Point of interest information identification model training, identification method, device and electronic equipment

Abstract

The application discloses a kind of point of interest information identification model training, identification method, device and electronic equipment, by the way of multi-task learning, when training point of interest information identification model, not only obtain the first point of interest information and semantic component label of training point of interest information identification model, but also obtain associated task label;After the feature extraction of first point of interest information, the first feature data obtained is input into second sub-model to identify point of interest information, and is input into associated semantic task model to process associated semantic task;The loss value obtained by comprehensively identifying point of interest information and associated semantic task processing two parts is used to update the parameters of point of interest information identification model.So, it can improve the semantic understanding and analysis capability of point of interest information identification model by the way of multi-task joint learning.The technical scheme of the application can be applied to the technical fields of big data, intelligent transportation system and intelligent vehicle-road cooperative system.

Description

Technical Field

[0001] This application relates to the field of big data technology, and in particular to a training and recognition method, apparatus and electronic device for a point of interest information recognition model. Background Technology

[0002] Points of interest (POIs) are generally landmarks, attractions, or any locations with physical significance in a geographic information system (GIS). PPI information provides convenient and practical guidance and is an essential element in many navigation applications. Various entity hierarchies may exist within PPI information, such as master-child relationships, which are crucial for building GIS and map knowledge graphs.

[0003] In related technologies, named entity recognition is generally used to identify entity membership relationships in point-of-interest (POI) information. However, current POI recognition models are often trained on complex corpora from various scenarios, lacking specificity for the POI recognition context, having weak semantic analysis capabilities, and resulting in low accuracy. Summary of the Invention

[0004] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0005] This application provides a training and recognition method, apparatus, and electronic device for a point of interest (POI) information recognition model, which can improve the accuracy of POI information recognition.

[0006] On one hand, embodiments of this application provide a training method for an interest point information recognition model, wherein the interest point information recognition model includes a first sub-model and a second sub-model; the training method includes:

[0007] Obtain first point of interest information, semantic component tags corresponding to words in the first point of interest information, and associated task tags of the first point of interest information; wherein, the semantic component tags are used to characterize the membership relationship of entities in the first point of interest information, and the associated task tags are used to characterize the task results of associated tasks of the first point of interest information.

[0008] The first point of interest information is input into the first sub-model for natural language processing to obtain the first feature data corresponding to the first point of interest information.

[0009] The first feature data is input into the second sub-model, and semantic component analysis is performed on the first interest point information to obtain the semantic component analysis results corresponding to the words in the first interest point information.

[0010] The first feature data is input into the association task model, and the first point of interest information is processed by the association task to obtain the association task processing result corresponding to the first point of interest information.

[0011] Based on the semantic component analysis results and the semantic component labels, as well as the association task processing results and the association task labels, the training loss value is determined;

[0012] Based on the loss value, the parameters of the first sub-model and the second sub-model are updated.

[0013] On the other hand, embodiments of this application also provide a method for identifying point-of-interest (POI) information, used to identify POI information through an POI information identification model, wherein the POI information identification model includes a first sub-model and a second sub-model; the identification method includes:

[0014] Obtain information on third points of interest;

[0015] The third point of interest information is input into the first sub-model for natural language processing to obtain the second feature data corresponding to the third point of interest information.

[0016] The second feature data is input into the second sub-model, and semantic component analysis is performed on the third point of interest information to obtain semantic component recognition results, wherein the semantic component recognition results are used to characterize the membership relationship of entities in the third point of interest information;

[0017] The interest point information recognition model is trained using the aforementioned interest point information recognition model training method.

[0018] On the other hand, embodiments of this application also provide a training device for an interest point information recognition model, wherein the interest point information recognition model includes a first sub-model and a second sub-model; the training device includes:

[0019] The acquisition module is used to acquire first point of interest information, semantic component tags corresponding to words in the first point of interest information, and associated task tags of the first point of interest information; wherein, the semantic component tags are used to characterize the membership relationship of entities in the first point of interest information, and the associated task tags are used to characterize the task results of associated tasks of the first point of interest information.

[0020] The first processing module is used to input the first point of interest information into the first sub-model for natural language processing to obtain the first feature data corresponding to the first point of interest information.

[0021] The analysis module is used to input the first feature data into the second sub-model, perform semantic component analysis on the first interest point information, and obtain the semantic component analysis results corresponding to the words in the first interest point information;

[0022] The second processing module is used to input the first feature data into the association task model, perform association task processing on the first point of interest information, and obtain the association task processing result corresponding to the first point of interest information.

[0023] The calculation module is used to determine the training loss value based on the semantic component analysis results and the semantic component labels, as well as the association task processing results and the association task labels;

[0024] An update module is used to update the parameters of the first sub-model and the second sub-model based on the loss value.

[0025] Furthermore, the association task model includes a name category prediction model; the second processing module is specifically used for:

[0026] The first feature data is input into the name category prediction model to predict the name category of the first point of interest information, thereby obtaining the name category prediction result corresponding to the first point of interest information.

[0027] Furthermore, the associated task model includes a name similarity discrimination model; the second processing module is specifically used for:

[0028] First target interest information and second target interest information are selected from the first interest information;

[0029] From the first feature data, determine the first target data obtained by processing the first target interest point information, and the second target data obtained by processing the second target interest point information;

[0030] The first target data and the second target data are input into the name similarity discrimination model to judge the name similarity between the first target interest information and the second target interest information, and the name similarity discrimination result between the first target interest information and the second target interest information is obtained.

[0031] Furthermore, the aforementioned calculation module is specifically used for:

[0032] Based on the semantic component analysis results and the semantic component labels, a first sub-loss value is determined for semantic component analysis of the first point of interest information;

[0033] Based on the associated task processing result and the associated task label, a second sub-loss value is determined for the associated task processing of the first point of interest information;

[0034] The first sub-loss value and the second sub-loss value are weighted and summed to obtain the training loss value.

[0035] Furthermore, the aforementioned calculation module is specifically used for:

[0036] The first value is obtained by multiplying the first sub-loss value and the first weight.

[0037] The second value is obtained by multiplying the second sub-loss value and the second weight.

[0038] The training loss value is obtained by summing the first value and the second value.

[0039] Wherein, the first weight is greater than the second weight.

[0040] Furthermore, the training device for the aforementioned point of interest information recognition model also includes:

[0041] The detection module is used to detect whether each piece of information about the first point of interest is missing a corresponding semantic component label or associated task label;

[0042] The jump module is used to skip the step of inputting the first feature data into the second sub-model to perform semantic component analysis on the first interest point information when the corresponding semantic component label is missing; or...

[0043] This is used to skip the step of inputting the first feature data into the association task model and performing association task processing on the first point of interest information when the corresponding associated task label is missing from the first point of interest information.

[0044] Furthermore, the training device for the above-mentioned point of interest information recognition model also includes a pre-training module, which is used for:

[0045] Acquire batches of second points of interest information;

[0046] The information of each second interest point is masked to obtain the training dataset of interest point information.

[0047] The first sub-model is pre-trained based on the training dataset of the interest point information and the second interest point information.

[0048] Furthermore, the aforementioned pre-trained module is used for at least one of the following:

[0049] The second point of interest information is subjected to random masking; or,

[0050] Masking is applied to several words in the second point of interest information; or,

[0051] Masking is applied to several entity information in the second point of interest information.

[0052] On the other hand, embodiments of this application also provide an interest point information identification device, used to identify interest point information through an interest point information identification model, wherein the interest point information identification model includes a first sub-model and a second sub-model; the device includes:

[0053] The second acquisition module is used to acquire information about the third point of interest.

[0054] The third processing module is used to input the third point of interest information into the first sub-model for natural language processing to obtain the second feature data corresponding to the third point of interest information.

[0055] The second analysis module is used to input the second feature data into the second sub-model, perform semantic component analysis on the third point of interest information, and obtain semantic component recognition results, wherein the semantic component recognition results are used to characterize the membership relationship of entities in the third point of interest information;

[0056] The interest point information recognition model is trained using the aforementioned interest point information recognition model training method.

[0057] Furthermore, the aforementioned second analysis module is specifically used for:

[0058] The third point of interest information is searched and matched in a preset dictionary to obtain the third feature data corresponding to the third point of interest information;

[0059] The second feature data and the third feature data are subjected to feature fusion processing to obtain the fourth feature data;

[0060] The fourth feature data is input into the second sub-model, and semantic component analysis is performed on the third interest point information to obtain the semantic component recognition result.

[0061] Furthermore, the aforementioned second analysis module is specifically used for:

[0062] The third point of interest information is segmented into words;

[0063] Based on the results of the word segmentation process, feature fusion processing is performed on the character feature vectors to obtain the second word feature vectors corresponding to the words in the third interest point information.

[0064] The first word feature vector and the second word feature vector are subjected to feature fusion processing to obtain the fourth feature data.

[0065] On the other hand, embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described training method or interest point information recognition method for the interest point information recognition model.

[0066] On the other hand, embodiments of this application also provide a computer-readable storage medium storing a program, which is executed by a processor to implement the above-described training method or interest point information recognition method for the interest point information recognition model.

[0067] On the other hand, embodiments of this application also provide a computer program product, which includes a computer program stored in a computer-readable storage medium. The processor of an electronic device reads the computer program from the computer-readable storage medium and executes the computer program, causing the electronic device to perform a training method or an interest point information recognition method that implements the above-described interest point information recognition model.

[0068] The embodiments of this application include at least the following beneficial effects: By employing a multi-task learning approach, when training the interest point information recognition model, not only are the first interest point information and its corresponding semantic component labels obtained for training the model, but also the associated task labels of the first interest point information are acquired. Then, after feature extraction from the first interest point information, the resulting first feature data is input into a second sub-model for interest point information recognition, and also input into an associated task model for associated semantic task processing. Next, the loss values ​​obtained from both the interest point information recognition and the associated semantic task processing are combined to update the parameters of the interest point information recognition model. Thus, through a multi-task joint learning approach, the semantic understanding and analysis capabilities of the interest point information recognition model can be improved. When using this model for subsequent interest point information recognition, this helps to improve the accuracy of the obtained recognition results.

