Media abstract generation method and device fusing large model and structured semantics

Abstract

The invention discloses a large model and structured semantic fused media abstract generation method and device, and the method comprises the steps: introducing structured semantic information of a source manuscript based on a large model workflow, taking the structured semantic information as a data basis for abstract generation, carrying out the knowledge injection of the structured semantic information, enhancing the semantic understanding capability of a model, and improving the abstract generation efficiency. Meanwhile, the external knowledge of the key information is expanded, so that the generation effect is improved by utilizing external knowledge enhancement; and finally, generating abstracts in batches, and performing key information checking and effect evaluation on the generated abstracts, thereby ensuring the accuracy and objectivity of the generated abstracts. Therefore, according to the method, the accuracy, coherence and information integrity of generating the abstract are greatly improved, so that the method is very suitable for large-scale application and popularization in the field of abstract generation.

Description

Technical Field

[0001] The present invention belongs to the technical field of information processing, and in particular relates to a method and device for generating a media summary by integrating a large model with structured semantics. Background Art

[0002] With the rapid development of the information age, the number of media manuscripts has exploded. How to quickly and accurately extract the core content of the manuscripts and generate high-quality summaries to improve user reading attention and expand the click volume and dissemination influence of the manuscripts has become an important challenge in the field of information processing.

[0003] Among them, traditional summary generation methods mostly rely on keyword extraction or simple sentence compression, which makes it difficult to fully capture the deep semantic structure and key information of the manuscript, especially when dealing with complex and long media manuscripts. The effect is particularly limited, therefore, it will reduce the accuracy of the generated summary; in addition, the latest summary generation methods based on large models, although they have strong semantic understanding and generalization capabilities, but due to the hallucination problem of the large model itself, and the lack of effective use of external knowledge and verification mechanism of generated content, the generated summary often lacks accuracy and objectivity; therefore, based on the aforementioned deficiencies, how to provide a generation method that can generate accurate media summaries has become a problem that needs to be solved urgently. Summary of the invention

[0004] The purpose of the present invention is to provide a method and device for generating a media summary by integrating a large model with structured semantics, so as to solve the problem of poor accuracy of generating summaries in the prior art.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] In a first aspect, a media summary generation method integrating a large model and structured semantics is provided, comprising:

[0007] Acquire text information of a specified media manuscript, and extract keyword information, subject information, entity information, relationship information and event information from the text information, so as to generate semantic structure information of the specified media manuscript based on the keyword information, the subject information, the entity information, the relationship information and the event information;

[0008] Performing internal knowledge enhancement processing on the semantic structure information to obtain enhanced knowledge information;

[0009] Performing external background expansion processing on entity information and relationship information in the semantic structure information to obtain background expansion information;

[0010] Constructing a structured prompt model based on the large model, and generating a structured prompt template based on the structured prompt model, the background extension information and the enhanced knowledge information;

[0011] Generating a plurality of candidate summary information according to the structured prompt template;

[0012] Performing verification processing on each candidate summary information, and sorting each candidate summary information according to the verification processing result to obtain a sorting sequence, wherein the verification processing includes quality verification processing and consistency verification processing with designated media manuscripts;

[0013] Based on the sorting sequence, optimal summary information of the designated media manuscript is determined, and the optimal summary information is output for display.

[0014] Based on the above disclosed content, the present invention first obtains the text information of the specified media manuscript, and then performs coarse-grained and fine-grained semantic characterization from five dimensions: keywords, topics, entities, relationships, and events, thereby constructing a complete manuscript semantic structure information to provide accurate data support for subsequent summary generation; then, a structured prompt model based on a large model is established, and the internal knowledge enhancement of the semantic structure information of the manuscript is performed, and the external knowledge of key information is expanded; then, a structured prompt template can be generated based on the aforementioned information and the structured prompt model; in this way, by Knowledge enhancement can enhance the semantic understanding of manuscripts, and use external resources to expand the background information of key entities and events, which can improve the comprehensiveness and accuracy of summaries. After generating a structured prompt template, the prompt template can be used to generate summaries in batches. At the same time, the present invention is supplemented by quality verification and consistency verification with the original text to evaluate the effect of generating summaries, and sorts the summaries according to the verification results. Based on this, the accuracy and objectivity of the generated summaries can be guaranteed. Finally, based on the sorted summaries, the optimal summary is determined and output for display, thereby completing the summary generation process for the specified media manuscript.

[0015] Through the above design, the present invention introduces the structured semantic information of the source manuscript based on the large model workflow as the data basis for summary generation, and then injects the structured semantic information into the knowledge to enhance the semantic understanding ability of the model, and at the same time expands the external knowledge of key information, thereby using external knowledge enhancement to improve the generation effect; finally, the summaries are generated in batches, and the generated summaries are checked for key information and evaluated for effect, so as to ensure the accuracy and objectivity of the generated summaries; in this way, the present invention greatly improves the accuracy, coherence and information integrity of the generated summaries, and is therefore very suitable for large-scale application and promotion in the field of summary generation.

[0016] In a possible design, keyword information, topic information, entity information, relationship information and event information in the text information are extracted, including:

[0017] Construct a media domain dictionary and a stop dictionary;

[0018] Based on the media domain dictionary and the stop dictionary, and using a keyword extraction model, performing keyword extraction processing on the text information to obtain the keyword information after the keyword extraction processing;

[0019] Using the LDA topic extraction model, performing topic extraction processing on the text information to obtain initial topic information;

[0020] Using a preset media theme tag, the initial theme information is subjected to theme mapping processing, so as to obtain theme information corresponding to the text information after the theme mapping processing;

[0021] The information extraction model is used to extract entity, relationship and event information from the text information to obtain entity information, relationship information and event information corresponding to the text information, wherein the information extraction model includes a yayi-uie model.

