Video recall method, device, electronic device and storage medium

By constructing heterogeneous graphs and training video recall models, the video recall rate is improved by using the association relationship between video titles, query statements and query paths, the video recall rate is solved, the problem of low recall rate in the existing technology is achieved, and more accurate video matching is achieved.

CN113761270BActive Publication Date: 2024-11-26TENCENT TECHNOLOGY (SHENZHEN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110536313.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-05-17
Publication Date
2024-11-26
Estimated Expiration
2041-05-17

AI Technical Summary

Technical Problem

The recall rate of existing video recall methods is low and cannot effectively meet users' needs to find interesting videos in massive videos.

Method used

By obtaining video title and query information, detecting the same keywords between the video title and the query statement, building heterogeneous graphs, training the preset video recall model based on heterogeneous graph, query statement and video title, and using query paths and keys The word represents the relationship between video and query statements and video recalls.

Benefits of technology

It improves the recall rate during video query and can more accurately match the video content expected by users.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113761270B_ABST
    Figure CN113761270B_ABST
Patent Text Reader

Abstract

The present application discloses a video recall method, device, electronic device and storage medium, including: obtaining a video title and query information corresponding to a sample video, wherein the query information includes a query statement and a query path, detecting the same keywords between the video title and the query statement, obtaining the same keywords, constructing a heterogeneous graph for characterizing the association relationship between the sample video, the query statement and the query path according to the query path and the same keywords, training a preset video recall model based on the heterogeneous graph, the query statement and the video title, and performing video recall through the trained video recall model. This scheme can improve the recall rate during video query.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a video recall method, device, electronic equipment and storage medium. Background Art

[0002] With the development of online video platforms, the number of videos continues to expand, and users need to spend a lot of time searching for videos of interest from a massive amount of videos.

[0003] The current video recall method is to recall videos based on the search terms input by the user. Specifically, the video is recalled using the semantic association between the query statement and the video title. However, the videos recalled by this method are often different from the videos expected by the user. That is, the current video recall method has a low recall rate. Summary of the invention

[0004] The present application provides a video recall method, device, electronic device and storage medium, which can improve the recall rate during video query.

[0005] This application provides a video recall method, including:

[0006] Obtaining a video title and query information corresponding to a sample video, wherein the query information includes a query statement and a query path;

[0007] Detecting the same keywords between the video title and the query statement to obtain the same keywords;

[0008] According to the query path and the same keyword, construct a heterogeneous graph for characterizing the association relationship between the sample video, the query statement and the query path;

[0009] A preset video recall model is trained based on the heterogeneous graph, query statements and video titles, and video recall is performed using the trained video recall model.

[0010] Accordingly, the present application also provides a video recall device, comprising:

[0011] An acquisition module, used to acquire a video title and query information corresponding to a sample video, wherein the query information includes a query statement and a query path;

[0012] A detection module, used to detect the same keywords between the video title and the query sentence, and obtain the same keywords;

[0013] A construction module, used to construct a heterogeneous graph for characterizing the association relationship between the sample video, the query statement and the query path according to the query path and the same keyword;

[0014] A training module, used for training a preset video recall model based on the heterogeneous graph, query statements and video titles;

[0015] The recall module is used to perform video recall using the trained video recall model.

[0016] Optionally, in some embodiments of the present invention, the training module includes:

[0017] A processing unit, configured to perform graph embedding processing on the heterogeneous graph to obtain a node vector corresponding to each node in the heterogeneous graph;

[0018] A word segmentation unit is used to perform word segmentation processing on the query statement and the video title respectively to obtain a query word set corresponding to the query statement and a title word set corresponding to the video title;

[0019] The training unit is used to train a preset video recall model based on the node vector, the query word set and the title word set to obtain a trained video recall model.

[0020] Optionally, in some embodiments of the present application, the training unit includes:

[0021] A first construction subunit is used to construct a first semantic feature corresponding to the query statement and a second semantic feature corresponding to the video title according to the first subnetwork, the query word set and the title word set in the video recall model;

[0022] A second construction subunit is used to construct a first heterogeneous graph feature corresponding to the query statement and a second heterogeneous graph feature corresponding to the video title based on the second subnetwork in the video recall model, the node vector and the same keyword;

[0023] The training subunit is used to train a preset video recall model according to the first semantic feature, the second semantic feature, the first heterogeneous graph feature and the second heterogeneous graph feature to obtain a trained video recall model.

[0024] Optionally, in some embodiments of the present application, the second construction subunit is specifically used for:

[0025] Inputting the node vector into the second subnetwork in the video recall model to obtain a vector feature corresponding to each node in the heterogeneous graph, wherein the vector feature carries an association weight between nodes;

[0026] Obtaining vector features corresponding to the same keyword;

[0027] According to the association weights between nodes, construct query features corresponding to each node vector under different query statements, and fuse the query features to obtain a first heterogeneous graph feature corresponding to the query statement;

[0028] Based on the vector features corresponding to the same keyword, a second heterogeneous graph feature corresponding to the video title is constructed.

[0029] Optionally, in some embodiments of the present application, the first construction subunit is specifically used for:

[0030] Inputting the first word vector into the first subnetwork of the video recall model to obtain a first word embedding representation corresponding to each query word in the query word set, and performing average pooling processing on the first word embedding representation to obtain a first semantic feature corresponding to the query sentence, and;

[0031] The second word vector is input into the second subnetwork of the video recall model to obtain a second word embedding representation corresponding to each title word in the title word set, and the second word embedding representation is average pooled to obtain a second semantic feature corresponding to the video title.

[0032] Optionally, in some embodiments of the present application, the training subunit is specifically used for:

[0033] fusing the first semantic feature and the first heterogeneous graph feature to obtain a query feature corresponding to the query statement, and;

[0034] fusing the second semantic feature and the second heterogeneous graph feature to obtain a title feature corresponding to the video title;

[0035] The preset video recall model is trained according to the query features and the title features to obtain a trained video recall model.

[0036] Optionally, in some embodiments of the present application, a construction unit is further included, and the construction unit is specifically used to:

[0037] retaining keywords in the query statement and the video title whose frequency is less than a preset value, and constructing a reference dictionary based on the retained keywords;

[0038] The text lengths of the query sentence and the video title are aligned according to the reference dictionary.

