Deep learning-based methods, systems, devices, and storage media for classifying interest points.

Abstract

This invention provides a method, system, device, and storage medium for interest point classification based on deep learning. The method includes: constructing a prefix tree based on entity names from a pre-defined interest point database, where each intermediate node contains only one character; placing interest point information corresponding to aliases in leaf nodes; obtaining at least two entity fields and corresponding candidate interest point numbers from a source file through entity recognition, and inputting the source file (identified as head and tail entities respectively) into a trained interest point neural network to detect whether a parent-child relationship threshold is met; matching the head and tail entities that meet the parent-child relationship threshold with the names of the interest points in the source file; and deduplicating the interest points based on their geographical information. This invention utilizes information from a tourism knowledge base and uses a pre-trained model for relationship classification to normalize interest points, significantly improving the deduplication rate.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, and more specifically, to a method, system, device, and storage medium for interest point classification based on deep learning. Background Technology

[0002] POI is an abbreviation for "Point of Interest." In Geographic Information Systems (GIS), a POI can be a building, a shop, a mailbox, a bus stop, etc. Traditional GIS data collection methods require cartographers to use sophisticated surveying instruments to obtain the latitude and longitude of a point of interest and then mark it. Because POI collection is a very time-consuming and labor-intensive task, the number of POIs in a GIS, to a certain extent, represents the value of the entire system. Each POI contains four aspects of information: name, category, coordinates, and classification. Comprehensive POI information is essential for enriching navigation maps. Timely POI information can remind users of road conditions, detailed information about surrounding buildings, and facilitate finding the places you need during navigation, allowing you to choose the most convenient and smooth route for path planning. Therefore, the number of POIs on a navigation map directly affects the usability of the navigation. POIs are classified into primary and secondary categories, each with corresponding industry codes and names. This facilitates the recording and differentiation of information collected.

[0003] The tourism industry knowledge base involves entities known as Points of Interest (POIs) in Geographic Information Systems (GIS). Common locations such as tourist attractions, hotels, shopping malls, and restaurants fall under the POI category. POIs play a crucial role in numerous internet-based electronic map services. The rise and vigorous development of e-commerce, local services, and the sharing economy are inseparable from the overall planning, management, and application of POI data.

[0004] Points of Interest (POIs) are not isolated nodes; different POIs are often intricately connected. For example, the POIs West Lake Scenic Area and Bai Causeway reflect an inclusion relationship, with West Lake Scenic Area as the parent node and Bai Causeway as the child node. Due to the complexity of POI data sources during the construction and iterative updates of the knowledge base—including publicly available official data, web crawler data, and numerous related platforms—the lack of identification of hierarchical relationships between POIs during batch import of multi-source data can lead to numerous duplicate POIs in the system, causing traffic splitting, or the presence of detached POIs, affecting client-side display and preventing users from fully understanding the tourist attractions, thus hindering traffic conversion.

[0005] Associating POI nodes that may have parent-child relationships helps improve the quality of knowledge bases in the tourism sector. Identifying hierarchical relationships typically relies on massive amounts of unstructured text, primarily consisting of POI introductions and descriptions, thus requiring natural language processing techniques. In recent years, deep learning and large-scale language models have been widely used in relation extraction tasks, significantly improving algorithm performance and becoming the mainstream choice for research and development.

[0006] Therefore, the present invention provides a method, system, device and storage medium for interest point classification based on deep learning. Summary of the Invention

[0007] To address the problems in existing technologies, the present invention aims to provide a method, system, device, and storage medium for classifying interest points based on deep learning. This overcomes the difficulties of existing technologies by utilizing information from a tourism knowledge base and using a pre-trained model for relationship classification, thereby achieving the normalization of interest points and significantly improving the deduplication and recognition rate of interest points.

[0008] Embodiments of the present invention provide a deep learning-based interest point classification method, comprising the following steps:

[0009] At least a prefix tree is constructed based on the names of entities in the preset interest point database, with each intermediate node containing only one character, and the interest point information corresponding to the alias is placed in the leaf node;

[0010] The source file is obtained by entity recognition. At least two entity fields and corresponding candidate interest point numbers are obtained from the source file. The source file, after being identified by marking any two entity fields as head entity and tail entity respectively, is input into a trained interest point neural network to detect whether the parent-child relationship threshold is met.

[0011] The head entity and tail entity that meet the parent-child relationship threshold are matched with the name of the point of interest to which the source file belongs;

[0012] If the geographic information of the points of interest corresponding to the head entity and the tail entity respectively meets the preset threshold, then duplicates are removed; otherwise, new intermediate nodes and leaf nodes are added based on the head entity and the tail entity.

[0013] Preferably, the step of constructing a prefix tree based at least on the names of entities in a preset interest point database, where each intermediate node contains only one character, and placing the interest point information corresponding to the alias into the leaf node, includes:

[0014] A prefix tree is constructed using the name and alias fields of all entities in the predefined point of interest database;

[0015] The alias strings are inserted one by one. The prefix tree contains no characters except for the root node, and the leaf nodes contain the terminator. Each intermediate node of the prefix tree contains only one character.

[0016] Place the point of interest information corresponding to the alias into the leaf node.

[0017] Preferably, the step of inserting alias strings one by one, wherein the prefix tree, except for the root node which contains no characters and the leaf nodes which contain a terminator, and each intermediate node of the prefix tree contains only one character, further includes:

[0018] Starting from the root node and proceeding to a certain node, the characters passed through are concatenated to represent the string corresponding to that node. From the root node to any leaf node, an entity alias is closed.

