidance information of the address text to be recognized, where the recognition guidance information includes at least one of basic information of a word included in the address text to be recognized, identification information of a target word, or feature information of a word;

and the address identification module is used for obtaining an address role identification result of the address text to be identified according to the address text to be identified and the identification guidance information.

In some possible embodiments, the apparatus further comprises: the address keyword lexicon building module is used for building an address keyword lexicon, wherein the address keyword lexicon comprises address keywords and identification information of the address keywords; the guidance information obtaining module is configured to: determining the address keywords hit by the address text to be recognized in the address keyword word bank based on the address keyword word bank, and determining the hit address keywords as target address words; and obtaining the identification information of the target address words based on the address keyword word bank.

In some possible embodiments, the basic information of the word includes at least one of part-of-speech information of the word or position information of the word in the address text to be recognized.

In some possible embodiments, the characteristic information of the word includes at least one of pinyin information of the word or morphological information of the word.

In some possible embodiments, the address identification module is configured to: acquiring text characteristics of the address text to be recognized; acquiring the information characteristics of the identification guidance information; fusing the text features and the information features to obtain fused features; and determining the address role recognition result of the address text to be recognized according to the fusion characteristics.

In some possible embodiments, the address identification module is configured to: and splicing the text features and the information features, and determining the spliced text features and the spliced information features as the fusion features.

In some possible embodiments, the obtaining of the address role recognition result of the address text to be recognized according to the address text to be recognized and the recognition guidance information is realized by an address recognition model, wherein the address recognition model includes a text feature extraction model, a fusion feature extraction model and a recognition result acquisition model which are sequentially cascaded; the address recognition device is configured to: respectively acquiring the information characteristics of the identification guidance information; extracting text features of the address text to be recognized through the text feature extraction model based on the address text to be recognized; extracting fusion features corresponding to the address texts to be recognized through the fusion feature extraction model based on the text features and the information features; and obtaining the address role recognition result of the address text to be recognized through the recognition result acquisition model based on the fusion characteristics.

In some possible embodiments, the address recognition model is a first recognition model or a second recognition model, wherein the text feature extraction model of the first recognition model is a model representing a Bert network based on a bidirectional encoder of a Transformer, the fused feature extraction model is a long-short term memory network, and the recognition result acquisition model is a conditional random field model; the text feature extraction model of the second recognition model is an attention mechanism model, the fused feature extraction model is a long-short term memory network, and the recognition result acquisition model is a conditional random field model.

Wherein, the address recognition model is obtained by training through a model training device.

In some possible embodiments, the model training apparatus may be specifically configured to:

acquiring a training data set; training a first initial neural network model corresponding to the first recognition model based on the training data set to obtain a trained first recognition model; and training the second initial neural network model by taking the first recognition model as a teacher model and taking the second initial neural network model corresponding to the second recognition model as a student model to obtain the second recognition model.

In some possible embodiments, the model training apparatus includes:

the model building module is used for building a third initial neural network model by adopting a first programming language and building a fourth initial neural network model with the same structure as the third initial neural network model by adopting a second programming language;

a model training module, configured to train the third initial neural network model until a training end condition is met, use a model obtained when the training end condition is met as a trained third recognition model, and store model parameters of the third recognition model;

a model determining module configured to obtain the address recognition model by using the stored model parameters of the third recognition model as model parameters of the fourth initial neural network model; the performance of the address recognition model corresponding to the second programming language is better than that of the third recognition model corresponding to the first programming language.

In a third aspect, an embodiment of the present application provides an electronic device, including a processor and a memory, where the processor and the memory are connected to each other;

the memory is used for storing computer programs;

the processor is configured to perform the method provided by the first aspect when the computer program is called.

In a fourth aspect, the present application provides a computer-readable storage medium, where a computer program is stored, where the computer program is executed by a processor to implement the method provided in the first aspect.

In a fifth aspect, embodiments of the present application provide a computer program product or a computer program comprising computer instructions stored in a computer-readable storage medium. The processor of the electronic device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the method provided by the first aspect.

In the embodiment of the application, firstly, the address text to be recognized is obtained, the recognition guidance information of the address text to be recognized is obtained, and then, the address role recognition result of the address text to be recognized is obtained according to the address text to be recognized and the recognition guidance information. By adopting the mode, the identification guide information comprises at least one item of basic information of words contained in the address text to be identified, identification information of the target address words or character characteristic information, wherein the identification information of the target address words can represent the identification result of the target address words, so that the finally obtained address role identification result of the address text to be identified can be combined with two factors of the information of the address text to be identified and the identification guide information, the identification accuracy of the address text to be identified is greatly improved, for addresses of different sources, the address texts to be identified of different sources can be effectively distinguished due to the existence of the identification guide information, and the identification accuracy is improved.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.

Fig. 1 is a schematic flowchart of an address identification method according to an embodiment of the present application;

fig. 2 is a schematic diagram of an address role recognition result of an address text to be recognized provided in an example of the present application;

FIG. 3 is a schematic structural diagram of an address identification model provided in an embodiment of the present application;

fig. 4 is a schematic diagram illustrating a principle of obtaining an address role recognition result of an address text to be recognized through an address recognition model according to an embodiment of the present application;

fig. 5 is a schematic diagram of an address role recognition result of an address text to be recognized, which is obtained through an address recognition model according to an embodiment of the present application;

fig. 6 is a schematic diagram illustrating an address role recognition result of an address text to be recognized according to an embodiment of the present application;

FIG. 7 is a schematic diagram of a second recognition model provided by embodiments of the present application;

FIG. 8 is a schematic diagram illustrating a migration learning process of an address recognition model according to an embodiment of the present application;

fig. 9 is a schematic structural diagram of an address identification and application system based on the address identification method provided in the embodiment of the present application;

FIG. 10 is a schematic structural diagram of an address recognition apparatus according to an embodiment of the present application;

fig. 11 is a schematic structural diagram of an electronic device provided in an embodiment of the present application.

Detailed Description

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.

The address identification method provided by the embodiment of the application can be applied to various fields such as artificial intelligence, big data and the like, such as the fields of human-computer interaction based on Natural Language Processing (NLP), Cloud computing in Cloud technology, artificial intelligence Cloud service and related data computing processing in the big data field, and aims to identify the address role identification result of the address text to be identified. The address text to be recognized may be understood as a text corresponding to the address information, and may be specifically determined based on an actual application scenario, which is not limited herein.

Artificial intelligence is a theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a manner similar to human intelligence. Artificial intelligence is the research of the design principle and the realization method of various intelligent machines, so that the machines have the functions of perception, reasoning and decision making.

Natural Language Processing (NLP) is an important direction in the fields of computer science and artificial intelligence. It studies various theories and methods that enable efficient communication between humans and computers using natural language. Natural language processing is a science integrating linguistics, computer science and mathematics. Therefore, the research in this field will involve natural language, i.e. the language that people use everyday, so it is closely related to the research of linguistics. Natural language processing techniques typically include text processing, semantic understanding, machine translation, robotic question answering, and the like.

The cloud technology is a hosting technology for unifying series resources such as hardware, software, network and the like in a wide area network or a local area network to realize the calculation, storage, processing and sharing of data. The address identification method provided by the embodiment of the application can be realized based on cloud computing (cloud computing) in cloud technology.

Cloud Computing refers to obtaining required resources in an on-demand and easily-extensible manner through a Network, and is a product formed by development and fusion of traditional computers and Network Technologies, such as Grid Computing (Grid Computing), Distributed Computing (Distributed Computing), Parallel Computing (Parallel Computing), Utility Computing (Utility Computing), Network Storage (Network Storage Technologies), Virtualization (Virtualization), Load balancing (Load Balance) and the like.

An artificial intelligence cloud Service is also generally called AIaaS (AI as a Service). The method is a service mode of an artificial intelligence platform, and particularly, the AIaaS platform splits several types of common artificial intelligence services, and provides independent or packaged services, such as text translation services, at a cloud.