[0022] In a possible design, the keyword extraction model includes: a TF-IDF feature extraction unit, a textrank network unit, a YAKE network unit, and a keyword classification recognition unit;

[0023] The keyword extraction process is performed on the text information based on the media domain dictionary and the stop dictionary and using a keyword extraction model to obtain the keyword information after the keyword extraction process, including:

[0024] Based on the media domain dictionary and the stop dictionary, and calling the word segmentation component, the text information is segmented to obtain a plurality of segmented words;

[0025] Using the TF-IDF feature extraction unit, the textrank network unit, and the YAKE network unit, feature extraction processing is performed on each segmentation word to obtain first feature information, second feature information, and third feature information of each segmentation word, respectively, wherein the first feature information of any segmentation word includes the TF-IDF feature value of the any segmentation word, the second feature information of any segmentation word includes the word importance feature value of the any segmentation word, and the third feature information of any segmentation word includes the position feature information and length feature information of the any segmentation word;

[0026] The first feature information, the second feature information and the third feature information of each segmentation word are input into a keyword classification and recognition unit for keyword recognition processing, so as to identify the keywords of the text information from multiple segmentation words and obtain the keyword information, wherein the keyword classification and recognition unit includes a random forest network.

[0027] In a possible design, background extension processing is performed on the entity information and the relationship information in the semantic structure information to obtain background extension information, including:

[0028] Calling a search plug-in to perform background search processing on the entity information and the relationship information to obtain initial extended information;

[0029] Acquire business scenario information of the specified media manuscript, and determine an information source list based on the business scenario information;

[0030] According to the information source list, the initial extended information is filtered to obtain filtered extended information;

[0031] The background extension information is generated based on the screening extension information.

[0032] In a possible design, each candidate summary information is checked, including:

[0033] For any candidate summary information, a consistency check is performed on the candidate summary information and the designated media manuscript to obtain a consistency check result;

[0034] Performing a quality check on any of the candidate summary information to obtain a quality check result;

[0035] A verification result of any candidate summary information is generated according to the quality verification result and the consistency verification result.

[0036] In a possible design, a consistency check is performed on any candidate summary information and the designated media manuscript to obtain a consistency check result, including:

[0037] Acquire original text entity information, original text subject information and original text event information of the specified media manuscript, as well as summary entity information, summary event information and summary subject information of any candidate summary information;

[0038] Using a gradient decision boosting tree, a consistency evaluation process is performed on the original text entity information, the original text topic information, the original text event information, the summary entity information, the summary event information, and the summary topic information, so as to obtain the consistency verification result after the consistency evaluation process;

[0039] Accordingly, performing a quality check on any candidate summary information to obtain a quality check result includes:

[0040] The quality assessment model is used to perform a quality check on any candidate summary information, so as to obtain the quality check result after the quality check, wherein the quality assessment model includes a CONNER tool.

[0041] In a possible design, several candidate summary information are generated according to the structured prompt template, including:

[0042] Based on the summary information generation model and according to the structured prompt template, a number of candidate summary information are generated;

[0043] Wherein, after sorting each candidate summary information according to the verification processing result to obtain a sorting sequence, the method further includes:

[0044] Obtain at least one candidate summary information from the sorted sequence as sample summary information;

[0045] The sample summary information is used to perform feedback optimization processing on the summary information generation model to obtain an optimized summary information generation model, so as to use the optimized summary information generation model to generate candidate summary information for the next designated media manuscript.

[0046] In a possible design, after extracting keyword information, topic information, entity information, relationship information and event information from the text information, the method further includes:

[0047] Outputting the keyword information, topic information, entity information, relationship information and event information to the human-computer interaction terminal, so that the user corresponding to the human-computer interaction terminal can modify the keyword information, topic information, entity information, relationship information and event information to obtain modified keyword information, modified topic information, modified entity information, modified relationship information and modified event information;

[0048] In response to a human-computer interaction operation with the human-computer interaction terminal, obtaining revised keyword information, revised subject information, revised entity information, revised relationship information, and revised event information;

[0049] The keyword extraction model is feedback optimized using the corrected keyword information, the LDA topic extraction model is feedback optimized using the corrected topic information, and the information extraction model is feedback optimized using the corrected entity information, the corrected relationship information and the corrected event information, so as to obtain an optimized keyword extraction model, an optimized LDA topic extraction model and an optimized information extraction model respectively, so as to use the optimized keyword extraction model, the optimized LDA topic extraction model and the optimized information extraction model to generate the keyword information, topic information, entity information, relationship information and event information of the next specified media manuscript.

[0050] In a second aspect, a media summary generation device integrating a large model and structured semantics is provided, comprising:

[0051] A semantic structure information generating unit, used to obtain text information of a specified media manuscript, and extract keyword information, topic information, entity information, relationship information and event information in the text information, so as to generate semantic structure information of the specified media manuscript based on the keyword information, the topic information, the entity information, the relationship information and the event information;

[0052] A knowledge enhancement unit, used for performing internal knowledge enhancement processing on the semantic structure information to obtain enhanced knowledge information;

[0053] An extension unit, used for performing external background extension processing on entity information and relationship information in the semantic structure information to obtain background extension information;

[0054] A structured prompt generating unit, used to construct a structured prompt model based on the large model, and generate a structured prompt template based on the structured prompt model, the background extension information and the enhanced knowledge information;

[0055] A summary generating unit, used for generating a plurality of candidate summary information according to the structured prompt template;

[0056] A summary generation unit is used to check and process each candidate summary information, and sort each candidate summary information according to the check and process result to obtain a sorting sequence, wherein the check and process includes quality check and consistency check with designated media manuscripts;

[0057] An output unit is used to determine the optimal summary information of the specified media manuscript based on the sorting sequence, and output the optimal summary information for display.

[0058] In a third aspect, another media summary generation device that integrates a large model and structured semantics is provided. Taking the device as an electronic device as an example, the device includes a memory, a processor and a transceiver that are communicatively connected in sequence, wherein the memory is used to store a computer program, the transceiver is used to send and receive messages, and the processor is used to read the computer program to execute the media summary generation method that integrates the large model and structured semantics as designed in the first aspect or any one of the first aspects.

