Method and device for generating event description information, storage medium, and electronic device

Abstract

The present invention discloses a method and device for generating event description information, a storage medium, and an electronic device. The method comprises: determining a currently released incremental event resource from an event resource stream; comparing the incremental event description information corresponding to the incremental event resource with the object event description information contained in an event resource cluster; clustering the incremental event resource into a target event resource subcluster and merging the event description information into the target event resource subcluster if the comparison result indicates that the incremental event description information and the object event description information contained in a target event resource subcluster in the event resource cluster meet clustering conditions; and generating hot event description information for the hot event based on the object event description information contained in the target event resource subcluster if the target object event corresponding to the target event resource subcluster is determined to be a hot event. The present invention solves the technical problem of poor timeliness in generating hot event description information.

Description

Technical Field

[0001] The present invention relates to the field of computers, and in particular to a method and device for generating event description information, a storage medium, and an electronic device. Background Art

[0002] In various information applications, it is usually necessary to generate corresponding event key information in a timely manner based on the hot events currently released on the Internet, and publish the hot events in the form of "hot search list" or "local list" to guide users to understand the current real-time hot spots based on the list.

[0003] The existing method of generating key event information is often to save hot event resources in batches, and perform multiple classification operations based on the batch-saved hot event resources to obtain multiple event clusters, and then determine the output hot information and the entries output on the hot list based on the classified event clusters.

[0004] This clustering method requires hot event resources to reach a certain scale before effective classification can be performed. It cannot cluster newly added event resources in real time, resulting in delayed classification. Consequently, the hot term entries generated based on the classification operation are less timely. In other words, the existing method suffers from the problem of poor timeliness in the generated description information.

[0005] To address the above-mentioned problems, no effective solutions have been proposed so far. Summary of the Invention

[0006] Embodiments of the present invention provide a method and apparatus for generating event description information, a storage medium, and an electronic device, so as to at least solve the technical problem of poor timeliness in generating hot event description information.

[0007] According to one aspect of an embodiment of the present invention, a method for generating event description information is provided, comprising: determining a currently released incremental event resource from an event resource stream; comparing the incremental event description information corresponding to the incremental event resource with the object event description information contained in an event resource cluster, wherein the event resource cluster includes at least one event resource subcluster obtained after streaming clustering of historical event resources, the historical event resource is an event resource released before the incremental event resource in the event resource stream, and the event resource subcluster includes at least two object event description information associated with the same object event; when the comparison result indicates that the incremental event description information and the object event description information contained in the target event resource subcluster in the event resource cluster meet the clustering conditions, clustering the incremental event resource into the target event resource subcluster, and merging the event description information into the target event resource subcluster; when the target object event corresponding to the target event resource subcluster is determined to be a hot event, generating hot event description information of the hot event based on the object event description information contained in the target event resource subcluster.

[0008] According to another aspect of an embodiment of the present invention, a device for generating event description information is also provided, including: a determination unit for determining a currently released incremental event resource from an event resource stream; a comparison unit for comparing the incremental event description information corresponding to the incremental event resource with the object event description information contained in an event resource cluster, wherein the event resource cluster includes at least one event resource subcluster obtained after streaming clustering of historical event resources, the historical event resource is an event resource released before the incremental event resource in the event resource stream, and the event resource subcluster includes at least two object event description information associated with the same object event; a clustering unit for clustering the incremental event resource into the target event resource subcluster and merging the event description information into the target event resource subcluster when the comparison result indicates that the incremental event description information and the object event description information contained in the target event resource subcluster in the event resource cluster meet the clustering conditions; a generation unit for generating hot event description information of the hot event based on the object event description information contained in the target event resource subcluster when the target object event corresponding to the target event resource subcluster is determined to be a hot event.

[0009] According to another aspect of the embodiments of the present invention, a computer-readable storage medium is provided, in which a computer program is stored. The computer program is configured to execute the above-mentioned method for generating event description information when running.

[0010] According to another aspect of an embodiment of the present application, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the above-described method for generating event description information.

[0011] According to another aspect of an embodiment of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the method for generating event description information through the computer program.

[0012] In an embodiment of the present invention, the following method is adopted: first, the currently released incremental event resource is determined from the event resource stream; the incremental event description information corresponding to the incremental event resource is compared with the object event description information contained in the event resource cluster; when the comparison result indicates that the incremental event description information and the object event description information contained in the target event resource subcluster in the event resource cluster meet the clustering condition, the incremental event resource is clustered into the target event resource subcluster, and the event description information is merged into the target event resource subcluster; when the target object event corresponding to the target event resource subcluster is determined to be a hot event, the hot event description information of the hot event is generated based on the object event description information contained in the target event resource subcluster. Then, real-time streaming clustering is performed based on the incremental event resources in the incremental event resource stream, and event description information is generated based on the resource cluster generated by the streaming clustering, thereby ensuring the timeliness of the event description information, that is, solving the existing technical problem of the timeliness of the generation of hot event description information. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0014] Figure 1 is a schematic diagram of a hardware environment for an optional method for generating event description information according to an embodiment of the present invention;

[0015] Figure 2 is a flowchart of an optional method for generating event description information according to an embodiment of the present invention;

[0016] Figure 3 is a schematic diagram of an optional method for generating event description information according to an embodiment of the present invention;

[0017] Figure 4is a schematic diagram of another optional method for generating event description information according to an embodiment of the present invention;

[0018] Figure 5 is a schematic diagram of another optional method for generating event description information according to an embodiment of the present invention;

[0019] Figure 6 is a schematic diagram of another optional method for generating event description information according to an embodiment of the present invention;

[0020] Figure 7 is a schematic diagram of another optional method for generating event description information according to an embodiment of the present invention;

[0021] Figure 8 is a flowchart of another optional method for generating event description information according to an embodiment of the present invention;

[0022] Figure 9 is a schematic structural diagram of an optional device for generating event description information according to an embodiment of the present invention;

[0023] Figure 10 FIG. 4 is a schematic structural diagram of an optional electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0025] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] According to one aspect of an embodiment of the present invention, a method for generating event description information is provided. As an optional implementation, the method for generating event description information can be applied to, but is not limited to, Figure 1 In the event description information generation system composed of the server 102 and the terminal device 104 shown in FIG. Figure 1As shown, the server 102 is connected to the terminal device 104 via the network 110. The above-mentioned network may include but is not limited to: a wired network, a wireless network, wherein the wired network includes: a local area network, a metropolitan area network and a wide area network, and the wireless network includes: Bluetooth, WIFI and other networks that realize wireless communication. The above-mentioned terminal devices may include but are not limited to at least one of the following: a mobile phone (such as an Android phone, an iOS phone, etc.), a laptop computer, a tablet computer, a PDA, a MID (Mobile Internet Devices), a PAD, a desktop computer, a smart TV, a car-mounted device, etc. An application with a hot event aggregation function may be installed on the above-mentioned terminal device, such as a social APP that can display a real-time hot search list, a search APP that can display hot events in the search interface, and an instant messaging APP that can display the progress of hot events in the news column. The terminal device is also provided with a display, a processor, and a memory. The display can be used to display the program interface with the hot event aggregation function, as well as the content of the event to be published and the event description information of the generated hot event. The processor can be used to pre-process the content of the event to be published, for example, splicing the acquired event image and event text content into an event resource; the memory is used to store the event resource to be published. It is understood that after the terminal device 104 acquires the event resource to be published, the terminal device 104 can then send the event resource to the server 102 via the network 110. After the server 102 receives the event resources sent by multiple terminal devices, it obtains the currently published incremental event resource from the event resource stream composed of the event resources sent by the multiple terminal devices, and performs a clustering operation based on the incremental event resource and multiple event clusters. When the clustering conditions are met, the incremental event resource is clustered with the target event. Furthermore, when the output conditions of the hot event are met, the description information of the hot event is output based on the hot event cluster; the terminal device 104 can receive the description information of the hot event returned by the server 102 via the network 110 and display the description information of the hot event. Furthermore, in response to a selection operation on the description information of a hot event, the detailed content of the hot event can be displayed. The server 102 can be a single server, a server cluster consisting of multiple servers, or a cloud server. The server includes a database and a processing engine. The database can include multiple event resource subclusters for clustering operations; the processing engine is used to perform clustering operations based on the acquired incremental event resources and event resource subclusters.

[0027] According to one aspect of an embodiment of the present invention, the above-mentioned event description information generation system can also perform the following steps: the terminal device 104 executes step S102 to generate an incremental event resource; then executes step S104 to send the incremental event resource to the server 102 through the network 110; the server 102 executes steps S106 to S112 to determine the currently released incremental event resource from the event resource stream; compares the incremental event description information corresponding to the incremental event resource with the object event description information contained in the event resource cluster; the event resource cluster includes at least one event resource subcluster obtained by streaming clustering of historical event resources, and the historical event resources For event resources published before the incremental event resources in the event resource stream, the event resource subcluster includes at least two object event description information associated with the same object event; if the comparison result indicates that the incremental event description information meets the object event description information contained in the target event resource subcluster in the event resource cluster, the incremental event resource is clustered into the target event resource subcluster, and the event description information is merged into the target event resource subcluster; if the target object event corresponding to the target event resource subcluster is determined to be a hot event, hot event description information of the hot event is generated based on the object event description information contained in the target event resource subcluster. Next, the server 102 executes step S114 to send the hot event description information; if the terminal device 104 receives the hot event description information, it displays the hot event description information. It is understandable that if the terminal device 104 is a device with sufficient computing and processing capabilities, the above steps S106 and S112 can also be performed in the terminal device 104.