Big data (Big data) refers to a data set which cannot be captured, managed and processed by a conventional software tool within a certain time range, and is a massive, high-growth-rate and diversified information asset which can have stronger decision-making power, insight discovery power and process optimization capability only by a new processing mode. With the advent of the cloud era, big data has attracted more and more attention. The big data is based on technologies such as a large-scale parallel processing database, data mining, a distributed file system, a distributed database, the cloud computing and the like, and the address identification method provided by the embodiment is effectively implemented.

Referring to fig. 1, fig. 1 is a schematic flowchart of an address identification method according to an embodiment of the present disclosure. The method can be executed by any electronic equipment, such as a server or a user terminal, or alternatively, the user terminal and the server are interactively completed, optionally, the method can be executed by the server, the user terminal can send the acquired address text to be recognized to the server, the server further acquires the recognition guidance information of the address text to be recognized, and the address role recognition result of the address text to be recognized is obtained according to the address text to be recognized and the recognition guidance information.

The server may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server or a server cluster providing basic cloud computing services such as cloud service, a cloud database, cloud computing, cloud functions, cloud storage, Network service, cloud communication, middleware service, domain name service, security service, CDN (Content Delivery Network), big data, an artificial intelligence platform, and the like. The user terminal may be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, and the like, and the user terminal and the server may be directly or indirectly connected through wired or wireless communication, but are not limited thereto.

As shown in fig. 1, an address identification method provided in an embodiment of the present application may include the following steps:

and step S1, acquiring the address text to be recognized.

In some possible embodiments, the address text to be recognized is a text obtained from address information, and the address information may include information of province, city, district, town street, village, road, house number, POI, building, room number, express delivery site, express delivery cabinet name, and the like. The acquisition of the address text to be recognized includes, but is not limited to, the address text to be recognized obtained in the processes of intelligent question answering, intelligent translation, semantic analysis and the like in the field of artificial intelligence, for example, a text obtained by text conversion of the acquired voice through map software, a text directly input through the map software, a text obtained by analysis of waybill information, and the like, which is not limited herein.

The POI (Point of Interest, abbreviated as POI) is a term in a geographic information system, generally refers to any geographic object that can be abstracted as a Point, is a certain landmark and a scenic spot in the geographic information system, and is used to mark places such as government departments represented by the place, commercial institutions of various industries (gas stations, department stores, supermarkets, restaurants, hotels, convenience stores, hospitals, and the like), tourist attractions (parks and public toilets), historic sites, traffic facilities (various stations, parking lots, speeding cameras, speed limit signs), and the like. The main purpose of the interest points is to describe the addresses of the things or events, so that the description capability and the query capability of the positions of the things or events can be greatly enhanced, and the accuracy and the speed of geographic positioning are improved.

Optionally, when the address recognition method provided in this embodiment of the application is executed by a user terminal, the user terminal may obtain a text input by a user as an address text to be recognized, or obtain the address text to be recognized after converting a user voice (a voice related to an address) into the text, or obtain the address text to be recognized based on the address text to be recognized that the user obtains from a network, big data, and the like based on the user terminal, where a specific obtaining manner may be determined based on actual application scene requirements, and is not limited herein.

Optionally, when the address identification method provided in this embodiment of the application is executed by a server, the server may obtain any text to be processed in an address database as an address text to be identified, or generate the address text to be identified based on a generation instruction sent by a user terminal and based on technologies such as cloud computing and big data, or obtain the address text to be identified from a storage space indicated by the acquisition instruction based on an acquisition instruction sent by the user terminal, where a specific obtaining manner may be determined based on actual application scene requirements, and is not limited herein. The storage space includes, but is not limited to, a cloud server, a cloud storage space, and the like, and is not limited herein.

Step S2, obtaining recognition guidance information of the address text to be recognized, where the recognition guidance information includes at least one of basic information of a word included in the address text to be recognized, identification information of a target address word, or feature information of a word, and the identification information of the target address word represents a recognition result of the target address word.

In some possible embodiments, it is necessary to obtain recognition guidance information of the address text to be recognized, where the recognition guidance information may be understood as a kind of supervision information corresponding to the address text to be recognized, and the recognition guidance information may be capable of understanding POIs included in the address text to be recognized from multiple angles, and the recognition guidance information may include at least one of basic information of each word in the address text to be recognized, identification information of a target address word, or feature information of a word.

In an alternative embodiment, the guidance information to be recognized may include basic information of a word, where the basic information of the word includes at least one of part-of-speech information of the word or position information of the word in the address text to be recognized.

For a word, the part-of-speech information of the word can be understood as sentence component information of the address text to be recognized, and the word attribute of the word in the address text is represented. Taking a certain participle in the address text to be recognized as an example, if the participle belongs to the Chinese language, the part-of-speech information of the word can be understood as that the participle is a noun, a verb, an adjective, a pronoun, a number word, an quantifier, a distinguishment word, an adverb, a preposition, a interword, an exclamation word, a pseudonym, and the like.

For a word, the position information of the word in the address text to be recognized represents the position of the word in the address text, and may be understood as word order information of the word in the address text to be recognized, a front-back position relation of each word in the address text to be recognized, and the like.

Optionally, the basic information of the word may also be a combination of part-of-speech information of the word and position information of the word in the address text to be recognized.

For example, when the address text to be recognized is recognized, the address text to be recognized may be segmented, and then part-of-speech analysis may be performed on the address text to be recognized according to the chinese part-of-speech information of the address text to be recognized. For example, taking the waybill address "bayobo xie-releasing to sunny district american cubic district in beijing city" as an example, the waybill address explicitly informs that the bayobo (noun) xie (verb) is released, and the key component of "bayobo" can be more easily recognized through the part-of-speech information of each participle of the waybill address, the position information of the word between the front and the back in the address text to be recognized, and the structure of combining the verb + noun + verb.

By the aid of the method and the device, the key information of the address text to be recognized can be understood from multiple angles by means of the basic information of the words of the address text to be recognized and the position information of the words in the address text to be recognized, and accordingly recognition accuracy is improved.

In one possible embodiment, the recognition guidance information includes identification information of the target address word; the method further comprises the following steps: constructing an address keyword word bank, wherein the address keyword word bank comprises address keywords and identification information of the address keywords; the acquiring of the identification guidance information of the address text to be identified includes: determining address keywords hit by the address text to be recognized in the address keyword word bank based on the address keyword word bank, and determining the hit address keywords as target address words; and obtaining the identification information of the target address words based on the address keyword word bank.

Alternatively, the recognition guidance information may include identification information of the target address word. The target address word may be understood as an address word having a special meaning in the address text to be recognized, and may specifically be some specific address words, such as address words that are easy to be recognized by mistake. The identification information of the target address word is the identification result obtained by identifying the target address word.

The specific acquisition mode of the target address word is not limited in the embodiment of the present application, and the address word may be an address word conforming to a specific word structure or an address word in a pre-constructed address keyword word bank, where the address keyword word bank may be a word bank corresponding to an address word which is obtained through statistical analysis and is prone to error.

For example, taking address role recognition as an example, for "open yuan" in an address "Handan' Yi Yuan street open yuan No. one yard No. 4-7", the address role recognition domain dictionary is expressed as a POI entity, "one yard" is expressed as a building, and another address "Haizhan zone Yongfeng Luo Xishan No. 5 building" with almost the same structure is expressed as a whole in the address role recognition domain dictionary as a POI entity, that is, for the same word, the "open yuan" and "Xishan Yi yard" have different meanings in different addresses, that is, the "open yuan" and the "Xishan Yi yard" belong to words with special meanings in address role recognition, that is, the "open yuan" and the "Xishan Yi yard" are target address words.

It is understood that the above is only an example, and the present embodiment is not limited thereto.

The identification information of the target address word can be obtained in the following manner:

and constructing an address keyword word bank, wherein the address keyword word bank comprises address keywords and identification information of the address keywords.

Alternatively, the address keyword lexicon may be understood as a domain dictionary repository, and in particular, the domain dictionary repository may include a large number of POI entity dictionaries (e.g., dictionaries containing 700w POI entities), village name dictionaries (e.g., dictionaries containing 180w village names), road name dictionaries (e.g., dictionaries containing 44w road names), and the like, without any limitation herein.