[0019] Preferably, the step of obtaining at least two entity fields and corresponding candidate interest point numbers from the source file through entity recognition, and inputting the source file after identifying any two entity fields as the head entity and tail entity respectively into a trained interest point class neural network to detect whether the parent-child relationship threshold is met includes:

[0020] Entity recognition is used to obtain at least two entity fields and the corresponding candidate interest point numbers from the source file.

[0021] Arrange and combine any two entity fields in the entity fields as the head entity and the tail entity respectively to obtain several element combinations as the interest points to be tested;

[0022] Based on each combination of elements, the corresponding entity fields in the source file are identified as the head entity and the tail entity, respectively, and the identified text is obtained accordingly;

[0023] The identified texts are input into a trained interest point neural network to detect whether each identified text satisfies the parent-child relationship threshold.

[0024] Preferably, the step of inputting the identified text into a trained interest point class neural network to detect whether each identified text satisfies a parent-child relationship threshold includes:

[0025] Assume the word sequence after word segmentation of the identified text and addition of positioning characters at the beginning is x. 1:T The latent vector sequence obtained after BERT feature extraction is as follows:

[0026] h 1:T =BERT(x 1:T );

[0027] Assuming the head and tail entities are located at positions i to j and k to m in the input sequence, respectively, their word-level vector representations are average-pooled and then fed into two fully connected layers. The output representations of the head and tail entities are as follows:

[0028]

[0029]

[0030] The latent vector corresponding to the position of the located character is obtained, and then processed through a fully connected layer to obtain:

[0031] h cls =FFN cls (h1), where CLS is the positioning character;

[0032] The latent vector of the located character is concatenated with the head and tail entity vectors, and then fed into a fully connected layer and an activation function layer to output a probability model in which the head entity is the parent interest point of the tail entity:

[0033]

[0034] The probabilistic model uses a binary cross-entropy function as the objective during the training phase. Assuming the true label of the sample is y, the loss is:

[0035]

[0036] During training, the AdamW algorithm is used to optimize the loss in batches until the function converges. An appropriate threshold is selected using development set data. During the inference phase, entity pairs whose output probability is greater than the threshold are predicted as parent-child relationships.

[0037] Preferably, the step of matching the head entity and tail entity that satisfy the parent-child relationship threshold with the name of the point of interest to which the source file belongs includes:

[0038] Determine whether the number of the point of interest to which the source text belongs matches the candidate number recalled by the head entity and / or tail entity. If yes, the head entity or tail entity is successfully matched. If not, remove the prefixes such as province, city, and administrative region, as well as the preset suffix fields from the names of the head entity and / or tail entity and the point of interest to which the source text belongs, and then perform matching again.

[0039] Preferably, when the geographic information of the points of interest corresponding to the head entity and the tail entity respectively meets a preset threshold, if so, deduplication is performed; otherwise, new intermediate nodes and leaf nodes are added based on the head entity and the tail entity, including:

[0040] Obtain the first geographic information of a successfully matched head or tail entity;

[0041] Obtain second geographic information of at least one candidate point of interest recalled from the head or tail entities that did not match successfully.

[0042] Determine whether the distance between the first geographic information and a second geographic information meets a preset threshold. If yes, then deduplicate the points of interest to be tested. If no, then add new intermediate nodes and leaf nodes based on the head and tail entities of the points of interest to be tested.

[0043] Embodiments of the present invention also provide a deep learning-based interest point classification system for implementing the above-described deep learning-based interest point classification method, wherein the deep learning-based interest point classification system includes:

[0044] The prefix tree generation module constructs a prefix tree based on the names of entities in the preset interest point database. Each intermediate node contains only one character, and the interest point information corresponding to the alias is placed in the leaf node.

[0045] The relationship detection module obtains at least two entity fields and corresponding candidate interest point numbers from the source file through entity recognition. The source file, after being identified by marking any two entity fields as the head entity and the tail entity respectively, is input into a trained interest point neural network to detect whether the parent-child relationship threshold is met.

[0046] The name matching module matches the head entity and tail entity that meet the parent-child relationship threshold with the name of the point of interest to which the source file belongs;

[0047] The point of interest deduplication module removes duplicates when the geographic information of the points of interest corresponding to the head entity and the tail entity meets a preset threshold; otherwise, it adds new intermediate nodes and leaf nodes based on the head entity and the tail entity.

[0048] Embodiments of the present invention also provide an interest point classification device based on deep learning, comprising:

[0049] processor;

[0050] A memory in which executable instructions of the processor are stored;

[0051] The processor is configured to perform the steps of the deep learning-based interest point classification method by executing the executable instructions.

[0052] Embodiments of the present invention also provide a computer-readable storage medium for storing a program that, when executed, implements the steps of the deep learning-based interest point classification method described above.

[0053] The purpose of this invention is to provide a method, system, device, and storage medium for classifying interest points based on deep learning. By utilizing information from a tourism knowledge base and using a pre-trained model for relationship classification, the invention achieves the normalization of interest points and greatly improves the deduplication rate of interest points. Attached Figure Description

[0054] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0055] Figure 1 This is a flowchart of the interest point classification method based on deep learning of the present invention.

