Real-time news summary data processing method and device based on large model

Abstract

The present invention provides a real-time news summary data processing method and device based on a large model, which relates to the technical field of real-time news summary processing, including: first verifying the authenticity of each real-time news event, and determining the summary summary and detail summary of each real-time news event that passes the first verification at the same time; performing a second verification on the summary summary and detail summary of each real-time news event respectively with six elements of news as constraints, and inputting the summary summary and detail summary of each real-time news event that passes the verification into a multimodal large model, so that the multimodal large model determines a target real-time news event from the real-time news events based on request information, and outputs a response to the request information according to the summary summary or detail summary of the target real-time news event; so as to alleviate the technical problem of low reliability of the response to the real-time news event.

Description

Technical Field

[0001] The present invention relates to the technical field of real-time news summary processing, and in particular to a real-time news summary data processing method and device based on a large model. Background Art

[0002] With the emergence of multimodal large models (multimodal large language models), AI can communicate smoothly with people and has shown leaps and bounds in semantic understanding and logical reasoning.

[0003] The current multimodal large model generally obtains answers to user questions through summary information of real-time news; the extracted summary information is only slightly consistent with the current request. For real-time news responses that rely on summary information, the accuracy of the summary information of real-time news events will affect the reliability of the answer results. Summary of the invention

[0004] The purpose of the present invention is to provide a real-time news summary data processing method and device based on a large model to alleviate the technical problem of low reliability of responses to real-time news events.

[0005] In a first aspect, an embodiment of the present invention provides a real-time news summary data processing method based on a large model, comprising:

[0006] Responding to the request information for the real-time news event, obtaining at least one real-time news event;

[0007] Performing a first verification on the authenticity of each of the real-time news events, and eliminating the real-time news events that do not meet the authenticity requirements;

[0008] For the real-time news events that have passed the first verification, simultaneously determining a summary summary and a detailed summary of each real-time news event that has passed the first verification; wherein each real-time news event that has passed the first verification includes a summary summary and a plurality of detailed summaries;

[0009] Taking the six elements of news as constraints, a second verification is performed on the summary and detail summary of each real-time news event that passes the first verification, and the summary and detail summary of each real-time news event that meets the authenticity requirements are input into the multimodal large model;

[0010] The request information is input into the multimodal big model, a target real-time news event is determined from the real-time news events based on the request information, and a response to the request information is output according to a general summary or a detailed summary of the target real-time news event based on the request information.

[0011] In combination with the first aspect, an embodiment of the present invention provides a first possible implementation of the first aspect, wherein, for the real-time news event that has passed the first verification, the step of simultaneously determining a summary summary and a detailed summary of each real-time news event that has passed the first verification includes:

[0012] While extracting entity key information from each real-time news event that has passed the first verification to generate a summary, each real-time news event that has passed the first verification is segmented to generate multiple detail summaries corresponding to each real-time news event that has passed the first verification.

[0013] In combination with the first aspect, an embodiment of the present invention provides a second possible implementation of the first aspect, wherein the step of segmenting each of the real-time news events that pass the first verification and generating a plurality of detail summaries corresponding to each of the real-time news events that pass the first verification includes:

[0014] Performing a first segmentation on each of the real-time news events that pass the first verification according to a preset number of characters or paragraphs to obtain an initial summary;

[0015] Performing semantic recognition on the initial abstracts to obtain at least one semantic topic corresponding to each initial abstract;

[0016] A second segmentation is performed on each of the initial summaries according to the number of the semantic topics to obtain a plurality of detail summaries of each real-time news event that passes the first verification; wherein each detail summary includes a semantic topic.

[0017] In combination with the first aspect, the embodiment of the present invention provides a third possible implementation of the first aspect, wherein the steps of performing a second verification on the summary and detail summary of each real-time news event that passes the first verification respectively with the six elements of news as constraints, and inputting the summary and detail summary of each real-time news event that meets the authenticity requirements into the multimodal large model include:

[0018] Determining the completeness of the summary of each real-time news event according to the types of the six news elements corresponding to the summary of each real-time news event that passes the first verification;

[0019] Obtaining and summing up the types of the six news elements respectively corresponding to each detail summary of each real-time news event that passes the first verification, and determining the completeness of the detail summary corresponding to each real-time news event;

[0020] The completeness is compared with a preset completeness threshold, real-time news events corresponding to summary summaries or detail summaries that do not meet the requirements are eliminated, and the summary summaries and detail summaries corresponding to the remaining real-time news events are input into the multimodal large model.

[0021] In combination with the first aspect, the embodiment of the present invention provides a fourth possible implementation of the first aspect, wherein the steps of performing a second verification on the summary and detail summary of each real-time news event that passes the first verification respectively with the six elements of news as constraints, and inputting the summary and detail summary of each real-time news event that meets the authenticity requirements into the multimodal large model include:

[0022] Determining the completeness of the summary of each real-time news event according to the types of the six news elements corresponding to the summary of each real-time news event that passes the first verification;

[0023] Obtaining and summing up the types of the six news elements respectively corresponding to each detail summary of each real-time news event that passes the first verification, and determining the completeness of the detail summary corresponding to each real-time news event;

[0024] If the completeness does not meet the preset completeness threshold, the summary and / or detail summary that does not meet the requirements is expanded by adding news elements so that the expanded summary and / or detail summary meets the completeness requirements; wherein the news elements are obtained from the summary and detail summary corresponding to the real-time news event that meets the completeness requirements;

[0025] The summary and detail summaries of each real-time news event that meets the completeness requirements are input into the multimodal big model.