[0039] Optionally, in some embodiments of the present application, the recall module is specifically used to:

[0040] When a video search operation is received, a video set to be recalled, a video search statement corresponding to the video search operation, and a video search path are obtained, wherein the video set to be recalled includes at least one video to be recalled, and the video to be recalled corresponds to a video title to be recalled;

[0041] The text similarity between the video title to be recalled and the video search sentence is calculated, and the video is recalled through the trained video recall model based on the text similarity.

[0042] Optionally, in some embodiments of the present application, the detection module is specifically used to:

[0043] Counting the frequency of each keyword appearing in the video title, and counting the frequency of each keyword appearing in the query statement;

[0044] Removing keywords with a frequency greater than a preset value from the video title to obtain a processed video title, and;

[0045] Removing keywords with a frequency greater than a preset value from the query statement to obtain a processed query statement;

[0046] The same keywords between the processed video title and the processed query statement are detected to obtain the same keywords.

[0047] After obtaining the video title and query information corresponding to the sample video, the application detects the same keywords between the video title and the query statement to obtain the same keywords, and then, based on the query path and the same keywords, constructs a heterogeneous graph for characterizing the association relationship between the sample video, the query statement and the query path. Finally, the preset video recall model is trained based on the heterogeneous graph, the query statement and the video title, and the video is recalled through the trained video recall model. The video recall method provided by the application uses the query path and the same keywords to construct a heterogeneous graph for characterizing the association relationship between the sample video, the query statement and the query path. Then, the preset video recall model is trained based on the heterogeneous graph, the query statement and the video title. When the video is subsequently recalled through the trained video recall model, not only the semantic association between the query statement and the video title can be used, but also the association between the video, the query statement and the query path can be used, thereby improving the recall rate during video query. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0049] Figure 1a is a schematic diagram of a scenario of a video recall method provided by this application;

[0050] Figure 1b It is a flowchart of the video recall method provided by this application;

[0051] Figure 1c is a heterogeneous graph of the video recall method provided by this application;

[0052] Figure 1d It is an architecture diagram of a video recall model in the video recall method provided by this application;

[0053] Figure 1e is a schematic diagram of the attention mechanism in the video recall method provided in this application;

[0054] Figure 1f is a structural diagram of a gating unit in a video recall method provided by the present application;

[0055] Figure 2 is another flow chart of the video recall method provided by the present application;

[0056] Figure 3 It is a structural schematic diagram of the video recall device provided by the present application;

[0057] Figure 4 It is a structural schematic diagram of the electronic device provided by this application. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in this application to clearly and completely describe the technical solutions in this application. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments in the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0059] Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making.

[0060] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. The basic technologies of artificial intelligence generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0061] Natural language processing (NLP) is an important direction in the fields of computer science and artificial intelligence. It studies various theories and methods that can achieve effective communication between people and computers using natural language. Natural language processing is a science that integrates linguistics, computer science, and mathematics. Therefore, research in this field will involve natural language, that is, the language people use in daily life, so it is closely related to the study of linguistics. Natural language processing technology usually includes text processing, semantic understanding, machine translation, robot question answering, knowledge graph and other technologies.

[0062] Machine Learning (ML) is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers 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 make computers intelligent. Its applications are spread across all areas of artificial intelligence.

[0063] Deep learning is the core part of machine learning, which usually includes artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning. Among them, deep learning is a new research direction in the field of machine learning. That is, deep learning is a method in machine learning based on characterizing and learning data. Observations (such as an image) can be represented in a variety of ways, such as a vector of intensity values ​​for each pixel, or more abstractly represented as a series of edges, areas of specific shapes, etc. Using certain specific representation methods makes it easier to learn tasks from examples (for example, face recognition or facial expression recognition). The benefit of deep learning is to replace manual feature acquisition with efficient algorithms for unsupervised or semi-supervised feature learning and hierarchical feature extraction.

[0064] With the research and advancement of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, unmanned driving, automatic driving, drones, robots, smart medical care, smart customer service, etc. I believe that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0065] The solution provided in the embodiments of the present application involves technologies such as natural language processing and deep learning of artificial intelligence, which is now illustrated by the following embodiments.

[0066] The present application provides a video recall method, device, electronic device and storage medium.

[0067] Among them, the video recall device can be specifically integrated in a server or a terminal. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal can be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, etc., but is not limited to this. The terminal and the server can be directly or indirectly connected via wired or wireless communication, and this application does not limit this.

[0068] For example, see Figure 1aThe video recall device is integrated on a server, and the server can obtain the video title and query information corresponding to the sample video, wherein the query information includes a query statement and a query path. Then, the server detects the same keywords between the video title and the query statement, and obtains the same keywords. Then, the server constructs a heterogeneous graph for characterizing the association relationship between the sample video, the query statement, and the query path according to the query path and the same keywords. Finally, the server can train a preset video recall model based on the heterogeneous graph, the query statement, and the video title, and perform video recall through the trained video recall model. For example, the server responds to a video recall request triggered by a terminal, recalls a target video through the trained video recall model, and sends the recalled target video to the terminal, and the terminal can display the target video through a web page or a mini program.

[0069] The video recall method provided in the present application utilizes query paths and the same keywords to construct a heterogeneous graph for characterizing the association relationship between sample videos, query statements, and query paths. Then, a preset video recall model is trained based on the heterogeneous graph, query statements, and video titles. When the video is subsequently recalled using the trained video recall model, not only the semantic association between the query statement and the video title can be utilized, but also the association relationship between the video, the query statement, and the query path can be utilized, thereby improving the recall rate during video query.

[0070] It should be noted that the order of description of the following embodiments is not intended to limit the priority order of the embodiments.

[0071] A video recall method comprises: obtaining a video title and query information corresponding to a sample video, detecting the same keywords between the video title and the query statement, obtaining the same keywords, constructing a heterogeneous graph for characterizing the association relationship between the sample video, the query statement and the query path according to the query path and the same keywords, training a preset video recall model based on the heterogeneous graph, the query statement and the video title, and performing video recall using the trained video recall model.

[0072] See also Figure 1b , Figure 1b The following is a flow chart of the video recall method provided by this application. The specific flow of the video recall method can be as follows:

[0073] 101. Obtain a video title and query information corresponding to a sample video.