[0069] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0070] The accompanying drawings are used to provide a further understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0071] Figure 1 A task architecture diagram for a point-of-interest (POI) information processing task provided in this application embodiment;

[0072] Figure 2 An executable provided for embodiments of this application Figure 1 The model architecture diagram of the task architecture shown;

[0073] Figure 3 A schematic diagram illustrating the processing flow of an interest point information recognition model provided in an embodiment of this application;

[0074] Figure 4 A schematic diagram of an implementation environment provided for an embodiment of this application;

[0075] Figure 5 A schematic diagram illustrating another implementation environment provided for an embodiment of this application;

[0076] Figure 6 A flowchart illustrating a training method for an interest point information recognition model provided in an embodiment of this application;

[0077] Figure 7 A schematic diagram of the processing flow of another point of interest information recognition model provided in an embodiment of this application;

[0078] Figure 8 A schematic diagram illustrating the pre-training process of an ALBERT model provided in this application embodiment;

[0079] Figure 9 This is a schematic diagram illustrating the training of an interest point information recognition model based on multi-task learning, as provided in an embodiment of this application.

[0080] Figure 10 A schematic diagram comparing the F1 scores of multiple models provided in this application embodiment;

[0081] Figure 11 A flowchart illustrating a point-of-interest (POI) information identification method provided in an embodiment of this application;

[0082] Figure 12 This application provides a schematic diagram of the interface of a terminal device according to an embodiment of the present application.

[0083] Figure 13 A schematic diagram of the interface of a map navigation software provided in an embodiment of this application;

[0084] Figure 14 A schematic diagram of the interface of another map navigation software provided in an embodiment of this application;

[0085] Figure 15A schematic diagram of the structure of a training device for an interest point information recognition model provided in an embodiment of this application;

[0086] Figure 16 This is a structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0087] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0088] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0089] Before providing a detailed description of the embodiments of this application, some of the nouns and terms involved in the embodiments of this application will be explained first. The nouns and terms involved in the embodiments of this application are subject to the following interpretations.

[0090] A Geographic Information System (GIS) is a computer system used to collect, store, manage, process, retrieve, analyze, and represent geospatial data. It is a system for analyzing and processing massive amounts of geographic data.

[0091] A Point of Interest (POI) is a landmark or scenic spot in a geographic information system, used to identify the government departments, commercial establishments of various industries (gas stations, department stores, supermarkets, restaurants, hotels, convenience stores, hospitals, etc.), tourist attractions (parks, public toilets), historical sites, and transportation facilities (various stations, parking lots, speed cameras, speed limit signs), etc.

[0092] Artificial Intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to have perception, reasoning, and decision-making capabilities. AI technology is a comprehensive discipline involving a wide range of fields, encompassing both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0093] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence. Its applications span all areas of artificial intelligence. Machine learning (deep learning) typically includes techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0094] Natural Language Processing (NLP) is an important field within computer science and artificial intelligence. It studies the theories and methods for enabling effective communication between humans and computers using natural language. NLP is a science that integrates linguistics, computer science, and mathematics. Since the natural language involved in this field is the language people use in daily life, it is closely related to linguistic research. NLP techniques typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs.

[0095] Big data refers to data sets that cannot be captured, managed, and processed within a certain timeframe using conventional software tools. It represents massive, rapidly growing, and diverse information assets that require new processing models to achieve stronger decision-making, insightful discovery, and process optimization capabilities. With the advent of the cloud era, big data has attracted increasing attention. Big data requires specialized technologies to effectively process large amounts of data within a tolerable timeframe. Technologies suitable for big data include massively parallel processing databases, data mining, distributed file systems, distributed databases, cloud computing platforms, the internet, and scalable storage systems.

[0096] Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and cryptographic algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked using cryptographic methods. Each block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying platform, a platform product service layer, and an application service layer. Blockchains can include public blockchains, consortium blockchains, and private blockchains. A public blockchain is one where anyone can access the blockchain network at any time to read data, send data, or compete for ledger entries. A consortium blockchain is one jointly managed by several organizations or institutions. A private blockchain has a degree of centralized control; the right to write to the ledger is controlled by a specific organization or institution, and data access and use are subject to strict permission management.

[0097] Intelligent Traffic Systems (ITS), also known as Intelligent Transportation Systems, effectively integrate advanced science and technology (information technology, computer technology, data communication technology, sensor technology, electronic control technology, automatic control theory, operations research, artificial intelligence, etc.) into transportation, service control, and vehicle manufacturing. This strengthens the connection between vehicles, roads, and users, thereby forming a comprehensive transportation system that ensures safety, improves efficiency, enhances the environment, and saves energy.

[0098] The training and recognition methods for the point of interest (POI) information recognition model provided in this application embodiment involve artificial intelligence (AI) technology, primarily natural language processing (NLP) and machine learning / deep learning technologies within the AI ​​field. Specifically, the method provided in this application embodiment can use NLP technology to process PPI information, thereby facilitating the identification of the membership relationships of entities within the PPI information. During the identification of the membership relationships of entities within the PPI information, a model trained through machine learning / deep learning can be used to perform feature data extraction, classification tasks, etc., and a training method for the PPI information recognition model is also provided.

[0099] The methods provided in this application embodiment can be executed in application scenarios such as big data, intelligent transportation systems, and intelligent vehicle-road cooperative systems. For example, in big data application scenarios, there is a need to statistically analyze the popularity of tourist attractions. In this scenario, the methods provided in this application embodiment can be used to statistically analyze the number of views of entities of the same type as those searching for points of interest on the internet, thereby obtaining the analysis results corresponding to each point of interest. In the application scenarios of intelligent transportation systems and intelligent vehicle-road cooperative systems, by analyzing the affiliation relationships of entities in each point of interest information, it is possible to generate a more refined geographic information system, thereby facilitating the provision of navigation guidance.

[0100] Of course, it should be noted that the above application scenarios are merely illustrative and do not imply any limitation on the actual application of the methods in the embodiments of this application. Those skilled in the art will understand that the methods provided in the embodiments of this application can be used to perform specified tasks in different application scenarios.

[0101] Furthermore, it should be noted that in various specific embodiments of this application, when processing data related to the identity or characteristics of the target object, such as the target object's information, behavioral data, historical data, and location information, is required, the target object's permission or consent will be obtained first. Moreover, the collection, use, and processing of this data will comply with the relevant laws, regulations, and standards of the relevant countries and regions. Additionally, when embodiments of this application require obtaining sensitive information about the target object, separate permission or consent from the target object will be obtained through pop-ups or redirection to a confirmation page. Only after obtaining the target object's separate permission or consent will the necessary target object-related data for the normal operation of the embodiments of this application be obtained.

[0102] Points of interest (POIs) are generally landmarks, attractions, or any physical locations in a geographic information system (GIS), such as schools, banks, restaurants, gas stations, hospitals, and supermarkets. POI information includes the name of the POI, such as "Experimental Middle School First Cafeteria" or "Civilized Community Convenience Supermarket." POI information provides convenient and practical guidance and is an essential element in many navigation applications.

[0103] In point-of-interest (POI) information, there may sometimes be multiple POIs, each of which is an entity. For example, the aforementioned "Experimental Middle School First Cafeteria" includes two POIs: "Experimental Middle School" and "First Cafeteria". Similarly, the POI information for "East Gate of Technology Exchange Building" also includes two POIs: "Technology Exchange Building" and "East Gate". Of course, the number of entities contained in each POI information is not fixed, and this application does not impose any restrictions on this. For POI information containing multiple POI entities, there may be various entity affiliation relationships, such as master-child relationships. Master-child relationships are used to characterize the dependency relationship between POIs. For example, in the POI information for "Railway Campus of University A", "University A" and "Railway Campus" are two entities, and "Railway Campus" belongs to "University A". In this case, "University A" and "Railway Campus" have a master-child relationship, with "University A" being the master of "Railway Campus" and conversely, "Railway Campus" being a child of "University A". In addition to the master-child relationship mentioned above, other entity affiliations can be set, such as the affiliation between a brand (which can also be a master) and a branch, etc., which will not be elaborated on here. These entity affiliations are very important for building geographic information systems and map knowledge graphs.

[0104] In related technologies, named entity recognition is generally used to identify entity membership relationships in point-of-interest (POI) information. However, current POI recognition models are often trained on complex corpora from various scenarios, lacking specificity for the POI recognition context, having weak semantic analysis capabilities, and resulting in low accuracy.

[0105] Based on this, embodiments of this application provide a training method, apparatus, electronic device, and storage medium for an interest point information recognition model. The training method employs a multi-task learning approach. When training the interest point information recognition model, it acquires not only the first interest point information and its corresponding semantic component labels used for training the model, but also the associated task labels of the first interest point information. Then, after feature extraction from the first interest point information, the resulting first feature data is input into a second sub-model for interest point information recognition and into an associated task model for semantic task processing. Finally, the loss values ​​obtained from both the interest point information recognition and the associated semantic task processing are combined to update the parameters of the interest point information recognition model. Thus, through a multi-task joint learning approach, the semantic understanding and analysis capabilities of the interest point information recognition model can be improved. This is beneficial for improving the accuracy of the recognition results when using the model for subsequent interest point information recognition.

[0106] The following section, in conjunction with the accompanying drawings, first introduces and explains the point of interest information processing task related to the application of the embodiments of this application.