[0026] In combination with the first aspect, an embodiment of the present invention provides a fifth possible implementation of the first aspect, wherein the step of inputting the request information into the multimodal macro model and determining a target real-time news event from the real-time news events based on the request information includes:

[0027] Performing semantic recognition on the input request information based on the multimodal large model to determine the target domain corresponding to the request information;

[0028] Performing word segmentation processing on the summary summary and the detail summary of each real-time news event input into the multimodal big model to obtain the word segmentation type of each real-time news event input into the multimodal big model; wherein the word segmentation type includes a subject, an object and a qualifier, and the qualifier is located before the subject and the object;

[0029] Determining the real domain of the subject and / or object defined by the qualifier according to the category of the qualifier;

[0030] Determine the real domain corresponding to each real-time news event according to the distribution of the real domains of the subject and / or the object in the summary and detail summaries of each real-time news event;

[0031] The target domain of the request information is matched with the real domain corresponding to each real-time news event, and the target real-time news event matching the request information is determined from the real-time news events input into the multimodal big model.

[0032] In combination with the first aspect, an embodiment of the present invention provides a sixth possible implementation of the first aspect, wherein the step of outputting a response to the request information according to a summary or a detailed summary of the target real-time news event based on the request information includes:

[0033] Performing semantic recognition on the input request information based on the multimodal large model to determine whether the request information contains a first keyword for representing a general intent and a second keyword for representing a detailed intent;

[0034] If only the first keyword exists, selecting a summary of the target real-time news event to output a response to the request information;

[0035] If only the second keyword exists, selecting the detail summary of the target real-time news event and outputting a response to the request information;

[0036] If the first keyword and the second keyword exist, a summary of the target real-time news event is selected to output a response to the request information.

[0037] In combination with the first aspect, the embodiment of the present invention provides a seventh possible implementation manner of the first aspect, wherein the step of first verifying the authenticity of each of the real-time news events and eliminating the real-time news events that do not meet the authenticity requirements includes:

[0038] Determining the authenticity score of each real-time news event according to the preset verification weight of each news element and the completeness of each news element in each real-time news event;

[0039] Comparing the authenticity score of each of the real-time news events with a preset score threshold to perform a first verification of the authenticity of the real-time news event;

[0040] The real-time news events whose authenticity scores are less than the preset score threshold are eliminated.

[0041] In combination with the first aspect, the embodiment of the present invention provides an eighth possible implementation of the first aspect, wherein, in response to request information for a real-time news event, the step of acquiring at least one real-time news event includes:

[0042] Receiving request information input by a user in a graphical user interface;

[0043] According to the request information, control the real-time search engine to crawl target content whose matching degree with the request information reaches a preset matching threshold value and the release time of the target content from at least one third-party website;

[0044] At least one real-time news event is determined according to the target content whose release time meets the preset time range requirement.

[0045] In a second aspect, an embodiment of the present invention further provides a real-time news summary data processing device based on a large model, comprising:

[0046] An acquisition module, in response to request information for real-time news events, acquires at least one real-time news event;

[0047] A first verification module performs a first verification on the authenticity of each of the real-time news events, and removes the real-time news events that do not meet the authenticity requirements;

[0048] A determination module, for the real-time news events that have passed the first verification, simultaneously determines a summary summary and a detailed summary of each real-time news event that has passed the first verification; wherein each real-time news event that has passed the first verification includes a summary summary and a plurality of detailed summaries;

[0049] The second verification module performs a second verification on the summary and detail summary of each real-time news event that passes the first verification, taking the six elements of news as constraints, and inputs the summary and detail summary of each real-time news event that meets the authenticity requirements into the multimodal large model;

[0050] A response module inputs the request information into the multimodal big model, determines a target real-time news event from the real-time news events based on the request information, and outputs a response to the request information according to a summary or a detailed summary of the target real-time news event based on the request information.

[0051] The embodiment of the present invention provides a real-time news summary data processing method and device based on a big model, which matches at least one real-time news event corresponding to the request information from a third party, screens out false news with incomplete elements based on the six elements of news, and simultaneously produces a summary summary and a detail summary for the remaining news, and then verifies the completeness of the summary summary and detail summary of each real-time news event with the six elements of news as constraints; the summary summary and detail summary with complete six elements of news are input into the multimodal big model, and the multimodal big model selects the corresponding summary to respond according to the summary or detail request intention for the real-time news event in the request information to ensure the reliability of the response.

[0052] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description and the drawings.

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0055] Figure 1 A flow chart of a real-time news summary data processing method based on a large model provided by an embodiment of the present invention;

[0056] Figure 2 A schematic diagram of functional modules of a large model-based real-time news summary data processing device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] When the current multimodal big model asks questions about real-time news events, it may get some "outdated answers" that do not match the current real-time news, resulting in poor accuracy in answering questions about real-time news events. For example, how many gold medals did country X win in the Olympics? This request information is real-time during the Olympic Games, and its gold medal count is subject to change.

[0059] In addition, the summary information extracted by the multimodal large model from real-time news is also limited in its closeness to real-time requests, which further reduces the accuracy of responding to request information using the summary information.

[0060] Based on this, an embodiment of the present invention provides a real-time news summary data processing method and device based on a large model. It can screen real-time news similar to the current request, and obtain two types of summary information of the real-time news, namely, detailed and general, from two channels. It can determine appropriate summary information for answering according to specific request information, thereby ensuring the accuracy of the response to the real-time news request.

[0061] To facilitate understanding of this embodiment, a real-time news summary data processing method based on a large model disclosed in an embodiment of the present invention is first introduced in detail. This method can be applied to intelligent devices such as controllers, host computers, servers, and touch-screen interactive terminals.

[0062] Figure 1 A flow chart of a real-time news summary data processing method based on a large model provided by an embodiment of the present invention.

[0063] Reference Figure 1 , the method comprises the following steps:

[0064] Step S102: acquiring at least one real-time news event in response to request information for the real-time news event.

[0065] Among them, the request information can be understood as a request statement for a real-time news event; the request statement may include multiple modal forms, such as voice, text, video, etc.; after converting each modal request statement into text, the statement text may include a first text information for locating the real-time news event, and a second text information for characterizing a corresponding request for the text information; based on the request information, at least one real-time news event can be retrieved accordingly.