[0056] Figure 2 A schematic diagram of the prefix tree in the implementation process of the deep learning-based interest point classification method of the present invention.

[0057] Figure 3 A schematic diagram of the BERT model during the implementation of the deep learning-based interest point classification method of this invention.

[0058] Figure 4 This is a schematic diagram of the module of the interest point classification system based on deep learning of the present invention.

[0059] Figure 5 This is a schematic diagram of the structure of the interest point classification device based on deep learning of the present invention.

[0060] Figure 6 This is a schematic diagram of the structure of a computer-readable storage medium according to an embodiment of the present invention. Detailed Implementation

[0061] The following specific examples illustrate the implementation methods of this application. Those skilled in the art can easily understand the other advantages and effects of this application from the content disclosed herein. This application can also be implemented or applied through other different specific embodiments, and various details in this application can be modified or changed according to different viewpoints and application systems without departing from the spirit of this application. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0062] The embodiments of this application will now be described in detail with reference to the accompanying drawings, so that those skilled in the art can easily implement the application. This application may be embodied in many different forms and is not limited to the embodiments described herein.

[0063] In this application, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics represented in connection with that embodiment or example, which are included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics represented may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate different embodiments or examples represented in this application, as well as features of different embodiments or examples.

[0064] Furthermore, the terms "first" and "second" are used for illustrative purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the representation of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0065] For the purpose of clearly describing this application, devices that are not relevant to the description are omitted, and the same or similar components throughout the specification are given the same reference numerals.

[0066] Throughout this specification, when it is said that a device is "connected" to another device, this includes not only "direct connection" but also "indirect connection" by placing other components in between. Furthermore, when it is said that a device "comprises" a certain constituent element, unless otherwise stated otherwise, this does not exclude other constituent elements, but rather implies that other constituent elements may be included.

[0067] When we say that a device is "above" another device, this can mean that it is directly above the other device, or it can mean that other devices are present in between. Conversely, when we say that a device is "directly" "above" another device, there are no other devices present in between.

[0068] Although the terms first, second, etc., are used in some instances herein to refer to various elements, these elements should not be limited by these terms. These terms are used only to distinguish one element from another. For example, first interface and second interface, etc., are used. Furthermore, as used herein, the singular forms “a,” “an,” and “the” are intended to also include the plural forms unless the context indicates otherwise. It should be further understood that the terms “comprising,” “including,” indicate the presence of features, steps, operations, elements, components, items, kinds, and / or groups, but do not exclude the presence, occurrence, or addition of one or more other features, steps, operations, elements, components, items, kinds, and / or groups. The terms “or” and “and / or” as used herein are interpreted as inclusive, or mean any one or any combination thereof. Thus, “A, B, or C” or “A, B, and / or C” means “any one of: A; B; C; A and B; A and C; B and C; A, B, and C.” Exceptions to this definition will only occur if the combination of elements, functions, steps, or operations is inherently mutually exclusive in some way.

[0069] The technical terms used herein are for reference only to specific embodiments and are not intended to limit the scope of this application. The singular form used herein includes the plural form unless the statement explicitly indicates otherwise. The word "comprising" as used in the specification means to specify a particular characteristic, region, integer, step, operation, element, and / or component, and does not exclude the presence or addition of other characteristics, regions, integers, steps, operations, elements, and / or components.

[0070] Although not explicitly defined, all terms, including technical and scientific terms used herein, shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. Terms defined in commonly used dictionaries shall be further interpreted as having a meaning consistent with the relevant technical literature and the content of this present application, and shall not be over-interpreted as having an ideal or overly formulaic meaning unless otherwise defined.

[0071] Figure 1 This is a flowchart of the deep learning-based interest point classification method of the present invention. Figure 1 As shown, an embodiment of the present invention provides an interest point classification method based on deep learning, including the following steps:

[0072] S110. Construct a prefix tree based on the names of entities in the preset interest point database, where each intermediate node contains only one character, and place the interest point information corresponding to the alias into the leaf node.

[0073] S120. Obtain at least two entity fields and corresponding candidate interest point numbers from the source file through entity recognition. Input the source file after identifying any two entity fields as head entity and tail entity respectively into the trained interest point neural network to detect whether the parent-child relationship threshold is met.

[0074] S130. Match the head entity and tail entity that meet the parent-child relationship threshold with the name of the point of interest to which the source file belongs.

[0075] S140. If the geographic information of the points of interest corresponding to the head entity and the tail entity meets the preset threshold, then deduplication is performed; otherwise, new intermediate nodes and leaf nodes are added based on the head entity and the tail entity.

[0076] This invention proposes a deep learning-based scheme for identifying POI hierarchical relationships. It improves upon the limitations of traditional template-based or statistical methods that are restricted by expert experience and manual feature engineering. Furthermore, it can be easily embedded into scheduled tasks to automatically scan the knowledge base daily and push the identification results to website operators, thereby reducing labor costs.

[0077] In a preferred embodiment, step S110 includes:

[0078] S111. Construct a prefix tree using the name and alias fields of all entities in the preset point of interest database.

[0079] S112. Insert the alias strings one by one. The prefix tree has no characters except for the root node, the leaf nodes contain the terminator, and each intermediate node of the prefix tree contains only one character.

[0080] S113. Place the interest point information corresponding to the alias into the leaf node, but this is not the only option.