[0107] Reference Figure 1 , Figure 1 This diagram illustrates a task architecture for point-of-interest (POI) information processing. Generally, POI information processing tasks can include multiple dimensions, such as... Figure 1 The examples shown primarily include three dimensions: semantic analysis, syntactic analysis, and lexical analysis of interest point information. Developers in this field can use this as a basis... Figure 1 The task architecture in the text is used to develop a business system capable of identifying points of interest information.

[0108] Specifically, in Figure 1In the task architecture, semantic analysis is mainly used to identify the functional and semantic tags of relevant words in point-of-interest (POI) information. For example, for POI information such as "Sinopec Senmei Gas Station," semantic analysis can identify which words are core words and which are secondary words, thus facilitating the identification of the main attributes of the location corresponding to the POI information. Semantic tag identification is mainly used to distinguish the semantic meanings represented by each word in the POI information, that is, what types of information each word represents. For example, for POI information such as "KFC Wanliu Shopping Center," semantic tag identification can identify that "KFC" and "Wanliu" represent "brand" type information, while "shopping" and "center" represent "business" type information. Syntactic analysis is the main task dimension applied by the method provided in this application embodiment; it can also be called the name understanding task. This task aims to analyze the semantic components in POI information and can include sub-tasks such as main component analysis, subordinate component analysis, and hierarchical identification. Principal component analysis aims to identify the main point and sub-points, while subordinate component analysis aims to identify aliases, branches, and descriptions. For example, for "Sinopec Senmei Gas Station," "Sinopec Senmei" can be identified as the main point, and "gas station" as the sub-point; for the name "KFC (Wanliu Shopping Center)," "KFC" can be identified as the main point, and "Wanliu Shopping Center" as the branch. Lexical analysis focuses more on processing words in the point of interest information. For example, after segmenting the point of interest information, the "word role" of each word can be determined. This process is similar to the aforementioned method of determining semantic tags and will not be elaborated here. In some cases, the point of interest information may not include the commonly used name of the actual main point, but rather synonyms or aliases of the commonly used name of the main point. For example, the commonly used point of interest information is "Tsinghua University School of Physics," but the actual search might be for "Tsinghua School of Physics." Therefore, lexical analysis can be used to expand the words in the point of interest information with synonyms, improving the comprehensiveness and accuracy of point of interest identification and retrieval. In the process of lexical analysis, it is generally necessary to segment the interest point information to obtain multiple words that make up the interest point information. Here, the segmentation method used can be full segmentation, fine-grained segmentation, or coarse-grained segmentation, etc. Different segmentation granularities may produce different segmentation results. The specific segmentation granularity can be selected according to the actual application requirements, and this application does not impose any restrictions on it.

[0109] Combination Figure 1 The diagram illustrates a task architecture for a point-of-interest (POI) information processing task. This application provides an executable... Figure 1 The task architecture model diagram shown is as follows: Figure 2 As shown. Correspondingly, in Figure 2The model architecture mainly consists of three parts. The semantic analysis task can include sub-tasks such as word weight calculation, word category prediction, and semantic similarity calculation. In terms of models, models capable of similarity learning, behavioral feature extraction, and linguistic feature extraction can be used, and multi-model fusion can be employed to improve the accuracy of the results. The lexical analysis task can utilize relevant word segmentation algorithms for segmentation at various granularities, and can perform word role matching based on various dictionaries or word role recognition based on models.

[0110] Syntactic analysis is the primary task applied to the methods provided in the embodiments of this application. Figure 2 As can be seen from this, the model layer in this part can adopt a language model combined with multi-task learning. Based on the feature data extracted through point of interest information, syntactic analysis is performed on the point of interest information to identify the main points, sub-points, and semantic components such as branches and aliases in the point of interest information.

[0111] It should be noted that, Figure 1 The task architecture shown Figure 2 The illustrated model architecture is only used to provide an overall overview of the point of interest (POI) information processing task and does not imply that the POI information recognition model in this application needs to process all task types or include all components. Furthermore, those skilled in the art can adjust the above task architecture or model architecture according to actual needs, such as deleting or adding certain technical content, and this application does not impose any restrictions on this.

[0112] In related technical fields, models generally need to be trained before being put into use to ensure that their parameters meet predetermined requirements. This application proposes a training method for an interest point information recognition model. This interest point information recognition model is mainly used for executing... Figure 1 For ease of description, the following is a brief introduction to an exemplary structure of the model and its recognition principle in application, which is used for syntactic analysis processing tasks.

[0113] Reference Figure 3The interest point information recognition model provided in this application embodiment mainly includes two parts, referred to as the first sub-model 310 and the second sub-model 320, respectively. The first sub-model 310 is used to extract feature data from the input interest point information, and the second sub-model 320 is used to identify the membership relationships of entities in the interest point information based on the extracted feature data, thereby outputting a prediction result. Specifically, in some embodiments, the first sub-model 310 can employ a language model, which can extract semantic features of the interest point information. For example, word embedding can be used to semantically encode the interest point information. Since the interest point information is continuous, the recognition model needs to output the recognition results corresponding to each word during prediction. Therefore, the second sub-model 320 in this application embodiment can employ a sequence labeling model. Sequence labeling models are highly efficient in natural language processing tasks. For example, for an input sequence of length N, when it is necessary to label each element to obtain an output of length N, using a sequence labeling model can greatly improve the speed and efficiency of model processing. For example, in some embodiments, the first sub-model 310 described above may employ the BERT model, the Long Short Term Memory (LSTM) model, and their derivative models such as the ALBERT model and the BiLSTM model; the second sub-model 320 described above may employ the Hidden Markov Model (HMM), the Conditional Random Field (CRF), etc., and this application does not impose specific limitations on this.

[0114] Of course, it should be noted that the interest point information recognition model provided in this application embodiment, in addition to the first and second sub-models mentioned above, may also include other parts, such as a model structure for filtering input data or optimizing and selecting intermediate feature data (e.g., a dropout layer). For example, the first sub-model in a certain interest point information recognition model is a BiLSTM model, the second sub-model is a Conditional Random Field, and it also includes an embedding layer and an encoding layer. In the embedding layer, each word of the input interest point information is transformed into a vector. These vectors are input into the BiLSTM model, and feature extraction is performed based on context information to obtain higher-dimensional feature vectors for the words. Then, the encoding layer can map the high-dimensional feature vectors into low-dimensional vectors that conform to the label dimensions, so that each word can be labeled based on the low-dimensional vectors by the Conditional Random Field. The labeled data is the recognition result output by the model.

[0115] The specific implementation methods of the embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0116] Reference Figure 4 , Figure 4This illustration shows a schematic diagram of an implementation environment for the training method of the point of interest information recognition model provided in this application embodiment. In this implementation environment, the main hardware and software components involved include a first terminal 410.

[0117] Specifically, the first terminal 410 may have a relevant map application installed and be configured with a point-of-interest (POI) information recognition model. This POI recognition model can obtain POI information from the map application and provide the corresponding recognition results back to the map application, enabling the map application to provide navigation results based on the recognition results given by the POI information recognition model. Figure 4 As shown, the point of interest information recognition model can be trained and applied in the first terminal 410.

[0118] Additionally, refer to Figure 5 , Figure 5 This is a schematic diagram of another implementation environment for the training method of the point of interest information recognition model provided in the embodiments of this application. The main hardware and software components involved in this implementation environment include a second terminal 510 and a server 520, wherein the second terminal 510 and the server 520 are connected in communication.

[0119] Specifically, the server 520 may have a relevant map application installed and configured with a point-of-interest (POI) recognition model. This map application can receive navigation requests and send the navigation results to the second terminal 510 for display and browsing. Specifically, the POI recognition model can obtain POI information from the map application and provide the corresponding recognition results back to the map application, enabling the map application to provide navigation results based on the recognition results given by the POI recognition model.

[0120] The first terminal 410 and the second terminal 510 in the above embodiments may include smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, and in-vehicle terminals, but are not limited to these.

[0121] Server 520 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0122] In addition, server 520 can also be a node server in a blockchain network.

[0123] The second terminal 510 and the server 520 can establish a communication connection via a wireless network or a wired network. This wireless or wired network uses standard communication technologies and / or protocols. The network can be the Internet or any other network, including but not limited to any combination of Local Area Network (LAN), Metropolitan Area Network (MAN), Wide Area Network (WAN), mobile, wired or wireless networks, private networks, or virtual private networks.

[0124] Of course, this is understandable. Figure 4 and Figure 5 The implementation environment described in this application is merely one of the optional application scenarios for the training method of the point of interest information recognition model provided in this embodiment. Actual applications are not fixed. Figure 4 and Figure 5 The software and hardware environment is shown. The method provided in this application embodiment can be applied to various technical fields, including but not limited to big data, intelligent transportation systems, and intelligent vehicle-road cooperative systems.

[0125] Reference Figure 6 , Figure 6 This is a flowchart illustrating a training method for a point of interest information recognition model provided in an embodiment of this application. The training method can be executed by the terminal alone or by the terminal in cooperation with the server. The training method includes, but is not limited to, the following steps 610 to 660.

[0126] Step 610: Obtain the first point of interest information, the semantic component labels corresponding to the words in the first point of interest information, and the associated task labels of the first point of interest information; wherein, the semantic component labels are used to represent the membership relationship of entities in the first point of interest information, and the associated task labels are used to represent the task results of the associated tasks of the first point of interest information.

[0127] In this step, training the point of interest (POI) information recognition model requires acquiring corresponding training data. This training data mainly consists of two parts: the first part is the data input into the model for processing and obtaining the corresponding prediction results; the second part is the label data corresponding to the first part. The label data is used to identify the true results of the prediction task. By comparing the prediction results and the label data, the accuracy of the model's predictions can be reflected, and this can be further used to update the model's parameters and determine the progress of the model training.

