Cboth rewriting system for digital human live broadcast

Abstract

A copywriting rewriting system for digital human live broadcast belongs to the technical field of computer programs, and expands a plurality of different expression modes which do not change original meaning according to the content of an original verbal skill copywriting based on a segment of existing live broadcast verbal skill copywriting, a target verbal skill length (or playing duration) and a non-repetition rate of the target verbal skill. And character contents are ensured not to be repeated. According to the invention, one-time off-line operation before broadcasting can be realized. In order to further ensure the safety of the content, the limited content generated by the method can be manually examined before broadcasting, so that the safety and the accuracy of the content are ensured, and the risk caused by forbidden words, inaccurate content or repeated texts and the like in the live broadcasting process is reduced to the minimum.

Description

Technical Field

[0001] The invention belongs to the technical field of computer programs and relates to a text rewriting system for digital human live broadcast. Background Art

[0002] In the live broadcast scenario of the digital human, the digital human will speak according to the prepared live broadcast copy. In the actual live broadcast scenario, the copy used cannot be repeated. Using repeated copy will cause the risk of being judged as recorded or low-quality content.

[0003] The more traditional method is usually to prepare a long enough copy of the script. These copies can be prepared manually or generated with the help of AI methods. In the case of a long live broadcast, hundreds of thousands of words of non-repetitive text need to be prepared. These methods are very labor-intensive or require continuous consumption of a large amount of AI resources, which is linearly related to the length of the target text. For example, generating a ten-hour live broadcast copy requires 60 times more AI resources than generating a ten-minute live broadcast copy.

[0004] Some existing text rewriting solutions are based on the idea of ​​word replacement and rewrite and expand an existing text. For the words in the original text, several candidate replacement words are used to replace and reorganize them to generate a new reorganized text. The main disadvantage of this method is that the same words will present different semantics in different contexts; and in some scenarios where phrases are fixed, using synonyms to replace them will destroy the original fixed collocation. After replacing the words, even if synonyms are used, it is easy to cause the reorganized copy to be incoherent. Summary of the invention

[0005] The present invention aims at solving the problems in the prior art and provides a text rewriting system for digital human live broadcast.

[0006] A text rewriting system for digital human live broadcasting includes the following steps: an existing speech text is input into a text unit splitting module of a speech text expansion module, and the text unit splitting module generates text unit 1, text unit 2, text unit 3 and text unit n, and then transmits them to a text rewriting module; the text rewriting module generates a number of candidate segments of text unit 1, a number of candidate segments of text unit 2, a number of candidate segments of text unit 3, and a number of candidate segments of text unit n, and the candidate segments of text unit n (n is a positive integer) are respectively transmitted to a candidate segment reorganization module, and the candidate segment reorganization module generates a number of reorganized new speech texts, and a target live broadcast duration and a target non-repetitiveness module are connected to the candidate segment reorganization module and the text unit splitting module.

[0007] A text rewriting system for digital human live broadcast is based on an existing live broadcast script, a target script length (or playback time) and a target script non-repetition rate. According to the content of the original script, several different expressions without changing the original meaning are expanded to ensure that the text content is not repeated.

[0008] The advantages of the present invention are that it can adaptively adjust different text unit splitting granularities and reorganization times according to the needs of actual scenarios, such as the target live broadcast duration and the target non-repetitiveness. The text unit splitting granularity can be from the phrase level to the sentence level, and finally to the paragraph level. The finer the splitting granularity, the better the text non-repetitiveness can be obtained. The coarser the splitting granularity, the better the text fluency and diversity of expression can be obtained. In actual use, based on a paragraph of original script of several thousand words, to obtain a non-repetitive text of hundreds of thousands of words that meets the requirements of dozens of hours of long-term live broadcast, the present invention can be used for a one-time rewrite.

[0009] Cost controllable: The present invention only needs to generate a limited number of replacement texts at one time, and there is no need to increase the consumption of AI resources as the length of the live broadcast increases.