[0028] In an embodiment of the present invention, the currently released incremental event resources are first determined from the event resource stream; the incremental event description information corresponding to the incremental event resources is compared with the object event description information contained in the event resource cluster; when the comparison result indicates that the incremental event description information and the object event description information contained in the target event resource subcluster in the event resource cluster meet the clustering conditions, the incremental event resources are clustered into the target event resource subcluster, and the event description information is merged into the target event resource subcluster; when the target object event corresponding to the target event resource subcluster is determined to be a hot event, hot event description information of the hot event is generated based on the object event description information contained in the target event resource subcluster, and then real-time streaming clustering is performed based on the incremental event resources in the incremental event resource stream, and event description information is generated based on the resource cluster generated by the streaming clustering, thereby ensuring the timeliness of the event description information, that is, solving the existing technical problem of the timeliness of the generation of hot event description information.

[0029] The above is only an example and is not limited in this embodiment.

[0030] As an optional implementation, Figure 2 As shown, the method for generating the above event description information includes the following steps:

[0031] S202, determining the currently published incremental event resource from the event resource stream;

[0032] S204, comparing the incremental event description information corresponding to the incremental event resource with the object event description information contained in the event resource cluster;

[0033] The event resource cluster includes at least one event resource sub-cluster obtained by stream clustering of historical event resources, the historical event resources being event resources published before the incremental event resources in the event resource stream, and the event resource sub-cluster including at least two object event description information associated with the same object event;

[0034] S206, if the comparison result indicates that the incremental event description information and the object event description information contained in the target event resource subcluster in the event resource cluster meet the clustering condition, cluster the incremental event resources into the target event resource subcluster, and merge the event description information into the target event resource subcluster;

[0035] S208 : When the target object event corresponding to the target event resource subcluster is determined to be a hot event, generate hot event description information of the hot event based on the object event description information included in the target event resource subcluster.

[0036] It should be noted that the above-mentioned event resources can be media resources that describe an event, and may include but are not limited to event text resources, event image resources, event video resources, etc. The above-mentioned event text resources, event image resources, and event video resources can be used to describe an event. The types of the above-mentioned event resources may also include but are not limited to news information, social copy posted on social media, etc. The above types of event resources are all examples, and this embodiment does not limit the specific resource types.

[0037] It should be noted that the above-mentioned implementation methods of the present application can be specifically applied to applications with hot event aggregation functions, such as social apps that can display real-time hot search lists, search apps that can display hot events in the search interface, and instant messaging apps that can display the progress of hot events in the news column.

[0038] The following further explains the streaming clustering method. Streaming clustering is a real-time online clustering method that can be specifically manifested by instantly acquiring incremental event resources generated online in real time and performing clustering operations sequentially, thereby making the clustering results more timely.

[0039] like Figure 3 As shown in Figure (a), the generated hot event description information can be the hot news title information in the hot search list. This product sorts all the mined events in reverse order of popularity and generates an authoritative hot event list for the entire network. The main product entrances include direct access, start page, home page, etc. Figure 3 As shown in Figure (b), the generated hot event description information can be the local hot news title information in the local hot search list. This product classifies all the mined events according to regional labels, and generates corresponding local hot lists for different cities, with the same entrance as the hot search list. As shown in Figure (c) in Figure 3, the generated hot event description information is displayed in a time-sensitive card in a search application. For relatively hot events, a time-sensitive card will be generated at the top of the search results page, including the header image and selected high-quality articles related to the event. Figure 3 As shown in Figure (d), the generated hot event description information can be displayed in the event context column in the search application. For events with multiple reversals or many subsequent developments, the event context will be used to connect the events of each node together, making it easier for users to quickly grasp the cause and effect of the event and the latest developments. Figure 3 As shown in Figure (e), the generated hot event description information can be displayed in the information of the search application. The main feed stream on the home page will mount some authoritative and hot events, as well as their corresponding graphic resources. Figure 3 As shown in Figure (f) in the figure, the generated hot event description information can be used to recommend keywords in the search box of the search application. Some hot events will be extracted as recommended search terms in the homepage and direct search box.

[0040] Optionally, before determining the currently published incremental event resource from the event resource stream, the event resources in the event resource stream may be filtered through the following steps:

[0041] S1, sequentially obtain candidate incremental event resources from the event resource stream;

[0042] S2: If the account characteristics of the publishing account of the candidate incremental event resource meet the account screening conditions, obtain the next candidate incremental event resource;

[0043] S3: If the event content characteristics of the candidate incremental event resource meet the content screening conditions, obtain the next candidate incremental event resource;

[0044] S4. When the account characteristics of the publishing account of the candidate incremental event resource do not meet the account screening conditions, and the event content characteristics of the candidate incremental event resource do not meet the content screening conditions, the candidate incremental event resource is used as the currently published incremental event resource.

[0045] Specifically, because search resources are a mix of user-generated content and professionally produced content, with professionally produced content primarily being published by official accounts, the quality of the content varies widely. Therefore, it's necessary to select high-quality resources for merging. Otherwise, fabricated event resources, or even those that contradict reality, may be introduced, impacting the quality and authority of the resulting events. Resource quality screening primarily focuses on both account and content dimensions.

[0046] From the account perspective, the main process of cleaning is as follows:

[0047] S1. Utilize the account classification system for professional content search to obtain account categories and eliminate accounts with weak news attributes (e.g., accounts related to horoscopes).

[0048] S2 uses a domain expertise model for professional content search to filter out accounts with weak authority in the field, preventing the infiltration of dirty data such as false news and biased commentary on current events, which could affect the authority of event generation.

[0049] S3 uses a professional account quality model for content search to filter out low-quality accounts (e.g., water army accounts) to avoid mixing in dirty data such as false news, which could affect the authority of event generation.

[0050] From the perspective of content cleaning, the main process is as follows:

[0051] S1, remove titles composed of purely invalid characters (including numbers, special symbols, English, etc.);

[0052] S2, remove abnormal titles that are too long or too short;

[0053] S3, filters out resources with low timeliness based on the first release time of the resources;

[0054] S4, using a professional content search model to filter out pornographic content (pornographic content refers to content resources that do not comply with platform regulations);

[0055] S5, using the clickbait model of professional content production search to filter clickbait resources;

[0056] S6, using the relevance model of professional production content search to filter out resources that are not relevant to the topic;

[0057] S7, using the resource category model of professional production content search, discards non-event vertical resources such as "constellation" and "weather".

[0058] Through the above-mentioned implementation mode of the present application, first determine the currently released incremental event resources from the event resource stream; use the incremental event description information corresponding to the incremental event resources to compare with the object event description information contained in the event resource cluster; when the comparison result indicates that the incremental event description information and the object event description information contained in the target event resource subcluster in the event resource cluster meet the clustering conditions, cluster the incremental event resources into the target event resource subcluster, and merge the event description information into the target event resource subcluster; when the target object event corresponding to the target event resource subcluster is determined to be a hot event, generate the hot event description information of the hot event based on the object event description information contained in the target event resource subcluster. Then, perform real-time streaming clustering based on the incremental event resources in the incremental event resource stream, and generate event description information based on the resource cluster generated by the streaming clustering, thereby ensuring the timeliness of the event description information, that is, solving the technical problem of the timeliness of the generation of the existing hot event description information.

[0059] As an optional method, the comparison of the incremental event description information corresponding to the incremental event resource with the object event description information contained in the event resource cluster includes:

[0060] S1, sequentially obtain candidate event resource subclusters from the event resource cluster as the current event resource subcluster;

[0061] S2, sequentially comparing the incremental event description information with the object event description information contained in the current event resource subcluster. As an optional implementation, the sequentially comparing the incremental event description information with the object event description information contained in the current event resource subcluster includes:

[0062] Repeat the following steps until the object event description information included in the current event resource subcluster is traversed:

[0063] S1,,get the current object event description information from the current event resource sub-cluster;

[0064] S2, extracting incremental description features of incremental description information and current description features of current object event description information in the target dual-tower feature extraction model;

[0065] S3, when the feature similarity between the incremental description feature and the current description feature is greater than or equal to a first threshold, updating the clustering coefficient of the current event resource subcluster and obtaining the next object event description information; when the feature similarity between the incremental description feature and the current description feature is less than the first threshold, obtaining the next object event description information;

[0066] S4: When the clustering coefficient of the current event resource subcluster is greater than a second threshold, determine that the current event resource subcluster is a target event resource subcluster that meets the clustering condition.

