Data processing, method, device, and storage medium for processing and providing entity objects

Abstract

An embodiment of the present application provides a method, device, and storage medium for data processing, processing and providing of entity objects. The method includes: respectively obtaining a source text set and a source picture set corresponding to a target object from multiple data sources; performing text summarization on the source text set to obtain text summary information of the target object; selecting a target picture from the source picture set that semantically matches the text summary information; and generating a description content of the target object according to the text summary information and the target picture. In the embodiment of the present application, source texts related to the target object from multiple data sources can be fused, text summary information corresponding to the target object can be mined, and then combined with source pictures related to the target object from multiple data sources, the description content of the target object can be generated quickly, concisely, and with high quality.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular, to a method, device, and storage medium for data processing, processing of entity objects, and provision thereof. Background Art

[0002] A knowledge graph, known as knowledge domain visualization or knowledge domain mapping map in the library and information science field, is a series of various different graphs that display the development process and structural relationships of knowledge. It uses visualization technology to describe knowledge resources and their carriers, and to mine, analyze, construct, draw, and display knowledge and the interconnections between them.

[0003] Generally, we can refer to a certain distinguishable and independently existing thing involved in a knowledge graph as an entity. For example, an entity can be an animal, a city, a fruit, a point of interest (POI) that people are interested in, and so on.

[0004] Currently, the description information of entities is usually manually compiled by the managers of the entities, which leads to problems such as low creation efficiency and poor quality of the description information of entities. Summary of the Invention

[0005] Multiple aspects of this application provide a method, device, and storage medium for data processing, processing of entity objects, and provision thereof, for quickly and high-quality creating or providing description content of target objects.

[0006] An embodiment of this application provides a data processing method, including:

[0007] Respectively obtaining a source text set and a source picture set corresponding to a target object from multiple data sources;

[0008] Performing text summarization on the source text set to obtain text summary information of the target object;

[0009] Selecting a target picture semantically matching the text summary information from the source picture set;

[0010] Generating description content of the target object according to the text summary information and the target picture.

[0011] An embodiment of this application also provides a data processing method, including:

[0012] Encoding source texts from different data sources using the same encoding parameters to obtain semantic encodings of the respective source texts;

[0013] Respectively obtaining the semantics of the respective source texts from the semantic encodings of the respective source texts;

[0014] Fuse the semantics of each source text according to the influence weight of each source text on the fused semantics to obtain the fused semantics;

[0015] Generate text summary information according to the fused semantics.

[0016] The embodiment of the present application also provides a processing method for an entity object, including:

[0017] Determine a plurality of network data sources associated with the entity object, where the network data sources include web page content and at least one data platform providing services related to the entity object;

[0018] Obtain a text data set and a picture data set of the entity object from the plurality of data sources;

[0019] Obtain the key text data of the entity object by performing data summarization on the text data set;

[0020] Screen key picture data matching the text summary information from the picture data set;

[0021] Generate a description content of the entity object according to the key text data and the key picture data, where the description content includes video data.

[0022] The embodiment of the present application also provides a providing method for an entity object, including:

[0023] Obtain a text data set and a picture data set of the entity object from a web page and a first data platform;

[0024] Obtain a text data set and a picture data set of the entity object from the plurality of data sources;

[0025] Obtain the key text data of the entity object by performing data summarization on the text data set;

[0026] Screen key picture data matching the text summary information from the picture data set;

[0027] Generate a description content of the entity object according to the key text data and the key picture data;

[0028] Based on a search request for an entity object, display the entity object on the client of the second data platform, and provide description content at the associated position of the entity object, where the description content includes video.

[0029] The embodiment of the present application also provides a computing device, including a memory and a processor;

[0030] The memory is used to store one or more computer instructions;

[0031] The processor is coupled to the memory and is configured to execute the one or more computer instructions for:

[0032] Obtain a source text set and a source picture set corresponding to a target object from multiple data sources respectively;

[0033] Perform text summarization on the source text set to obtain text summary information of the target object;

[0034] Select a target picture from the source picture set that semantically matches the text summary information;

[0035] Generate description content of the target object based on the text summary information and the target picture.

[0036] An embodiment of the present application further provides a computing device, including a memory and a processor;

[0037] The memory is used to store one or more computer instructions;

[0038] The processor is coupled to the memory and is configured to execute the one or more computer instructions for:

[0039] Encode source texts from different data sources using the same encoding parameters to obtain semantic encodings of the source texts;

[0040] Respectively obtain the semantics of the source texts from the semantic encodings of the source texts;

[0041] Fuse the semantics of the source texts according to the influence weights of the source texts on the fused semantics to obtain the fused semantics;

[0042] Generate text summary information according to the fused semantics.

[0043] An embodiment of the present application further provides a computing device, including a memory and a processor;

[0044] The memory is used to store one or more computer instructions;

[0045] The processor is coupled to the memory and is configured to execute the one or more computer instructions for:

[0046] Determine multiple network data sources associated with an entity object, where the network data sources include web page content and at least one data platform providing services related to the entity object;

[0047] Obtain a text data set and a picture data set of the entity object from the multiple data sources;

[0048] Obtain key text data of the entity object by performing data summarization on the text data set;

[0049] Screen key picture data matching the text summary information from the picture dataset;

[0050] Generate description content of the entity object according to the key text data and the key picture data, where the description content includes video data.

[0051] An embodiment of the present application further provides a computing device, including a memory and a processor;

[0052] The memory is used to store one or more computer instructions;

[0053] The processor is coupled to the memory and is used to execute the one or more computer instructions for:

[0054] Obtain the text dataset and the picture dataset of the entity object from the web page and the first data platform;

[0055] Obtain the text dataset and the picture dataset of the entity object from the multiple data sources;

[0056] Obtain the key text data of the entity object by performing data summarization on the text dataset;

[0057] Screen key picture data matching the text summary information from the picture dataset;

[0058] Generate description content of the entity object according to the key text data and the key picture data;

[0059] Based on a search request for the entity object, display the entity object on the client of the second data platform, and provide description content at the associated position of the entity object, where the description content includes video.

[0060] An embodiment of the present application further provides a computer-readable storage medium storing computer instructions, which when executed by one or more processors, cause the one or more processors to execute the foregoing data processing method, entity object processing method, or entity object providing method.

[0061] In an embodiment of the present application, the source text set and the source picture set corresponding to the target object can be obtained from multiple data sources respectively; perform text summarization on the source text set to obtain the text summary information of the target object; generate the description content of the target object according to the text summary information and the source picture set. Accordingly, in an embodiment of the present application, the source text related to the target object from multiple data sources can be fused, the text summary information corresponding to the target object can be mined, and then combined with the source pictures related to the target object from multiple data sources, the description content of the target object can be generated or provided quickly, refinedly, and with high quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The drawings described herein are provided to further understand the present application and form a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application. In the drawings:

[0063] Figure 1a FIG. [X] is a schematic flowchart of a data processing method provided for an exemplary embodiment of the present application;

[0064] Figure 1b FIG. [X] is a schematic logical diagram of a data processing method provided for an exemplary embodiment of the present application;

[0065] Figure 2 FIG. [X] is a schematic diagram of a query interface provided for an exemplary embodiment of the present application;

[0066] Figure 3 FIG. [X] is a schematic logical diagram of a text summarization process provided for an exemplary embodiment of the present application;

[0067] Figure 4 FIG. [X] is a schematic diagram of an application scenario provided for an exemplary embodiment of the present application;

[0068] Figure 5 FIG. [X] is a schematic logical diagram of a semantic matching process provided for an exemplary embodiment of the present application;

[0069] Figure 6 FIG. [X] is a schematic flowchart of another data processing method provided for another exemplary embodiment of the present application;

[0070] Figure 7 FIG. [X] is a schematic flowchart of a method for processing an entity object provided for another exemplary embodiment of the present application;

[0071] Figure 8 FIG. [X] is a schematic flowchart of a method for providing an entity object provided for another exemplary embodiment of the present application;

[0072] Figure 9 FIG. [X] is a schematic diagram of the structure of a computing device provided for another exemplary embodiment of the present application;

[0073] Figure 10 FIG. [X] is a schematic diagram of the structure of another computing device provided for another exemplary embodiment of the present application;

[0074] Figure 11 FIG. [X] is a schematic diagram of the structure of yet another computing device provided for another exemplary embodiment of the present application;

[0075] Figure 12 FIG. [X] is a schematic diagram of the structure of yet another computing device provided for another exemplary embodiment of the present application. Detailed implementation manners

[0076] To make the objectives, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be clearly and completely described below in conjunction with the specific embodiments and corresponding drawings of the present application. Apparently, the described embodiments are only a part rather than all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present application.

[0077] Currently, there are problems such as low creation efficiency and poor quality in the description information of entities. In view of this, an embodiment of the present application provides a solution. One of the basic ideas is: the source text set and source picture set corresponding to the target object can be respectively obtained from multiple data sources; a text summary is performed on the source text set to obtain the text summary information of the target object; according to the text summary information and the source picture set, the description content of the target object is generated. Accordingly, in the embodiment of the present application, the source texts related to the target object from multiple data sources can be integrated, the text summary information corresponding to the target object can be mined, and then combined with the source pictures related to the target object from multiple data sources, the description content of the target object can be generated quickly, concisely and with high quality.

[0078] The technical solutions provided by the embodiments of the present application are described in detail below in conjunction with the drawings.

[0079] Figure 1a As shown in the figure, which is a schematic flowchart of a data processing method provided for an exemplary embodiment of the present application. The data processing method provided in this embodiment can be executed by a data processing device, which can be implemented as software or as a combination of software and hardware, and the data processing device can be integrally provided in a computing device. As Figure 1a shown, the data processing method includes:

[0080] Step 100: Respectively obtain the source text set and source picture set corresponding to the target object from multiple data sources;

[0081] Step 101: Perform a text summary on the source text set to obtain the text summary information of the target object;

[0082] Step 102: Select the target picture semantically matching the text summary information from the source picture set;

[0083] Step 103: Generate the description content of the target object according to the text summary information and the target picture.

[0084] The data processing method provided in this embodiment can be applied to various scenarios that require creating descriptive content for an object. This embodiment does not limit the application scenarios. Among them, in different application scenarios, the target object may be ever-changing and diverse. The target object can be a person, a commodity, a geographical location, an event, a relationship, etc. This embodiment also does not limit the type of the target object.

[0085] In addition, the target object can be any entity in the knowledge graph. Of course, this embodiment is not limited to this. The target object can also be a node or an edge in other relational networks, or an independent entity, etc.

[0086] In step 100, the source text set corresponding to the target object can be obtained from multiple data sources.

[0087] Among them, the multiple data sources can be several fixed data sources. Of course, they can also be data sources screened from a vast amount of network resources. Taking the POI (point of interest) in the knowledge graph as an example, the multiple data sources can include tourism platforms, social platforms, news platforms, and encyclopedia platforms, etc. Of course, these are only exemplary.

[0088] In practical applications, techniques such as web crawlers can be used to obtain the source text set corresponding to the target object from multiple data sources. However, this embodiment is not limited to this.

[0089] Among them, the source text can be text related to the target object. In different data sources, the content perspectives of the source text may not be exactly the same. Still taking POI as an example, the content perspectives of the source text include but are not limited to news manuscripts, user evaluations, encyclopedia knowledge, travel guides, introduction manuscripts, etc. The source text may come from UGC (user generated content) or PGC (professionally-generated content).

[0090] Based on the obtained source text set, in step 101, the source text set can be subjected to text summarization to obtain the text summary information of the target object.

[0091] Among them, performing text summarization on the source text set means performing text summarization on each source text in the source text set. The technical solution of text summarization will be described in detail later.