[0010] The effect is controllable: the text content generated by the present invention is based on the original copy, and the content will not deviate from the original text. And with the help of the large language model's ability to understand and generate text, the reorganized text also has good fluency and can be directly used in the subsequent automated process of digital human live broadcast without manual intervention and modification. Compared with the replacement word scheme, it can avoid destroying the fixed phrase collocation (including product names that appear during the live broadcast), context and context in the original text, and can also avoid replacing a polysemous word and causing semantic deviation and incoherence. The large language model can generate a rewrite of a section of original text based on the understanding of the section. The generated rewritten text will not have the problem of incoherence caused by the replacement word method. Since the segmentation of the text unit will be divided according to the granularity of phrases, half sentences separated by commas, whole sentences or paragraphs, the components in the sentence will not be chopped up, and the reorganization between candidate segments will not cause fluency problems.

[0011] Content security: Since the amount of text generated is limited, the expansion of the amount of text is mainly achieved by reorganization. This system can be implemented as a one-time offline operation before broadcasting. To further ensure the security of the content, the limited content generated by the present invention can also be manually reviewed before broadcasting to ensure the security and accuracy of the content, and to minimize the risks caused by banned words, inaccurate content or repeated text during the live broadcast. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are 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. As shown in the figure:

[0013] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0014] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0015] Example 1: Figure 1 As shown, a text rewriting system for digital human live broadcasting includes the following steps: an existing speech text is input into the text unit splitting module of the speech text expansion module, and the text unit splitting module generates text unit 1, text unit 2, text unit 3 and text unit n, and then transmits them to the text rewriting module; the text rewriting module generates several candidate segments of text unit 1, several candidate segments of text unit 2, several candidate segments of text unit 3, and several candidate segments of text unit n, and several candidate segments of text unit n (n is a positive integer) are respectively transmitted to the candidate segment reorganization module, and the candidate segment reorganization module generates several reorganized new speech texts, and the target live broadcast duration and target non-repetition degree module are connected to the candidate segment reorganization module and the text unit splitting module.

[0016] Target live broadcast duration and target non-repetition module: Based on the length of the original script, the target broadcast duration, and the speed of the digital human’s oral broadcast, the number of times the script text needs to be reorganized can be directly calculated:

[0017]

[0018] The target non-repetition degree means that any text segment of length k does not appear twice within m reorganizations. Then, based on the length of the text unit, the number of text units in a text segment of length k can be calculated:

[0019]

[0020] where n k文本单元 is the number of text segments contained in the k-word segment, is the average length of text segments, and Ceil is rounded up. Since candidates can be selected and combined arbitrarily between text units, in order to ensure that there are enough candidates to ensure that texts of consecutive k-word lengths are not repeated in m reorganizations, the number of candidates for this text needs to meet the following requirements:

[0021]

[0022] where n 候选 is the number of candidates available for each text unit. Therefore, according to the target non-repetitive degree k and the number of reorganizations m, a set of n that satisfy the above inequality can be calculated. 候选 and

[0023] Assume that the target is repeated m=100 times, and the non-repetitiveness is that there is no repetition in a segment of k=100 consecutive words. Pick n 候选 =5 (i.e., using the large language model to generate 4 candidates for each text unit, plus the original expression, there are a total of 5 candidates), the number of candidate combinations is The minimum requirement of 100 reorganizations can be met.

[0024] Example 2: Figure 1 As shown, a text rewriting system for digital human live broadcast reduces the consumption of manpower and AI resources when generating a large number of live broadcast texts. By generating candidates and combinations, a large number of texts without repeated fragments are generated in a low-cost manner, ensuring compliance with the semantics of the original text, reducing the risk of hallucinations in the generation process of large language models, and solving the problems of incoherent new texts and semantic deviations caused by replacing words.

[0025] Based on an existing live broadcast script, the target script length (or playback time) and the target script non-repetition rate, expand the content of the original script with several different expressions that do not change the original meaning to ensure that the text content is not repeated.

