Video title generation method and apparatus, computer device, and storage medium

Abstract

The application provides a video title generation method and device, computer equipment and a storage medium, and belongs to the technical field of computers. The method comprises the following steps: obtaining a content label of a video; in response to the content label indicating that the video belongs to a tutorial video, identifying an original title of the video to obtain a style label of the original title; and in response to the style label indicating that the original title does not conform to a presentation style, generating a target title conforming to the presentation style based on the original title and the video. The application is based on natural language processing technology of artificial intelligence, fully utilizes the original title and the video content of the tutorial video, automatically and efficiently generates a tutorial title for the tutorial video, can more accurately reflect the core content taught by the tutorial video, facilitates users to intuitively find tutorial interest points from the title, enhances the attraction of the tutorial video to users, improves the play rate of the tutorial video, and thus improves the propagation efficiency of the tutorial video.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a video title generation method, apparatus, computer device, and storage medium. Background Technology

[0002] Video is an important form of information sharing and acquisition for users. Users can record tutorial videos to share skills and upload them to video sharing platforms so that other users can learn and master these skills. For example, tutorial videos include food preparation videos, mobile phone maintenance videos, and so on.

[0003] Video titles play a crucial role in attracting users to click and watch videos. Currently, video titles are generally user-defined; users who record videos can edit a title before uploading the video and upload both the video and title to the video sharing platform.

[0004] For tutorial videos, user-defined video titles often fail to accurately reflect the key information of the tutorial, making the videos difficult to find, resulting in low play rates and inefficient dissemination. Therefore, there is an urgent need for a method to automatically generate video titles to improve the play rate and dissemination efficiency of tutorial videos. Summary of the Invention

[0005] This application provides a video title generation method, apparatus, computer device, and storage medium, which can improve the playback rate of tutorial videos and increase dissemination efficiency. The technical solution is as follows.

[0006] On the one hand, a method for generating video titles is provided, the method comprising:

[0007] Obtain the content tags of the video, which are used to indicate whether the video belongs to a tutorial video;

[0008] In response to the content tag indicating that the video belongs to a tutorial video, the original title of the video is identified to obtain the style tag of the original title. The style tag is used to indicate whether the original title conforms to the expression style of a tutorial title.

[0009] In response to the style tag indicating that the original title does not conform to the expression style, a target title that conforms to the expression style is generated based on the original title and the video.

[0010] On one hand, a video title generation device is provided, the device comprising:

[0011] The tutorial video recognition module is used to obtain the content tags of the video, and the content tags are used to indicate whether the video belongs to the tutorial video.

[0012] The title style recognition module is used to identify the original title of the video in response to the content tag indicating that the video belongs to the tutorial video, and obtain the style tag of the original title. The style tag is used to indicate whether the original title conforms to the expression style of the tutorial title.

[0013] The title generation module is used to generate a target title that conforms to the expression style based on the original title and the video, in response to the style tag indicating that the original title does not conform to the expression style.

[0014] In one optional implementation, the tutorial video recognition module includes:

[0015] A video feature extraction unit is used to extract features from the video to obtain feature vectors of the video in multiple dimensions.

[0016] The tutorial video recognition unit is used to determine the content tags of the video based on the feature vectors of the video in multiple dimensions.

[0017] In another alternative implementation, the video feature extraction unit includes:

[0018] The content acquisition subunit is used to acquire multiple image frames and multiple audio frames of the video;

[0019] An image feature extraction subunit is used to determine the feature vector of the image dimension of the video based on the multiple image frames;

[0020] An audio feature extraction subunit is used to determine the feature vector of the audio dimension of the video based on the multiple audio frames;

[0021] The text feature extraction subunit is used to determine the text dimension feature vector of the video based on the text information corresponding to the video.

[0022] In another alternative implementation, the image feature extraction subunit is used for:

[0023] Feature extraction is performed on the multiple image frames respectively to obtain multiple feature vectors of the multiple image frames;

[0024] The feature vectors of the multiple image frames are aggregated to obtain the feature vector of the image dimension of the video.

[0025] In another alternative implementation, the apparatus further includes:

[0026] The text recognition module is used to perform text recognition on the multiple image frames to obtain subtitle text;

[0027] The speech recognition module is used to perform speech recognition on the multiple audio frames to obtain the dialogue text;

[0028] The text information determination module is used to determine the original title, the subtitle text, and the dialogue text as the text information.

[0029] In another alternative implementation, the tutorial video recognition unit is used for:

[0030] The feature vectors of the video in multiple dimensions are fused to obtain the multimodal vector of the video;

[0031] Based on the multimodal vectors, the content tags of the video are determined.

[0032] In another alternative implementation, the title generation module includes:

[0033] The tutorial main keyword determination unit is used to determine the tutorial main keywords corresponding to the video in response to the style tag indicating that the original title does not conform to the expression style.

[0034] The title generation unit is used to generate the target title based on at least one of the tutorial main words, the video type of the video, and the multimodal vector of the video, as well as the original title;

[0035] The tutorial subject terms are used to represent the tutorial theme of the video, and the multimodal vector is obtained by fusing the feature vectors of the video in multiple dimensions.

[0036] In another alternative implementation, the tutorial subject word determination unit is used for:

[0037] In response to the style tag indicating that the original title does not conform to the expression style, the main keywords of the tutorial are determined from the original title;

[0038] In response to the absence of the tutorial subject word in the original title, the tutorial subject word is determined from the subtitle text and the dialogue text of the video.

[0039] In another alternative implementation, the title generation unit includes:

[0040] The word segmentation subunit is used to divide the main words of the tutorial and the original title into multiple words;

[0041] The title word determination subunit is used to input the video type, the multimodal vector, and the multiple word segments into the title generation model, and determine multiple title words through the title generation model. The title generation model is used to generate a title that conforms to the expression style.

[0042] The title generation subunit is used to compose the target title from the plurality of title words.

[0043] In another alternative implementation, the title word determines the sub-unit for:

[0044] The video type, the multimodal vector, and the multiple word segments are input into the title generation model to obtain the attention weights of the multiple word segments;

[0045] In response to the first segmentation among the plurality of segmented words having an attention weight greater than an attention threshold, the first segmentation is determined as the title word;

[0046] In response to the fact that the attention weights of the multiple word segments are all less than or equal to the attention threshold, the word output by the title generation model is determined as the title word.

[0047] In another alternative implementation, the apparatus further includes:

[0048] The video type determination module is used to determine the video type of the video based on the feature vectors of the video in multiple dimensions.

[0049] In another alternative implementation, the title style recognition module is used for:

[0050] In response to the content tag indicating that the video belongs to a tutorial video, the original title is input into the style recognition model to obtain the style tag output by the style recognition model;

[0051] The style recognition model is trained using a first type of title as positive samples and at least one of a second or third type of title as negative samples. The first type of title conforms to the expression style, and the corresponding video belongs to the tutorial video category. The second type of title is the original title of a video that does not belong to the tutorial video category. The third type of title does not conform to the expression style, and the corresponding video belongs to the tutorial video category.

[0052] On one hand, a computer device is provided, the computer device including a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to implement the video title generation method described in any of the above optional implementations.

[0053] On one hand, a computer-readable storage medium is provided, wherein a computer program is stored in the computer-readable storage medium, the computer program being loaded and executed by a processor to implement the video title generation method described in any of the above optional implementations.

[0054] On one hand, a computer program product or computer program is provided, which includes computer program code stored in a computer-readable storage medium. A processor of a computer device reads the computer program code from the computer-readable storage medium and executes the computer program code, causing the computer device to perform the video title generation method described in any of the above optional implementations.

[0055] This application provides a scheme for generating titles for tutorial videos. The scheme identifies the original title of the tutorial video and, if the original title does not conform to the style of a tutorial title, automatically and efficiently generates a title that conforms to that style based on the original title and the video content. Because the generated title fully utilizes both the original title and the video content, it more accurately reflects the core content taught in the tutorial video, allowing users to intuitively identify the tutorial's points of interest from the title. This enhances the tutorial video's appeal to users, increases its play rate, and ultimately improves its dissemination efficiency. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a schematic diagram of an implementation environment provided in an embodiment of this application;

[0058] Figure 2 This is a flowchart of a video title generation method provided in an embodiment of this application;

[0059] Figure 3 This is a flowchart of a video title generation method provided in an embodiment of this application;

[0060] Figure 4 This is a schematic diagram of a video recognition tutorial provided in an embodiment of this application;

[0061] Figure 5 This is a schematic diagram illustrating a method for identifying title styles, as provided in an embodiment of this application.