[0128] Specifically, in this embodiment, when training the interest point information recognition model, a multi-task learning approach is adopted, combining it with an association task model to jointly participate in the prediction task, thereby updating the relevant parameters of the interest point information recognition model. Here, in multi-task learning, one part of the learning task is the interest point information recognition task, that is, the task that the trained interest point information recognition model needs to perform, and the other part of the learning task is the task that the association task model is responsible for performing. Therefore, correspondingly, when obtaining training data in step 610, the interest point information used for training can first be obtained, denoted as the first interest point information; for the task performed by the interest point information recognition model, the semantic component labels corresponding to the words in the first interest point information can be obtained, and these semantic component labels are used to characterize the membership relationship of entities in the first interest point information. Here, the data format of semantic component tags and the membership relationships they represent can be flexibly set as needed, and this application does not impose any restrictions on this. For example, exemplaryly, in some embodiments, vectors can be used as the data format of semantic component tags. For instance, when the semantic component tag is (1, 0, 0), it means that the word corresponding to the semantic component tag belongs to the entity (or part of the entity) of the main point in the first point of interest information; when the semantic component tag is (0, 1, 0), it means that the word corresponding to the semantic component tag belongs to the entity (or part of the entity) of the sub-point in the first point of interest information; when the semantic component tag is (0, 0, 1), it means that the word corresponding to the semantic component tag belongs to the entity (or part of the entity) of the branch in the first point of interest information.

[0129] It should be noted that, due to the potential differences in the granularity of word segmentation, in this embodiment, each word in the first point of interest information may represent an individual entity or be part of an entity. In actual processing, multiple consecutive words with the same semantic component label can be identified as a single entity. In other words, when labeling the semantic component of words, the semantic component labels of each entity in the first point of interest information can be labeled first. Then, for the same entity that has been split into multiple words, the semantic component labels of the resulting words are all based on the original entity's semantic component labels. For example, if "Tsinghua University" is the main point in a certain point of interest information, and word segmentation of "Tsinghua University" yields the words "Tsinghua" and "University," then the semantic component labels corresponding to these two words are both labeled as main point type labels.

[0130] In this step, in addition to obtaining the semantic component tags mentioned above, the associated task tags of the first point of interest information are also obtained. Here, the associated task tags are used to characterize the task results of the associated tasks of the first point of interest information, and can be used to measure the accuracy of the task performed by the associated task model. This application does not limit the specific data format of the tags; for example, they can be any one or more of numerical values, vectors, matrices, or tensors. The specific meaning represented by the associated task tags needs to be determined according to the specific associated task type. This part will be introduced and explained in subsequent embodiments and will not be elaborated here.

[0131] It should be noted that in this step, the relevant information and data can be obtained locally or from the cloud, and this application embodiment does not limit this.

[0132] Step 620: Input the first point of interest information into the first sub-model for natural language processing to obtain the first feature data corresponding to the first point of interest information.

[0133] In this step, the acquired first point of interest information can be input into the first sub-model of the point of interest information recognition model. The first sub-model performs natural language processing on the first point of interest information to extract its corresponding feature data, which is denoted as the first feature data. In this embodiment, the data format of the feature data is not limited; for example, the feature data may include one or more combinations of numerical values, vectors, matrices, or tensors.

[0134] Specifically, in some embodiments, when the first sub-model performs natural language processing on the first point of interest information, it can map the characters or words in the first point of interest information to a vector space with a unified lower dimension, thereby obtaining the corresponding first feature data. Strategies for generating such mappings include neural networks, dimensionality reduction of word co-occurrence matrices, probabilistic models, and interpretable knowledge base methods.

[0135] It should be noted that, in the embodiments of this application, the first feature data of the first point of interest information can be feature data extracted at the character level or feature data extracted at the word level. For example, taking the feature data extracted at the character level as an example, in subsequent processing, since it is necessary to identify its semantic components at the word level through the second sub-model, an intermediate processing step can be added to convert the character-level first feature data into word-level first feature data. Specifically, the first point of interest information can be segmented into word groups, which include multiple words. For each word, the first feature data corresponding to the characters contained within it can be fused to obtain the first feature data corresponding to the word. Here, when performing word segmentation, in some embodiments, a dictionary-based word segmentation algorithm can be used, first dividing the sentence into words according to the dictionary, and then finding the best combination of words; in some embodiments, a character-based word segmentation algorithm can also be used, first dividing the sentence into individual characters, and then combining the characters into words to find the optimal combination. This application does not limit this approach. Furthermore, when fusing feature data, any processing method can be used, such as splicing, superimposing, and then normalizing.

[0136] Step 630: Input the first feature data into the second sub-model, perform semantic component analysis on the first interest point information, and obtain the semantic component analysis results corresponding to the words in the first interest point information.

[0137] In this step, after obtaining the first feature data, it can be input into the second sub-model to perform semantic component analysis on the first point of interest information, thereby obtaining the semantic component analysis results corresponding to each word. Here, the semantic component analysis results are used to characterize the membership relationships of entities in the first point of interest information predicted by the model. Its data format and corresponding meaning are the same as the aforementioned semantic component labels, and will not be repeated here.

[0138] Step 640: Input the first feature data into the association task model, perform association task processing on the first point of interest information, and obtain the association task processing result corresponding to the first point of interest information.

[0139] In this step, as before, this embodiment of the application uses a multi-task learning approach, combining the association task model with the prediction task. Therefore, here, the first feature data is also input into the association task model, and the association task model performs association task processing based on the feature data extracted by the interest point information recognition model, thereby obtaining the association task processing result corresponding to the first interest point information. Similarly, here, the data format and corresponding meaning of the association task processing result are the same as the aforementioned association task label.

[0140] Specifically, in the embodiments of this application, the associated task model can be any learning model related to natural language processing tasks, and this application does not impose any restrictions on it.

[0141] It should be noted that, in this embodiment, the association task model itself may also include a feature extraction component. However, in application, the association task model does not need to extract feature data of the first point of interest information through its own feature extraction component, but instead uses the feature data extracted by the first sub-model of the point of interest information recognition model. In this way, the accuracy of the association task model in performing the association task can be used as a reference to measure the effectiveness of the first sub-model in extracting feature data, thereby providing feedback to the point of interest information recognition model to update its own parameters.

[0142] Step 650: Determine the training loss value based on the semantic component analysis results and semantic component labels, as well as the association task processing results and association task labels.

[0143] In this step, after obtaining the semantic component analysis results and the association task processing results, the training loss value can be determined based on the performance of the interest point information recognition model and the association task model. Specifically, for the multi-task learning in this embodiment, the training loss value also includes two parts. One part is the loss value of the interest point information recognition task, which can be determined through the semantic component analysis results and semantic component labels. The other part is the loss value of the task performed by the association task model, which can also be determined through the association task processing results and association task labels. Based on these two parts of the loss value, the overall training loss value can be determined.

[0144] Specifically, for various models in the field of artificial intelligence, their prediction accuracy can be measured by a loss function. The loss function is defined on a single training data point and measures the prediction error of that training data. Specifically, the loss value of that training data is determined by the label of that single training data point and the model's prediction result for that training data. However, in actual training, a training dataset contains many training data points. Therefore, a cost function is generally used to measure the overall error of the training dataset. The cost function is defined on the entire training dataset and is used to calculate the average prediction error of all training data, providing a better measure of the model's prediction performance. For general machine learning models, the aforementioned cost function, plus a regularization term to measure model complexity, can serve as the training objective function. Based on this objective function, the loss value of the entire training dataset can be calculated. Many types of loss functions are commonly used, such as 0-1 loss, squared loss, absolute loss, logarithmic loss, and cross-entropy loss, which can all be used as loss functions for machine learning models. These will not be elaborated upon here. In this embodiment, any one of these loss functions can be selected to determine the training loss value, that is, the loss value between the prediction result and the label. The prediction results here are either semantic component analysis results or association task processing results, and the labels are either semantic component labels or association task labels.

[0145] It is understood that in the embodiments of this application, when determining the loss value of the task of identifying point of interest information and the loss value of the task performed by the associated task model, the same loss function or different loss functions may be used, and this application does not impose any restrictions on this.

[0146] Step 660: Update the parameters of the first and second sub-models based on the loss values.

[0147] The loss value obtained in step 650 can be used to evaluate the accuracy of the model's predictions, enabling backpropagation training of the model and updating its internal parameters. Therefore, in this step, after obtaining the training loss value, the parameters of the first and second sub-models in the interest point information recognition model can be updated. Specifically, the parameters can be updated iteratively multiple times using the backpropagation algorithm. When the number of iterations reaches a preset number, the training is considered complete, and a trained interest point information recognition model is obtained. Alternatively, in some embodiments, a validation dataset can be pre-set. During training, the model parameters are iterated once or a predetermined number of times, and the current recognition accuracy of the model is verified using the validation dataset. When the recognition accuracy reaches a preset standard, the model training is considered complete. The method for collecting the validation dataset is the same as that for the training data, and will not be repeated here.

[0148] It is understood that in this embodiment, through multi-task learning, when training the interest point information recognition model, not only are the first interest point information and its corresponding semantic component labels obtained for training the model, but also the associated task labels of the first interest point information are acquired. Then, after feature extraction from the first interest point information, the resulting first feature data is input into a second sub-model for interest point information recognition, and also input into an associated task model for associated semantic task processing. Next, the loss values ​​obtained from both the interest point information recognition and the associated semantic task processing are combined to update the parameters of the interest point information recognition model. In this way, the semantic understanding and analysis capabilities of the interest point information recognition model can be improved through multi-task joint learning. This is beneficial for improving the accuracy of the recognition results when using the model for subsequent interest point information recognition.