And then, based on the address keyword lexicon, searching whether the address text to be identified has address keywords matched with the address keyword lexicon, and if so, determining the hit address keywords as target keywords.

And finally, determining the identification information of the target address word according to the identification information corresponding to the hit address keyword based on the address keyword word bank.

Taking the address text to be recognized as ' the number 5 building of the number one institute of west mountain of yongsheng luo of hai lake area ', the address text to be recognized can be matched with the address keywords in the address keyword lexicon to obtain the address keyword of the number one institute of west mountain ', and the number one institute of west mountain ' is taken as the target address word to obtain the identification information of the target address word of the number one institute of west mountain '.

It is understood that the above description is only an example, and the present implementation is not limited thereto.

According to the method and the device, the target address words in the address text to be recognized can be effectively determined based on the constructed address keyword lexicon, and because the target address words are address words with special meanings, the identification information of the target address words is recognized based on the address keyword lexicon, the recognition accuracy can be improved more accurately, and the situation that the target address words are recognized wrongly is avoided.

In an alternative embodiment, the guidance information to be recognized may include character information of a word, the character information of the word including at least one of pinyin information of the word or morphological information of the word.

The pinyin information of the character is the pinyin information of each character contained in the address text to be identified. The morphological information of a word may be understood as stroke information of the word.

In practical application, for the address text to be recognized, at least one word may have a wrongly recognized word, which causes a recognition error of the address text to be recognized, and affects the address character recognition result to avoid the occurrence of the error.

For example, when the address text to be recognized is recognized, the address text to be recognized may be segmented, then the feature information of each character in the address text to be recognized may be obtained, and the feature information of the character is used to recognize the address text to be recognized.

For example, by using the pinyin information of the words, the correct understanding (actually, retail) of the address "the Xinhe bridge in Qingpu district of Shanghai 689 sells" the shoes and clothes for sale "the middle" can be solved, so that the "shoes and clothes for sale" the Tianxing "can be identified as a complete POI entity.

For another example, the Chinese language model n-gram can be used for obtaining the form information of the Chinese stroke horizontal, vertical, left falling, right falling and hook characters, so that the accurate understanding and recognition of the shape of the character, such as the accurate understanding and recognition of the "Tianshida-Mao" in the second building of the restaurant on the opposite side of the Haihe Dachai sea in the Haihe district of Beijing can be solved (actually, the character is the "Tianshixiao-Mao").

Through the embodiment, for the address text to be recognized, the situation of wrongly-recognized characters exists, so that the accuracy of the address character recognition result of the final address text to be recognized is influenced, the real meaning expressed by the address text to be recognized can be more accurately understood by using the pinyin information and the shape information of the characters, the ambiguity of wrongly-recognized characters is avoided, and the recognition accuracy of the address text to be recognized is improved.

And step S3, obtaining the address role recognition result of the address text to be recognized according to the address text to be recognized and the recognition guidance information.

After the address text to be recognized is obtained through the above-mentioned step S1 and the recognition guidance information is obtained through the step S2, the result of the recognition of the address character of the address text to be recognized can be obtained based on the address text to be recognized and the recognition guidance information.

Optionally, the address role Recognition result of the address text to be recognized may be understood as a Named Entity Recognition result, Named Entity Recognition (NER), also referred to as proper name Recognition and Named Entity, which refers to an Entity having a specific meaning in the Recognition text, and mainly includes a name of a person, a name of a place, a name of an organization, a proper noun, and the like, and characters such as time, quantity, currency, proportional numerical values, and the like. It refers to things that can be identified by proper nouns (names), and a named entity generally represents only one specific individual, including names of people, places, etc. For example, for the address "beijing haiji district beijing university", the named entity recognition result of the address may be beijing haiji district (division) beijing university (POI entity), i.e., the named entity recognition result may indicate which word in the address "beijing haiji district beijing university" is a city, which word is a division, and which word is a POI entity.

Alternatively, the address role recognition result of the address text to be recognized can be understood as the address role tag corresponding to the named entity recognition. Wherein, the address role label can be understood as: and the label of the address text to be recognized is given after the named entity of the address text to be recognized is recognized, and the label is used as a basis for dividing and recognizing the boundary of the subsequent address role. The address role label consists of two parts: the BIS segmentation labels and category labels. In the BIS segmentation labels, B represents the role start, I represents the role middle part, S represents the role including only the current content, and the category labels may be divided into 32 categories as required, such as cities, divisions, streets, roads, POI entities, and the like. For example, as for the address "beijing university of haijiang district beijing", the corresponding BIS division label and category label and the final recognition result are shown in fig. 2, it can be seen that the BIS division labels of the address "beijing university of haijiang district beijing" are north (B-CITY), kyoto (I-CITY), munich (I-CITY) sea (B-DIS), lake (I-DIS), district (I-DIS), north (B-POI), kyoto (I-POI), and large (I-POI) mathematics (I-POI) in this order. The category labels of the address "beijing university of haichi district of beijing city" are beijing city (city), haichi district (division), and beijing university (POI entity) in that order.

Here, dis (division inter partitions) represents a region, and poi (point of interfaces) represents a point of interest. The BIS tag B-DIS of "sea" represents the beginning of the district in the "hai lake district beijing university" as the address text to be recognized, "the BIS tag I-DIS of" lake "represents the middle part of the district in the" hai lake district beijing university "as the address text to be recognized," lake "represents the beginning of the POI in the" hai lake district beijing university "as the address text to be recognized," bei "represents the middle part of the POI in the" hai lake district beijing university "as the" beijing "as the address text to be recognized.

It should be noted that the address character recognition result of the address text to be recognized shown in fig. 2 includes address elements such as cities, regions, and POIs, but this is not limited in this embodiment of the application, and in practical applications, different address character recognition results are obtained according to a specific address text to be recognized, and category labels including different address elements are obtained.

For example, for an address role recognition result corresponding to an address text to be recognized, "new station township girlful", the result may be: new (B-STRT) station (I-STRT) town (I-STRT) village (I-STRT) small (B-VILG) Chinese (I-VILG) family (I-VILG). At this time, the address text "new station, town and country maigre" to be recognized includes address elements of the town and the country. Wherein, the new BIS label B-STRT represents that the new address is the start of the village in the address text to be identified, namely the new village and small girl of the new station, the BIS label I-STRT of the station represents that the station is the middle part of the village in the address text to be identified, namely the new station village and small girl, the small BIS label B-VILG represents that the small address is the start of the village in the address text to be identified, namely the new station village and small girl, and the big BIS label I-VILG represents that the small address is the middle part of the village in the address text to be identified, namely the new station village and small girl.

Alternatively, the specific meaning of the BIS segmentation label may also be in the manner shown in table 1.

TABLE 1

Labelling	Means of
		B	Header of current word as geographical named entity
I	The current word being internal to the geographically-named entity
		E	The current word being the end of the geographical named entity
S	The current word isGeographical named entities
		O	The current word not being a geographically named entity or component

In an optional embodiment, the obtaining of the address character recognition result of the address text to be recognized according to the address text to be recognized and the recognition guidance information may be implemented as follows: acquiring text characteristics of the address text to be recognized; acquiring the information characteristics of the identification guidance information; fusing the text features and the information features to obtain fused features; and determining an address role recognition result of the address text to be recognized according to the fusion characteristics.

Optionally, text features of the address text to be recognized and information features of the recognition guide information may be extracted, then the text features of the address text to be recognized and the information features of the recognition guide information are fused to obtain fusion features after the text features and the information features are fused, and finally, the address role recognition result of the address text to be recognized is determined according to the fusion features.

For example, both textual features and informational features may be feature vectors. The text feature may be a semantic vector (may be simply referred to as a text feature vector) corresponding to the address text to be recognized, and the information feature may be at least one of a basic feature vector, a domain feature vector, and a derived feature vector corresponding to the recognition guidance information. The basic characteristic vector is the information characteristic of basic information of a word, the field characteristic vector is the information characteristic of identification information of a target address word, and the derived characteristic vector is the information characteristic of the characteristic information of a word.