[0066] Step S104, the authenticity of each real-time news event is verified for the first time, and the real-time news events that do not meet the authenticity requirements are eliminated.

[0067] It should be noted that the current real-time news events are captured only because they are related to the request information, and the authenticity of such real-time news events themselves cannot be guaranteed. The embodiment of the present invention verifies the news elements corresponding to each real-time news event to determine the authenticity of each real-time news event to be involved in the subsequent steps, and removes those that do not meet the authenticity requirements;

[0068] Among them, each news event includes at most six news elements, such as time, people, place, cause, process, and result; as an optional embodiment, verification rules for each news element and a verification weight corresponding to each news element can be pre-configured, that is, whether each news element is verified or not has a different effect on the authenticity of the real-time news event, and corresponding verification weights and verification rules are set for each news element according to the degree of influence of each news element on the authenticity of the real-time news event.

[0069] Step S106, for the real-time news events that have passed the first verification, a general summary and a detailed summary of each real-time news event that has passed the first verification are determined simultaneously.

[0070] Among them, each real-time news event that passes the first verification includes a general summary and multiple detailed summaries. Here, each real-time news event that passes the first verification, that is, the authenticity meets the requirements, is subjected to dual-path summary generation, generating general summary information and detailed summary information respectively; the general summary can be used to represent the overall overview of the real-time news event, while the detailed summary is to split the real-time news event into several detailed summary information, each detailed summary information only expresses a small link in the real-time news event, and the overall content of the real-time news event cannot be known from a detailed summary.

[0071] Step S108, using the six elements of news as constraints, conduct a second verification on the summary and detail summary of each real-time news event that passes the first verification, and input the summary and detail summary of each real-time news event that meets the authenticity requirements into the multimodal large model.

[0072] Among them, the real-time news events that pass the first verification can determine the news facts that meet the authenticity requirements. In the process of summarizing the news facts in the aforementioned embodiment, if there is an abnormality in the summary segmentation, the generated summary information will have a large error, which will lead to the low accuracy of the subsequent response based on the summary information; therefore, the embodiment of the present invention performs a second verification on each summary with the six elements of news as constraints; that is, if the summary with complete six elements of news meets the second verification requirements, if the summary with incomplete six elements of news cannot participate in the subsequent steps.

[0073] Step S110, inputting the request information into the multimodal big model, determining the target real-time news event from the real-time news events based on the request information, and outputting a response to the request information according to a summary or a detailed summary of the target real-time news event.

[0074] Among them, the target real-time news event is a real-time news event that matches the request information on the basis of satisfying the authenticity of the news event and the accuracy of summary splitting in the aforementioned embodiment; the text content extracted from the request information includes the specific intention of its request, that is, the general intention or the detailed intention, and then selects the corresponding summary for response.

[0075] In a preferred embodiment of actual application, at least one real-time news event corresponding to the request information is matched from a third party, false news with incomplete elements is screened out based on the six elements of news, and a summary summary and a detail summary are simultaneously produced for the remaining news. The completeness of the summary summary and detail summary of each real-time news event is then verified with the six elements of news as constraints; the summary summary and detail summary with complete six elements of news are input into the multimodal big model, and the multimodal big model selects the corresponding summary to respond based on the summary or detail request intention for the real-time news event in the request information to ensure the reliability of the response.

[0076] In some embodiments, in response to request information of multiple modes, text extraction is performed, and at least one real-time news event is obtained based on the extracted text; here, the subject of step S102 includes:

[0077] Step 1.1), receiving the request information input by the user in the graphical user interface.

[0078] Here, the user can input the requested information by interacting with the graphical user interface; for example, the user can input the requested information in the corresponding modal form, such as voice, video, text, etc., in the graphical user interface, or input the requested information in the graphical user interface by single clicking, long pressing, or double clicking a specific mark on the graphical user interface; in actual applications, for multimodal input information, the multimodal large model converts it into text form, performs text extraction, and obtains at least one real-time news event accordingly.

[0079] Step 1.2), based on the request information, control the real-time search engine to crawl target content whose matching degree with the request information reaches a preset matching threshold from at least one third-party website, as well as the release time of the target content.

[0080] It should be noted that the target content crawled from the third-party website according to the request information may include multi-modal forms, such as crawling a video, audio, picture, text, etc. that closely matches the request information from the third-party website;

[0081] Here, the multimodal big model can perform text extraction on the multimodal request information. The first text information and the second text information included in the extracted request information can locate the target content that matches the first text information to the required degree from the website. The multimodal big model then performs text extraction on the multimodal target content. Each text content after the target content is extracted has its corresponding release time.

[0082] For example, the text content corresponding to the request information is how many gold medals country X has won in the Olympics, wherein the first text information used for positioning is "how many gold medals country X has won in the Olympics", and the second text information is used to represent the request corresponding to the real-time news as "how many"; based on the text content of the request information, at least one real-time news event matching "the number of gold medals country X has won in the Olympics" may be crawled, and the multimodal large model then performs text extraction on each real-time news event to obtain the text content corresponding to each real-time news event and the release time of the text content.

[0083] The crawling rules can be set in advance, and the search results will be filtered through the search engine to obtain the search results that meet the preset time range requirements. For example, the target content of 5 websites that meet the preset time range requirements is retrieved.

[0084] Step 1.3), at least one real-time news event is determined based on the target content whose release time meets the preset time range requirement.

[0085] It is understandable that the news located by "Number of Olympic gold medals of country X" has real-time requirements, that is, the news content in different time periods, that is, the number of gold medals may change. Therefore, in order to ensure the reliability of the response, the embodiment of the present invention determines the real-time news event that matches the request information according to the release time within the preset time range; for example, the target content whose release time is within the current preset time range (one day, one hour, half an hour, etc.) and whose text content extracted by the multimodal large model matches the text content extracted from the request information is selected as the real-time news event.

