Virtual live broadcast control method and system
By constructing a personalized live broadcast theme text automatic generation method with knowledge-related relationships, the problem of single virtual live broadcast method and insufficient intimacy of virtual anchors is solved, the personalization and humor ability of virtual anchors is improved, and the fun and interactiveness of live broadcasts is enhanced. .
Patent Information
- Application Number
- CN202210529215.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-16
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2042-05-16
Smart Images

Figure CN115348458B_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present specification relate to the field of computer technology, and in particular to a virtual live broadcast control method. Background Art
[0002] With the rise of live streaming platforms, more and more users are using live streaming to promote various information in the live streaming rooms of live streaming platforms, such as selling goods on the live streaming platforms, teaching on the live streaming platforms, performing on the live streaming platforms, etc.
[0003] With the evolution and development of virtual reality / virtual reality / digital human technology, it has become a trend to use digital people or virtual characters to replace real anchors to overcome the training costs of real anchors, service time limits, and supporting hardware requirements for live broadcast rooms, so that live broadcasts can be conducted without time or space restrictions. In addition, in order to enhance the fun and interactivity of live broadcasts, virtual image live broadcasts, as a very important part of live broadcast services, have occupied an increasingly large proportion of live broadcast services in recent years.
[0004] However, the live broadcasting mode of the live broadcasting room in the current live broadcasting platform is fixed and single, which cannot better meet the needs of users. Therefore, an effective method is urgently needed to solve such problems. Summary of the invention
[0005] In view of this, an embodiment of this specification provides a virtual live broadcast control method. One or more embodiments of this specification also involve a virtual live broadcast control system, a computing device, a computer-readable storage medium and a computer program to solve the technical defects existing in the prior art.
[0006] According to a first aspect of an embodiment of this specification, a virtual live broadcast control method is provided, including:
[0007] Set up virtual characters to broadcast information on live broadcast topics in the virtual live broadcast room;
[0008] Acquire key information related to the live broadcast theme, and extract keywords contained in the key information based on a pre-built keyword map;
[0009] Extracting associated words related to the keyword from a pre-constructed associated word graph;
[0010] Acquire a hot word related to the associated word, wherein a first degree of association between the hot word and the associated word is greater than a second degree of association between the hot word and the keyword;
[0011] Based on the keywords, the associated words and the hot words, generating a live broadcast theme text related to the live broadcast theme, wherein the live broadcast theme text includes at least two paragraphs, and the first paragraph consists of the keywords and the associated words, and the second paragraph consists of the hot words and the associated words;
[0012] A rendering model of the virtual character is obtained and the virtual character is driven to broadcast according to the live broadcast theme text.
[0013] According to a second aspect of an embodiment of this specification, a virtual live broadcast control system is provided, including:
[0014] A setting module is configured to set a virtual character to broadcast information on a live broadcast theme in a virtual live broadcast room;
[0015] A first acquisition module is configured to acquire key information related to the live broadcast theme and extract keywords contained in the key information based on a pre-built keyword map;
[0016] An extraction module, configured to extract associated words related to the keyword from a pre-built associated word graph;
[0017] A second acquisition module is configured to acquire a hot word related to the associated word, wherein a first degree of association between the hot word and the associated word is greater than a second degree of association between the hot word and the keyword;
[0018] A generating module is configured to generate a live broadcast theme text related to the live broadcast theme based on the keywords, the associated words and the hot words, wherein the live broadcast theme text comprises at least two paragraphs, and the first paragraph consists of the keywords and the associated words, and the second paragraph consists of the hot words and the associated words;
[0019] The driving module is configured to obtain the rendering model of the virtual character and drive the virtual character to broadcast according to the live broadcast theme text.
[0020] According to a third aspect of an embodiment of this specification, a computing device is provided, including:
[0021] Memory and processor;
[0022] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement any step of the virtual live broadcast control method.
[0023] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided, which stores computer-executable instructions, and when the instructions are executed by a processor, the steps of any one of the virtual live broadcast control methods are implemented.
[0024] According to a fifth aspect of the embodiments of this specification, a computer program is provided, wherein when the computer program is executed in a computer, the computer is caused to execute the steps of the above-mentioned virtual live broadcast control method.
[0025] An embodiment of the present specification sets a virtual character to broadcast information on a live broadcast theme in a virtual live broadcast room, obtains key information related to the live broadcast theme, and extracts keywords contained in the key information based on a pre-constructed keyword map, extracts associated words related to the keywords in a pre-constructed associated word map, obtains hot words related to the associated words, wherein a first degree of association between the hot words and the associated words is greater than a second degree of association between the hot words and the keywords, generates a live broadcast theme text related to the live broadcast theme based on the keywords, the associated words and the hot words, wherein the live broadcast theme text includes at least two paragraphs, and the first paragraph consists of the keywords and the associated words, and the second paragraph consists of the hot words and the associated words, obtains a rendering model of the virtual character and drives the virtual character to broadcast according to the live broadcast theme text.
[0026] In the embodiment of this specification, when setting a virtual character to broadcast information on a live broadcast theme in a virtual live broadcast room, key information related to the live broadcast theme can be obtained, and then a live broadcast theme text can be jointly constructed based on the keywords in the key information, the associated words related to the keywords, and the hot words related to the associated words, and the virtual character can be driven to broadcast the live broadcast theme text. By generating and broadcasting the live broadcast theme text in real time, it is conducive to enhancing the fun and interactivity of the live broadcast, and at the same time, it is conducive to achieving the diversity of live broadcast methods, so as to better meet the needs of users for different types of information. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 is a flow chart of a virtual live broadcast control method provided by an embodiment of this specification;
[0028] Figure 2a It is a schematic diagram of a construction result of a knowledge graph provided by an embodiment of this specification;
[0029] Figure 2b It is a schematic diagram of a live broadcast theme text construction process provided by an embodiment of this specification;
[0030] Figure 2c It is a schematic diagram of a live text construction process provided by an embodiment of this specification;
[0031] Figure 2d It is a schematic diagram of another live text construction process provided by an embodiment of this specification;
[0032] Figure 3 is a processing flow chart of a virtual live broadcast control method provided by an embodiment of this specification;
[0033] Figure 4 It is a structural diagram of a virtual live broadcast control system provided by an embodiment of this specification;
[0034] Figure 5 It is a structural block diagram of a computing device provided by an embodiment of this specification. DETAILED DESCRIPTION
[0035] Many specific details are described in the following description to facilitate a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of this specification, so this specification is not limited to the specific implementation disclosed below.