[0149] It should be noted that the association task model used in the embodiments of this application can be either pre-trained or initialized and trained together with the interest point information recognition model in this application. In other words, when the association task model used in the training of the interest point information recognition model is a pre-trained model, it is not necessary to update or change the parameters of the association task model during the training of the interest point information recognition model; when the association task model used in the training of the interest point information recognition model is an untrained model, the interest point information recognition model and the association task model can be regarded as a whole model, and the parameters of the interest point information recognition model and the association task model can be updated by the obtained loss value.

[0150] Furthermore, in this embodiment, if a pre-trained association task model is selected, to improve the correlation between the association task model and the interest point information recognition model, the association task model can be trained using training data from the interest point information's related domain. For example, in some embodiments, the same batch of interest point information can be obtained, the association task model can be trained first to obtain a trained association task model, and then the interest point information and the trained association task model can be used to train the interest point information recognition model. This allows the association task model to better fit the application domain of the interest point information, thereby more effectively reflecting the interest point information recognition model's ability to extract feature data, which is beneficial for improving the prediction accuracy of the trained interest point information recognition model.

[0151] The implementation of the associated task model in this application embodiment will be explained and illustrated below with some specific application examples.

[0152] For example, in some embodiments, the association task model includes a name category prediction model; inputting first feature data into the association task model, performing association task processing on the first point of interest information, and obtaining the association task processing result corresponding to the first point of interest information includes:

[0153] The first feature data is input into the name category prediction model to predict the name category of the first point of interest information, and the name category prediction result corresponding to the first point of interest information is obtained.

[0154] In this embodiment, the association task model may include a name category prediction model. This model predicts the name category to which the first point of interest (POI) belongs based on its first feature data. Here, the name category characterizes the type of location the POI belongs to. For example, the name category may include classifications such as "school," "residential building," "shop," "restaurant," and "hospital." The specific name category can be flexibly adjusted as needed, and this application does not impose any restrictions on it.

[0155] The name category prediction model itself performs a classification task, that is, it classifies the information of the first point of interest into categories, thereby obtaining the corresponding name category prediction results. For example, the name category prediction model can predict whether "Experimental Primary School Staff Quarters" belongs to the name category of "school" or the name category of "residential building". Here, different name categories can be distinguished using any data format, and this application does not impose any restrictions on this.

[0156] It is understandable that in the field of machine learning, classification tasks, as one of the fundamental tasks, have a wide variety of models and algorithms that can be used. Therefore, in this embodiment, the classification model algorithm selected for the name category prediction model is not limited. Furthermore, since the name category prediction model in this embodiment is based on feature data extracted from the first sub-model for prediction, its prediction accuracy is related to the feature extraction capability of the first sub-model to a certain extent. Therefore, the accuracy of the name category prediction association task performed by the name category prediction model, i.e., the accuracy of the name category prediction result, can be used to update the parameters of the interest point information recognition model.

[0157] For example, in other embodiments, the association task model includes a name similarity discrimination model; inputting first feature data into the association task model, performing association task processing on the first point of interest information, and obtaining the association task processing result corresponding to the first point of interest information includes:

[0158] First target interest point information and second target interest point information are randomly selected from the first interest point information;

[0159] Determine first target data obtained by processing first target interest point information and second target data obtained by processing second target interest point information from first feature data;

[0160] The first target data and the second target data are input into the name similarity discrimination model to judge the name similarity between the first target interest point information and the second target interest point information, and the name similarity discrimination result of the first target interest point information and the second target interest point information is obtained.

[0161] In this embodiment, the association task model may further include a name similarity discrimination model, which can be used to determine the similarity between two or more points of interest (POIs). For ease of description, taking the discrimination of the similarity between two POIs as an example, firstly, two POIs can be randomly selected from the first POIs, denoted as the first target POI and the second target POI. For the first target POI, the corresponding first target data is determined from the first feature data; similarly, the second target data corresponding to the second target POI is also determined from the first feature data. Then, the name similarity discrimination model can determine the name similarity discrimination result between the first target POI and the second target POI based on the first target data and the second target data.

[0162] Specifically, in some embodiments, the similarity between point-of-interest (POI) information can be represented as a percentage. For example, when the name similarity between the first and second POI information is 90%, it indicates a high degree of similarity and a high likelihood of correlation. Conversely, when the name similarity between the first and second POI information is 20%, it indicates a low degree of similarity and a higher probability of no correlation. When determining the name similarity discrimination result based on the first and second target data, the name similarity discrimination model can directly calculate the difference between the first and second target data to obtain a difference value. This difference value can be, for example, the difference between numerical values, the Euclidean distance between vectors, or the norm difference between matrices. Then, based on the difference between the first and second target data, the name similarity discrimination result is determined through the difference value and a predetermined function. This function ensures a negative correlation between the name similarity discrimination result and the difference value. This application does not limit the specific functional relationship.

[0163] It is understood that, in the embodiments of this application, since the first target data and the second target data are extracted by the first sub-model, the accuracy of the name similarity discrimination model is related to the feature extraction capability of the first sub-model to a certain extent. Therefore, the accuracy of the name similarity discrimination model in performing the association task of name similarity discrimination, that is, the accuracy of the name similarity discrimination result, can be used to feed back and update the parameters of the interest point information recognition model.

[0164] It should be noted that the above examples of association task models are only used to illustrate the implementation of association task models in the embodiments of this application, and do not imply any limitation on their specific implementation. Furthermore, in practical applications, the number of association task models used to train the point of interest information recognition model can be one or more. When multiple association task models are used, each association task model corresponds to one association task.

[0165] In some embodiments, the training loss value is determined based on the semantic component analysis results and semantic component labels, as well as the association task processing results and association task labels, including:

[0166] Based on the semantic component analysis results and semantic component labels, determine the first sub-loss value for semantic component analysis of the first point of interest information;

[0167] Based on the results of the associated task processing and the associated task labels, determine the second sub-loss value for the associated task processing of the first point of interest information;

[0168] The first and second sub-loss values ​​are weighted and summed to obtain the training loss value.

[0169] In this embodiment, when determining the training loss value, the loss value for semantic component analysis of the first point of interest information can first be determined based on the semantic component analysis results and semantic component labels using the corresponding loss function, and is denoted as the first sub-loss value. Simultaneously, the loss value for association task processing of the first point of interest information can be determined based on the association task processing results and association task labels using the corresponding loss function, and is denoted as the second sub-loss value. Then, the first and second sub-loss values ​​can be weighted and summed to obtain the training loss value. Here, the weights corresponding to the first and second sub-loss values ​​can be flexibly set as needed, and this application does not impose any restrictions on this.

[0170] It should be noted that, as mentioned above, in the embodiments of this application, the number of associated task models used to train the interest point information recognition model may be multiple. When multiple associated task models are used, the sum or weighted sum of the loss values ​​corresponding to each associated task model can be used as the overall loss value of the associated task.

[0171] In some embodiments, the first sub-loss value and the second sub-loss value are weighted and summed to obtain the training loss value, including:

[0172] The first value is obtained by multiplying the first sub-loss value and the first weight.

[0173] The second value is obtained by multiplying the second sub-loss value and the second weight.

[0174] The training loss value is obtained by summing the first and second values.

[0175] The first weight is greater than the second weight.

[0176] In this embodiment, when weighting the first sub-loss value obtained from the task of identifying interest points and the second sub-loss value obtained from the task of association, the weights corresponding to the two can be set first: the weight corresponding to the first sub-loss value is denoted as the first weight, and the weight corresponding to the second sub-loss value is denoted as the second weight. Here, since the model to be trained in this embodiment is an interest point identification model, in order to reduce the situation where the influence of the association task is too large, resulting in low accuracy of the interest point identification model after training, the first weight can be set to be greater than the second weight. In this way, the training task can be more biased towards the influence of the interest point identification task, thereby maximizing the influence of the main task on the training during multi-task learning. After setting the two weights, the product of the first sub-loss value and the first weight can be determined to obtain the first value; and the product of the second sub-loss value and the second weight can be determined to obtain the second value. Thus, the training loss value can be obtained based on the sum of the first value and the second value.

[0177] In some embodiments, the method further includes:

[0178] Detect whether each primary point of interest is missing its corresponding semantic component label or associated task label;

[0179] If the semantic component label corresponding to the first point of interest information is missing, skip the step of inputting the first feature data into the second sub-model to perform semantic component analysis on the first point of interest information; or...

[0180] If the first point of interest information lacks a corresponding associated task label, skip the step of inputting the first feature data into the associated task model to process the first point of interest information for associated task processing.

[0181] In this embodiment, because a multi-task learning approach is used during training, some tasks may lack corresponding labels when acquiring training data. For example, the acquired first point of interest information includes "East Gate of Tsinghua University," "High School Affiliated to Tsinghua University," and "Third Canteen of Fudan University." The labels related to these three points of interest include the associated task labels for "East Gate of Tsinghua University" and "High School Affiliated to Tsinghua University" (e.g., labels representing the similarity between their names), as well as the semantic component labels for "High School Affiliated to Tsinghua University" and "Third Canteen of Fudan University." Therefore, during training, to improve the utilization rate of labels in the acquired training data and considering the potential difficulty in performing tasks due to missing labels, a task decoupling strategy can be adopted for training. Specifically, it can detect whether each first point of interest is missing a corresponding semantic component label or associated task label. If the corresponding semantic component label is missing, the loss value obtained from semantic component analysis cannot be determined. Therefore, step 630 can be skipped for this training data. Similarly, if the corresponding associated task label is missing, the loss value obtained from performing associated tasks cannot be determined. Therefore, step 640 can be skipped for this training data.

[0182] It should be further explained that since the number of association task models used to train the interest point information recognition model may be multiple, if the corresponding association task label is missing, it is only necessary to skip the step of inputting the first feature data into the association task model that is missing the association task label. For association tasks that do not have missing labels, they can be executed as usual. Furthermore, if a first interest point information is missing its corresponding semantic component label and all association task labels, that first interest point information can be removed from the training data.