According to the embodiment, the identification result corresponding to the address text to be identified can be obtained according to the fusion characteristic after the text characteristic and the information characteristic are fused, the address identification only considering the information of the address text to be identified is avoided, the address identification mode for the address text to be identified is comprehensively considered through the fusion characteristic after the text characteristic and the information characteristic are fused, more accurate address identification can be carried out according to the basic information of words, whether the words contain the target address words and the characteristic information of the words, and the accuracy of the address identification is improved.

The fusion process of the text feature and the information feature can be realized in the following manner.

In an optional embodiment, the fusing the text feature and the information feature to obtain a fused feature includes: and splicing the text features and the information features, and determining the spliced text features and the spliced information features as the fusion features.

Optionally, the text feature corresponding to the address text to be recognized and the information feature corresponding to the recognition guidance information may be spliced to obtain a spliced text feature and a spliced feature corresponding to the information feature, and the spliced feature is determined as the fusion feature.

Optionally, the text features and the recognition guidance information have different importance degrees for the address role recognition result, so that when the text features and the information features are spliced, different weights corresponding to the text features and the information features can be set. For example, a text feature vector corresponding to the text feature may be concatenated with at least one of the base feature vector, the domain feature vector, and the derived feature vector. The base feature vector, the domain feature vector, and the derived feature vector may be set to have different weights. In an actual scene, weights with different sizes can be set according to the importance degree of the influence of the text feature vector, the basic feature vector, the field feature vector and the derived feature vector on the address role recognition result.

Through the steps S1-S3, the recognition result corresponding to the address text to be recognized can be obtained.

According to the embodiment, firstly, the address text to be recognized is obtained, the recognition guide information of the address text to be recognized is obtained, and then the address role recognition result of the address text to be recognized is obtained according to the address text to be recognized and the recognition guide information. By adopting the above mode, the identification guidance information includes at least one item of basic information of words contained in the address text to be identified, identification information of the target address words, or characteristic information of the words, wherein the identification information of the target address words can represent the identification result of the target address words, so that the finally obtained address role identification result of the address text to be identified can be combined with two factors of the information of the address text to be identified and the identification guidance information, and the identification accuracy of the address text to be identified is greatly improved.

In an optional embodiment, the obtaining of the address role identification result of the address text to be identified according to the address text to be identified and the identification guidance information is realized by an address identification model, wherein the address identification model comprises a text feature extraction model, a fusion feature extraction model and an identification result acquisition model which are sequentially cascaded; the obtaining of the address role identification result of the address text to be identified according to the address text to be identified and the identification guidance information includes: respectively acquiring the information characteristics of the identification guidance information; extracting text features of the address text to be recognized through the text feature extraction model based on the address text to be recognized; extracting fusion features corresponding to the address texts to be recognized through the fusion feature extraction model based on the text features and the information features; and obtaining the address role recognition result of the address text to be recognized through the recognition result acquisition model based on the fusion characteristics.

Optionally, the address recognition is performed on the address text to be recognized to obtain a recognition result corresponding to the address text to be recognized, and the recognition result can be realized through an address recognition model. As for the address recognition model, as shown in fig. 3, the address recognition model includes a text feature extraction model, a fusion feature extraction model, and a recognition result obtaining model, which are sequentially cascaded.

The text feature extraction model may be implemented by using a neural network structure, a specific network structure of the text feature extraction model is not limited in the embodiments of the present application, and may be selected and configured according to actual requirements, and the text feature extraction model includes, but is not limited to, a Bidirectional Encoder representation from transforms based on a converter transform (Bert model), a lightweight version Bert model (a Lite Bert, abbreviated as ALBERT), a neural network structure based on attention mechanism attribute, a natural language processing NLP model based on a self-attention (self-attribute) neural network structure, and the like.

The fusion feature extraction model may be implemented by using a Neural Network structure, a specific Network structure of the fusion feature extraction model is not limited in the embodiment of the present application, and the fusion feature extraction model includes, but is not limited to, a Long Short-Term Memory Neural Network (LSTM), a Bi-directional Long Short-Term Memory (Bi-directional LSTM) Neural Network combined by a forward LSTM and a backward LSTM, a Recurrent Neural Network (RNN for Short), a Gated Recurrent Unit (GRU), and other Neural Network models.

The recognition result obtaining model may be implemented by using a neural network structure, a specific network structure of the recognition result obtaining model is not limited in the embodiment of the present application, and the recognition result obtaining model includes, but is not limited to, a Conditional Random Field (CRF).

Alternatively, the information features of the identification guidance information may be acquired separately, and then the text features of the address text to be identified may be extracted through the text feature extraction model. And extracting the text features and the information features through a fusion feature extraction model to obtain fusion features corresponding to the address text to be recognized. And finally, acquiring a model through the recognition result, and recognizing the fusion characteristics to obtain a recognition result corresponding to the address text to be recognized.

In the address role recognition, the CRF model may be used for the address role recognition. However, because the CRF model has insufficient text comprehension capability, different description modes can be provided for the same place, but the CRF model cannot effectively identify the same place in different description modes as the same place. For another example, for describing complex long addresses, the CRF model cannot effectively understand all the address roles therein. That is to say, when the CRF model is applied to address identification, the problem of weak generalization ability exists for different source addresses, which results in poor address role identification effect, and the specific expression is in the following 2 aspects:

(1) the ability to understand different source addresses and complex long addresses is insufficient.

Due to the fact that the CRF model is not enough for text understanding, generalization ability for different source addresses is weak when the CRF model is applied. For example, "China telecom from the happy road in the city of Tusangwa county to the west second intersection" and "100 meter telecom from the west of the city of the happy avenue in Kyoto province are descriptions of the same place, but the CRF model cannot effectively identify the two addresses. For another example, for a complex long address, such as "3, 7, 8, east 120 m village and small two places" in the G area of the trade center of new village, 7, 8, of the flood river, north street, in Jilin province, the CRF model cannot effectively understand all the address roles therein.

(2) The text understanding of the map vertical domain is not sufficient.

The address text has fewer words compared with the general NLP text such as news, the semantics of the address text is highly abstract and summarized, and the address text has uniqueness in a vertical field. If the address character is easily identified wrongly only from the perspective of the address text. For example, the address "Handan City Yi Yuan street Haoyuan No. 4-7", wherein "Yiyuan" represents a first district, "Yi Yao" represents a building, and in the address "Yongfeng Lu Yi Yao No. 5 building" with almost the same structure, "Xishan Yi Yao" as a whole represents a district name. The CRF models are not effectively distinguishable from each other in recognition.

In order to solve the above problems, the address recognition model can more accurately obtain the address role recognition result of the address text to be recognized, and the address recognition model in the application can be used for address recognition, so that the address recognition model can recognize address texts with different sources and different characteristics to understand a complex address. The address recognition model of the present application is detailed below:

for the address identification model, Bert + Bilstm + CRF can be used as the network structure of the address role identification model.

In the embodiment of the application, a method of a Bert model, a Bilstm network and a CRF model is adopted, general semantic features are extracted by utilizing pre-training of a Bert layer mass corpus knowledge, the method is suitable for semi-open addresses of different sources, and on the basis, the Bilstm network and the CRF model are utilized to search an optimal sequence by learning the transition probability among output labels so as to understand complex long addresses.

As shown in fig. 4, the network structure of the address recognition model may be a Bert model + a blstm neural network + a CRF model, and the specific process of obtaining the recognition result corresponding to the address text to be recognized through the address recognition model is described by taking the address text to be recognized as "new township small girls" as an example.

