CN110166802B

CN110166802B - Bullet screen processing method and device and storage medium

Info

Publication number: CN110166802B
Application number: CN201910372154.6A
Authority: CN
Inventors: 余自强
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2019-05-06
Filing date: 2019-05-06
Publication date: 2022-11-01
Anticipated expiration: 2039-05-06
Also published as: CN110166802A

Abstract

The invention relates to the technical field of networks, and discloses a bullet screen processing method, a bullet screen processing device and a storage medium, wherein the method comprises the following steps: acquiring a bullet screen text, and converting the bullet screen text into corresponding word vectors; inputting a word vector corresponding to the bullet screen text into a pre-trained pivot recognition model to obtain a pivot probability value corresponding to the bullet screen text, wherein the pivot recognition model is obtained based on pivot sample training, and the pivot probability value is used for expressing the probability that the bullet screen text is a pivot bullet screen; and determining the bullet screen text with the pivot probability value larger than the set threshold value as a pivot bullet screen so that the client displays the pivot bullet screen according to the set display mode when receiving the pivot bullet screen. According to the technical scheme provided by the embodiment of the invention, a large amount of bullet screen texts can be automatically and quickly judged through the pivot identification model, the pivot bullet screen in the bullet screen texts is identified, the problem that the pivot bullet screen cannot be effectively identified in the prior art is solved, and the identification accuracy is high.

Description

Bullet screen processing method and device and storage medium

Technical Field

The present invention relates to the field of network technologies, and in particular, to a bullet screen processing method, apparatus, and storage medium.

Background

With the development of internet technology, people increasingly like to open barrages when watching dramas or movies on the internet, and enjoy the fun brought by the interactive experience. However, the quality of the barrage is not guaranteed more and more, and besides some barrages with low quality or offensive speech, the barrage showing the scenario often appears, which may destroy the viewing experience of the user on the whole video, and especially when watching a suspense type film, the penetrating barrage may greatly reduce the viewing experience. Although a video website or a video player provides a function of closing a barrage, some barrages of drama are already displayed to users when a video is just played, and at the moment, the users have no time to open the function of closing the barrage, so that many users directly select to close the barrage before watching the video, or perform speech attack on the barrages through the barrages, so that the environment of the whole barrage becomes worse, and the user experience is seriously influenced. In addition, there is also a need to interact through the barrage when reading novels, cartoons, listening to story audio, and thus there is also a problem of telepresence of the barrage. Therefore, a method for effectively identifying the drama-through bullet screen is needed.

Disclosure of Invention

The embodiment of the invention provides a bullet screen processing method, a bullet screen processing device, a terminal, a server and a storage medium, and aims to solve the problem that a dramatic bullet screen cannot be identified in the prior art.

In a first aspect, an embodiment of the present invention provides a bullet screen processing method, including:

acquiring a bullet screen text, and converting the bullet screen text into a corresponding word vector matrix;

inputting a word vector matrix corresponding to the bullet screen text into a pre-trained pivot recognition model to obtain a pivot probability value corresponding to the bullet screen text, wherein the pivot recognition model is obtained based on pivot bullet screen sample training, and the pivot probability value is used for representing the probability that the bullet screen text is a pivot bullet screen;

and determining the bullet screen text with the pivot probability value larger than a set threshold value as a pivot bullet screen so that the client displays the pivot bullet screen according to a set display mode when receiving the pivot bullet screen.

Optionally, the method further comprises:

responding to a bullet screen acquisition request of a client, and sending a bullet screen file to the client, wherein the bullet screen file comprises: a spoiler pop-up screen and a non-spoiler pop-up screen.

Optionally, the method further comprises:

responding to a barrage acquisition request of a client, and sending a barrage file to the client, wherein the barrage file comprises: the client side determines the display effect of the full-play bullet screen according to the full-play probability value corresponding to the full-play bullet screen.

Optionally, the pivot recognition model is trained by:

obtaining a pivot bullet screen sample set, wherein each pivot bullet screen sample comprises: the bullet screen text is marked with a pivot mark, and the pivot mark is used for representing whether the bullet screen text is a pivot bullet screen or not;

for each full play bullet screen sample in the full play bullet screen sample set, updating the vocabulary belonging to a preset type in the full play bullet screen sample into a replacement identifier corresponding to the preset type;

for each perspective pop-up screen sample in the perspective pop-up screen sample set, converting the updated perspective pop-up screen sample into a corresponding word vector matrix;

and training the pivot recognition model based on the word vector matrix and the pivot identification corresponding to the pivot samples in the pivot bullet screen sample set.

Optionally, the converting the updated drama-popup screen sample into a corresponding word vector matrix specifically includes:

converting each word contained in the updated perspective bullet screen sample into a corresponding word vector, wherein one replacement identifier in the updated perspective bullet screen sample corresponds to one word, and obtaining a word vector matrix corresponding to the perspective bullet screen sample based on the word vector corresponding to each word; alternatively, the first and second electrodes may be,

the updated full play bullet screen sample is subjected to word segmentation processing to obtain a plurality of words, one replacement identifier in the updated full play bullet screen corresponds to one word, the obtained words are respectively converted into corresponding word vectors, and a word vector matrix corresponding to the full play bullet screen sample is obtained based on the word vectors corresponding to the words.

Optionally, the converting the bullet screen text into a corresponding word vector matrix specifically includes:

updating the vocabulary belonging to the preset type in the bullet screen text into a replacement identifier corresponding to the preset type;

and converting the updated bullet screen text into a corresponding word vector matrix.

Optionally, the converting the updated barrage text into a corresponding word vector matrix specifically includes:

converting each word contained in the updated bullet screen text into a corresponding word vector, wherein one replacement identifier in the updated bullet screen text corresponds to one word, and a word vector matrix corresponding to the bullet screen text is obtained based on the word vector corresponding to each word; alternatively, the first and second electrodes may be,

carrying out word segmentation processing on the updated bullet screen text to obtain a plurality of vocabularies, wherein one replacement identifier in the updated bullet screen text corresponds to one vocabulary, converting the obtained vocabularies into corresponding word vectors respectively, and obtaining a word vector matrix corresponding to the bullet screen text based on the word vectors corresponding to the vocabularies.

Optionally, each spoilt pop-up screen sample comprises: and the bullet screen text marked with the drama identification is manually marked or the bullet screen text marked with the drama identification is uploaded by the client.

Optionally, the obtaining of the drama-popup screen sample set specifically includes:

and receiving the bullet screen text which is marked with the pivot identification and uploaded by the client, and adding the bullet screen text into the pivot bullet screen sample set.

Optionally, the obtaining of the spoiled barrage sample set specifically includes:

acquiring a drama telesopic barrage sample set corresponding to any scenario type, wherein the drama telesopic barrage sample in the drama telesopic barrage sample set corresponding to any scenario type is a barrage text generated in the playing process of multimedia resources belonging to any scenario type;

the training of the spoiler identification model based on the word vector matrix and the spoiler identification corresponding to the spoiler bullet screen samples in the spoiler bullet screen sample set specifically comprises:

and training a pivot recognition model corresponding to any plot type based on the word vector matrix and the pivot identification corresponding to the pivot samples in the pivot bullet screen sample set.

Optionally, the method further comprises: determining the plot type of the multimedia resource to which the barrage text belongs;

the inputting of the word vector matrix corresponding to the bullet screen text into a pre-trained pivot recognition model specifically includes:

and inputting the word vector matrix corresponding to the bullet screen text into a perspective identification model corresponding to the plot type of the multimedia resource to which the bullet screen text belongs.

Optionally, the spoiler identification model includes an input layer, a first preset number of convolution pooling layers, a second preset number of full-connected layers, and an output layer, which are connected in sequence, where each convolution pooling layer includes a convolution layer and a pooling layer, which are connected in sequence;

the method for obtaining the pivot probability value corresponding to the bullet screen text comprises the following steps of inputting a word vector matrix corresponding to the bullet screen text into a pre-trained pivot recognition model, and specifically comprises the following steps:

inputting a word vector matrix corresponding to the bullet screen text into an input layer of the pivot recognition model;

inputting the word vector matrix passing through the input layer into a first preset number of convolution pooling layers which are sequentially connected so as to perform convolution pooling processing on the word vector matrix for a first preset number of times, and obtaining a convolution pooling processing result;

inputting the convolution pooling processing result into a second preset number of full-connection layers which are connected in sequence to obtain a full-connection processing result;

and inputting the full-connection processing result into the output layer to obtain a perspective probability value corresponding to the bullet screen text.

In a second aspect, an embodiment of the present invention provides a bullet screen processing method, including:

responding to a barrage display operation input through a control interface of a client, and sending a barrage acquisition request to a server;

receiving a bullet screen file returned by the server, wherein the bullet screen file comprises a pivot bullet screen and a non-pivot bullet screen;

and when the set display mode is to display the pivot bullet screen according to the appointed display effect, displaying the non-pivot bullet screen, and displaying the pivot bullet screen according to the appointed display effect.

Optionally, the bullet screen file further includes a pivot probability value corresponding to the pivot bullet screen, and the pivot probability value is determined based on a pivot identification model.

Correspondingly, the displaying the drama popup according to the specified display effect specifically includes:

according to the appointed display effect, obtaining a weakening display parameter corresponding to the perspective probability value of the perspective bullet screen, wherein the weakening display parameter is negatively related to the perspective probability value;

and performing weakening display on the dramatic popup according to the obtained weakening display parameters.

Optionally, the method further comprises:

and determining the display mode according to the bullet screen display operation.

Optionally, the method further comprises:

responding to a drama pop-up screen marking operation aiming at the displayed pop-up screen text, and marking the pop-up screen text as a drama pop-up screen;

and sending the bullet screen text marked as the drama full-blown bullet screen to the server so that the server adds the bullet screen text marked as the drama full-blown bullet screen to a drama full-blown bullet screen sample set for training the drama full-blown recognition model.

In a third aspect, an embodiment of the present invention provides a bullet screen processing apparatus, including:

the system comprises a preprocessing module, a word vector matrix and a display module, wherein the preprocessing module is used for acquiring a bullet screen text and converting the bullet screen text into a corresponding word vector matrix;

the recognition module is used for inputting a word vector matrix corresponding to the bullet screen text into a pre-trained pivot recognition model to obtain a pivot probability value corresponding to the bullet screen text, the pivot recognition model is obtained based on pivot bullet screen sample training, and the pivot probability value is used for representing the probability that the bullet screen text is a pivot bullet screen;

and the judging module is used for determining the bullet screen text with the pivot probability value larger than the set threshold value as the pivot bullet screen so that the client displays the pivot bullet screen according to the set display mode when receiving the pivot bullet screen.