[0036] The terms used in one or more embodiments of this specification are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of this specification. The singular forms of "a", "said" and "the" used in one or more embodiments of this specification and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.
[0037] It should be understood that although the terms first, second, etc. may be used to describe various information in one or more embodiments of this specification, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of this specification, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".
[0038] First, the terms involved in one or more embodiments of this specification are explained.
[0039] Digital human: A virtual character with a digital appearance and the ability to express through language, facial expressions and body movements.
[0040] Live broadcast room: Each live broadcast window corresponds to a live broadcast room. The live broadcast room is unique in a service platform.
[0041] Virtual live broadcast room: A digital human anchor and digital live broadcast room built using digital drive technology, which provides live broadcast content in an interactive way of simulated human live broadcast. It can also be understood as a specific type of live broadcast room in the live broadcast room (a type where a virtual character acts as the anchor of the live broadcast room to broadcast live), for example, it can include but is not limited to game live broadcast rooms, film and television live broadcast rooms, life live broadcast rooms, comprehensive live broadcast rooms, etc. The virtual live broadcast room here can be any live broadcast room, and the virtual live broadcast room may include digital human anchors, scenes, live broadcast texts and other components.
[0042] Pre-trained language model: A model trained using large-scale unlabeled corpus and self-supervision methods. It is usually used as an initialization model for downstream tasks to improve the performance of downstream tasks.
[0043] Data to text generation (data2text): Given some structured data, such as tables, knowledge graphs, etc., generate corresponding text containing structured data content.
[0044] Slot filling: Filling the empty spaces in the text to make the text content complete. The filling involved in the embodiments of this specification usually requires a text template with slots and a corresponding generation / filling method.
[0045] Mining high-frequency popular online terms: Obtaining short phrases and sentences that appear frequently in the corpus through statistical or sequence labeling models and other methods.
[0046] Human-computer interaction: Similar to the interaction between real people, human-computer interaction refers to communication between people and virtual anchors in the virtual live broadcast room through methods such as question and answer and dialogue.
[0047] In live broadcast, the main responsibility of the anchor is to introduce and explain the product or news content to the users in the live broadcast room, including answering questions about the product or detailed information about the news. In addition to these contents to attract users to click and watch, if the content can be highlighted in terms of fun and intimacy, it can attract users to watch and drive them to stay and watch subjectively. For example, some occasional marketing methods in live broadcasts, including humor, novelty, and eye-catching methods, allow users to stop and watch. With the evolution and development of virtual reality / virtual reality / digital human technology, it has become a trend to use digital people or virtual characters to replace real anchors to overcome the training costs of real anchors, service time restrictions, and supporting hardware requirements of live broadcast rooms, so as to be able to broadcast live without time or space restrictions. In virtual live broadcast rooms, virtual anchors can currently achieve relatively objective and professional live broadcast capabilities, but they are inferior to real anchors in terms of intimacy with users, character personalization, and fun. Specifically, when virtual anchors broadcast live, if they want to achieve the effect of real anchors, there are the following problems:
[0048] 1. Generally, users of live broadcast rooms have a more natural affinity towards real anchors than virtual anchors. This is because in human cognition, virtual anchors are robots without personality and do not have the conditions to be human. Therefore, how to enhance users' intimacy with virtual anchors is a point that needs to be improved at present;
[0049] 2. Compared with virtual anchors, real anchors can not only explain products or news professionally during live broadcasts, but also create interesting content in real time, which is tied to the character's personality. Differentiated personalities and highly creative expressions of interesting content (such as telling jokes, etc.) are very lacking in current virtual anchors. So how to make virtual anchors also have a certain degree of personalization and humor is also a point for improvement.
[0050] Based on this, the embodiment of this specification proposes a method for automatically generating personalized live broadcast theme texts around knowledge association. In order to deal with the two problems mentioned above, the embodiment of this specification uses large-scale humorous corpus, personalized joke corpus and corresponding multiple modeling methods, and uses artificial intelligence technology to help virtual anchors build the ability to automatically generate live broadcast theme text libraries, and cooperate with interactive methods to support virtual anchors in the live broadcast process. The ability to tell jokes or talk shows in a specific style, ultimately enables virtual anchors to have personalization and humor capabilities to a certain extent, and narrows the distance between users and virtual anchors.
[0051] In this specification, a virtual live broadcast control method is provided. This specification also relates to a virtual live broadcast control system, a computing device, a computer-readable storage medium, and a computer program, which are described in detail one by one in the following embodiments.
[0052] Figure 1 A flowchart of a virtual live broadcast control method provided according to an embodiment of the present specification is shown, which specifically includes the following steps.
[0053] Step 102, setting a virtual character to broadcast information on the live broadcast theme in the virtual live broadcast room.
[0054] Specifically, the live broadcast room is usually used for live broadcast of commodities or games, but in case of a special live broadcast theme, a virtual character in the virtual live broadcast room can be set up to broadcast around the live broadcast theme. The live broadcast theme can be a medium, medium-large or large event, competition or conference.
[0055] In order to ensure the comprehensiveness and completeness of the information related to the live broadcast theme, a virtual character is set up to broadcast different types of information related to the live broadcast theme during the live broadcast around the live broadcast theme.
[0056] The virtual live broadcast control method provided in the embodiments of this specification sets a virtual character to perform live broadcast, which can ensure that the virtual character can broadcast live without interruption throughout the day, which can not only increase the fun of the virtual character's live broadcast, but also fill the time gap of the real character's live broadcast, allowing the audience to watch different types of information related to the live broadcast theme at any time.
[0057] Step 104, obtaining key information related to the live broadcast theme, and extracting keywords contained in the key information based on a pre-constructed keyword map.