[0067] In this embodiment, the comparison and clustering operation are performed in a streaming clustering manner. Specifically, a double traversal operation may be included, where the current event resource subcluster is first obtained from the event resource subcluster, and then the event description information of the incremental event resource and the object event description information in the current event resource subcluster are compared in sequence. When the comparison result is greater than the first target threshold, it is determined that the comparison is successful, and the clustering coefficient is updated (for example, the clustering coefficient is increased by one for each successful comparison). In the process of comparing with the object event description information in sequence, when the clustering coefficient exceeds the second target threshold, the current subcluster is determined to be the target subcluster, and the current subcluster and the event description information of the incremental event resource are clustered.

[0068] As an optional method, after the event description information and object event description information of the incremental event resources are enhanced based on the knowledge graph, the information features of the enhanced event description information and object event description information can be extracted respectively, and cosine similarity calculation can be performed. When the calculated similarity is greater than the first target threshold, it is determined that the event description information and the object event description information are successfully compared, and the clustering coefficient is updated.

[0069] The above clustering coefficient can be a counting unit, that is, it is used to record the number of successful comparisons in the process of sequential comparison in subclusters. When the number of successful comparisons exceeds the second target threshold, the current subcluster is judged to be the target subcluster, and a clustering operation is performed based on the target subcluster.

[0070] Through the above-mentioned implementation mode of the present application, the following steps are repeatedly performed until the object event description information included in the current event resource subcluster is traversed: obtaining the current object event description information from the current event resource subcluster; extracting the incremental description features of the incremental description information and the current description features of the current object event description information in the target dual-tower feature extraction model; when the feature similarity between the incremental description features and the current description features is greater than or equal to the first threshold, updating the clustering coefficient of the current event resource subcluster and obtaining the next object event description information; when the feature similarity between the incremental description features and the current description features is less than the first threshold, obtaining the next object event description information; when the clustering coefficient of the current event resource subcluster is greater than the second threshold, determining that the current event resource subcluster is the target event resource subcluster that meets the clustering conditions, thereby using the event description information of the incremental event resource to be sequentially compared with different subclusters and the object event description information in the subclusters, so that there is no need to perform clustering operations based on a complete data set, but the incremental data will be flexibly passed into the data set in sequence, and each data will only be clustered once. This can improve the timeliness of clustering operations and solve the technical problem that existing clustering operations require a lot of time to enrich the data required for offline clustering.

[0071] As an optional implementation, before determining the currently published incremental event resource from the event resource stream, the above further includes:

[0072] S1, training the initial dual-tower feature extraction model using a sample set until a convergence condition is met, wherein the sample set includes search index text and search result text generated based on a network search log set, and the search result text is the text that is hit in the text set searched by the search index text;

[0073] S2, in each training process, inputting the search index text in the sample set into the first feature extraction model in the initial dual-tower feature extraction model to obtain search information features, and inputting the search result text in the sample set into the second feature extraction model in the initial dual-tower feature extraction model to obtain target information features;

[0074] S3, when the feature similarity between the search information feature and the target information feature does not meet the convergence condition, adjusting the model parameters of the initial dual-tower feature extraction model;

[0075] S4, when the feature similarity between the search information feature and the target information feature meets the convergence condition, the initial dual-tower feature extraction model is determined as the target dual-tower feature extraction model.

[0076] It can be understood that in the above embodiment of the present application, before extracting reference description features of reference description information using the target dual-tower feature extraction model, the initial dual-tower feature extraction model needs to be trained using a sample set.

[0077] The above dual-tower feature extraction model structure can be Figure 4 As shown. Specifically, the above sample set can be obtained from the posterior click signal of the search log, wherein the search log includes the search terms searched by the user, multiple search result texts corresponding to the search terms, and the target result text selected by the user from the multiple search results. For example, if a user searches for a query and clicks on an article on the result page, it can be considered that the article is related to the query. By using the data generated in the search log as a training sample, the accuracy of the model feature extraction and the accuracy of the cosine similarity calculation can be improved.

[0078] In the above-mentioned embodiment of the present application, the search term is input into the first model in the dual-tower feature extraction model to obtain the Query Embedding, that is, the vector representation corresponding to the search term; and the corresponding target result text is input into the second model in the dual-tower feature extraction model to obtain the Title Embedding, that is, the vector representation corresponding to the target result text. The cosine similarity between the two is calculated, and then it is determined whether the convergence condition is currently met based on the obtained cosine similarity. When the convergence condition is met, the initial dual-tower feature extraction model is determined as the target dual-tower feature extraction model. It can be understood that the two sub-models in the dual-tower feature extraction model share parameters during the training process, and the model structures of the two are the same. In the actual feature extraction process, the trained first model for generating the Query Embedding can be used as the model used in the actual feature extraction process.

[0079] Through the above-mentioned implementation mode of the present application, the initial dual-tower feature extraction model is trained using a sample set until the convergence condition is met, wherein the sample set includes search index text and search result text generated based on a network search log set, and the search result text is the text hit in the text set searched by the search index text; in each training process, the search index text in the sample set is input into the first feature extraction model in the initial dual-tower feature extraction model to obtain the search information feature, and the search result text in the sample set is input into the second feature extraction model in the initial dual-tower feature extraction model to obtain the target information feature; when the feature similarity between the search information feature and the target information feature does not meet the convergence condition, the model parameters of the initial dual-tower feature extraction model are adjusted; when the feature similarity between the search information feature and the target information feature meets the convergence condition, the initial dual-tower feature extraction model is determined as the target dual-tower feature extraction model, thereby introducing a large amount of prior context pre-training knowledge, so that the distribution of the vector output by the model is no longer sparse, and the accuracy and confidence of the feature representation are also greatly improved.

[0080] As an optional implementation manner, before obtaining the current object event description information from the current event resource sub-cluster, the following steps may also be included:

[0081] S1, searching for a reference event entity that matches the incremental event entity contained in the incremental event description information in the entity knowledge graph, wherein the entity knowledge graph is used to indicate the relationship between multiple event entities;

[0082] S2, using the reference event entity to replace the incremental event entity, and obtaining the incremental event description information after the entity replacement.

[0083] It should be noted that in this embodiment, the entity knowledge graph is used to indicate the relationship between multiple event entities. The signal of the entity knowledge graph can be used to enhance the indication of event description information. By using the upper and lower information of entities in the knowledge graph and the relationship between entities, we can normalize the description of events and thereby improve the accuracy of the aggregation operation.

[0084] Specifically, the above knowledge graph may at least include person relationships between person entities; hierarchical relationships between geographical entities; hierarchical relationships between object entities; and similarity relationships between action entities.

[0085] The following describes a comparison method based on person entity relationships. Assume that the knowledge graph records that Xiao Wang and Xiao Li are husband and wife. Event description information A includes the following: Xiao Li is suspected of using a second account to establish a persona; event description information B includes the following: Xiao Wang's wife is suspected of using the second account to establish a persona. If these two events are directly compared, without the person relationship "Xiao Wang's wife is Xiao Li," it is difficult to determine that event description information A and event description information B are the same event. Therefore, during the comparison process, event description information A can be rewritten as "Xiao Wang's wife is suspected of using a second account to establish a persona" and then compared with event description information B "Xiao Wang's wife is suspected of using the second account to establish a persona." This allows events A and B to be accurately determined to be the same event, achieving precise clustering.

[0086] The following is an explanation of the comparison method based on geographic entity relationships. Suppose the knowledge graph records that City C is a prefecture-level city in Province B. The content included in the event description information D is: A traffic accident occurred in City C; the content included in the event description information E is: A traffic accident in Province B caused multiple injuries. If the two events are directly compared, due to the lack of the geographical hierarchical relationship of "City C is a prefecture-level city in Province B", the above event description information D and event description information E are difficult to be identified as the same event. Therefore, during the comparison process, the event description information D can be rewritten as "A traffic accident occurred in City C, Province B", and then compared with E "A traffic accident in Province B caused multiple injuries". Then, events D and E can be accurately determined to be the same event, achieving precise clustering.

[0087] Through the above implementation of the present application, before performing description information comparison, a reference event entity that matches the incremental event entity contained in the incremental event description information is found in the entity knowledge graph; the incremental event entity is replaced with the reference event entity to obtain the incremental event description information after the entity replacement, and the signal of the knowledge graph is introduced for knowledge enhancement. The upper and lower information of the entities in the knowledge graph and the relationship between the entities are used to correct the comparison process, thereby improving the accuracy of the comparison results and thus improving the clustering effect. As an optional implementation, after determining the currently released incremental event resources from the event resource stream, the above also includes:

[0088] S1, extracting event elements from incremental event resources, wherein event elements include: event action elements and event entity elements;

[0089] S2, splicing the event elements to obtain multiple candidate event description information;

[0090] S3, obtaining the splicing confidence corresponding to each of the multiple candidate event description information;

[0091] S4, determining event description information from multiple candidate event description information according to the splicing confidence.