Optionally, the system further comprises a response module, configured to: responding to a bullet screen acquisition request of a client, and sending a bullet screen file to the client, wherein the bullet screen file comprises: the pop-up screen and the non-pop-up screen, or the pop-up screen file comprises: the method comprises the steps of determining a full-play bullet screen, a non-full-play bullet screen and full-play probability values corresponding to the full-play bullet screen by the client, so that the client determines weakening display parameters when the full-play bullet screen is displayed according to the full-play probability values corresponding to the full-play bullet screen.

Optionally, a training module is further included for:

for each perspective bullet screen sample in the perspective bullet screen sample set, updating vocabularies belonging to a preset type in the perspective bullet screen samples into a replacement identifier corresponding to the preset type;

Optionally, each spoilt pop-up screen sample comprises: and the bullet screen text marked with the pivot identification is manually marked or the bullet screen text marked with the pivot identification is uploaded by the client.

Optionally, the training module is specifically configured to: converting each word contained in the updated perspective bullet screen sample into a corresponding word vector, wherein one replacement identifier in the updated perspective bullet screen sample corresponds to one word, and obtaining a word vector matrix corresponding to the perspective bullet screen sample based on the word vector corresponding to each word; or, the training module is specifically configured to: performing word segmentation on the updated perspective bullet screen sample to obtain a plurality of words, wherein one replacement identifier in the updated perspective bullet screen corresponds to one word, the obtained words are respectively converted into corresponding word vectors, and a word vector matrix corresponding to the perspective bullet screen sample is obtained based on the word vectors corresponding to the words.

Optionally, the preprocessing module is specifically configured to: updating the vocabulary belonging to the preset type in the bullet screen text into a replacement identifier corresponding to the preset type; and converting the updated bullet screen text into a corresponding word vector matrix.

Optionally, the preprocessing module is specifically configured to: converting each word contained in the updated bullet screen text into a corresponding word vector, wherein one replacement identifier in the updated bullet screen text corresponds to one word, and a word vector matrix corresponding to the bullet screen text is obtained based on the word vector corresponding to each word; or, the preprocessing module is specifically used for carrying out word segmentation on the updated bullet screen text to obtain a plurality of vocabularies, one replacement identifier in the updated bullet screen text corresponds to one vocabulary, and the obtained vocabularies are converted into corresponding word vectors respectively, based on the word vectors corresponding to the vocabularies, the word vector matrix corresponding to the bullet screen text is obtained.

Optionally, the training module is specifically configured to: and receiving the bullet screen text which is marked with the pivot identification and uploaded by the client, and adding the bullet screen text into the pivot bullet screen sample set.

Optionally, the training module is specifically configured to: acquiring a drama telesopic barrage sample set corresponding to any scenario type, wherein the drama telesopic barrage sample in the drama telesopic barrage sample set corresponding to any scenario type is a barrage text generated in the playing process of multimedia resources belonging to any scenario type; and training a pivot recognition model corresponding to any plot type based on the word vector matrix and the pivot identification corresponding to the pivot samples in the pivot bullet screen sample set.

Optionally, a type identification module is further included, configured to: determining the plot type of the multimedia resource to which the barrage text belongs;

correspondingly, the identification module is specifically configured to: and inputting the word vector matrix corresponding to the bullet screen text into a pivot recognition model corresponding to the plot type of the multimedia resource to which the bullet screen text belongs to obtain a pivot probability value corresponding to the bullet screen text.

Optionally, the pivot recognition model includes an input layer, a first preset number of convolution pooling layers, a second preset number of full-connected layers, and an output layer, which are connected in sequence, where each convolution pooling layer includes a convolution layer and a pooling layer, which are connected in sequence. Correspondingly, the identification module is specifically configured to: inputting a word vector matrix corresponding to the bullet screen text into an input layer of the pivot recognition model; inputting the word vector matrixes passing through the input layer into a first preset number of convolution pooling layers which are sequentially connected, so as to perform convolution pooling processing on the word vector matrixes for a first preset number of times, and obtain convolution pooling processing results; inputting the convolution pooling processing result into a second preset number of full-connection layers which are connected in sequence to obtain a full-connection processing result; and inputting the full-connection processing result into the output layer to obtain a perspective probability value corresponding to the bullet screen text.

In a fourth aspect, an embodiment of the present invention provides a bullet screen processing apparatus, including:

the bullet screen request module is used for responding to bullet screen display operation input through a control interface of the client and sending a bullet screen acquisition request to the server;

the bullet screen receiving module is used for receiving bullet screen files returned by the server, and the bullet screen files comprise a pivot bullet screen and a non-pivot bullet screen;

and the bullet screen display module is used for displaying the non-play-through bullet screen when the set display mode is used for displaying the play-through bullet screen according to the appointed display effect, and displaying the play-through bullet screen according to the appointed display effect.

Optionally, the pop-up screen file further includes a spoiler probability value corresponding to the spoiler pop-up screen, and the spoiler probability value is determined based on the spoiler identification model.

Correspondingly, the barrage display module is specifically configured to: according to the appointed display effect, obtaining a weakening display parameter corresponding to the perspective probability value of the perspective bullet screen, wherein the weakening display parameter is negatively related to the perspective probability value; and performing weakening display on the dramatic popup according to the obtained weakening display parameters.

Optionally, a mode determining module is further included, configured to: and determining the display mode according to the bullet screen display operation.

Optionally, the system further comprises a perspective barrage marking module, configured to: responding to a zooming-out bullet screen marking operation aiming at the displayed bullet screen text, and marking the bullet screen text as a zooming-out bullet screen; and sending the bullet screen text marked as the drama full-blown bullet screen to the server so that the server adds the bullet screen text marked as the drama full-blown bullet screen to a drama full-blown bullet screen sample set for training the drama full-blown recognition model.

In a fifth aspect, an embodiment of the present invention provides a server, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of any one of the methods in the first aspect when executing the computer program.

In a sixth aspect, an embodiment of the present invention provides a terminal, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the methods in the second aspect when executing the computer program.

In a seventh aspect, an embodiment of the present invention provides a computer-readable storage medium, on which computer program instructions are stored, which, when executed by a processor, implement the steps of any one of the above-mentioned methods.

According to the technical scheme provided by the embodiment of the invention, the pivot identification model is trained by utilizing the pivot sample marked with the pivot bullet screen, so that the pivot identification model can identify the pivot bullet screen, a large number of bullet screen texts can be quickly judged based on the pre-trained pivot identification model, the pivot bullet screen is identified, then, the identified pivot bullet screen is filtered or specially processed, particularly, the bullet screen texts sent by a user in real time can be quickly judged whether the pivot bullet screen is the pivot bullet screen or not, and then, the pivot bullet screen is displayed to other users in real time, so that the interaction instantaneity is enhanced, and the experience of the user in watching the network video is improved.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the embodiments of the present invention will be briefly described below, and it is obvious that the drawings described below are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings without creative efforts.

Fig. 1 is a schematic view of an application scenario of a bullet screen processing method according to an embodiment of the present invention;

fig. 2 is a flowchart illustrating a training method of a perspective recognition model according to an embodiment of the present invention;

fig. 3A is an operation diagram of a reporting drama pop-up screen by a client according to an embodiment of the present invention;

fig. 3B is an operation diagram of a reporting drama pop-up screen by a client according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a convolutional neural network;

fig. 5 is a schematic flowchart of a bullet screen processing method according to an embodiment of the present invention;

fig. 6 is a schematic flowchart of a bullet screen processing method according to an embodiment of the present invention;

fig. 7 is a schematic flowchart of a bullet screen processing method according to an embodiment of the present invention;

fig. 8A is an operation diagram of a setting presentation mode provided by a client according to an embodiment of the present invention;

fig. 8B is an operation diagram of a setting presentation mode provided by a client according to an embodiment of the present invention;

fig. 8C is an operation diagram of a setting presentation mode provided by the client according to an embodiment of the present invention;

fig. 9 is a schematic flowchart of a bullet screen processing method according to an embodiment of the present invention;

fig. 10 is a schematic structural diagram of a bullet screen processing device according to an embodiment of the present invention;

fig. 11 is a schematic structural diagram of a bullet screen processing device according to an embodiment of the present invention;

fig. 12 is a schematic structural diagram of a server according to an embodiment of the present invention;

fig. 13 is a schematic structural diagram of a terminal according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application.

For convenience of understanding, terms referred to in the embodiments of the present application are explained below:

barrage (barrage), refers to a commentary subtitle that pops up when watching a video over a network. In the process of watching videos, the user can watch the barrage published by other users and can also send the barrage, and the sent barrage can also be seen by other users, so that interaction is realized in the process of watching videos.

Deep learning, wherein the concept of deep learning is derived from the research of an artificial neural network, and a multilayer perceptron comprising a plurality of hidden layers is a deep learning structure. Deep learning forms a more abstract class or feature of high-level representation properties by combining low-level features to discover a distributed feature representation of the data. Deep learning is a new field in machine learning research, and the motivation is to establish and simulate a neural network for analyzing and learning of human brain, and to interpret data such as images, sounds, texts and the like by simulating the mechanism of human brain. Common deep learning models include: convolutional Neural Networks (CNN), cyclic Neural Networks (RNN), long Short-Term Memory Networks (LSTM), deep Neural Networks (DNN), deep Belief Networks (DBNs), and the like. There are two ways that data can propagate in a neural network, one along the path from input to output, called forward propagation, and the other from output back to input, called backward propagation. In the forward propagation process, input information is processed layer by layer through a neural network and transmitted to an output layer, errors between output values and expectations are described through a loss function, backward propagation is carried out, partial derivatives of the loss function to the weight of each neuron are calculated layer by layer, weight gradient data of the loss function to weight vectors are formed and serve as the basis for updating weight parameters, and training of the neural network is completed in the process of continuously updating the weight parameters.