[0086] On the basis of the above-mentioned embodiment, the determined real-time news event is indeed more relevant to the requested information. However, with the rise of self-media, the real-time news events published on the Internet are mixed with true and false ones, and the accuracy of the response based on the real-time news events with poor authenticity will also be affected; therefore, step S104 of the embodiment of the present invention verifies the authenticity of the real-time news events matched in the above-mentioned embodiment through the integrity of the six elements of news, and ensures the authenticity of the real-time news events participating in the subsequent response steps; specifically, it includes:

[0087] In step 2.1), the authenticity score of each real-time news event is determined based on the preset verification weight of each news element and the completeness of each news element in each real-time news event.

[0088] Here, according to the criticality of each news element or according to the influence of each news element on the authenticity of the news event, a corresponding verification rule is configured for each news element; for example, the more critical the news element is, the higher the completeness requirement is; generally speaking, in order of criticality from high to low, the news elements can be sorted as follows: time, result, person, place, process and cause; that is, the completeness requirement of the news element time is relatively high. Taking the text information corresponding to a real-time news event "The tenth competition day of the A Olympic Games - August 5 - will produce 19 Olympic gold medals" as an example, it can be determined whether there is a time news element in the text information, and the completeness of the time representation in the text information can be verified. The time includes year, month, day, hour, minute and second, and the summary information contains month and day, so the completeness of the time news element can be verified.

[0089] Based on the foregoing embodiment, the verification weight corresponding to the more critical news elements is also higher; the product of the completeness of each news element in the text information corresponding to each real-time news event and its verification weight is added together to obtain the authenticity score of each real-time news event.

[0090] In step 2.2), the authenticity score of each real-time news event is compared with the preset score threshold to perform the first verification of the authenticity of the real-time news event.

[0091] Here, if the authenticity score of the corresponding text information of the real-time news event is higher than the preset score threshold, the real-time news event is deemed to meet the authenticity requirements, and such news event is a news fact; otherwise, the authenticity of the real-time news event does not meet the requirements, and such news event is classified as a news clue, which lacks the completeness of news elements and is likely to be false.

[0092] In step 2.3), real-time news events whose authenticity scores are less than the preset score threshold are eliminated.

[0093] News clues that may be false are eliminated from real-time news events to ensure the authenticity of real-time news events involved in subsequent steps, that is, real-time news events involved in subsequent steps are all news facts.

[0094] In some embodiments, step S106 may simultaneously determine two types of summary information corresponding to the real-time news event through two paths, namely, a detailed summary and a general summary, specifically including:

[0095] In step 3.1), while extracting entity key information from each real-time news event that has passed the first verification to generate a summary, each real-time news event that has passed the first verification is segmented to generate multiple detail summaries corresponding to each real-time news event that has passed the first verification.

[0096] Here, the text content after the multimodal big model performs text extraction on the real-time news event is subjected to entity key information extraction operation; wherein, entity key information can be understood as entity information with actual meanings such as subject, predicate, object, qualifier, etc., and the information after the undisputed auxiliary words, conjunctions, etc. are removed from the text content; it should be noted that while generating the summary, the text content after the multimodal big model performs text extraction on the real-time news event is also segmented to generate a detailed summary;

[0097] For example, the text content can be segmented and detailed summaries can be generated by the following steps:

[0098] Step 3.1.1), each real-time news event that passes the first verification is first segmented according to a preset number of characters or paragraphs to obtain an initial summary.

[0099] Among them, the preset number of characters can be pre-set to 5 characters, 10 characters, 20 characters, 30 characters, 50 characters, etc. as the segmentation unit, punctuation marks can be used as segmentation trigger marks or a single paragraph can be used as a segmentation unit, and the text content after the multimodal large model extracts text from real-time news events is first segmented to obtain multiple initial summaries.

[0100] In step 3.1.2), semantic recognition is performed on the initial abstracts to obtain at least one semantic topic corresponding to each initial abstract.

[0101] Here, by performing semantic recognition on each initial summary, the semantic topic corresponding to each initial summary can be determined; wherein the semantic topic can be understood as the expressed views, intentions and topics in each initial summary; each initial summary may correspond to multiple semantic topics.

[0102] For example, if the initial summary after the first split is "The Sports Bureau of A held a press conference on the same day to introduce the details of the delegation's visit to A. According to the introduction, the delegation will depart from B to A on the 31st. During the visit to A, they will attend a welcome event hosted by the B government, interact with athletes and young people in A, participate in large-scale galas, visit landmark attractions, and visit the community, so that citizens of A will have the opportunity to appreciate the Olympic athletes up close";

[0103] After the second split, we can get semantic theme 1: A Sports Bureau held a press conference on the same day to introduce the details of the delegation's visit to A. Semantic theme 2: The delegation will depart from B to A on the 31st; semantic theme 3: During the visit to A on the 31st, they will attend the welcoming event hosted by B; semantic theme 4: Interact with A athletes and young people on the 31st; semantic theme 5: Participate in a large-scale gala on the 31st; semantic theme 6: Visit landmark attractions on the 31st; semantic theme 7: Visit the community on the 31st; semantic theme 8: Let A citizens have the opportunity to appreciate the Olympic athletes up close.

[0104] Step 3.1.3), performing a second segmentation on each initial summary according to the number of semantic topics, and obtaining multiple detailed summaries of each real-time news event that passes the first verification;

[0105] Among them, based on the above-mentioned embodiment, a second segmentation is performed on each initial summary, so that each detail summary after segmentation only includes one semantic topic, and multiple detail summaries belonging to the same real-time news event are generated.

[0106] For example, eight detailed summaries belonging to the same real-time news event may be generated based on the aforementioned eight semantic topics.