[0059] In a fourth aspect, a storage medium is provided, on which instructions are stored. When the instructions are executed on a computer, the method for generating a media summary by integrating a large model with structured semantics as in the first aspect or any possible design of the first aspect is executed.

[0060] In a fifth aspect, a computer program product comprising instructions is provided, which, when executed on a computer, causes the computer to execute the media summary generation method by integrating the large model with structured semantics as described in the first aspect or any possible design of the first aspect.

[0061] Beneficial effects:

[0062] (1) The present invention sorts out the key dimensions for semantic characterization of article content, constructs a complete structured understanding model system, and integrates multi-dimensional information to establish a complete manuscript semantic structure information. This can lay the foundation for the structured semantic guidance of subsequent summary generation and the consistency evaluation of generation quality.

[0063] (2) The present invention analyzes the summary generation business scenario, sorts out the key generation requirements, injects the internal semantic information of the manuscript, and expands the external knowledge background of the key information; then, based on this, a structured prompt template is constructed; in this way, the model's global semantic understanding of the manuscript can be enhanced, and at the same time, the detailed generation requirements are complied with, thereby ensuring the comprehensiveness of the generation effect.

[0064] (3) The present invention generates summaries in batches, and simultaneously combines the consistency evaluation of key information and the evaluation of generation quality to integrate the effect ranking of summaries, thereby ensuring the accuracy and objectivity of summary generation.

[0065] (4) A human-computer collaboration mechanism is introduced in multiple stages such as confirmation of extracted information and screening of candidate summaries, which not only ensures the ultimate practicality of the summary, but also allows the accumulation of feedback data to form a closed loop of data to continuously optimize the algorithm effect of summary generation; in this way, the present invention creates a closed-loop summary generation method and enhances the ability to be applied in specific scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1A schematic diagram of the steps of a method for generating a media summary by integrating a large model with structured semantics provided in an embodiment of the present invention;

[0067] Figure 2 A process for constructing a media domain dictionary and a deactivated dictionary provided in an embodiment of the present invention;

[0068] Figure 3 A training flow chart of a keyword extraction model provided by an embodiment of the present invention;

[0069] Figure 4 A flowchart for extracting semantic structure information provided by an embodiment of the present invention;

[0070] Figure 5 A flow chart of generating a structured prompt template provided by an embodiment of the present invention;

[0071] Figure 6 A schematic diagram of training the gradient decision boosting tree provided in an embodiment of the present invention to check the consistency of candidate summary information;

[0072] Figure 7 A flowchart of the verification process of candidate summary information provided by an embodiment of the present invention;

[0073] Figure 8 A schematic diagram of feedback optimization provided by an embodiment of the present invention;

[0074] Fig. 9 A schematic diagram of the structure of a media summary generation device integrating a large model and structured semantics provided in an embodiment of the present invention;

[0075] Fig.10 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0076] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in combination with the drawings and the description of the embodiments or the prior art. Obviously, the following description of the structure of the drawings is only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0077] It should be understood that although the terms first, second, etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another unit. For example, a first unit can be referred to as a second unit, and similarly, a second unit can be referred to as a first unit without departing from the scope of the exemplary embodiments of the present invention.

[0078] It should be understood that the term "and / or" that may appear in this article is merely a description of the association relationship between associated objects, indicating that three relationships may exist. For example, A and / or B can represent three situations: A exists alone, B exists alone, and A and B exist at the same time. The term " / and" that may appear in this article describes another type of association object relationship, indicating that two relationships may exist. For example, A / and B can represent two situations: A exists alone, and A and B exist alone. In addition, the character " / " that may appear in this article generally indicates that the previous and next associated objects are in an "or" relationship.

[0079] Example:

[0080] See also Figure 1 As shown, the media summary generation method integrating the big model and structured semantics provided in this embodiment introduces the structured semantic information of the source manuscript as the data basis for summary generation based on the big model workflow, and then injects knowledge into the structured semantic information to enhance the semantic understanding ability of the model, and at the same time expands the external knowledge of key information, thereby using external knowledge enhancement to improve the generation effect; finally, batch generate summaries, and perform key information verification and effect evaluation on the generated summaries to ensure the accuracy and objectivity of the generated summaries; in this way, the method greatly improves the accuracy, coherence and information integrity of the generated summaries, and is therefore very suitable for large-scale application and promotion in the field of summary generation; among them, for example, the method can be but not limited to running on the summary generation end side, and optionally, the summary generation end side can be but not limited to a personal computer, a tablet computer or a smart phone. It can be understood that the aforementioned execution subject does not constitute a limitation on the embodiments of the present application, and accordingly, the operation steps of the method can be but not limited to the following steps S1 to S7.

[0081] S1. Obtain text information of a specified media manuscript, and extract keyword information, topic information, entity information, relationship information and event information from the text information, so as to generate semantic structure information of the specified media manuscript based on the keyword information, the topic information, the entity information, the relationship information and the event information. In this embodiment, different methods may be used to obtain text information for different specified media manuscripts, wherein text information may be directly extracted from text manuscripts, while audio information may be first extracted from audio and video manuscripts and speech recognition may be performed, and then the speech information may be uniformly converted into text information. In this way, after obtaining the text information of the specified media manuscript, semantic structured information may be constructed.

[0082] In specific implementation, this embodiment performs coarse-grained and fine-grained semantic characterization from five dimensions: keywords, topics, entities, relationships, and events, so as to construct complete manuscript semantic structure information; wherein, the process of constructing the semantic structure information can be but is not limited to as shown in the following steps S11 to S15.