The loss function (loss function) is a function that maps the value of a random event or its associated random variable to a non-negative real number to represent the "risk" or "loss" of the random event. In application, the loss function is usually associated with the optimization problem as a learning criterion, i.e. the model is solved and evaluated by minimizing the loss function. For example, in machine learning, a loss function is used for parameter estimation (parametric estimation) of a model, and a loss value obtained based on the loss function can be used to describe a difference degree between a predicted value and an actual value of the model. Common loss functions include a mean square error loss function, a Support Vector Machine (SVM) hinge loss function, a cross entropy loss function, and the like.

The modified linear unit (ReLU) is an activation function of the neural network, the expression capability of the ReLU is stronger than that of other activation functions, and the convergence rate of the model can be maintained in a stable state.

Word vector is the form of converting the text into a mathematical vector for subsequent processing by the model.

Unknown words (UNK), i.e. words that are not included in the word segmentation vocabulary or that have a low frequency of occurrence in the whole corpus but have to be segmented, mainly include various proper nouns (names of people, places, names of enterprises, etc.), abbreviations, newly added words, etc.

A named entity generally refers to an entity with a specific meaning or strong reference in the text, and generally includes a name of a person, a name of a place, a name of an organization, a date and time, a proper noun, and the like. The concept of named entities can be very broad, and any special piece of text that is needed by a business can be called a named entity.

Named Entity Recognition (NER), a basic task of natural language processing, aims to extract Named Entities from unstructured input text. The discriminant Model CRF is the current mainstream Model of the NER, and its objective function not only considers the input state feature function, but also includes the label transfer feature function.

Chinese word segmentation is the process of dividing a Chinese character sequence into several independent words, i.e. recombining continuous character sequences into word sequences according to a certain standard.

Stop Words (Stop Words) refer to the automatic filtering of some Words or phrases before or after processing natural language data (or text) in order to save storage space and improve search efficiency in information retrieval. Stop words are all manually input and are not automatically generated, and the generated stop words form a stop word list.

Word2vec, is a group of related models used to generate Word vectors. These models are shallow, two-layer neural networks that are trained to reconstruct linguistic word text. The network is represented by words and the input words in adjacent positions are guessed, and the order of the words is unimportant under the assumption of the bag-of-words model in word2 vec. After training is completed, the word2vec model can be used to map each word to a vector, which can be used to represent word-to-word relationships, and the vector is a hidden layer of the neural network.

one-hot encoding, also known as unique hot encoding or one-bit efficient encoding. The method is to use an N-bit status register to encode N states, each state having its own independent register bit and only one of which is active at any one time. The one-hot vector is a vector represented by one-hot coding.

Sigmoid, a relatively common function in machine learning, has an output between 0 and 1, and in the binary task, the output of Sigmoid is often used as the event probability.

A Client (Client), also called Client, refers to a program corresponding to a server and providing local services to clients. Except for some application programs which only run locally, the application programs are generally installed on common clients and need to be operated together with a server. After the internet has developed, the more common clients include web browsers used on the world wide web, email clients for receiving and sending emails, and client software for instant messaging. For this kind of application, a corresponding server and a corresponding service program are required in the network to provide corresponding services, such as database services, e-mail services, etc., so that a specific communication connection needs to be established between the client and the server to ensure the normal operation of the application program.

Any number of elements in the drawings are by way of example and not by way of limitation, and any nomenclature is used for distinction only and not by way of limitation.

In a specific practical process, the quality of the barrage is not guaranteed more and more, and besides some barrages with low quality or offensive speech, the barrage which reveals the scenario often appears, so that the viewing experience of the user on the whole video is probably destroyed, and particularly when the user watches suspense videos, the viewing experience can be greatly reduced by penetrating the barrage. In order to solve the problems caused by the play pop-up screen, some video websites add a manual screening function or a user reporting function, background personnel manually screen out the play pop-up screen, or the user actively reports the play pop-up screen to filter out the play pop-up screen, but the manual screening mode has the disadvantages of huge operation cost, low screening accuracy and easy screen missing. In addition, a method for identifying the full play barrage by matching keywords is provided, but the barrage often contains strong semantic information or approximate vocabularies, preset keywords cannot cover all situations, and the identification accuracy is low. Therefore, the existing method for identifying the drama-transparency barrage has high operation cost and low identification accuracy, and cannot effectively identify and filter the drama-transparency barrage.

Therefore, the inventor of the present application considers that a neural network is trained by using a pivot bullet screen sample marked with whether the pivot bullet screen is a pivot bullet screen, and a pivot identification model capable of identifying the pivot bullet screen is obtained. After acquiring the bullet screen text sent by the user, converting the bullet screen text into a corresponding word vector matrix, inputting the word vector matrix corresponding to the bullet screen text into a pre-trained pivot recognition model to obtain a pivot probability value corresponding to the bullet screen text, and determining the bullet screen text with the pivot probability value being larger than a set threshold value as a pivot bullet screen. Through the perspective identification model, a large amount of barrage texts can be rapidly judged, the perspective barrage is identified, then the identified perspective barrage is filtered or specially processed, particularly, the barrage texts sent by the user immediately can be rapidly judged whether to be the perspective barrage, and then the perspective barrage is displayed for other users immediately, so that the interaction instantaneity is enhanced, and the experience of watching network videos by the user is promoted.

After introducing the design concept of the embodiment of the present application, some simple descriptions are provided below for application scenarios to which the technical solution of the embodiment of the present application can be applied, and it should be noted that the application scenarios described below are only used for describing the embodiment of the present application and are not limited. In specific implementation, the technical scheme provided by the embodiment of the application can be flexibly applied according to actual needs.

Fig. 1 is a schematic view of an application scenario of the bullet screen processing method according to the embodiment of the present application. The application scenario shown in fig. 1 includes at least one terminal 101, a multimedia resource server 102, and a barrage information server 103. Among them, the terminal 101 has a client installed therein, and the client is served by a multimedia resource server 102 and a bullet screen information server 103. The client may be a multimedia resource client such as a browser client, a video application client, an audio application client, and a reading client, where the client is served by the multimedia resource server 102, for example, a cooling client, an internet music client, and an online reading client, and may also be a player or a reader capable of providing a playing service without networking, and the client may be used to play the multimedia resource stored in the storage space of the terminal 101, and may also be a player or a reader that has an independent playing function and can use the multimedia resource service provided by the multimedia resource server 102, for example, thousands of mutes, storm videos, and so on, which is not specifically limited in this embodiment of the present application.

A user can access the multimedia resource server 102 through a client installed in the terminal 101 to use a multimedia service provided by the multimedia resource server 102. For example, the terminal 101 may access the multimedia resource server 102 through a video application client, and may also access a web portal of the multimedia resource server 102 through a browser client. In the process of using the multimedia service provided by the multimedia resource server 102, the user may also access the bullet screen information server 103 through a client installed in the terminal 101, obtain bullet screen information corresponding to the currently used multimedia service provided by the bullet screen information server 103, display the bullet screen information through a display interface of the client, and send the bullet screen information to the bullet screen information server 103.

The multimedia resource server 102 is used for providing multimedia services, which may refer to video services, audio services, picture services, reading services, question answering services, and the like, and multimedia resources include, but are not limited to, video, audio, text, pictures, and the like. Taking the multimedia resource server 102 as a video server as an example, the video services provided by the multimedia resource server 102 may include services such as live video broadcasting, online video playing, video downloading, and the like, and for the multimedia resource server 102, the services provided by the multimedia resource server may not be a single service, for example, for the multimedia resource server, the video server may not only be limited to a video service, but also provide other types of multimedia services such as an audio service, and for the audio server, the audio server may not only be limited to an audio service, but also provide more types of multimedia services such as a video service, and of course, the multimedia resource server may also provide functions such as forwarding, commenting, and the like, which is not specifically limited in this embodiment of the present application. The video online playing service may refer to converting a certain movie into a video data stream, and providing the video data stream to the terminal 101 through a video client or a web portal for online playing or offline downloading.

It should be noted that the multimedia resource server 102 may refer to a single server or a server cluster composed of a plurality of servers, and each service may be implemented by the same server or by different servers in the server cluster, which is not specifically limited in this embodiment of the present application.

The bullet screen information server 103 is used for providing bullet screen information services, which may include: multimedia resource retrieval service and barrage service. The multimedia resource retrieval service can be used in combination with the barrage service, that is, multimedia information is converted to enable the multimedia resource to correspond to barrage information in the barrage service, and a multimedia information database is provided, where the multimedia information database can be used to store information required for conversion, such as conversion rules, correspondence between multimedia information, and the like, for conversion, so as to provide accurate barrage information service for different platforms or clients, and certainly, description information of the multimedia resource itself, such as multimedia playing duration, and the like, can also be stored in the multimedia information database. The bullet screen service means that the bullet screen information server can collect bullet screen information and provide bullet screen information corresponding to the multimedia resource currently played by the client. Specifically, the bullet screen information server 103 may serve a plurality of multimedia resource servers 102, collect and store bullet screen information sent by different platform users and different clients, and provide the collected bullet screen information to the terminal 101 for display, so as to expand the functions of the multimedia resource servers 102. The collected and stored data is bullet screen information, the bullet screen information at least comprises bullet screen texts, and the bullet screen information further comprises one or more of unique user identification of a bullet screen sender, bullet screen sending time and bullet screen interaction information. The unique user identifier may be an identifier supported by the bullet screen information server and used for uniquely identifying a bullet screen sender. The bullet screen sending time may be a time point when the user actually issues the bullet screen content, or a display time point of the bullet screen content in the multimedia resource, which is not specifically limited in this embodiment of the present application. And the barrage interaction information can be evaluation information, approval information, bad comment information, reward information, gift sending information and the like of other users on the barrage content. In addition, the bullet screen information can have a multimedia identifier for identifying the multimedia resource corresponding to the bullet screen information, the bullet screen information can also have a bullet screen identifier for uniquely identifying the bullet screen information, and each bullet screen identifier corresponds to one bullet screen information.

It should be noted that the bullet screen information server 103 may refer to a single server or a server cluster composed of a plurality of servers, and each service may be implemented by the same server or by different servers in the server cluster, which is not specifically limited in this embodiment of the present application.

Of course, the multimedia resource server 102 and the barrage information server 103 shown in fig. 1 may be arranged in the same server or a server cluster.