[0062] Figure 6 This is a schematic diagram illustrating how to generate a title when the original title does not contain the main subject of the tutorial;

[0063] Figure 7This is a schematic diagram illustrating how a title is generated when the original title contains the main subject of the tutorial;

[0064] Figure 8 This is a schematic diagram illustrating a model training and application provided in an embodiment of this application;

[0065] Figure 9 This is a block diagram of a video title generation device provided in an embodiment of this application;

[0066] Figure 10 This is a block diagram of a terminal provided in an embodiment of this application;

[0067] Figure 11 This is a block diagram of a server provided in an embodiment of this application. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0069] The terms "first," "second," "third," and "fourth," etc., used in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0070] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0071] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0072] Computer vision (CV) is the science of enabling machines to "see." More specifically, it refers to machine vision, which uses cameras and computers to replace human eyes in tasks such as target recognition, tracking, and measurement, and further performs image processing to create images more suitable for human observation or transmission to instruments. As a scientific discipline, computer vision studies related theories and technologies, attempting to build artificial intelligence systems capable of extracting information from images or multidimensional data. Computer vision technologies typically include image processing, image recognition, image semantic understanding, image retrieval, optical character recognition (OCR), video processing, video semantic understanding, and video content / behavior recognition.

[0073] Key technologies in speech technology include Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and voiceprint recognition. Enabling computers to hear, see, speak, and feel is the future direction of human-computer interaction, with speech emerging as one of the most promising methods.

[0074] Natural Language Processing (NLP) is an important field within computer science and artificial intelligence. It studies the theories and methods for enabling effective communication between humans and computers using natural language. NLP is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language people use in daily life—and thus it has a close relationship with linguistic research. NLP techniques typically include text processing, semantic understanding, machine translation, question answering, and knowledge graphs.

[0075] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instruction-based learning.

[0076] The solutions provided in this application involve technologies such as computer vision, speech technology, natural language processing, machine learning, and deep learning in artificial intelligence, which are specifically illustrated in the following embodiments.

[0077] Figure 1 This is a schematic diagram of an implementation environment provided in an embodiment of this application. See also... Figure 1 The implementation environment includes terminal 110 and server 120.

[0078] Terminal 110 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, smart TV, smart car, etc., but is not limited to these. Optionally, terminal 110 has a client application that displays video titles and plays videos. For example, this client can be a video client, a browser client, or a social media client.

[0079] Server 120 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0080] Terminal 110 obtains video titles from server 120 through a connection with server 120 and displays the video titles to attract users to watch the videos corresponding to the video titles. Terminal 110 and server 120 can be connected directly or indirectly via wired or wireless communication, and this application does not impose any restrictions on this connection.

[0081] Figure 2 This is a flowchart illustrating a video title generation method provided in an embodiment of this application. This video title generation method can be executed by a terminal; alternatively, it can be executed by a server; or it can be executed through interaction between the terminal and the server. In this embodiment, a brief description is given using a server as the execution subject. See [link to documentation]. Figure 2 This embodiment includes the following steps.

[0082] 201. The server retrieves the video's content tags, which are used to indicate whether the video belongs to the tutorial video category.

[0083] The video includes both images and audio, providing users with an intuitive viewing experience from both visual and auditory perspectives. It should be noted that this application does not impose a limit on the video's duration; the video can be a short video with a duration less than or equal to a duration threshold, or a long video with a duration greater than the duration threshold. The duration threshold is a duration value greater than 0, for example, a duration threshold of 11 seconds or 15 seconds.

[0084] Tutorial videos are videos with instructional and guiding functions, used to guide users in understanding knowledge or instructing them to master a certain skill. For example, the process of analyzing the network structure of a neural network, the solution method for a certain math problem, the writing skills of argumentative essays, popular science knowledge about history, how to peel a kiwi, how to cook a certain dish, how to use tea sets, or how to assemble furniture can all be recorded as tutorial videos.

[0085] Optionally, the content tag can be 0 or 1. If the video's content tag is 0, then the video is not a tutorial video; if the content tag is 1, then the video is a tutorial video.

[0086] 202. In response to the content tag indicating that the video belongs to a tutorial video, the server identifies the original title of the video and obtains the style tag of the original title. The style tag is used to indicate whether the original title conforms to the expression style of a tutorial title.

[0087] If the video is a tutorial video, the server continues to identify whether the original title of the video follows the style of a tutorial title. If the video is not a tutorial video, the server ends the processing of that video and does not perform subsequent steps, in order to reduce meaningless processing, improve processing efficiency, reduce the consumption of computing resources, and improve resource utilization.

[0088] Optionally, the original title is a user-defined title for the uploaded video. The tutorial title should clearly indicate that the video has an instructional function and what its core content is. For example, a tutorial titled "Master a Skill for Cleaning Up Computer Junk Files" clearly indicates that the video teaches how to clean up computer junk files. Similarly, a tutorial titled "Delicious and Visually Appealing Fruit Salad, Learn in 5 Minutes!" clearly indicates that the video teaches how to make a fruit salad. The commonality among multiple tutorial titles lies in their stylistic expression.

[0089] Optionally, the style tag can be 0 or 1. If the style tag of the title is 0, the title does not conform to the style of a tutorial title; if the style tag is 1, the title conforms to the style of a tutorial title.

[0090] 203. In response to the style tag indicating that the original title of the video does not conform to the above expression style, the server generates a target title that conforms to the expression style based on the original title and the video.

[0091] If the original title of the video does not conform to the above style of expression, the server will generate a target title that conforms to the style of expression based on the original title and the video content.

[0092] This application provides a scheme for generating titles for tutorial videos. The scheme identifies the original title of the tutorial video and, if the original title does not conform to the style of a tutorial title, automatically and efficiently generates a title that conforms to that style based on the original title and the video content. Because the generated title fully utilizes both the original title and the video content, it more accurately reflects the core content taught in the tutorial video, allowing users to intuitively identify the tutorial's points of interest from the title. This enhances the tutorial video's appeal to users, increases its play rate, and ultimately improves its dissemination efficiency.

[0093] Figure 3 This is a flowchart illustrating a video title generation method provided in an embodiment of this application. This video title generation method can be executed by a terminal; alternatively, it can be executed by a server; or it can be executed through interaction between the terminal and the server. In this embodiment, the server is used as the execution subject for illustration; see [link to documentation]. Figure 3 This embodiment includes the following steps.

[0094] 301. The server extracts features from the video to obtain feature vectors of the video in multiple dimensions.

[0095] The video includes images and audio. Optionally, the video also includes subtitles to help viewers understand the video content. In this embodiment, an example of a video including images, audio, and subtitles is used for illustration.

[0096] Because videos have features in multiple dimensions such as images, audio, and text, the server simulates human intelligence to analyze the video from these multiple dimensions, obtaining feature vectors in multiple dimensions. This results in a richer and more accurate representation of the video. Based on this richer and more accurate video representation, video recognition and title generation can be performed, further improving the quality of title generation.

[0097] Accordingly, the steps of the server extracting features from the video to obtain the feature vector of the video in multiple dimensions include steps 3011 to 3015.

[0098] 3011. The server retrieves multiple image frames and multiple audio frames from the video.

[0099] Optionally, the server may acquire a subset of image frames from the video as the aforementioned plurality of image frames. For example, the server may acquire one image frame at intervals of a first duration, based on the video's timeline. This first duration can be flexibly configured, such as 1 second, 3 seconds, or 4 seconds. Alternatively, the server may divide all image frames from the video into multiple acquisition units, each unit comprising a subset of image frames, and the server acquires one image frame from each acquisition unit. This application does not limit the method by which the server acquires a subset of image frames from the video. Alternatively, the server may extract all image frames from the video as the aforementioned plurality of image frames. This application does not limit this method.

[0100] Optionally, the server may acquire some or all of the audio frames of the video by acquiring multiple image frames in a similar manner; this application does not limit this.