[0074] Video refers to various technologies that capture, record, process, store, transmit and reproduce a series of static images in the form of electrical signals. The development of network technology has also enabled video documentary clips to exist on the Internet in the form of streaming media and can be received and played by computers. Video data is a stream of images that changes over time and contains richer information and content that cannot be expressed by other media. Transmitting information in the form of video can express the content to be transmitted intuitively, vividly, realistically and efficiently.

[0075] The query information includes a query statement and a query path. The sample video may be a video played on a video website or a video inserted in a web page, for example, various film and television videos, live videos, program videos, and short videos, etc. Moreover, the sample video may be obtained from a video website or from a video database.

[0076] Query information refers to the query information generated by the user's query on the video. The query information includes query statements and query paths. The query statements are the statements used when querying the video. For example, the user enters "Zhang San XX TV series" to query video A, then "Zhang San XX TV series" is the query statement corresponding to video A. In addition, the user enters the search engine through the website to query video A, then the corresponding query path is: search engine-search result page-video A.

[0077] 102. Detect the same keywords between the video title and the query statement to obtain the same keywords.

[0078] Keywords are derived from the English word "keywords", which refers specifically to the words used by a single media when making and using indexes. Keyword search is one of the main methods of online search indexes. It refers to the specific names of products, services, companies, etc. that visitors want to know about.

[0079] The video title is a text describing the video content, such as "xx engineer made his own helicopter", and its corresponding keywords can be "xx engineer", "made-in-house" and "helicopter", and the query statement is the statement used by the user to query the video, such as "xx engineer, helicopter", "helicopter" or "xx engineer". Taking the query statement "xx engineer, helicopter" as an example, it can be seen that the query is expected to have the same keywords as the video title "xx engineer" and "helicopter", so "xx engineer" and "helicopter" are determined as the same keywords between the video title and the query statement, that is, the same keywords.

[0080] It should be noted that, in order to improve the recall ability of the subsequent model, keywords with high frequency but no practical significance, such as prepositions and modal particles, often appear in the video title and the query statement. Therefore, in some embodiments, these keywords with high frequency can be removed to improve the recall ability of the subsequent model. That is, the step of "detecting the same keywords between the video title and the query statement to obtain the same keywords" can specifically include:

[0081] (11) Count the frequency of each keyword appearing in the video title, and count the frequency of each keyword appearing in the query statement;

[0082] (12) removing keywords whose frequency is greater than a preset value in the video title to obtain a processed video title, and removing keywords whose frequency is greater than a preset value in the query statement to obtain a processed query statement;

[0083] (13) Detect the same keywords between the processed video title and the processed query statement and obtain the same keywords.

[0084] For example, N-Gram can be used to count the frequency of each keyword in the video title and the frequency of each keyword in the query statement. It should be noted that N-Gram is an algorithm based on a statistical language model. Its basic idea is to perform a sliding window operation of size N on the content of the text according to bytes, forming a sequence of byte fragments of length N. Each byte fragment is called a gram. The frequency of occurrence of all grams is counted and filtered according to a pre-set threshold to form a key gram list, which is the vector feature space of the text. Each gram in the list is a feature vector dimension. The model is based on the assumption that the appearance of the Nth word is only related to the previous N-1 words, but not to any other words. The probability of the whole sentence is the product of the probability of occurrence of each word. These probabilities can be obtained by directly counting the number of times N words appear at the same time in the corpus. Commonly used are Uni-Gram, Bi-Gram and Tri-Gram. In this application, the word frequencies corresponding to Uni-gram, Bi-gram and Thri-gram are 20, 2 and 2 respectively.

[0085] 103. According to the query path and the same keywords, a heterogeneous graph is constructed to characterize the relationship between sample videos, query statements and query paths.

[0086] A heterogeneous graph is a special information network that contains multiple types of objects or multiple types of connections. A heterogeneous graph, represented by G = [V, E], consists of an object set V and a connection set E. There can be more than one type of node and edge in a heterogeneous graph, thus allowing different types of nodes to have different dimensions of features or attributes.

[0087] In this application, the query path can also be called a meta-path. A meta-path is a composite relationship connecting two objects. It is a widely used semantic capture structure. For example, for a movie video, its meta-path can be movie video-actor-movie video. This meta-path describes the relationship between the two movie videos. For another example, a meta-path of movie video-director-movie video means that the two movies are directed by the same director. Heterogeneity is an intrinsic property of heterogeneous graphs, that is, various types of nodes and edges. For example, different types of nodes have different features, and their features may fall in different feature spaces.

[0088] It can be seen that the present application can construct a heterogeneous graph representing the association relationship between sample videos, query statements and query paths based on query paths and the same keywords, wherein the video title of the sample video, the query statement and the same keywords are nodes in the heterogeneous graph, such as Figure 1c shown.

[0089] 104. A preset video recall model is trained based on the heterogeneous graph, query statements and video titles, and video recall is performed using the trained video recall model.

[0090] For search products, the order of recall results often affects the user experience. Existing models often only use a single calculation method to calculate the correlation between the recall results and the query, or calculate the semantic correlation. The semantic model can calculate the semantic correlation between the query and the title to determine the order of the recall results. However, this ignores the fact that the videos recalled by this method are often different from the videos expected by the user. Therefore, this application uses semantic information and graph structure information to train the video recall model, so that when the subsequent video recall model is used, it can use semantic information and graph structure information for video recall, thereby improving the recall rate during video query.

[0091] It should be noted that since the heterogeneous graph, video title and query statement all carry a lot of high-dimensional information, it is not convenient to use this information during model training. Therefore, it is necessary to convert the high-dimensional information into low-dimensional information. That is, optionally, in some embodiments, the step of "training a preset video recall model based on the heterogeneous graph, query statement and video title" may specifically include:

[0092] (21) Performing graph embedding processing on the heterogeneous graph to obtain a node vector corresponding to each node in the heterogeneous graph;

[0093] (22) performing word segmentation processing on the query statement and the video title respectively to obtain a query word set corresponding to the query statement and a title word set corresponding to the video title;

[0094] (23) The preset video recall model is trained based on the node vector, the query word set and the title word set to obtain a trained video recall model.

[0095] Among them, graph embedding, also known as network representation learning, is to map the nodes in a network into a low-dimensional vector based on the characteristics of the network, so as to quantitatively measure the similarity between nodes for more convenient application.