Of course, the method provided in the embodiment of the present application is not limited to be used in the application scenario shown in fig. 1, and may also be used in other possible application scenarios, and the embodiment of the present application is not limited. The functions that can be implemented by each device in the application scenario shown in fig. 1 will be described in the following method embodiments, and will not be described in detail herein.

To further illustrate the technical solutions provided by the embodiments of the present application, the following detailed description is made with reference to the accompanying drawings and the detailed description. Although the embodiments of the present application provide the method operation steps as shown in the following embodiments or figures, more or less operation steps may be included in the method based on the conventional or non-inventive labor. In steps where no necessary causal relationship exists logically, the order of execution of the steps is not limited to that provided by the embodiments of the present application.

The following describes the technical solution provided in the embodiment of the present application with reference to the application scenario shown in fig. 1.

Referring to fig. 2, an embodiment of the present application provides a method for training a pivot recognition model, which specifically includes the following steps:

s201, obtaining a pivot bullet screen sample set, wherein each pivot bullet screen sample comprises: and the pop-up screen text is marked with the pivot identification, and the pivot identification is used for representing whether the pop-up screen text is a pivot pop-up screen or not.

During specific implementation, a large amount of bullet screen information is stored in the bullet screen information server, bullet screen texts can be obtained from the bullet screen information, and a pivot mark used for representing whether the bullet screen is a pivot bullet screen or not is marked on each bullet screen text in an artificial marking mode, for example, the bullet screen text with the pivot mark of "1" is the pivot bullet screen, and the bullet screen text with the pivot mark of "0" is a non-pivot bullet screen. In the embodiment of the present application, a pivot popup refers to a popup text containing pivot information. And adding the bullet screen text marked with the pivot identification as a pivot bullet screen sample into a pivot bullet screen sample set.

When the method is specifically implemented, the client also provides a function of reporting the drama pop-up screen. Referring to fig. 3A, a user reports a bullet screen by clicking bullet screen text 302 displayed on a client play interface 301. Referring to fig. 3B, in response to the report operation for the bullet screen text 302, the client displays a report window 303 on the playing interface 301, then the user selects a drama option 304 in the report window, and then clicks a submit button 305, thereby completing the drama through bullet screen marking operation for the bullet screen text 302. The client responds to the operation of the showing bullet screen text for marking the bullet screen text into the bullet screen, marks the bullet screen text into the bullet screen, and sends the bullet screen text marked into the bullet screen to the bullet screen information server. And the bullet screen information server receives bullet screen texts which are marked with the pivot identifications and uploaded by the client, and adds the bullet screen texts into the pivot bullet screen sample set.

By means of the drama televisual reporting function of the client, the drama televisual barrages reported by the user through the client are automatically collected and added into the drama televisual barrage sample set, the problem that manual marking is high in cost is solved, the sample marking accuracy is improved, and the high-quality drama televisual barrage sample set is obtained in a low-cost mode.

In the embodiment of the present application, the drama popup sample is mainly a popup text displayed during playing of a multimedia resource with a scenario or an episode, such as a multimedia resource of a movie, a tv show, a novel, a cartoon, a story audio, and the like, but is not limited to the above-listed contents.

S202, for each pivot bullet screen sample in the pivot bullet screen sample set, updating the vocabulary belonging to the preset type in the pivot bullet screen sample into a replacement identifier corresponding to the preset type.

Taking a drama through barrage sample of a video as an example, a corresponding barrage text can relate to a plurality of words or character names related to a scenario, the words and the barrage text belong to videos with high drama relevance, but the frequency of the words appearing in the whole drama through barrage sample set is low, for example, the word "Xie Erduo" generally appears only in the life big explosion of a television drama, while the frequency of the words appearing in a massive barrage sample set "Xie Erduo" becomes extremely low, and the words can be treated as unlisted words when a word vector matrix is generated, so that the information can be lost when a model is trained, and the previously trained drama through recognition model cannot effectively recognize the words. Therefore, for such vocabularies which have a high degree of association with the scenario information of a certain multimedia resource and do not have a general meaning, it is necessary to replace the vocabularies with a replacement identifier having a general concept, for example, the character name may be replaced with "operator >", the key scenario words may be replaced with "result >", "roll >" and the like, so that the character name in any scenario may be replaced with "operator >", and the key scenario words may be replaced with "result >" or "roll >", so that training can be performed on the scenario information during training, and the finally obtained scenario recognition model can recognize the vocabularies related to the scenario information, and can perform scenario recognition on the multimedia resources with different characters and scenarios, and can meet semantic processing of special nouns in different multimedia resources, thereby improving accuracy and universality of the scenario recognition model. For this purpose, the preset types in step S202 include a character name and a key storyline, and a unique corresponding replacement identifier is configured for each preset type.

In specific implementation, named entity recognition can be performed on the drama-popup-screen sample, words belonging to the named entities are recognized, the preset types corresponding to the recognized words are determined, and the words belonging to the named entities are updated to the replacement identifications corresponding to the preset types. For example, the drama perspective barrage sample "Teddy is dead last" to perform named entity recognition, the "Teddy" is recognized as a named entity of the character name class, "dead" is a named entity of the key drama word class, the "Teddy" in the drama perspective barrage sample is updated to < operator >, the "dead" is updated to < result >, and the updated drama perspective barrage sample is obtained as "< operator > and last < result >. For example, in the same manner, a drama bullet-screen sample "the sister of the man is haunted ghost" is updated to "the sister of < actor >", a drama bullet-screen sample "the man and the woman are together" is updated to "the actor > and < actor > < result >", and a drama bullet-screen sample "the lawyer is the mother of the victim" is updated to "the actor > is the mother of the actor >.

Specifically, the named entities included in each preset type may be preset, for example, the named entities included in the character names may be character pronouns common in the bullet screen texts, such as "male owner", "female one", "male two", etc., words such as "fielder", "dead" are often included in the bullet screen text of the suspense movie, words such as "who and who are together" are included in the love story movie, and therefore the named entities included in the key story words may be "fielder", "dead", "together", etc. The named entities included in each preset type are not limited to the examples listed above, and the commonly used vocabulary describing the names of characters or scenarios can be determined as the named entities through the statistic pivot barrage, and the named entities corresponding to each preset type can be continuously added or adjusted based on newly added barrage text or multimedia resources in the process of using the pivot recognition model.

Further, the multimedia resource corresponding to the drama through barrage sample can be determined through the multimedia identifier corresponding to the drama through barrage sample, and related information such as a scenario introduction, a character introduction and the like corresponding to the multimedia resource is obtained, for example, for the movie introduction, the character introduction and the like of a movie or a television series, for a story introduction, a chapter directory and the like of a novel, the character name and the scenario keyword are extracted from the obtained related information, and then the extracted character name and the scenario keyword are added to the named entity included in the corresponding preset type.

And S203, aiming at each pivot bullet screen sample in the pivot bullet screen sample set, converting the updated pivot bullet screen sample into a corresponding word vector matrix.

As a possible implementation manner, step S203 specifically includes: and aiming at each perspective bullet screen sample in the perspective bullet screen sample set, carrying out word segmentation on the updated perspective bullet screen sample to obtain a plurality of words, respectively converting the obtained words into corresponding word vectors, and obtaining a word vector matrix corresponding to the perspective bullet screen sample based on the word vectors corresponding to the words.

The words in the embodiments of the present application refer to fixed phrases consisting of one or more words. For example, in Chinese, the words "this", "is", and the words "mother" and "lawyer" are also words. For example, the word "today" is a word, and the word "ice creat" is also a word.

It should be noted that, when the updated drama-perspective barrage sample is converted into the corresponding word vector matrix in the above manner, one replacement identifier in the updated drama-perspective barrage corresponds to one word. For example, the updated drama popup sample is "< operator > last < result >", and when the word segmentation processing is performed, "< operator >" is recognized as a vocabulary and "< result >" is recognized as a vocabulary.

In specific implementation, the term segmentation processing can be performed on the perspective barrage sample through a term segmentation tool so as to divide each perspective barrage sample into a plurality of words. Taking a chinese drama bullet screen sample as an example, the drama bullet screen sample can be segmented by a chinese corpus training segmentation tool (for example, jieba segmentation), for example, the drama bullet screen sample "the lawyer is the mother of the victim" the updated result "this < actor > is the mother of < actor >" is subjected to segmentation processing, and the obtained vocabulary includes: "this", "< operator >", "is exactly", "< operator >", "is", "mom". In the word segmentation process, irrelevant information such as punctuations and stop words can be optionally removed to reduce the number of words obtained through word segmentation processing and improve processing efficiency, for example, words such as "these" and "are included in the word segmentation result, and the stop words specifically include which words, which can be determined by those skilled in the art according to actual needs and by combining experience, and the embodiment of the present application is not limited.

The embodiment of the application only takes Chinese as an example, and in practical application, the drama popup screen sample is not limited to Chinese, can be in other languages such as English, japanese, german and the like, and can also be a mixed text containing multiple languages. And aiming at different languages, performing word segmentation processing by adopting corresponding word segmentation tools, wherein the specific word segmentation process is the prior art in the field and is not described in detail.

In specific implementation, for each perspective bullet screen sample in the perspective bullet screen sample set, a word vector generation model can be used to convert a plurality of words obtained through word segmentation into a word vector, then the word vectors corresponding to the words are sequentially filled into each row of a preset vector matrix to obtain a word vector matrix corresponding to the perspective bullet screen sample, and if the word vectors corresponding to the words cannot fill the vector matrix, all rows in the vector matrix in which the word vectors are not filled are set to zero. The length of the word vector is equal to the number of columns of the vector matrix, and the number of rows of the vector matrix is not less than the number of words contained in the drama bullet screen sample.

In the embodiment of the present application, the Word vector generation model is a model capable of generating a corresponding Word vector based on a text, and may be Word2vec or glove, for example. The length of the word vector and the number of columns and rows of the vector matrix may be determined according to the length of the pivot bullet screen sample and the type of the selected pivot recognition model in combination with practical experience, which is not limited in the embodiment of the present application. In practical applications, the drama popup sample is generally short text with a short length, for example, generally not more than 100 words, and in this case, the number of rows in the vector matrix may be 100.