[0092] In the above-mentioned embodiment of the present application, after determining the currently released incremental event resource from the event resource stream, the event elements in each incremental event resource can be extracted by the above-mentioned method, and multiple event description information corresponding to the incremental event resource can be determined based on the event elements.

[0093] Specifically, the above-mentioned event elements may include event action elements and event entity elements, the above-mentioned event action elements may be participles whose part of speech is a verb, and the above-mentioned entity elements may be participles whose part of speech is a noun. By extracting the event action elements and event entity elements respectively, the subject, object and trigger word of an event may be further obtained. At the same time, according to the definition of an event, the three elements are the basic features of an event. Therefore, these three elements may be spliced together, i.e. (subject + trigger word + object), using a subject-verb-object order similar to that in Chinese syntax, to form a complete event title, i.e., event description information. Since the extraction results of the model are not necessarily completely valid, in this embodiment, by fine-tuning the general 4-layer TinyBert semantic model, the event title is subjected to event classification, i.e., when multiple event description information corresponding to the same incremental event resource is obtained, the confidence score of each event description information is obtained. And the event description information with the highest score is used as the final event description information of the incremental event resource.

[0094] By performing event classification on event titles, the following effects can be achieved: event titles that are too poorly expressed can be filtered out by scores; subsequent event title screening is convenient. Specifically, when multiple target event resource subclusters are obtained, the event description information of each subcluster can be determined and output based on the scores of the object event description information in each target event resource subcluster.

[0095] The scoring method in the above TinyBert semantic model includes: the completeness of the subject, predicate, and object structure of an event is positively correlated with the final score, and the higher the completeness, the higher the score; the fluency of the expression of an event is positively correlated with the final score, and the more fluent the expression, the higher the score of the corresponding event description information; the amount of information of an event is positively correlated with the final score, including the more information that can satisfy users to obtain effective information, the higher the score of the corresponding event description information.

[0096] In the aforementioned implementation of the present application, the aforementioned method is used to obtain event description information corresponding to incremental event resources. This event description information is then compared with multiple event resource subclusters to achieve clustering, and the number of event description information contained in the cluster is determined. If the number exceeds a given threshold, a hot event is output, and the title of the hot event is the event description information with the highest event score in the cluster.

[0097] Through the above-mentioned implementation mode of the present application, event elements in incremental event resources are extracted, wherein the event elements include: event action elements, event entity elements; the event elements are spliced to obtain multiple candidate event description information; the splicing confidence corresponding to each of the multiple candidate event description information is obtained; the event description information is determined from the multiple candidate event description information according to the splicing confidence, so as to extract the event elements of the incremental event separately, and determine the description information with the highest splicing confidence from the multiple event description information obtained by splicing as the event description information of the event, thereby improving the accuracy of extracting event description information from event resources.

[0098] As an optional implementation, the above-mentioned extraction of event elements from the incremental event resource includes:

[0099] S1: Extract the character features of the text content in the incremental event resources in the first annotation sub-model of the sequence annotation model, and extract the part-of-speech features and attention weights of each word in the text content of the incremental event resources in the second annotation sub-model of the sequence annotation model;

[0100] S2: Based on the character features of the text content in the incremental event resource, the part-of-speech features of each segmentation, and the attention weight of the segmentation, determine the first confidence level that the i-th segmentation in the incremental event resource is an event action element, where i is a natural number greater than or equal to 1 and less than or equal to N, and N is the number of segmentations contained in the incremental event resource;

[0101] S3, constructs a query vector using the event action element and the text content in the incremental event resource;

[0102] S4, inputting the query vector into the entity extraction network to obtain a second confidence corresponding to each character in the text content in the incremental event resource, wherein the second confidence is used to indicate the probability value of the character being an event entity element associated with the event action element in the query vector.

[0103] In the above-mentioned method of the present application, a sequence labeling model is used to extract event action elements, and an entity extraction network is used to extract event entity elements.

[0104] As an optional implementation, the above-mentioned determination of the first confidence that the i-th segmentation in the incremental event resource is an event action element based on the character features of the text content in the incremental event resource, the part-of-speech features of each segmentation, and the attention weight of the segmentation includes:

[0105] The character features of the character corresponding to the i-th participle are fused with the attention weight associated with the i-th participle to obtain the grammatical features of the i-th participle;

[0106] The grammatical features of the i-th segmentation, the part-of-speech features of the i-th segmentation, and the character features of the character corresponding to the i-th segmentation are input into the random field network to obtain the first confidence that the i-th segmentation is an event action element.

[0107] The above method is described in detail below. Figure 5 The method of extracting event action elements using the above sequence labeling model is explained.

[0108] exist Figure 5 In the sequence tagging model shown in the figure, the word granularity signal is passed through Albert-Layer to obtain the semantic granularity feature signal hidden-states, that is, Figure 5 The char-representation in [Xiaoming created the theme song for the movie] is then flattened by character. For example, the sentence "Xiaoming created the theme song for the movie" is segmented into [Xiaoming created the theme song for the movie], and the corresponding POS features are [n / p / n / v / q / n]. Flattening [Xiaoming created the theme song for the movie] by the corresponding number of characters yields [nnpnnvvqnnn]. This ensures that the number of POS and characters on the input is consistent.

[0109] Based on this, the computational process of fusing text semantic features with grammatical features allows the text semantic features to carry grammatical features. First, the initially embedded part-of-speech features are used as input to the Graph-Generate-Layer. The input part-of-speech features X are transformed through three independent linear operations to obtain the Q, K, and V matrices:

[0110] Q=W Q X

[0111] K=W K X

[0112] V=W V X

[0113] Then use the Q, K, and V matrices as inputs for the self-attention mechanism. The specific formula is as follows:

[0114]

[0115] The above formula is used to calculate the part-of-speech granularity features (i.e., the above-mentioned part-of-speech features), namely the Pos-Representation in the figure and the attention matrix Attention-Score corresponding to the part-of-speech granularity. This matrix is then considered as the dependency graph W of the text at the grammatical level. This dependency graph W and the obtained semantic feature Char-Representation are used as two inputs to the inductive graph convolutional network layer Graph-Conv-Layer, thereby strengthening the projection of syntactic structure features in the word vector space features to obtain fused features. The specific calculation is as follows:

[0116]

[0117] The above formula is equivalent to fusing and activating semantic features based on grammatical dependencies. Finally, the character features, part-of-speech features, and fused features are combined into a vector dimension to form the final feature. The final event action element is predicted using the Conditional Random Field (CRF) layer.

[0118] Specifically, based on the first confidence level that the i-th word in the incremental event resource is an event action element, the word with the highest first confidence level can be determined as the above-mentioned event action element from multiple word segments in the incremental event resource; in another way, the word with a first confidence level greater than the target threshold can be determined as the above-mentioned event action element.

[0119] The following describes a method for extracting event entity elements using an entity extraction network. The above-mentioned query vector is constructed using event action elements and text content in incremental event resources, including:

[0120] S1, constructing a query statement based on the event-action element, where the query statement is used to request the entity extraction network to output an answer that matches the query statement;

[0121] S2, concatenates the sentence vector corresponding to the query statement, the separator vector, and the text vector corresponding to the text content to obtain the query vector.

[0122] Specifically, if Figure 6As shown, the event action element is used as the question, the original text content as the paragraph, and the arguments (subject and object terms) as the expected answer. Through multiple rounds of questioning, the subject and object terms for each trigger word are obtained. For example, to extract the subject of the event action element "creation," a question sentence is constructed: "What is the subject of the trigger word 'creation'?" This is concatenated with the original sentence using the special symbol [SEP] to obtain a Q&A pair input, such as: "What is the subject of the trigger word 'creation'? [SEP] Xiao Ming composed the theme song for the movie."

[0123] Optionally, the query vector is input into the entity extraction network to obtain the second confidence corresponding to each character in the text content in the incremental event resource:

[0124] Obtaining a first answer confidence and a second answer confidence corresponding to each character, wherein the first answer confidence is used to indicate a probability value that the character matched therewith is the first character of the entity associated with the query vector, and the second answer confidence is used to indicate a probability value that the character matched therewith is the last character of the answer entity associated with the query vector;

[0125] Repeat the following operations until all characters in the Chinese text content of the incremental event resource are traversed, and a first sequence of first answer confidences corresponding to each character and a second sequence of second answer confidences corresponding to each character are generated:

[0126] Get the reference character from each character as the current character;

[0127] Get the first answer confidence of the current character and the sequence of the second answer confidence;

[0128] Multiplying the first answer confidence by each second answer confidence in the second answer confidence sequence to obtain a product confidence sequence corresponding to each second answer confidence;

[0129] When traversing each character in the text content in the incremental event resource and generating multiple product confidence sequences, obtaining the highest product confidence in the multiple product confidence sequences as the target product confidence;

[0130] The target characters corresponding to the target product confidence and the characters between the target characters are combined to obtain the event entity element.