[0058] Specifically, the key information includes key news information with a first narrative style; and the key news information with a first narrative style may be information containing an objective introduction, narration, description, explanation or objective evaluation of general news, encyclopedia news, event news, etc.
[0059] A keyword graph is a knowledge graph consisting of multiple keywords as nodes and the relationships between keywords as edges. The edges in the keyword graph have a pointing relationship, which can be used to represent the existence of a relationship between the two nodes connected by the edge, and can also be used to determine the path formed by adjacent nodes in the keyword graph.
[0060] In actual applications, the construction of keyword maps often relies on keywords with objective common sense, and the keyword maps can be updated regularly by constantly mining new keywords. For example, from "I saw that athlete A's action in the competition can do a double somersault and turn 1440 degrees," we can extract keywords such as "athlete A, action, double somersault and turn 1440 degrees", and there can be a directional association relationship between these keywords such as "athlete A->action->double somersault and turn 1440 degrees".
[0061] Among them, since the viewing users in the live broadcast room can send comment information on the live broadcast content during the live broadcast viewing process, the key information in the embodiments of this specification can be obtained by extracting information from the comment information, or by extracting information from the live broadcast content. The specific information can be determined according to actual needs and is not limited here.
[0062] After obtaining key information related to the live broadcast theme, the keywords contained in the key information can be extracted based on the pre-constructed keyword map. Specifically, the keywords contained in the keyword map can be compared with the key information. If the comparison is consistent, it can be determined that the key information contains the consistent keyword.
[0063] After extracting the keywords, since the keywords will be used to construct the live broadcast theme text later, in the process of constructing the live broadcast theme text, other keywords in the keyword map that are associated with the keyword can also be determined, and the keyword and the keywords that are associated with the keyword can be used to jointly construct the live broadcast theme text to achieve the diversity of the generated live broadcast theme text.
[0064] Taking the virtual live broadcast control method provided in the embodiments of this specification as an example of being applied to an e-commerce virtual live broadcast scenario, if the live broadcast theme is a skiing competitive activity, the key news information of the first narrative style can be news, events, encyclopedias, and other information about the skiing activity.
[0065] Since the key information includes key news information with a first narrative style, when a virtual character is set to broadcast news information in the first narrative style, the virtual character can be driven to objectively broadcast the news information in the first narrative style according to the first broadcasting style, thereby reflecting the objectivity and authenticity of the news information in the first narrative style.
[0066] Step 106: extracting associated words related to the keyword from the pre-constructed associated word graph.
[0067] Specifically, the associated words include interesting associated words with the second narrative style, and the interesting associated words of the second narrative style can be words containing subjective, colloquial or even non-real content such as jokes, humorous segments, talk shows, etc. Taking the virtual live broadcast control method provided in the embodiment of this specification as an example of being applied to an e-commerce virtual live broadcast scenario, if the live broadcast theme is a skiing competitive event, the interesting keywords of the second narrative style can be words in a small joke or segment related to the competitive event in the skiing competitive event.
[0068] An associative word graph is a knowledge graph consisting of multiple associative words as nodes and the associative relationships between the associative words as edges. The edges in the associative word graph have a pointing relationship, which can be used to represent the existence of an associative relationship between the two nodes connected by the edge, and can also be used to determine the path formed by adjacent nodes in the associative word graph.
[0069] In practical applications, the construction of a graph of associated words can be carried out by analyzing and extracting words from online texts and speech recognition texts (ASR texts) of entertainment programs to mine interesting associated words. For example, from the sentence “We can obtain 1440 resource values in one day,” we can extract associated words such as “one day, obtain, 1440 resource values,” and there can be a directional association relationship between these associated words such as “one day->obtain->1440 resource values.”
[0070] After extracting the keywords contained in the key information, you can continue to extract the associated words related to the keywords in the associated word map.
[0071] In specific implementation, the step of extracting associated words related to the keyword in a pre-constructed associated word graph includes: determining the word unit corresponding to each node in the pre-constructed associated word graph; semantically matching the keyword with the word unit, and determining the associated words related to the keyword in the associated word graph based on the matching results.
[0072] Specifically, since each node in the associated word map corresponds to an associated word, when extracting associated words related to keywords from the associated word map, the extraction can be performed by semantic matching. For example, if the keyword is "athlete A", the keywords associated with the keyword in the keyword map include "action" and "two-way somersault 1440 degrees", etc. In this case, when determining the associated words of the keyword, first determine that the associated word of "two-way somersault 1440 degrees" can be "1440 resource values" (both contain 1440, and the semantics of 1440 are the same), and the associated word of "1440 resource values" in the associated word map can be "acquisition", so the determined associated words of "athlete A" include "1440 resource values" and "acquisition", etc.
[0073] After extracting the associated words, since the associated words are needed to construct the live broadcast theme text later, in the process of constructing the live broadcast theme text, other associated words in the associated word map that have an associated relationship with the associated word can also be determined, and the associated word and the associated words that have an associated relationship with the associated word can be used to jointly construct the live broadcast theme text to achieve the diversity of the generated live broadcast theme text.
[0074] Step 108: Obtain hot words related to the associated words, wherein a first degree of association between the hot words and the associated words is greater than a second degree of association between the hot words and the keyword.
[0075] Specifically, hot words are words that are popular or frequently used on the Internet, usually Internet terms.
[0076] In a specific implementation, the step of obtaining hot words related to the associated words includes: obtaining network hot information according to a preset time period; inputting the network hot information into a sequence labeling model for processing to generate a sequence labeling result corresponding to the network hot information; and determining the hot words related to the associated words in the network hot information according to the sequence labeling result.
[0077] Specifically, since the hot words in the network will continue to change over time, when obtaining the hot words related to the associated words, the network hot information can be continuously re-acquired according to a certain time period, and the acquired network hot words can be processed through the sequence labeling model. Specifically, the hot words in the network hot information are sequence-labeled through the sequence labeling model, and then the hot words related to the associated words are determined among the words labeled as hot words.
[0078] Alternatively, firstly, entities are extracted or matched from the online hot rankings through the entity dictionary to obtain the corresponding hot words. The timeliness of these hot words can be updated and replaced according to the changing speed of the online hot words, such as monthly or annual updates. In particular, some old or inappropriate memes may have a counterproductive effect if added to the live broadcast theme text, so they will be updated regularly.