For example, assuming that the length of the word vector is 100, the number of columns and the number of rows of the vector matrix are both 100, the vocabulary corresponding to the dramatic bullet screen sample "the lawyer is the mother of the victim" includes: the 6 vocabularies are respectively converted into 100-dimensional word vectors by using a word vector generation model, the obtained 6 word vectors are sequentially filled into the first 6 rows of a 100 x 100-dimensional vector matrix, all the 7 th row to the 100 th row of the vector matrix are set to be zero, and finally the drama pop-up bullet screen sample is converted into a 100 x 100 word vector matrix.

In practical application, when the number of samples contained in the pivot bullet screen sample set is large enough, the pivot recognition model with a good recognition effect can be obtained without considering the meaning of a vocabulary level and information of grammar, syntax, structure and the like of a language when the pivot recognition model is trained. On this basis, as another possible implementation manner, step S203 specifically includes: and aiming at each perspective bullet screen sample in the perspective bullet screen sample set, converting each word contained in the updated perspective bullet screen sample into a corresponding word vector, and obtaining a word vector matrix corresponding to the perspective bullet screen sample based on the word vector corresponding to each word.

It should be noted that, in the embodiment of the present application, for a drama barrage sample in chinese, a word refers to a word, such as "this", "law", "teacher", and the like. For the drama barrage sample in english, a word refers to a word (word) in common meaning, such as "today", "ice", "create", and the like.

It should be noted that, when the updated drama-popup screen sample is converted into the corresponding word vector matrix in the above manner, one replacement identifier in the updated drama-popup screen sample corresponds to one word. For example, the updated drama popup sample is "< operator > last < result >", where "< operator >" is considered as a word and converted into a word vector, "< result >" is considered as a word and converted into a word vector.

In specific implementation, word segmentation processing is not required to be performed on the perspective bullet screen sample, each word contained in the perspective bullet screen sample can be directly mapped into one-hot vectors to obtain a word vector corresponding to each word, then the word vector of each word is sequentially filled into each row of the vector matrix to obtain a word vector matrix corresponding to the perspective bullet screen sample, and if the word vector corresponding to the word cannot fill the vector matrix, all rows in the vector matrix in which the word vector is not filled are set to zero. The length of the word vector is equal to the number of columns of the vector matrix, and the number of rows of the vector matrix is not less than the number of words contained in the drama bullet screen sample. For the replacement identifiers, one-hot vectors corresponding to the replacement identifiers may be specified in advance, for example, the first 100 vectors in the one-hot space may be allocated to the replacement identifiers.

For example, assuming that the length of the word vector is 100, the number of columns and the number of rows of the vector matrix are both 100, the perusal bullet screen sample "the lawyer is the mother of the victim" the updated result is that "the < actor > is the mother of the < actor >, and the words contained therein: the ' this ', ' operator > ', ' just ', ' operator >, ' of ', ' mother ' and ' mother ' are respectively converted into a 100-dimensional word vector, so that the dramatic bullet screen sample ' the lawyer is the mother of the victim ' corresponds to 8 word vectors in total, the 8 word vectors are sequentially filled into the first 8 rows of the vector matrix, all the 9 th row to the 100 th row of the vector matrix are set to zero, and finally the dramatic bullet screen sample is converted into a 100 x 100 word vector matrix.

By the method, the word-level-based dramatic-perspective recognition model can be constructed, word segmentation processing is not needed, processing efficiency can be improved, and recognition accuracy of the dramatic-perspective recognition model is further improved. In addition, the minimum unit forming the language is a word no matter what kind of voice, so that a plurality of languages are supported more easily based on a word-level perspective recognition model, which is important for constructing a cross-language system.

And S204, training a pivot recognition model based on a word vector matrix and a pivot identification corresponding to pivot samples in the pivot bullet screen sample set.

In particular implementations, the pivot recognition model may be a neural network. In the embodiment of the application, the pivot identification model is trained, namely, the pivot bullet screen sample in the pivot bullet screen sample set is used for updating the weight parameters of the neural network. Wherein, each scouring process comprises the following steps: inputting a word vector matrix corresponding to a perspective pop-up screen sample in a perspective pop-up screen sample set into a neural network as an input value to obtain a predicted value corresponding to the perspective pop-up screen sample, wherein the predicted value represents the probability that the perspective pop-up screen sample is a perspective pop-up screen, calculating a loss value between a perspective mark corresponding to the perspective pop-up screen sample and the predicted value through a loss function, then calculating the gradient of each weight parameter in the neural network based on back propagation of the loss function, and updating each weight parameter in the neural network based on the gradient. And circularly executing the training steps until a neural network meeting the requirements is obtained, and obtaining a final perspective recognition model.

The neural network used for training in the embodiments of the present application includes, but is not limited to, the following types: convolutional Neural Networks (CNN), cyclic Neural Networks (RNN), long Short-Term Memory Networks (LSTM), deep Neural Networks (DNN), deep Belief Networks (DBNs), and the like.

The following explains a specific training process of the perspective recognition model by taking a convolutional neural network as an example.

Referring to fig. 4, the convolutional neural network includes an input layer, a first preset number of convolutional pooling layers, a second preset number of fully-connected layers, and an output layer, which are sequentially connected, wherein each convolutional pooling layer includes one convolutional layer and one pooling layer, which are sequentially connected. In the actual application process, the values of the first preset number and the second preset number may be determined by those skilled in the art according to the specific requirements of the perspective recognition model, such as recognition efficiency and recognition accuracy, in combination with actual experience, and the embodiment of the present application is not limited.

The process of training the dramatic recognition model based on the convolutional neural network shown in fig. 4 includes: inputting a word vector matrix corresponding to a dramatic transparent bullet screen sample into a convolutional neural network, calculating through a weight parameter of each layer of the convolutional neural network to obtain a predicted value, specifically, inputting the word vector matrix corresponding to the dramatic transparent bullet screen sample into an input layer of a dramatic transparent recognition model, inputting the word vector matrix passing through the input layer into a first preset number of convolutional pooling layers which are sequentially connected, performing convolution pooling processing on the word vector matrix for a first preset number of times to obtain a convolution pooling processing result, inputting the convolution pooling processing result into a second preset number of fully-connected layers which are sequentially connected to obtain a fully-connected processing result, inputting the fully-connected processing result into the output layer to obtain a dramatic transparent bullet screen corresponding to the bullet screen sample, if the dramatic probability value is greater than a set threshold value, determining the predicted value of the bullet screen sample as a dramatic transparent bullet screen, and otherwise determining the predicted value of the bullet screen sample as a non-dramatic transparent bullet screen. And then calculating the loss between the predicted value and the dramatic diathermy identification corresponding to the dramatic diathermy bullet screen sample, calculating the partial derivative (namely gradient) of the loss function to each weight parameter by using chain derivation, and updating the weight parameters of each layer in the convolutional neural network layer by layer according to a gradient descent formula so as to continuously optimize the convolutional neural network until the convolutional neural network meeting the requirement is obtained, namely the final dramatic diathermy identification model is obtained.

Illustratively, a convolutional neural network includes one input layer, three convolutional pooling layers, three fully-connected layers, and one output layer. The process of training the perspective recognition model comprises the following steps: the word vector matrix corresponding to the dramatic bullet screen sample is input into an input layer of a convolutional neural network, the matrix of the input layer is sequentially transferred to three convolutional pooling layers for processing, each convolutional pooling layer consists of a convolutional layer and a pooling layer, the convolutional layers in the three convolutional pooling layers can respectively use 3 x 3 convolutional kernels of 32 channels, 64 channels and 128 channels for convolution operation and are activated by using a ReLU function, and the pooling layers in the three convolutional pooling layers can uniformly use 2 x 2 maximum pooling operation. Each convolutional layer convolves the matrix input to the convolutional layer by the following convolution formula:

wherein the input matrix X of the convolutional layer is composed of X_{i·t+r,j·t+s}The matrix formed by convolution layers has an output matrix Y of Y_ijA matrix of filters W representing convolution formed by W_rsAnd p is the row length of the W matrix, q is the column length of the W matrix, m is the row length of the X matrix, n is the column length of the X matrix, t is the step length of the convolution operation, and b is the offset value after convolution.

Matrix Y input pooling layer for convolutional layer outputThe pooling layer pools the matrix Y output from the convolutional layer by the pooling formula given below, resulting in a matrix of z_ijThe matrix Z formed:

z_ij＝max(y_{i·t+r,j·t+s})，

r＝0,1,2,3,…,p-1，

s＝0,1,2,3,…,q-1，

after convolution and pooling processing of the three convolution pooling layers, a matrix output by the third convolution pooling layer is expanded into 12800-dimensional vectors, the expanded 12800-dimensional vectors are input into the full-connection layer, and full-connection processing is performed, wherein the number of neurons contained in the three full-connection layers is 12800, 4096 and 1. And finally, calculating the probability value of the perspective bullet screen sample as the perspective bullet screen through the output layer based on the value output by the full connection layer to obtain a numerical value between 0 and 1, and specifically, calculating the probability value of the perspective bullet screen sample as the perspective bullet screen through a sigmoid function. And calculating the probability value of the sprite screen sample and the loss of the actual sprite identification based on a loss function, calculating the gradient of each layer of neuron through the back propagation of the loss function, and updating each weighting parameter based on the gradient to obtain a final sprite identification model.

Because the pop-up screen text is generally short text without strong context information and the calculation speed of the convolutional neural network is high, the pop-up screen can be quickly identified based on the obtained pop-up recognition model trained by the convolutional neural network.

The pivot recognition model obtained by training through the method can output a pivot probability value based on the word vector matrix of the input bullet screen text, wherein the pivot probability value is used for representing the probability that the input bullet screen text is a pivot bullet screen. And based on a pre-trained pivot recognition model, the bullet screen text which is stored in the bullet screen information server and is not marked with the pivot identification can be recognized.

Referring to fig. 5, an embodiment of the present application provides a bullet screen processing method, which is applicable to a bullet screen information server in the scene shown in fig. 1, and specifically includes the following steps:

s501, acquiring the bullet screen text, and converting the bullet screen text into a corresponding word vector matrix.