[0096] Optionally, the DeepWalk algorithm can be used to perform graph embedding processing on the social network graph. DeepWalk is a graph structure data mining algorithm that combines the random walk and Word2vec algorithms. The algorithm can learn the hidden information of the network and can represent the nodes in the graph as a vector containing potential information.

[0097] Optionally, GCN (Graph Convolutional networks) can also be used to perform graph embedding processing on heterogeneous graphs. The graph convolutional neural network algorithm takes the global neighborhood of each node as input, iteratively aggregates the neighborhood embedding of the node by defining a convolution algorithm on the graph, and uses the embedding obtained in the previous iteration and its embedded function to obtain a new embedding. Aggregating embeddings of only local neighborhoods makes it scalable, and multiple iterations allow learning to embed a node to describe the global neighborhood.

[0098] Whether using the DeepWalk algorithm, the graph convolutional neural network algorithm, or other graph embedding methods for feature extraction, a node feature set consisting of node features of all nodes in the graph can be obtained. Feature extraction of each node in the heterogeneous graph obtains the node vector corresponding to each node in the heterogeneous graph. The node vector includes the attributes of the node itself and the topology or connection relationship between the nodes.

[0099] In addition, the language model can be used to perform word segmentation on the query sentence and the video title, and then the node vector, the query word set and the title word set are used to train the preset video recall model to obtain the trained video recall model.

[0100] It should be noted that before using the language model to perform word segmentation processing on the query statement and the video title, a reference dictionary can be pre-built, and the text lengths of the query statement and the video title are aligned based on the reference dictionary. Then, the language model is used to perform word segmentation processing on the query statement and the video title. That is, optionally, in some embodiments, it also includes:

[0101] (31) retaining keywords in the query statement and the video title whose frequency is less than a preset value, and constructing a reference dictionary based on the retained keywords;

[0102] (32) Align the text lengths of query statements and video titles according to the reference dictionary.

[0103] For example, specifically, retain the keywords with a frequency less than 2 in the query statement and the video title, and build a reference dictionary based on the retained keywords. Then, align the text lengths of the query statement and the video title based on the reference dictionary, and uniformly use a preset identifier (such as <unk>) represents a word that is not in the reference dictionary. The query sentence and / or video title whose text length is longer than the preset text length is truncated, and the text length shorter than the preset text length is padded with the preset mark. For example, if the preset text length is 5, the query sentence "Zhang San" and the video title "Zhang San takes a plane" are aligned, then the query sentence is adjusted to "Zhang San <unk> <unk> <unk>”.

[0104] Furthermore, this application designs an end-to-end video recall model that integrates semantic information and structural information. The architecture diagram of the model is as follows: Figure 1d As shown, the first sub-network uses a neural network to capture the semantic features of the query statement and the semantic features of the video title, and the second sub-network constructs the heterogeneous graph features corresponding to the query statement and the heterogeneous graph features corresponding to the video title by querying the path node vector and the same keywords. Then, the two parts of information are fused through a gate mechanism to output the predicted value of the correlation between the query statement and the video title. Finally, the video recall model is trained according to the actual value of the correlation between the pre-annotated query statement and the video title. That is, optionally, in some embodiments, the step of "training the preset video recall model based on the node vector, the query word set and the title word set to obtain the trained video recall model" may specifically include:

[0105] (41) constructing a first semantic feature corresponding to the query statement and a second semantic feature corresponding to the video title according to the first subnetwork, the query word set, and the title word set in the video recall model;

[0106] (42) constructing a first heterogeneous graph feature corresponding to the query statement and a second heterogeneous graph feature corresponding to the video title based on the second sub-network, the node vector and the same keyword in the video recall model;

[0107] (43) A preset video recall model is trained according to the first semantic feature, the second semantic feature, the first heterogeneous graph feature, and the second heterogeneous graph feature to obtain a trained video recall model.

[0108] With respect to step (31), the first sub-network may be used to obtain word embedding vectors corresponding to the query word set and the title word set, and then average pooling processing may be performed on the obtained word embedding vectors to finally obtain semantic features corresponding to the query statement and semantic features corresponding to the video title. That is, optionally, in some embodiments, the step of "constructing a first semantic feature corresponding to the query statement and a second semantic feature corresponding to the video title according to the first sub-network, the query word set and the title word set in the video recall model" may specifically include:

[0109] The first word vector is input into the first sub-network of the video recall model to obtain the first word embedding representation corresponding to each query word in the query word set, and the first word embedding representation is average pooled to obtain the first semantic feature corresponding to the query statement, and the second word vector is input into the second sub-network of the video recall model to obtain the second word embedding representation corresponding to each title word in the title word set, and the second word embedding representation is average pooled to obtain the second semantic feature corresponding to the video title.

[0110] For step (32), the heterogeneous graph and the same keyword are used as the input of the second sub-network. The heterogeneous graph attention neural network in the second sub-network can learn the feature representation of each node vector. Then, the vector features of the same keyword in the heterogeneous graph can be obtained. Finally, the node vectors are fused through the attention pooling mechanism to obtain the heterogeneous graph features of the query statement. Specifically, under each query path, the vector features of each node are learned by the attention mechanism. The attention mechanism is as follows: Figure 1e As shown. Then, for each node on the graph, different vector features based on the query path are learned, and different vector representations of the same node are aggregated into a vector representation containing structural features (i.e., heterogeneous graph features corresponding to the query statement) through the attention mechanism. For the video title, the heterogeneous graph features for describing the video title can be constructed based on the vector features corresponding to the same keywords, that is, the vector features corresponding to the same keywords are spliced ​​to obtain the second heterogeneous graph features corresponding to the video title. That is, optionally, in some embodiments, the step of "constructing the first heterogeneous graph features corresponding to the query statement and the second heterogeneous graph features corresponding to the video title based on the second sub-network, the heterogeneous graph and the same keywords in the video recall model" can specifically include:

[0111] (51) inputting the node vector into the second subnetwork in the video recall model to obtain the vector feature corresponding to each node in the heterogeneous graph;

[0112] (52) Obtaining vector features corresponding to the same keyword;

[0113] (53) constructing query features corresponding to each node vector under different query statements according to the association weights between nodes, and fusing the query features to obtain a first heterogeneous graph feature corresponding to the query statement;

[0114] (54) Based on the vector features corresponding to the same keywords, a second heterogeneous graph feature corresponding to the video title is constructed.