[0107] As an optional embodiment, step S108 can be implemented by performing a second verification on the summary summary and detail summary generated by the real-time news event that have passed the first verification, and selecting the summary generation operation that is correct, that is, the generated summary that also meets the integrity requirements of the six elements of news to participate in the subsequent steps, which can be specifically implemented by the following steps, including:

[0108] Step 4.1), according to the types of the six news elements corresponding to the summary of each real-time news event that passes the first verification, determine the completeness of the summary of each real-time news event.

[0109] Here, the six elements of news are searched from the summary respectively to determine the completeness of the summary corresponding to each real-time news event. For example, the number of news element types included in the summary is calculated by ratio with the six news elements to obtain the completeness of the summary.

[0110] Step 4.2), obtain the types of the six news elements corresponding to each detail summary of each real-time news event that passes the first verification and add them up to determine the completeness of the detail summary corresponding to each real-time news event.

[0111] It should be noted that each real-time news event may include multiple detail summaries. If the complete six news elements can be found in all the detail summaries included in each real-time news event, the detail summary corresponding to the real-time news event is considered complete. In actual applications, the ratio of the sum of the number of news element types corresponding to all the detail summaries contained in each real-time news event to the six news elements is calculated to obtain the completeness of the detail summary.

[0112] In step 4.3), the completeness is compared with the preset completeness threshold, and the real-time news events corresponding to the summary or detail summary that do not meet the requirements are eliminated, and the summary and detail summary corresponding to the remaining real-time news events are input into the multimodal large model.

[0113] If the number of news elements pieced together from the summary of all the details of the current real-time news event does not meet the requirements, the current real-time news event may have an erroneous operation in the summary generation process, resulting in the loss of one or some news elements. At this time, the current real-time news event that does not meet the requirements cannot participate in the subsequent response steps to ensure the accuracy of the response.

[0114] As another optional embodiment, step S108 can be performed by performing a second verification on the summary summary and detail summary generated by the real-time news event that passed the first verification, and expanding and compensating the summary corresponding to the real-time news event where there is an error in the summary generation operation, that is, the generated summary does not meet the requirements of the integrity of the six elements of news, so that the summary of the real-time news event after expansion and compensation can meet the requirements of the integrity of the six elements of news, and then participate in the subsequent steps, which can be specifically implemented by the following steps, including:

[0115] Step 5.1), according to the types of the six news elements corresponding to the summary of each real-time news event that passes the first verification, determine the completeness of the summary of each real-time news event.

[0116] Here, the six elements of news are searched from the summary respectively to determine the completeness of the summary corresponding to each real-time news event. For example, the number of news element types included in the summary is calculated by ratio with the six news elements to obtain the completeness of the summary.

[0117] Step 5.2), obtain the types of the six news elements corresponding to each detail summary of each real-time news event that passes the first verification and add them up to determine the completeness of the detail summary corresponding to each real-time news event.

[0118] It should be noted that each real-time news event may include multiple detail summaries. If the complete six news elements can be found in all the detail summaries included in each real-time news event, the detail summary corresponding to the real-time news event is considered complete. In actual applications, the ratio of the sum of the number of news element types corresponding to all the detail summaries contained in each real-time news event to the six news elements is calculated to obtain the completeness of the detail summary.

[0119] In step 5.3), if the completeness does not meet the preset completeness threshold, the summary and / or detailed summary that does not meet the requirements is expanded by adding news elements so that the expanded summary and / or detailed summary meets the completeness requirements.

[0120] If the number of news elements pieced together from the summary of all details of the current real-time news event does not meet the requirement, the current real-time news event may have an error in the summary generation process, resulting in the loss of one or some news elements. In this case, the summary of the current real-time news event that does not meet the requirement and does not meet the completeness requirement is expanded;

[0121] Here, the news elements to be expanded are determined according to the types of news elements that do not meet the complete requirements in each real-time news event; for example, the types of news elements that do not meet the complete requirements in the current real-time news event are location and process, and the news elements to be expanded are determined to be location and process, and the summary of the current real-time news event after expansion meets the requirements of the completeness of the six news elements;

[0122] The added news elements are obtained from the summary and detail summaries corresponding to the real-time news events that meet the completeness requirements.

[0123] In step 5.4), the summary and detail summaries of each real-time news event that meets the completeness requirements are input into the multimodal large model.

[0124] Here, the dual-path summaries corresponding to each real-time news event that meets the completeness requirements are input together so that the multimodal large model can make corresponding responses.

[0125] In some embodiments, in step S110, the multimodal big model selects a target real-time news event that best matches the request information from at least one real-time news event determined to be input in the aforementioned embodiments based on the input request information, including:

[0126] In step 6.1, semantic recognition is performed on the input request information based on the multimodal large model to determine the target domain corresponding to the request information.

[0127] The target domain can be understood as the domain to which the real-time news event targeted by the request information belongs. The multimodal large model can obtain the real corresponding real-time news event domain by semantically identifying the corresponding text content after extracting the request information. However, in actual applications, the capture target of the real-time news event is determined based on the first text information corresponding to the request information, so there may be real-time news events that do not match the target domain.

[0128] In step 6.2, the summary and detail summaries of each real-time news event input into the multimodal big model are segmented to obtain the segmentation type of each real-time news event input into the multimodal big model.

[0129] Among them, the types of segmentation include subject, object and determiner, and the determiner is placed before the subject and object. It can be understood that the multimodal large model has the ability of segmentation and semantic recognition.

[0130] Step 6.3), based on the category of the qualifier, determine the real domain of the subject and / or object defined by the qualifier.

[0131] Here, according to the category to which the qualifier belongs, the real domain of the subject and / or object it qualifies can be obtained, that is, the real domain corresponding to the abstract.

[0132] Step 6.4), according to the distribution of the real domains of the subject and / or object in the summary and detail summary of each real-time news event, determine the real domain corresponding to each real-time news event.