[0101] The aforementioned server reduces the number of image or audio frames that need to be processed subsequently by acquiring only a portion of the image or audio frames, thereby improving processing efficiency, reducing the consumption of computing resources, and increasing resource utilization.

[0102] 3012. The server determines the feature vector of the image dimension of the video based on multiple image frames.

[0103] In one optional implementation, the server determines the feature vector of each image frame separately, and then aggregates the feature vectors of multiple image frames into a single feature vector, achieving dimensionality reduction from frame-level features to video-level features. Accordingly, step 3012 includes: the server extracting features from multiple image frames to obtain multiple feature vectors for each image frame; and aggregating the multiple feature vectors from the multiple image frames to obtain the image-dimensional feature vector of the video.

[0104] Optionally, see Figure 4 The server uses EfficientNet 401 to determine the feature vector of each image frame. Then, it inputs the feature vectors of multiple image frames into NeXtVLAD 402 (a neural network for aggregating frame-level features). Based on NeXtVLAD, it performs temporal fusion of multiple frames to obtain a video-level image representation vector, which is the image-dimensional feature vector of the video. It should be noted that the server can also use other neural networks for feature extraction to obtain the video's image-dimensional feature vector; this application does not impose any restrictions on this.

[0105] The above technical solution, by aggregating feature vectors from multiple image frames, obtains a lower-dimensional vector representation in the image dimension while ensuring the accuracy of feature representation. Subsequent processing based on this lower-dimensional vector representation can improve processing efficiency, reduce computational resource consumption, and improve resource utilization.

[0106] In another alternative implementation, the server can extract features from multiple image frames to obtain multiple feature vectors for each image frame. These feature vectors can then be concatenated to obtain the feature vectors for the image dimensions of the video.

[0107] 3013. The server determines the feature vector of the audio dimension of the video based on multiple audio frames.

[0108] Optionally, the server determines the feature vector of the audio dimension through steps similar to those used to determine the feature vector of the image dimension. That is, the server extracts features from multiple audio frames to obtain multiple feature vectors of multiple audio frames; and aggregates the multiple feature vectors of multiple audio frames to obtain the feature vector of the audio dimension of the video.

[0109] Optionally, see also Figure 4 The server uses VGGish (a network for extracting semantic feature vectors from audio) 403 to determine the feature vector of each audio frame; then it inputs the feature vectors of multiple audio frames into NeXtVLAD (a neural network for aggregating frame-level features) 404. Based on NeXtVLAD, the multiple frames are temporally fused to obtain a video-level audio representation vector, which is the feature vector of the audio dimension of the video.

[0110] It should be noted that, optionally, after the server extracts features from multiple audio frames and obtains multiple feature vectors for the multiple audio frames, it can also concatenate the above multiple feature vectors to obtain the feature vector of the audio dimension of the video.

[0111] 3014. The server retrieves the text information corresponding to the video.

[0112] Optionally, the text information includes at least one of subtitle text, dialogue text, and the original title of the video. In this application, the example of text information including subtitle text, dialogue text, and the original title is used for illustration. In embodiments where the text information includes one or two of the above three, the process of the server processing the text information is the same as the process of text information processing in this application, and will not be described in detail here.

[0113] The dialogue text is obtained through speech recognition. The subtitle text is displayed in text form during video playback. The subtitle text includes at least one of a first subtitle text and a second subtitle text. The first subtitle text is added after the video recording is completed to facilitate understanding, while the second subtitle text is the text captured during the video recording process.

[0114] Accordingly, the steps for the server to obtain the text information corresponding to the video include: the server performing text recognition on multiple image frames to obtain subtitle text; the server performing speech recognition on multiple audio frames to obtain dialogue text; and the server determining the original title, subtitle text, and dialogue text as text information. Optionally, the server performs text recognition on image frames based on OCR technology to obtain subtitle text; and performs speech recognition on audio frames based on ASR technology to obtain dialogue text.

[0115] It should be noted that the above embodiment uses the example of the server determining the original title, subtitle text, and dialogue text as text information. Optionally, to achieve a more accurate and concise text expression, the server extracts keywords to retain key information in the original title, subtitle text, and dialogue text. Accordingly, the step of the server determining the original title, subtitle text, and dialogue text as text information can be replaced with the following steps: the server extracts keywords from the original title, subtitle text, and dialogue text; and the extracted keywords are determined as the text information corresponding to the video.

[0116] The above technical solution extracts text information through keyword extraction, reducing meaningless text and resulting in more accurate and concise text expression. Based on the text information composed of keywords, video recognition and title generation are performed, which can reduce the interference of meaningless text and further improve the accuracy of video recognition and title generation.

[0117] Another point to note is that the above embodiments use the example of a server performing text recognition on an image frame to obtain subtitle text. In other embodiments, the video also includes a subtitle file for storing the first subtitle text. Optionally, the server obtains the first subtitle text from this subtitle file; by performing text recognition on the image frame, the second subtitle text is obtained, thereby reducing the computational resources consumed by text recognition, improving resource utilization, and increasing the efficiency of subtitle text acquisition.

[0118] 3015. The server determines the text dimension feature vector of the video based on the text information corresponding to the video.

[0119] In one optional implementation, taking the text information corresponding to the video, including subtitle text, dialogue text, and the original title of the video, as an example, the server concatenates the subtitle text, dialogue text, and original title to obtain the concatenated text; and performs feature extraction on the concatenated text to obtain the feature vector of the text dimension of the video.

[0120] In another optional implementation, taking the text information corresponding to the video, including keywords extracted from subtitle text, dialogue text, and original title, as an example, the server concatenates the extracted keywords to obtain the concatenated text; and performs feature extraction on the concatenated text to obtain the feature vector of the text dimension of the video.

[0121] Optionally, see [link to relevant documentation] Figure 4 The server uses the Concat function 405 to concatenate the original title, ASR keywords, and OCR keywords to obtain the concatenated text. This concatenated text is then input into ALBERT (a lightweight semantic understanding model) 406, where ALBERT constructs a deep representation of the text dimension, resulting in a text-dimensional feature vector. The ASR keywords are those extracted by the server from the dialogue text, and the OCR keywords are those extracted from the subtitle text. It should be noted that the server can also use other semantic understanding models, such as BERT (a model for semantic understanding), to determine the text-dimensional feature vector; this application does not impose any restrictions on this.

[0122] 302. The server determines the content tags and video type of the video based on the feature vectors of the video in multiple dimensions. The content tags are used to indicate whether the video belongs to the tutorial video category.

[0123] Among them, video type refers to the type of video content, such as sports and fitness, TV series, lifestyle, food, archaeology, computer / mobile phone, cute pets, parenting, or automobiles.

[0124] In one alternative implementation, see [link to relevant documentation]. Figure 4 The server fuses the feature vectors of the video across multiple dimensions to obtain the video's multimodal vector; based on this multimodal vector, the content tags and video type of the video are determined.

[0125] The above technical solution fuses feature vectors from multiple dimensions into a multimodal vector, enabling the multimodal vector to comprehensively represent information from various dimensions of the video, ensuring the accuracy of vector representation, and thus improving the accuracy of video recognition based on this multimodal vector.

[0126] The server's steps for fusing feature vectors from multiple dimensions of a video to obtain its multimodal vector include: concatenating the feature vectors from multiple dimensions to obtain a concatenated feature vector; and passing the concatenated feature vector through at least one fully connected layer to obtain the multimodal vector. The fully connected layer is used for feature fusion to reduce the dimensionality of the feature vector, facilitating the output layer's final calculation and output based on the reduced-dimensionality feature vector, thus improving recognition efficiency. The server sequentially passes the concatenated feature vector through at least one fully connected layer, with each fully connected layer reducing the dimensionality of the feature vector by a certain amount. This multi-level dimensionality reduction minimizes information loss and ensures the accuracy of the multimodal vector's representation of the video.