[0092] By performing text summarization on the source text set, the content of the source texts related to the target object in multiple data sources can be fused and refined to obtain refined and high-quality summary sentences.

[0093] In practical applications, according to the output requirements of the description content or according to the knowledge point specifications of the target object, the text summary information may include one or more text summary statements. The text summary information can present information in dimensions such as the highlight knowledge, key attributes, or cultural background of the target object, etc.

[0094] In step 100, a set of source pictures corresponding to the target object can also be obtained from multiple data sources.

[0095] It should be noted that the set of source pictures and the set of source texts can be from an incompletely identical data source set. That is, any data source used in this embodiment can provide only the source text corresponding to the target object or only the source pictures.

[0096] Similarly, the source pictures may come from UGC or may come from PGC. In addition, techniques such as web crawlers can also be used to discover and obtain source pictures from multiple data sources.

[0097] In this embodiment, the source pictures in the set of source pictures are all source pictures related to the target object.

[0098] Based on the text summary information and the set of source pictures obtained through the above processing process, in step 102, the description content of the target object can be generated according to the text summary information and the set of source pictures.

[0099] In some cases, the number of source pictures in the set of source pictures may be relatively small. In this case, step 102 can add the entire set of source pictures to the description content.

[0100] In other cases, the number of source pictures in the set of source pictures is rich. In this case, in step 102, target pictures that are semantically matched with the text summary information can be further selected from the set of source pictures and then added to the description content. Further explanations will be provided later for this type of situation.

[0101] Among them, the description content of the target object can be a smooth video, for example, a 1-minute short video. Of course, it can also be a page combining text and pictures, or a set of photo albums, etc. This embodiment does not limit the presentation form of the description content.

[0102] Figure 1b It is a logical schematic diagram of a data processing method provided by an exemplary embodiment of the present application. As Figure 1b shown, the aforementioned data processing device can be integrated in the Figure 1b server in. Based on this, the data processing method provided in this embodiment can be for query users of the target object ( Figure 1b characterized as the user side in), or for maintainers of the target object ( Figure 1b characterized as the merchant side in).

[0103] In the case of querying a user for a target object, after step 103, a query interface may also be displayed; in response to a query instruction for the target object occurring in the query interface, the description content of the target object is displayed in the query interface.

[0104] Among them, the query interface can be deployed in a dedicated application and interact with the query user in the form of a Q&A interface; it can also be embedded in existing tourism platforms or social platforms, etc., to provide query services for the query user. Figure 2 This is a schematic diagram of a query interface provided by an exemplary embodiment of the present application. Figure 2 The query interface in [reference] adopts the form of a Q&A interface.

[0105] This enables the query user to quickly and comprehensively understand the target object through the query interface, and no longer needs to search everywhere for information related to the target object, which can greatly optimize the query experience of the query user.

[0106] Reference Figure 1b , in the case of the maintainer of the target object, before step 100, a content creation interface may also be displayed; in response to a content creation instruction for the target object occurring in the content creation interface, steps 100 - 103 are executed to generate the description content of the target object; the description content of the target object is displayed in the content creation interface.

[0107] Among them, the content creation interface can be deployed in a dedicated application; it can also be embedded in existing tourism platforms or social platforms, etc., to provide content maintenance services for the maintainer of the target object.

[0108] This enables the maintainer to quickly and concisely create the description content of the target object through the content creation interface. The description content of the target object will be used to display to the query user, so as to better introduce and promote the target object.

[0109] Accordingly, in this embodiment, source texts related to the target object from multiple data sources can be fused, and the text summary information corresponding to the target object can be mined. Then, combined with source pictures related to the target object from multiple data sources, the description content of the target object can be generated quickly, concisely, and of high quality. And it can provide a high-quality service experience for the query user or maintainer of the target object.

[0110] In the above or following embodiments, the source texts from different data sources can be input into a text summary model; in the text summary model, the source texts from different data sources are semantically fused to generate the text summary information of the target object.

[0111] In this embodiment, the source text set corresponding to the target object can be used as the input of the text summary model, and the output of the text summary model will be the text summary information of the target object.

[0112] Among them, the semantic fusion process can be understood as a process of jointly determining the text summary field according to the semantics understood from multiple source texts. Based on the semantic fusion process, the text summary information of the target object can be refined.

[0113] In this embodiment, the text summary model can adopt an Encoder-Decoder architecture. The idea of the Encoder-Decoder architecture is usually to encode the input source sequence into an intermediate vector (this is the encoding process), and this intermediate vector is an encoding of a specific length, and then restore it into an output target sequence through this intermediate vector (this is the decoding process). Adopting the Encoder-Decoder architecture can quickly and accurately implement the summary processing of the text.

[0114] Of course, this embodiment is not limited to this, and the text summary model can also adopt other model architectures that can implement the text summary function.

[0115] In some implementation manners, multiple source texts can be spliced with each other to generate a spliced text, and the spliced text is input into the text summary model. In the text summary model, the spliced text can be encoded to obtain the semantic encoding of the spliced text; the semantics in the spliced text can be obtained from the semantic encoding of the spliced text, and according to the obtained semantics, the text summary information of the target object is generated.

[0116] In this semantic fusion manner, the semantics of multiple source texts can be spliced to generate the text summary information of the target object.

[0117] In other implementation manners, multiple source texts can be synchronously input into the text summary model. In the text summary model, the source texts from different data sources can be encoded using the same encoding parameters to obtain the semantic encodings of the respective source texts; the semantics of the respective source texts are respectively obtained from the semantic encodings of the respective source texts; according to the influence weights of the respective source texts on the fused semantics, the semantics of the respective source texts are fused to obtain the fused semantics; according to the fused semantics, the text summary information of the target object is generated.

[0118] In this implementation manner, during the process of encoding multiple source texts separately, the encoding parameters are shared, which enables the input source texts to be extended, and the number of source texts can be increased or decreased as needed. At the same time, the number of parameters of the text summary model can be saved, and the training time for the encoding process can be saved. The semantic encoding can include the semantics of the source text and context structure information, etc. Among them, sharing encoding parameters means encoding multiple source texts using the same encoding parameters; the extensibility of the source text means that the number of source texts can be increased or decreased as needed.

[0119] In addition, the influence weights on the fused semantics are configured for multiple source texts, and a voting mechanism can be used to jointly determine the fused semantics by multiple source texts. The voting mechanism can be understood as each source text has an influence on the fused semantics, but the degree of influence is not exactly the same. In this semantic fusion method, the information redundancy in the text summary information can be reduced, and the context word order in the text summary information can be effectively guaranteed. The text summary information obtained is both comprehensive and concise.

[0120] The influence weights of multiple source texts on the fused semantics can be customized. For example, a higher influence weight can be set for source texts from a travel platform, while a lower influence weight can be set for source texts from a social platform. For another example, the influence weights of multiple source texts on the fused semantics can be equal, for example, the influence rights of n source texts can all be 1 / n, where n is a positive integer.

[0121] Of course, the influence weights of multiple source texts on the fused semantics can also be determined by the text summary model. This requires providing the text summary model with knowledge and experience for determining the influence weights during the text summary model training process. For example, the influence weights of each source text and the text attribute information that affects the influence weights can be annotated in the training text of the text summary model, so that the text summary model can learn the influence relationship between text attributes and influence weights, so that in the application process of the text summary model, the influence weights corresponding to each of the multiple source texts can be determined based on the text attribute analysis of the input source text. Moreover, in this embodiment, the method of determining the influence weights of multiple source texts on the fused semantics is not limited to this.

[0122] The text summary model outputs fields according to time. The output fields of multiple output times can be combined in time order to form text summary information.

[0123] Taking time t as an example, time t can be any output time of the text summary model. Using the aforementioned voting mechanism, in this embodiment, the semantic fusion process at time t can be:

[0124] According to the semantics of each source text, the probability of each word to be taken as the semantic decoding field at time t is calculated respectively; according to the influence weight of each source text on the fused semantics, for the first word, the probability of taking the first word as the semantic decoding field at time t in each source text is weighted summed to obtain the probability of taking the first word as the model output field at time t, where the first word is any word among the words to be taken.

[0125] Among them, for the first word, an exemplary process of weighted calculation of the probability of taking the first word in each source text as the semantic decoding field at time t can be expressed as: , where p(w) is the probability of word w being the model output field at time t; pn (w) is the probability that the word w is the decoded field at time t under the nth source text; n is the number of source texts; among them, the influence weights of multiple source texts on the fused semantics are all 。

[0126] Accordingly, the probability that each word to be selected is the output field of the model at time t can be determined. This probability is jointly determined by the semantics of multiple source texts, and this probability can be understood as the result of semantic fusion.

[0127] Among them, the word to be selected can be a word in the word list preset for the data processing process of this embodiment, or a word in any source text, which is not limited here.

[0128] On this basis, according to the probability that each word to be selected is the output field of the model at time t, the output field of the text summarization model at time t can be determined. For example, the word to be selected with the highest probability as the output field of the model at time t can be used as the output field of the text summarization model at time t.

[0129] Accordingly, the output field of each output moment of the text summarization model can be determined, thereby generating text summary information.

[0130] Of course, in this embodiment, other implementation manners can also be used to perform semantic fusion on multiple source texts, and this embodiment is not limited thereto.

[0131] In addition, in this embodiment, an attention mechanism can be added to the text summarization model to more accurately understand the semantics of the source text, thereby improving the accuracy of the text summary information.

[0132] In the above or following embodiments, background knowledge text of the target object can also be added during the process of performing text summarization on the source text set to improve the comprehensiveness of the text summary information.

[0133] Among them, the background knowledge text may include key attribute information of the target object. The key attribute information may at least include information that can reflect the highlights or key features of the target object. For different types of target objects, the key attribute information may be different. For example, for a POI, the key attribute information may include locations for filming movies and TV shows, places where celebrities check in, etc. For a product, the key attribute information may include brand, style, sales volume, etc. In practical applications, information that is not desired to be omitted but is desired to be known by the query user can be added to the background knowledge text.

[0134] For example, when the target object is an entity in a knowledge graph, the knowledge graph can be used as the background knowledge text of the target object.

[0135] In this embodiment, the background knowledge text of the target object can be input into the text summarization model. In the text summarization model, semantic fusion is performed on the semantics of the background knowledge text and the semantics of the source text mentioned in the foregoing embodiment to generate the text summary information of the target object. Among them, the semantic recognition process of the background knowledge text can refer to the semantic recognition process of the source text in the foregoing embodiment, which will not be elaborated here.

[0136] In order to perform semantic fusion on the semantics of the background knowledge text and the semantics of the source text mentioned in the foregoing embodiment, in one implementation, semantics can be extracted from the source text set (which can be the fused semantics mentioned in the foregoing embodiment); semantics can be extracted from the background knowledge text; according to the respective fusion weights corresponding to the source text set and the background knowledge text, the semantics of the source text set and the semantics of the background knowledge text are fused to obtain the fused semantics, and according to the fused semantics, the text summary information of the target object is generated.

[0137] Among them, the respective fusion weights corresponding to the source text set and the background knowledge text can be custom-set. For example, it can be set that the fusion weight corresponding to the source text set is higher than the fusion weight corresponding to the background knowledge text. Another example is that it can be set that the fusion weight corresponding to the source text set is equal to the fusion weight corresponding to the background knowledge text.

[0138] Of course, the respective fusion weights corresponding to the source text set and the background knowledge text can also be determined by the text summarization model. This requires providing the knowledge and experience for determining the fusion weights to the text summarization model during the training process of the text summarization model. For example, the respective fusion weights corresponding to the source text set and the background knowledge text and the text attribute information that affects the weights can be marked in the training text of the text summarization model for the text summarization model to learn the influence relationship between the text attributes and the fusion weights, so that during the application process of the text summarization model, the respective fusion weights corresponding to the input source text set and the background knowledge text can be determined according to the text attribute analysis of the source text set and the background knowledge text. Moreover, in this embodiment, the method for determining the respective fusion weights corresponding to the source text set and the background knowledge text is not limited to this.