[0183] The training method of the point of interest information recognition model provided in this application will be described in detail below with reference to some more specific embodiments.

[0184] Reference Figure 7This application provides an interest point information recognition model, wherein the first sub-model of the model adopts the ALBERT model, and the second sub-model adopts the Conditional Random Field. Interest point information is input into the interest point information recognition model, and the ALBERT model processes it to obtain context-dependent dynamic word vectors, which are the first feature data in this application embodiment. As mentioned above, in this application embodiment, the word vectors can be converted into word vectors by concatenation or weighting based on the word segmentation results obtained from segmenting the interest point information. In the interest point information recognition model of this application embodiment, in addition to applying the word vectors obtained above, a dictionary is used to search and match the interest point information to determine the domain knowledge features corresponding to the interest point information. Here, the dictionary is used to store at least one word file, and any word file is used to store multiple words under the corresponding semantic category. For example, the interest point dictionary can store main point files, sub-point files, branch files, business district files, brand files, role files, etc., and the domain knowledge features can be represented by one or more vectors. For example, when using a vector to represent the domain knowledge features corresponding to point-of-interest (POI) information, each element in the vector can correspond to a word file. When a search and matching operation is performed on the POI information, and it is found that the POI information matches a certain word file, the value of the element position corresponding to that word file can be set to 1. Conversely, if a search and matching operation is performed on the POI information, and it is found that the POI information does not match a certain word file, the value of the element position corresponding to that word file can be set to 0. In this way, the domain knowledge features corresponding to the POI information can be determined. Of course, in some embodiments, the value of the element position corresponding to the word file can also be determined based on the specific words matched within the word file; this application does not impose any restrictions on this. Furthermore, since some POI information contains parentheses, which generally indicate sub-points or branch names, whether the words in the POI information are within parentheses can also be used as a dimension of the domain knowledge features.

[0185] After obtaining the domain knowledge features, the first feature data and domain knowledge features corresponding to each word can be fused according to the word segmentation results. The specific fusion method is similar to the aforementioned embodiments and will not be repeated here. After fusion, the obtained fused feature data can be input into a conditional random field (CRF), which can label the fused feature data corresponding to each word. (Refer to...) Figure 7Specifically, in this embodiment, the BMES annotation method can be used. In this method, the label at the beginning of each entity name in the point of interest information is represented by B, the label in the middle of each entity name is represented by M, the label at the end of each entity name is represented by E, the label for independent entity names is represented by S, and the label for pre-defined non-entity names is represented by O. Thus, the information between each adjacent label "B" and label "E" represents a single entity. In this embodiment, when identifying the semantic components of point of interest information using this model, in some embodiments, the model can be set to only annotate entities with specified semantic components, or the labels can be further subdivided. For example, the labels corresponding to words with semantic components belonging to the main point can be marked with the main point symbol, and the labels corresponding to words with semantic components belonging to the sub-points can be marked with the sub-point symbol, thereby achieving the identification of the semantic components of words.

[0186] In this embodiment of the application, when... Figure 7 When training the model shown, in some embodiments, the ALBERT model can be pre-trained using a corpus related to point-of-interest (POI) information to obtain an ALBERT model suitable for handling POI information-related tasks. For details, please refer to... Figure 8 , Figure 8 This document illustrates a flowchart of the pre-training process for an ALBERT model. Pre-training the ALBERT model first requires a large amount of corpus related to interest point information. In this embodiment, this corpus can be obtained in two ways: firstly, it can be collected from relevant interest point information databases; secondly, it can be obtained based on the aforementioned dictionary. Here, the obtained interest point information is denoted as second interest point information. Based on the second interest point information, the language model can be pre-trained. For example, using… Figure 8 Taking the ALBERT model as an example, firstly, the acquired second point of interest information can be subjected to multi-granularity masking. For instance, in some embodiments, the second point of interest information can be randomly masked; in other embodiments, the point of interest information can be segmented, with the segmentation granularity being either word-level or entity-level. Then, several words or entity information can be masked. Of course, in the embodiments of this application, multiple masking methods can also be combined, which can greatly improve the semantic modeling capability of the pre-trained ALBERT model and facilitate the extraction of more accurate feature data.

[0187] After pre-training the ALBERT model, the interest point information recognition model can be trained as a whole using the training method provided in this embodiment. Specifically, please refer to... Figure 9 In this embodiment of the application, when... Figure 7The model shown is trained by combining two related tasks: name category prediction and name similarity discrimination. During training, the loss values ​​for the conditional random field sequence labeling task (i.e., the interest point information recognition task), the name category prediction task, and the name similarity discrimination task are calculated separately. These loss values ​​are then weighted to obtain the final loss value. The specific loss function is shown below:

[0188] Loss=α*L ner +β*L class +γ*L sim

[0189] In the formula, Loss represents the overall loss value, α represents the weighted average of the loss values ​​corresponding to the interest point information recognition task, and L... ner L represents the loss value for the task of identifying points of interest; β represents the weighted weight of the loss value for the task of predicting name and category; L represents the weighted weight of the loss value for the task of predicting name and category. class L represents the loss value for the name category prediction task; γ represents the weighted weight of the loss value for the name similarity discrimination task. sim This represents the loss value corresponding to the name similarity discrimination task.

[0190] Specifically, in this embodiment of the application, to address the difficulty in obtaining labeled corpora for each task during multi-task learning, a corpus decoupling strategy is proposed, which allows skipping certain tasks. In this case, when calculating the loss function, if a sample in the current task lacks a corresponding label, it needs to be removed from the loss function. Therefore, the final loss function is expressed as follows:

[0191]

[0192] In the formula, This represents the loss value corresponding to the task of identifying interest point information after decoupling from the corpus. This represents the loss value for the name category prediction task after corpus decoupling; This represents the loss value for the name similarity discrimination task after corpus decoupling.

[0193] To compare and verify the effectiveness of the training method for the interest point information recognition model provided in the embodiments of this application, the scheme of recognizing interest point information using the BiSLTM model was used as the benchmark scheme in the embodiments of this application. The performance of the models trained based on various model strategies was verified, and precision, recall, and F1 score were used as evaluation indicators. The specific experimental results are as follows:

[0194] Table 1

[0195]

[0196] Table 1 lists the following schemes from top to bottom: The scheme with the "ALBERT model" strategy replaces the BiSLTM model in the baseline scheme with the ALBERT model; the scheme with the "ALBERT model + word vectors" strategy adds the means of converting the ALBERT model output into word vectors; the scheme with the "ALBERT model + word vectors + CRF" strategy adds the means of using a conditional random field as a decoder to obtain sequence labeling data; the scheme with the "POI-ALBERT model + word vectors + CRF" strategy adds the means of pre-training the ALBERT model within the interest point information domain; and the scheme with the "POI-ALBERT model + word vectors + CRF + domain knowledge features" strategy adds the means of combining domain knowledge features for recognition. As can be seen from the data in Table 1, the scheme in this application uses the contextual features of interest point information extracted by the model and combines them with the domain knowledge features obtained by dictionary matching as the feature data used by the model for prediction, which can significantly improve the accuracy of recognition. Furthermore, by using the ALBERT model pre-trained in the interest point information domain and combining it with the conditional random field for recognition, a better recognition effect can be achieved compared to the BiSLTM model.

[0197] Table 2

[0198]

[0199] Table 3

[0200]

[0201] Furthermore, in this embodiment, the relevant indicator data of the model obtained after adding "multi-task learning", "LOSS weight tuning" and "corpus decoupling" are compared. Figure 10 The table shows a comparison of the F1 scores of models trained after various model strategies. Combined with Tables 2 and 3, it can be seen that the training method for the interest point information recognition model provided in this embodiment, by employing multi-task learning and corpus decoupling strategies, and combining the weighted loss values ​​of each task to train the interest point information recognition model, can significantly improve the model's recognition accuracy. Figure 10 As can be seen, the F1 score of the model obtained by adding "multi-task learning", "LOSS weight tuning" and "corpus decoupling" in this embodiment of the application is significantly improved, which is equivalent to the prediction scheme combining BiSLTM and CRF models.

[0202] Reference Figure 11 , Figure 11This is a flowchart illustrating a point-of-interest (POI) information recognition method provided in an embodiment of this application. Similar to the training method for the aforementioned POI information recognition model, this POI information recognition method can be executed by the terminal alone or by the terminal in cooperation with a server. The POI information recognition method includes, but is not limited to, the following steps 1110 to 1130.

[0203] Step 1110: Obtain information on the third point of interest;

[0204] Step 1120: Input the third point of interest information into the first sub-model for natural language processing to obtain the second feature data corresponding to the third point of interest information;

[0205] Step 1130: Input the second feature data into the second sub-model, perform semantic component analysis on the third point of interest information, and obtain the semantic component recognition result, wherein the semantic component recognition result is used to characterize the membership relationship of entities in the third point of interest information;

[0206] Among them, the point of interest information recognition model is composed of, for example Figure 6 The interest point information recognition model shown was trained using the training method described.

[0207] The foregoing embodiments described the training method for the point of interest (POI) information recognition model provided in this application. After the POI information recognition model is trained, it can be used in POI information recognition tasks to obtain the desired POI information recognition results.

[0208] Specifically, in this embodiment, the interest point information to be identified can be denoted as the third interest point information. After obtaining the third interest point information, it can be input into the first sub-model of the aforementioned trained interest point information recognition model for natural language processing to obtain the feature data corresponding to the third interest point information, denoted as the second feature data. Then, the second feature data is input into the second sub-model, and semantic component analysis is performed on the third interest point information through the second sub-model to obtain the semantic component recognition result. This semantic component recognition result can represent the membership relationship of entities in the third interest point information. Its specific data form and meaning can be referred to the aforementioned semantic component label, and will not be repeated here.