[0115] Among them, the vector feature carries the association weights between nodes. In some embodiments, the self-attention mechanism can be used to construct the first heterogeneous graph feature corresponding to the query statement. Specifically, first, the attention weights of each word in the query statement and other words k in the sentence are established, and then the attention weight vector is normalized and linearly weighted with all the information of the query statement (such as the position of each word, etc.). In this method, each word in the sentence can establish an association relationship with other words at any distance in the sentence. The goal is to be able to have a reference to the weights of different words (i.e., association weights) when generating words.

[0116] After obtaining the first semantic feature, the second semantic feature, the first heterogeneous graph feature, and the second heterogeneous graph feature, the video recall model can be trained based on the first semantic feature, the second semantic feature, the first heterogeneous graph feature, and the second heterogeneous graph feature. Specifically, the semantic feature and the graph structure feature of the query statement can be fused, and the semantic feature and the graph structure feature of the video title can be fused. Then, the video recall model is trained based on the fusion result. That is, optionally, in some embodiments, the step of "training the preset video recall model according to the first semantic feature, the second semantic feature, the first heterogeneous graph feature, and the second heterogeneous graph feature to obtain the trained video recall model" can specifically include:

[0117] (61) fusing the first semantic feature and the first heterogeneous graph feature to obtain a query feature corresponding to the query statement, and fusing the second semantic feature and the second heterogeneous graph feature to obtain a title feature corresponding to the video title;

[0118] (62) The preset video recall model is trained according to the query features and the title features to obtain a trained video recall model.

[0119] For details, please refer to Figure 1f The first semantic feature and the first heterogeneous graph feature are fused through a gating unit, and the second semantic feature and the second heterogeneous graph feature are fused. The gating unit can adaptively adjust the proportion of semantic information and graph structure information. If the similarity between the query statement and the video title is high, the model will pay more attention to semantic information because the semantic matching model has a better advantage in calculating similarity. If the similarity between the query statement and the video title is low, the model will pay more attention to structural information and is more inclined to find the target video through the nodes in the heterogeneous graph.

[0120] It can be understood that after obtaining the trained video recall model, when a video search operation is received, a set of videos to be recalled, a video search statement corresponding to the video search operation, and a video search path are obtained, wherein the set of videos to be recalled includes at least one video to be recalled, and the video to be recalled corresponds to a video title to be recalled. Then, the text similarity between the video title to be recalled and the video search statement can be calculated, and the video recall is performed based on the text similarity through the trained video recall model. For example, when the text similarity is less than 50%, the target video is searched through the nodes in the heterogeneous graph; when the text similarity is greater than or equal to 50%, the target video is recalled according to semantic features.

[0121] After obtaining the video title and query information corresponding to the sample video, the present application detects the same keywords between the video title and the query statement to obtain the same keywords, and then, based on the query path and the same keywords, constructs a heterogeneous graph for characterizing the association relationship between the sample video, the query statement, and the query path. Finally, the preset video recall model is trained based on the heterogeneous graph, the query statement, and the video title, and the video is recalled through the trained video recall model. The video recall method provided by the present application uses the query path and the same keywords to construct a heterogeneous graph for characterizing the association relationship between the sample video, the query statement, and the query path. Then, the preset video recall model is trained based on the heterogeneous graph, the query statement, and the video title. When the video is subsequently recalled through the trained video recall model, not only the semantic association between the query statement and the video title can be used, but also the association between the video, the query statement, and the query path can be used, thereby improving the recall rate during video query.

[0122] The method described in the embodiment will be further described in detail below with examples.

[0123] In this embodiment, the video recall device is specifically integrated into a server as an example for description.

[0124] See also Figure 2 , a video recall method, the specific process can be as follows:

[0125] 201. The server obtains a video title and query information corresponding to a sample video.

[0126] The sample video may be a video played on a video website or a video inserted in a web page, etc. For example, it may be various film and television videos, live videos, program videos, short videos, etc., and the sample video may be obtained from a video website or from a video database.

[0127] The query information refers to the query information generated when a user queries the video. The query information includes a query statement and a query path. The query statement is the statement used when querying the video.

[0128] 202. The server detects the same keywords between the video title and the query statement and obtains the same keywords.

[0129] Keywords are derived from the English word "keywords", which refers specifically to the words used by a single media when making and using indexes. Keyword search is one of the main methods of online search indexes. It refers to the specific names of products, services, companies, etc. that visitors want to know about.

[0130] For example, specifically, the server can count the frequency of occurrence of each keyword in the video title, and count the frequency of occurrence of each keyword in the query statement, and remove the keywords with a frequency greater than a preset value in the video title to obtain a processed video title, and remove the keywords with a frequency greater than a preset value in the query statement to obtain a processed query statement. Finally, the server detects the same keywords between the processed video title and the processed query statement to obtain the same keywords.

[0131] 203. The server constructs a heterogeneous graph for representing the relationship between the sample video, the query statement, and the query path according to the query path and the same keyword.

[0132] 204. The server trains a preset video recall model based on the heterogeneous graph, the query statement and the video title, and performs video recall using the trained video recall model.

[0133] Since heterogeneous graphs, video titles and query statements all carry a lot of high-dimensional information, it is not convenient to use this information during model training. Therefore, it is necessary to convert high-dimensional information into low-dimensional information. Specifically, the server performs graph embedding processing on the heterogeneous graph to obtain a node vector corresponding to each node in the heterogeneous graph. Then, the server can perform word segmentation processing on the query statement and the video title respectively to obtain a query word set corresponding to the query statement and a title word set corresponding to the video title. Finally, the server can train a preset video recall model based on the node vector, the query word set and the title word set to obtain a trained video recall model. After obtaining the trained video recall model, when the server receives a video search operation, the server obtains the video set to be recalled, the video search statement corresponding to the video search operation, and the video search path. Then, the server can calculate the text similarity between the video title to be recalled and the video search statement, and perform video recall based on the text similarity through the trained video recall model.