[0081] In a preferred embodiment, step S112 further includes:

[0082] Starting from the root node and proceeding to a certain node, the characters passed through can be concatenated to represent the string corresponding to that node. From the root node to any leaf node, an entity alias is closed, but this is not the only limit.

[0083] In a preferred embodiment, step S120 includes:

[0084] S121. Obtain at least two entity fields and corresponding candidate interest point numbers from the source file through entity recognition.

[0085] S122. Arrange and combine any two entity fields in the entity field as the head entity and the tail entity respectively to obtain several element combinations as the interest points to be tested.

[0086] S123. Based on each element combination, the corresponding entity fields in the source file are identified as the head entity and the tail entity, respectively, and the recognized text is obtained.

[0087] S124. Input the identified text into the trained interest point neural network respectively, and detect whether each identified text satisfies the parent-child relationship threshold, but not limited to this.

[0088] In a preferred embodiment, step S124 includes:

[0089] S1241. Assume the word sequence after segmenting the recognized text and adding positioning characters at the beginning is x. 1:T The latent vector sequence obtained after BERT feature extraction is as follows:

[0090] h 1:T =BERT(x 1:T ).

[0091] S1242. Assuming the head entity and tail entity are located at positions i to j and k to m in the input sequence, respectively, their word-level vector representations are average-pooled and then fed into two fully connected layers. The output representations of the head and tail entities are as follows:

[0092]

[0093]

[0094] S1243. Obtain the latent vector corresponding to the position of the located character, and process it through a fully connected layer to obtain:

[0095] h cls =FFN cls (h1), where CLS is the positioning character.

[0096] S1244. Concatenate the latent vector of the located character with the head and tail entity vectors, and then pass it into a fully connected layer and an activation function layer to output a probability model where the head entity is the parent interest point of the tail entity:

[0097]

[0098] S1245. In the training phase, the probabilistic model uses the binary cross-entropy function as the objective. Assuming the true label of the sample is y, the loss is:

[0099]

[0100] S1246. During training, the AdamW algorithm is used to optimize the loss in batches until the function converges. An appropriate threshold is selected using the development set data. During the inference phase, entity pairs whose output probability is greater than the threshold are predicted as parent-child relationships, but this is not a limitation.

[0101] In a preferred embodiment, step S130 includes:

[0102] S131. Determine whether the number of the interest point to which the source text belongs matches the candidate number recalled by the head entity and / or tail entity. If yes, proceed to step S133; otherwise, proceed to step S132.

[0103] S132. After removing the prefixes such as province, city, and administrative region, as well as the preset suffix fields, from the names of the head entity and / or tail entity and the interest point to which the source text belongs, return to step S131.

[0104] If the head entity or tail entity is successfully matched in step S133, proceed to step S140, but this is not the only option.

[0105] In a preferred embodiment, step S140 includes:

[0106] S141. Obtain the first geographic information of a successfully matched head or tail entity.

[0107] S142. Obtain the second geographic information of at least one candidate point of interest from the head or tail entity that was not successfully matched.

[0108] S143. Determine whether the distance between the first geographic information and a second geographic information meets the preset threshold. If yes, then deduplicate the points of interest to be tested. If no, then add new intermediate nodes and leaf nodes based on the head entity and tail entity of the points of interest to be tested, but not limited to this.

[0109] The purpose of this invention is to provide a deep learning-based interest point classification method that utilizes information from a tourism knowledge base and uses a pre-trained model for relationship classification to normalize interest points and greatly improve the deduplication and recognition rate of interest points.

[0110] Figure 2 A schematic diagram of the prefix tree in the implementation process of the deep learning-based interest point classification method of the present invention. Figure 3 A schematic diagram of the BERT model during the implementation of the deep learning-based interest point classification method of this invention. (See diagram below.) Figure 2 , 3 As shown, the POI hierarchical relationship identification framework proposed in this invention mainly includes four modules: entity identification, relationship classification, name matching, and geographic verification.

[0111] I. Entity Recognition Module

[0112] Most Points of Interest (POIs) have several important attributes, including name, alias, latitude and longitude, address, and description. The description, which usually contains a large amount of text, is the main field used to extract parent and child POIs. The entity recognition module is configured to use a pipeline extraction strategy, which first identifies entity references that may represent POIs from the text, and then classifies the relationships between entities.

[0113] Entity reference fragment identification is achieved through a trie search. First, a prefix tree is constructed using the name and alias fields of all entities in the knowledge base. Except for the root node (which contains no characters) and leaf nodes (which contain an end character), each intermediate node in the prefix tree contains only one character. The characters traversed from the root node to any given node, when concatenated, represent the string corresponding to that node; therefore, all successor nodes of each node in the tree share the same prefix. In the prefix tree, a single entity alias is closed from the root node to any leaf node.

[0114] By using a pre-built alias prefix tree for forward maximum matching, entity names located in the knowledge base can be retrieved from POI description text. Since the leaf nodes of the prefix tree can store an array of POI IDs with that string as an alias, i.e., an inverted index can be built, the prefix tree matching returns not only entity mentions but also the POI IDs (points of interest codes) corresponding to each mention.

[0115] II. Relationship Classification Module

[0116] The relationship classification module is configured to combine all entity mentions output by the previous module in pairs to obtain candidate entity pairs. In this step, a neural network is used to determine the relationships between the entity pairs. Since the hierarchical relationship is inverse, only one type needs to be identified, so this task can be regarded as binary classification.