[0079] For example, if the determined associated words include "1440 resource values" and "obtain", and the words marked as hot words include "can this be said", then the hot words related to the associated words can be determined based on semantics, namely "can this be said".
[0080] Among them, the correlation between hot words and related words is greater than the correlation between hot words and keywords.
[0081] After extracting the popular words, the live broadcast theme text can be constructed by using the keywords, related words and popular words to achieve the diversity of the generated live broadcast theme text.
[0082] Step 110, based on the keywords, the associated words and the hot words, generate a live broadcast theme text related to the live broadcast theme, wherein the live broadcast theme text includes at least two paragraphs, and the first paragraph consists of the keywords and the associated words, and the second paragraph consists of the hot words and the associated words.
[0083] Specifically, since the virtual characters usually express the text content of the live broadcast theme text corresponding to the live broadcast theme, it eventually becomes a live broadcast stream. The playback engine provides multiple playback capabilities such as sequence, timing, loop, and insertion. Whether it is based on the pre-edited live broadcast theme text or the temporary decision-making live broadcast content produced by the events or behaviors in the live broadcast room, it can be expressed in a timely manner.
[0084] Therefore, when using a virtual character to broadcast a live broadcast on a live broadcast theme, a live broadcast theme text for the virtual character to broadcast can be generated based on real-time or pre-acquired keywords, related words, and popular words, thereby driving the virtual character to broadcast the live broadcast theme text.
[0085] In specific implementation, the step of generating a live broadcast theme text related to the live broadcast theme based on the keywords, the associated words and the popular words includes: constructing a knowledge graph corresponding to the keywords based on the association relationship between the keywords, the associated words and the popular words; generating a live broadcast theme text related to the live broadcast theme and having a second narrative style based on word units contained in the path formed by adjacent nodes in the knowledge graph.
[0086] Furthermore, based on the association relationship between the keywords, the associated words and the popular words, the step of constructing a knowledge graph corresponding to the keywords includes: taking the keywords, the associated words and the popular words as nodes, and taking the association relationship between the keywords, the associated words and the popular words as edges, to construct the knowledge graph corresponding to the keywords.
[0087] Specifically, after obtaining keywords, associated words, and popular words, the keywords, associated words, and popular words can be used as nodes, and the associations between the keywords, associated words, and popular words can be used as edges to construct a knowledge graph corresponding to the keywords. The edges between the nodes in the knowledge graph can also have a pointing relationship. A schematic diagram of the construction result of a knowledge graph provided in an embodiment of this specification is as follows Figure 2b shown.
[0088] Therefore, when using it, starting from the given keyword (athlete A), by retrieving the directional relationship between the edges of each node in the knowledge graph, the path formed by the keyword and its adjacent nodes in the knowledge graph can be obtained. The path can be specifically: athlete A->action->two somersaults 1440 degrees->1440 resource values<-acquisition.
[0089] Then, based on the word units contained in the path, the live broadcast theme text related to the live broadcast theme can be generated.
[0090] In an embodiment of the present application, a neural network model can be used to generate a live broadcast theme text. Therefore, based on the keywords, the associated words and the popular words, a live broadcast theme text related to the live broadcast theme is generated. Specifically, the keywords, the associated words and the popular words can be input into a text generation model for processing to generate a live broadcast theme text related to the live broadcast theme.
[0091] Specifically, the text generation model is the data2text (data to text generation) model. By giving some structured data, such as tables, knowledge graphs, etc., it can generate corresponding text containing structured data content. Therefore, after the embodiment of this specification generates a knowledge graph based on keywords, related words and popular words, the knowledge graph can be input into the text generation model for processing to generate the corresponding live broadcast theme text, and the live broadcast theme text can have a second narrative style.
[0092] The text generation model is trained in the following manner: a reference text having a second narrative style is obtained; the reference text is used as sample data and input into a pre-trained text generation model for training to obtain the text generation model, wherein the pre-trained text generation model masks keywords contained in the reference text and predicts the masked keywords, so as to adjust the model parameters of the pre-trained text generation model based on the prediction results.
[0093] Specifically, in order to enable the trained text generation model to output a live broadcast theme text with a second narrative style, when training the text generation model to be trained, a reference text with a second narrative style can be obtained first. The reference text can be in the form of a knowledge graph, a table or text, etc., and then the reference text is used as sample data and input into the pre-trained text generation model for training to obtain the trained text generation model. During the training process, the pre-trained text generation model can mask the keywords contained in the reference text and predict the masked keywords, so as to adjust the model parameters of the pre-trained text generation model based on the prediction results and the actual results of the masked keywords, thereby ensuring the accuracy of the training results.
[0094] Alternatively, the step of generating a live broadcast theme text related to the live broadcast theme based on the keywords, the associated words and the popular words can also be achieved in the following way: obtaining a reference text with a second narrative style, and constructing a text generation template related to the live broadcast theme based on the reference text, wherein the text generation template contains word slots to be filled; based on the keywords, the associated words and the popular words, the word slots to be filled are filled to generate a live broadcast theme text related to the live broadcast theme.
[0095] Furthermore, the step of constructing a text generation template related to the live broadcast theme based on the reference text includes: obtaining a template type of a preset text generation template; inputting the reference text into a semantic analysis model for processing to obtain a corresponding semantic analysis result; determining a target reference text that matches the template type based on the semantic analysis result; and determining a word slot to be filled in the target reference text based on the target reference text to generate a text generation template related to the live broadcast theme.
[0096] Specifically, when constructing the live broadcast theme text using keywords, related words, and popular words, a text generation template can be pre-constructed. In order to make the constructed live broadcast theme text have the second narrative style, when constructing the text generation template, its template type can be determined as humorous and funny. And the reference text for constructing the live broadcast theme text can also have the second narrative style. Therefore, the reference text can be a text obtained by semantic recognition of a talk show, or it can be a talk show script. Then, the target reference text that meets the template type in the reference text, that is, the paragraph, can be identified through the semantic classification model, and some of the words therein can be deleted. The position of the deleted words in the paragraph becomes the word slot to be filled. At this point, the paragraph that deletes some words and retains the word slot to be filled can be used as a text generation template related to the live broadcast theme text.