[0083] S11. Construct a media domain dictionary and a stop dictionary. In specific applications, for keyword extraction, this embodiment constructs a media domain dictionary and a stop dictionary to perform word segmentation of text information, and performs keyword recognition of word segmentation words based on the integrated model of TextRank+TFIDF+YAKE. The construction process of the media domain dictionary and the stop dictionary is as follows: (1) obtaining multiple media source texts; (2) using the Jieba word segmentation component to perform word segmentation on each media source text; (3) using 1gram, 2gram or 3gram models (i.e., triple models) to combine word segmentation to obtain multiple combined words; (4) performing frequency statistics and threshold screening (i.e., performing word frequency statistics) to obtain candidate stop words and media domain proper nouns; (5) combining common stop words to perform word fusion and perform visual display to obtain actual media domain nouns and actual stop words during human-computer interaction; (6) constructing a media domain dictionary based on actual media domain nouns, and constructing a stop dictionary based on actual stop words; specifically, the construction process of the aforementioned dictionary can be but is not limited to referring to Figure 2 shown.

[0084] Thus, after constructing the media domain dictionary and the disabled dictionary based on the aforementioned step S11, keyword extraction processing can be performed based on this and in combination with the keyword extraction model, and the process can be but is not limited to as shown in the following step S12.

[0085] S12. Based on the media domain dictionary and the stop dictionary, and using the keyword extraction model, the text information is subjected to keyword extraction processing to obtain the keyword information after the keyword extraction processing; in the specific implementation, it has been explained above that the keyword extraction model is an integrated model based on textrank+tfidf+yake; therefore, the aforementioned keyword extraction model may include but is not limited to: TF-IDF feature extraction unit, textrank network unit, YAKE network unit and keyword classification recognition unit; based on this, it can be taken as an example that the following steps S12a~S12c can be used but is not limited to be used to extract the aforementioned keyword information.

[0086] S12a. Based on the media domain dictionary and the stop dictionary, the word segmentation component is called to perform word segmentation processing on the text information to obtain multiple segmentation words; in specific applications, the example word segmentation component can be but not limited to the jieba word segmentation component (which is an open source Chinese word segmentation tool); of course, other word segmentation components can also be used, and this is not limited to the aforementioned examples.

[0087] After completing the segmentation of the aforementioned text information based on the jieba segmentation component, the media domain dictionary and the stop dictionary, the aforementioned keyword extraction model can be used to identify keywords, and the process can be but is not limited to the following steps S12b and S12c.

[0088] S12b. Using the TF-IDF feature extraction unit, the textrank network unit, and the YAKE network unit, feature extraction processing is performed on each segmentation word to obtain first feature information, second feature information, and third feature information of each segmentation word, respectively, wherein the first feature information of any segmentation word includes the TF-IDF feature value of the any segmentation word, the second feature information of any segmentation word includes the word importance feature value of the any segmentation word, and the third feature information of any segmentation word includes the position feature information and length feature information of the any segmentation word; in this embodiment, T The F-IDF feature value is the word frequency-inverse text frequency index value, which is used to evaluate the importance of a word to a file set or one of the files in a corpus; the textrank network unit is a graph-based sorting algorithm for text, which can establish a graph model based on multiple word segments, and use a voting mechanism to sort the important components in the text, so as to obtain the importance of each word segment; similarly, the YAKE network unit uses indicators such as word position, contextual relationship, full-text word frequency, and length to calculate the candidate word score. In the specific implementation, this embodiment selects indicators such as word position and length as its corresponding feature values.

[0089] In this way, after obtaining different word features of each segmented word, the keyword classification and recognition unit can be used to perform keyword recognition, and the process can be but not limited to the following step S12c.

[0090] S12c. Input the first feature information, the second feature information and the third feature information of each segmented word into the keyword classification and recognition unit for keyword recognition processing, so as to identify the keywords of the text information from multiple segmented words and obtain the keyword information; in this embodiment, the keyword classification and recognition unit may include but is not limited to a random forest network; in this way, the aforementioned keyword extraction model extracts different word features, and then classifies and recognizes keywords based on the random forest network and the extracted multiple word features.

[0091] Optionally, the training process of the aforementioned keyword extraction model may be, but is not limited to: (1) obtaining a number of training media texts; (2) performing JIEBA word segmentation based on a custom dictionary (i.e., based on the aforementioned media domain dictionary and stop dictionary); (3) manually screening out keywords as label information; (4) using the keyword extraction model to extract the first feature information, second feature information, and third feature information of each word segmentation; (5) using a random forest network and the aforementioned extracted feature information to perform keyword classification and recognition training, and using the model output information and the aforementioned label information to adjust the model parameters until the model converges to obtain a trained keyword extraction model; wherein, the training process can be found in Figure 3 shown.

[0092] After the keyword information of the text information is extracted based on the aforementioned steps S12a to S12c, the subject information may be extracted, and the process may be but not limited to that shown in the following step S13.

[0093] S13. Use the LDA topic extraction model to perform topic extraction processing on the text information to obtain initial topic information; in this embodiment, the initial topic information is extracted based on the LDA topic extraction model trained in the media field (which is Latent Dirichlet Allocation (LDA), a topic model (topic model l), which can give the topic of each document in the form of probability distribution); then, the initial topic information is mapped through preset media topic tags to obtain more detailed topic information; wherein, the topic mapping process can be but is not limited to as shown in the following step S14.

[0094] S14. Use preset media topic tags to perform topic mapping processing on the initial topic information, so as to obtain the topic information corresponding to the text information after the topic mapping processing; in this embodiment, three levels of tags can be set in order from large to small to perform topic mapping, so as to obtain more detailed topic information; of course, the number of levels of tag settings can be specifically set according to actual use, and is not limited to the aforementioned examples.

[0095] After the subject information of the text information is extracted through the aforementioned steps S13 and S14, entity, relationship and event information can be extracted, and the process can be but not limited to as shown in the following step S15.

[0096] S15. Utilize the information extraction model to extract entity, relationship and event information from the text information to obtain entity information, relationship information and event information corresponding to the text information; in specific implementation, the information extraction model may include, but is not limited to, the yayi-ui model, that is, the yayi-ui model is used to extract entity, relationship and event information; wherein, the yayi-ui model is a large information extraction model based on Yayi 2.0, which is a unified information extraction framework for both Chinese and English, and can complete the integration of entity, relationship and event information extraction; of course, other information extraction models may also be used, and are not limited to the aforementioned examples.