[0117] See Figure 2 In this invention, we predict whether the head entity is the parent node of the tail entity. For example, the description of the POI "Hangzhou West Lake" includes phrases such as "...West Lake and its surrounding area have more than 100 parks and scenic spots... Famous scenic spots include Broken Bridge, Leifeng Pagoda, Xiling Seal Engraving Society...". Assuming that entity recognition yields four mentions: "West Lake", "Broken Bridge", "Leifeng Pagoda", and "Xiling Seal Engraving Society", there are 12 possible entity pair arrangements, such as (West Lake, Broken Bridge), (West Lake, Leifeng Pagoda), etc. It's important to note that "West Lake" is the parent POI of "Broken Bridge," therefore (West Lake, Broken Bridge) is a positive example, while (Broken Bridge, West Lake) is a negative example.

[0118] See Figure 3Entity-to-relation classification is implemented using R-BERT. This model can utilize the large amount of linguistic knowledge obtained by BERT during the pre-training stage to improve the prediction effect. The skeleton of R-BERT is the same as that of BERT, which is the Transformer bidirectional encoder structure. It fully extracts contextual semantic information using a multi-head self-attention mechanism. Relation extraction – R-BERT is an important task in natural language processing. Relation extraction tasks are divided into pipeline extraction and joint line extraction. Among them, the pipeline extraction method generally divides relation extraction into two processes: first, entity extraction is performed, and then the entities are classified to obtain the relationship between entities, thus transforming relation extraction into a classification problem. Joint line extraction extracts entities and relations at the same time. Joint line relation extraction will become an important direction for relation extraction. The R-BERT model is a relation classification model based on the pre-trained BERT model. The core parts of the model are roughly as follows: (1) Use the pre-trained BERT to perform word feature representation. (2) Extract the feature information of the special marker symbol "[CLS]" in the pre-trained BERT classification task, the feature information of the subject, and the feature information of the object for relation classification. The R-BERT model utilizes entity information for classification, proposing a novel relationship classification model.

[0119] For input, two special symbols, "$" and "#", are needed to identify the positions of the head and tail entities in the original text. For example, when classifying (West Lake, Broken Bridge), the input text is: "88 $West Lake $ and its surrounding areas have more than 100 parks and scenic spots 88 Famous scenic spots include #Broken Bridge#, Leifeng Pagoda, Xiling Seal Engraving Society 88", and the output label is 1. Assuming the word sequence after word segmentation and adding [CLS] to the head of the text is x 1:T The latent vector sequence obtained after BERT feature extraction is as follows:

[0120] h 1:T =BERT(x 1:T )

[0121] Assuming the head and tail entities are located at positions i to j and k to m in the input sequence, respectively, their word-level vector representations are average-pooled and then fed into two fully connected layers to output the representations of the head and tail entities.

[0122]

[0123]

[0124] The hidden vector corresponding to the [CLS] position is obtained and processed through a fully connected layer.

[0125] h cls =FFNcls (h1)

[0126] The [CLS] vector is concatenated with the head and tail entity vectors, and then passed to a fully connected layer and a sigmoid layer to output the probability that the head entity is the parent POI of the tail entity.

[0127]

[0128] The model uses a binary cross-entropy function as the objective during the training phase. Assuming the true label of the sample is y, the loss is...

[0129]

[0130] During training, the AdamW algorithm is used to optimize the loss in batches until the function converges. An appropriate threshold is selected using development set data. During the inference phase, entity pairs whose output probability is greater than the threshold are predicted as parent-child relationships.

[0131] III. Name Matching Module

[0132] The name matching module is configured to match entity pairs predicted as hierarchical relationships in the previous step with the names of the POIs to which the source text belongs, ensuring that either the head entity or the tail entity corresponds to that POI. Since the entity recognition module returns the POI ID with that reference as an alias while recognizing the entity mention, let's take the head entity as an example:

[0133] (1) If the id of the POI to which the source text belongs is exactly in the candidate id of the header entity recall, then the match is successful.

[0134] (2) If (1) fails to match, remove the prefixes such as province, city, administrative region and suffixes such as "scenic area", "landscape area" and "resort area" from the names of the POI to which the header entity and the source text belong. If the remaining parts are the same, the match is successful.

[0135] For example, POI1xxx7 is named "Hangzhou West Lake," and its description reads "West Lake and its surrounding area have more than 100 parks and scenic spots, including the Broken Bridge, Leifeng Pagoda, and Xiling Seal Engraving Society." We can identify the triple (West Lake, Broken Bridge, parent-child) from this description. Furthermore, the prefix tree retrieves candidate IDs POI1xxx5 and POI1xxx6 for "West Lake." Although the ID for "Hangzhou West Lake" is not among them, its name, after removing "Hangzhou," is completely consistent with the head entity. Therefore, the head entity successfully matches the POI to which the source text belongs. The same logic applies to the tail entity.

[0136] IV. Geographic Validation Module

[0137] For the triples received by the name matching module, if one endpoint matches successfully, it is also necessary to consider whether all candidate POIs recalled by the other endpoint can be considered as the parent or child of the POI to which the source text belongs. For example, in the triple (West Lake, Broken Bridge, Parent and Child), suppose there are POIs with "Broken Bridge" as an alias, including the Broken Bridge in Hangzhou and the Broken Bridge Scenic Area in Hulunbuir, but only the former is a child POI of "West Lake in Hangzhou". In this step, the present invention can verify the locational relationship between the POI that is successfully matched in the name matching module and the candidate POIs at the other end of the triple, following the following logic:

[0138] (1) The POIs corresponding to the head and tail entities must be located in the same province or city.