[0139] Still taking the t moment in the foregoing embodiment as an example, in this implementation, the first probability of each candidate word as the semantic decoding field at the t moment can be determined according to the semantics of the source text set; the second probability of each candidate word as the semantic decoding field at the t moment can be determined according to the semantics of the background knowledge text. Among them, the processing process of determining the first probability of each candidate word as the semantic decoding field at the t moment according to the semantics of the source text set can refer to the relevant description in the foregoing embodiment, which will not be elaborated here.

[0140] Among them, the first probability is also the probability that each candidate word, calculated based on the semantics of each source text and the influence weight of each source text on the fused semantics, serves as the semantic fusion field of the source text at time t; the second probability is also the probability that each candidate word, determined according to the background knowledge text, serves as the semantic decoding field of the background knowledge text at time t.

[0141] It should be noted that in this embodiment, the first probability that each candidate word in the source text set serves as the semantic decoding field at time t corresponds to the probability that each candidate word serves as the model output field at time t in the foregoing embodiment. It should be understood that in this embodiment, the probability that each candidate word determined in the foregoing embodiment serves as the model output field at time t still needs to participate in the fusion with the second probability that each candidate word in the background knowledge text serves as the semantic decoding field at time t, rather than directly determining the output field of the text summarization model at time t.

[0142] Among them, the process of determining the second probability that each candidate word serves as the semantic decoding field at time t according to the semantics of the background knowledge text is a single-text processing process. The text summarization model can encode the background knowledge text during the encoding process to obtain the semantic encoding of the background knowledge text; and can obtain the semantics of the background knowledge text from the semantic encoding of the background knowledge text during the decoding process, and then calculate the second probability that each candidate word serves as the semantic decoding field at time t according to the obtained semantics. The attention mechanism can also be added during the encoding and decoding processes.

[0143] In addition, the decoding processes for the source text set and the background knowledge text can share decoding parameters, and moreover, the attention mechanism can also use the same output hidden state to align with each source text in the background knowledge text and the source text set respectively to determine the attention distribution information for each text.

[0144] On the basis of determining the above first probability and second probability, the weighted sum of the first probability and the second probability can be performed on each candidate word according to the respective fusion weights corresponding to the source text set and the background knowledge text to determine the probability that each candidate word serves as the model output field at time t.

[0145] Accordingly, the probability that each candidate word serves as the model output field at time t can be determined. This probability is jointly determined by the semantics of the source text set and the background knowledge text, and this probability can be understood as the result of semantic fusion.

[0146] Among them, the candidate word can be a word in the word list preset for the data processing process of this embodiment, or a word in the background knowledge text or any source text, which is not limited here.

[0147] On this basis, the output field of the text summarization model at time t can be determined according to the probability of each candidate word as the output field of the model at time t. For example, the candidate word with the highest probability of being the output field of the model at time t can be used as the output field of the text summarization model at time t.

[0148] Accordingly, the output field at each output time of the text summarization model can be determined, thereby generating text summary information.

[0149] Of course, in this embodiment, other implementation manners can also be used to perform semantic fusion on multiple source texts, and this embodiment is not limited thereto. For example, without calculating the above-mentioned first probability, the weight distribution between each source text and the background knowledge text can be directly determined, and the probabilities of each candidate word determined from each source text set and the background knowledge text as the semantic decoding field at time t are directly weighted and summed to obtain the probability of each candidate word as the output field of the model at time t, and then text summary information is generated.

[0150] In the above or following embodiments, a target picture semantically matching the text summary information can be selected from the source picture set.

[0151] In this embodiment, a screening scheme for the source picture set is provided.

[0152] As mentioned in the foregoing embodiments, the source picture set may come from UGC or PGC. This results in uneven quality of the source pictures in the source picture set. And there may be several source pictures in the source picture set that are far from the text summary information in terms of content.

[0153] To improve the quality of the description content, in this embodiment, a target picture semantically matching the text summary information can be selected from the source picture set for generating the description content.

[0154] Further, in this embodiment, before selecting the target picture semantically matching the text summary information, the visual quality of each source picture in the source picture set can be evaluated separately, and the source pictures meeting the preset visual standards can be used as candidate pictures.

[0155] In one implementation manner, the source picture set can be input into a visual evaluation model. In the visual evaluation model, the visual features of each source picture in the source picture set in at least one visual dimension are determined respectively; according to the visual features of each source picture in at least one visual dimension, the visual quality parameters of each source picture are calculated respectively; and the source pictures with visual quality parameters meeting the preset standards are used as candidate pictures.

[0156] Among them, the visual dimension can be picture content, structure, light, clarity, and abstract vision, etc. Of course, the visual dimension is not limited thereto.

[0157] In practical applications, the visual evaluation model can adopt a convolutional neural network (CNN) to determine the visual features of each source image in at least one visual dimension. For example, the convolutional layer of a deep convolutional neural network can be used to extract features from each source image, and based on the extracted high-dimensional features, the visual features in at least one visual dimension can be predicted respectively.

[0158] By synthesizing the visual features in at least one visual dimension, visual quality evaluation can be performed on each source image to obtain the visual quality parameters corresponding to each source image. The visual quality parameters can be in the form of scores, of course, and are not limited to this.

[0159] In this implementation, before using the visual evaluation model, the visual evaluation model can be trained. The training process of the visual evaluation model can include:

[0160] Obtain at least one set of sample images, where each set of sample images is labeled with the arrangement order of the sample images within the group in the visual quality dimension, the visual quality parameters of each sample image, and the visual features of each sample image in at least one visual dimension; input at least one set of sample images into the visual evaluation model to train the visual evaluation model.

[0161] Among them, the arrangement order of the visual quality dimension is introduced into at least one set of sample images, which enables the visual evaluation model to not only output the visual quality parameters for each source image in the source image set, but also ensure that the output visual quality parameters conform to the arrangement order of each source image in the source image set in the visual quality dimension. For example, the source image set contains 3 source images A, B, and C, and the visual quality of the 3 source images decreases in turn. After considering the arrangement order of the 3 source images in the visual quality dimension, the scores output by the visual evaluation model may be A - 90 points, B - 80 points, and C - 70 points. It can also be understood that the visual evaluation model uses the same index benchmark to output the visual evaluation indicators of each source image in the source image set.

[0162] This can change a scoring task into a ranking task, so as to more objectively determine the visual quality of each source image in the source image set.

[0163] Based on the selected alternative images, in this embodiment, a target image that is semantically matched with the text summary information can be selected from the alternative images.

[0164] The process of semantic matching can be: respectively perform content analysis on multiple alternative images; according to the content analysis results of the multiple alternative images, select the alternative images whose content analysis results are semantically matched with the text summary information from the multiple alternative images as the target images.

[0165] The process of semantic matching refers to the process of performing content analysis on the alternative images, and it is also the process of further screening the alternative images according to the text summary.

[0166] In practical applications, LSTM (Long Short-Term Memory) can be used for content parsing. The content parsing result can be a short text or a judgment result such as "whether it matches". Based on the content parsing result, the target pictures that are semantically matched with the text summary information can be determined.

[0167] The semantic matching relationship between the target pictures and the text summary information can be represented by adding text summary information identifiers to the target pictures. Of course, it can also be represented by means such as an index table. This embodiment does not make any limitations in this regard.

[0168] In some cases, the text summary information may contain multiple text summary statements. In this embodiment, the alternative pictures whose content parsing results are semantically matched with the target text summary statement can be selected from multiple alternative pictures as the target pictures associated with the target text summary statement; the target text summary statement is any one of the multiple text summary statements.

[0169] That is, for each text summary statement, the associated target pictures are determined respectively.

[0170] In this case, it can be understood that the target pictures are grouped according to the level of the text summary statements. Of course, in this embodiment, the grouping level can be further refined. For example, it can be refined to the fields in the text summary statement, that is, the target pictures associated with each field are determined.

[0171] On this basis, in this embodiment, the description content of the target object can be generated according to the semantic matching relationship between the text summary information and the target pictures.

[0172] The process of generating the description content of the target object can be understood as the process of sorting the target pictures.

[0173] For the above several different grouping methods of the target pictures, different implementation methods can be used to generate the description content of the target object.

[0174] For the case where the target pictures are not grouped, the target pictures that are semantically matched with the text summary information can be sorted according to the display requirements. The display requirements can be smooth lighting, smooth clarity, etc., and the display requirements can be configured according to the actual situation.

[0175] For the case of grouping target images at the text summary statement level, based on the association relationship between the text summary statements and the target images, the target image groups associated with each of the multiple text summary statements can be determined; according to the display order of the multiple text summary statements, perform inter-group sorting on the multiple target image groups; on the basis of the inter-group sorting, perform intra-group sorting on each target image group according to the display requirements to determine the display order of the multiple target images;

[0176] Associate the display order of the multiple text summary statements with the display order of the multiple target images to generate the description content of the target object.

[0177] In this case, inter-group sorting can be performed on the multiple target images according to the display order of each text summary statement in the text summary information, and intra-group sorting can also be performed according to the display requirements. Combining the results of the intra-group sorting and the inter-group sorting, the final sorting result of the target images can be obtained. Among them, the execution order of the inter-group sorting and the intra-group sorting is not limited in this embodiment.

[0178] In addition, the display period of each group of target images can correspond to the display period of its corresponding text summary statement. For example, within the display period of text summary statement A, only the target images associated with text summary statement A are displayed. From another perspective, it can also be that within the display period of target image group a, the text summary statements associated with target image group a are continuously displayed.

[0179] For the case of grouping target images at the field level, the processing solution in the case of grouping target images at the text summary statement level can be referred to. The difference is that the inter-group order of the target images will be adapted to the display order of the fields.

[0180] Figure 3 It is a logical schematic diagram of a text summary process provided by an exemplary embodiment of the present application. Refer to Figure 3 , for the target object, multiple source texts from different data sources are respectively input into the encoder. The encoding parameters of each encoder are shared, and an attention mechanism is added during the encoding process of each encoder to generate the semantic encoding of each source text. Among them, the semantic encoding of each source text contains semantics, context structure information, and attention distribution information.

[0181] In addition, the background knowledge text of the target object will also be input into the encoder, and an attention mechanism can also be added to the encoder to generate the semantic encoding of the background knowledge text. Among them, the semantic encoding of the background knowledge text contains semantics, context structure information, and attention distribution information.

[0182] During the decoding process of the decoder, semantics can be obtained from the semantic encodings of each source text and the background knowledge text, respectively, and semantic fusion is performed.

[0183] Figure 3 Among them, weights are assigned at two levels. One level is the weight assignment for the fusion between the background knowledge text and the source text set. The fusion weight of the background knowledge text is λ, while the fusion weight of the source text set is 1 - λ. The other level is the weight assignment for the influence between each source text. p1 - pn are respectively assigned as .

[0184] In the decoder, according to the above weight assignment, semantic fusion can be performed on each source text and the background knowledge text to determine the probability of each word to be taken as the output field at each output moment.

[0185] And at each output moment, the word to be taken with the highest probability can be selected as the output field, and then the text summary information of the target object can be formed.

[0186] Figure 4 This is a schematic diagram of an application scenario provided by an exemplary embodiment of the present application. In Figure 4 In the provided application scenario, the target object is a POI. For this POI, a source text set can be obtained from multiple data sources, Figure 4 Several exemplary source texts are shown in it: review summary, entity description, user review list, etc.