As shown in fig. 4, the address text "new station town country mai" to be recognized is input into the Bert model, so as to obtain text features (semantic vectors) corresponding to the "new station town country mai", and then information features (including at least one of basic feature vectors (such as the basic features in fig. 4), domain feature vectors (such as the domain features in fig. 4), and derived feature vectors (such as the derived features in fig. 4) corresponding to the recognition guidance information of the "new station town country mai" are obtained. Then, the obtained text features and information features are input into a blstm neural network (in fig. 4, L in the blstm neural network is a network element of the blstm neural network), and the text features and the information features are fused through a full Connected layer (FC for short), so that fusion features corresponding to the "new station, town and country small girls" are obtained. And finally, inputting the obtained fusion characteristics corresponding to the small farmers in towns and villages of the new station into a CRF model to obtain the identification result corresponding to the small farmers in towns and villages of the new station. It can be seen that the BIS segmentation labels corresponding to the "small girls in towns and towns of the new station" are sequentially as follows: new (B-STRT) station (I-STRT) town (I-STRT) village (I-STRT) small (B-VILG) Country (I-VILG) family (I-VILG), wherein STRT (street) represents village and VILG (village) represents village. "new" BIS tag B-STRT indicates "new" as the start of the town in the address text to be recognized "new station town country little girl", the "station" BIS tag I-STRT indicates "station" as the middle part of the town in the address text to be recognized "new station town country little girl", the "small" BIS tag B-VILG indicates "small" as the start of the village in the address text to be recognized "new station town country little girl", and the "girl" BIS tag I-VILG indicates "small" as the middle part of the village in the address text to be recognized "new station town little girl".

As shown in fig. 5, taking the address text to be recognized as "hai lake district beijing university" as an example, address recognition is performed by using an address recognition model, the address text to be recognized "hai lake district beijing university" is input into a Bert model, text features (semantic vectors) corresponding to "hai lake district beijing university" are obtained, and then information features (including at least one of a basic feature vector (such as the basic feature in fig. 5), a domain feature vector (such as the domain feature in fig. 5), and a derivative feature vector (such as the derivative feature in fig. 5) corresponding to the recognition guidance information of "hai lake district beijing university" are obtained. Then, the obtained text features and information features are input into a Bilstm neural network (in fig. 5, L in the Bilstm neural network is a network element of the Bilstm neural network), and the text features and the information features are fused through a full Connected layer (FC for short), so that fusion features corresponding to "beijing university in hai lake area" are obtained. And finally, inputting the obtained fusion characteristics corresponding to the "Beijing university in Haihu district" into a CRF model to obtain an identification result corresponding to the "Beijing university in Haihu district". It can be seen that the BIS segmentation labels corresponding to "beijing university in hai lake district" are: sea (B-DIS) lake (I-DIS) area (B-POI) North (B-POI) Beijing (I-POI) university (I-POI), wherein DIS (division into dispersions) represents a region and POI (Point of Interes) represents a point of interest. The BIS tag B-DIS of "sea" represents the beginning of the district in the "hai lake district beijing university" as the address text to be recognized, "the BIS tag I-DIS of" lake "represents the middle part of the district in the" hai lake district beijing university "as the address text to be recognized," lake "represents the beginning of the POI in the" hai lake district beijing university "as the address text to be recognized," bei "represents the middle part of the POI in the" hai lake district beijing university "as the" beijing "as the address text to be recognized.

It should be noted that, the address recognition model firstly extracts the features of the address text to be recognized by using the Bert model as a baseline model, so as to obtain the text features of the address text to be recognized. The Bert model is pre-trained on the basis of mass corpus knowledge, so that a large amount of general knowledge of Chinese texts is fused. Meanwhile, because the Bert model has a 12-layer Transformer structure, the Bert model can be finely adjusted on the basis of a small amount of special Fine-tuning samples by relying on the strong learning capability of the Bert model, and the Bert model has strong generalization capability, thereby effectively adapting to address texts with different sources and different characteristics and understanding complex addresses.

On the basis of text features of the address text to be recognized extracted by the Bert model, a Bilstm network + CRF model is introduced to further learn the features and learn the prior knowledge of the map vertical field, so that the defect that the native Bert model cannot solve the problem of optimal condition probability in sequence labeling is overcome. Invalid sequences can be effectively avoided by learning transition probability among the labels. For example, by constructing a domain dictionary to explicitly inform the model that "open element" is the cell name and "Xishan Yi Hospital" is the cell name, the model can be made to have the specific cognitive ability of "domain expert" in the learning process.

The following details the process of improving the vertical domain address text recognition effect by introducing basic features, domain features and derivative features into the address recognition model.

Optionally, word segmentation, part of speech, word position basis features may be introduced. And (4) identifying special components in the address by introducing information such as address proper names and the like in combination with a special word segmentation tool for the map. By introducing Chinese part-of-speech information, the grammatical ability of the text is improved, for example, the user explicitly informs to put (verb) the thank you (verb) structure in a waybill address' the duel camp American cubic district, the heaven city, the sunny district, the American district, the Xiong cell (noun), and the key component of the honor cell can be more easily identified by an address identification model.

Optionally, domain dictionary knowledge may be introduced to improve recognition accuracy. The domain dictionary can be a mined 700wPOI entity dictionary, a dictionary of 180w village nouns, a 44w road dictionary and the like, and candidate characters hit by the domain dictionary can be subjected to disambiguation by being used as prior knowledge and transmitted to an address recognition model. For example, the "open yuan" in the address "Handan' Yi Yuan street open yuan one yard No. 4-7" is represented as a POI entity in the domain dictionary, the "one yard" represents a building, and another address "Haizu Yongfeng Lu one yard No. 5 building" in the "Xishan one yard" with almost the same structure represents a POI entity in the dictionary as a whole, thereby making the address recognition result more accurate.

Optionally, pinyin and wrongly written character derivative features can be introduced to solve the problem of wrongly written characters in the address. For example, the correct understanding (actual retail) of the address "the Wen Xin bridge in Qingpu district 689 of Shanghai city for selling the shoes and clothes for sale" the middle "can be solved by using the pinyin feature, so that the" the shoes and clothes for sale "can be identified as a complete POI entity. For another example, the character shape of the n-gram of the Chinese strokes of horizontal, vertical, left falling, right falling and hook can be used to solve the problem of the shape of a character which is similar to a wrong character, such as the accurate understanding of the second floor of a restaurant in the opposite restaurant of the Haihe Dajie Dachai sea of Beijing (actually, the Angel mansion).

After the address recognition model is obtained in the above manner, the recognition accuracy of the address recognition model can be improved. For example, as shown in fig. 6, taking an address text to be recognized as "the skyway area north aster" as an example, the address recognition model is used to perform address recognition on the "the skyway area north aster", and the BIS segmentation label corresponding to the "the skyway area north aster" can be correctly recognized through the mode of the Bert model, the blstm network and the CRF model: a sky (B-DIS) bridge (I-DIS) region (I-DIS) Qing (B-POI) river (I-POI) north (I-POI) aster (I-POI). The method avoids obtaining the wrong BIS segmentation label corresponding to the clear river north aster in the overpass area only through the primary Bert model: a sky (B-DIS) bridge (I-DIS) region (I-DIS) Qing (B-POI) river (B-POI) north (I-POI) aster (I-POI).

According to the embodiment, the text features of the address text to be recognized can be extracted through the address recognition model based on the text feature extraction model, the information features of the recognition guidance information of the address text to be recognized are combined, the text features and the information features are extracted through the fusion feature extraction model to obtain the fusion features corresponding to the address text to be recognized, finally, the fusion features are recognized through the recognition result acquisition model to obtain the address role recognition result of the address text to be recognized, and the address recognition mode through the address recognition model can fully consider the prior knowledge of basic features, domain features and derivative features in the address recognition field, so that the accuracy of the address recognition result is greatly improved.

In an alternative embodiment, the address recognition model is a first recognition model or a second recognition model, wherein the text feature extraction model of the first recognition model is a model representing a Bert network based on a bidirectional encoder of a Transformer, the fused feature extraction model is a long-short term memory network, and the recognition result acquisition model is a conditional random field model; the text feature extraction model of the second recognition model is an attention mechanism model, the fused feature extraction model is a long-short term memory network, and the recognition result acquisition model is a conditional random field model.