[0127] Optionally, the server determines the content tag and video type based on multimodal vectors and a recognition network. In some embodiments, the output of the recognition network is a probability distribution of whether a video belongs to a tutorial video and a probability distribution of the video belonging to each video type. Based on the output of the recognition network, the server further determines the content tag and video type: If the probability that the video belongs to a tutorial video is greater than a first threshold, the server determines the content tag of the video to be 1, i.e., the video belongs to a tutorial video; if the probability that the video belongs to a tutorial video is less than or equal to the first threshold, the server determines the content tag of the video to be 0, i.e., the video does not belong to a tutorial video; the server determines the highest probability from the probability distribution of the video belonging to each video type; and determines the video type corresponding to the highest probability as the video type to which the video belongs. The first threshold can be flexibly configured, for example, the first threshold is 0.55 or 0.6, etc. Assuming the first threshold is 0.55, if the model outputs a probability of 0.875 that the video belongs to a tutorial video, then the video is determined to belong to a tutorial video. In other embodiments, the output of the recognition network is a content tag and a video type, and the server can directly obtain the content tag and video type output by the recognition network.

[0128] It should be noted that in steps 301 and 302 above, the server determines the content tag and video type through a series of video recognition processes. Optionally, in other embodiments, the server determines the video content tag and video type based on pre-stored correspondences between videos and content tags, and between videos and video types. Optionally, the aforementioned correspondences are pre-determined and stored by the server through steps 301 and 302; or, the aforementioned correspondences are stored based on manual annotation, and this application does not impose any restrictions on this.

[0129] 303. In response to a content tag indicating that a video belongs to a tutorial video, the server identifies the original title of the video and obtains the style tag of the original title. The style tag is used to indicate whether the original title conforms to the expression style of a tutorial title.

[0130] Optionally, the server uses a style recognition model to determine whether the original title conforms to the style of a tutorial title. In some embodiments, the input to the style recognition model is the title, and the output is the probability that the title conforms to the style of a tutorial title. Based on the output of the style recognition model, the server further determines the style tag of the title: if the probability output by the style recognition model is greater than a second threshold, the server determines the style tag of the title to be 1, that is, the title conforms to the style of a tutorial title; if the probability output by the style recognition model is less than or equal to the second threshold, the server determines the style tag of the title to be 0, that is, the title does not conform to the style of a tutorial title. The second threshold can be flexibly configured, for example, the second threshold is 0.55 or 0.6, etc. In other embodiments, the output of the style recognition model is a style tag; the server inputs the original title into the style recognition model to obtain the style tag output by the style recognition model.

[0131] Optionally, see Figure 5 The style recognition model, based on ALBERT, performs semantic understanding on the input title and then determines whether the title conforms to the style of a tutorial title. The process of the server recognizing the original title of the video and obtaining the style tag of the original title is also known as the process of the video title 501 passing through ALBERT 502 to obtain the result 503 of whether the title conforms to the style of a tutorial title.

[0132] If the server determines that the original title conforms to the style of a tutorial title, it will not execute the step of generating a title that conforms to the style described above, thereby reducing the computing resources consumed in generating the title and improving resource utilization. If the server determines that the original title does not conform to the style of a tutorial title, it will continue to execute the following steps to generate a title that conforms to the style described above.

[0133] 304. The server responds to the style tag indicating that the original title does not conform to the style of expression by identifying the main subject of the tutorial from the original title.

[0134] The subject of the tutorial is used to indicate the theme of the video tutorial. For example, in the title "Teach you a skill to clean up computer junk files," the subject of the tutorial is "computer junk files"; in the title "Delicious and visually appealing fruit salad, learn how to make it easily in 5 minutes!", the subject of the tutorial is "fruit salad."

[0135] Optionally, the server determines the main keywords of the tutorial from the original title using a sequence labeling model. This sequence labeling model has the function of labeling each word in the input text with an attribute label. Optionally, the attribute label is 0 or 1, where 0 indicates that the word is not a main keyword of the tutorial, and 1 indicates that the word is a main keyword of the tutorial. The server inputs the original title into the sequence labeling model to obtain the attribute label of each word in the original title; the word with an attribute label of 1 is identified as the main keyword of the tutorial.

[0136] It should be noted that the subject matter of the tutorial represents the theme of the video and is crucial for generating a title that conforms to the aforementioned style of expression. If the original title contains the subject matter, for example, the original title "Delicious and Visually Appealing Fruit Salad" contains the subject matter "fruit salad," the server can refer to this original title and generate a title related to "fruit salad," reflecting the tutorial theme of the video and achieving effective title generation. If the original title does not contain the subject matter, for example, the original title "Very Practical Tips" does not contain a meaningful subject matter, simply referring to this original title cannot achieve effective title generation. Therefore, it is necessary to further combine the video content to determine the subject matter, and then combine the subject matter to generate a title that conforms to the aforementioned style of expression. Based on the above, considering the accuracy, efficiency, and resource utilization of title generation, when the original title does not contain the subject matter, the title is generated through steps 305 to 306; when the original title contains the subject matter, the title is generated through step 307.

[0137] 305. The server responded that the tutorial subject words were not found in the original title. The tutorial subject words were determined from the video's subtitle text and dialogue text.

[0138] Optionally, the server determines the main keywords of the tutorial from the subtitle and dialogue texts using a sequence labeling model. The process by which the server determines the main keywords of the tutorial from the subtitle and dialogue texts is the same as the process by which the server determines the main keywords of the tutorial from the original title, and will not be repeated here. Optionally, the subtitle and dialogue texts are the results obtained by the server through text recognition of image frames and speech recognition of audio frames in step 3014 above.

[0139] 306. The server generates a target title that matches the style of the tutorial title based on the main keywords of the tutorial, the video type, the multimodal vector, and the original title.

[0140] Optionally, the server divides the tutorial's main keywords and the original title into multiple word segments; it inputs the video type, multimodal vector, and multiple word segments into a title generation model, which determines multiple title words; the target title is composed of these title words. The title generation model generates a title that matches the style of the tutorial title. This model is capable of generating a title that matches the video content and type, and intuitively represents the tutorial's theme, based on the video type, the multimodal vector comprehensively representing information from various dimensions of the video, the corresponding tutorial main keywords, and the original title. The video type is determined by the server in step 302 along with the video's content tags, and the multimodal vector is the result obtained by fusing the video's feature vectors across multiple dimensions in step 302.

[0141] The above technical solution, on the one hand, can generate a title that represents the tutorial theme even when the original title lacks the main tutorial keywords, by combining these keywords extracted from other text in the video, thus improving the accuracy of title generation. On the other hand, by combining the video's multimodal vectors to generate the target title, it strengthens the utilization of the video's multi-dimensional information, resulting in a higher degree of matching between the generated title and the video content, further improving the accuracy of title generation. Furthermore, it considers the differences between different video types, combining the video type with the title generation, making the generated title more consistent with the title expression style of that type of tutorial video, thereby further improving the accuracy of title generation. Moreover, the above title generation model is applicable even when the original title lacks the main tutorial keywords, improving the universality of the title generation model and reducing invalid title generation, thus improving the usability of the titles.

[0142] It's important to note that the title generation model determines a title word each time based on video type, multimodal vectors, and multiple word segments. This process is repeated to obtain multiple title words. During the determination of each title word, the server also considers previously determined title words or a combination of previously determined title words to ensure the semantic accuracy of multiple consecutive title words. Furthermore, the title generation model uses a target vocabulary, and its output represents the probability that each word in the target vocabulary is a title word in the current determination.

[0143] Furthermore, during the process of determining each title word, the server uses an attention mechanism to determine the attention weight of each of the multiple word segments. The attention weight of each word represents the degree of attention the server pays to that word during the title word determination process, and also indicates the degree of influence that word has on the determined title word. In the title word determination process, word segments with higher attention weights are better able to represent the key information of the tutorial video, such as the tutorial theme. Therefore, word segments with higher attention weights can be directly used as the determined title words, instead of determining title words from the target vocabulary based on the word segment vector representation. Since the above word segments are the main tutorial words determined from the video's subtitle and dialogue text, they have a higher degree of relevance to the video. Therefore, using these word segments as title words can make the generated title match the video content of the tutorial video more closely, thereby further improving the accuracy of title generation.

[0144] Accordingly, the steps by which the server determines each title word through the title generation model include: the server inputs the video type, multimodal vector, and multiple word segments into the title generation model to obtain the attention weights of the multiple word segments; in response to the first word segment having an attention weight greater than the attention threshold, the first word segment is determined as the title word; in response to the attention weights of multiple word segments being less than or equal to the attention threshold, the word output by the title generation model is determined as the title word. The first word segment is the one with the highest attention weight. The attention weights can be flexibly configured, for example, to 0.8, 0.85, or 0.9.