[0097] In this way, text generation templates of various template types can be quickly expanded, which is conducive to maintaining the diversity of live broadcast theme texts generated based on different text generation templates.
[0098] Step 112, obtaining a rendering model of the virtual character and driving the virtual character to broadcast according to the live broadcast theme text.
[0099] Specifically, when a live broadcast is performed on a live broadcast theme using a virtual character, the virtual character may be driven to perform the live broadcast by broadcasting the live broadcast theme text of the first type of information and / or the second type of information.
[0100] In the specific implementation, since different virtual characters may have different character styles, for example, some virtual characters may have a humorous style, and some virtual characters may have a mature and intellectual style, the live broadcast theme text can be stylized and rewritten according to the character style of the virtual anchor. Specifically, after generating the live broadcast theme text related to the live broadcast theme, the text content of the live broadcast theme text can also be adjusted according to the preset character style of the virtual character to generate a target live broadcast theme text that matches the preset character style.
[0101] In summary, when the live broadcast theme text is a paragraph, the schematic diagram of the live broadcast theme text construction process provided in the embodiment of this specification is as follows: Figure 2aAs shown, it mainly includes: the construction of the original humor corpus, the training of the pre-trained language model, the generation of jokes, and personalized rewriting. Among them, the construction of the original corpus is mainly achieved by collecting content sources such as humorous jokes, crosstalk manuscripts, and talk show manuscripts publicly available on the Internet, while the pre-trained language model uses a large-scale language model and a large-scale training corpus mined. By masking some phrases or words in the text of the corpus, the model predicts the words corresponding to the masked position and learns how to construct a context with humorous language characteristics. For the generation of jokes, it is based on the learned language model, combined with the mined keywords, related words, and hot words to generate the corresponding jokes, and through personalized rewriting, the generated jokes are rewritten in a specified style while keeping the semantic expression content basically unchanged, and finally stored in the joke database for interactive use by online talk shows.
[0102] Among them, there are mainly two methods for generating jokes. The first is to generate corresponding jokes through the data2text method given keywords and corresponding associated words and popular words; the second method is to construct a speech template by analyzing the common techniques of talk show jokes, and finally complete the construction of the joke by filling in the slots through given keywords, associated words and popular words.
[0103] Among them, a schematic diagram of a live text construction process provided in an embodiment of this specification is as follows Figure 2c As shown, the process is as follows:
[0104] 1. First, extract the corresponding key information from the network information.
[0105] 2. The obtained key information can be compared with the keywords in the keyword map to determine the keywords contained in the key information. Then, the keywords can be semantically matched with the associated words in the associated word map to determine the associated words related to the keywords. Then, the associated words can be semantically matched with the pre-acquired network hot words to determine the network hot words related to the associated words.
[0106] Among them, the construction of the keyword graph often relies on the knowledge graph based on objective common sense, and the keyword graph can be updated and replaced regularly.
[0107] For the construction of the associated word map, interesting related information can be mined by analyzing and extracting models from online texts and ASR texts of entertainment TV programs. For example, from "We can obtain 200 resource values a day.", we can extract "obtain -> {} resource value".
[0108] For the construction of Internet hot words, we can extract or match entities from the Internet hot rankings through a sequence annotation model and an entity dictionary to obtain the corresponding Internet hot words.
[0109] 3. Based on the existing keywords, related words and Internet buzzwords, the corresponding live broadcast text is generated through the data2text generation model.
[0110] In addition, another schematic diagram of a live text construction process provided in an embodiment of this specification is as follows: Figure 2d As shown, the process is as follows:
[0111] 1. Obtain keywords, related words, and Internet hot words through the above methods.
[0112] 2. By analyzing the common methods of talk show segments, corresponding speech templates are sorted out, and according to the characteristics of the speech templates, template types are summarized. The live texts that meet the speech templates are identified from talk show programs and talk show scripts through semantic classification models, and some phrases are deleted. The positions of the deleted words in the live text become the word slots to be filled. The segments that delete some words and retain the word slots to be filled can be used as speech templates related to the live theme text.
[0113] In one embodiment, there are two main types of templates:
[0114] a. Association: Given a keyword G, construct a related word B that has a certain correlation with the keyword G and can visualize the relationship, such as playing ice hockey and mopping the floor;
[0115] b. Expected violation. Given a keyword G, give an associated word B that is related to G but violates the normal cognition of G. For example, for the number 1440, the associated word related to the athlete may be "two-week somersault with a 1440-degree turn", while the associated word related to it but violates the normal cognition of it may be "able to obtain 1440 resource values in two weeks".
[0116] 3. Starting from the given keywords, fill in the keywords, related words, and Internet buzzwords into the appropriate positions of the speech template. The specific implementation is as follows:
[0117] The key words are athlete A; the associated words are action, two somersaults with a turn of 1440 degrees, and being able to obtain 1440 resource values in two weeks; the Internet buzzword is "Can this be said?"
[0118] Speech template: I saw [k1] and [k2] in the competition can do [k3]. It’s nothing special. I [k5]. [k6].
[0119] Generated live text: I saw athlete A doing a double somersault and turning 1440 degrees in the competition. It's nothing special. I can get 1440 resource points in two weeks. Is this something I can say?
[0120] In addition, after the live text is generated, it can be personalized and rewritten. Personalized rewriting of live text means that given a live text D1 and a specified text W with a certain style, a live text D2 is generated through a style transfer model while keeping the basic semantics of the live text unchanged. The style of the live text D2 is transferred to the style corresponding to the given text W. The specific process is as follows:
[0121] 1. Style Decoupling: For a given style text B (Style Text), an encoder and a style decoupling module are used to convert the text into a content-independent stylized vector (Style Embedding).
[0122] 2. Graph-enhanced plan generation: Extract the corresponding semantically represented keywords from the given segment A, and use the graph network to predict the order in which these keywords appear in the live text A, thereby rearranging these keywords. Finally, the decoder and the stylized vector obtained in the first step are used to generate the corresponding stylized target text C (Stylized Target Text).