Alternatively, the address recognition model may be the first recognition model or the second recognition model.

For the first recognition model, the text feature extraction model of the first recognition model may be a model representing a Bert network based on a bidirectional encoder of a Transformer, the fused feature extraction model may be a long-short term memory network, and the recognition result acquisition model is a conditional random field model.

For the second recognition model, the text feature extraction model of the second recognition model may be an attention mechanism model, the fused feature extraction model may be a long-short term memory network, and the recognition result obtaining model may be a conditional random field model.

For example, the first recognition model may be a model of a Bert + Bilstm + CRF structure as shown in FIG. 5, and the second recognition model may be a network model of an attribute + Bilstm + CRF structure as shown in FIG. 7.

In the example shown in FIG. 7, the second recognition model comprises an attention mechanism network, a Bilstm network and a CRF model, which are cascaded in sequence. Where the blstm in this example includes each LSTM layer (i.e., L as shown in the figure) and a fully connected layer connected to each LSTM layer, FC as shown in the figure,

the attention network is used for extracting text characteristics of an address text to be recognized, for the example, the address to be recognized is the university of Beijing in Haizu, at the moment, the attention network is used for extracting text characteristics of the text of the university of Beijing in Haizu, basic characteristics in the example are basic information of words of the university of Beijing in Haizu, domain characteristics refer to identification information of target address words of the university of Beijing in Haizu, and derived characteristics are characteristic information of words of the university of Beijing in Haizu. And then inputting the obtained text characteristics and information characteristics (basic characteristics, field characteristics and derivative characteristics) of the Haihai district Beijing university into a BilSTM network, fusing the text characteristics and the information characteristics, and fusing the text characteristics and the information characteristics through a full connection layer FC to obtain the fusion characteristics corresponding to the Haihai district Beijing university. And finally, inputting the obtained fusion characteristics corresponding to the "Beijing university in Haihu district" into a CRF model to obtain an identification result corresponding to the "Beijing university in Haihu district". It can be seen that the BIS segmentation labels (address role identification results) corresponding to "beijing university in hai lake region" are: sea (B-DIS) lake (I-DIS) area (B-POI) North (B-POI) Beijing (I-POI) university (I-POI), wherein DIS (division into dispersions) represents a region and POI (Point of Interes) represents a point of interest.

In an alternative embodiment, the address recognition model is a second recognition model, wherein the second recognition model is trained by: acquiring a training data set; training a first initial neural network model corresponding to the first recognition model based on the training data set to obtain a trained first recognition model; and training the second initial neural network model by taking the first recognition model as a teacher model and taking the second initial neural network model corresponding to the second recognition model as a student model to obtain the second recognition model.

Optionally, for the second recognition model, the second recognition model may be trained in the following manner:

step 1, a training data set is obtained, the training data set comprises a plurality of training data, each data comprises a sample address text and a real class label corresponding to the sample address text, the real class label represents a sample recognition result (real recognition result), the training data in the training data set are input into a first initial neural network model corresponding to a first recognition model one by one, the first initial neural network model is trained, and during training, recognition guide information of the sample address text is obtained; inputting the information characteristics of the sample address text and the guide information into a first initial neural network model to obtain a predicted identification result; and calculating a loss value based on the real category label and the predicted recognition result until the loss value is converged to obtain a trained first recognition model.

And 2, taking the trained first recognition model as a teacher model, taking a second initial neural network model corresponding to the second recognition model as a student model, and training the second initial neural network model to enable the second initial neural network model to perform transfer learning on the first recognition model to obtain a second recognition model.

In particular, knowledge of the first recognition model is transferred to the second recognition model by knowledge distillation, and the first recognition model and the second recognition model may be homogeneous or heterogeneous.

The Bert + Bilstm + CRF address recognition model obtained in the above way can well solve the problem of address text character recognition, but has higher requirements on recognition speed in certain resource-limited scenes, and the efficiency needs to be processed at a high-quality level. In order to solve the problem, a knowledge distillation scheme can be adopted, wherein Bert + Bilstm + CRF is used as a teacher model, and attention + Bilstm + CRF is used as a student model for knowledge transfer, so that the efficiency can be improved by 10 times on the basis of partial precision loss.

For example, referring to fig. 8, a specific technical architecture is shown in fig. 8. The method comprises the steps of taking Bert + Bilstm + CRF shown in the figure 5 as a teacher model, selecting attention + Bilstm + CRF shown in the figure 7 as a student model to carry out knowledge transfer, transferring the knowledge learned by the Bert + Bilstm + CRF model to an attention + Bilstm + CRF small model, and improving the on-line service speed. For the description of the specific principle and structure of Bert + Bilstm + CRF, refer to the description of FIG. 5, and for the description of attention + Bilstm + CRF, refer to the description of FIG. 7, and detailed description thereof is omitted.

When the Bert + Bilstm + CRF is used as a teacher model for migration learning, a Logit model (Logit mode) can be cascaded behind a Bilstm network, and a group of target vectors can be output through the Logit model. When the attention + Bilstm + CRF is subjected to the transfer learning, a group of sample vectors is output through the Bilstm, so that the attention + Bilstm + CRF can learn the knowledge contained in the group of target vectors through the transfer learning (this process is the fitting shown in FIG. 8), for example, the similarity between the group of sample vectors and the group of target vectors satisfies a certain threshold (e.g., 90%), and finally the transfer learning is completed.

Through the embodiment, the knowledge learned by the first recognition model can be migrated to the second recognition model through migration learning, so that the method is suitable for an address recognition scene with higher requirement on the calculation efficiency, and the address recognition efficiency is improved.

The following describes the training process of the address recognition model.

In a possible embodiment, the address recognition model is trained by: constructing a third initial neural network model, wherein the third initial neural network model is programmed by adopting a first programming language; training the third initial neural network model until a training end condition is met, and taking the model obtained after the training end condition is met as a trained third recognition model; storing the model parameters of the third recognition model; adopting a second programming language to construct a fourth initial neural network model with the same structure as the third initial neural network model, and taking the stored model parameters of the third recognition model as the model parameters of the fourth initial neural network model to obtain the address recognition model; the performance of the address recognition model corresponding to the second programming language is better than that of the third recognition model corresponding to the first programming language.

Optionally, the first programming language is adopted for programming, a third initial neural network model is constructed, then the third initial neural network model is trained until the third initial neural network model meets the training end condition, and then the model meeting the training end condition is used as the trained third recognition model. The training end condition may be that the recognition accuracy of the third initial neural network model reaches a certain threshold (e.g., 90%).

And then storing the model parameters of the trained third initial neural network model.

Adopting a second programming language to construct a fourth initial neural network model with the same structure as the third initial neural network model, taking the stored model parameters of the third recognition model as the model parameters of the fourth initial neural network model, and programming the fourth initial neural network model to obtain an address recognition model;

specifically, under the same cpu condition, the recognition speed of the address recognition model corresponding to the second programming language is higher than that of the third recognition model corresponding to the first programming language.

For example, the first programming language may be python and the second programming language may be c + +. In order to improve the performance of the address recognition model, when the method is applied on an actual line, a mode of a model network and model parameters of a third recognition model stored after python training is not adopted any more, but only the model parameters of the third recognition model are stored, the model network part of the third recognition model can be rewritten by c + + to obtain the address recognition model after c + + rewriting, meanwhile, hardware acceleration is carried out by means of a bottom MKL library of Intel, and the efficiency can be further improved by 5 times.

Optionally, in an embodiment, a knowledge distillation scheme may also be adopted, wherein the Bert + Bilstm + CRF is used as a teacher model, and the attention + Bilstm + CRF is used as a student model for knowledge migration, so that the efficiency can be improved by 10 times on the basis of losing part of the precision, and meanwhile, in order to further improve the performance to the utmost, a mode of storing a model network and model weights of attention + Bilstm + CRF after python training is no longer adopted in the actual on-line application, but only the model weights of attention + Bilstm + CRF are stored, the model network part of attention + Bilstm + CRF is rewritten by c + +, and hardware acceleration is performed by means of an underlying MKL library of Intel, and the efficiency can be further improved by 5 times. In this way, efficiency can be greatly improved.