[0145] To make the title generation process clearer, the following will combine... Figure 6 Please provide an explanation. See also... Figure 6The aforementioned title generation model is based on the Transformer's Encoder-Decoder structure, with an added Attention mechanism. The encoder 601 converts text into vector representations, and the decoder 602 converts these vector representations into text output for a specific task. In this application, the specific task refers to a title conforming to the aforementioned style. The server identifies the main subject words of the video; it then segments the original title and the identified main subject words into multiple words; these words are converted into vector representations, resulting in words 1 to n; words 1 to n are encoded by the encoder 601 to obtain representations of words 1 to n; for each title word, the server, based on the Attention mechanism and the decoder 602, determines the latent vector and representations of words 1 to n of the previous title word, and determines the attention weight for each word segment; if the attention weight of the word with the largest attention weight is greater than the attention threshold, ... The segmented word is then identified as the title word for this determination, which is copied from the input text. If the attention weight of each segmented word is less than or equal to the attention threshold, the decoder 602 determines the hidden vector on which the title word for this determination is based, based on the input multimodal vector, the video type vector, the determined previous title word, the vector obtained by weighted summation of word 1 to word n ​​based on the attention weight, and the hidden vector on which the previous title word was determined. Based on the hidden vector, the vector obtained by weighted summation of word 1 to word n ​​based on the attention weight, and the determined previous title word, the title word for this determination is obtained.

[0146] It should be noted that words 1 to n above are individual vector representations of each word segment. Word 1 to n represent the semantically related vector representations of each word segment, and word n ​​represents the generation based on the first n-1 word vector representations of word n. Here, n and m are both positive integers, and their specific values ​​can be the same or different. Furthermore, since there is no previously determined title word in the process of determining the first title word, the starting character "" is used. <s>"Replace the previously determined title word as the input of decoder 602."

[0147] It should be noted that the above embodiments determine whether to directly copy word segments as title words by setting an attention threshold. Optionally, in some embodiments, the server can also use a neural network to process the attention vector to obtain the copy probability of each word segment, and determine the word segments with a copy probability greater than the copy threshold as title words. This application does not limit the specific method for determining whether to directly copy word segments as title words. The copy threshold can be flexibly configured, for example, a copy threshold of 0.7 or 0.8, etc.

[0148] Another point to note is that the above embodiments are designed to determine titles with higher accuracy. Of course, the server may not directly copy the segmented words into title words; it can directly determine the words output by the title generation model as title words. Furthermore, the server may not use an attention mechanism to determine title words; this application does not impose any restrictions on this.

[0149] Another point to note is that the above embodiments use four aspects—tutorial main keywords, video type, multimodal vector, and original title—to generate titles with higher accuracy. In other embodiments, the server may also generate titles based on one or more of the tutorial main keywords, video type, multimodal vector, and original title; this application does not impose any restrictions on this. The process of generating titles based on one or more of the tutorial main keywords, video type, multimodal vector, and original title is similar to the title generation process in step 306, and will not be listed here individually.

[0150] 307. The server responds to the presence of tutorial keywords in the original title, and generates a target title that matches the style of the tutorial title based on the video type and the original title.

[0151] Step 307 is similar to step 306. To make the title generation process clearer, the following will combine... Figure 7 Please provide an explanation. See also... Figure 7 The original title of the video is segmented into multiple words. These segments are then converted into vector representations, resulting in words 1 to n. Words 1 to n are encoded by encoder 701 to obtain word representations 1 to n. For each title word, the server uses an attention mechanism, based on the latent vector of the previous title word determined by decoder 702 and the word representations 1 to n, to determine the attention weight of each segment. If the attention weight of the segment with the largest attention weight is greater than the attention threshold, then that segment is determined as the title word for this determination, i.e., copied from the input text. If the attention weight of each segment is less than or equal to the attention threshold, decoder 702 determines the latent vector on which the current title word is based, based on the type vector of the input video type, the determined previous title word, the vector obtained by weighted summation of word representations 1 to n based on attention weights, and the latent vector on which the previous title word was determined. Based on this latent vector, the vector obtained by weighted summation of word representations 1 to n based on attention weights, and the determined previous title word, the current title word is obtained.

[0152] The above technical solution, since the original title already contains the main subject words that can represent the theme of the tutorial, can generate a more accurate tutorial title by combining the original title. Therefore, there is no need to combine the video content to generate the title, thereby reducing the computing resources consumed in title generation, improving resource utilization, and improving the efficiency of title generation when the main subject words of the tutorial are in the original title.

[0153] This application provides a scheme for generating titles for tutorial videos. The scheme identifies the original title of the tutorial video and, if the original title does not conform to the style of a tutorial title, automatically and efficiently generates a title that conforms to that style based on the original title and the video content. Because the title generation fully utilizes both the original title and the video content, it more accurately reflects the core content taught in the tutorial video, allowing users to intuitively identify the tutorial's points of interest from the title. This reduces the cost for users to discover these points, enhances the tutorial video's appeal, increases its play rate, and thus improves its dissemination efficiency. Furthermore, compared to manually writing tutorial titles, the title generation efficiency is higher, meeting the needs of writing titles for a large number of newly added tutorial videos. In addition, this application provides personalized optimizations to the tutorial title generation process. By effectively identifying the tutorial video to generate the title, compared to the common method of automatic title generation, it reduces the generation of invalid titles, lowers the probability of generating invalid titles, reduces the waste of computing resources, and improves resource utilization.

[0154] It should be noted that if the original title contains the main keywords of the tutorial, the server can also generate a title based on one or more of the main keywords of the tutorial, the video type, the multimodal vector, and the original title. This application does not impose any restrictions on this.

[0155] Another point to note is that after the server generates a tutorial title for the tutorial video, it stores the correspondence between the tutorial video and the tutorial title. In some embodiments, in response to a terminal's request, the server returns the tutorial video and tutorial title to the terminal, allowing the terminal to display the tutorial video and tutorial title. Optionally, the terminal displays the tutorial video's cover image; the terminal can display the tutorial title below the video cover image or display the tutorial title on the video cover image, which is not limited in this application. In other embodiments, the server provides the tutorial title corresponding to the tutorial video to a recommendation service, enabling the recommendation service to quickly and effectively recommend tutorial videos based on the tutorial title, increasing the exposure of tutorial videos, further improving the playback rate of tutorial videos, and improving the dissemination efficiency of tutorial videos. In other embodiments, the server provides the tutorial title corresponding to the tutorial video to a search service, enabling the search service to query the tutorial title based on search keywords, quickly and effectively determine tutorial videos matching the search keywords, and display the tutorial videos to users, thereby increasing the exposure of tutorial videos, further improving the playback rate of tutorial videos, and improving the dissemination efficiency of tutorial videos.

[0156] Another point to note is that, in some embodiments, after the server generates a tutorial title for the tutorial video, it can flexibly display the original title of the video or the tutorial title to the user based on the user's interests. Optionally, the server can display the tutorial titles of some tutorial videos and the original titles of another set of tutorial videos to the user, obtain the click-through rate of the user on both sets of videos, and determine the user's interest as either the tutorial title or the original title based on the click-through rate. Of course, the server can also combine metrics such as the user's click-through rate, viewing time, collection rate, like rate, and sharing rate of the two sets of tutorial videos to determine user interests; this application does not impose any limitations on this. In other embodiments, the server can obtain multiple title words with high probability ranking from a target vocabulary to generate multiple tutorial titles; based on the user's interests, it dynamically selects one of the multiple tutorial titles to display to the user. The process by which the server determines the user's interest in different tutorial titles is similar to the process by which the server determines the user's interest in the original title and the tutorial title, and will not be elaborated further here.

[0157] Another point to note is that the above embodiments are described using a server as the execution subject. In some embodiments, the execution subject of this video title generation method can also be a terminal, such as a desktop computer, laptop computer, tablet computer, or smartphone. The terminal can be equipped with a video playback client, a video processing client, or a video title generation client. The terminal can generate tutorial titles for tutorial videos through the client. The process of the terminal generating tutorial titles for tutorial videos is the same as the process of the server generating tutorial titles for tutorial videos, and will not be described again here.