[0134] After the server of the present application obtains the video title and query information corresponding to the sample video, the server detects the same keywords between the video title and the query statement to obtain the same keywords. Then, the server constructs a heterogeneous graph for characterizing the association relationship between the sample video, the query statement, and the query path based on the query path and the same keywords. Finally, the server trains a preset video recall model based on the heterogeneous graph, the query statement, and the video title, and performs video recall through the trained video recall model. The video recall method provided by the present application uses the query path and the same keywords to construct a heterogeneous graph for characterizing the association relationship between the sample video, the query statement, and the query path. Then, the preset video recall model is trained based on the heterogeneous graph, the query statement, and the video title. When the video is subsequently recalled through the trained video recall model, not only the semantic association between the query statement and the video title can be used, but also the association between the video, the query statement, and the query path can be used, thereby improving the recall rate during video query.

[0135] In order to better implement the video recall method of the present application, the present application also provides a video recall device based on the above (recall device for short). The meanings of the terms are the same as those in the above video recall method, and the specific implementation details can refer to the description in the method embodiment.

[0136] See also Figure 3 , Figure 3 The schematic diagram of the structure of the video recall device provided in the present application, wherein the recall device may include an acquisition module 301, a detection module 302, a construction module 303, a training module 304 and a recall module 305, which may be specifically as follows:

[0137] The acquisition module 301 is used to acquire the video title and query information corresponding to the sample video.

[0138] The query information includes the query statement and the query path. The sample video can be a video played on a video website or a video inserted in a web page, for example, various film and television videos, live videos, program videos, short videos, etc., and the sample video can be obtained from a video website or a video database.

[0139] Query information refers to the query information generated by the user's query of the video. The query information includes the query statement and the query path. The query statement is the statement used when querying the video.

[0140] The detection module 302 is used to detect the same keywords between the video title and the query sentence to obtain the same keywords.

[0141] A construction module 303 is used to construct a heterogeneous graph for representing the association relationship between the sample video, the query statement and the query path according to the query path and the same keyword;

[0142] In this application, the query path can also be called a meta-path. A meta-path is a compound relationship connecting two objects. It is a widely used semantic capture structure. For example, for a movie video, its meta-path can be movie video-actor-movie video. This meta-path describes the relationship between the two movie videos. For another example, a meta-path of movie video-director-movie video means that the two movies are directed by the same director. Heterogeneity is an intrinsic property of heterogeneous graphs, that is, various types of nodes and edges. For example, different types of nodes have different features, and their features may fall in different feature spaces.

[0143] The training module 304 is used to train the preset video recall model based on the heterogeneous graph, the query statement and the video title.

[0144] Since the heterogeneous graph, video title and query statement all carry a lot of high-dimensional information, it is not convenient to use this information during model training. Therefore, it is necessary to convert the high-dimensional information into low-dimensional information. Specifically, the training module 304 performs graph embedding processing on the heterogeneous graph to obtain the node vector corresponding to each node in the heterogeneous graph. Then, the training module 304 can perform word segmentation processing on the query statement and the video title respectively to obtain the query word set corresponding to the query statement and the title word set corresponding to the video title. Finally, the training module 304 can train the preset video recall model based on the node vector, the query word set and the title word set to obtain the trained video recall model.

[0145] Optionally, in some embodiments, the training module 304 may specifically include:

[0146] A processing unit, used for performing graph embedding processing on the heterogeneous graph to obtain a node vector corresponding to each node in the heterogeneous graph;

[0147] A word segmentation unit is used to perform word segmentation processing on the query sentence and the video title respectively to obtain a query word set corresponding to the query sentence and a title word set corresponding to the video title;

[0148] The training unit is used to train a preset video recall model based on the node vector, the query word set and the title word set to obtain a trained video recall model.

[0149] Optionally, in some embodiments, the training unit may specifically include:

[0150] A first construction subunit is used to construct a first semantic feature corresponding to the query statement and a second semantic feature corresponding to the video title according to the first subnetwork, the query word set and the title word set in the video recall model;

[0151] A second construction subunit is used to construct a first heterogeneous graph feature corresponding to the query statement and a second heterogeneous graph feature corresponding to the video title based on the second subnetwork, the node vector and the same keyword in the video recall model;

[0152] The training subunit is used to train a preset video recall model according to the first semantic feature, the second semantic feature, the first heterogeneous graph feature and the second heterogeneous graph feature to obtain a trained video recall model.

[0153] Optionally, in some embodiments, the second construction subunit can be specifically used to: input the node vector into the second subnetwork in the video recall model to obtain the vector features corresponding to each node in the heterogeneous graph, the vector features carrying the association weights between nodes; obtain the vector features corresponding to the same keyword; according to the association weights between nodes, construct the query features corresponding to each node vector under different query statements, and fuse the query features to obtain the first heterogeneous graph features corresponding to the query statements; based on the vector features corresponding to the same keyword, construct the second heterogeneous graph features corresponding to the video title.

[0154] Optionally, in some embodiments, the first construction subunit can be specifically used to: input the first word vector into the first subnetwork of the video recall model to obtain a first word embedding representation corresponding to each query word in the query word set, and perform average pooling processing on the first word embedding representation to obtain a first semantic feature corresponding to the query statement; and input the second word vector into the second subnetwork of the video recall model to obtain a second word embedding representation corresponding to each title word in the title word set, and perform average pooling processing on the second word embedding representation to obtain a second semantic feature corresponding to the video title.

[0155] Optionally, in some embodiments, the training subunit can be specifically used to: fuse the first semantic feature and the first heterogeneous graph feature to obtain the query feature corresponding to the query statement, and fuse the second semantic feature and the second heterogeneous graph feature to obtain the title feature corresponding to the video title, and train the preset video recall model according to the query feature and the title feature to obtain the trained video recall model.

[0156] Optionally, in some embodiments, a construction unit is also included, which can be specifically used to: retain keywords in query statements and video titles whose frequency is less than a preset value, and construct a reference dictionary based on the retained keywords, and align the text lengths of query statements and video titles based on the reference dictionary.

[0157] The recall module 305 is used to perform video recall using the trained video recall model.

[0158] After obtaining the trained video recall model, when the recall module 305 receives a video search operation, it obtains a set of videos to be recalled, a video search statement corresponding to the video search operation, and a video search path. Then, the recall module 305 can calculate the text similarity between the title of the video to be recalled and the video search statement, and perform video recall based on the text similarity through the trained video recall model. That is, optionally, in some embodiments, the recall module 305 can be specifically used for: when receiving a video search operation, obtaining a set of videos to be recalled, a video search statement corresponding to the video search operation, and a video search path, the set of videos to be recalled includes at least one video to be recalled, and the video to be recalled corresponds to a video title to be recalled; calculating the text similarity between the title of the video to be recalled and the video search statement, and performing video recall based on the text similarity through the trained video recall model.