It should be noted that, for the performances of the first recognition model, the second recognition model and the third recognition model, the test can be performed through F1 values (the harmonic mean of the accurate value and the recall ratio). The F1 value is an index used in statistics to measure the accuracy of the model. The method gives consideration to the accuracy and the recall rate of the classification model. The F1 score can be viewed as a harmonic mean of model accuracy and recall with a maximum of 1 and a minimum of 0.

Based on the model training method provided by the embodiment of the application, the address recognition model with high recognition accuracy can be trained. As shown in table 2, method 1 is a conventional address recognition method, and method 2 is a method for performing address recognition by using the first recognition model of the present application. As can be seen from table 2, the F1 value of the model training method provided in the embodiment of the present application is higher, which indicates that the accuracy of the address recognition model trained by the model training method provided in the embodiment of the present application is higher.

TABLE 2 comparison of Performance of Address recognition models obtained by different training methods

Method	F1 value
		Method
1	81.1％
		Method 2	94.9％

Optionally, when the application is actually deployed, the efficiency can be greatly improved on the basis of loss tolerable precision through a knowledge migration and hardware acceleration mode so as to adapt to the requirement in a resource-limited scene, the time consumption can be reduced from 80ms to 3ms on the basis of keeping the F1 value at 90.9% on a test set, and the role identification requirement of the address text in different scenes can be met.

Referring to fig. 9, which is a schematic structural diagram of an address identification and application system based on the address identification method in the present application, the address identification of the system may be implemented by an address identification model (i.e., address role identification shown in the figure), and optionally, as shown in fig. 9, the address identification model may be a model based on Bert + blstm + CRF.

The address identification and application system is mainly divided into an online address segmentation part and an offline resource model part. The main framework of the address identification and application system is shown in fig. 9, wherein the offline resource model is mainly used in the construction process of the address identification model, and the address segmentation part is mainly used in the application of the address identification.

For the resource model, the method is mainly used for pre-training the address recognition model, and mainly comprises the steps of feature construction of the address recognition model, model optimization and training set manufacturing. And processing the basic data, specifically, building a field dictionary library, carrying out standardization processing on the field dictionary and the field dictionary, and building a waste word list which is not suitable for address role recognition. The words in the waste word list can be understood as the words which do not belong to the address recognition field in the address recognition and the words which are updated and eliminated.

After an address recognition model is trained, address recognition can be carried out by using an address segmentation part, for a given address text to be recognized, normalization and basic word segmentation are carried out firstly, then feature generation is carried out by using feature engineering, then prediction is carried out by using a core model (Bert + Bilstm + CRF), and rule base verification (such as determining whether a target address word exists in a rule base) needs to be carried out aiming at an address under a special scene.

The address role identification module adopting the Bert + Bilstm + CRF scheme is a core. In the embodiment of the application, a method of a Bert model, a Bilstm network and a CRF model is adopted, general semantic features are extracted by utilizing pre-training of a Bert layer mass corpus knowledge, the method is suitable for semi-open addresses of different sources, and on the basis, the Bilstm network and the CRF model are utilized to search an optimal sequence by learning the transition probability among output labels so as to understand complex long addresses.

Finally, various services are provided on the application level by combining the results of the address role recognition, such as administrative division recognition (PCD recognition), door address recognition (doorplate address recognition), extraction of a plurality of address keywords (multi-key extraction), address level calculation, processing for generating long and short addresses, addresses in a public security system, and the like.

Specifically, in an application scenario actually involving address identification, the address identification method can be applied to analysis of waybill addresses submitted by users in geocode logistics sorting services, mining of cattle address secret numbers, production and manufacturing of map POI data, processing of addresses in a public security system and the like.

Through the embodiment, the boundary front edge deep learning model Bert can be used as a basis, and the text features in the map field can be extracted and abstracted by using the latest NLP result. A Bilstm network and a CRF model are introduced to a model optimization side to modify a native Bert network so as to improve the effect of an address role recognition result (such as sequence labeling: BIS segmentation labels), and prior knowledge is introduced from the angles of basic text features, domain dictionary features and depth derivative features to improve the generalization capability at a feature optimization side, so that the problem of adapting to different address types, entity ambiguity, spoken words, wrongly-written characters and other difficult points is effectively solved. Meanwhile, in order to solve the problem of high time consumption of high-complexity model prediction, model migration compression is carried out through a knowledge distillation (migration learning) scheme, the prediction time consumption is greatly shortened on the premise of ensuring the role recognition effect, and the resource consumption is reduced. And finally, the double improvement of the address role recognition effect and efficiency is realized.

It should be noted that, in practical application, after the address character recognition result of the address text to be recognized is obtained in the above manner, the multimedia information corresponding to the address character recognition result of the address text to be recognized may also be displayed according to practical requirements, for example, the positioning information, the navigation information, and the like of the address text to be recognized are displayed in the map software.

Referring to fig. 10, fig. 10 is a schematic structural diagram of an address recognition apparatus according to an embodiment of the present application. The address recognition apparatus 1 provided in the embodiment of the present application includes:

the text to be recognized acquisition module 11 is used for acquiring an address text to be recognized;

a guidance information obtaining module 12, configured to obtain recognition guidance information of the address text to be recognized, where the recognition guidance information includes at least one of basic information of a word included in the address text to be recognized, identification information of a target word, or feature information of a word;

and the address identification module 13 is configured to obtain an address role identification result of the address text to be identified according to the address text to be identified and the identification guidance information.

In some possible embodiments, the address identification module is configured to: acquiring text characteristics of the address text to be recognized; acquiring the information characteristics of the identification guide information; fusing the text features and the information features to obtain fused features; and determining the address role recognition result of the address text to be recognized according to the fusion characteristics.

acquiring a training data set; training a first initial neural network model corresponding to the first recognition model based on the training data set to obtain a trained first recognition model; and taking the first recognition model as a teacher model, taking a second initial neural network model corresponding to the second recognition model as a student model, and training the second initial neural network model to obtain the second recognition model.

In some possible embodiments, the model training apparatus includes:

The two devices, namely the address recognition device and the model training device, can be deployed on the same electronic equipment, and can also be deployed on different electronic equipment.

In a specific implementation, the address recognition apparatus 1 may execute the implementation manners provided in the steps in fig. 1 through the built-in functional modules, which may specifically refer to the implementation manners provided in the steps, and are not described herein again.

In the embodiment of the application, firstly, the address text to be recognized is obtained, the recognition guidance information of the address text to be recognized is obtained, and then, the address role recognition result of the address text to be recognized is obtained according to the address text to be recognized and the recognition guidance information. By adopting the above mode, the identification guidance information includes at least one item of basic information of words contained in the address text to be identified, identification information of the target address words, or characteristic information of the words, wherein the identification information of the target address words can represent the identification result of the target address words, so that the finally obtained address role identification result of the address text to be identified can be combined with two factors of the information of the address text to be identified and the identification guidance information, and the identification accuracy of the address text to be identified is greatly improved.

Referring to fig. 11, fig. 11 is a schematic structural diagram of an electronic device provided in an embodiment of the present application. As shown in fig. 11, the electronic device 1000 in the present embodiment may include: the processor 1001, the network interface 1004, and the memory 1005, and the electronic device 1000 may further include: a user interface 1003, and at least one communication bus 1002. Wherein a communication bus 1002 is used to enable connective communication between these components. The user interface 1003 may include a Display screen (Display) and a Keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a standard wireless interface. The network interface 1004 may optionally include a standard wired interface, a wireless interface (e.g., WI-FI interface). The memory 1004 may be a high-speed RAM memory or a non-volatile memory (e.g., at least one disk memory). The memory 1005 may optionally be at least one memory device located remotely from the processor 1001. As shown in fig. 11, a memory 1005, which is a kind of computer-readable storage medium, may include therein an operating system, a network communication module, a user interface module, and a device control application program.