[0131] Specifically, the input text obtained in the above steps, "What is the subject of the trigger word "creation"? [SEP] Xiao Ming created the theme song for the movie", is used to obtain the hidden-state of the word granularity through Bert, and then input into the Binary-Pointer layer. This layer will calculate a probability value for each character that is the beginning and end of the Answer segment. Through two layers of loop traversal, the starting and ending positions corresponding to the maximum value of the product of the head and tail probabilities are obtained to determine the Answer extraction result. The model structure and working diagram are shown in Figure 6. Assuming that the product of the starting probability of "Xiao" and the ending probability of "Ming" is the largest, the subject word of the event action element "creation" is determined to be "Xiao Ming".

[0132] Through the above-mentioned implementation mode of the present application, the character features of the text content in the incremental event resource are extracted in the first annotation sub-model in the sequence annotation model, and the part-of-speech features of each segmentation in the text content in the incremental event resource and the attention weight of the segmentation are extracted in the second annotation sub-model in the sequence annotation model; based on the character features of the text content in the incremental event resource, the part-of-speech features of each segmentation and the attention weight of the segmentation, the first confidence that the i-th segmentation in the incremental event resource is an event action element is determined, where i is a natural number greater than or equal to 1 and less than or equal to N, and N The method is as follows: the number of word segments contained in the incremental event resources; the query vector is constructed using the event action elements and the text content in the incremental event resources; the query vector is input into the entity extraction network to obtain the second confidence corresponding to each character in the text content in the incremental event resources; the event extraction process is then decomposed into multiple extraction steps such as event action element recognition and event entity recognition, and machine reading comprehension technology is introduced to model the problem of pairing event entity elements and event action elements as a reading comprehension problem with event action elements as queries and semantic roles as answers, which effectively solves the technical problem of poor event extraction accuracy in the process of existing models handling multi-event extraction.

[0133] As an optional implementation manner, in the case where the target object event corresponding to the target event resource subcluster is determined to be a hot event, generating the hot event description information of the hot event based on the object event description information contained in the target event resource subcluster includes at least one of the following:

[0134] S1, when the quantity characteristic of the event resources included in the target event resource subcluster meets the quantity condition of the hot event, the target object event corresponding to the target event resource subcluster is determined as the hot event;

[0135] S2: When the growth rate characteristics of the newly added event resources in the target event resource sub-cluster meet the growth rate conditions of hot events, the target object event corresponding to the target event resource sub-cluster is determined as a hot event.

[0136] In this embodiment, there are at least two ways to determine hot events from the object events corresponding to multiple resource subclusters. In the first way, the hot events can be determined based on whether the quantity characteristics of the event resources included in the target event resource subcluster meet the quantity conditions of the hot events. For example, when the number of event resources included in the target event resource subcluster is greater than a certain threshold, the object event corresponding to the target event resource subcluster is determined as a hot event. In the second way, the hot events can be determined based on the growth rate characteristics of the newly added event resources in the target event resource subcluster. For example, when the number of newly added event resources per unit time in the target event resource subcluster is greater than a certain threshold, the object event corresponding to the target event resource subcluster is determined as a hot event. In another way, the hot events can also be determined by combining the above two ways. For example, the quantity characteristics and growth rate characteristics of the event resources in each resource subcluster can be respectively assigned quantity points and growth rate points, and the hot events can be determined based on the weighted average of the quantity points and growth rate points corresponding to the event resources in each resource subcluster.

[0137] As an optional implementation manner, when the quantity characteristic of the event resources included in the target event resource subcluster meets the quantity condition of the hot events, determining the target object event corresponding to the target event resource subcluster as a hot event includes at least one of the following:

[0138] S1, when the number of event resources included in the target event resource subcluster is greater than or equal to a resource quantity threshold, determining the target object event corresponding to the target event resource subcluster as a hot event;

[0139] S2: When the number of event resources included in the target event resource subcluster is greater than the number of event resources included in any other event resource subcluster in the event resource cluster, the target object event corresponding to the target event resource subcluster is determined as a hot event.

[0140] It can be understood that in this embodiment, the method of determining hot events based on the quantity characteristics of object events in the resource subcluster can be: Method 1, setting a quantity threshold, when the number of object events in the current resource subcluster is greater than the quantity threshold, the target object event corresponding to the target event resource subcluster is determined as a hot event; Method 2, that is, setting the number of hot events output in each output cycle, and determining the time resource subcluster whose number of event resources is greater than the number of event resources included in any other event resource subcluster in the event resource cluster as the target subcluster, and determining the object event corresponding to the target subcluster as the hot event; Method 3, it can also be combined with the quantity threshold and the output quantity limit in the output cycle to determine the subcluster whose number of object events is greater than the threshold and greater than all other subclusters as the target subcluster.

[0141] As an optional manner, when the growth rate characteristics of the newly added event resources in the target event resource subcluster meet the growth rate conditions of the hot event, determining the target object event corresponding to the target event resource subcluster as a hot event includes at least one of the following:

[0142] S1, when the rate of newly added event resources of the target event resource sub-cluster is greater than or equal to the rate threshold, the target object event corresponding to the target event resource sub-cluster is determined as a hot event;

[0143] S2: When the event resource addition rate of the target event resource subcluster is greater than or equal to the event resource addition rate of any other event resource subcluster in the event resource cluster, the target object event corresponding to the target event resource subcluster is determined as a hot event.

[0144] It can be understood that in this embodiment, the method of determining hot events based on the growth rate characteristics of object events in the resource sub-cluster can be: Method 1, setting a growth rate threshold, when the growth rate of object events in the current resource sub-cluster is greater than the growth rate threshold, the target object event corresponding to the target event resource sub-cluster is determined as a hot event; Method 2, that is, setting the number of hot events output in each output cycle, and determining the time resource sub-cluster whose event resource growth rate is greater than the growth rate of event resources included in any other event resource sub-cluster in the event resource cluster as the target sub-cluster, and determining the object event corresponding to the target sub-cluster as the hot event; Method 3, it can also be combined with the growth rate threshold and the output number limit in the output cycle to determine the sub-cluster whose object event growth rate is greater than the threshold and greater than all other sub-clusters as the target sub-cluster.

[0145] Through the above-mentioned implementation mode of the present application, when the target object event corresponding to the target event resource subcluster is determined to be a hot event, the hot event description information of the hot event is generated based on the object event description information contained in the target event resource subcluster, including at least one of the following: when the quantity characteristics of the event resources included in the target event resource subcluster meet the quantity conditions of the hot event, the target object event corresponding to the target event resource subcluster is determined to be a hot event; when the growth rate characteristics of the newly added event resources in the target event resource subcluster meet the growth rate conditions of the hot event, the target object event corresponding to the target event resource subcluster is determined to be a hot event, thereby accurately determining the hot event to be output from multiple data dimensions, thereby improving the accuracy of the hot event output.

[0146] As an optional implementation, after comparing the incremental event description information corresponding to the incremental event resource with the object event description information included in the event resource cluster, the method further includes: repeatedly performing the following steps until the isolated event description information included in the isolated event resource cluster is traversed:

[0147] S1, obtaining current isolated event description information from an isolated event resource cluster, wherein the isolated event resource cluster includes event description information of isolated event resources that do not meet clustering conditions;

[0148] S2, comparing the incremental event description information with the current isolated event description information;

[0149] S3, when the comparison result indicates that the incremental event description information and the current isolated event description information meet the merging condition, merging the incremental event resource with the isolated event resource corresponding to the current isolated event description information to obtain an incremental event cluster;

[0150] S4, adding the incremental event cluster as an event resource subcluster to the event resource cluster;

[0151] S5. When the comparison result indicates that the incremental event description information and the current isolated event description information do not meet the merging condition, the incremental event resource is added to the isolated event resource cluster.

[0152] As an optional implementation, after adding the incremental event resource to the isolated event resource cluster, the above further includes:

[0153] S1, traverse the isolated event resource cluster and obtain any two isolated event description information as the isolated event description information pair to be compared;

[0154] S2, when the isolated event description information pairs meet the merging conditions, merging the isolated event resources corresponding to the isolated event description information pairs to obtain an event cluster;

[0155] S3: Add the event cluster as an event resource subcluster to the event resource cluster.

[0156] It can be understood that in this embodiment, after obtaining the event description information corresponding to the incremental event resource and comparing the event description information with the event resource sub-cluster, if no event resource sub-cluster that can be merged is found, the incremental event resource and the corresponding event description information are added to the isolated event resource cluster.

[0157] like Figure 7 As shown, the similarity between a new event description information and a cluster is determined by traversing all the event description information in the cluster. If more than half of the event description information is similar to the current new event title, then the cluster and the new event description information are considered similar and can be merged. If the new event description information cannot find a suitable cluster to merge with, it will enter the isolated resource pool (the so-called isolated resource pool is the collection of all event titles that cannot be merged) for merging. That is, traverse the event description information of all isolated resource pools. If there is an isolated event title that can be merged, then this event title and the new event title can be merged to form a cluster, and the cluster is transferred to the event resource cluster set. If there is no isolated event title that can be merged, then the new event title will be left in the isolated resource pool as a new isolated resource.