[0187] According to the source text set, text summarization can be performed to generate an intelligent summary (reason to go) as the text description information corresponding to this POI.

[0188] And based on this text description information, a video corresponding to the target object can be further generated as the description content of the target object.

[0189] Figure 5 This is a logical schematic diagram of a semantic matching process provided by an exemplary embodiment of the present application. As Figure 5 shown, aesthetic evaluation can be respectively performed on each source picture in the source picture set to screen out alternative pictures that meet the aesthetic evaluation criteria.

[0190] For the alternative pictures, content understanding can be further performed, and the content understanding result is matched with the text summary information, and then the source picture that matches the text summary information is used as the target picture.

[0191] On this basis, the target pictures can be sorted and associated with the text summary information in terms of display time, and then a video is generated as the description content of the target object.

[0192] In the above or following embodiments, if there are multiple target pictures and the text summary information contains multiple text summary statements, a user interaction function can be provided before generating the description content. The user can perform a content selection operation to generate a content selection instruction.

[0193] In this embodiment, step 103 may include:

[0194] Display a plurality of target pictures and a plurality of text summary statements;

[0195] If a content selection instruction is received, select the target pictures and text summary statements that match the content selection instruction from the plurality of target pictures and the plurality of text summary statements;

[0196] Generate a description content of the target object according to the target pictures and text summary statements that match the content selection instruction.

[0197] Among them, as mentioned above, the user can be the query user of the target object (such as Figure 1b the user terminal in), or the maintainer of the target object (such as Figure 1b the merchant terminal in).

[0198] In this embodiment, the target pictures and text summary information can be displayed in the terminal device used by the user.

[0199] For the maintainer of the target object, this type of user can perform a content selection operation on their terminal device, select some or all of the target pictures and text summary statements, so that a content selection instruction is generated on the terminal device of this type of user. In this embodiment, based on the content selection instruction, the target pictures and text summary statements required by this type of user can be filtered out to generate a description content of the target object.

[0200] For the query user of the target object, this type of user can perform a content selection operation on their terminal device, select the text summary statements and target pictures they are interested in, so that a content selection instruction is generated on the terminal device of this type of user. In this embodiment, based on the content selection instruction, the target pictures and text summary statements that this type of user is interested in can be filtered out to generate a description content of the target object.

[0201] In addition, in this embodiment, a user interaction function can also be provided separately for the target pictures, or a user interaction function can be provided separately for the text summary information; when the number of one of the text summary statements included in the target pictures or text summary information is 1, a user interaction function can also be provided. The scenario conditions of the user interaction function in this embodiment are not limited.

[0202] Figure 6 It is a schematic flowchart of another data processing method provided for another exemplary embodiment of the present application. As Figure 6 shown, the method includes:

[0203] Step 600, encode the source texts from different data sources using the same encoding parameters to obtain the semantic encodings of the respective source texts;

[0204] Step 601: Obtain the semantics of each source text from their semantic encodings respectively.

[0205] Step 602: Fuse the semantics of each source text according to the influence weights of each source text on the fused semantics to obtain the fused semantics.

[0206] Step 603: Generate text summary information based on the fused semantics.

[0207] The data processing method provided in this embodiment can be applied to various scenarios that require text summary processing, especially scenarios that require text summary processing for texts from multiple data sources. This embodiment does not limit the application scenarios.

[0208] In this embodiment, multiple source texts can be synchronously used as the basis for text summary.

[0209] In this embodiment, during the process of encoding multiple source texts respectively, the encoding parameters are shared, which enables the input source texts to be expandable, and the number of source texts can be increased or decreased as needed. Meanwhile, the number of parameters can be saved. The semantic encoding may include the semantics and context structure information of the source text, etc.

[0210] In addition, influence weights on the fused semantics are configured for multiple source texts. A voting mechanism can be adopted, and the fused semantics are jointly determined by multiple source texts. In this semantic fusion method, information redundancy in the text summary information can be reduced, and the context word order in the text summary information can be effectively guaranteed. The text summary information obtained accordingly is both comprehensive and refined.

[0211] Among them, the influence weights of multiple source texts on the fused semantics can be customarily set. For example, a higher influence weight can be set for the source text from a certain tourism platform, while a lower influence weight can be set for the source text from a certain social platform. For another example, the influence weights of multiple source texts on the fused semantics can be equal. For instance, the influence rights of n source texts can all be 1 / n, where n is a positive integer.

[0212] Of course, if a text summarization model is used to implement the data processing solution in this embodiment, the influence weights of multiple source texts on the fused semantics can also be determined by the text summarization model. This requires providing the knowledge and experience for determining the influence weights to the text summarization model during the training process of the text summarization model. For example, the influence weights of each source text and the text attribute information that affects the influence weights can be marked in the training text of the text summarization model, so that the text summarization model can learn the influence relationship between the text attributes and the influence weights, and thus, during the application process of the text summarization model, the influence weights corresponding to each of the input source texts can be determined based on the analysis of the text attributes of the input source texts. Moreover, in this embodiment, the method for determining the influence weights of multiple source texts on the fused semantics is not limited to this.

[0213] In this embodiment, the output fields can be output according to time instants. The output fields at multiple output time instants can be combined in time order to form text summary information.

[0214] Taking the t-th instant as an example, the t-th instant can be any output instant. The semantic fusion process at the t-th instant can be as follows:

[0215] According to the semantics of each source text, calculate the probability of each candidate word as the semantic decoding field at the t-th instant for each source text; according to the influence weights of each source text on the fused semantics, for the first word, weight and sum the probabilities of taking the first word as the semantic decoding field under each source text to obtain the probability of taking the first word as the output field at the t-th instant, where the first word is any one of the candidate words.

[0216] Among them, an exemplary process of weighting and summing the probabilities of taking the first word as the semantic decoding field under each source text for the first word can be expressed as: , where p(w) is the probability of the word w as the output field at the t-th instant; p n (w) is the probability of the word w as the decoding field at the t-th instant under the n-th source text; n is the number of source texts; among them, the influence weights of multiple source texts on the fused semantics are all .

[0217] Accordingly, the probabilities of each candidate word as the output field at the t-th instant can be determined. This probability is jointly determined by the semantics of multiple source texts, and this probability can be understood as the result of semantic fusion.

[0218] Among them, the candidate words can be the words in the word list preset for the data processing process of this embodiment, or the words in any source text, which is not limited herein.

[0219] Based on this, the output field at time t can be determined according to the probability of each word to be taken as the output field at time t. For example, the word to be taken with the highest probability as the output field at time t can be used as the output field at time t.

[0220] Accordingly, the output field at each output time can be determined, thereby generating text summary information.

[0221] In addition, in this embodiment, an attention mechanism can be added during the encoding process to more accurately understand the semantics of the source text, thereby improving the accuracy of the text summary information.

[0222] In the above or following embodiments, background knowledge text corresponding to the source text set can also be added during the process of performing text summarization on the source text set to improve the comprehensiveness of the text summary information. The source text set and the background knowledge text can be directed to the same target object.

[0223] Among them, the background knowledge text can include key attribute information of the target object. The key attribute information can at least include information that can reflect the highlights or key features of the target object. For different types of target objects, the key attribute information may be different. For example, for a POI, the key attribute information may include filming locations of movies and TV shows, celebrity check-in locations, etc. For a product, the key attribute information may include brand, style, sales volume, etc. In practical applications, information that is not desired to be omitted but is desired to be known by the query user can be added to the background knowledge text.

[0224] For example, when the target object is an entity in a knowledge graph, the knowledge graph can be used as the background knowledge text of the target object.

[0225] In this embodiment, semantic fusion of the source text set and the background knowledge text can be performed to generate text summary information of the target object.

[0226] In order to perform semantic fusion of the source text set and the background knowledge text, in one implementation, semantics can be extracted from the source text set; semantics can be extracted from the background knowledge text; according to the respective fusion weights corresponding to the source text set and the background knowledge text, the semantics of the source text set and the semantics of the background knowledge text are fused to obtain the fused semantics, and according to the fused semantics, text summary information of the target object is generated.

[0227] Among them, the respective fusion weights corresponding to the source text set and the background knowledge text can be customarily set. For example, it can be set that the fusion weight corresponding to the source text set is higher than the fusion weight corresponding to the background knowledge text. Or for another example, it can be set that the fusion weight corresponding to the source text set is equal to the fusion weight corresponding to the background knowledge text.

[0228] Of course, if the data processing solution of this embodiment is implemented using a text summarization model, the fusion weights corresponding to the source text set and the background knowledge text can also be determined by the text summarization model. This requires providing the knowledge and experience for determining the fusion weights to the text summarization model during the training process of the text summarization model. For example, the fusion weights corresponding to the source text set and the background knowledge text, as well as the text attribute information that affects the weights, can be marked in the training text of the text summarization model for the text summarization model to learn the influence relationship between the text attributes and the fusion weights. Thus, during the application process of the text summarization model, the fusion weights corresponding to the source text set and the background knowledge text can be determined by analyzing the text attributes of the input source text set and background knowledge text. Moreover, in this embodiment, the method for determining the fusion weights corresponding to the source text set and the background knowledge text is not limited to this.

[0229] Still taking the t moment in the foregoing embodiment as an example, in this implementation manner, the first probability of each candidate word as the semantic decoding field at the t moment can be determined according to the semantics of the source text set; and the second probability of each candidate word as the semantic decoding field at the t moment can be determined according to the semantics of the background knowledge text.

[0230] Among them, for the process of determining the first probability of each candidate word as the semantic decoding field at the t moment according to the semantics of the source text set, reference can be made to the relevant description in the foregoing embodiment, and details are not elaborated here.

[0231] It should be noted that in this embodiment, the first probability of each candidate word in the source text set as the semantic decoding field at the t moment corresponds to the probability of each candidate word as the output field at the t moment in the foregoing embodiment. It should be understood that in this embodiment, the probability of each candidate word determined as the output field at the t moment in the foregoing embodiment still needs to participate in the fusion with the second probability of each candidate word as the semantic decoding field at the t moment in the background knowledge text, rather than directly determining the output field at the t moment.

[0232] Among them, the process of determining the second probability of each candidate word as the semantic decoding field at the t moment according to the semantics of the background knowledge text is a single-text processing process. During the encoding process, the background knowledge text can be encoded to obtain the semantic encoding of the background knowledge text; and during the decoding process, the semantics of the background knowledge text can be obtained from the semantic encoding of the background knowledge text, and then the second probability of each candidate word as the semantic decoding field at the t moment can be calculated according to the obtained semantics. An attention mechanism can also be added during the encoding and decoding processes.

[0233] In addition, the decoding processes of the source text set and the background knowledge text can share decoding parameters, and moreover, the attention mechanism can also use the same output hidden state to align with each source text in the background knowledge text and the source text set respectively to determine the attention distribution information under each text.

[0234] Based on the determination of the above first probability and second probability, the weighted sum of the first probability and the second probability can be performed on each candidate word according to the respective fusion weights corresponding to the source text set and the background knowledge text, so as to determine the probability of each candidate word as the output field at time t.

[0235] Accordingly, the probability of each candidate word as the output field at time t can be determined. This probability is jointly determined by the semantics of the source text set and the background knowledge text, and this probability can be understood as the result generated by semantic fusion.

[0236] Among them, the candidate word can be a word in the word list preset for the data processing process of this embodiment, or a word in the background knowledge text or any source text, which is not limited herein.

[0237] On this basis, the output field at time t can be determined according to the probability of each candidate word as the output field at time t. For example, the candidate word with the highest probability as the output field at time t can be used as the output field at time t.

[0238] Accordingly, the output field at each output time can be determined, thereby generating text summary information.