In the electronic device 1000 shown in fig. 11, the network interface 1004 may provide a network communication function; the user interface 1003 is an interface for providing input to a user; and the processor 1001 may be used to invoke a device control application stored in the memory 1005 to implement:

acquiring an address text to be identified;

It should be understood that in some possible embodiments, the processor 1001 may be a Central Processing Unit (CPU), and the processor may be other general purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and the like. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The memory may include both read-only memory and random access memory, and provides instructions and data to the processor. The portion of memory may also include non-volatile random access memory. For example, the memory may also store device type information.

In a specific implementation, the electronic device 1000 may execute, through each built-in functional module thereof, the implementation manner provided in each step in fig. 1, which may be specifically referred to as the implementation manner provided in each step, and is not described herein again.

An embodiment of the present application further provides a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and the computer program is executed by a processor to implement the method provided in each step in fig. 1, which may specifically refer to the implementation manner provided in each step, and is not described herein again.

The computer readable storage medium may be an internal storage unit of the task processing device provided in any of the foregoing embodiments, for example, a hard disk or a memory of an electronic device. The computer readable storage medium may also be an external storage device of the electronic device, such as a plug-in hard disk, a Smart Memory Card (SMC), a Secure Digital (SD) card, a flash memory card (flash card), and the like, provided on the electronic device. The computer readable storage medium may further include a magnetic disk, an optical disk, a read-only memory (ROM), a Random Access Memory (RAM), and the like. Further, the computer readable storage medium may also include both an internal storage unit and an external storage device of the electronic device. The computer-readable storage medium is used for storing the computer program and other programs and data required by the electronic device. The computer readable storage medium may also be used to temporarily store data that has been output or is to be output.

Embodiments of the present application provide a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the electronic device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the method provided by the steps of fig. 1.

The terms "first", "second", and the like in the claims and in the description and drawings of the present application are used for distinguishing between different objects and not for describing a particular order. Furthermore, the terms "include" and "have," as well as any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, article, or electronic device that comprises a list of steps or elements is not limited to only those steps or elements recited, but may alternatively include other steps or elements not expressly listed or inherent to such process, method, article, or electronic device. Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. It is explicitly and implicitly understood by one skilled in the art that the embodiments described herein can be combined with other embodiments. The term "and/or" as used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

Those of ordinary skill in the art will appreciate that the elements and algorithm steps of the examples described in connection with the embodiments disclosed herein may be embodied in electronic hardware, computer software, or combinations of both, and that the components and steps of the examples have been described in a functional general in the foregoing description for the purpose of illustrating clearly the interchangeability of hardware and software. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.

The above disclosure is only for the purpose of illustrating the preferred embodiments of the present application and is not to be construed as limiting the scope of the present application, which is therefore intended to be covered by the present application with all equivalent modifications made to the claims of the present application.

Claims

1. An address identification method, the method comprising:

acquiring an address text to be identified;

acquiring identification guide information of the address text to be identified, wherein the identification guide information comprises at least one item of basic information of words contained in the address text to be identified, identification information of target address words or characteristic information of words, and the identification information of the target address words represents identification results of the target address words;

according to the address text to be recognized and the recognition guidance information, obtaining an address role recognition result of the address text to be recognized through an address recognition model, wherein the address recognition model is obtained through training in the following mode:

constructing a third initial neural network model, wherein the third initial neural network model is programmed by adopting a first programming language;

training the third initial neural network model until a training end condition is met, and taking the model obtained after the training end condition is met as a trained third recognition model;

storing model parameters of the third recognition model;

adopting a second programming language to construct a fourth initial neural network model with the same structure as the third initial neural network model, and taking the stored model parameters of the third recognition model as the model parameters of the fourth initial neural network model to obtain the address recognition model;

wherein the performance of the address recognition model corresponding to the second programming language is better than the performance of the third recognition model corresponding to the first programming language.

2. The method according to claim 1, wherein the recognition guidance information includes identification information of a target address word; the method further comprises the following steps:

constructing an address keyword word bank, wherein the address keyword word bank comprises address keywords and identification information of the address keywords;

the acquiring of the identification guidance information of the address text to be identified includes:

determining address keywords hit by the address text to be recognized in the address keyword word bank based on the address keyword word bank, and determining the hit address keywords as target address words;

and obtaining the identification information of the target address word based on the address keyword word bank.

3. The method according to claim 1 or 2, wherein the basic information of the word comprises at least one of part-of-speech information of the word or position information of the word in the address text to be recognized.

4. The method of claim 1 or 2, wherein the characteristic information of the word includes at least one of pinyin information of the word or morphological information of the word.

5. The method according to claim 1 or 2, wherein obtaining the address character recognition result of the address text to be recognized according to the address text to be recognized and the recognition guidance information comprises:

acquiring text characteristics of the address text to be recognized;

acquiring information characteristics of the identification guidance information;

fusing the text features and the information features to obtain fused features;

and determining an address role recognition result of the address text to be recognized according to the fusion characteristics.

6. The method of claim 5, wherein said fusing the textual features and the informational features to obtain fused features comprises:

and splicing the text features and the information features, and determining the spliced text features and the spliced information features as the fusion features.

7. The method according to claim 1 or 2, wherein the address recognition model comprises a text feature extraction model, a fusion feature extraction model and a recognition result acquisition model which are sequentially cascaded;

the obtaining of the address role recognition result of the address text to be recognized according to the address text to be recognized and the recognition guidance information comprises the following steps:

respectively acquiring information characteristics of each piece of identification guidance information;

extracting the text features of the address text to be recognized through the text feature extraction model based on the address text to be recognized;

extracting fusion features corresponding to the address text to be recognized through the fusion feature extraction model based on the text features and the information features;

and obtaining a model through the identification result based on the fusion characteristics to obtain an address role identification result of the address text to be identified.

8. The method of claim 7, wherein the address recognition model is a first recognition model or a second recognition model, wherein the text feature extraction model of the first recognition model is a model representing a Bert network based on a bidirectional encoder of a Transformer, the fused feature extraction model is a long-short term memory network, and the recognition result obtaining model is a conditional random field model;

the text feature extraction model of the second recognition model is an attention mechanism model, the fusion feature extraction model is a long-term and short-term memory network, and the recognition result acquisition model is a conditional random field model.

9. The method of claim 8, wherein the address recognition model is a second recognition model, and wherein the second recognition model is trained by:

acquiring a training data set;

training a first initial neural network model corresponding to the first recognition model based on the training data set to obtain the trained first recognition model;

and taking the first recognition model as a teacher model, taking a second initial neural network model corresponding to the second recognition model as a student model, and training the second initial neural network model to obtain the second recognition model.

10. An address identification apparatus, the apparatus comprising:

the instruction information acquisition module is used for acquiring identification instruction information of the address text to be identified, wherein the identification instruction information comprises at least one item of basic information of words contained in the address text to be identified, identification information of target words or characteristic information of words;

the address recognition module is used for obtaining an address role recognition result of the address text to be recognized through an address recognition model according to the address text to be recognized and the recognition guidance information, wherein the address recognition model is obtained by training in the following way:

training the third initial neural network model until a training end condition is met, and taking the model meeting the training end condition as a trained third recognition model;

storing model parameters of the third recognition model;

wherein the address recognition model corresponding to the second programming language has better performance than the third recognition model corresponding to the first programming language.

11. The apparatus of claim 10, further comprising:

the address keyword lexicon building module is used for building an address keyword lexicon, wherein the address keyword lexicon comprises address keywords and identification information of the address keywords;

the guidance information obtaining module is configured to: determining address keywords hit by the address text to be recognized in the address keyword word bank based on the address keyword word bank, and determining the hit address keywords as target address words; and obtaining the identification information of the target address words based on the address keyword word bank.

12. An electronic device comprising a processor and a memory, the processor and the memory being interconnected;

the memory is used for storing a computer program;

the processor is configured to perform the method of any of claims 1 to 9 when the computer program is invoked.

13. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program which is executed by a processor to implement the method of any one of claims 1 to 9.