[0209] In some embodiments, semantic component analysis is performed on the third point of interest information based on the second feature data to obtain a semantic component recognition result, including:

[0210] The third point of interest information is searched and matched in the preset dictionary to obtain the third feature data corresponding to the third point of interest information;

[0211] The second and third feature data are fused to obtain the fourth feature data.

[0212] The fourth feature data is input into the second sub-model, and semantic component analysis is performed on the third interest point information to obtain the semantic component recognition result.

[0213] In the embodiments of this application, reference is made to Figure 7 The model structure shown, when identifying the third point of interest information, can not only use the second feature data extracted by the first sub-model, but also combine it with a dictionary to determine the domain knowledge features corresponding to the third point of interest information, thus obtaining the third feature data. Then, feature fusion processing can be performed on the second and third feature data to obtain the fourth feature data. Based on the fourth feature data, semantic component analysis is performed on the third point of interest information through the second sub-model to obtain the semantic component recognition result. This effectively improves the accuracy of the obtained recognition result. Specifically, the fusion processing here can adopt any processing form such as weighted averaging or concatenation; this application does not limit this.

[0214] Taking vector-based feature data as an example, assuming that the language model used in this embodiment is in the form of output character feature vectors, that is, the second feature data includes character feature vectors, then when fusing the second and third feature data, firstly, for the third feature data, it is obtained through dictionary search matching, generally in the form of word feature vectors, which is denoted as the first word feature vector in this embodiment. For the second feature data, the third interest point information can be segmented to obtain each word in the third interest point information. Then, based on the segmentation results, the character feature vectors can be fused to obtain the word feature vectors corresponding to the words in the third interest point information, which are denoted as the second word feature vectors. Next, the first word feature vector and the second word feature vector can be fused to obtain the fourth feature data.

[0215] It is understood that the point of interest information identification method provided in this application embodiment can be applied to application scenarios such as big data, intelligent transportation systems, and intelligent vehicle-road cooperative systems.

[0216] For example, refer to Figure 12 , Figure 12 The image shows a schematic diagram of the interface of a terminal device that displays point-of-interest information. Figure 12 In the terminal page shown, when searching for information about a point of interest (POI), the terminal device can receive the input query information. For example, when the input information is a POI belonging to a main point stored in a geographic information system, the terminal device can display POI information including information on each branch under that main point. For example... Figure 12If the input "KFC" is the main point among multiple points of interest, the terminal device can obtain the point of interest information corresponding to multiple branches that have a physical affiliation with the main point, display them for selection, and thus facilitate guidance and navigation.

[0217] For example, refer to Figure 13 and Figure 14 , Figure 13 and Figure 14 The image shows a schematic diagram of the interface of a map navigation software. Figure 13 In China, existing map navigation software typically displays all of a point of interest (POI) when showing its information. For example, for... Figure 13 A Universal Resort in the image contains multiple Points of Interest (POIs). When displaying these POIs, each is prefixed with "Universal Resort," such as "Universal Resort Tour Bus Parking Area." This is inconvenient to use and prone to incomplete display. However, the POI information identification method provided in this embodiment can identify the entity hierarchical relationships between POIs. Therefore, when searching for POI information for the main POI "Universal Resort," only the names of its sub-POIs can be displayed. (Refer to...) Figure 14 In map navigation software, the main point information in the labels of other points of interest belonging to the main point "Universal Resort" can be omitted, making the interface display simpler, more readable, and more convenient in providing geographical information guidance.

[0218] Reference Figure 15 , Figure 15 This is a schematic diagram of the structure of a training device for an interest point information recognition model provided in an embodiment of this application. The interest point information recognition model includes a first sub-model and a second sub-model; the training device for the interest point information recognition model includes:

[0219] The acquisition module 1510 is used to acquire first point of interest information, semantic component labels corresponding to words in the first point of interest information, and associated task labels of the first point of interest information; wherein, the semantic component labels are used to represent the membership relationship of entities in the first point of interest information, and the associated task labels are used to represent the task results of the associated tasks of the first point of interest information.

[0220] The first processing module 1520 is used to input the first point of interest information into the first sub-model for natural language processing to obtain the first feature data corresponding to the first point of interest information.

[0221] The analysis module 1530 is used to input the first feature data into the second sub-model, perform semantic component analysis on the first interest point information, and obtain the semantic component analysis results corresponding to the words in the first interest point information;

[0222] The second processing module 1540 is used to input the first feature data into the association task model, perform association task processing on the first interest point information, and obtain the association task processing result corresponding to the first interest point information.

[0223] The calculation module 1550 is used to determine the training loss value based on the semantic component analysis results and semantic component labels, as well as the association task processing results and association task labels.

[0224] The update module 1560 is used to update the parameters of the first and second sub-models based on the loss value.

[0225] Furthermore, the association task model includes a name category prediction model; the aforementioned second processing module is specifically used for:

[0226] The first feature data is input into the name category prediction model to predict the name category of the first point of interest information, and the name category prediction result corresponding to the first point of interest information is obtained.

[0227] Furthermore, the association task model includes a name similarity discrimination model; the aforementioned second processing module is specifically used for:

[0228] First target interest point information and second target interest point information are randomly selected from the first interest point information;

[0229] Determine first target data obtained by processing first target interest point information and second target data obtained by processing second target interest point information from first feature data;

[0230] The first target data and the second target data are input into the name similarity discrimination model to judge the name similarity between the first target interest point information and the second target interest point information, and the name similarity discrimination result of the first target interest point information and the second target interest point information is obtained.

[0231] Furthermore, the aforementioned calculation module is specifically used for:

[0232] Based on the semantic component analysis results and semantic component labels, determine the first sub-loss value for semantic component analysis of the first point of interest information;

[0233] Based on the results of the associated task processing and the associated task labels, determine the second sub-loss value for the associated task processing of the first point of interest information;

[0234] The first and second sub-loss values ​​are weighted and summed to obtain the training loss value.

[0235] Furthermore, the aforementioned calculation module is specifically used for:

[0236] The first value is obtained by multiplying the first sub-loss value and the first weight.

[0237] The second value is obtained by multiplying the second sub-loss value and the second weight.

[0238] The training loss value is obtained by summing the first and second values.

[0239] The first weight is greater than the second weight.

[0240] Furthermore, the training device for the aforementioned point of interest information recognition model also includes:

[0241] The detection module is used to detect whether each primary point of interest information is missing corresponding semantic component labels or associated task labels;

[0242] The jump module is used to skip the step of inputting the first feature data into the second sub-model to perform semantic component analysis on the first point of interest information when the corresponding semantic component label is missing; or...

[0243] This is used to skip the step of inputting the first feature data into the association task model and performing association task processing on the first point of interest information when the corresponding association task label is missing from the first point of interest information.

[0244] Furthermore, the training device for the aforementioned point of interest information recognition model also includes a pre-training module, which is used for:

[0245] Acquire batches of second points of interest information;

[0246] The information of each second point of interest is masked to obtain the training dataset of point of interest information.

[0247] The first sub-model is pre-trained based on the training dataset and the second interest point information.

[0248] Furthermore, the aforementioned pre-trained module is used for at least one of the following:

[0249] Apply random masking to the second point of interest information; or,

[0250] Masking is applied to several words in the second point of interest information; or,

[0251] Masking is applied to several entity information in the second point of interest information.

[0252] It is understandable that, such as Figure 6 The content of the training method embodiment of the interest point information recognition model shown is applicable to the training device embodiment of this interest point information recognition model. The specific functions implemented by the training device embodiment of this interest point information recognition model are the same as those shown below. Figure 6 The training implementation of the interest point information recognition model shown is the same, and the beneficial effects achieved are the same as those shown. Figure 6 The training implementation of the point of interest information recognition model shown also achieves the same beneficial effects.

[0253] On the other hand, embodiments of this application also provide an interest point information identification device, used to identify interest point information through an interest point information identification model, the interest point information identification model including a first sub-model and a second sub-model; the device includes:

[0254] The second acquisition module is used to acquire information about the third point of interest.

[0255] The third processing module is used to input the third point of interest information into the first sub-model for natural language processing to obtain the second feature data corresponding to the third point of interest information.

[0256] The second analysis module is used to input the second feature data into the second sub-model, perform semantic component analysis on the third point of interest information, and obtain semantic component recognition results. The semantic component recognition results are used to characterize the membership relationship of entities in the third point of interest information.

[0257] The interest point information recognition model is trained using the training method described above.

[0258] Furthermore, the aforementioned second analysis module is specifically used for:

[0259] The third point of interest information is searched and matched in the preset dictionary to obtain the third feature data corresponding to the third point of interest information;

[0260] The second and third feature data are fused to obtain the fourth feature data.

[0261] The fourth feature data is input into the second sub-model, and semantic component analysis is performed on the third interest point information to obtain the semantic component recognition result.

[0262] Furthermore, the aforementioned second analysis module is specifically used for:

[0263] The information on the third point of interest is segmented into words;

[0264] Based on the results of word segmentation, feature fusion processing is performed on the character feature vectors to obtain the second word feature vectors corresponding to the words in the third interest point information.

[0265] The first word feature vector and the second word feature vector are fused to obtain the fourth feature data.

[0266] It is understandable that, such as Figure 11The content of the example of the point of interest information recognition method shown is applicable to the example of this point of interest information recognition device. The specific functions implemented by this point of interest information recognition device are the same as those shown in the example. Figure 11 The method for identifying point of interest information shown is the same as the embodiment described above, and achieves the same beneficial effects. Figure 11 The beneficial effects achieved by the example of the point of interest information identification method shown are also the same.

[0267] Reference Figure 16 This application also discloses an electronic device, including:

[0268] At least one processor 1610;

[0269] At least one memory 1620 is used to store at least one program;

[0270] When at least one program is executed by at least one processor 1610, such that at least one processor 1610 performs as follows: Figure 6 The training method embodiment of the interest point information recognition model shown is or Figure 11 The example shown is a method for identifying points of interest.