[0159] After the acquisition module 301 of the present application acquires the video title and query information corresponding to the sample video, the detection module 302 detects the same keywords between the video title and the query statement to obtain the same keywords. Then, the construction module 303 constructs a heterogeneous graph for characterizing the association relationship between the sample video, the query statement and the query path according to the query path and the same keywords. Finally, the training module 304 trains the preset video recall model based on the heterogeneous graph, the query statement and the video title, and the recall module 305 performs video recall through the trained video recall model. The video recall method provided by the present application uses the query path and the same keywords to construct a heterogeneous graph for characterizing the association relationship between the sample video, the query statement and the query path. Then, the preset video recall model is trained based on the heterogeneous graph, the query statement and the video title. When the video is subsequently recalled through the trained video recall model, not only the semantic association between the query statement and the video title can be used, but also the association between the video, the query statement and the query path can be used, thereby improving the recall rate during video query.

[0160] In addition, the present application also provides an electronic device, such as Figure 4 As shown, it shows a schematic diagram of the structure of the electronic device involved in this application, specifically:

[0161] The electronic device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art will appreciate that Figure 4 The electronic device structure shown in the figure does not constitute a limitation on the electronic device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0162] The processor 401 is the control center of the electronic device, and uses various interfaces and lines to connect various parts of the entire electronic device. By running or executing software programs and / or modules stored in the memory 402, and calling data stored in the memory 402, the processor 401 performs various functions of the electronic device and processes data, thereby monitoring the electronic device as a whole. Optionally, the processor 401 may include one or more processing cores; preferably, the processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 401.

[0163] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 402 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices. Accordingly, the memory 402 may also include a memory controller to provide the processor 401 with access to the memory 402.

[0164] The electronic device also includes a power supply 403 for supplying power to each component. Preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, so as to manage charging, discharging, power consumption and other functions through the power management system. The power supply 403 can also include one or more DC or AC power supplies, recharging systems, power failure detection circuits, power converters or inverters, power status indicators and other arbitrary components.

[0165] The electronic device may further include an input unit 404, which may be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control.

[0166] Although not shown, the electronic device may further include a display unit, etc., which will not be described in detail herein. Specifically in this embodiment, the processor 401 in the electronic device will load the executable files corresponding to the processes of one or more application programs into the memory 402 according to the following instructions, and the processor 401 will run the application programs stored in the memory 402, thereby realizing various functions, as follows:

[0167] Obtain the video title and query information corresponding to the sample video, detect the same keywords between the video title and the query statement, obtain the same keywords, and construct a heterogeneous graph for characterizing the relationship between the sample video, the query statement, and the query path based on the query path and the same keywords. Train the preset video recall model based on the heterogeneous graph, query statement, and video title, and perform video recall using the trained video recall model.

[0168] The specific implementation of the above operations can be found in the previous embodiments, which will not be described in detail here.

[0169] After obtaining the video title and query information corresponding to the sample video, the present application detects the same keywords between the video title and the query statement to obtain the same keywords, and then, based on the query path and the same keywords, constructs a heterogeneous graph for characterizing the association relationship between the sample video, the query statement, and the query path. Finally, the preset video recall model is trained based on the heterogeneous graph, the query statement, and the video title, and the video is recalled through the trained video recall model. The video recall method provided by the present application uses the query path and the same keywords to construct a heterogeneous graph for characterizing the association relationship between the sample video, the query statement, and the query path. Then, the preset video recall model is trained based on the heterogeneous graph, the query statement, and the video title. When the video is subsequently recalled through the trained video recall model, not only the semantic association between the query statement and the video title can be used, but also the association between the video, the query statement, and the query path can be used, thereby improving the recall rate during video query.

[0170] A person of ordinary skill in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be completed by instructions, or by controlling related hardware through instructions. The instructions may be stored in a computer-readable storage medium and loaded and executed by a processor.

[0171] To this end, the present application provides a storage medium in which a plurality of instructions are stored, and the instructions can be loaded by a processor to execute the steps in any of the video recall methods provided in the present application. For example, the instructions can execute the following steps:

[0172] Obtain the video title and query information corresponding to the sample video, detect the same keywords between the video title and the query statement, obtain the same keywords, and construct a heterogeneous graph for characterizing the relationship between the sample video, the query statement, and the query path based on the query path and the same keywords. Train the preset video recall model based on the heterogeneous graph, query statement, and video title, and perform video recall using the trained video recall model.

[0173] The specific implementation of the above operations can be found in the previous embodiments, which will not be described in detail here.

[0174] The storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0175] Since the instructions stored in the storage medium can execute the steps in any video recall provided in the present application, the beneficial effects that can be achieved by any video recall method provided in the present application can be achieved. Please refer to the previous embodiments for details and will not be repeated here.

[0176] The above is a detailed introduction to a video recall method, device, electronic device and storage medium provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for technical personnel in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.< / unk> < / unk> < / unk> < / unk>

Claims

1. A video recall method, It is characterized in that include: Obtaining a video title and query information corresponding to a sample video, wherein the query information includes a query statement; Detecting the same keywords between the video title and the query statement to obtain the same keywords; According to the video title, the query statement and the same keyword, a heterogeneous graph is constructed for characterizing the association relationship between the sample video and the query statement, wherein the video title, the query statement and the same keyword are nodes in the heterogeneous graph, and the video title and the query statement are associated through the same keyword; A preset video recall model is trained based on the heterogeneous graph, query statements and video titles, and video recall is performed using the trained video recall model.

2. The method according to claim 1, It is characterized in that The training of the preset video recall model based on the heterogeneous graph, the query statement and the video title includes: Performing graph embedding processing on the heterogeneous graph to obtain a node vector corresponding to each node in the heterogeneous graph; Perform word segmentation processing on the query statement and the video title respectively to obtain a query word set corresponding to the query statement and a title word set corresponding to the video title; The preset video recall model is trained based on the node vector, the query word set and the title word set to obtain a trained video recall model.