[0026] The steps are as follows:

[0027] The first step is to split the original text into smaller units, such as phrases, sentences or paragraphs.

[0028] In the second step, for each basic unit that has been split, the large language model uses its AI capabilities to understand and generate texts to generate several new candidate texts with the same content as the original text but in different expressions. Compared with solutions based on word replacement, the large language model can better understand longer texts and more accurately understand the semantics of the original text; and with the generative capabilities of the large language model, the newly generated text can also have good fluency.

[0029] The third step, after generating several new candidate texts for each text unit, you can select a candidate from the candidate text fragments each time to reorganize a new speech copy. The meaning expressed by this reorganized speech copy is the same as the original text, and it uses a new expression that is not repeated in the original text. By combining different candidate fragments, a large number of non-repetitive texts can be combined. In the scenario of long-term live broadcast, the above reorganization steps can be repeated many times to reorganize the speech copy whose length meets the target live broadcast duration. The number of non-repetitive texts that can be obtained is exponentially related to the number of text units split out. There are many ways to select candidate fragments. You can use the most basic random sampling, or you can stagger the selection based on the statistical combination of adjacent fragments.

[0030] Example 3: Take the following short live broadcast copy as an example: Family members, welcome to our live broadcast room! Today I have brought you a lot of benefits and good things to share. Don’t forget to follow me so that you won’t miss any exciting content! The anchor is waiting for you here, let’s have fun together!

[0031] The following candidates can be generated by replacing words:

[0032] [Family|everyone|buddies|brothers|dear ones], [welcome|glad to see|sincerely invite|warmly welcome|welcome to][come|come to|enter|come to|come to][our|my|our|own|our] live broadcast room! Today [I|we|I] [brought|brought][to|everyone|you all][a lot|a lot|a lot of] benefits[and|as well as] good things[to share|share with everyone|share], don't forget[click|poke|click][once] to follow[oh|yeah], [like this|like this|like this][just like this|just like this][will not|will not][miss|miss out|miss out][any|anything|everything][wonderful|great] content[here|here]! The host is [here|in the live broadcast room|here][waiting|waiting for][everyone|everybody|everyone], [let’s|together|together together][let’s have fun|have fun|live it up][let’s|come on|ah]!

[0033] By combining the replacement words, it is possible to compose the following incoherent new text:

[0034] Don’t forget to follow us, so you don’t miss out on all the great content!

[0035] like Figure 1 As shown, a copywriting rewriting system for digital human live broadcast includes the following steps:

[0036] For the following short original live broadcast copy:

[0037] Family, welcome to our live broadcast room! Today I have brought you a lot of benefits and good things to share. Don’t forget to follow us so that you won’t miss any exciting content! The host is waiting for you here, let’s have fun together!

[0038] The length is 75 words. A new live text with a length of 1,000 words and no repeated 50 consecutive words is required.

[0039] Phase 1: Text unit segmentation.

[0040] According to the length of about 15 words, separated by punctuation marks, the text is divided into the following six units:

[0041] Family members, welcome to our live broadcast room!

[0042] Today I bring you a lot of benefits and good things to share.

[0043] Don’t forget to follow us.

[0044] This way you won’t miss any exciting content!

[0045] The host is waiting for you here.

[0046] Let’s have fun together!

[0047] Phase 2: Calling a large language model to rewrite text.

[0048] For each text unit, 5 candidates are generated:

[0049] "Family, come and join our live broadcast room!"; "Welcome, family, to our live broadcast room!"; "Family, don't forget to come to our live broadcast room!"; "Family, the live broadcast room is ready and waiting for you!"; "Welcome everyone, family, to our live broadcast room!"

[0050] “Today I want to share a lot of benefits and good things with my family,”; “Today in the live broadcast room, I want to bring you a lot of benefits and good things,”; “Today I want to show some benefits and good things in front of my family,”; “Today in the live broadcast room, I want to bring some benefits and good things to my family,”; “Today in the live broadcast room, I want to bring you a lot of benefits and good things,”

[0051] "Remember to follow us,"; "Don't forget to follow us,"; "Follow us, don't forget,"; "Remember to follow,"; "Don't forget to follow, remember,"

[0052] "This way you won't miss any good content!"; "This way you won't miss any wonderful moments!"; "This way you won't miss any exciting content!"; "This way you won't miss any interesting content!"; "This way you won't miss any wonderful programs!"