[0158] Another point to note is that before the server executes steps 301 to 307 above, it also trains a relevant model to generate titles based on the trained model. For example, see... Figure 8 The server generates a tutorial video discrimination model 802, a style recognition model 803, and a title generation model 804 based on a video library 801 with labeled tutorial titles. The tutorial video discrimination model identifies whether a video belongs to a tutorial video; the style recognition model determines whether the original title conforms to the style of a tutorial title; and the title generation model generates a title that conforms to the aforementioned style. Once a title conforming to the style is generated, it can be used. The storage format of the videos in video library 801 is shown in Table 1.

[0159] Table 1

[0160] video Is it a tutorial video? Video type Original title Annotated tutorial title Video 1 yes Exercise and fitness Original title 1 Tutorial Title 1 Video 2 no TV drama Original title 2 Video 3 yes Life Original title 3 Tutorial Title 3 Video 4 yes gourmet food Original title 4 Tutorial Title 4 Video 5 no archeology Original title 5 Video 6 yes Computer / Mobile Phone Original title 6 Tutorial Title 6

[0161] The process by which the server application tutorial video discrimination model 802 determines whether a video belongs to the category of tutorial videos and determines the video type is also referred to as steps 301 to 302. The structure of the tutorial video discrimination model 802 is as follows: Figure 4 As shown, the server performs target learning simultaneously on two tasks: determining whether a video belongs to a tutorial video and determining the video type, based on the videos in the video library 801, and trains the tutorial video discrimination model 802.

[0162] The process by which the server application style recognition model 803 determines whether the original title conforms to the expression style of the tutorial title is also step 303. The style recognition model is trained using the first type of title as positive samples and at least one of the second and third types of titles as negative samples. The first type of title conforms to the expression style, and the corresponding video belongs to the tutorial video category. The second type of title is the original title of a video that does not belong to the tutorial video category. The third type of title does not conform to the expression style, and the corresponding video belongs to the tutorial video category. In other words, a positive sample (a tutorial title) is a tutorial title labeled "Is it a tutorial video? == Yes". A negative sample is an original title labeled "Is it a tutorial video? == No", or an original title labeled "Is it a tutorial video? == Yes" and "Original title != labeled tutorial title".

[0163] The process of generating a title using the server application title generation model 804 is also known as steps 305 to 306 or step 307. The structure of title generation model 804 is as follows: Figure 6 or Figure 7 As shown, the server learns from the video library 801 based on the network structure of the title generation model 804 and continuously updates the parameters in the network structure to obtain the title generation model 804.

[0164] In addition, the server also uses a sequence labeling model to determine the main keywords of the tutorial from the text. This sequence labeling model is trained based on manually labeled main keywords of the tutorial in the text.

[0165] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0166] Figure 9 This is a block diagram of a video title generation device provided in an embodiment of this application. See also... Figure 9 The device includes: a tutorial video recognition module 901, a title style recognition module 902, and a title generation module 903.

[0167] The tutorial video recognition module 901 is used to obtain the content tags of the video, which are used to indicate whether the video belongs to the tutorial video category.

[0168] The title style recognition module 902 is used to identify the original title of the video in response to the content tag indicating that the video belongs to the tutorial video, and obtain the style tag of the original title. The style tag is used to indicate whether the original title conforms to the expression style of the tutorial title.

[0169] The title generation module 903 is used to generate a target title that conforms to the above expression style based on the original title and the video, in response to the style tag indicating that the original title does not conform to the above expression style.

[0170] This application provides a scheme for generating titles for tutorial videos. The scheme identifies the original title of the tutorial video and, if the original title does not conform to the style of a tutorial title, automatically and efficiently generates a title that conforms to that style based on the original title and the video content. Because the generated title fully utilizes both the original title and the video content, it more accurately reflects the core content taught in the tutorial video, allowing users to intuitively identify the tutorial's points of interest from the title. This enhances the tutorial video's appeal to users, increases its play rate, and ultimately improves its dissemination efficiency.

[0171] In one alternative implementation, the tutorial video recognition module 901 includes:

[0172] The video feature extraction unit is used to extract features from the video and obtain feature vectors of the video in multiple dimensions.

[0173] The tutorial video recognition unit is used to determine the content tags of a video based on its feature vectors in multiple dimensions.

[0174] In another alternative implementation, the video feature extraction unit includes:

[0175] The content acquisition subunit is used to acquire multiple image frames and multiple audio frames of the video;

[0176] The image feature extraction subunit is used to determine the feature vector of the image dimension of the video based on multiple image frames;

[0177] The audio feature extraction subunit is used to determine the feature vector of the audio dimension of the video based on multiple audio frames.

[0178] The text feature extraction subunit is used to determine the text dimension feature vector of the video based on the text information corresponding to the video.

[0179] In another alternative implementation, the image feature extraction subunit is used for:

[0180] Feature extraction is performed on multiple image frames to obtain multiple feature vectors for the multiple image frames;

[0181] By aggregating multiple feature vectors from multiple image frames, we obtain the feature vectors of the video's image dimensions.

[0182] In another alternative implementation, the device further includes:

[0183] The text recognition module is used to perform text recognition on multiple image frames to obtain subtitle text;

[0184] The speech recognition module is used to recognize speech from multiple audio frames to obtain the dialogue text;

[0185] The text information determination module is used to determine the original title, subtitle text, and dialogue text as text information.

[0186] In another alternative implementation, the tutorial video recognition unit is used for:

[0187] The feature vectors of the video in multiple dimensions are fused to obtain the multimodal vector of the video;

[0188] Based on multimodal vectors, determine the content tags of the video.

[0189] In another alternative implementation, the title generation module 903 includes:

[0190] The tutorial main keyword determination unit is used to determine the tutorial main keywords corresponding to the video in response to style tag indications that the original title does not conform to the above expression style;

[0191] The title generation unit is used to generate a target title based on at least one of the following: the main keywords of the tutorial, the video type of the video, and the multimodal vector of the video, as well as the original title.

[0192] Among them, the tutorial subject words are used to represent the tutorial theme of the video, and the multimodal vector is obtained by fusing the feature vectors of the video in multiple dimensions.

[0193] In another alternative implementation, the tutorial subject word determination unit is used for:

[0194] In response to style tag indications that the original title does not conform to the above style of expression, the main keywords of the tutorial are determined from the original title;

[0195] Since the tutorial subject words are not present in the original title, the tutorial subject words are determined from the video's subtitle text and dialogue text.

[0196] In another alternative implementation, the title generation unit includes:

[0197] Word segmentation is used to divide the main text of the tutorial and the original title into multiple words.

[0198] The title word determination subunit is used to input video type, multimodal vector and multiple word segments into the title generation model, and the title generation model determines multiple title words. The title generation model is used to generate a title that conforms to the above expression style.

[0199] The title generation sub-unit is used to compose a target title from multiple title words.

[0200] In another alternative implementation, the title word determines the sub-unit, used for:

[0201] The video type, multimodal vector, and multiple word segments are input into the title generation model to obtain the attention weights of multiple word segments;

[0202] If the attention weight of the first segment among multiple segmented words is greater than the attention threshold, the first segmented word is identified as the title word.

[0203] If the attention weights of multiple word segments are all less than or equal to the attention threshold, the words output by the title generation model are determined as title words.

[0204] In another alternative implementation, the device further includes:

[0205] The video type determination module is used to determine the video type based on the video's feature vectors in multiple dimensions.

[0206] In another alternative implementation, the title style recognition module 902 is used for:

[0207] In response to the content tag indicating that the video belongs to the tutorial video, the original title is input into the style recognition model to obtain the style tag output by the style recognition model;

[0208] The style recognition model is trained using the first type of title as positive samples and at least one of the second and third types of titles as negative samples. The first type of title conforms to the above expression style and the corresponding video belongs to the tutorial video category. The second type of title is the original title of the video that does not belong to the tutorial video category. The third type of title does not conform to the above expression style and the corresponding video belongs to the tutorial video category.

[0209] It should be noted that the video title generation device provided in the above embodiments is only illustrated by the division of the above functional modules when generating video titles. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the video title generation device and the video title generation method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0210] In this application embodiment, the computer device may be configured as a terminal or a server. If the computer device is configured as a terminal, the terminal acts as the execution entity to implement the technical solutions provided in this application embodiment. If the computer device is configured as a server, the server acts as the execution entity to implement the technical solutions provided in this application embodiment. Alternatively, the technical solutions provided in this application embodiment may be implemented through interaction between the terminal and the server, and this application embodiment does not limit this approach.