[0158] In this embodiment, resources in the isolated event resource cluster will be retrieved regularly, and the retrieved new event clusters will be put back into the resource cluster for aggregation. The reason is that during the entire online clustering process, all cluster information is dynamically changing, and the event titles contained in the same cluster may not be exactly the same at different times. This results in some event titles that are close to the similarity threshold being mistakenly filtered by the system in the early stage and thus cannot be merged into the cluster. For example:

[0159] Event A includes: a=1, and b=2;

[0160] Event B includes: a=1;

[0161] Event C includes: b=2;

[0162] In the process of judging the description information of different events based on the similarity of text features, event A can be merged with event B, and event A can also be merged with event C, but event B and event C cannot be merged. Assuming that the order of entering the entire system is C→B→A, then first of all, event C has neither a cluster that can be merged nor an isolated resource that can be merged, so event C becomes an isolated resource;

[0163] Then, event B is input. Since event B and event C cannot be merged, that is, no cluster can be generated, there is no cluster to merge with, and no isolated resource to merge with, so event B also becomes an isolated resource.

[0164] Finally, event A enters the system. Since there's no suitable cluster for A to merge into, it enters the isolated resource pool. When traversing all events in the isolated resource pool, suppose A encounters B first. A and B merge, forming cluster (A, B), and enter the resource cluster set. Event description C is left behind and cannot be merged into cluster (A, B). The same situation occurs in another traversal method, where A encounters C first.

[0165] Based on the above reasons, in the above implementation of the present application, isolated resources in the isolated event resource cluster are periodically obtained and re-entered into the system for clustering operations, thereby solving the problem of "resource omission" caused by the characteristic that each event in streaming clustering is compared only once, thereby improving the accuracy of event clustering.

[0166] As an optional method, after the above implementation of this application, expired resources will be eliminated and decommissioned in the following way: the event resource subclusters contained in the event resource cluster are screened according to the timeliness conditions, and the event resource subclusters that do not meet the timeliness conditions are deleted from the event resource cluster.

[0167] Specifically, for example:

[0168] Event D includes: May 5, a=1;

[0169] Event E includes: July 5, a=1.

[0170] Because event D and event E occurred at different times, they are not the same event and should not be clustered. However, because they both contain the similar text "a=1," they are still likely to be clustered when comparing and clustering using text features. This affects the accuracy of the clustering operation and the efficiency of event output.

[0171] To avoid this, the above-mentioned embodiment of the present application also performs timeliness filtering on event resources. This means that the timeliness of resources within the system is regularly checked, and expired resources are filtered out and eliminated. This reduces the storage pressure on the system, improves clustering accuracy, and also improves the timeliness of event description information.

[0172] The following combination Figure 8 Describe the complete process of this application.

[0173] In this example, a complete event mining system based on streaming resources is constructed based on search scenarios. By combining event extraction and online aggregation, events contained in resources are mined. The specific process can be divided into five steps: filtering and classifying resources, extracting event elements from resources, streaming clustering of incremental resources, periodic recovery of existing resources, and elimination of expired resources.

[0174] S1, filtering and classification of resources;

[0175] like Figure 8 As shown in the figure, due to the uneven quality of resource content in the search scenario, it is necessary to filter the acquired incremental articles to select high-quality resources for merging. In this step, resource quality screening is mainly carried out from two dimensions: account and content.

[0176] S2, extraction of event elements of resources;

[0177] The above method may further include steps S21, trigger word identification; S2, semantic role identification; and S23, event classification.

[0178] Among them, step S21, trigger word recognition can be achieved through a sequence tagging model. As shown in Figure 5, the word granularity signal is passed through the Albert-Layer to obtain the semantic granularity feature signal hidden-states, that is, Figure 5 The char-representation in [I Love China] is then flattened by character. For example, the sentence "I love China" is segmented into [I love China], and the corresponding POS feature is [nvn]. [I love China] is flattened by the corresponding number of characters to obtain [nvnn]. This ensures that the number of POS and characters on the input is consistent.

[0179] Based on this, the computational process of fusing text semantic features with grammatical features allows the text semantic features to carry grammatical features. First, the initially embedded part-of-speech features are used as input to the Graph-Generate-Layer. The input part-of-speech features X are transformed through three independent linear operations to obtain the Q, K, and V matrices:

[0180] Q=W Q X

[0181] K=W K X

[0182] V=W V X

[0183] Then use the Q, K, and V matrices as inputs for the self-attention mechanism. The specific formula is as follows:

[0184]

[0185] The词性granularity features (i.e., the above-mentioned词性features) are calculated according to the above formula, that is, Pos-Representation in the figure and the attention matrix Attention-Score corresponding to the词性granularity. Then, this matrix is regarded as the dependency graph W of the text at the syntactic level, and this dependency graph W and the obtained semantic feature Char-Representation are used as two inputs to the inductive graph convolutional network layer Graph-Conv-Layer to strengthen the projection of syntactic structure features in the word vector space feature to obtain a fused feature. The specific calculation is as follows:

[0186]

[0187] The above formula is equivalent to performing an information fusion and activation on the semantic features according to the syntactic dependencies. Finally, the character features,词性features, and fused features are jointly combined in terms of vector dimensions as the final features. The final trigger word result is predicted through the conditional random field CRF-Layer layer.

[0188] Step S22, semantic role recognition can be achieved through a question-answering model. In the case where multiple trigger words are recognized in step S22, the corresponding semantic roles can be recognized according to each trigger word. For example, when the trigger word "receive" is received, questions such as "Who is the subject of receiving?" and "What is the object of receiving?" can be constructed and concatenated with the original text as inputs to determine the entity words corresponding to different trigger words.

[0189] As Figure 6 shown, if the subject of the trigger word "create" is to be extracted, the Question sentence "What is the subject of the trigger word 'create'?" is constructed and concatenated with the original sentence through the special symbol [SEP] as the QA question-answering pair input, in the form of: "What is the subject of the trigger word 'create'? [SEP] Xiaoming created the theme song for the movie". The input end first obtains the word-level hidden-state through Bert, and then inputs it to the Binary-Pointer layer. This layer calculates a probability value for each word to be the start and end of the Answer segment. By traversing through two layers of loops, the start and end positions corresponding to the maximum product of the start and end probabilities are obtained to determine the Answer extraction result. The model structure and working schematic are as Figure 6 shown. The product of the start probability of "small" and the end probability of "ming" is the largest, so the subject word of the trigger word "create" is "Xiaoming".

[0190] As shown in step S23, the event is classified. Specifically, through the above two steps, the three elements of the subject, object and trigger word of an event can be obtained. At the same time, according to the definition of an event, the three elements are the basic features of an event. Therefore, we can directly splice these three elements together, that is, (subject + trigger word + object), similar to the subject, predicate and object in Chinese syntax, to form a complete event title. Since the extraction results of the model are not necessarily 100% valid, we also fine-tuned the general 4-layer TinyBert semantic model to perform event classification on the event title. Purpose 1: For event titles that are too poorly expressed, filter and eliminate them. Purpose 2: Facilitate subsequent event title screening. Scoring principle: The score is generated according to the model. The more complete the subject, predicate and object structure of an event, the higher the score; the more fluent the expression of an event, the higher the score; the higher the amount of information an event has, the more it can satisfy users to obtain effective information, and the higher the score.

[0191] S3, streaming clustering of incremental resources;

[0192] In traditional clustering methods, articles must first be collected into a collection. Clustering can only be performed after the collection reaches a certain number of articles. In this case, the candidate data for clustering is a fixed set of articles. Articles are clustered repeatedly within this set until the average distance within a cluster is minimized and the average distance between clusters is maximized. After the clustering results are output, the original collection of articles must be cleared (otherwise, the collection will contain an increasing number of outdated articles) and a new round of collection and clustering must be started.

[0193] In this embodiment, unlike traditional offline clustering, streaming clustering is an online clustering method that does not provide a complete data set. Instead, it flexibly passes incremental data into the data set in sequence, and each data point is clustered only once. This greatly ensures the timeliness of clustering without wasting a lot of time to enrich the data required for offline clustering.

[0194] The above step S3 may specifically include: S31, obtaining the title of the current incremental event obtained in S2; S32, performing a clustering operation using the title of the current incremental event and the event cluster; S33, determining whether to output the title of the hot event based on the size of the event cluster; S34, determining whether to put the title of the current incremental event into the isolated resource pool based on the clustering situation;

[0195] Specifically, the above step S31 may include: obtaining the event title expression of the high-quality resource through the step of "extracting the event element of the resource". At this time, the event can be passed to the resource cluster aggregation module to merge the resource clusters.