[0139] (2) The distance between the POIs corresponding to the head and tail entities cannot exceed 5 kilometers.

[0140] Finally, duplicate POI pairs that pass the verification are removed. If the hierarchical relationship is not in the current knowledge base, it is output as newly discovered knowledge.

[0141] This invention proposes a POI hierarchical relationship detection method that fully utilizes information from a tourism knowledge base and uses a pre-trained model for relationship classification, resulting in a significant improvement in performance compared to traditional methods. This invention has been applied to the POI relationship recognition project in the ticketing department, achieving an accuracy rate exceeding 92% after implementation.

[0142] One embodiment of the present invention is as follows:

[0143] First, generate the entity alias prefix tree:

[0144] Build an inverted index from aliases to POI IDs using the full data in the knowledge base.

[0145] Initialize a prefix tree, insert the alias strings one by one, and put the POI id corresponding to the alias into the leaf node.

[0146] Then, the relationship classification model is trained, and the specific steps are as follows:

[0147] A batch of text was manually labeled, including the location of entity pairs in the text and the relation type of entity pairs. The labeled data was divided into training and validation sets. The loss of the R-BERT model on the training data was calculated, and the AdamW optimization algorithm was used to minimize the objective function. The model parameters were iteratively updated in batches until convergence. The AdamW optimizer provides an adaptive gradient method, and it is an improved algorithm based on Adam + L2 regularization. Using Adam to optimize the loss with L2 regularization is not effective. If an L2 regularization term is introduced, the gradient of the regularization term will be added when calculating the gradient. Adam is a first-order optimization algorithm that can replace the traditional stochastic gradient descent (SGD) process. It can iteratively update the neural network weights based on the training data. Adam was initially proposed by Diederik Kingma of OpenAI and Jimmy Ba of the University of Toronto in their 2015 ICLR paper (Adam: A Method for Stochastic Optimization).

[0148] The online identification process using the method of this invention is as follows:

[0149] (1) Scan the knowledge base and use the entity recognition module to retrieve the entity mentions and their corresponding candidate POI IDs for the introductory text of each POI.

[0150] (2) For the results identified in (1), input them into the relation classification module to generate candidate entity pairs and use the R-BERT model to classify the entity pair relations;

[0151] (3) Input the entity pairs predicted as parent-child relationships in (2) into the name matching module, and match the id and name of the POI to which the description text belongs with the id and name of the head entity and the tail entity respectively;

[0152] (4) Input the successful matching results in (3) into the geographic verification module to verify the similarity between the candidate POI at the other end of the triplet and the current POI in terms of administrative division and geographical location. POI pairs that meet the business logic are output as the final result.

[0153] (5) If the POI parent-child relationship output by (4) is not in the current knowledge base, then push it to the event system.

[0154] Figure 4 This is a schematic diagram of the modules of the deep learning-based interest point classification system of the present invention. Figure 4 As shown, embodiments of the present invention also provide a deep learning-based interest point classification system for implementing the above-described deep learning-based interest point classification method. The deep learning-based interest point classification system includes:

[0155] The prefix tree generation module 51 constructs a prefix tree based on the names of entities in the preset interest point database. Each intermediate node contains only one character, and the interest point information corresponding to the alias is placed in the leaf node.

[0156] The relationship detection module 52 obtains at least two entity fields and corresponding candidate interest point numbers from the source file through entity recognition. The source file, after being labeled as the head entity and tail entity respectively, is input into the trained interest point neural network to detect whether the parent-child relationship threshold is met.

[0157] The name matching module 53 matches the head entity and tail entity that meet the parent-child relationship threshold with the name of the point of interest to which the source file belongs.

[0158] The interest point deduplication module 54 deduplicates the points of interest when the geographic information of the points of interest corresponding to the head entity and the tail entity meets the preset threshold. If so, it deduplicates the points of interest; otherwise, it adds new intermediate nodes and leaf nodes based on the head entity and the tail entity.

[0159] In a preferred embodiment, the prefix tree generation module 51 is configured to construct a prefix tree using the name and alias fields of all entities in a preset point of interest database. Alias ​​strings are inserted one by one; the prefix tree contains no characters except for the root node, leaf nodes contain a terminator, and each intermediate node contains only one character. The point of interest information corresponding to the alias is placed in the leaf node, but this is not a limitation.

[0160] In a preferred embodiment, the prefix tree generation module 51 is further configured to start from the root node and go to a certain node, and the characters passed through are connected to represent the string corresponding to the node. The path from the root node to any leaf node closes an entity alias, but is not limited to this.

[0161] In a preferred embodiment, the relationship detection module 52 is configured to obtain at least two entity fields and corresponding candidate interest point numbers from the source file through entity recognition. Any two entity fields are arranged as the head entity and tail entity, respectively, to obtain several element combinations as interest points to be tested. Based on each element combination, the corresponding entity fields in the source file are identified as the head entity and tail entity, respectively, to obtain the recognized text. The recognized texts are then input into a trained interest point neural network to detect whether each recognized text satisfies a parent-child relationship threshold, but is not limited to this.