[0053] "Family, I am waiting for your arrival here," "Family in the studio, I am waiting for you here," "Dear family, I am waiting for your presence here," "Dear family, I am looking forward to your appearance here," "Family, I am looking forward to your arrival in the studio,"

[0054] "Come on, let's get excited together!"; "Let's have fun together!"; "Let's heat up the atmosphere together!"; "Let's have a carnival together!"; "Let's join the carnival together!"

[0055] Together with the original text unit, each text unit can have 6 candidate segments.

[0056] Phase III: candidate fragment reassembly.

[0057] Based on the target length of 1000 words and the original length of 75 words, it is calculated that 14 reorganizations are required (1000 / 75 and rounded up). The simplest random sampling is adopted: randomly selecting a candidate each time, and the following new live copy can be reorganized:

[0058] Family, welcome to our live broadcast room! Today I have brought you a lot of benefits and good things to share. Don’t forget to follow us so that you won’t miss any exciting content! The host is waiting for you here, let’s have fun together!

[0059] Family, welcome to our live broadcast room! Today I will bring you a lot of benefits and good things in the broadcast room. Pay attention and don’t forget to pay attention, so that you won’t miss any good content! The anchor is waiting for everyone here, let’s have a carnival together!

[0060] Dear friends, don't forget to come to our live studio! Today I will bring you a lot of benefits and goodies in the live studio. Remember to follow us so that you won't miss any exciting content! Dear friends in the live studio, I'm waiting for you here, let's join the carnival together!

[0061] Welcome, my dear family, to our live broadcast room! Today I want to share a lot of benefits and good things with my family. Remember to follow me so that you won’t miss any wonderful moments! Dear family, I am waiting for your presence here, let’s have a carnival together!

[0062] Welcome, my dear friends, to our live studio! Today I have brought you a lot of benefits and good things to share. Remember to follow me so that you won’t miss any interesting content! My dear friends in the studio, I am waiting for you here, let’s join the carnival together!

[0063] My dear friends, the live broadcast room is ready and waiting for you! Today I will bring you a lot of benefits and goodies in the live broadcast room. Don’t forget to follow me so that you won’t miss any interesting content! My dear friends, I am looking forward to your appearance here, let’s have a carnival together!

[0064] Family members, don't forget to come to our live broadcast room! Today in the live broadcast room, I will bring some benefits and good things to the family members, don't forget to pay attention, remember. In this way, you won't miss any wonderful programs! The anchor is waiting for everyone here, let's have fun together!

[0065] Family, come and join our broadcast room! Today I have brought you a lot of benefits and good things to share. Don't forget to follow us. Remember. This way you won't miss any wonderful programs! The host is waiting for you here, let's have fun together!

[0066] Welcome, my dear friends, to our live studio! Today I will bring you a lot of benefits and goodies in the studio. Don't forget to follow me so that you won't miss any good content! Dear friends, I am waiting for your presence here. Let's have fun together!

[0067] My dear friends, the live broadcast room is ready and waiting for you! Today in the live broadcast room, I will bring you a lot of benefits and goodies. Don't forget to pay attention, remember. In this way, you won't miss any exciting content! My dear friends, I am here waiting for you to come, come, let's get excited together!

[0068] My dear friends, the live studio is ready for you! Today in the live studio, I will bring you a lot of benefits and goodies. Remember to follow us so that you won’t miss any good content! My dear friends, I am looking forward to your arrival in the live studio. Let’s have fun together!

[0069] Family members, don't forget to come to our live broadcast room! Today I have brought you a lot of benefits and good things to share. Remember to follow us so that you won't miss any exciting content! The anchor is waiting for everyone here, let's have fun together!