[0133] Among them, each real-time news event includes multiple summaries, and the detailed summaries also include multiple ones, then each summary may correspond to different real domains according to its qualifier category; here, according to the corresponding number of real domains of each summary in each real-time news event, the real domain corresponding to each real-time news event is determined; for example, the real domain with the largest number of corresponding real domains of each summary in each real-time news event is used as the real domain corresponding to each real-time news event.

[0134] In step 6.5), the target domain of the requested information is matched with the real domain corresponding to each real-time news event, and the target real-time news event matching the requested information is determined from the real-time news events input into the multimodal large model.

[0135] Since the first text information after the text content is extracted from the request information may point to multiple news types or multiple application scenario types at the same time, the embodiment of the present invention can determine the target real-time news event corresponding to the target news type / target scenario / real field through the category of the qualifier in the summary information;

[0136] For example, real-time news events may include other activities or competitions where gold medals were won. The gold medals won in the Olympic Games to which they belong can be determined by the qualifier Olympic Games before the name of the gold medal in the summary information "August 5, the tenth competition day of the Olympic Games A, will produce 19 Olympic gold medals", rather than the gold medals won in the game scene. That is, the summary information of other real-time news events where gold medals were won in the game scene will be removed;

[0137] For another example, "Xiao Li" may exist in multiple fields, such as celebrity artist Xiao Li, athlete Xiao Li, etc. If the information requested is "How many gold medals did Xiao Li win in the Olympics", based on the summary information "Xiao Li, a diving athlete, won gold medals in the Olympics", it can be known from the qualifiers before the subject and object that Xiao Li, who participated in the Olympics and won the gold medal, is a diving athlete. Therefore, the summary information of news events such as celebrity artists who also won gold medals but whose winners are Xiao Li are eliminated.

[0138] On the basis of the above-mentioned embodiment, step S110 may also select a corresponding summary or detailed summary of the target real-time news event to answer according to the intention of the request information, so as to further improve the accuracy of the answer, including:

[0139] In step 7.1, semantic recognition is performed on the input request information based on the multimodal large model to determine whether the request information contains a first keyword for representing the general intent and a second keyword for representing the detailed intent.

[0140] For example, "What is the significance of winning the Olympic Games?", "What is the significance of protecting the ecological environment?", "What is the summary of the course of this Olympic Games?"; it can be seen that "significance" and "summary" are the first keywords that represent the intention of summarization. It can be seen that such requests are seeking a more general response, and the use of a summary can make the response result more complete;

[0141] For example, "How many gold medals does Country X have in the Olympics?" and "When does the Olympic diving competition start?" It can be seen that "quantity" and "time" are the second keywords that represent the intention of details. It can be seen that such requests are seeking a more detailed response. The application of detail summaries can make the response to detail requests more accurate.

[0142] Step 7.2), if only the first keyword exists, a summary of the target real-time news event is selected to output a response to the request information.

[0143] As described in the above embodiment, it is possible to select a summary summary for response based on the first keyword present in the request information.

[0144] Step 7.3), if only the second keyword exists, a detailed summary of the target real-time news event is selected to output a response to the requested information.

[0145] As described in the above embodiment, it is possible to select a detailed summary for response based on the second keyword present in the request information.

[0146] In step 7.4), if the first keyword and the second keyword exist, a summary of the target real-time news event is selected to output a response to the request information.

[0147] In actual application, if there are keywords with both general intentions and detailed intentions, choose the relatively more complete summary to answer the question in order to ensure the accuracy of the answer result.

[0148] It should be noted that if a detailed summary is used to respond to a general request, the detailed summary is obtained after the real-time news event is carefully split, and there may be omissions or deletions of the real-time news event, which makes the general response incomplete; if a general summary is used to respond to a detailed request, it is necessary to search the general summary for the detailed request, and the time to respond will be relatively longer; therefore, the embodiment of the present invention can ensure the accuracy of the response while ensuring a shorter response time.

[0149] In some embodiments, Figure 2 As shown, the embodiment of the present invention provides a real-time news summary data processing device based on a large model, comprising:

[0150] An acquisition module, in response to request information for real-time news events, acquires at least one real-time news event;

[0151] A first verification module performs a first verification on the authenticity of each of the real-time news events, and removes the real-time news events that do not meet the authenticity requirements;

[0152] A determination module, for the real-time news events that have passed the first verification, simultaneously determines a summary summary and a detailed summary of each real-time news event that has passed the first verification; wherein each real-time news event that has passed the first verification includes a summary summary and a plurality of detailed summaries;

[0153] The second verification module performs a second verification on the summary and detail summary of each real-time news event that passes the first verification, taking the six elements of news as constraints, and inputs the summary and detail summary of each real-time news event that meets the authenticity requirements into the multimodal large model;

[0154] A response module inputs the request information into the multimodal big model, determines a target real-time news event from the real-time news events based on the request information, and outputs a response to the request information according to a summary or a detailed summary of the target real-time news event based on the request information.

[0155] Furthermore, the determination module is also specifically used to extract entity key information from each real-time news event that has passed the first verification to generate a summary, and at the same time, segment each real-time news event that has passed the first verification to generate multiple detail summaries corresponding to each real-time news event that has passed the first verification.

[0156] Furthermore, the determination module is also specifically used to perform a first segmentation on each of the real-time news events that have passed the first verification according to a preset number of characters or paragraphs to obtain an initial summary; perform semantic recognition on the initial summary to obtain at least one semantic topic corresponding to each initial summary; perform a second segmentation on each of the initial summaries according to the number of the semantic topics to obtain multiple detail summaries of each of the real-time news events that have passed the first verification; wherein each detail summary includes a semantic topic.