In specific implementation, the bullet screen text without the marked pivot identifier is acquired from the bullet screen information collected by the bullet screen information server, and the vocabulary in the acquired bullet screen text belonging to the preset type is updated to the replacement identifier corresponding to the preset type. For example, named entity recognition is performed on the bullet screen text, namely the lawyer is the mother of the victim, named entities of the lawyer and the victim in the character name class are recognized, the lawyer and the victim in the dramatic breakthrough bullet screen sample are updated to < actor >, and the updated dramatic breakthrough bullet screen sample is obtained to be the mother of < actor >. And then, converting the updated bullet screen text into a corresponding word vector matrix.

If the word-level-based pivot recognition model is adopted for recognition, the updated barrage text can be converted into a corresponding word vector matrix in the following way: and converting each word contained in the updated bullet screen text into a corresponding word vector, wherein one replacement identifier in the updated bullet screen text corresponds to one word, and obtaining a word vector matrix corresponding to the bullet screen text based on the word vector corresponding to each word. The specific implementation manner of this step may refer to the specific implementation manner of step S203, and is not described again.

If the vocabulary-level-based perspective recognition model is adopted for recognition, the updated barrage text can be converted into a corresponding word vector matrix in the following way: performing word segmentation processing on the updated bullet screen text to obtain a plurality of words, wherein one replacement identifier in the updated bullet screen text corresponds to one word, the obtained words are respectively converted into corresponding word vectors, and a word vector matrix corresponding to the bullet screen text is obtained based on the word vectors corresponding to the words. The specific implementation manner of this step may refer to the specific implementation manner of step S203, and is not described again.

Furthermore, before named entity recognition, related information such as a plot introduction and a figure introduction of the multimedia resource corresponding to the barrage text can be acquired, the named entity extracted from the acquired related information is combined with a preset named entity to perform named entity recognition, so that semantic processing on proper nouns in different multimedia resources is met, and recognition accuracy of the named entity is improved.

S502, inputting the word vector matrix corresponding to the bullet screen text into a pre-trained pivot recognition model to obtain a pivot probability value corresponding to the bullet screen text.

For example, if the pivot recognition model is obtained by training using the convolutional neural network shown in fig. 4, step S202 specifically includes: inputting a word vector matrix corresponding to the bullet screen text into an input layer of the perspective identification model; inputting the word vector matrix passing through the input layer into a first preset number of convolution pooling layers which are sequentially connected so as to carry out convolution pooling processing on the word vector matrix for a first preset number of times, and obtaining a convolution pooling processing result; inputting the convolution pooling processing result into a second preset number of full connection layers which are connected in sequence to obtain a full connection processing result; and inputting the full-connection processing result into an output layer to obtain a perspective probability value corresponding to the bullet screen text. The specific architecture and processing process of the pivot recognition model can refer to the content in the training process of the pivot recognition model, and are not described in detail.

S503, determining the bullet screen text with the zoom probability value larger than the set threshold value as the zoom bullet screen.

The threshold set in this step may be specifically determined by a person skilled in the art according to the recognition accuracy of the perspective recognition model and by combining with his own experience, and the embodiment of the present application is not limited.

When the bullet screen text is determined to be a drama through bullet screen, the bullet screen text is marked with a corresponding drama through identification "1". In addition, the bullet screen text with the rating probability value not greater than the set threshold is determined as a non-rating bullet screen, and for the non-rating bullet screen, the rating identification may not be marked or the rating identification thereof may be marked as "0" to distinguish the non-rating bullet screen from the rating identification. And storing the pivot bullet screen and the non-pivot bullet screen identified by the pivot identification model into bullet screen files of corresponding multimedia resources in the bullet screen information server. When the client receives the pivot bullet screen sent by the bullet screen information server, the pivot bullet screen can be displayed according to a set display mode, and the specific implementation mode is subsequently introduced.

According to the bullet screen processing method, a large amount of bullet screen texts can be automatically and quickly judged through the bullet screen identification model, and the bullet screen of the bullet screen can be identified.

On the basis of any of the above embodiments, in order to further improve the recognition accuracy of the perspective recognition model, the multimedia resources may be divided according to the scenario types. Taking video as an example, the video can be divided into a plurality of scenario types such as suspicion, love, action, and the like. Taking a novel as an example, the novel can be divided into a plurality of plot types such as a sentiment novel, a reasoning novel and a swordsman novel. Because the plot keywords contained in the multimedia resources of the same plot type are relatively similar, the difficulty of named entity identification is reduced, and further the efficiency of model training and the identification accuracy of the perspective identification model are improved.

For each scenario type, acquiring a barrage text generated in the playing process of a multimedia resource belonging to the scenario type, marking a full-play identifier for the barrage text, adding the marked barrage text as a full-play barrage sample into a full-play barrage sample set corresponding to the scenario type, and then training a full-play identification model corresponding to the scenario type based on the full-play barrage sample in the full-play barrage sample set corresponding to the scenario type, wherein the specific training method is the same as the steps shown in fig. 2, and is not repeated.

Further, when the pop-up screen sample is obtained in a user reporting mode, the client responds to the pop-up screen marking operation aiming at the displayed pop-up screen text, obtains the multimedia identification of the multimedia resource currently played by the client, marks the multimedia identification and the pop-up identification on the pop-up screen text, and then sends the pop-up screen text to the pop-up screen information server. And the bullet screen information server receives bullet screen texts which are marked with the multimedia identifications and the play pivot identifications and uploaded by the client, determines the scenario types corresponding to the bullet screen texts according to the multimedia identifications, and adds the bullet screen texts to the play pivot bullet screen sample sets corresponding to the corresponding scenario types.

Based on a pivot recognition model trained for each scenario type, referring to fig. 6, an embodiment of the present application further provides a bullet screen processing method, which specifically includes the following steps:

s601, acquiring the bullet screen text, and converting the bullet screen text into a corresponding word vector matrix.

The step S501 may be referred to in the detailed implementation of this step, and is not described again.

S602, determining the scenario type of the multimedia resource to which the barrage text belongs.

Specifically, the barrage text is input when the user views the multimedia resource through the client, and the client sends the barrage text and the multimedia identifier of the currently played multimedia resource to the barrage information server together, so that the multimedia resource to which the barrage text belongs can be determined through the multimedia identifier corresponding to the barrage text, and the scenario type of the multimedia resource is further determined.

And S603, inputting the word vector matrix corresponding to the bullet screen text into a pivot recognition model corresponding to the scenario type of the multimedia resource to which the bullet screen text belongs.

And S604, determining the bullet screen text with the zooming probability value larger than the set threshold value as a zooming bullet screen.

Aiming at the barrage texts of the multimedia resources with different scenario types, the corresponding pivot recognition model of the scenario type is adopted to recognize the barrage texts, so that the recognition accuracy of the pivot barrage is improved.

On the basis of any embodiment, when a user needs to open the bullet screen function, the bullet screen display operation for opening the bullet screen can be input through the control interface of the client, the client acquires a corresponding bullet screen file from the bullet screen information server, and the bullet screen text in the bullet screen file is displayed. Referring to fig. 7, an embodiment of the present application provides a bullet screen processing method, which specifically includes the following steps:

s701, the client responds to the bullet screen display operation input through the control interface of the client and sends a bullet screen acquisition request to the bullet screen information server.

In specific implementation, after receiving the barrage display operation, the client acquires the multimedia identifier of the multimedia resource currently played by the client, and adds the multimedia identifier to the barrage acquisition request.

S702, the bullet screen information server responds to the bullet screen obtaining request of the client and sends a bullet screen file to the client, wherein the bullet screen file comprises: a pantographic barrage and a non-pantographic barrage.

In specific implementation, after acquiring a bullet screen acquisition request sent by a client, a bullet screen information server responds to the bullet screen acquisition request, determines a multimedia resource currently played by the client according to a multimedia identifier in the bullet screen acquisition request, and returns a bullet screen file of the multimedia resource to the client.

And S703, the client receives the bullet screen file returned by the bullet screen information server.

And S704, the client displays the bullet screen text according to the set display mode.

In specific implementation, when the set display mode is that the full play barrage is not displayed, the non-full play barrage is displayed, and the full play barrage is not displayed; and when the set display mode is to display the play-through bullet screen according to the specified display effect, displaying the non-play-through bullet screen and displaying the play-through bullet screen according to the specified display effect.

As shown in fig. 8A, a user may input a bullet screen display operation by clicking a bullet screen button 802 on a control interface 801 of a client, the client responds to the bullet screen display operation, sends a bullet screen acquisition request to a bullet screen information server, the client receives a bullet screen file returned by the bullet screen information server, and displays a bullet screen text 804 in the bullet screen file on a play interface 803, where the display effect may refer to fig. 8B. When the user clicks the bullet screen button 802 on the control interface 801 of the client again, the bullet screen is closed, and the client responds to the input bullet screen display operation to stop displaying the bullet screen text.

In practical application, the client can be set as default bullet screen closing, that is, after the client is started each time, the client does not display the bullet screen, the user opens the bullet screen by clicking the bullet screen button 802, and the bullet screen is closed by clicking the bullet screen button 802 again. Certainly, the user can modify the default parameters of the client and set the default parameters as default parameters for opening the bullet screen, that is, after the client is started each time, the client automatically sends a bullet screen acquisition request to the bullet screen information server, the client receives the bullet screen file returned by the bullet screen information server, a bullet screen text in the bullet screen file is displayed on a playing interface, and the user can close the bullet screen by clicking a bullet screen button.

As shown in fig. 3A, when the user views the multimedia resource, the control interface is generally hidden to increase the area for displaying the multimedia resource, at this time, the user may display the control interface by waking up, for example, clicking any area in the playing interface, and the client may display the control interface 801 in the manner shown in fig. 8A or fig. 8B. If the user does not perform any operation within the preset time, the client continues to hide the control interface 801.

In specific implementation, the default display mode of the client may be not to display the drama popup. The user inputs bullet screen display operation by clicking a bullet screen button on a client control interface, the client responds to the bullet screen display operation, sends a bullet screen acquisition request to a bullet screen information server, the client receives a bullet screen file returned by the bullet screen information server, only displays a non-dramatic bullet screen in the bullet screen file on a play interface, and does not display a dramatic bullet screen.