[0097] Thus, through the aforementioned steps S11 to S15, the semantic information extraction of the five dimensions of keywords, topics, entities, relationships, and events can be completed; wherein, in this embodiment, the semantic information of the aforementioned five dimensions can also be visualized, that is, output to the human-computer interaction terminal, so that the information can be manually corrected and the corrected keyword, topic, entity, relationship, and event information can be returned; then, the semantic structure information of the specified media manuscript can be constructed based on the aforementioned modified five-dimensional information. In this embodiment, the aforementioned semantic structure information can be constructed in the form of a graph, for example, but not limited to, and the entire construction process of the semantic structure information can be, but not limited to, referring to Figure 4 shown.

[0098] After obtaining the semantic structure information, internal and external knowledge enhancement can be performed on it to improve the semantic understanding of the manuscript and enhance the comprehensiveness and accuracy of the generated summary; wherein, the internal and external knowledge enhancement process can be but is not limited to as shown in the following steps S2 and S3.

[0099] S2. Perform internal knowledge enhancement processing on the semantic structure information to obtain enhanced knowledge information. In this embodiment, the internal knowledge enhancement process is as follows: (1) obtaining an internal database, wherein the internal database stores media text information downloaded from a number of trusted websites; (2) performing expansion processing on the semantic structure information based on the media text information to obtain enhanced knowledge information after the expansion processing. In this way, the internal knowledge enhancement of the semantic structure information can be completed through the aforementioned steps. Based on this, internal knowledge enhancement is equivalent to using locally stored trusted media text information to expand the semantic structure information (such as information expansion based on entities and events, etc.), thereby enhancing the semantic understanding ability of the manuscript.

[0100] After the internal knowledge enhancement is completed, the external knowledge enhancement may be performed, and the process is shown in the following step S3.

[0101] S3. Perform external background expansion processing on the entity information and relationship information in the semantic structure information to obtain background expansion information; in specific applications, for example, but not limited to, the following steps S31 to S34 can be used to complete the external knowledge enhancement of the semantic structure information.

[0102] S31. Call the search plug-in to perform background search processing on the entity information and the relationship information to obtain initial extended information; in specific applications, the search plug-in can be, but is not limited to, a commonly used search engine, that is, using a search engine to perform background search of entity and relationship information on the Internet to obtain initial extended information.

[0103] After the initial extended information is obtained, the information may be screened based on its source to ensure the reliability of the information, wherein the information screening process may be, but is not limited to, as shown in the following steps S32 and S33.

[0104] S32. Obtain the business scenario information of the specified media manuscript, and determine the information source list based on the business scenario information; in this embodiment, the information source list corresponding to different business scenarios is pre-stored in the example summary generation terminal, which can be called when used; wherein, the information source list stores reliable information source URLs or websites; therefore, after determining the information source list of the specified media manuscript based on the business scenario, the initial extended information can be screened based on this, and the process can be but not limited to as shown in the following step S33.

[0105] S33. According to the information source list, the initial extended information is filtered and processed to obtain filtered extended information; in this embodiment, if the information source website or URL of any initial extended information does not exist in the aforementioned information source list, then any initial extended information is filtered out, otherwise, the any initial extended information is retained; in this way, after completing the filtering of the initial extended information, the background extended information can be generated based on the filtered extended information, and the process can be but is not limited to as shown in the following step S34.

[0106] S34. Based on the filtered extended information, the background extended information is generated; in this embodiment, the text content on the website corresponding to the filtered extended information is obtained, and the obtained text content is used as the background extended information; at the same time, in order to ensure the correctness of the background extended information, this embodiment also provides a human-computer interaction process, that is, the background extended information is sent to the human-computer interaction end for manual confirmation, and after receiving the confirmation notification, it is used as the final background extended information.

[0107] In this way, based on the aforementioned steps S2 and S3 and the corresponding sub-steps, after completing the internal and external knowledge enhancement of the semantic structure information, a structured prompt template can be generated with the help of a structured prompt model; wherein, the generation process of the structured prompt template can be but is not limited to as shown in the following step S4.

[0108] S4. Construct a structured prompt model based on a large model, and generate a structured prompt template based on the structured prompt model, the background extension information and the enhanced knowledge information; in the specific implementation, the example can be but limited to the Qwen-14B-based large model (which is the open source Tongyi Qianwen 14 billion parameter model under Alibaba Cloud), and according to business needs, it covers five parts of requirements, namely, factual requirements, accuracy requirements, relevance requirements, generation effect requirements, and error verification, to construct a structured prompt model; among them, in the structured prompt, the factual requirements ensure the objectivity of the generated summary in terms of events, the accuracy requirements ensure the accuracy of the key information involved, the relevance requirements ensure the consistency of the subject and the comprehensiveness of the details of the generated content, and the error verification ensures the smooth usability of the generated results; in this embodiment, the generation process of the structured prompt template can be but is not limited to referring to Figure 5 As shown; of course, the aforementioned structured prompt model can also be pre-set on the summary generation end side and can be called when used.

[0109] Therefore, after generating a structured prompt template through the aforementioned structured prompt model and combining enhanced knowledge information and background extension information, the structured prompt template can be used to generate several candidate summary information of the specified media manuscript; wherein, the generation process of the candidate summary information can be but is not limited to as shown in the following step S5.

[0110] S5. Generate a number of candidate summary information according to the structured prompt template; in specific implementation, the example may be but not limited to based on the summary information generation model, and generate a number of candidate summary information according to the structured prompt template; and the example summary information generation model may be but not limited to the Qwen-14B model; in this way, after generating a number of candidate summary information, a candidate summary information verification process may be introduced to ensure the accuracy and objectivity of the generated summary; wherein, the candidate summary information verification process may be but not limited to as shown in the following step S6.

[0111] S6. Verify each candidate summary information, and sort each candidate summary information according to the verification result to obtain a sorting sequence, wherein the verification process includes quality verification and consistency verification with the specified media manuscript; in the specific implementation, any candidate summary information is taken as an example for specific explanation, and the process can be but not limited to the following steps S61 to S63.