[0162] In a preferred embodiment, the relation detection module 52 is further configured to assume that the word sequence after segmenting the identified text and adding a positioning character at the beginning is x. 1:T The latent vector sequence obtained after BERT feature extraction is as follows:

[0163] h 1:T =BERT(x 1:T )

[0164] Assuming the head and tail entities are located at positions i to j and k to m in the input sequence, respectively, their word-level vector representations are average-pooled and then fed into two fully connected layers. The output representations of the head and tail entities are as follows:

[0165]

[0166]

[0167] The latent vector corresponding to the position of the located character is obtained, and then processed through a fully connected layer to obtain:

[0168] h cls =FFN cls (h1), where CLS is the positioning character.

[0169] The latent vector of the located character is concatenated with the head and tail entity vectors, and then fed into a fully connected layer and an activation function layer to output a probability model in which the head entity is the parent interest point of the tail entity:

[0170]

[0171] The probabilistic model uses a binary cross-entropy function as the objective during the training phase. Assuming the true label of the sample is y, the loss is:

[0172]

[0173] During training, the AdamW algorithm is used to optimize the loss in batches until the function converges. An appropriate threshold is selected using the development set data. During the inference phase, entity pairs whose output probability is greater than the threshold are predicted as parent-child relationships, but this is not a limitation.

[0174] In a preferred embodiment, the name matching module 53 is configured to determine whether the number of the point of interest to which the source text belongs matches the candidate number recalled by the head entity and / or tail entity. If yes, the head entity or tail entity is successfully matched. If no, the names of the head entity and / or tail entity and the point of interest to which the source text belongs are each modified by removing prefixes such as province, city, and administrative region, as well as preset suffix fields, and then the matching is performed again.

[0175] In a preferred embodiment, the point of interest deduplication module 54 is configured to obtain a first geographic information corresponding to a successfully matched head entity or tail entity; obtain second geographic information of at least one recalled candidate point of interest from the head entity or tail entity that did not match; determine whether the distance between the first geographic information and a second geographic information meets a preset threshold; if so, deduplication of the point of interest to be tested is performed; if not, new intermediate nodes and leaf nodes are added based on the head entity and tail entity of the point of interest to be tested, but this is not a limitation.

[0176] The purpose of this invention is to provide a deep learning-based interest point classification method that utilizes information from a tourism knowledge base and uses a pre-trained model for relationship classification to normalize interest points and greatly improve the deduplication and recognition rate of interest points.

[0177] This invention also provides a deep learning-based interest point classification device, including a processor and a memory storing executable instructions for the processor. The processor is configured to execute steps of a deep learning-based interest point classification method via the executable instructions.

[0178] As shown above, this embodiment of the deep learning-based interest point classification system of the present invention utilizes information from the tourism knowledge base and uses a pre-trained model to classify relationships, thereby normalizing interest points and greatly improving the deduplication recognition rate of interest points.

[0179] Those skilled in the art will understand that various aspects of the present invention can be implemented as systems, methods, or program products. Therefore, various aspects of the present invention can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software aspects, collectively referred to herein as a "circuit," "module," or "platform."

[0180] Figure 5 This is a schematic diagram of the interest point classification device based on deep learning according to the present invention. See below for reference. Figure 5 To describe an electronic device 600 according to this embodiment of the present invention. Figure 5 The electronic device 600 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0181] like Figure 5 As shown, the electronic device 600 is presented in the form of a general-purpose computing device. The components of the electronic device 600 may include, but are not limited to: at least one processing unit 610, at least one storage unit 620, a bus 630 connecting different platform components (including storage unit 620 and processing unit 610), a display unit 640, etc.

[0182] The storage unit stores program code, which can be executed by the processing unit 610 to perform the steps described in the above-described section on the electronic prescription transfer processing method according to various exemplary embodiments of the present invention. For example, the processing unit 610 can perform actions such as... Figure 1 The steps are shown in the figure.

[0183] Storage unit 620 may include a readable medium in the form of a volatile storage unit, such as random access memory (RAM) 6201 and / or cache memory 6202, and may further include a read-only memory (ROM) 6203.

[0184] Storage unit 620 may also include a program / utility 6204 having a set (at least one) program module 6205, such program module 6205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0185] Bus 630 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the multiple bus structures.

[0186] Electronic device 600 can also communicate with one or more external devices 700 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 600, and / or with any device that enables electronic device 600 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 650. Furthermore, electronic device 600 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 660. Network adapter 660 can communicate with other modules of electronic device 600 via bus 630. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.

[0187] This invention also provides a computer-readable storage medium for storing a program, which, when executed, implements the steps of a deep learning-based interest point classification method. In some possible implementations, various aspects of the invention can also be implemented as a program product comprising program code that, when run on a terminal device, causes the terminal device to perform the steps described in the above-described electronic prescription processing method section of this specification according to various exemplary embodiments of the invention.

[0188] As shown above, this embodiment of the deep learning-based interest point classification system of the present invention utilizes information from the tourism knowledge base and uses a pre-trained model to classify relationships, thereby normalizing interest points and greatly improving the deduplication recognition rate of interest points.

[0189] Figure 5 This is a schematic diagram of the structure of the computer-readable storage medium of the present invention. (Reference) Figure 5 As shown, a program product 800 for implementing the above-described method according to an embodiment of the present invention is described. This product may employ a portable compact disc read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the program product of the present invention is not limited thereto. In this document, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0190] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0191] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0192] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0193] In summary, the purpose of this invention is to provide a method, system, device, and storage medium for classifying interest points based on deep learning. By utilizing information from a tourism knowledge base and using a pre-trained model for relationship classification, the invention achieves the normalization of interest points and greatly improves the deduplication and recognition rate of interest points.