In practical application, some users want to display the pivot barrage, and the users can input corresponding barrage display operation through the control interface to switch the display mode to display the pivot barrage according to a specified display effect. As shown in fig. 8C, by clicking the pop-up setting button 805, the pop-up setting interface 806 is popped up, and by clicking the presentation mode switching button 807 in the pop-up setting interface 806, the presentation mode is switched from "not to present the drama pop-up" to present the drama pop-up according to the designated presentation effect ". When the set display mode is used for displaying the full-play pop-up screen according to the appointed display effect, the non-full-play pop-up screen is displayed according to the common display mode, and the full-play pop-up screen is displayed according to the appointed display effect so as to distinguish the full-play pop-up screen from the non-full-play pop-up screen.

In the embodiment of the present application, the display effect may be to display the bullet screen text with a set transparency, display the bullet screen text with a set font size or font, display the bullet screen text with a set bullet screen rolling speed (or bullet screen retention time), display the bullet screen text in a set bullet screen display area, and the like. For example, a non-drama barrage is displayed in an opaque manner, while a drama barrage is displayed with a certain transparency; the non-dramatic transparent bullet screen is displayed in a 10-size and Song style, and the dramatic transparent bullet screen is displayed in a 6-size and regular style; the bullet screen retention time of the non-dramatic perspective bullet screen is 6 seconds, namely the non-dramatic perspective dark bullet screen disappears after 6 seconds, and the bullet screen retention time of the dramatic perspective bullet screen is 3 seconds, namely the non-dramatic perspective dark bullet screen disappears after 3 seconds; the non-splash-through barrage can be displayed at any position in the barrage display area, and the splash-through barrage can be displayed only in the set barrage display area, for example, the edge position of the barrage display area; and so on. The above-mentioned display effects are merely exemplary, and in practical applications, other display effects may be used to display the drama pop-up screen, and the embodiment of the present application is not limited.

In the embodiment of the application, the specified display effect can be fixedly set by the client, or multiple display effects can be provided by the client, and one of the display effects is selected by the user as the specified display effect. For each presentation effect, the user may set parameters of the presentation effect through the client, for example, when the presentation effect is to present the bullet screen text with the set transparency, the user may set the transparency.

Further, the bullet screen file stored by the bullet screen information server may further include a pivot probability value corresponding to the pivot bullet screen. And after the bullet screen information server acquires the bullet screen acquisition request sent by the client, returning the corresponding bullet screen file of the multimedia resource to the client.

Based on the perspective value corresponding to the perspective pop-up screen, when the set display mode is that the perspective pop-up screen is displayed according to the appointed display effect, the client side obtains the weakening display parameter corresponding to the perspective value of the perspective pop-up screen according to the appointed display effect, the weakening display parameter is in negative correlation with the perspective value, and the perspective pop-up screen is subjected to weakening display according to the obtained weakening display parameter.

In the embodiment of the application, the weakening display parameter may be a parameter of a bullet screen text which can be weakened for display, such as transparency, a font size, a font style, a bullet screen rolling speed, bullet screen retention time, a bullet screen display area and the like.

Specifically, the weakening display parameter corresponding to the perspective probability value can be calculated by the following formula: o =1-p, where o is the fade display parameter and p is the sharpness probability value. For example, p =0.8, when the presentation effect may be to present the bullet screen text at a set transparency, the display parameter is weakened to transparency, and the transparency of the dramatic bullet screen is set to 0.2; when the bandwagon effect can be when showing the barrage text with the barrage retention time of setting for, the reduction shows the parameter and is barrage retention time, if the barrage retention time of non-drama perspective barrage is 6 seconds, then sets up the barrage retention time of drama perspective barrage into 0.2 times of the barrage retention time of non-drama perspective barrage, and the barrage retention time of drama perspective barrage is 1.2 seconds promptly. For other display effects, the weakened display parameters during the showing of the pivot bullet screen can be calculated based on the pivot probability value according to the enumerated modes, and are not repeated.

Of course, the weakening display parameter corresponding to the penetration probability value can also be calculated by other formulas, for example, o = (1-p) × α, where α is an adjustment parameter. The adjustment parameter may be preset by a person skilled in the art according to the perspective probability of the perspective barrage output by the perspective identification model in combination with actual experience, or may be set by the user.

According to the bullet screen processing method, the server automatically and quickly judges the collected bullet screen text through the pivot identification model, and a pivot bullet screen and a non-pivot bullet screen are identified. When the client needs to acquire the bullet screen text, the identified bullet screen file and the identified non-bullet screen file are sent to the client, so that the client can filter the play through bullet screen and only display the non-play through bullet screen, and thus, a user does not need to worry about being played through when enjoying interactive fun brought by the bullet screen, and user experience is improved. In addition, for a user who wants to show the pivot bullet screen, a special display effect is set for the pivot bullet screen, and the corresponding weakening display parameter when the pivot bullet screen is shown is determined according to the pivot probability value of the pivot bullet screen so as to weaken the showing of the pivot bullet screen.

Referring to fig. 9, an embodiment of the present application further provides a bullet screen processing method, which specifically includes the following steps:

s901, the client A responds to the bullet screen input operation and sends the input bullet screen text to the bullet screen information server.

In specific implementation, a user can perform barrage input operation through the client a when watching a multimedia resource, namely, inputting a barrage text, the client a responds to the barrage input operation to obtain a multimedia identifier of the multimedia resource currently played by the client a, marks the multimedia identifier on the barrage text input by the user, and sends the barrage text marked with the multimedia identifier to the barrage information server.

S902, the bullet screen information server receives the bullet screen text sent by the client A, and converts the bullet screen text into a corresponding word vector matrix.

S903, inputting the word vector matrix corresponding to the bullet screen text into a pre-trained pivot recognition model by the bullet screen information server to obtain a pivot probability value corresponding to the bullet screen text.

In specific implementation, if different perspective recognition models are trained for different scenario types, the scenario type to which the multimedia resource corresponding to the barrage text belongs can be recognized according to the multimedia identifier corresponding to the barrage text, the perspective recognition model corresponding to the recognized scenario type is obtained, the word vector matrix corresponding to the barrage text is input into the obtained perspective recognition model corresponding to the scenario type, and the perspective probability value corresponding to the barrage text is obtained.

And S904, the bullet screen information server determines whether the bullet screen text is the full play bullet screen according to the full play probability value, and marks a corresponding full play identification for the bullet screen text.

In this step, if the pivot probability value is greater than the bullet screen text with the set threshold, the bullet screen text is determined to be the pivot bullet screen, otherwise, the bullet screen text is determined to be the non-pivot bullet screen.

And S905, the bullet screen information server sends the bullet screen text marked with the pivot identification to other clients.

In this step, the other clients refer to the client that is playing the same multimedia resource as the client a and needs to acquire the barrage.

And S906, the other clients receive the bullet screen text sent by the bullet screen information server and display the bullet screen text according to the set display mode.

Specifically, other clients determine whether the bullet screen text is a full play bullet screen according to the full play identification of the bullet screen text, and if the bullet screen text is a non-full play bullet screen, the bullet screen text is displayed; if the pop-up screen text is a pivot pop-up screen, the pop-up screen text is displayed according to a display mode set by each client, for example, if the set display mode is not to display the pivot pop-up screen, the pivot pop-up screen is not displayed, and if the set display mode is to display the pivot pop-up screen according to a specified display effect, the pivot pop-up screen is displayed according to the display effect set by the client. For further implementation of the display effect, reference may be made to the specific implementation of step S704, which is not described in detail herein.

The bullet screen processing method of the embodiment of the application automatically and quickly judges the bullet screen text sent by the client in real time through the pivot identification model, marks the pivot identification used for representing whether the bullet screen text is the pivot bullet screen or not on the bullet screen text based on the judgment result, then sends the bullet screen text marked with the pivot identification to other clients, and the other clients determine whether the bullet screen text is the pivot bullet screen or not according to the pivot identification. Therefore, the bullet screen processing method provided by the embodiment of the application can quickly identify whether the bullet screen text sent by the user is a full play bullet screen or not, and sends the identification result and the bullet screen text to other clients in real time, so that the bullet screen text can be shared among the users in real time, the interaction instantaneity is enhanced, and the user experience is improved.

As shown in fig. 10, based on the same inventive concept as the bullet screen processing method, the embodiment of the present application further provides a bullet screen processing apparatus 100, including: a preprocessing module 1001, an identification module 1002 and a judgment module 1003.

The preprocessing module 1001 is configured to obtain a bullet screen text and convert the bullet screen text into a corresponding word vector matrix.

The recognition module 1002 is configured to input a word vector matrix corresponding to the bullet screen text into a pre-trained pivot recognition model to obtain a pivot probability value corresponding to the bullet screen text, where the pivot recognition model is obtained based on pivot sample training, and the pivot probability value is used to indicate a probability that the bullet screen text is a pivot bullet screen.

The determining module 1003 is configured to determine the bullet screen text with the pivot probability value being greater than the set threshold as the pivot bullet screen, so that the client displays the pivot bullet screen according to the set display mode when receiving the pivot bullet screen.

Optionally, the system further comprises a response module, configured to: responding to a bullet screen acquisition request of the client, and sending a bullet screen file to the client, wherein the bullet screen file comprises: a spoiler pop-up screen and a non-spoiler pop-up screen.

Optionally, the bullet screen file further includes a pivot probability value corresponding to the pivot bullet screen, so that the client determines a display effect of the pivot bullet screen according to the pivot probability value corresponding to the pivot bullet screen.

Optionally, a training module is further included for:

for each pivot bullet screen sample in the pivot bullet screen sample set, updating vocabularies belonging to a preset type in the pivot bullet screen sample into a replacement identifier corresponding to the preset type;

for each perspective bullet screen sample in the perspective bullet screen sample set, converting the updated perspective bullet screen sample into a corresponding word vector matrix;

and training a drama clarity identification model based on the word vector matrix and the drama clarity identification corresponding to the drama clarity bullet screen samples in the drama clarity bullet screen sample set.

Further, each spoil barrage sample comprises: and the bullet screen text marked with the pivot identification is manually marked or the bullet screen text marked with the pivot identification is uploaded by the client.

Optionally, the training module is specifically configured to: and converting each word contained in the updated perspective bullet screen sample into a corresponding word vector, wherein one replacement identifier in the updated perspective bullet screen sample corresponds to one word, and a word vector matrix corresponding to the perspective bullet screen sample is obtained based on the word vector corresponding to each word. Or, the training module is specifically configured to: performing word segmentation on the updated perspective bullet screen sample to obtain a plurality of words, wherein one replacement identifier in the updated perspective bullet screen corresponds to one word, converting the obtained words into corresponding word vectors respectively, and obtaining a word vector matrix corresponding to the perspective bullet screen sample based on the word vectors corresponding to the words.