[0196] The above step S32 may include: in order to improve the accuracy of merging, the signal of the knowledge graph is introduced for knowledge enhancement. By using the upper and lower information of the entities in the knowledge graph and the relationship between the entities, we can normalize the description of the event and perform similarity calculation based on the normalized title.

[0197] In the process of similarity calculation, the trained deep learning model Twin-Tower BERT is used to calculate the vector representation of the event title, and then the similarity is calculated using the vector representation of the title in the cluster and the vector representation of the current title. In the training process of the deep learning model Twin-Tower BERT, a large amount of prior contextual pre-training knowledge is introduced. The distribution of vectors is no longer sparse, and the accuracy and confidence of the representation are greatly improved. The schematic diagram of the above-mentioned deep learning model Twin-Tower BERT is as follows Figure 4 As shown in Figure 2, the data for model training is obtained from the posterior click signal of the search log. For example, if a user searches for a query and clicks on an article in the result page, it can be considered that the article is relevant to the query.

[0198] Then, the vectorized expression of the event title can be used to determine the similarity between two event titles in the same vector space. If the similarity exceeds a given threshold, it is determined that the two event titles can be merged.

[0199] As for the similarity judgment between a new event title and a cluster, the method is to traverse all event titles in the cluster. If more than half of the event titles are similar to the new event title, then the cluster and the new event title are considered similar and can be merged.

[0200] The above step S33 may include: if the new event title can find a suitable cluster to merge, then it is necessary to determine the number of event titles contained in the cluster. If the number exceeds a given threshold, the hot event will be output externally, and the event title will be the title with the highest event score in the cluster.

[0201] The above step S34 may include: if the new event title cannot find a suitable cluster to merge, then enter the isolated resource pool (the so-called isolated resource pool is the collection of all event titles that cannot be merged) for merging, that is, traverse the event titles of all isolated resource pools. If there is an isolated event title that can be merged, then this event title and the new event title can be merged to form a cluster, and the cluster can be transferred to the resource cluster set; if there is no isolated event title that can be merged, then the new event title will be left in the isolated resource pool as a new isolated resource.

[0202] S4, periodic recovery of stock resources;

[0203] In this embodiment, isolated resources in the isolated event resource cluster are periodically obtained and re-entered into the system for clustering operations, thereby solving the problem of "resource omission" caused by the characteristic that each event is only compared once in streaming clustering, thereby improving the accuracy of event clustering.

[0204] S5, the elimination of expired resources.

[0205] In the above embodiment of the present application, event resources are also filtered for timeliness, that is, the timeliness of resources in the system is regularly checked, and expired resources are filtered and eliminated. This reduces the storage pressure of the system, improves the clustering accuracy, and also improves the timeliness of the event description information.

[0206] Through the above-mentioned implementation mode of the present application, the title extraction of event resources and the clustering and output of hot events are realized through the five steps of filtering and classifying resources, extracting event elements of resources, streaming clustering of incremental resources, periodic recovery of existing resources, and elimination and exit of expired resources, thereby achieving the technical effect of accurately outputting event titles and accurately and efficiently clustering events.

[0207] It should be noted that for the aforementioned method embodiments, for simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should be aware that the present invention is not limited by the order of the actions described, because according to the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0208] According to another aspect of the embodiment of the present invention, there is also provided a device for generating event description information for implementing the above-mentioned method for generating event description information. Figure 9 As shown, the device includes:

[0209] A determining unit 902 is configured to determine the currently published incremental event resource from the event resource stream;

[0210] Comparison unit 904 is configured to compare the incremental event description information corresponding to the incremental event resource with the object event description information contained in the event resource cluster, wherein the event resource cluster includes at least one event resource subcluster obtained by stream clustering historical event resources, the historical event resource being an event resource published before the incremental event resource in the event resource stream, and the event resource subcluster including at least two object event description information associated with the same object event;

[0211] A clustering unit 906 is configured to cluster the incremental event resources into the target event resource subcluster and merge the event description information into the target event resource subcluster if the comparison result indicates that the incremental event description information and the object event description information contained in the target event resource subcluster in the event resource cluster meet the clustering condition;

[0212] The generating unit 908 is configured to generate hot event description information of the hot event based on the object event description information included in the target event resource subcluster when the target object event corresponding to the target event resource subcluster is determined to be a hot event.

[0213] Optionally, in this embodiment, the embodiments to be implemented by the above-mentioned various unit modules can refer to the above-mentioned various method embodiments, which will not be repeated here.

[0214] According to another aspect of the embodiment of the present invention, an electronic device for implementing the above-mentioned method for generating event description information is also provided. The electronic device may be Figure 10 The terminal device or server shown in FIG. This embodiment is described by taking the electronic device as a terminal device as an example. Figure 10 As shown, the electronic device includes a memory 1002 and a processor 1004. The memory 1002 stores a computer program, and the processor 1004 is configured to execute the steps in any of the above method embodiments through the computer program.

[0215] Optionally, in this embodiment, the electronic device may be located in at least one network device among a plurality of network devices of a computer network.

[0216] Optionally, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0217] S1, determine the currently published incremental event resource from the event resource stream;

[0218] S2, comparing the incremental event description information corresponding to the incremental event resource with the object event description information contained in the event resource cluster, wherein the event resource cluster includes at least one event resource subcluster obtained by streaming clustering of historical event resources, the historical event resource being an event resource published before the incremental event resource in the event resource stream, and the event resource subcluster including at least two object event description information associated with the same object event;

[0219] S3, if the comparison result indicates that the incremental event description information and the object event description information contained in the target event resource subcluster in the event resource cluster meet the clustering condition, clustering the incremental event resource into the target event resource subcluster and merging the event description information into the target event resource subcluster;

[0220] S4. When the target object event corresponding to the target event resource subcluster is determined to be a hot event, generate hot event description information of the hot event based on the object event description information included in the target event resource subcluster.

[0221] Alternatively, those skilled in the art will appreciate that Figure 10 The structure shown is for illustration only, and the electronic device may also be a vehicle-mounted terminal, a smart phone (such as an Android phone, an iOS phone, etc.), a tablet computer, a PDA, a mobile Internet device (MID), a PAD, or other terminal devices. Figure 10 It does not limit the structure of the above electronic device. For example, the electronic device may also include Figure 10 More or fewer components (such as network interfaces, etc.) as shown in, or with Figure 10 Different configurations shown.

[0222] Among them, the memory 1002 can be used to store software programs and modules, such as the program instructions / modules corresponding to the method and device for generating event description information in the embodiment of the present invention. The processor 1004 executes various functional applications and data processing by running the software programs and modules stored in the memory 1002, that is, realizes the above-mentioned method for generating event description information. The memory 1002 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1002 may further include a memory remotely located relative to the processor 1004, and these remote memories may be connected to the terminal via a network. Examples of the above-mentioned networks include but are not limited to the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Among them, the memory 1002 can be used specifically but not limited to store image recognition models for identifying entities, event information corresponding to entity objects, and other information. As an example, if Figure 10 As shown, the memory 1002 may include, but is not limited to, the determining unit 902, the comparing unit 904, the clustering unit 906, and the generating unit 908 in the device for generating the event description information. In addition, it may also include, but is not limited to, other module units in the device for generating the event description information, which will not be described in detail in this example.

[0223] Optionally, the transmission device 1006 is configured to receive or send data via a network. Specific examples of the network may include a wired network and a wireless network. In one embodiment, the transmission device 1006 includes a network interface controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In one embodiment, the transmission device 1006 is a radio frequency (RF) module, which is configured to communicate with the Internet wirelessly.

[0224] In addition, the electronic device further includes: a display 1008 and a connection bus 1010 for connecting various module components in the electronic device.

[0225] In other embodiments, the terminal device or server may be a node in a distributed system, wherein the distributed system may be a blockchain system, and the blockchain system may be a distributed system formed by connecting multiple nodes through network communication. The nodes may form a peer-to-peer (P2P) network, and any computing device, such as a server, terminal, or other electronic device, may become a node in the blockchain system by joining the peer-to-peer network.

[0226] According to one aspect of the present application, a computer program product is provided, comprising a computer program / instructions containing program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication component and / or installed from a removable medium. When the computer program is executed by a central processing unit, the various functions provided in the embodiments of the present application are performed.

[0227] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0228] According to one aspect of the present application, a computer-readable storage medium is provided, and a processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the above-mentioned method for generating event description information.

[0229] Optionally, in this embodiment, the computer-readable storage medium may be configured to store a computer program for performing the following steps:

[0230] S1, determine the currently published incremental event resource from the event resource stream;

[0231] S2, comparing the incremental event description information corresponding to the incremental event resource with the object event description information contained in the event resource cluster, wherein the event resource cluster includes at least one event resource subcluster obtained by streaming clustering of historical event resources, the historical event resource being an event resource published before the incremental event resource in the event resource stream, and the event resource subcluster including at least two object event description information associated with the same object event;

[0232] S3, if the comparison result indicates that the incremental event description information and the object event description information contained in the target event resource subcluster in the event resource cluster meet the clustering condition, clustering the incremental event resource into the target event resource subcluster and merging the event description information into the target event resource subcluster;

[0233] S4. When the target object event corresponding to the target event resource subcluster is determined to be a hot event, generate hot event description information of the hot event based on the object event description information included in the target event resource subcluster.