[0239] Of course, in this embodiment, other implementation manners can also be used to perform semantic fusion on multiple source texts, and this embodiment is not limited thereto. For example, without calculating the above first probability, the weight distribution between each source text and the background knowledge text can be directly determined, and the probability of each candidate word determined by each source text set and the background knowledge text as the semantic decoding field at time t can be directly weighted and summed to obtain the probability of each candidate word as the output field at time t, and then text summary information is generated.

[0240] Figure 7 It is a schematic flowchart of a method for processing an entity object provided for another exemplary embodiment of the present application. As Figure 7 shown, the method includes:

[0241] Step 700, determine a plurality of network data sources associated with the entity object, where the network data sources include web page content and at least one data platform providing services related to the entity object;

[0242] Step 701, obtain a text data set and a picture data set of the entity object from a plurality of data sources;

[0243] Step 702, obtain key text data of the entity object by performing data summarization on the text data set;

[0244] Step 703, screen key picture data matching the text summary information from the picture data set;

[0245] Step 704, generate the description content of the entity object according to the key text data and key picture data, and the description content includes video data.

[0246] In the method for processing entity objects provided in this embodiment, it can be applied to various scenarios that require creating description content for entity objects. This embodiment does not limit the application scenarios. Among them, in different application scenarios, entity objects may vary widely and be diverse. Entity objects can be people, goods, geographical locations, events, relationships, etc., and this embodiment also does not limit the types of entity objects. In addition, entity objects can be any objects in the knowledge graph.

[0247] In this embodiment, multiple network data sources managed by the entity object can be determined. The network data sources include but are not limited to web page content and at least one data platform that provides services related to the entity object. For different types of entity objects, the types of network data sources may not be exactly the same. For example, for entity objects of the geographical location type, multiple network data sources may include various tourism platforms, encyclopedia web page content, etc.

[0248] Among them, for the technical details of steps 701 - 704, reference can be made to Figure 1a the descriptions in the related embodiments of the provided data processing method. To save space, it will not be elaborated here, but this should not cause a loss of the protection scope of this application.

[0249] Figure 8 It is a schematic flowchart of a method for providing an entity object provided in another exemplary embodiment of this application. As Figure 8 shown, the method includes:

[0250] Step 800, obtain the text data set and picture data set of the entity object from the web page and the first data platform;

[0251] Step 801, obtain the text data set and picture data set of the entity object from multiple data sources;

[0252] Step 802, obtain the key text data of the entity object by performing data summarization on the text data set;

[0253] Step 803, screen the key picture data that matches the text summary information from the picture data set;

[0254] Step 804, generate the description content of the entity object according to the key text data and key picture data;

[0255] Step 805, based on a search request for the entity object, display the entity object on the client of the second data platform, and provide the description content at the associated position of the entity object, and the description content includes video.

[0256] In the method for providing an entity object provided in this embodiment, it can be applied to various entity object search scenarios. This embodiment does not limit the application scenarios. Among them, in different application scenarios, the entity objects may be ever-changing and diverse. The entity object can be a person, a commodity, a geographical location, an event, a relationship, etc., and this embodiment does not limit the type of the entity object either. In addition, the entity object can be any object in the knowledge graph.

[0257] In this embodiment, description content can be generated for the entity object in advance. For the generation process of the description content, reference can be made to Figure 1a the description in the related embodiments of the provided data processing method. To save space, it will not be elaborated here, but this should not cause a loss of the protection scope of this application.

[0258] In this embodiment, the description content of the entity object can be displayed on a second data platform outside the first data platform. For example, if the first data platform is a tourism platform, in this embodiment, a dedicated entity object search platform can be deployed, and the description content of the entity object can be displayed on the entity object search platform.

[0259] Accordingly, in this embodiment, a search entry for the entity object outside the first data platform can be provided for the user, so as to more flexibly respond to the user's search needs, rather than being limited by the search capabilities of the first data platform.

[0260] In addition, in this embodiment, a display search interface can be provided in the second data platform to respond to the entity object search operation issued by the user. The interface format of the search interface is not limited in this embodiment, and the search interface can adopt a question-and-answer format, a search bar format, etc.

[0261] For other details not covered in this embodiment, reference can be made to Figure 1a the description in the related embodiments of the provided data processing method. To save space, it will not be elaborated here, but this should not cause a loss of the protection scope of this application.

[0262] It should be noted that the execution subject of each step of the method provided in the above embodiment can be the same device, or the method can also be executed by different devices as the execution subject. For example, the execution subject of steps 100 to 102 can be device A; or, the execution subject of steps 100 and 101 can be device A, and the execution subject of step 102 can be device B; and so on.

[0263] In addition, in some of the processes described in the above embodiments and the accompanying drawings, which include a plurality of operations that occur in a specific order, it should be clearly understood that these operations may not be executed in the order in which they appear herein or may be executed in parallel. The serial numbers of the operations, such as 100, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions such as "first" and "second" in this text are used to distinguish different messages, devices, words, etc., and do not represent a sequence, nor do they limit that "first" and "second" are of different types.

[0264] Figure 9 Schematic diagram of a computing device provided by another exemplary embodiment of the present application. As Figure 9 shown, the computing device includes: a memory 90 and a processor 91;

[0265] The memory 90 is used to store one or more computer instructions;

[0266] The processor 91 is coupled to the memory 90 and is used to execute one or more computer instructions for:

[0267] Obtaining a source text set and a source picture set corresponding to a target object from multiple data sources respectively;

[0268] Performing text summarization on the source text set to obtain text summary information of the target object;

[0269] Selecting a target picture that semantically matches the text summary information from the source picture set;

[0270] Generating a description content of the target object according to the text summary information and the target picture.

[0271] In an alternative embodiment, when the processor 91 performs text summarization on the source text set to obtain text summary information of the target object, it is used for:

[0272] Inputting the source text from different data sources into a text summarization model;

[0273] In the text summarization model, performing semantic recognition on the source text from different data sources to obtain the semantics of each source text;

[0274] Performing semantic fusion on the semantics of each source text to generate text summary information of the target object.

[0275] In an alternative embodiment, when the processor 91 performs semantic recognition on the source text from different data sources to obtain the semantics of each source text, it is used for:

[0276] Encode the source texts from different data sources using the same encoding parameters to obtain the semantic encodings of the source texts;

[0277] Respectively obtain the semantics of the source texts from the semantic encodings of the source texts.

[0278] In an alternative embodiment, the text summarization model includes a neural network model. When the processor 91 encodes the source texts from different data sources, it is used for:

[0279] Use the attention mechanism of the neural network model to encode the source texts from different data sources.

[0280] In an alternative embodiment, when the processor 91 performs semantic fusion on the semantics included in each source text to generate the text summary information of the target object, it is used for:

[0281] Fuse the semantics of each source text according to the influence weights of each source text on the fused semantics to obtain the fused semantics;

[0282] Generate the text summary information of the target object according to the fused semantics.

[0283] In an alternative embodiment, when the processor 91 fuses the semantics of each source text according to the influence weights of each source text on the fused semantics to obtain the fused semantics, it is used for:

[0284] Respectively calculate the probabilities of each candidate word as the semantic decoding field at time t according to the semantics of each source text;

[0285] According to the influence weights of each source text on the fused semantics, for the first word, calculate the probability of the first word as the output field of the model at time t according to the probability of taking the first word as the semantic decoding field at time t under each source text, where the first word is any one of the candidate words, and t is any output time of the text summarization model.

[0286] In an alternative embodiment, when the processor 91 generates the text summary information of the target object according to the fused semantics, it is used for:

[0287] Take the candidate word with the highest probability as the output field of the model at time t as the output field of the text summarization model at time t to generate the text summary information.

[0288] In an alternative embodiment, the processor 91 is further used for:

[0289] Input the background knowledge text of the target object into the text summarization model, and the background knowledge text contains the key attribute information of the target object;

[0290] In the text summary model, semantic recognition is performed on the background knowledge text to obtain the semantics of the background knowledge text;

[0291] When generating the text summary information of the target object according to the fused semantics, it is used for:

[0292] Semantically fuse the fused semantics corresponding to the source text set and the semantics of the background knowledge text to generate the text summary information of the target object.

[0293] In an alternative embodiment, when the processor 91 semantically fuses the fused semantics corresponding to the source text set and the semantics of the background knowledge text to generate the text summary information of the target object, it is used for:

[0294] Based on the semantics of each source text and the influence weight of each source text on the fused semantics, calculate the probability of each candidate word as the semantic fusion field of the source text at time t;

[0295] According to the background knowledge text, determine the probability of each candidate word as the semantic decoding field of the background knowledge text at time t;

[0296] According to the respective fusion weights of the source text set and the background knowledge text, calculate the probability of each candidate word as the model output field at time t based on the probability of each candidate word as the semantic fusion field of the source text at time t and the probability of each candidate word as the semantic decoding field of the background knowledge text at time t;

[0297] Take the word with the highest probability as the model output field at time t as the output field of the text summary model at time t to generate the text summary information;

[0298] Where t is any output time of the text summary model.

[0299] In an alternative embodiment, when the processor 91 selects the target picture that semantically matches the text summary information from the source picture set, it is used for:

[0300] Perform visual quality evaluation on each source picture in the source picture set, and take the source pictures that meet the preset visual standards as alternative pictures;

[0301] Select the target picture that semantically matches the text summary information from the alternative pictures.

[0302] In an alternative embodiment, when the processor 91 performs visual quality evaluation on each source picture in the source picture set and takes the source pictures that meet the preset visual standards as alternative pictures, it is used for:

[0303] Input the source picture set into the visual evaluation model;

[0304] In a visual evaluation model, determine the visual features of each source picture in the source picture set respectively under at least one visual dimension;

[0305] Calculate the visual quality parameters of each source picture respectively according to the visual features of each source picture under at least one visual dimension;

[0306] Use the source pictures whose visual quality parameters meet the preset criteria as candidate pictures.

[0307] In an optional embodiment, during the training process of the visual evaluation model, the processor 91 is used for:

[0308] Obtain at least one group of sample pictures, and each group of sample pictures is labeled with the arrangement order between the sample pictures in the group in the visual quality dimension, the visual quality parameters of each sample picture, and the visual features of each sample picture under at least one aesthetic dimension;

[0309] Input at least one group of sample pictures into the aesthetic evaluation model to train the visual evaluation model.

[0310] In an optional embodiment, when there are multiple candidate pictures, when the processor 91 selects a target picture that semantically matches the text summary information from the candidate pictures, it is used for:

[0311] Perform content analysis on multiple candidate pictures respectively;

[0312] Select, from multiple candidate pictures, the candidate pictures whose content understanding results semantically match the text summary information as the target pictures.

[0313] In an optional embodiment, if the text summary information contains multiple text summary statements, when the processor 91 selects, from multiple candidate pictures, the candidate pictures whose content understanding results semantically match the text summary information as the target pictures, it is used for:

[0314] Determine the candidate pictures whose content understanding results semantically match the target text summary statement as the target pictures associated with the target text summary statement;

[0315] The target text summary statement is any one of the multiple text summary statements.

[0316] In an optional embodiment, when the processor 91 generates the description content of the target object according to the semantic matching relationship between the text summary information and the target picture, it is used for:

[0317] Based on the association relationship between the text summary statement and the target picture, determine the target picture groups respectively associated with multiple text summary statements;

[0318] Sort the multiple groups of target pictures in order among the groups according to the display order of the multiple text summary statements;

[0319] On the basis of the sorting among the groups, sort each group of target pictures according to the display requirements to determine the display order of the multiple target pictures;

[0320] Associate the display order of the multiple text summary statements with the display order of the multiple target pictures to generate the description content of the target object.