[0123] 3. When using it, you only need to give the text corpus that does not have a specific style and the corpus with a certain style type. Through this model, you can achieve arbitrary switching between styles.
[0124] After the live broadcast text is generated, the virtual character can be driven to interact and broadcast with the live broadcast viewers based on the jokes. For example, the user clicks on the keyword of the live broadcast text he wants to hear, and based on the keywords clicked by most users, one or at least two related jokes are retrieved. Then, according to the status of the live broadcast room (such as the number of people, user distribution, group portrait, etc.), the live broadcast text to be broadcast is dynamically decided, and the virtual character is driven to broadcast it.
[0125] This solution produces live texts through a variety of artificial intelligence methods, and relies on large-scale pre-trained models and rich knowledge mined from the Internet, thereby avoiding the repetitiveness and uniqueness of live texts. Or through the learning of large-scale pre-trained models, and with the help of slot filling methods, on the premise of building a highly accurate associated knowledge graph, it avoids problems such as incoherent language and inconsistent logic caused by the generation method. Through intelligent learning of a large number of jokes, talk shows and other content, as well as summarized talk show common skills, the produced live texts can have a certain text style in the form of language expression, and through the constructed associated knowledge, the live texts can maintain a certain degree of fun. Through intelligent stylized migration, the virtual anchor can continue to narrate live texts with a certain style, thereby helping the virtual anchor to establish personalized characteristics and avoid the loss of affinity due to lack of personality.
[0126] In the embodiments of this specification, the intelligent live text production schemes are diverse, so even if the same data is given, the live text produced can maintain a certain diversity, and the stylized migration can rewrite the same live text into a variety of different styles, thereby maintaining the diversity of the live text. In addition, the language logic relies on the slot filling method and the associated knowledge graph and other data with strong structured features, so that the model can produce and modify the live text within a limited range, thereby controlling the generated live text semantic expression within a controllable range to a certain extent, avoiding unpredictable results. Through the learning of large-scale corpus and the construction of associated knowledge graphs, the produced live text has a certain text style in the form of language expression, relying on the graph to ensure good humor and fun. The live text has a certain strong personality language expression (talk show interpretation style), so as to create a personality with a specific persona for the virtual anchor. While stereotyping the character, it can lay a good foundation for deeper penetration of the character's personality in the future and enhance the anchor's affinity. The triggering method of the live broadcast text relies on the communication and interaction process between the user and the virtual anchor. The talk show segments are triggered by clicking, entering keywords in the comment box, etc., thereby enhancing the user experience and giving the user a better sense of participation. The interaction with the virtual anchor also greatly increases the intimacy between the user and the virtual anchor.
[0127] An embodiment of the present specification sets a virtual character to broadcast information on a live broadcast theme in a live broadcast room, obtains key information related to the live broadcast theme, and extracts keywords contained in the key information based on a pre-constructed keyword map, extracts associated words related to the keywords in a pre-constructed associated word map, obtains hot words related to the associated words, wherein a first degree of association between the hot words and the associated words is greater than a second degree of association between the hot words and the keywords, generates a live broadcast theme text related to the live broadcast theme based on the keywords, the associated words and the hot words, wherein the live broadcast theme text includes at least two paragraphs, and the first paragraph consists of the keywords and the associated words, and the second paragraph consists of the hot words and the associated words, generates a rendering model of the virtual character and drives the virtual character to broadcast the live broadcast theme text.
[0128] In the embodiment of this specification, when setting up a virtual character to broadcast information on the live broadcast theme in the live broadcast room, key information related to the live broadcast theme can be obtained, and then the live broadcast theme text can be jointly constructed based on the keywords in the key information, the associated words related to the keywords, and the hot words related to the associated words, and the virtual character can be driven to broadcast the live broadcast theme text. The method of generating and broadcasting the live broadcast theme text in real time is conducive to enhancing the fun and interactivity of the live broadcast, and at the same time, it is conducive to realizing the diversity of live broadcast methods, so as to better meet the needs of users for different types of information.
[0129] The following combination Figure 3 Taking the application of the virtual live broadcast control method provided in this specification in the skiing competition scene as an example, the virtual live broadcast control method is further explained. Figure 3 A processing flow chart of a virtual live broadcast control method provided by an embodiment of the present specification is shown, which specifically includes the following steps.
[0130] Step 302, setting a virtual character to broadcast information about the skiing competition activity in the live broadcast room.
[0131] Step 304, obtaining key information related to the skiing competition activity, and extracting keywords contained in the key information based on a pre-constructed keyword map.
[0132] Step 306, determining the word unit corresponding to each node in the pre-constructed associated word graph.
[0133] Step 308, semantically match the keyword with the word unit, and determine the associated words related to the keyword in the associated word map according to the matching results.
[0134] Step 310, obtaining network hot information according to a preset time period.
[0135] Step 312: input the network hot information into the sequence labeling model for processing to generate a sequence labeling result corresponding to the network hot information.
[0136] Step 314: determine the hot words related to the associated words in the network hot information according to the sequence labeling result.
[0137] The first correlation between the hot word and the associated word is greater than the second correlation between the hot word and the keyword.
[0138] Step 316, taking keywords, associated words and hot words as nodes, and taking the associations among keywords, associated words and hot words as edges, to construct a knowledge graph corresponding to the keywords.
[0139] Step 318, based on the word units contained in the path formed by adjacent nodes in the knowledge graph, generate a live broadcast theme text related to the skiing competition activity and having a second narrative style.
[0140] The live broadcast theme text includes at least two paragraphs, and the first paragraph consists of keywords and associated words, and the second paragraph consists of popular words and associated words.
[0141] Step 320, generating a rendering model of the virtual character and driving the virtual character to broadcast the live broadcast theme text.
[0142] In the embodiment of this specification, when setting up a virtual character to broadcast information about skiing competitive activities in a live broadcast room, key information related to skiing competitive activities can be obtained, and then a live broadcast theme text is constructed based on the keywords in the key information, the associated words related to the keywords, and the hot words related to the associated words, and the virtual character is driven to broadcast the live broadcast theme text. By generating and broadcasting the live broadcast theme text in real time, it is conducive to enhancing the fun and interactivity of the live broadcast, and at the same time, it is conducive to achieving the diversity of live broadcast methods, so as to better meet the user's needs for different types of information.