3. The method according to claim 2, It is characterized in that The step of training a preset video recall model based on the node vector, the query word set and the title word set to obtain a trained video recall model includes: Constructing a first semantic feature corresponding to the query statement and a second semantic feature corresponding to the video title according to the first subnetwork, the query word set and the title word set in the video recall model; Based on the second subnetwork in the video recall model, the node vector and the same keyword, construct a first heterogeneous graph feature corresponding to the query statement and a second heterogeneous graph feature corresponding to the video title; A preset video recall model is trained according to the first semantic feature, the second semantic feature, the first heterogeneous graph feature, and the second heterogeneous graph feature to obtain a trained video recall model.

4. The method according to claim 3, It is characterized in that The step of constructing a first heterogeneous graph feature corresponding to the query statement and a second heterogeneous graph feature corresponding to the video title based on the second subnetwork in the video recall model, the node vector, and the same keyword includes: Inputting the node vector into the second subnetwork in the video recall model to obtain a vector feature corresponding to each node in the heterogeneous graph, wherein the vector feature carries an association weight between nodes; Obtaining vector features corresponding to the same keyword; According to the association weights between nodes, construct query features corresponding to each node vector under different query statements, and fuse the query features to obtain a first heterogeneous graph feature corresponding to the query statement; Based on the vector features corresponding to the same keyword, a second heterogeneous graph feature corresponding to the video title is constructed.

5. The method according to claim 3, It is characterized in that The step of constructing a first semantic feature corresponding to the query statement and a second semantic feature corresponding to the video title according to the first subnetwork, the query word set and the title word set in the video recall model includes: Inputting each query word in the query word set into the first subnetwork of the video recall model to obtain a first word embedding representation corresponding to each query word in the query word set, and performing average pooling processing on the first word embedding representation to obtain a first semantic feature corresponding to the query statement, and; Each title word in the title word set is input into the first subnetwork of the video recall model to obtain a second word embedding representation corresponding to each title word in the title word set, and average pooling is performed on the second word embedding representation to obtain a second semantic feature corresponding to the video title.

6. The method according to claim 3, It is characterized in that The step of training a preset video recall model according to the first semantic feature, the second semantic feature, the first heterogeneous graph feature, and the second heterogeneous graph feature to obtain a trained video recall model includes: fusing the first semantic feature and the first heterogeneous graph feature to obtain a query feature corresponding to the query statement, and; fusing the second semantic feature and the second heterogeneous graph feature to obtain a title feature corresponding to the video title; The preset video recall model is trained according to the query features and the title features to obtain a trained video recall model.

7. The method according to claim 2, It is characterized in that Before the query statement and the video title are segmented to obtain a query word set corresponding to the query statement and a title word set corresponding to the video title, the method further includes: retaining keywords in the query statement and the video title whose frequency is less than a preset value, and constructing a reference dictionary based on the retained keywords; The text lengths of the query sentence and the video title are aligned according to the reference dictionary.

8. The method according to any one of claims 1 to 7, It is characterized in that The video recall is performed using the trained video recall model, including: When a video search operation is received, a video set to be recalled and a video search statement corresponding to the video search operation are obtained, wherein the video set to be recalled includes at least one video to be recalled, and the video to be recalled corresponds to a video title to be recalled; The text similarity between the video title to be recalled and the video search sentence is calculated, and the video is recalled through the trained video recall model based on the text similarity.

9. The method according to any one of claims 1 to 7, It is characterized in that The detecting the same keywords between the video title and the query statement to obtain the same keywords includes: Counting the frequency of each keyword appearing in the video title, and counting the frequency of each keyword appearing in the query statement; removing keywords with a frequency greater than a preset value from the video title to obtain a processed video title, and; Removing keywords whose frequency is greater than a preset value from the query statement to obtain a processed query statement; The same keywords between the processed video title and the processed query statement are detected to obtain the same keywords.

10. A video recall device, It is characterized in that include: An acquisition module, used to acquire a video title and query information corresponding to a sample video, wherein the query information includes a query statement; A detection module, used to detect the same keywords between the video title and the query sentence, and obtain the same keywords; A construction module, used to construct a heterogeneous graph for characterizing the association relationship between the sample video and the query statement according to the video title, the query statement and the same keyword, wherein the video title, the query statement and the same keyword are nodes in the heterogeneous graph, and the video title and the query statement are associated through the same keyword; A training module, used for training a preset video recall model based on the heterogeneous graph, query statements and video titles; The recall module is used to perform video recall using the trained video recall model.

11. The device according to claim 10, It is characterized in that The training module includes: A processing unit, configured to perform graph embedding processing on the heterogeneous graph to obtain a node vector corresponding to each node in the heterogeneous graph; A word segmentation unit is used to perform word segmentation processing on the query statement and the video title respectively to obtain a query word set corresponding to the query statement and a title word set corresponding to the video title; The training unit is used to train a preset video recall model based on the node vector, the query word set and the title word set to obtain a trained video recall model.

12. The device according to claim 11, It is characterized in that The training unit comprises: A first construction subunit is used to construct a first semantic feature corresponding to the query statement and a second semantic feature corresponding to the video title according to the first subnetwork, the query word set and the title word set in the video recall model; A second construction subunit is used to construct a first heterogeneous graph feature corresponding to the query statement and a second heterogeneous graph feature corresponding to the video title based on the second subnetwork in the video recall model, the node vector and the same keyword; The training subunit is used to train a preset video recall model according to the first semantic feature, the second semantic feature, the first heterogeneous graph feature and the second heterogeneous graph feature to obtain a trained video recall model.

13. The device according to claim 12, It is characterized in that The second construction subunit is specifically used for: Inputting the node vector into the second subnetwork in the video recall model to obtain a vector feature corresponding to each node in the heterogeneous graph, wherein the vector feature carries an association weight between nodes; Obtaining vector features corresponding to the same keyword; According to the association weights between nodes, construct query features corresponding to each node vector under different query statements, and fuse the query features to obtain a first heterogeneous graph feature corresponding to the query statement; Based on the vector features corresponding to the same keyword, a second heterogeneous graph feature corresponding to the video title is constructed.

14. An electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, in, When the processor executes the program, the steps of the video recall method as described in any one of claims 1 to 8 are implemented.

15. A computer-readable storage medium, It is characterized in that A computer program is stored thereon, wherein when the computer program is executed by a processor, the steps of the video recall method according to any one of claims 1 to 8 are implemented.