[0157] Furthermore, the second verification module is also specifically used to determine the completeness of the summary of each real-time news event based on the types of the six news elements corresponding to the summary of each real-time news event that has passed the first verification; obtain the types of the six news elements corresponding to each detail summary of each real-time news event that has passed the first verification and add them up to determine the completeness of the detail summary corresponding to each real-time news event; compare the completeness with a preset completeness threshold, eliminate the real-time news events corresponding to the summary or detail summary that do not meet the requirements, and input the summary and detail summaries corresponding to the remaining real-time news events into the multimodal large model.

[0158] Furthermore, the second verification module is also specifically used to determine the completeness of the summary of each real-time news event based on the types of the six news elements corresponding to the summary of each real-time news event that has passed the first verification; obtain the types of the six news elements corresponding to each detail summary of each real-time news event that has passed the first verification and add them up to determine the completeness of the detail summary corresponding to each real-time news event; if the completeness does not meet the preset completeness threshold, expand the summary and / or detail summary that does not meet the requirements by adding news elements, so that the expanded summary and / or detail summary meets the completeness requirements; wherein the news elements are obtained from the summary and detail summary corresponding to the real-time news event that meets the completeness requirements; and input the summary and detail summary of each real-time news event that meets the completeness requirements into the multimodal large model.

[0159] Furthermore, the response module is also specifically used to perform semantic recognition on the input request information based on the multimodal big model to determine the target domain corresponding to the request information; perform word segmentation processing on the summary and detail summary of each real-time news event input into the multimodal big model to obtain the word segmentation type of each real-time news event input into the multimodal big model; wherein the word segmentation type includes subject, object and qualifier, and the qualifier is located before the subject and the object; according to the category of the qualifier, determine the real domain of the subject and / or object defined by the qualifier; according to the distribution of the real domain of the subject and / or object in the summary and detail summary of each real-time news event, determine the real domain corresponding to each real-time news event; match the target domain of the request information with the real domain corresponding to each real-time news event, and determine the target real-time news event matching the request information from the real-time news events input into the multimodal big model.

[0160] Furthermore, the response module is also specifically used to perform semantic recognition on the input request information based on the multimodal large model, and determine whether there is a first keyword for representing the general intention and a second keyword for representing the detailed intention in the request information; if only the first keyword exists, a general summary of the target real-time news event is selected to output a response to the request information; if only the second keyword exists, a detailed summary of the target real-time news event is selected to output a response to the request information; if the first keyword and the second keyword exist, a general summary of the target real-time news event is selected to output a response to the request information.

[0161] Furthermore, the first verification module is also specifically used to determine the authenticity score of each real-time news event based on the preset verification weight of each news element and the completeness of each news element in each real-time news event; compare the authenticity score of each real-time news event with a preset score threshold to perform a first verification on the authenticity of the real-time news event; and eliminate the real-time news events whose authenticity score is less than the preset score threshold.

[0162] Furthermore, the acquisition module is also specifically used to receive request information input by a user in a graphical user interface; based on the request information, control the real-time search engine to crawl target content whose matching degree with the request information reaches a preset matching threshold, as well as the release time of the target content from at least one third-party website; and determine at least one real-time news event based on the target content whose release time meets the preset time range requirement.

[0163] It can be understood that the specific operation methods of each functional module in this embodiment can refer to the detailed description of the corresponding steps in the above method embodiment, and will not be repeated here.

[0164] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0165] In addition, in the description of the embodiments of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, and it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0166] In the description of the present invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0167] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the aforementioned embodiments, those of ordinary skill in the art should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the aforementioned embodiments within the technical scope disclosed by the present invention, or can easily conceive of changes, or make equivalent replacements for some of the technical features therein. Such modifications, changes or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the protection scope of the present invention.

Claims

1. A real-time news summary data processing method based on a large model, characterized in that: include: Responding to the request information for the real-time news event, obtaining at least one real-time news event; Performing a first verification on the authenticity of each of the real-time news events, and eliminating the real-time news events that do not meet the authenticity requirements; For the real-time news events that have passed the first verification, simultaneously determining a summary summary and a detailed summary of each real-time news event that has passed the first verification; wherein each real-time news event that has passed the first verification includes a summary summary and a plurality of detailed summaries; Taking the six elements of news as constraints, a second verification is performed on the summary and detail summary of each real-time news event that passes the first verification, and the summary and detail summary of each real-time news event that meets the authenticity requirements are input into the multimodal large model; Inputting the request information into the multimodal macromodel, determining a target real-time news event from the real-time news events based on the request information, and outputting a response to the request information according to a summary or a detailed summary of the target real-time news event; The steps of performing a second verification on the summary and detail summary of each real-time news event that passes the first verification respectively with the six elements of news as constraints, and inputting the summary and detail summary of each real-time news event that meets the authenticity requirements into the multimodal large model include: Determining the completeness of the summary of each real-time news event according to the types of the six news elements corresponding to the summary of each real-time news event that passes the first verification; Obtaining and summing up the types of the six news elements respectively corresponding to each detail summary of each real-time news event that passes the first verification, and determining the completeness of the detail summary corresponding to each real-time news event; The completeness is compared with a preset completeness threshold, and real-time news events corresponding to summary summaries or detailed summaries that do not meet the requirements are eliminated, and the summary summaries and detailed summaries corresponding to the remaining real-time news events are input into the multimodal macro model; The step of inputting the request information into the multimodal macro model and determining a target real-time news event from the real-time news events based on the request information comprises: Performing semantic recognition on the input request information based on the multimodal large model to determine the target domain corresponding to the request information; Performing word segmentation processing on the summary summary and the detail summary of each real-time news event input into the multimodal big model to obtain the word segmentation type of each real-time news event input into the multimodal big model; wherein the word segmentation type includes a subject, an object and a qualifier, and the qualifier is located before the subject and the object; Determining the real domain of the subject and / or object defined by the qualifier according to the category of the qualifier; Determine the real domain corresponding to each real-time news event according to the distribution of the real domains of the subject and / or the object in the summary and detail summaries of each real-time news event; Matching the target domain of the request information with the real domain corresponding to each real-time news event, and determining the target real-time news event matching the request information from the real-time news events input into the multimodal big model; The six elements of news include time, people, place, cause, process and result.