[0112] S61. For any candidate summary information, a consistency check is performed on the candidate summary information and the designated media manuscript to obtain a consistency check result; in specific applications, examples include but are not limited to first obtaining the original entity information, original subject information and original event information of the designated media manuscript, as well as the summary entity information, summary event information and summary subject information of any candidate summary information; then, a gradient decision boosting tree (i.e., GBDT model) is used to perform a consistency evaluation process on the original entity information, the original subject information, the original event information, the summary entity information, the summary event information and the summary subject information, so as to obtain the consistency check result after the consistency evaluation process.

[0113] Among them, see Figure 6 As shown, the training process of the above-mentioned gradient decision boosting tree for consistency verification is given as follows: first, the corresponding entity, topic and event information is extracted from the original manuscript, and then, the corresponding entity, topic and event information is extracted from the generated summary information (a plurality of training information can be pre-set as long as the information is used); then, the accuracy of the entities, topics and events corresponding to the summary and the entities, topics and events of the original manuscript (i.e., the similarity between the two) and the recall rate of the entities, topics and events are used as indicators to train the gradient decision boosting tree model; at the same time, during the training process, the manual consistency score is used as label data, and combined with the score output by the model to adjust the parameters of the model; in this way, the training is continuously carried out according to this principle until the model converges, and the trained gradient decision boosting tree can be obtained; finally, the trained gradient decision boosting tree can be used to complete the consistency check of any candidate summary information with the specified media manuscript to obtain the corresponding score.

[0114] After the consistency check is completed, a quality check may be performed, and the process may be, but is not limited to, as shown in the following step S62.

[0115] S62. Perform a quality check on any of the candidate summary information to obtain a quality check result; in specific applications, for example, it can be but not limited to using a quality assessment model, and perform a quality check on any of the candidate summary information to obtain the quality check result after the quality assessment; optionally, for example, the quality assessment model can be but not limited to including a CONNER tool; wherein the CONNER tool is a quality assessment tool for AI-generated content, which assesses the quality of AI-generated content from six aspects: factuality, relevance, coherence, informativeness, helpfulness, and effectiveness; wherein the aforementioned CONNER tool is a commonly used tool for quality assessment of AI-generated content, and its principle will not be repeated here.

[0116] After obtaining the quality check result of any candidate summary information based on the aforementioned step S62, the aforementioned consistency check result can be combined to generate a check result corresponding to any candidate summary information, and the process can be but not limited to as shown in the following step S63.

[0117] S63. Generate a verification result of any candidate summary information according to the quality verification result and the consistency verification result; in this embodiment, the verification result of any candidate summary information can be obtained by, for example but not limited to, using a weighted fusion method, that is, the aforementioned consistency verification and quality verification results are both corresponding scores, therefore, multiplying their respective weights and then adding them together to obtain the final verification result.

[0118] Thus, based on the aforementioned steps S61 to S63, the consistency and quality scoring of each candidate summary information can be completed; wherein the aforementioned candidate summary information verification process can be but is not limited to referring to Figure 7 As shown; at the same time, after obtaining the final verification result, the final verification result (i.e., the verification score) can be used for sorting, that is, the candidate summary information is sorted in descending order according to the final evaluation score to obtain a sorted sequence; and after obtaining the sorted sequence, the optimal summary information of the aforementioned specified media manuscript can be determined, and the process can be but not limited to as shown in the following step S7.

[0119] S7. Based on the sorting sequence, determine the optimal summary information of the specified media manuscript, and output the optimal summary information for display; in this embodiment, the example may be but is not limited to taking the first n candidate summary information in the sorting sequence as the optimal summary information, wherein n is a positive integer, which may be but is not limited to 1, 3 or 5; of course, the specific value may be set according to actual use, and is not limited to the aforementioned example.

[0120] Furthermore, in this embodiment, after generating the sorting sequence, at least one candidate summary information can be obtained from the sorting sequence as sample summary information; wherein the sorting sequence can be sent to the human-computer interaction terminal so that the user can select one or more candidate summary information from the sorting sequence as sample summary information, so that feedback optimization of the summary information generation model can be performed based on this in the future, that is, the sample summary information is used to perform feedback optimization processing on the summary information generation model to obtain the optimized summary information generation model, so that the optimized summary information generation model can be used to generate candidate summary information of the next specified media manuscript; in this way, fine-tuning optimization of the model can be achieved, thereby continuously improving the generation effect.

[0121] Similarly, in the aforementioned semantic structure information extraction stage, keyword information, topic information, entity information, relationship information and event information can also be output to the human-computer interaction end, so that the user corresponding to the human-computer interaction end can modify the keyword information, topic information, entity information, relationship information and event information to obtain modified keyword information, modified topic information, modified entity information, modified relationship information and modified event information; then, in response to the human-computer interaction operation with the human-computer interaction end, the modified keyword information, modified topic information, modified entity information, modified relationship information and modified event information can be obtained; finally, the modified keyword information can be used to perform feedback optimization processing on the keyword extraction model, and the modified topic information can be used to optimize the keyword extraction model. The LDA topic extraction model is subjected to feedback optimization processing, and the information extraction model is subjected to feedback optimization processing using the corrected entity information, the corrected relationship information and the corrected event information, so as to obtain an optimized keyword extraction model, an optimized LDA topic extraction model and an optimized information extraction model respectively, so as to use the optimized keyword extraction model, the optimized LDA topic extraction model and the optimized information extraction model to generate the keyword information, topic information, entity information, relationship information and event information of the next specified media manuscript; based on this, the information extraction algorithm (i.e., the aforementioned models) can also be fine-tuned and optimized based on the manually confirmed corrected semantic information data, so as to continuously improve the effect of information extraction; optionally, the schematic diagram of the aforementioned feedback optimization can be seen in Figure 8 shown.