[0234] Optionally, in this embodiment, a person of ordinary skill in the art may understand that all or part of the steps in the various methods of the above embodiments may be completed by instructing the hardware related to the terminal device through a program, and the program may be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0235] If the integrated units in the above embodiments are implemented in the form of software functional units and sold or used as independent products, they can be stored in the above-mentioned computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or 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 a number of instructions for causing one or more computer devices (such as personal computers, servers, or network devices) to execute all or part of the steps of the above-mentioned methods in various embodiments of the present invention.

[0236] In the above embodiments of the present invention, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0237] In the several embodiments provided in this application, it should be understood that the disclosed client can be implemented in other ways. Among them, the device embodiments described above are merely illustrative. For example, the division of the above-mentioned units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0238] The units described above as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0239] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0240] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for generating event description information, characterized in that: include: Determine the currently published incremental event resource from the event resource stream; Comparing the incremental event description information corresponding to the incremental event resource with the object event description information contained in the event resource cluster, wherein the event resource cluster includes at least one event resource subcluster obtained by streaming clustering of historical event resources, the historical event resources being event resources published before the incremental event resource in the event resource stream, and the event resource subcluster including at least two object event description information associated with the same object event; If the comparison result indicates that the incremental event description information and the object event description information contained in the target event resource subcluster in the event resource cluster meet the clustering condition, clustering the incremental event resources into the target event resource subcluster and merging the event description information into the target event resource subcluster; In a case where the target object event corresponding to the target event resource subcluster is determined to be a hot event, hot event description information of the hot event is generated based on the object event description information included in the target event resource subcluster.

2. The method according to claim 1, characterized in that The comparing the incremental event description information corresponding to the incremental event resource with the object event description information contained in the event resource cluster includes: sequentially acquiring candidate event resource subclusters from the event resource cluster as current event resource subclusters; The incremental event description information is sequentially compared with the object event description information contained in the current event resource subcluster.

3. The method according to claim 2, characterized in that The sequentially comparing the incremental event description information with the object event description information contained in the current event resource subcluster includes: Repeat the following steps until the object event description information included in the current event resource subcluster is traversed: Acquire current object event description information from the current event resource subcluster; Extracting the incremental description features of the incremental event resources and the current description features of the current object event description information in the target dual-tower feature extraction model; When the feature similarity between the incremental description feature and the current description feature is greater than or equal to a first threshold, updating the clustering coefficient of the current event resource subcluster and obtaining the next object event description information; When the feature similarity between the incremental description feature and the current description feature is less than the first threshold, acquiring next object event description information; In a case where the clustering coefficient of the current event resource subcluster is greater than a second threshold, the current event resource subcluster is determined to be the target event resource subcluster that meets the clustering condition.

4. The method according to claim 3, characterized in that Before determining the currently published incremental event resource from the event resource stream, the method further includes: Training the initial dual-tower feature extraction model using a sample set until a convergence condition is met, wherein the sample set includes search index text and search result text generated based on a network search log set, and the search result text is text that is hit in the text set searched by the search index text; In each training process, the search index text in the sample set is input into the first feature extraction model in the initial dual-tower feature extraction model to obtain search information features, and the search result text in the sample set is input into the second feature extraction model in the initial dual-tower feature extraction model to obtain target information features; When the feature similarity between the search information feature and the target information feature does not meet the convergence condition, the model parameters of the initial dual-tower feature extraction model are adjusted; when the feature similarity between the search information feature and the target information feature meets the convergence condition, the initial dual-tower feature extraction model is determined as the target dual-tower feature extraction model.

5. The method according to claim 3, characterized in that Before obtaining the current object event description information from the current event resource sub-cluster, the method further includes: Searching for a reference event entity that matches the incremental event entity included in the incremental event description information in an entity knowledge graph, wherein the entity knowledge graph is used to indicate the relationship between multiple event entities; The reference event entity is used to replace the incremental event entity to obtain incremental event description information after entity replacement.

6. The method according to claim 1, characterized in that After determining the currently published incremental event resource from the event resource stream, the method further includes: Extracting event elements from the incremental event resource, wherein the event elements include: event action elements and event entity elements; splicing the event elements to obtain multiple candidate event description information; Obtaining the splicing confidence corresponding to each of the plurality of candidate event description information; The event description information is determined from the plurality of candidate event description information according to the splicing confidence.

7. The method according to claim 6, characterized in that Extracting the event element from the incremental event resource includes: Extracting character features of the text content in the incremental event resource in the first annotation sub-model in the sequence annotation model, and extracting part-of-speech features of each word in the text content in the incremental event resource and the attention weight of the word in the second annotation sub-model in the sequence annotation model; Determine, based on the character features of the text content in the incremental event resource, the part-of-speech features of each segmentation, and the attention weight of the segmentation, a first confidence that the i-th segmentation in the incremental event resource is an event action element, where i is a natural number greater than or equal to 1 and less than or equal to N, and N is the number of segmentations contained in the incremental event resource; Constructing a query vector using the event action element and text content in the incremental event resource; The query vector is input into an entity extraction network to obtain a second confidence corresponding to each character in the text content in the incremental event resource, wherein the second confidence is used to indicate a probability value of the character being an event entity element associated with the event action element in the query vector.

8. The method according to claim 7, characterized in that The determining, based on the character features of the text content in the incremental event resource, the part-of-speech features of each segmentation, and the attention weight of the segmentation, of a first confidence that the i-th segmentation in the incremental event resource is an event action element includes: Fusing the character features of the characters corresponding to the i-th participle and the attention weight associated with the i-th participle to obtain the grammatical features of the i-th participle; The grammatical features of the i-th segmentation, the part-of-speech features of the i-th segmentation, and the character features of the character corresponding to the i-th segmentation are input into a random field network to obtain the first confidence that the i-th segmentation is an event action element.

9. The method according to claim 1, characterized in that When the target object event corresponding to the target event resource subcluster is determined to be a hot event, generating the hot event description information of the hot event based on the object event description information included in the target event resource subcluster includes at least one of the following: In a case where a quantity characteristic of event resources included in the target event resource subcluster meets a quantity condition of hot events, the target object event corresponding to the target event resource subcluster is determined as a hot event; In the case that the growth rate characteristics of the newly added event resources in the target event resource subcluster meet the growth rate conditions of hot events, the target object event corresponding to the target event resource subcluster is determined as a hot event.

10. The method according to claim 1, characterized in that After comparing the incremental event description information corresponding to the incremental event resource with the object event description information contained in the event resource cluster, the method further includes: Repeat the following steps until all the isolated event description information included in the isolated event resource cluster is traversed: Acquiring current isolated event description information from the isolated event resource cluster, wherein the isolated event resource cluster includes event description information of isolated event resources that do not meet the clustering condition; Comparing the incremental event description information with the current isolated event description information; if the comparison result indicates that the incremental event description information and the current isolated event description information meet a merging condition, merging the incremental event resource with the isolated event resource corresponding to the current isolated event description information to obtain an incremental event cluster; Adding the incremental event cluster as the event resource subcluster to the event resource cluster; If the comparison result indicates that the incremental event description information and the current isolated event description information do not meet the merging condition, the incremental event resource is added to the isolated event resource cluster.

11. The method according to claim 10, characterized in that After adding the incremental event resource to the isolated event resource cluster, the method further includes: Traversing the isolated event resource cluster, obtaining any two isolated event description information as isolated event description information pairs to be compared; When the isolated event description information pairs meet the merging condition, the isolated event resources corresponding to the isolated event description information pairs are merged to obtain an event cluster; and the event cluster is added to the event resource cluster as the event resource subcluster.

12. A device for generating event description information, characterized in that: include: A determination unit, used to determine the currently published incremental event resource from the event resource stream; a comparison unit, configured to compare the incremental event description information corresponding to the incremental event resource with the object event description information contained in the event resource cluster, wherein the event resource cluster includes at least one event resource subcluster obtained by streaming clustering of historical event resources, the historical event resources being event resources published before the incremental event resource in the event resource stream, and the event resource subcluster including at least two object event description information associated with the same object event; a clustering unit configured to cluster the incremental event resources into the target event resource subcluster and merge the event description information into the target event resource subcluster if the comparison result indicates that the incremental event description information and the object event description information contained in the target event resource subcluster in the event resource cluster meet a clustering condition; The generating unit is configured to generate hot event description information of the hot event based on the object event description information included in the target event resource subcluster when the target object event corresponding to the target event resource subcluster is determined to be a hot event.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein the method according to any one of claims 1 to 11 is executed when the program is executed.

14. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.

15. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to execute the method according to any one of claims 1 to 11 through the computer program.

Images