[0143] Corresponding to the above method embodiment, this specification also provides a virtual live broadcast control system embodiment. Figure 4 FIG. 1 shows a schematic diagram of a virtual live broadcast control system provided by an embodiment of the present specification. Figure 4 As shown, the system includes:
[0144] The setting module 402 is configured to set the virtual character to broadcast information on the live broadcast theme in the live broadcast room;
[0145] The first acquisition module 404 is configured to acquire key information related to the live broadcast theme and extract keywords contained in the key information based on a pre-built keyword graph;
[0146] An extraction module 406 is configured to extract associated words related to the keyword from a pre-built associated word graph;
[0147] A second acquisition module 408 is configured to acquire a hot word related to the associated word, wherein a first degree of association between the hot word and the associated word is greater than a second degree of association between the hot word and the keyword;
[0148] The generating module 410 is configured to generate a live broadcast theme text related to the live broadcast theme based on the keywords, the associated words and the hot words, wherein the live broadcast theme text comprises at least two paragraphs, and the first paragraph consists of the keywords and the associated words, and the second paragraph consists of the hot words and the associated words;
[0149] The driving module 412 is configured to generate a rendering model of the virtual character and drive the virtual character to broadcast the live broadcast theme text.
[0150] Optionally, the key information includes key news information having a first narrative style, and the associated words include interesting associated words having a second narrative style.
[0151] Optionally, the generation module 410 is further configured to: construct a knowledge graph corresponding to the keyword based on the association relationship between the keyword, the associated words and the popular words; and generate a live broadcast theme text related to the live broadcast theme and having a second narrative style based on the word units contained in the path formed by adjacent nodes in the knowledge graph.
[0152] Optionally, the generation module 410 is further configured to: use the keywords, the associated words and the hot words as nodes, and use the associations between the keywords, the associated words and the hot words as edges to construct a knowledge graph corresponding to the keywords.
[0153] Optionally, the second acquisition module 408 is further configured to: acquire network hot information according to a preset time period;
[0154] The network hot information is input into a sequence labeling model for processing to generate a sequence labeling result corresponding to the network hot information; and hot words related to the associated words in the network hot information are determined according to the sequence labeling result.
[0155] Optionally, the extraction module 406 is further configured to: determine the word unit corresponding to each node in the pre-constructed associated word graph; semantically match the keyword with the word unit, and determine the associated words related to the keyword in the associated word graph based on the matching results.
[0156] Optionally, the generation module 410 is further configured to: obtain a reference text having a second narrative style, and construct a text generation template related to the live broadcast theme based on the reference text, wherein the text generation template contains word slots to be filled; based on the keywords, the associated words and the popular words, fill the word slots to be filled to generate a live broadcast theme text related to the live broadcast theme.
[0157] Optionally, the generation module 410 is further configured to: obtain a template type of a preset text generation template; input the reference text into a semantic analysis model for processing to obtain a corresponding semantic analysis result; determine a target reference text that matches the template type based on the semantic analysis result; determine the word slots to be filled in the target reference text based on the target reference text to generate a text generation template related to the live broadcast theme.
[0158] Optionally, the generation module 410 is further configured to: input the keywords, the associated words and the hot words into a text generation model for processing, and generate a live broadcast theme text related to the live broadcast theme.
[0159] Optionally, the text generation model is trained in the following manner: obtaining a reference text having a second narrative style; using the reference text as sample data and inputting it into a pre-trained text generation model for training to obtain the text generation model, wherein the pre-trained text generation model masks keywords contained in the reference text and predicts the masked keywords, so as to adjust model parameters of the pre-trained text generation model based on the prediction results.
[0160] Optionally, the virtual live broadcast control system further includes an adjustment module configured to adjust the text content of the live broadcast theme text according to a preset character style of the virtual character, and generate a target live broadcast theme text matching the preset character style.
[0161] The above is a schematic scheme of a virtual live broadcast control system of this embodiment. It should be noted that the technical scheme of the virtual live broadcast control system and the technical scheme of the virtual live broadcast control method described above are of the same concept, and the details not described in detail in the technical scheme of the virtual live broadcast control system can be found in the description of the technical scheme of the virtual live broadcast control method described above.
[0162] Figure 5 The structure block diagram of a computing device 500 provided according to an embodiment of the present specification is shown. The components of the computing device 500 include but are not limited to a memory 510 and a processor 520. The processor 520 is connected to the memory 510 via a bus 530, and the database 550 is used to store data.
[0163] The computing device 500 also includes an access device 540 that enables the computing device 500 to communicate via one or more networks 560. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 540 may include one or more of any type of network interface (e.g., a network interface card (NIC)) whether wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a World Wide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, a near field communication (NFC) interface, and the like.
[0164] In one embodiment of the present specification, the above components of the computing device 500 and Figure 5 Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Figure 5 The computing device structure block diagram shown is only for the purpose of illustration, and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.
[0165] The computing device 500 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or PC. The computing device 500 may also be a mobile or stationary server.
[0166] Among them, the processor 520 is used to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the above-mentioned virtual live broadcast control method.
[0167] The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the above virtual live broadcast control method belong to the same concept, and the details of the technical scheme of the computing device that are not described in detail can be found in the description of the technical scheme of the above virtual live broadcast control method.
[0168] An embodiment of the present specification also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the above-mentioned virtual live broadcast control method.
[0169] The above is a schematic scheme of a computer-readable storage medium of this embodiment. It should be noted that the technical scheme of the storage medium and the technical scheme of the virtual live broadcast control method described above are of the same concept, and the details not described in detail in the technical scheme of the storage medium can be found in the description of the technical scheme of the virtual live broadcast control method described above.
[0170] An embodiment of the present specification also provides a computer program, wherein when the computer program is executed in a computer, the computer is caused to execute the steps of the above-mentioned virtual live broadcast control method.
[0171] The above is a schematic scheme of a computer program of this embodiment. It should be noted that the technical scheme of the computer program and the technical scheme of the virtual live broadcast control method described above are of the same concept, and the details not described in detail in the technical scheme of the computer program can be found in the description of the technical scheme of the virtual live broadcast control method described above.