[0211] If the computer device is configured as a terminal. Figure 10 This illustration shows a structural block diagram of a terminal 1000 provided in an exemplary embodiment of this application. The terminal 1000 may be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, smart TV, smart vehicle, etc. The terminal 1000 may also be referred to as user equipment, portable terminal, laptop terminal, desktop terminal, or other names.

[0212] Typically, terminal 1000 includes a processor 1001 and a memory 1002.

[0213] Processor 1001 may include one or more processing cores, such as a quad-core processor, an octa-core processor, etc. Processor 1001 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 1001 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 1001 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content required to be displayed on the screen. In some embodiments, processor 1001 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0214] The memory 1002 may include one or more computer-readable storage media, which may be non-transitory. The memory 1002 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 1002 are used to store at least one piece of program code, which is executed by the processor 1001 to implement the video title generation method provided in the method embodiments of this application.

[0215] In some embodiments, the terminal 1000 may also optionally include a peripheral device interface 1003 and at least one peripheral device. The processor 1001, memory 1002, and peripheral device interface 1003 can be connected via a bus or signal line. Each peripheral device can be connected to the peripheral device interface 1003 via a bus, signal line, or circuit board. Specifically, the peripheral device includes at least one of the following: a radio frequency circuit 1004, a display screen 1005, a camera assembly 1006, an audio circuit 1007, a positioning assembly 1008, and a power supply 1009.

[0216] Peripheral device interface 1003 can be used to connect at least one I / O (Input / Output) related peripheral device to processor 1001 and memory 1002. In some embodiments, processor 1001, memory 1002 and peripheral device interface 1003 are integrated on the same chip or circuit board; in some other embodiments, any one or two of processor 1001, memory 1002 and peripheral device interface 1003 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0217] The radio frequency (RF) circuit 1004 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The RF circuit 1004 communicates with communication networks and other communication devices via electromagnetic signals. The RF circuit 1004 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals back into electrical signals. Optionally, the RF circuit 1004 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, etc. The RF circuit 1004 can communicate with other terminals via at least one wireless communication protocol. This wireless communication protocol includes, but is not limited to: the World Wide Web, metropolitan area networks, intranets, various generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area networks, and / or WiFi (Wireless Fidelity) networks. In some embodiments, the RF circuit 1004 may also include circuitry related to NFC (Near Field Communication), which is not limited in this application.

[0218] Display screen 1005 is used to display a UI (User Interface). This UI may include graphics, text, icons, videos, and any combination thereof. When display screen 1005 is a touch display screen, it also has the ability to collect touch signals on or above its surface. These touch signals can be input as control signals to processor 1001 for processing. In this case, display screen 1005 can also be used to provide virtual buttons and / or a virtual keyboard, also known as soft buttons and / or a soft keyboard. In some embodiments, there may be one display screen 1005, disposed on the front panel of terminal 1000; in other embodiments, there may be at least two display screens, disposed on different surfaces of terminal 1000 or in a folded design; in still other embodiments, display screen 1005 may be a flexible display screen, disposed on a curved or folded surface of terminal 1000. Furthermore, display screen 1005 may be configured as a non-rectangular, irregular shape, i.e., a non-rectangular screen. The display screen 1005 can be made of materials such as LCD (Liquid Crystal Display) and OLED (Organic Light-Emitting Diode).

[0219] The camera assembly 1006 is used to acquire images or videos. Optionally, the camera assembly 1006 includes a front-facing camera and a rear-facing camera. Typically, the front-facing camera is located on the front panel of the terminal, and the rear-facing camera is located on the back of the terminal. In some embodiments, there are at least two rear-facing cameras, which are any one of a main camera, a depth-sensing camera, a wide-angle camera, and a telephoto camera, to achieve background blurring by fusion of the main camera and the depth-sensing camera, panoramic shooting by fusion of the main camera and the wide-angle camera, VR (Virtual Reality) shooting, or other fusion shooting functions. In some embodiments, the camera assembly 1006 may also include a flash. The flash can be a single-color temperature flash or a dual-color temperature flash. A dual-color temperature flash refers to a combination of a warm-light flash and a cool-light flash, which can be used for light compensation at different color temperatures.

[0220] The audio circuit 1007 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, converting the sound waves into electrical signals that are input to the processor 1001 for processing, or input to the radio frequency circuit 1004 for voice communication. For stereo sound acquisition or noise reduction purposes, multiple microphones may be used, each positioned at a different location on the terminal 1000. The microphone may also be an array microphone or an omnidirectional microphone. The speaker is used to convert electrical signals from the processor 1001 or the radio frequency circuit 1004 into sound waves. The speaker may be a conventional diaphragm speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can convert electrical signals not only into audible sound waves but also into inaudible sound waves for purposes such as distance measurement. In some embodiments, the audio circuit 1007 may also include a headphone jack.

[0221] The positioning component 1008 is used to determine the current geographical location of the terminal 1000 in order to enable navigation or LBS (Location Based Service). The positioning component 1008 can be a positioning component based on the US GPS (Global Positioning System), China's BeiDou system, or Russia's Galileo system.

[0222] Power supply 1009 is used to power the various components in terminal 1000. Power supply 1009 can be AC ​​power, DC power, a disposable battery, or a rechargeable battery. When power supply 1009 includes a rechargeable battery, the rechargeable battery can be a wired rechargeable battery or a wireless rechargeable battery. A wired rechargeable battery is a battery that is charged via a wired line, and a wireless rechargeable battery is a battery that is charged via a wireless coil. The rechargeable battery can also be used to support fast charging technology.

[0223] In some embodiments, the terminal 1000 further includes one or more sensors 1010. The one or more sensors 1010 include, but are not limited to: an accelerometer 1011, a gyroscope 1012, a pressure sensor 1013, a fingerprint sensor 1014, an optical sensor 1015, and a proximity sensor 1016.

[0224] Those skilled in the art will understand that Figure 10 The structure shown does not constitute a limitation on terminal 1000 and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0225] If the computer device is configured as a server Figure 11 This is a block diagram of a server provided in an embodiment of this application. The server 1100 can vary significantly due to different configurations or performance. It may include one or more Central Processing Units (CPUs) 1101 and one or more memories 1102. The memories 1102 store computer programs, which are loaded and executed by the processors 1101 to implement the video title generation methods provided in the various method embodiments described above. Of course, the server may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. The server may also include other components for implementing device functions, which will not be elaborated upon here.

[0226] In an exemplary embodiment, a computer-readable storage medium is also provided, which stores a computer program that can be executed by a processor in a computer device to perform the video title generation method described above. For example, the computer-readable storage medium may be a ROM (Read-Only Memory), RAM (Random Access Memory), CD-ROM (Compact Disc Read-Only Memory), magnetic tape, floppy disk, or optical data storage device, etc.

[0227] This application also provides a computer program product or computer program that includes computer program code stored in a computer-readable storage medium. A processor of a computer device reads the computer program code from the computer-readable storage medium and executes the computer program code, causing the computer device to perform the video title generation method in the above-described method embodiments.

[0228] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0229] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.< / s>

Claims

1. A method for generating video titles, characterized in that, The method includes: Obtain the content tags of the video, which are used to indicate whether the video belongs to a tutorial video; In response to the content tag indicating that the video belongs to a tutorial video, the original title of the video is identified to obtain the style tag of the original title. The style tag is used to indicate whether the original title conforms to the expression style of a tutorial title. In response to the style tag indicating that the original title does not conform to the expression style, the tutorial subject words corresponding to the video are determined; based on the tutorial subject words and the original title, a target title conforming to the expression style is generated, or, based on at least one of the video type and the multimodal vector of the video, the tutorial subject words, and the original title, a target title conforming to the expression style is generated; wherein, the tutorial subject words are used to represent the tutorial theme of the video, and the multimodal vector is obtained by fusing the feature vectors of the video in multiple dimensions.

2. The method according to claim 1, characterized in that, The acquisition of video content tags includes: Feature extraction is performed on the video to obtain feature vectors of the video in multiple dimensions; Based on the feature vectors of the video in multiple dimensions, the content tags of the video are determined.