[0321] In an alternative embodiment, the description content includes a video.

[0322] In an alternative embodiment, the processor 91 is further configured to:

[0323] Display a query interface;

[0324] In response to a query instruction for the target object occurring in the query interface, display the description content of the target object in the query interface.

[0325] In an alternative embodiment, the processor 91 is further configured to:

[0326] Display a content creation interface;

[0327] In response to a content creation instruction for the target object occurring in the content creation interface, execute operations such as separately obtaining the source text set and source picture set corresponding to the target object from multiple data sources and subsequent operations to generate the description content of the target object;

[0328] In the content creation interface, display the description content of the target object.

[0329] In an alternative embodiment, the target object includes an entity in a knowledge graph.

[0330] In an alternative embodiment, if there are multiple target pictures and the text summary information contains multiple text summary statements, when the processor 91 generates the description content of the target object according to the text summary information and the target pictures, it is configured to:

[0331] Display the multiple target pictures and the multiple text summary statements;

[0332] If a content selection instruction is received, select the target pictures and text summary statements that match the content selection instruction from the multiple target pictures and the multiple text summary statements;

[0333] Generate the description content of the target object according to the target pictures and text summary statements that match the content selection instruction.

[0334] It should be noted that for the technical details in the above embodiments of the computing device, reference can be made to Figure 1aDescriptions in the relevant embodiments of the data processing method shown are not repeated here for the sake of brevity, but this should not result in any loss of the protection scope of the present application.

[0335] Further, as Figure 9 shown, the computing device further includes other components such as a communication component 92 and a power supply component 93. Figure 9 Only some components are schematically shown in Figure 9 for illustration purposes, and it does not mean that the computing device only includes

[0336] Correspondingly, an embodiment of the present application further provides a computer-readable storage medium storing a computer program, and when the computer program is executed, it can implement the steps executable by the computing device in the above method embodiment.

[0337] Figure 10 is a schematic structural diagram of another computing device provided for another exemplary embodiment of the present application. As Figure 10 shown, the computing device includes: a memory 100 and a processor 101;

[0338] The memory 100 is used to store one or more computer instructions;

[0339] The processor 101 is coupled to the memory 100 and is used to execute one or more computer instructions for:

[0340] Encoding source texts from different data sources using the same encoding parameters to obtain semantic encodings of the source texts;

[0341] Respectively obtaining the semantics of the source texts from the semantic encodings of the source texts;

[0342] Fusing the semantics of the source texts according to the influence weights of the source texts on the fused semantics to obtain the fused semantics;

[0343] Generating text summary information according to the fused semantics.

[0344] In an optional embodiment, when encoding source texts from different data sources, the processor 101 is used for:

[0345] Encoding source texts from different data sources using an attention mechanism.

[0346] In an optional embodiment, when fusing the semantics of the source texts according to the influence weights of the source texts on the fused semantics to obtain the fused semantics, the processor 101 is used for:

[0347] Calculating the probabilities of each candidate word as the semantic decoding field at time t under each source text according to the semantics of the source texts;

[0348] According to the influence weights of each source text on the fused semantics, for the first word, the probabilities of taking the first word as the semantic decoding field at time t under each source text are weighted and summed to obtain the probability of taking the first word as the model output field at time t, where the first word is any word among the words to be taken, and t is any output time.

[0349] In an alternative embodiment, when generating text summary information according to the fused semantics, the processor 101 is configured to:

[0350] Take the word to be taken with the highest probability as the output field at time t to generate text summary information.

[0351] In an alternative embodiment, the processor 101 is further configured to:

[0352] Obtain the background knowledge text corresponding to the source text set, where the background knowledge text contains key attribute information of the target object;

[0353] Extract the semantics of the background knowledge text;

[0354] When generating text summary information according to the fused semantics, it is configured to:

[0355] Perform semantic fusion on the fused semantics and the semantics of the background knowledge text to generate text summary information.

[0356] In an alternative embodiment, when the processor 101 performs semantic fusion on the fused semantics and the semantics of the background knowledge text to generate text summary information, it is configured to:

[0357] Based on the semantics of each source text and the influence weights of each source text on the fused semantics, calculate the probability of each word to be taken as the semantic fusion field of the source text at time t;

[0358] According to the background knowledge text, determine the probability of each word to be taken as the semantic decoding field of the background knowledge text at time t;

[0359] According to the respective fusion weights corresponding to the source text set and the background knowledge text, calculate the probability of each word to be taken as the model output field at time t based on the probability of each word to be taken as the semantic fusion field of the source text at time t and the probability of each word to be taken as the semantic decoding field of the background knowledge text at time t;

[0360] Take the word with the highest probability as the output field at time t as the output field at time t to generate text summary information;

[0361] where t is any output time.

[0362] It should be noted that for the technical details in the above embodiments of the computing device, reference can be made to Figure 6Descriptions in the relevant embodiments of the data processing method shown are not repeated here for the sake of brevity, but this should not cause loss of the protection scope of this application.

[0363] Further, as Figure 10 shown, the computing device further includes: other components such as a communication component 102 and a power supply component 103. Figure 10 Only some components are schematically shown in Figure 10 shown, and it does not mean that the computing device only includes

[0364] Correspondingly, an embodiment of this application further provides a computer-readable storage medium storing a computer program, and when the computer program is executed, it can implement the steps executable by the computing device in the above method embodiment.

[0365] Figure 11 is a schematic structural diagram of another computing device provided by another exemplary embodiment of this application. As Figure 11 shown, the computing device includes: a memory 110 and a processor 111;

[0366] The memory 110 is used to store one or more computer instructions;

[0367] The processor 111 is coupled to the memory 110 and is used to execute one or more computer instructions for:

[0368] Determine multiple network data sources associated with the entity object, and the network data sources include web page content and at least one data platform providing services related to the entity object;

[0369] Obtain a text data set and a picture data set of the entity object from multiple data sources;

[0370] Obtain the key text data of the entity object by performing data summarization on the text data set;

[0371] Screen key picture data matching the text summary information from the picture data set;

[0372] Generate a description content of the entity object according to the key text data and the key picture data, and the description content includes video data.

[0373] In an optional embodiment, when the processor 111 performs text summarization on the text data set to obtain the key text data of the entity object, it is used for:

[0374] Input the text data from different data sources into the text summarization model;

[0375] In the text summarization model, perform semantic recognition on the text data from different data sources to obtain the semantics of each text data;

[0376] Semantically fuse the semantics of each text data to generate the key text data of the entity object.

[0377] In an alternative embodiment, when the processor 111 performs semantic recognition on text data from different data sources to obtain the semantics of each text data, it is used for:

[0378] Encode the text data from different data sources using the same encoding parameters to obtain the semantic encodings of each text data;

[0379] Respectively obtain the semantics of each text data from the semantic encodings of each text data.

[0380] In an alternative embodiment, the text summary model includes a neural network model. When the processor 111 encodes the text data from different data sources, it is used for:

[0381] Encode the text data from different data sources using the attention mechanism of the neural network model.

[0382] In an alternative embodiment, when the processor 111 semantically fuses the semantics included in each text data to generate the key text data of the entity object, it is used for:

[0383] Fuse the semantics of each text data according to the influence weights of each text data on the fused semantics to obtain the fused semantics;

[0384] Generate the key text data of the entity object according to the fused semantics.

[0385] In an alternative embodiment, when the processor 111 fuses the semantics of each text data according to the influence weights of each text data on the fused semantics to obtain the fused semantics, it is used for:

[0386] Respectively calculate the probabilities of each candidate word as the semantic decoding field at time t according to the semantics of each text data;

[0387] According to the influence weights of each text data on the fused semantics, for the first word, calculate the probability of the first word as the output field of the model at time t according to the probabilities of taking the first word as the semantic decoding field at time t under each text data, where the first word is any one of the candidate words, and t is any output time of the text summary model.

[0388] In an alternative embodiment, when the processor 111 generates the key text data of the entity object according to the fused semantics, it is used for:

[0389] Take the candidate word with the highest probability as the output field of the text summary model at time t to generate the key text data.

[0390] In an alternative embodiment, the processor 111 is further configured to:

[0391] Input the background knowledge text of the entity object into the text summarization model, where the background knowledge text contains the key attribute information of the entity object;

[0392] In the text summarization model, perform semantic recognition on the background knowledge text to obtain the semantics of the background knowledge text;

[0393] When generating the key text data of the entity object according to the fused semantics, it is used for:

[0394] Perform semantic fusion on the fused semantics corresponding to the text data set and the semantics of the background knowledge text to generate the key text data of the entity object.

[0395] In an alternative embodiment, when the processor 111 performs semantic fusion on the fused semantics corresponding to the text data set and the semantics of the background knowledge text to generate the key text data of the entity object, it is used for:

[0396] Based on the semantics of each text data and the influence weight of each text data on the fused semantics, calculate the probability of each candidate word as the semantic fusion field of the text data at time t;

[0397] According to the background knowledge text, determine the probability of each candidate word as the semantic decoding field of the background knowledge text at time t;

[0398] According to the respective fusion weights corresponding to the text data set and the background knowledge text, calculate the probability of each candidate word as the output field of the model at time t based on the probability of each candidate word as the semantic fusion field of the text data at time t and the probability of each candidate word as the semantic decoding field of the background knowledge text at time t;

[0399] Take the word with the highest probability as the output field of the text summarization model at time t to generate the key text data;

[0400] Where t is any output time of the text summarization model.

[0401] In an alternative embodiment, when the processor 111 selects the target picture that semantically matches the key text data from the picture data set, it is used for:

[0402] Perform visual quality assessment on each picture data in the picture data set, and take the picture data that meets the preset visual standard as the candidate picture;

[0403] Select the target picture that semantically matches the key text data from the candidate pictures.

[0404] In an alternative embodiment, when the processor 111 respectively performs visual quality assessment on each piece of picture data in the picture data set and uses the picture data that meets the preset visual criteria as alternative pictures, it is used for:

[0405] Input the picture data set into the visual assessment model;

[0406] In the visual assessment model, respectively determine the visual features of each piece of picture data in the picture data set under at least one visual dimension;

[0407] According to the visual features of each piece of picture data under at least one visual dimension, respectively calculate the visual quality parameters of each piece of picture data;

[0408] Use the picture data whose visual quality parameters meet the preset criteria as alternative pictures.

[0409] In an alternative embodiment, during the training process of the visual assessment model, the processor 111 is used for:

[0410] Obtain at least one set of sample pictures, and each set of sample pictures is labeled with the arrangement order of the sample pictures within the group in the visual quality dimension, the visual quality parameters of each sample picture, and the visual features of each sample picture under at least one aesthetic dimension;

[0411] Input at least one set of sample pictures into the aesthetic assessment model to train the visual assessment model.

[0412] In an alternative embodiment, when there are multiple alternative pictures, when the processor 111 selects a target picture that semantically matches the key text data from the alternative pictures, it is used for:

[0413] Respectively perform content parsing on multiple alternative pictures;

[0414] According to the content parsing results of multiple alternative pictures, select the alternative pictures whose content understanding results semantically match the key text data from multiple alternative pictures as the target pictures.

[0415] In an alternative embodiment, if the key text data contains multiple text summary statements, when the processor 111 selects the alternative pictures whose content understanding results semantically match the key text data from multiple alternative pictures as the target pictures according to the content understanding results of multiple alternative pictures, it is used for:

[0416] Determine the alternative pictures whose content understanding results semantically match the target text summary statement as the target pictures associated with the target text summary statement;

[0417] The target text summary statement is any one of the multiple text summary statements.