[0172] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0173] The computer instructions include computer program codes, which may be in source code form, object code form, executable files or some intermediate forms, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.
[0174] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of this specification are not limited by the order of the actions described, because according to the embodiments of this specification, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the embodiments of this specification.
[0175] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0176] The preferred embodiments of this specification disclosed above are only used to help explain this specification. The optional embodiments do not describe all the details in detail, nor do they limit the invention to only the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of the embodiments of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the embodiments of this specification, so that technicians in the relevant technical field can well understand and use this specification. This specification is only limited by the claims and their full scope and equivalents.
Claims
1. A virtual live broadcast control method, comprising: Set up virtual characters to broadcast information on live broadcast topics in the virtual live broadcast room; Acquire key information related to the live broadcast theme, and extract keywords contained in the key information based on a pre-constructed keyword map, wherein the key information includes key news information with a first narrative style; Extracting associated words that semantically match the keyword from a pre-constructed associated word graph, wherein the associated words include interesting associated words with a second narrative style; Acquire a hot word related to the associated word, wherein a first degree of association between the hot word and the associated word is greater than a second degree of association between the hot word and the keyword, wherein the hot word is a word with high popularity or frequency of use in the network; Based on the keywords, the associated words and the hot words, generating a live broadcast theme text related to the live broadcast theme, wherein the live broadcast theme text includes at least two paragraphs, and the first paragraph consists of the keywords and the associated words, and the second paragraph consists of the hot words and the associated words; A rendering model of the virtual character is obtained and the virtual character is driven to broadcast according to the live broadcast theme text.
2. The virtual live broadcast control method according to claim 1, wherein generating a live broadcast theme text related to the live broadcast theme based on the keywords, the associated words and the hot words comprises: Based on the association relationship between the keyword, the associated word and the hot word, construct a knowledge graph corresponding to the keyword; Based on the word units contained in the path formed by adjacent nodes in the knowledge graph, a live broadcast theme text related to the live broadcast theme and having a second narrative style is generated.
3. According to the virtual live broadcast control method of claim 2, the step of constructing a knowledge graph corresponding to the keyword based on the association relationship between the keyword, the associated word, and the hot word comprises: The keywords, the associated words and the hot words are taken as nodes, and the associations among the keywords, the associated words and the hot words are taken as edges to construct a knowledge graph corresponding to the keywords.
4. According to the virtual live broadcast control method of claim 1, the step of obtaining hot words related to the associated words comprises: Obtain popular network information according to a preset time period; Inputting the network hot information into a sequence labeling model for processing to generate a sequence labeling result corresponding to the network hot information; Determine a hot word related to the associated word in the network hot information according to the sequence labeling result.
5. According to the virtual live broadcast control method of claim 1, extracting the associated words that semantically match the keyword from the pre-constructed associated word graph comprises: Determine the word unit corresponding to each node in the pre-constructed associated word graph; The keyword is semantically matched with the word unit, and the associated words related to the keyword in the associated word map are determined according to the matching result.
6. The virtual live broadcast control method according to claim 1, wherein generating a live broadcast theme text related to the live broadcast theme based on the keywords, the associated words and the hot words comprises: Acquire a reference text with a second narrative style, and construct a text generation template related to the live broadcast theme based on the reference text, wherein the text generation template includes a word slot to be filled; Based on the keywords, the associated words and the hot words, the word slots to be filled are filled to generate a live broadcast theme text related to the live broadcast theme.
7. The virtual live broadcast control method according to claim 6, wherein the step of constructing a text generation template related to the live broadcast theme based on the reference text comprises: Get the template type of the preset text generation template; Inputting the reference text into a semantic analysis model for processing to obtain a corresponding semantic analysis result; Determine a target reference text matching the template type according to the semantic analysis result; The word slots to be filled in the target reference text are determined based on the target reference text to generate a text generation template related to the live broadcast theme.
8. The virtual live broadcast control method according to claim 1, wherein generating a live broadcast theme text related to the live broadcast theme based on the keywords, the associated words and the hot words comprises: The keywords, the associated words and the hot words are input into a text generation model for processing to generate a live broadcast theme text related to the live broadcast theme.
9. According to the virtual live broadcast control method of claim 8, the text generation model is trained in the following manner: Obtain reference texts with a secondary narrative style; The reference text is used as sample data and input into a pre-trained text generation model for training to obtain the text generation model, wherein: The pre-trained text generation model masks the keywords included in the reference text and predicts the masked keywords, so as to adjust the model parameters of the pre-trained text generation model based on the prediction results.
10. The virtual live broadcast control method according to claim 1, after generating the live broadcast theme text related to the live broadcast theme, further comprising: According to the preset character style of the virtual character, the text content of the live broadcast theme text is adjusted to generate a target live broadcast theme text that matches the preset character style.
11. A virtual live broadcast control system, comprising: A setting module is configured to set a virtual character to broadcast information on a live broadcast theme in a virtual live broadcast room; A first acquisition module is configured to acquire key information related to the live broadcast theme and extract keywords contained in the key information based on a pre-built keyword map, wherein the key information includes key news information with a first narrative style; An extraction module is configured to extract associated words that semantically match the keyword from a pre-constructed associated word graph, wherein the associated words include interesting associated words with a second narrative style; A second acquisition module is configured to acquire a hot word related to the associated word, wherein a first degree of association between the hot word and the associated word is greater than a second degree of association between the hot word and the keyword, wherein the hot word is a word with high popularity or frequency of use in the network; A generating module is configured to generate a live broadcast theme text related to the live broadcast theme based on the keywords, the associated words and the hot words, wherein the live broadcast theme text comprises at least two paragraphs, and the first paragraph consists of the keywords and the associated words, and the second paragraph consists of the hot words and the associated words; The driving module is configured to obtain the rendering model of the virtual character and drive the virtual character to broadcast according to the live broadcast theme text.
12. A computing device comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the virtual live broadcast control method described in any one of claims 1 to 10 are implemented.
13. A computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the steps of the virtual live broadcast control method described in any one of claims 1 to 10.