[0194] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1.A point of interest classification method based on deep learning, characterized in that, The method comprises the following steps: constructing a prefix tree based on the names of entities in a preset interest point database, each intermediate node containing only one character, and placing the interest point information corresponding to the alias in the leaf node; obtaining at least two entity fields and corresponding candidate interest point numbers from the source file through entity recognition, inputting the source file in which any two of the entity fields are identified as head entity and tail entity respectively into a trained interest point neural network, and detecting whether the parent-child relationship threshold is met; matching the head entity and the tail entity that meet the parent-child relationship threshold with the name of the interest point to which the source file belongs respectively; judging whether the geographic information of the interest point corresponding to the head entity and the tail entity respectively meets a preset threshold, and if so, removing the duplicate, and if not, adding new intermediate nodes and leaf nodes according to the head entity and the tail entity. 2.The deep learning-based point of interest classification method of claim 1, wherein, The method of constructing a prefix tree based on the names of entities in a preset interest point database, each intermediate node containing only one character, and placing the interest point information corresponding to the alias in the leaf node comprises: constructing a prefix tree based on the names of all entities and alias fields in the preset interest point database; inserting the alias string piece by piece, the prefix tree not containing characters except the root node and containing a terminator in the leaf node, each intermediate node of the prefix tree containing only one character; placing the interest point information corresponding to the alias in the leaf node. 3.The deep learning-based point of interest classification method of claim 2, wherein, The method of inserting the alias string piece by piece, the prefix tree not containing characters except the root node and containing a terminator in the leaf node, each intermediate node of the prefix tree containing only one character, further comprises: from the root node to a certain node, the characters passed represent the string corresponding to the node, and from the root node to any leaf node, an entity alias is closed. 4.The deep learning-based point of interest classification method of claim 1, wherein, The method of obtaining at least two entity fields and corresponding candidate interest point numbers from the source file through entity recognition, inputting the source file in which any two of the entity fields are identified as head entity and tail entity respectively into a trained interest point neural network, and detecting whether the parent-child relationship threshold is met comprises: obtaining at least two entity fields and corresponding candidate interest point numbers from the source file through entity recognition; arranging and combining any two of the entity fields as head entity and tail entity respectively to obtain a plurality of element combinations as interest points to be tested; identifying the corresponding entity fields in the source file as head entity and tail entity respectively according to each of the element combinations to obtain recognition texts respectively; inputting the recognition texts into a trained interest point neural network to detect whether each of the recognition texts meets the parent-child relationship threshold. 5.The deep learning-based point of interest classification method of claim 1, wherein, The method of matching the head entity and the tail entity that meet the parent-child relationship threshold with the name of the interest point to which the source file belongs respectively comprises: determining whether the number of the interest point to which the source text belongs hits the candidate number recalled by the head entity and / or the tail entity, if yes, the head entity or the tail entity matches successfully, if not, removing the province, city, administrative district prefix and preset suffix field from the name of the head entity and / or the tail entity and the interest point to which the source text belongs respectively, and then performing matching again. 6.The deep learning-based point of interest classification method of claim 4, wherein, The method further includes: determining whether the geographic information of the interest point corresponding to the head entity and the tail entity respectively satisfies a preset threshold, if yes, deduplicating, if not, adding new intermediate nodes and leaf nodes according to the head entity and the tail entity, including: obtaining first geographic information corresponding to one of the head entity or the tail entity that matches successfully; obtaining second geographic information of at least one candidate interest point recalled for one of the head entity or the tail entity that does not match successfully; determining whether the distance between the first geographic information and one of the second geographic information satisfies a preset threshold, if yes, deduplicating the interest point under test, if not, adding new intermediate nodes and leaf nodes according to the head entity and the tail entity of the interest point under test. The method further includes: constructing a prefix tree based on at least the name of the entity of the preset interest point database, each intermediate node containing only one character, and placing the alias corresponding interest point information to the leaf node; obtaining at least two entity fields and corresponding candidate interest point numbers from the source file through entity recognition; inputting the source file after identifying any two of the entity fields as the head entity and the tail entity respectively into the interest point neural network trained to detect whether the parent-child relationship threshold is satisfied; matching the head entity and the tail entity satisfying the parent-child relationship threshold with the name of the interest point to which the source file belongs respectively; determining whether the geographic information of the interest point corresponding to the head entity and the tail entity respectively satisfies a preset threshold, if yes, deduplicating, if not, adding new intermediate nodes and leaf nodes according to the head entity and the tail entity. The method further includes: a processor; a memory having executable instructions of the processor stored therein; and the processor configured to execute the executable instructions to perform the steps of the interest point classification method based on deep learning of any one of claims 1 to 6. The program, when executed by the processor, implements the steps of the interest point classification method based on deep learning of any one of claims 1 to 6. 7.A point of interest classification system based on deep learning, configured to implement the point of interest classification method based on deep learning of claim 1, characterized in that, ​ ​ ​ ​ ​ 8.A point of interest classification device based on deep learning, characterized by, ​ ​ ​ ​ 9. A computer readable storage medium for storing a program, characterized in that, ​

Images