3. The method according to claim 2, characterized in that, The step of extracting features from the video to obtain feature vectors of the video in multiple dimensions includes: Acquire multiple image frames and multiple audio frames of the video; Based on the multiple image frames, determine the feature vector of the image dimension of the video; Based on the multiple audio frames, determine the feature vector of the audio dimension of the video; Based on the text information corresponding to the video, the feature vector of the text dimension of the video is determined.

4. The method according to claim 3, characterized in that, Determining the feature vector of the video's image dimensions based on the multiple image frames includes: Feature extraction is performed on the multiple image frames respectively to obtain multiple feature vectors of the multiple image frames; The feature vectors of the multiple image frames are aggregated to obtain the feature vector of the image dimension of the video.

5. The method according to claim 3, characterized in that, Before determining the text-dimensional feature vector of the video based on the text information corresponding to the video, the method further includes: Text recognition is performed on the multiple image frames to obtain subtitle text; Speech recognition is performed on the multiple audio frames to obtain the dialogue text; The original title, the subtitle text, and the dialogue text are identified as the text information.

6. The method according to claim 2, characterized in that, The step of determining the content tags of the video based on its feature vectors across multiple dimensions includes: The feature vectors of the video in multiple dimensions are fused to obtain the multimodal vector of the video; Based on the multimodal vectors, the content tags of the video are determined.

7. The method according to claim 1, characterized in that, In response to the style tag indicating that the original title does not conform to the expression style, the method for determining the tutorial subject words corresponding to the video includes: In response to the style tag indicating that the original title does not conform to the expression style, the main keywords of the tutorial are determined from the original title; In response to the absence of the tutorial subject word in the original title, the tutorial subject word is determined from the subtitle text and the dialogue text of the video.

8. The method according to claim 1, characterized in that, The process of generating a target title that conforms to the stated style based on at least one of the video type and the multimodal vector of the video, the main keywords of the tutorial, and the original title includes: The main words of the tutorial and the original title are divided into multiple word segments; The video type, the multimodal vector, and the multiple word segments are input into the title generation model, and multiple title words are determined by the title generation model. The title generation model is used to generate a title that conforms to the expression style. The target title is composed of the multiple title words.

9. The method according to claim 8, characterized in that, The process involves inputting the video type, the multimodal vector, and the multiple word segments into a title generation model, and determining multiple title words through the title generation model, including: The video type, the multimodal vector, and the multiple word segments are input into the title generation model to obtain the attention weights of the multiple word segments; In response to the first segmentation among the plurality of segmented words having an attention weight greater than an attention threshold, the first segmentation is determined as the title word; In response to the fact that the attention weights of the multiple word segments are all less than or equal to the attention threshold, the word output by the title generation model is determined as the title word.

10. The method according to claim 1, characterized in that, Before generating the target title that conforms to the stated style, the method further includes: The video type is determined based on the feature vectors of the video in multiple dimensions.

11. The method according to claim 1, characterized in that, In response to the content tag indicating that the video belongs to a tutorial video, the original title of the video is identified to obtain the style tag of the original title, including: In response to the content tag indicating that the video belongs to a tutorial video, the original title is input into the style recognition model to obtain the style tag output by the style recognition model; The style recognition model is trained using a first type of title as positive samples and at least one of a second or third type of title as negative samples. The first type of title conforms to the expression style, and the corresponding video belongs to the tutorial video category. The second type of title is the original title of a video that does not belong to the tutorial video category. The third type of title does not conform to the expression style, and the corresponding video belongs to the tutorial video category.

12. A video title generation device, characterized in that, The device includes: The tutorial video recognition module is used to obtain the content tags of the video, and the content tags are used to indicate whether the video belongs to the tutorial video. The title style recognition module is used to identify the original title of the video in response to the content tag indicating that the video belongs to the tutorial video, and obtain the style tag of the original title. The style tag is used to indicate whether the original title conforms to the expression style of the tutorial title. The title generation module includes: a tutorial subject word determination unit, used to determine the tutorial subject words corresponding to the video in response to the style tag indicating that the original title does not conform to the expression style; and a title generation unit, used to generate a target title conforming to the expression style based on the tutorial subject words and the original title, or, based on at least one of the video type and the multimodal vector of the video, the tutorial subject words, and the original title; wherein the tutorial subject words are used to represent the tutorial theme of the video, and the multimodal vector is obtained by fusing the feature vectors of the video in multiple dimensions.

13. The apparatus according to claim 12, characterized in that, The tutorial video recognition module includes: A video feature extraction unit is used to extract features from the video to obtain feature vectors of the video in multiple dimensions. The tutorial video recognition unit is used to determine the content tags of the video based on the feature vectors of the video in multiple dimensions.

14. The apparatus according to claim 13, characterized in that, The video feature extraction unit includes: The content acquisition subunit is used to acquire multiple image frames and multiple audio frames of the video; An image feature extraction subunit is used to determine the feature vector of the image dimension of the video based on the multiple image frames; An audio feature extraction subunit is used to determine the feature vector of the audio dimension of the video based on the multiple audio frames; The text feature extraction subunit is used to determine the text dimension feature vector of the video based on the text information corresponding to the video.

15. The apparatus according to claim 14, characterized in that, The image feature extraction subunit is used for: Feature extraction is performed on the multiple image frames respectively to obtain multiple feature vectors of the multiple image frames; The feature vectors of the multiple image frames are aggregated to obtain the feature vector of the image dimension of the video.

16. The apparatus according to claim 14, characterized in that, The device further includes: The text recognition module is used to perform text recognition on the multiple image frames to obtain subtitle text; The speech recognition module is used to perform speech recognition on the multiple audio frames to obtain the dialogue text; The text information determination module is used to determine the original title, the subtitle text, and the dialogue text as the text information.

17. The apparatus according to claim 13, characterized in that, The tutorial video recognition unit is used for: The feature vectors of the video in multiple dimensions are fused to obtain the multimodal vector of the video; Based on the multimodal vectors, the content tags of the video are determined.

18. The apparatus according to claim 12, characterized in that, The tutorial main word determination unit is used for: In response to the style tag indicating that the original title does not conform to the expression style, the main keywords of the tutorial are determined from the original title; In response to the absence of the tutorial subject word in the original title, the tutorial subject word is determined from the subtitle text and the dialogue text of the video.

19. The apparatus according to claim 12, characterized in that, The title generation unit includes: The word segmentation subunit is used to divide the main words of the tutorial and the original title into multiple words; The title word determination subunit is used to input the video type, the multimodal vector, and the multiple word segments into the title generation model, and determine multiple title words through the title generation model. The title generation model is used to generate a title that conforms to the expression style. The title generation subunit is used to compose the target title from the plurality of title words.

20. The apparatus according to claim 19, characterized in that, The title word determination subunit is used for: The video type, the multimodal vector, and the multiple word segments are input into the title generation model to obtain the attention weights of the multiple word segments; In response to the first segmentation among the plurality of segmented words having an attention weight greater than an attention threshold, the first segmentation is determined as the title word; In response to the fact that the attention weights of the multiple word segments are all less than or equal to the attention threshold, the word output by the title generation model is determined as the title word.

21. The apparatus according to claim 12, characterized in that, The device further includes: The video type determination module is used to determine the video type of the video based on the feature vectors of the video in multiple dimensions.

22. The apparatus according to claim 12, characterized in that, The title style recognition module is used for: In response to the content tag indicating that the video belongs to a tutorial video, the original title is input into the style recognition model to obtain the style tag output by the style recognition model; The style recognition model is trained using a first type of title as positive samples and at least one of a second or third type of title as negative samples. The first type of title conforms to the expression style, and the corresponding video belongs to the tutorial video category. The second type of title is the original title of a video that does not belong to the tutorial video category. The third type of title does not conform to the expression style, and the corresponding video belongs to the tutorial video category.

23. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program, which is loaded and executed by the processor to implement the video title generation method as described in any one of claims 1-11.

24. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which is loaded and executed by a processor to implement the video title generation method as described in any one of claims 1-11.

25. A computer program product, characterized in that, The computer program product includes computer program code, which, when executed by a processor, implements the video title generation method as described in any one of claims 1-11.

Images