[0271] It is understandable that, such as Figure 6 The training method embodiment of the interest point information recognition model shown. Figure 11 The content of the point of interest information recognition method embodiments shown are all applicable to the embodiments of this electronic device. The specific functions implemented by the embodiments of this electronic device are the same as those shown. Figure 6 The training method embodiment of the interest point information recognition model shown. Figure 11 The method for identifying point of interest information shown is the same as the embodiment described above, and achieves the same beneficial effects. Figure 6 The training method embodiment of the interest point information recognition model shown. Figure 11 The beneficial effects achieved by the example of the point of interest information identification method shown are also the same.

[0272] This application also discloses a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to implement, for example... Figure 6 The training method embodiment of the interest point information recognition model shown is or Figure 11 The example shown is a method for identifying points of interest.

[0273] Understandable, Figure 6 The training method embodiment of the interest point information recognition model shown. Figure 11 The content of the point of interest information recognition method embodiments shown is applicable to the embodiments of this computer-readable storage medium. The specific functions implemented by the embodiments of this computer-readable storage medium are the same as those in the embodiments of this computer-readable storage medium. Figure 6 The training method embodiment of the interest point information recognition model shown. Figure 11 The example of the point of interest information identification method shown is the same, and the beneficial effects achieved are the same. Figure 6 The training method embodiment of the interest point information recognition model shown. Figure 11 The beneficial effects achieved by the example of the point of interest information identification method shown are also the same.

[0274] This application also discloses a computer program product or computer program, which includes computer instructions stored in the aforementioned computer-readable storage medium. Figure 16 The processor of the illustrated electronic device can read the computer instructions from the aforementioned computer-readable storage medium, and the processor executes the computer instructions, causing the electronic device to perform... Figure 6 The training method embodiment of the interest point information recognition model shown. Figure 11 The example shown is a method for identifying points of interest.

[0275] Understandable, Figure 6 The training method embodiment of the interest point information recognition model shown. Figure 11 The content of the point of interest information recognition method embodiments shown is applicable to this computer program product or computer program embodiment, and the specific functions implemented by this computer program product or computer program embodiment are the same as those shown. Figure 6 The training method embodiment of the interest point information recognition model shown. Figure 11 The example of the point of interest information identification method shown is the same, and the beneficial effects achieved are the same. Figure 6 The training method embodiment of the interest point information recognition model shown. Figure 11 The beneficial effects achieved by the example of the point of interest information identification method shown are also the same.

[0276] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0277] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional technology for an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0278] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0279] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable storage medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0280] More specific examples (a non-exhaustive list) of computer-readable storage media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable storage media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0281] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0282] In the foregoing description of this specification, the references to terms such as "one embodiment," "another embodiment," or "some embodiments," etc., indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0283] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0284] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for identifying point-of-interest (POI) information, characterized in that, This method is used to identify point-of-interest (POI) information using a POI information identification model, wherein the POI information identification model includes a first sub-model and a second sub-model; the identification method includes: Obtain information on third points of interest; The third point of interest information is input into the first sub-model for natural language processing to obtain the second feature data corresponding to the third point of interest information. The third point of interest information is searched and matched in a preset dictionary to obtain the third feature data corresponding to the third point of interest information; The second feature data and the third feature data are subjected to feature fusion processing to obtain the fourth feature data; The fourth feature data is input into the second sub-model to perform semantic component analysis on the third point of interest information, thereby obtaining the semantic component recognition result; wherein, the semantic component recognition result is used to characterize the membership relationship of entities in the third point of interest information; The interest point information recognition model is trained through the following steps: Obtain first point of interest information, semantic component tags corresponding to words in the first point of interest information, and associated task tags of the first point of interest information; wherein, the semantic component tags are used to characterize the membership relationship of entities in the first point of interest information, and the associated task tags are used to characterize the task results of associated tasks of the first point of interest information. The first point of interest information is input into the first sub-model for natural language processing to obtain the first feature data corresponding to the first point of interest information. The first feature data is input into the second sub-model, and semantic component analysis is performed on the first interest point information to obtain the semantic component analysis results corresponding to the words in the first interest point information. The first feature data is input into the association task model, and the first point of interest information is processed by the association task to obtain the association task processing result corresponding to the first point of interest information. Based on the semantic component analysis results and the semantic component labels, as well as the association task processing results and the association task labels, the training loss value is determined; Based on the loss value, the parameters of the first sub-model and the second sub-model are updated.

2. The method for identifying point-of-interest information according to claim 1, characterized in that, The association task model includes a name category prediction model; the step of inputting the first feature data into the association task model, performing association task processing on the first point of interest information, and obtaining the association task processing result corresponding to the first point of interest information includes: The first feature data is input into the name category prediction model to predict the name category of the first point of interest information, thereby obtaining the name category prediction result corresponding to the first point of interest information.

3. The method for identifying point-of-interest information according to claim 1, characterized in that, The association task model includes a name similarity discrimination model; the step of inputting the first feature data into the association task model, performing association task processing on the first point of interest information, and obtaining the association task processing result corresponding to the first point of interest information includes: First target interest point information and second target interest point information are randomly selected from the first interest point information; From the first feature data, determine the first target data obtained by processing the first target interest point information, and the second target data obtained by processing the second target interest point information; The first target data and the second target data are input into the name similarity discrimination model to judge the name similarity between the first target interest information and the second target interest information, and the name similarity discrimination result between the first target interest information and the second target interest information is obtained.

4. The method for identifying point-of-interest information according to any one of claims 1-3, characterized in that, The step of determining the training loss value based on the semantic component analysis results and the semantic component labels, as well as the association task processing results and the association task labels, includes: Based on the semantic component analysis results and the semantic component labels, a first sub-loss value is determined for semantic component analysis of the first point of interest information; Based on the associated task processing result and the associated task label, a second sub-loss value is determined for the associated task processing of the first point of interest information; The first sub-loss value and the second sub-loss value are weighted and summed to obtain the training loss value.

5. The method for identifying point-of-interest information according to claim 4, characterized in that, The weighted summation of the first sub-loss value and the second sub-loss value to obtain the training loss value includes: The first value is obtained by multiplying the first sub-loss value and the first weight. The second value is obtained by multiplying the second sub-loss value and the second weight. The training loss value is obtained by summing the first value and the second value. Wherein, the first weight is greater than the second weight.

6. The method for identifying point-of-interest information according to claim 1, characterized in that, The method further includes: Detect whether each of the first points of interest information is missing a corresponding semantic component label or associated task label; When the first point of interest information lacks a corresponding semantic component label, skip the step of inputting the first feature data into the second sub-model to perform semantic component analysis on the first point of interest information; or... If the first point of interest information lacks a corresponding associated task label, skip the step of inputting the first feature data into the associated task model and performing associated task processing on the first point of interest information.

7. The method for identifying point-of-interest information according to claim 1, characterized in that, The method further includes: Acquire batches of second points of interest information; The information of each second interest point is masked to obtain the training dataset of interest point information. The first sub-model is pre-trained based on the training dataset of the interest point information and the second interest point information.

8. The method for identifying point-of-interest information according to claim 7, characterized in that, The masking process for each of the second interest point information includes at least one of the following: The second point of interest information is subjected to random masking; or, Masking is applied to several words in the second point of interest information; or, Masking is applied to several entity information in the second point of interest information.

9. The method for identifying point-of-interest information according to claim 1, characterized in that, The second feature data includes character feature vectors, and the third feature data includes first word feature vectors; feature fusion processing is performed on the second feature data and the third feature data to obtain fourth feature data, including: The third point of interest information is segmented into words; Based on the results of the word segmentation process, feature fusion processing is performed on the character feature vectors to obtain the second word feature vectors corresponding to the words in the third interest point information. The first word feature vector and the second word feature vector are subjected to feature fusion processing to obtain the fourth feature data.

10. A point-of-interest (POI) information identification device, characterized in that, The device is used to identify point-of-interest (POI) information using a point-of-interest information identification model, wherein the POI information identification model includes a first sub-model and a second sub-model; the identification device includes: The second acquisition module is used to acquire information about the third point of interest. The third processing module is used to input the third point of interest information into the first sub-model for natural language processing to obtain the second feature data corresponding to the third point of interest information. The second analysis module is used to search and match the third point of interest information in a preset dictionary to obtain the third feature data corresponding to the third point of interest information; to perform feature fusion processing on the second feature data and the third feature data to obtain the fourth feature data; to input the fourth feature data into the second sub-model to perform semantic component analysis on the third point of interest information to obtain the semantic component recognition result; wherein, the semantic component recognition result is used to characterize the membership relationship of entities in the third point of interest information; The interest point information recognition model is trained through the following steps: Obtain first point of interest information, semantic component tags corresponding to words in the first point of interest information, and associated task tags of the first point of interest information; wherein, the semantic component tags are used to characterize the membership relationship of entities in the first point of interest information, and the associated task tags are used to characterize the task results of associated tasks of the first point of interest information. The first point of interest information is input into the first sub-model for natural language processing to obtain the first feature data corresponding to the first point of interest information. The first feature data is input into the second sub-model, and semantic component analysis is performed on the first interest point information to obtain the semantic component analysis results corresponding to the words in the first interest point information. The first feature data is input into the association task model, and the first point of interest information is processed by the association task to obtain the association task processing result corresponding to the first point of interest information. Based on the semantic component analysis results and the semantic component labels, as well as the association task processing results and the association task labels, the training loss value is determined; Based on the loss value, the parameters of the first sub-model and the second sub-model are updated.

11. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the point of interest information identification method according to any one of claims 1 to 9.

12. A computer-readable storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the point of interest information identification method according to any one of claims 1 to 9.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the point of interest information identification method according to any one of claims 1 to 9.

Images