2. The method according to claim 1, characterized in that For the real-time news events that have passed the first verification, the step of simultaneously determining a summary summary and a detailed summary of each real-time news event that has passed the first verification includes: While extracting entity key information from each real-time news event that has passed the first verification to generate a summary, each real-time news event that has passed the first verification is segmented to generate multiple detail summaries corresponding to each real-time news event that has passed the first verification.

3. The method according to claim 2, characterized in that The step of segmenting each of the real-time news events that pass the first verification to generate a plurality of detail summaries corresponding to each of the real-time news events that pass the first verification comprises: Performing a first segmentation on each of the real-time news events that pass the first verification according to a preset number of characters or paragraphs to obtain an initial summary; Performing semantic recognition on the initial abstracts to obtain at least one semantic topic corresponding to each initial abstract; A second segmentation is performed on each of the initial summaries according to the number of the semantic topics to obtain a plurality of detail summaries of each real-time news event that passes the first verification; wherein each detail summary includes a semantic topic.

4. The method according to claim 1, characterized in that The steps of performing a second verification on the summary and detail summary of each real-time news event that passes the first verification respectively with the six elements of news as constraints, and inputting the summary and detail summary of each real-time news event that meets the authenticity requirements into the multimodal large model include: Determining the completeness of the summary of each real-time news event according to the types of the six news elements corresponding to the summary of each real-time news event that passes the first verification; Obtaining and summing up the types of the six news elements respectively corresponding to each detail summary of each real-time news event that passes the first verification, and determining the completeness of the detail summary corresponding to each real-time news event; If the completeness does not meet the preset completeness threshold, the summary and / or detail summary that does not meet the requirements is expanded by adding news elements so that the expanded summary and / or detail summary meets the completeness requirements; wherein the news elements are obtained from the summary and detail summary corresponding to the real-time news event that meets the completeness requirements; The summary and detail summaries of each real-time news event that meets the completeness requirements are input into the multimodal big model.

5. The method according to claim 1, characterized in that The step of outputting a response to the request information according to the summary or detail summary of the target real-time news event comprises: Performing semantic recognition on the input request information based on the multimodal large model to determine whether the request information contains a first keyword for representing a general intent and a second keyword for representing a detailed intent; If only the first keyword exists, selecting a summary of the target real-time news event to output a response to the request information; If only the second keyword exists, selecting the detail summary of the target real-time news event and outputting a response to the request information; If the first keyword and the second keyword exist, a summary of the target real-time news event is selected to output a response to the request information.

6. The method according to claim 1, characterized in that The steps of first verifying the authenticity of each of the real-time news events and eliminating the real-time news events that do not meet the authenticity requirements include: Determining the authenticity score of each real-time news event according to the preset verification weight of each news element and the completeness of each news element in each real-time news event; Comparing the authenticity score of each of the real-time news events with a preset score threshold to perform a first verification of the authenticity of the real-time news event; The real-time news events whose authenticity scores are less than the preset score threshold are eliminated.

7. The method according to claim 1, characterized in that In response to request information for real-time news events, the step of obtaining at least one real-time news event includes: Receiving request information input by a user in a graphical user interface; According to the request information, control the real-time search engine to crawl target content whose matching degree with the request information reaches a preset matching threshold value and the release time of the target content from at least one third-party website; At least one real-time news event is determined according to the target content whose release time meets the preset time range requirement.

8. A real-time news summary data processing device based on a large model, characterized in that: include: An acquisition module, in response to request information for real-time news events, acquires at least one real-time news event; A first verification module performs a first verification on the authenticity of each of the real-time news events, and removes the real-time news events that do not meet the authenticity requirements; A determination module, for the real-time news events that have passed the first verification, simultaneously determines a summary summary and a detailed summary of each real-time news event that has passed the first verification; wherein each real-time news event that has passed the first verification includes a summary summary and a plurality of detailed summaries; The second verification module performs a second verification on the summary and detail summary of each real-time news event that passes the first verification, taking the six elements of news as constraints, and inputs the summary and detail summary of each real-time news event that meets the authenticity requirements into the multimodal large model; a response module, inputting the request information into the multimodal macro model, determining a target real-time news event from the real-time news events based on the request information, and outputting a response to the request information according to a summary or a detailed summary of the target real-time news event; The second verification module is further used to determine the completeness of the summary of each real-time news event according to the types of the six news elements corresponding to the summary of each real-time news event that has passed the first verification; obtain the types of the six news elements corresponding to each detail summary of each real-time news event that has passed the first verification and add them up to determine the completeness of the detail summary corresponding to each real-time news event; compare the completeness with a preset completeness threshold, eliminate the real-time news events corresponding to the summary or detail summary that do not meet the requirements, and input the summary and detail summary corresponding to the remaining real-time news events into the multimodal large model; The response module is also used to perform semantic recognition on the input request information based on the multimodal big model to determine the target domain corresponding to the request information; perform word segmentation processing on the summary and detail summary of each real-time news event input into the multimodal big model to obtain the word segmentation type of each real-time news event input into the multimodal big model; wherein the word segmentation type includes subject, object and qualifier, and the qualifier is located before the subject and the object; according to the category of the qualifier, determine the real domain of the subject and / or object defined by the qualifier; according to the distribution of the real domain of the subject and / or object in the summary and detail summary of each real-time news event, determine the real domain corresponding to each real-time news event; match the target domain of the request information with the real domain corresponding to each real-time news event, and determine the target real-time news event matching the request information from the real-time news events input into the multimodal big model; The six elements of news include time, people, place, cause, process and result.

Images