[0418] In an alternative embodiment, when generating the description content of the entity object according to the semantic matching relationship between the key text data and the target picture, the processor 111 is configured to:

[0419] Determine the target picture groups associated with the respective text summary statements based on the association relationship between the text summary statements and the target picture;

[0420] Sort the multiple target picture groups among groups according to the display order of the multiple text summary statements;

[0421] On the basis of the inter-group sorting, sort each target picture group within the group according to the display requirements to determine the display order of the multiple target pictures;

[0422] Associate the display order of the multiple text summary statements with the display order of the multiple target pictures to generate the description content of the entity object.

[0423] It should be noted that for the technical details in the foregoing embodiments of the computing device, reference may be made to Figure 7 the description in the relevant embodiments of the processing method of the entity object shown. For the sake of brevity, it will not be repeated here, but this should not cause any loss to the protection scope of the present application.

[0424] Furthermore, as Figure 11 shown, the computing device further includes: a communication component 112, a power supply component 113, and other components. Figure 11 Only some components are schematically shown in Figure 11 and it does not mean that the computing device only includes

[0425] the components shown.

[0425] Correspondingly, an embodiment of the present application further provides a computer-readable storage medium storing a computer program, and when the computer program is executed, it can implement each step executable by the computing device in the foregoing method embodiment.

[0426] Figure 12 This is a schematic structural diagram of another computing device provided by another exemplary embodiment of the present application. As Figure 12 shown, the computing device includes: a memory 120 and a processor 121;

[0427] The memory 120 is used to store one or more computer instructions;

[0428] The processor 121 is coupled to the memory 120 and is configured to execute one or more computer instructions for:

[0429] Obtain the text data set and the picture data set of the entity object from the web page and the first data platform;

[0430] Obtain the text data set and the picture data set of the entity object from multiple data sources;

[0431] By performing data summarization on the text dataset, key text data of the entity object is obtained;

[0432] Filter key image data matching the text summary information from the image dataset;

[0433] Generate a description content of the entity object based on the key text data and the key image data;

[0434] Based on a search request for the entity object, display the entity object on the client of the second data platform, and provide the description content at the associated position of the entity object, where the description content includes videos.

[0435] In an optional embodiment, when the processor 121 obtains the key text data of the entity object by performing data summarization on the text dataset, it is used for:

[0436] Input the text dataset obtained from the web page and the first data platform into the text summary model;

[0437] In the text summary model, perform semantic recognition on the text dataset obtained from the web page and the first data platform to obtain the semantics of each text data;

[0438] Perform semantic fusion on the semantics of each text data to generate the key text data of the entity object.

[0439] In an optional embodiment, when the processor 121 performs semantic recognition on the text data obtained from the web page and the first data platform to obtain the semantics of each text data, it is used for:

[0440] Use the same encoding parameters to encode the text data obtained from the web page and the first data platform to obtain the semantic encoding of each text data;

[0441] Respectively obtain the semantics of each text data from the semantic encoding of each text data.

[0442] In an optional embodiment, the text summary model includes a neural network model. When the processor 121 encodes the text data obtained from the web page and the first data platform, it is used for:

[0443] Use the attention mechanism of the neural network model to encode the text data obtained from the web page and the first data platform.

[0444] In an optional embodiment, when the processor 121 performs semantic fusion on the semantics included in each text data to generate the key text data of the entity object, it is used for:

[0445] Fuse the semantics of each text data according to the influence weight of each text data on the fused semantics to obtain the fused semantics;

[0446] Generate key text data of the entity object according to the fused semantics.

[0447] In an optional embodiment, when the processor 121 fuses the semantics of each text data according to the influence weight of each text data on the fused semantics to obtain the fused semantics, it is used for:

[0448] Calculate the probability of each candidate word as the semantic decoding field at time t according to the semantics of each text data.

[0449] According to the influence weight of each text data on the fused semantics, for the first word, calculate the probability of the first word as the model output field at time t according to the probability of taking the first word as the semantic decoding field at time t under each text data, where the first word is any one of the candidate words, and t is any output time of the text summarization model.

[0450] In an optional embodiment, the processor 121 is further used for:

[0451] Input the background knowledge text of the entity object into the text summarization model, where the background knowledge text contains the key attribute information of the entity object.

[0452] In the text summarization model, perform semantic recognition on the background knowledge text to obtain the semantics of the background knowledge text.

[0453] When generating the key text data of the entity object according to the fused semantics, it is used for:

[0454] Perform semantic fusion on the fused semantics corresponding to the text data set and the semantics of the background knowledge text to generate the key text data of the entity object.

[0455] In an optional embodiment, when the processor 121 performs semantic fusion on the fused semantics corresponding to the text data set and the semantics of the background knowledge text to generate the key text data of the entity object, it is used for:

[0456] Calculate the probability of each candidate word as the semantic fusion field of the text data at time t based on the semantics of each text data and the influence weight of each text data on the fused semantics.

[0457] Determine the probability of each candidate word as the semantic decoding field of the background knowledge text at time t according to the background knowledge text.

[0458] According to the respective fusion weights corresponding to the text data set and the background knowledge text, calculate the probability of each candidate word as the model output field at time t according to the probability of each candidate word as the semantic fusion field of the text data at time t and the probability of each candidate word as the semantic decoding field of the background knowledge text at time t.

[0459] The word with the highest probability among the model output fields at time t is used as the output field of the text summary model at time t to generate key text data;

[0460] where t is any output time of the text summary model.

[0461] It should be noted that for the technical details in the above embodiments of the computing device, reference can be made to Figure 8 the description in the relevant embodiments of the method for providing the entity object shown. For the sake of brevity, it will not be repeated here, but this should not cause any loss to the protection scope of this application.

[0462] Furthermore, as Figure 12 shown, the computing device further includes: other components such as a communication component 122 and a power supply component 123. Figure 12 Only some components are schematically shown in Figure 12 and it does not mean that the computing device only includes

[0463]

[0464] Figures 9 - 12

[0465] Figures 9 - 12

[0465] Among them, Figures 9 - 12The communication component therein is configured to facilitate communication, in a wired or wireless manner, between the device where the communication component is located and other devices. The device where the communication component is located can access a wireless network based on a communication standard, such as WiFi, 2G, 3G, 4G / LTE, 5G and other mobile communication networks, or a combination thereof. In an exemplary embodiment, the communication component receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component further includes a Near Field Communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on Radio Frequency Identification (RFID) technology, Infrared Data Association (IrDA) technology, Ultra Wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

[0466] Wherein, Figures 9 - 12 The power supply component therein provides power for various components of the device where the power supply component is located. The power supply component can include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power for the device where the power supply component is located.

[0467] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0468] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and the combination of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices produce a device for realizing the functions specified in one or more processes in the flowchart and / or one or more blocks in the block diagram.

[0469] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including an instruction device that realizes the functions specified in one or more processes in the flowchart and / or one or more blocks in the block diagram.

[0470] These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0471] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0472] Memory may include non-permanent storage in a computer-readable medium, in the form of random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0473] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0474] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0475] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, various modifications and changes can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A data processing method, characterized in that, Applied to scenarios where descriptive information needs to be created for an object, including: Separately obtaining a source text set and a source picture set corresponding to the target object for which descriptive information needs to be created from multiple data sources; In a text summarization model, performing semantic recognition on source texts from different data sources to obtain the semantics of each source text; Fusing the semantics of each source text according to the influence weights of each source text on the fused semantics to obtain the fused semantics, including: calculating the probability of each candidate word as the semantic decoding field at time t for each source text according to the semantics of each source text; according to the influence weights of each source text on the fused semantics, for the first word, calculating the probability of the first word as the model output field at time t according to the probability of taking the first word as the semantic decoding field at time t for each source text, where the first word is any one of the candidate words, and t is any output time of the text summarization model; Generating text summary information of the target object according to the fused semantics; Selecting a target picture that semantically matches the text summary information from the source picture set; Generating descriptive content of the target object according to the text summary information and the target picture.

2. The method according to claim 1, wherein The performing semantic recognition on source texts from different data sources to obtain the semantics of each source text includes: Encoding source texts from different data sources with the same encoding parameters to obtain semantic encodings of each source text; Separately obtaining the semantics of each source text from the semantic encodings of each source text.

3. The method according to claim 2, wherein The text summarization model includes a neural network model, and the encoding source texts from different data sources includes: Encoding source texts from different data sources by using the attention mechanism of the neural network model.

4. The method according to claim 1, wherein The generating text summary information of the target object according to the fused semantics includes: Taking the candidate word with the highest probability as the model output field at time t as the output field of the text summarization model at time t to generate the text summary information.

5. The method according to claim 1, wherein It further includes: Inputting the background knowledge text of the target object into the text summarization model, where the background knowledge text contains key attribute information of the target object; Performing semantic recognition on the background knowledge text in the text summarization model to obtain the semantics of the background knowledge text; The generating text summary information of the target object according to the fused semantics includes: Performing semantic fusion on the fused semantics corresponding to the source text set and the semantics of the background knowledge text to generate the text summary information of the target object.

6. The method according to claim 5, wherein The performing semantic fusion on the fused semantics corresponding to the source text set and the semantics of the background knowledge text to generate the text summary information of the target object includes: Calculating the probability of each candidate word as the semantic fusion field of the source text at time t based on the semantics of each source text and the influence weights of each source text on the fused semantics; Determining the probability of each candidate word as the semantic decoding field of the background knowledge text at time t according to the background knowledge text; Calculate the probability of each candidate word as the output field of the model at time t according to the respective fusion weights corresponding to the source text set and the background knowledge text, based on the probability of each candidate word as the semantic fusion field of the source text at time t and the probability of it as the semantic decoding field of the background knowledge text at time t; Take the word with the highest probability as the output field of the model at time t as the output field of the text summary model at time t to generate the text summary information; where t is any output time of the text summary model.

7. The method according to claim 1, characterized in that Select a target picture semantically matching the text summary information from the source picture set, including: Perform visual quality assessment on each source picture in the source picture set, and take the source pictures meeting the preset visual criteria as alternative pictures; Select a target picture semantically matching the text summary information from the alternative pictures.

8. The method according to claim 7, wherein The performing visual quality assessment on each source picture in the source picture set and taking the source pictures meeting the preset visual criteria as alternative pictures includes: Input the source picture set into a visual assessment model; In the visual assessment model, respectively determine the visual features of each source picture in the source picture set under at least one visual dimension; Calculate the visual quality parameters of each source picture according to the visual features of each source picture under at least one visual dimension; Take the source pictures with visual quality parameters meeting the preset criteria as alternative pictures.

9. The method according to claim 8, characterized in that The training process of the visual assessment model is: Obtain at least one set of sample pictures, where each set of sample pictures is labeled with the arrangement order of the sample pictures within the group in the visual quality dimension, the visual quality parameters of each sample picture, and the visual features of each sample picture under at least one aesthetic dimension; Input the at least one set of sample pictures into the visual assessment model to train the visual assessment model.

10. The method according to claim 7, wherein There are multiple alternative pictures, and the selecting a target picture semantically matching the text summary information from the alternative pictures includes: Perform content analysis on multiple alternative pictures respectively; According to the content analysis results of the multiple alternative pictures, select the alternative pictures whose content understanding results are semantically matching the text summary information from the multiple alternative pictures as the target pictures.

11. The method according to claim 10, wherein If the text summary information contains multiple text summary statements, the selecting the alternative pictures whose content understanding results are semantically matching the text summary information from the multiple alternative pictures as the target pictures includes: Determine the alternative pictures whose content understanding results are semantically matching the target text summary statement as the target pictures associated with the target text summary statement; The target text summary statement is any one of the multiple text summary statements.