[0122] Thus, through the media summary generation method integrating the big model and structured semantics described in detail in the aforementioned steps S1 to S7, the present invention introduces the structured semantic information of the source manuscript based on the big model workflow as the data basis for summary generation, and then injects the structured semantic information into the knowledge to enhance the semantic understanding ability of the model, and at the same time expands the external knowledge of key information, thereby improving the generation effect by enhancing the external knowledge; finally, the summaries are generated in batches, and the generated summaries are checked for key information and evaluated for effect, so as to ensure the accuracy and objectivity of the generated summaries; in this way, the method greatly improves the accuracy, coherence and information integrity of the generated summaries, and is therefore very suitable for large-scale application and promotion in the field of summary generation.

[0123] like Fig. 9 As shown, the second aspect of this embodiment provides a hardware device for implementing the media summary generation method integrating a large model and structured semantics described in the first aspect of the embodiment, including:

[0124] A semantic structure information generating unit is used to obtain text information of a specified media manuscript and extract keyword information, topic information, entity information, relationship information and event information in the text information, so as to generate semantic structure information of the specified media manuscript based on the keyword information, the topic information, the entity information, the relationship information and the event information.

[0125] The knowledge enhancement unit is used to perform internal knowledge enhancement processing on the semantic structure information to obtain enhanced knowledge information.

[0126] The expansion unit is used to perform external background expansion processing on the entity information and relationship information in the semantic structure information to obtain background expansion information.

[0127] The structured prompt generating unit is used to construct a structured prompt model based on the large model, and generate a structured prompt template based on the structured prompt model, the background extension information and the enhanced knowledge information.

[0128] The summary generating unit is used to generate a plurality of candidate summary information according to the structured prompt template.

[0129] The summary generation unit is used to check and process each candidate summary information, and sort each candidate summary information according to the checking and processing result to obtain a sorting sequence, wherein the checking and processing includes quality checking and consistency checking with designated media manuscripts.

[0130] The output unit is used to determine the optimal summary information of the specified media manuscript based on the sorting sequence, and output the optimal summary information for display.

[0131] The working process, working details and technical effects of the device provided in this embodiment can be found in the first aspect of the embodiment and will not be described in detail here.

[0132] like Fig.10 As shown, the third aspect of this embodiment provides another media summary generation device that integrates a large model and structured semantics. Taking the device as an electronic device as an example, it includes: a memory, a processor and a transceiver that are communicatively connected in sequence, wherein the memory is used to store computer programs, the transceiver is used to send and receive messages, and the processor is used to read the computer program to execute the media summary generation method that integrates a large model and structured semantics as described in the first aspect of the embodiment.

[0133] For example, the memory may include, but is not limited to, random access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out memory (FIFO) and / or first-in-last-out memory (FILO), etc. Specifically, the processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). At the same time, the processor may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in the standby state.

[0134] In some embodiments, the processor may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. For example, the processor may not be limited to a microprocessor of the STM32F105 series, a reduced instruction set computer (RISC) microprocessor, an X86 or other architecture processor, or a processor with an integrated embedded neural network processor (NPU); the transceiver may be, but is not limited to, a wireless fidelity (WIFI) wireless transceiver, a Bluetooth wireless transceiver, a general packet radio service technology (Genera l Packet Rad ioService, GPRS) wireless transceiver, a ZigBee protocol (a low-power LAN protocol based on the IEEE 802.15.4 standard, ZigBee) wireless transceiver, a 3G transceiver, a 4G transceiver and / or a 5G transceiver, etc. In addition, the device may also include, but is not limited to, a power module, a display screen and other necessary components.

[0135] The working process, working details and technical effects of the electronic device provided in this embodiment can be found in the first aspect of the embodiment and will not be described in detail here.

[0136] A fourth aspect of this embodiment provides a storage medium storing instructions for the method for generating a media summary by fusing a large model and structured semantics as described in the first aspect of the embodiment, that is, the storage medium stores instructions, and when the instructions are executed on a computer, the method for generating a media summary by fusing a large model and structured semantics as described in the first aspect of the embodiment is executed.

[0137] The storage medium refers to a carrier for storing data, which may include but is not limited to a floppy disk, a CD, a hard disk, a flash memory, a USB flash drive and / or a memory stick, etc. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0138] The working process, working details and technical effects of the storage medium provided in this embodiment can be found in the first aspect of the embodiment and will not be described in detail here.

[0139] A fifth aspect of the present embodiment provides a computer program product comprising instructions, which, when executed on a computer, causes the computer to execute the method for generating a media summary by integrating a large model with structured semantics as described in the first aspect of the embodiment, wherein the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0140] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A media summary generation method integrating a large model and structured semantics, characterized in that: include: Acquire text information of a specified media manuscript, and extract keyword information, subject information, entity information, relationship information and event information from the text information, so as to generate semantic structure information of the specified media manuscript based on the keyword information, the subject information, the entity information, the relationship information and the event information; Performing internal knowledge enhancement processing on the semantic structure information to obtain enhanced knowledge information; Performing external background expansion processing on entity information and relationship information in the semantic structure information to obtain background expansion information; Constructing a structured prompt model based on the large model, and generating a structured prompt template based on the structured prompt model, the background extension information and the enhanced knowledge information; Generating a plurality of candidate summary information according to the structured prompt template; Performing verification processing on each candidate summary information, and sorting each candidate summary information according to the verification processing result to obtain a sorting sequence, wherein the verification processing includes quality verification processing and consistency verification processing with designated media manuscripts; Based on the sorting sequence, optimal summary information of the designated media manuscript is determined, and the optimal summary information is output for display.

2. The method according to claim 1, characterized in that Extracting keyword information, topic information, entity information, relationship information and event information from the text information includes: Construct a media domain dictionary and a stop dictionary; Based on the media domain dictionary and the stop dictionary, and using a keyword extraction model, performing keyword extraction processing on the text information to obtain the keyword information after the keyword extraction processing; Using the LDA topic extraction model, performing topic extraction processing on the text information to obtain initial topic information; Using a preset media theme tag, the initial theme information is subjected to theme mapping processing, so as to obtain theme information corresponding to the text information after the theme mapping processing; The information extraction model is used to extract entity, relationship and event information from the text information to obtain entity information, relationship information and event information corresponding to the text information, wherein the information extraction model includes a yayi-uie model.