[0070] Family, welcome to our live broadcast room! Today in the broadcast room, I will bring you a lot of benefits and goodies. Pay attention and don’t forget to pay attention, so that you won’t miss any interesting content! The anchor is waiting for you here, come, let’s get excited together!

[0071] Dear friends, don't forget to come to our live studio! Today I will bring you a lot of benefits and goodies in the live studio. Don't forget to follow us so that you won't miss any exciting content! Dear friends in the live studio, I'm waiting for you here, let's have fun together!

[0072] The longest repeated segment is "Don't forget to follow, remember. This way you won't miss any wonderful programs! The host is waiting for everyone here, let's have fun, let's have fun together!", with a length of 46 characters (40 characters excluding punctuation).

[0073] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A copywriting system for digital human live broadcast, characterized in that: The method includes the following steps: an existing speech copy is input into the text unit splitting module of the speech copy expansion module, and the text unit splitting module generates text unit 1, text unit 2, text unit 3 and text unit n, and then transmits them to the text rewriting module; the text rewriting module generates several candidate segments of text unit 1, several candidate segments of text unit 2, several candidate segments of text unit 3, and several candidate segments of text unit n, and several candidate segments of text unit n (n is a positive integer) are respectively transmitted to the candidate segment reorganization module, and the candidate segment reorganization module generates several reorganized new speech copies, and the target live broadcast duration and target non-repetition degree module are connected to the candidate segment reorganization module and the text unit splitting module.

2. A text rewriting system for digital human live broadcast according to claim 1, characterized in that: Target live broadcast duration and target non-repetition module: According to the length of the original script, the target broadcast duration and the speed of the digital human speaking, the number of times the script text needs to be reorganized m is directly calculated: The target non-repetition degree means that within m reorganizations, any text segment of length k does not appear twice. According to the length of the text unit, the number of text units in the text segment of length k is calculated: where n k The text unit is the number of text segments contained in the k-word segment. is the average length of the text segment, Ceil is rounded up, and candidates are randomly selected and combined between text units. In order to ensure that there are enough candidates to ensure that texts of consecutive k-word lengths are not repeated in m reorganizations, the number of candidates for this text needs to meet the following requirements: where n 候选 The number of candidates available for each text unit is calculated based on the target non-repetitiveness k and the number of reorganizations m to obtain a set of n that satisfies the above inequality. 候选 and The target is repeated m = 100 times, and the non-repetitiveness is that there is no repetition in a continuous segment of k = 100 words. Pick n 候选 =5 (i.e., using the large language model to generate 4 candidates for each text unit, plus the original expression, there are a total of 5 candidates), the number of candidate combinations is The minimum requirement of 100 reorganizations can be met.

3. A copywriting system for digital human live broadcast, characterized in that: Based on an existing live broadcast script, the target script length (or playback time) and the target script non-repetition rate, expand the content of the original script with several different expressions that do not change the original meaning to ensure that the text content is not repeated.

4. A text rewriting system for digital human live broadcast according to claim 3, characterized in that: The steps are as follows: The first step is to split the original text into smaller units, phrases, sentences or paragraphs. In the second step, for each of the split basic units, the AI ​​capabilities of the large language model to understand and generate texts are used to generate several new candidate texts with the same content as the original text but in different expressions. The third step, after generating several new candidate texts for each text unit, one candidate can be selected from the candidate text fragments each time to reorganize a new speech copy. The meaning expressed by this reorganized speech copy is the same as the original text, and a new expression method that is not repeated in the original text is used. By combining different candidate fragments, a large number of non-repetitive texts can be combined. In the scenario of long-term live broadcast, the above reorganization steps can be repeated many times to reorganize speech copy whose length meets the target live broadcast duration. The number of non-repetitive texts that can be obtained is exponentially related to the number of text units split. There are many ways to select candidate fragments. The most basic random sampling can be used, or the selection can be staggered according to the statistical combination of adjacent fragments.

Images