12. The method according to claim 11, wherein The generating the description content of the target object according to the semantic matching relationship between the text summary information and the target picture includes: Based on the association relationship between the text summary statements and the target pictures, determine the target picture groups respectively associated with the multiple text summary statements; Sort the multiple target picture groups inter - group according to the display order of the multiple text summary statements; On the basis of sorting among groups, perform in-group sorting on each target picture group according to the display requirements to determine the display order of multiple target pictures; Associate the display order of the multiple text summary statements with the display order of the multiple target pictures to generate the description content of the target object.

13. The method according to claim 12, wherein The description content includes videos.

14. The method according to claim 1, characterized in that It also includes: Display a query interface; In response to a query instruction for the target object that occurs in the query interface, display the description content of the target object in the query interface.

15. The method according to claim 1, characterized in that, It also includes: Display a content creation interface; In response to a content creation instruction for the target object that occurs in the content creation interface, execute the operations of respectively obtaining the source text set and source picture set corresponding to the target object from multiple data sources and subsequent operations to generate the description content of the target object; In the content creation interface, display the description content of the target object.

16. The method according to claim 1, wherein The target object includes entities in a knowledge graph.

17. The method according to claim 1, wherein If there are multiple target pictures and the text summary information contains multiple text summary statements, generating the description content of the target object according to the text summary information and the target pictures includes: Display multiple target pictures and the multiple text summary statements; If a content selection instruction is received, select the target pictures and text summary statements that match the content selection instruction from the multiple target pictures and the multiple text summary statements; Generate the description content of the target object according to the target pictures and text summary statements that match the content selection instruction.

18. A data processing method, characterized in that It includes: In a text summary model, use the same encoding parameters to encode source texts from different data sources to obtain the semantic encodings of each source text; Respectively obtain the semantics of each source text from the semantic encodings of each source text; Fuse the semantics of each source text according to the influence weights of each source text on the fused semantics to obtain the fused semantics, including: respectively calculating the probability of each candidate word as the semantic decoding field at time t according to the semantics of each source text; according to the influence weights of each source text on the fused semantics, for the first word, calculate the probability of the first word as the model output field at time t according to the probability of taking the first word as the semantic decoding field at time t under each source text, where the first word is any one of the candidate words and t is any output time of the text summary model; Generate text summary information according to the fused semantics.

19. The method according to claim 18, wherein The encoding of source texts from different data sources includes: Use an attention mechanism to encode source texts from different data sources.

20. The method according to claim 18, characterized in that, The generation of text summary information according to the fused semantics includes: Take the candidate word with the highest probability as the model output field at time t as the output field at time t to generate the text summary information.

21. The method according to claim 18, wherein It also includes: Obtain the background knowledge text corresponding to the source texts from different data sources, and the background knowledge text contains key attribute information of the target object; Extract the semantics of the background knowledge text; The generation of text summary information according to the fused semantics includes: Semantically fuse the semantics of the fused semantics and the background knowledge text to generate text summary information.

22. The method according to claim 21, wherein The semantically fusing the semantics of the fused semantics and the background knowledge text to generate text summary information includes: Calculate the probability of each candidate word as the semantic fusion field of the source text at time t based on the semantics of each source text and the influence weight of each source text on the fused semantics. Determine the probability of each candidate word as the semantic decoding field of the background knowledge text at time t according to the background knowledge text. According to the respective fusion weights of the source text set and the background knowledge text, calculate the probability of each candidate word as the model output field at time t based on the probability of each candidate word as the semantic fusion field of the source text at time t and the probability of each candidate word as the semantic decoding field of the background knowledge text at time t. Take the word with the highest probability as the output field at time t as the output field at time t to generate the text summary information. Where t is any output time.

23. A method for processing an entity object, characterized in that, Applied to scenarios where description information needs to be created for an object, including: Determine multiple network data sources associated with the entity object for which description information needs to be created, where the network data sources include web content and at least one data platform providing services related to the entity object. Obtain the text data set and picture data set of the entity object from the multiple network data sources. In the text summary model, perform semantic recognition on the text data from different network data sources to obtain the semantics of each text data. Fuse the semantics of each text data according to the influence weight of each text data on the fused semantics to obtain the fused semantics, including: respectively calculate the probability of each candidate word as the semantic decoding field at time t under each text data according to the semantics of each text data; according to the influence weight of each text data on the fused semantics, for the first word, calculate the probability of the first word as the model output field at time t according to the probability of taking the first word as the semantic decoding field at time t under each text data, where the first word is any one of the candidate words, and t is any output time of the text summary model. Generate the key text data of the entity object according to the fused semantics. Screen the key picture data matching the key text data from the picture data set. Generate the description content of the entity object according to the key text data and the key picture data, where the description content includes video data.

24. A method for providing an entity object, characterized in that, Applied to scenarios where description information needs to be created for an object, including: Obtain the text data set and picture data set of the entity object for which description information needs to be created from the web page and the first data platform. In the text summary model, perform semantic recognition on the text data set obtained from the web page and the first data platform to obtain the semantics of each text data. Fuse the semantics of each text data according to the influence weight of each text data on the fused semantics to obtain the fused semantics, including: calculate the probability of each candidate word as the semantic decoding field at time t under each text data according to the semantics of each text data; according to the influence weight of each text data on the fused semantics, for the first word, calculate the probability of the first word as the model output field at time t according to the probability of taking the first word as the semantic decoding field at time t under each text data, where the first word is any one of the candidate words, and t is any output time of the text summarization model; Generate the key text data of the entity object according to the fused semantics; Screen the key picture data matching the key text data from the picture dataset; Generate the description content of the entity object according to the key text data and the key picture data; Based on the search request for the entity object, display the entity object on the client of the second data platform and provide the description content at the associated position of the entity object, and the description content includes video.

25. The method according to claim 24, wherein The semantic recognition of the text data obtained from the web page and the first data platform to obtain the semantics of each text data includes: Use the same encoding parameters to encode the text data obtained from the web page and the first data platform to obtain the semantic encoding of each text data; Respectively obtain the semantics of each text data from the semantic encoding of each text data.

26. The method according to claim 25, wherein The text summarization model includes a neural network model, and the encoding of the text data obtained from the web page and the first data platform includes: Use the attention mechanism of the neural network model to encode the text data obtained from the web page and the first data platform.

27. The method according to claim 24, wherein It also includes: Input the background knowledge text of the entity object into the text summarization model, and the background knowledge text contains the key attribute information of the entity object; In the text summarization model, perform semantic recognition on the background knowledge text to obtain the semantics of the background knowledge text; The generating the key text data of the entity object according to the fused semantics includes: Perform semantic fusion on the fused semantics corresponding to the text dataset and the semantics of the background knowledge text to generate the key text data of the entity object.

28. The method according to claim 27, wherein The performing semantic fusion on the fused semantics corresponding to the text dataset and the semantics of the background knowledge text to generate the key text data of the entity object includes: Calculate the probability of each candidate word as the semantic fusion field of the text data at time t based on the semantics of each text data and the influence weight of each text data on the fused semantics; Determine the probability of each candidate word as the semantic decoding field of the background knowledge text at time t according to the background knowledge text; According to the respective fusion weights corresponding to the text dataset and the background knowledge text, calculate the probability of each candidate word as the model output field at time t according to the probability of each candidate word as the semantic fusion field of the text data at time t and the probability of being the semantic decoding field of the background knowledge text at time t; Take the word with the highest probability among the model output fields at time t as the output field of the text summary model at time t to generate the key text data; where t is any output time of the text summary model.

29. A computing device, characterized in that, Applied to scenarios where descriptive information needs to be created for an object, including a memory and a processor; The memory is used to store one or more computer instructions; The processor is coupled to the memory and is used to execute the one or more computer instructions for: Respectively obtain the source text set and the source picture set corresponding to the target object for which descriptive information needs to be created from multiple data sources; In the text summary model, perform semantic recognition on the source texts from different data sources to obtain the semantics of each source text; Fuse the semantics of each source text according to the influence weight of each source text on the fused semantics to obtain the fused semantics, including: respectively calculate the probability of each candidate word as the semantic decoding field at time t under each source text according to the semantics of each source text; according to the influence weight of each source text on the fused semantics, for the first word, calculate the probability of the first word as the model output field at time t according to the probability of taking the first word as the semantic decoding field at time t under each source text, where the first word is any one of the candidate words, and t is any output time of the text summary model; Generate the text summary information of the target object according to the fused semantics; Generate the description content of the target object according to the text summary information and the source picture set.

30. A computing device, characterized in that, Including a memory and a processor; The memory is used to store one or more computer instructions; The processor is coupled to the memory and is used to execute the one or more computer instructions for: In the text summary model, use the same encoding parameters to encode the source texts from different data sources to obtain the semantic encodings of each source text; Respectively obtain the semantics of each source text from the semantic encodings of each source text; Fuse the semantics of each source text according to the influence weight of each source text on the fused semantics to obtain the fused semantics, including: respectively calculate the probability of each candidate word as the semantic decoding field at time t under each source text according to the semantics of each source text; according to the influence weight of each source text on the fused semantics, for the first word, calculate the probability of the first word as the model output field at time t according to the probability of taking the first word as the semantic decoding field at time t under each source text, where the first word is any one of the candidate words, and t is any output time of the text summary model; Generate text summary information according to the fused semantics.

31. A computing device, characterized in that, Applied to scenarios where descriptive information needs to be created for an object, including a memory and a processor; The memory is used to store one or more computer instructions; The processor is coupled to the memory and is used to execute the one or more computer instructions for: Determine multiple network data sources associated with the entity object for which descriptive information needs to be created, and the network data sources include web page content and at least one data platform providing services related to the entity object; Obtain a text data set and a picture data set of an entity object from the multiple network data sources; In a text summarization model, perform semantic recognition on text data from different network data sources to obtain the semantics of each text data; Fuse the semantics of each text data according to the influence weight of each text data on the fused semantics to obtain the fused semantics, including: respectively calculate the probability of each candidate word as the semantic decoding field at time t according to the semantics of each text data; according to the influence weight of each text data on the fused semantics, for the first word, calculate the probability of the first word as the model output field at time t according to the probability of taking the first word as the semantic decoding field at time t under each text data, where the first word is any one of the candidate words, and t is any output time of the text summarization model; Generate key text data of the entity object according to the fused semantics; Screen key picture data matching the key text data from the picture data set; Generate a description content of the entity object according to the key text data and the key picture data, and the description content includes video data.

32. A computing device, characterized in that, Applied to scenarios that need to create description information for an object, including a memory and a processor; The memory is used to store one or more computer instructions; The processor is coupled with the memory and is used to execute the one or more computer instructions for: Obtain a text data set and a picture data set of an entity object that needs to create description information from a web page and a first data platform; In a text summarization model, perform semantic recognition on the text data set obtained from the web page and the first data platform to obtain the semantics of each text data; Fuse the semantics of each text data according to the influence weight of each text data on the fused semantics to obtain the fused semantics, including: respectively calculate the probability of each candidate word as the semantic decoding field at time t according to the semantics of each text data; according to the influence weight of each text data on the fused semantics, for the first word, calculate the probability of the first word as the model output field at time t according to the probability of taking the first word as the semantic decoding field at time t under each text data, where the first word is any one of the candidate words, and t is any output time of the text summarization model; Generate key text data of the entity object according to the fused semantics; Screen key picture data matching the key text data from the picture data set; Generate a description content of the entity object according to the key text data and the key picture data; Based on a search request for an entity object, display the entity object on the client of a second data platform and provide description content at an associated position of the entity object, and the description content includes video.

33. A computer-readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed by one or more processors, cause the one or more processors to execute the data processing method described in any one of claims 1-22, the processing method of the entity object described in claim 23, or the providing method of the entity object described in any one of claims 24-28.

Images