3. The method according to claim 2, characterized in that The keyword extraction model includes: a TF-IDF feature extraction unit, a textrank network unit, a YAKE network unit and a keyword classification and recognition unit; The keyword extraction process is performed on the text information based on the media domain dictionary and the stop dictionary and using a keyword extraction model to obtain the keyword information after the keyword extraction process, including: Based on the media domain dictionary and the stop dictionary, and calling the word segmentation component, the text information is segmented to obtain a plurality of segmented words; Using the TF-IDF feature extraction unit, the textrank network unit, and the YAKE network unit, feature extraction processing is performed on each segmentation word to obtain first feature information, second feature information, and third feature information of each segmentation word, respectively, wherein the first feature information of any segmentation word includes the TF-IDF feature value of the any segmentation word, the second feature information of any segmentation word includes the word importance feature value of the any segmentation word, and the third feature information of any segmentation word includes the position feature information and length feature information of the any segmentation word; The first feature information, the second feature information and the third feature information of each segmentation word are input into a keyword classification and recognition unit for keyword recognition processing, so as to identify the keywords of the text information from multiple segmentation words and obtain the keyword information, wherein the keyword classification and recognition unit includes a random forest network.

4. The method according to claim 1, characterized in that: Performing background extension processing on the entity information and the relationship information in the semantic structure information to obtain background extension information includes: Calling a search plug-in to perform background search processing on the entity information and the relationship information to obtain initial extended information; Acquire business scenario information of the specified media manuscript, and determine an information source list based on the business scenario information; According to the information source list, the initial extended information is filtered to obtain filtered extended information; The background extension information is generated based on the screening extension information.

5. The method according to claim 1, characterized in that Verify and process each candidate summary information, including: For any candidate summary information, a consistency check is performed on the candidate summary information and the designated media manuscript to obtain a consistency check result; Performing a quality check on any of the candidate summary information to obtain a quality check result; A verification result of any candidate summary information is generated according to the quality verification result and the consistency verification result.

6. The method according to claim 5, characterized in that Performing consistency check on any candidate summary information and the designated media manuscript to obtain consistency check results, including: Acquire original text entity information, original text subject information and original text event information of the specified media manuscript, as well as summary entity information, summary event information and summary subject information of any candidate summary information; Using a gradient decision boosting tree, a consistency evaluation process is performed on the original text entity information, the original text topic information, the original text event information, the summary entity information, the summary event information, and the summary topic information, so as to obtain the consistency verification result after the consistency evaluation process; Accordingly, performing a quality check on any candidate summary information to obtain a quality check result includes: The quality assessment model is used to perform a quality check on any candidate summary information, so as to obtain the quality check result after the quality check, wherein the quality assessment model includes a CONNER tool.

7. The method according to claim 1, characterized in that According to the structured prompt template, several candidate summary information are generated, including: Based on the summary information generation model and according to the structured prompt template, a number of candidate summary information are generated; Wherein, after sorting each candidate summary information according to the verification processing result to obtain a sorting sequence, the method further includes: Obtain at least one candidate summary information from the sorted sequence as sample summary information; The sample summary information is used to perform feedback optimization processing on the summary information generation model to obtain an optimized summary information generation model, so as to use the optimized summary information generation model to generate candidate summary information for the next designated media manuscript.

8. The method according to claim 2, characterized in that: After extracting keyword information, topic information, entity information, relationship information and event information from the text information, the method further includes: Outputting the keyword information, topic information, entity information, relationship information and event information to the human-computer interaction terminal, so that the user corresponding to the human-computer interaction terminal can modify the keyword information, topic information, entity information, relationship information and event information to obtain modified keyword information, modified topic information, modified entity information, modified relationship information and modified event information; In response to a human-computer interaction operation with the human-computer interaction terminal, obtaining revised keyword information, revised subject information, revised entity information, revised relationship information, and revised event information; The keyword extraction model is fed back and optimized using the corrected keyword information, the LDA topic extraction model is fed back and optimized using the corrected topic information, and the information extraction model is fed back and optimized using the corrected entity information, the corrected relationship information and the corrected event information, so as to obtain an optimized keyword extraction model, an optimized LDA topic extraction model and an optimized information extraction model respectively, so as to use the optimized keyword extraction model, the optimized LDA topic extraction model and the optimized information extraction model to generate the keyword information, topic information, entity information, relationship information and event information of the next specified media manuscript.

9. A media summary generation device integrating a large model and structured semantics, characterized in that: include: A semantic structure information generating unit, used to obtain text information of a specified media manuscript, and extract keyword information, topic information, entity information, relationship information and event information in the text information, so as to generate semantic structure information of the specified media manuscript based on the keyword information, the topic information, the entity information, the relationship information and the event information; A knowledge enhancement unit, used for performing internal knowledge enhancement processing on the semantic structure information to obtain enhanced knowledge information; An extension unit, used for performing external background extension processing on entity information and relationship information in the semantic structure information to obtain background extension information; A structured prompt generating unit, used to construct a structured prompt model based on the large model, and generate a structured prompt template based on the structured prompt model, the background extension information and the enhanced knowledge information; A summary generating unit, used for generating a plurality of candidate summary information according to the structured prompt template; A summary generation unit is used to check and process each candidate summary information, and sort each candidate summary information according to the check and process result to obtain a sorting sequence, wherein the check and process includes quality check and consistency check with designated media manuscripts; An output unit is used to determine the optimal summary information of the specified media manuscript based on the sorting sequence, and output the optimal summary information for display.

10. A computer program product comprising instructions, characterized in that When the instructions are executed on a computer, the computer is caused to execute the method for generating a media summary by integrating a large model with structured semantics as described in any one of claims 1 to 8.

Images