Optionally, the preprocessing module is specifically configured to: and converting each word contained in the updated bullet screen text into a corresponding word vector, wherein one replacement identifier in the updated bullet screen text corresponds to one word, and a word vector matrix corresponding to the bullet screen text is obtained based on the word vector corresponding to each word. Or the preprocessing module is specifically configured to perform word segmentation on the updated bullet screen text to obtain a plurality of words, where one replacement identifier in the updated bullet screen text corresponds to one word, convert the obtained words into corresponding word vectors respectively, and obtain a word vector matrix corresponding to the bullet screen text based on the word vectors corresponding to the words.

Optionally, the training module is specifically configured to: acquiring a perspective barrage sample set corresponding to any scenario type, wherein perspective barrage samples in the perspective barrage sample set corresponding to any scenario type are barrage texts generated in the playing process of multimedia resources belonging to any scenario type; and training a pivot recognition model corresponding to any plot type based on a word vector matrix and pivot identification corresponding to the pivot samples in the pivot bullet screen sample set.

Optionally, the system further comprises a type identification module, configured to: determining the plot type of the multimedia resource of the barrage text;

accordingly, the identifying module 1002 is specifically configured to: and inputting the word vector matrix corresponding to the bullet screen text into a pivot recognition model corresponding to the plot type of the multimedia resource of the bullet screen text to obtain a pivot probability value corresponding to the bullet screen text.

Optionally, the pivot recognition model includes an input layer, a first preset number of convolution pooling layers, a second preset number of full-connected layers, and an output layer, which are connected in sequence, and each convolution pooling layer includes a convolution layer and a pooling layer, which are connected in sequence. Accordingly, the identifying module 1002 is specifically configured to: inputting a word vector matrix corresponding to the bullet screen text into an input layer of the perspective identification model; inputting the word vector matrix passing through the input layer into a first preset number of convolution pooling layers which are sequentially connected, so as to perform convolution pooling processing on the word vector matrix for a first preset number of times, and obtain a convolution pooling processing result; inputting the convolution pooling processing result into a second preset number of full-connection layers which are connected in sequence to obtain a full-connection processing result; and inputting the full-connection processing result into an output layer to obtain a perspective probability value corresponding to the bullet screen text.

The bullet screen processing device and the bullet screen processing method provided by the embodiment of the application adopt the same inventive concept, can obtain the same beneficial effects, and are not described again.

As shown in fig. 11, based on the same inventive concept as the bullet screen processing method, the embodiment of the present application further provides a bullet screen processing apparatus 110, including: the system comprises a bullet screen requesting module 1101, a bullet screen receiving module 1102 and a bullet screen displaying module 1103.

A bullet screen request module 1101, configured to send a bullet screen acquisition request to the server in response to a bullet screen display operation input through a control interface of the client;

the bullet screen receiving module 1102 is configured to receive bullet screen files returned by the server, where the bullet screen files include a pivot bullet screen and a non-pivot bullet screen;

the bullet screen display module 1103 is configured to display the non-pivot bullet screen and the pivot bullet screen according to a specified display effect when the set display mode is to display the pivot bullet screen according to the specified display effect.

Optionally, the bullet screen file further includes a corresponding pivot probability value of the pivot bullet screen, and the pivot probability value is determined based on the pivot recognition model.

Correspondingly, the bullet screen display module 1103 is specifically configured to: according to the appointed display effect, obtaining a weakening display parameter corresponding to the spoiler probability value of the spoiler bullet screen, wherein the weakening display parameter is negatively correlated with the spoiler probability value; and performing weakening display on the dramatic bullet screen according to the obtained weakening display parameters.

Optionally, the method further comprises a mode determination module, configured to: and determining a display mode according to the bullet screen display operation.

Optionally, the system further comprises a perspective barrage marking module, configured to: responding to the zooming bullet screen marking operation aiming at the displayed bullet screen text, and marking the bullet screen text as a zooming bullet screen; and sending the bullet screen text marked as the drama bullet screen to the server so that the server adds the bullet screen text marked as the drama bullet screen to a drama bullet screen sample set for training the drama identification model.

Based on the same inventive concept as the bullet screen processing method, the embodiment of the present application further provides a server, which may specifically be the bullet screen information server shown in fig. 1, or may be a server independent of the bullet screen information server. As shown in fig. 12, the server 120 may include a processor 1201 and a memory 1202.

The Processor 1201 may be a general-purpose Processor, such as a Central Processing Unit (CPU), a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other Programmable logic device, discrete Gate or transistor logic, or discrete hardware components, that may implement or perform the methods, steps, and logic blocks disclosed in the embodiments of the present Application. A general purpose processor may be a microprocessor or any conventional processor or the like. The steps of a method disclosed in connection with the embodiments of the present application may be directly implemented by a hardware processor, or may be implemented by a combination of hardware and software modules in a processor.

Memory 1202, which is a non-volatile computer-readable storage medium, may be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. The Memory may include at least one type of storage medium, and may include, for example, a flash Memory, a hard disk, a multimedia card, a card-type Memory, a Random Access Memory (RAM), a Static Random Access Memory (SRAM), a Programmable Read Only Memory (PROM), a Read Only Memory (ROM), a charged Erasable Programmable Read Only Memory (EEPROM), a magnetic Memory, a magnetic disk, an optical disk, and so on. The memory is any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited to such. The memory 1202 in the embodiments of the present application may also be circuitry or any other device capable of performing a storage function for storing program instructions and/or data.

Based on the same inventive concept as the bullet screen processing method, the embodiment of the present application further provides a terminal, which may be a desktop computer, a portable computer, a smart phone, a tablet computer, a Personal Digital Assistant (PDA), or the like. As shown in fig. 13, the terminal 130 may include a processor 1301 and a memory 1302.

The Processor 1301 may be a general-purpose Processor, such as a Central Processing Unit (CPU), a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other Programmable logic device, discrete Gate or transistor logic device, or discrete hardware components, that may implement or perform the methods, steps, and logic blocks disclosed in the embodiments of the present Application. A general purpose processor may be a microprocessor or any conventional processor or the like. The steps of a method disclosed in connection with the embodiments of the present application may be directly implemented by a hardware processor, or may be implemented by a combination of hardware and software modules in a processor.

Memory 1302, which is a non-volatile computer-readable storage medium, may be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. The Memory may include at least one type of storage medium, and may include, for example, a flash Memory, a hard disk, a multimedia card, a card-type Memory, a Random Access Memory (RAM), a Static Random Access Memory (SRAM), a Programmable Read Only Memory (PROM), a Read Only Memory (ROM), a charged Erasable Programmable Read Only Memory (EEPROM), a magnetic Memory, a magnetic disk, an optical disk, and so on. The memory is any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited to such. The memory 1202 in the embodiments of the present application may also be circuitry or any other device capable of performing a storage function for storing program instructions and/or data.

An embodiment of the present application provides a computer-readable storage medium, which is used for storing computer program instructions for the electronic device, and which includes a program for executing the bullet screen processing method.

The computer storage media may be any available media or data storage device that can be accessed by a computer, including but not limited to magnetic memory (e.g., floppy disks, hard disks, magnetic tape, magneto-optical disks (MOs), etc.), optical memory (e.g., CDs, DVDs, BDs, HVDs, etc.), and semiconductor memory (e.g., ROMs, EPROMs, EEPROMs, non-volatile memory (NAND FLASH), solid State Disks (SSDs)), etc.

The above embodiments are only used to describe the technical solutions of the present application in detail, but the above embodiments are only used to help understanding the method of the embodiments of the present application, and should not be construed as limiting the embodiments of the present application. Modifications and substitutions that may be readily apparent to those skilled in the art are intended to be included within the scope of the embodiments of the present application.

Claims

1. A bullet screen processing method is characterized by comprising the following steps:

acquiring a bullet screen text, updating the vocabulary belonging to a preset type in the bullet screen text into a replacement identifier corresponding to the preset type, and converting the updated bullet screen text into a corresponding word vector matrix;

determining the pop-up screen text with the sprite probability value larger than a set threshold value as a sprite pop-up screen so that the client can display the sprite pop-up screen according to a set display mode when receiving the sprite pop-up screen; the display mode comprises a font, a font size, a bullet screen rolling speed and bullet screen retention time;

the updating the vocabulary belonging to the preset type in the bullet screen text into the replacement identifier corresponding to the preset type comprises the following steps:

acquiring related information corresponding to the bullet screen text, and performing named entity recognition on the bullet screen text by combining a preset named entity according to the named entity extracted from the related information to recognize words belonging to the named entity; the related information at least comprises a plot introduction and a character introduction of the multimedia resource corresponding to the barrage text;

and determining a preset type corresponding to the vocabulary belonging to the named entity, and updating the vocabulary belonging to the named entity into a replacement identifier corresponding to the preset type.

2. The method of claim 1, further comprising:

responding to a bullet screen acquisition request of a client, and sending a bullet screen file to the client, wherein the bullet screen file comprises: the pop-up screen and the non-pop-up screen, or the pop-up screen file comprises: the method comprises the steps of determining a full-play bullet screen, a non-full-play bullet screen and full-play probability values corresponding to the full-play bullet screen by the client, so that the client determines weakening display parameters when the full-play bullet screen is displayed according to the full-play probability values corresponding to the full-play bullet screen.

3. Method according to claim 1 or 2, characterized in that the spoiler recognition model is trained by:

obtaining a pivot bullet screen sample set, wherein each pivot bullet screen sample comprises: the method comprises the steps that bullet screen texts marked with a pivot mark are manually marked or bullet screen texts marked with the pivot mark are uploaded by a client, wherein the pivot mark is used for representing whether the bullet screen texts are pivot bullet screens or not;

4. The method according to claim 3, wherein the converting the updated drama-through bullet screen samples into the corresponding word vector matrix specifically comprises:

performing word segmentation on the updated perspective bullet screen sample to obtain a plurality of words, wherein one replacement identifier in the updated perspective bullet screen corresponds to one word, the obtained words are respectively converted into corresponding word vectors, and a word vector matrix corresponding to the perspective bullet screen sample is obtained based on the word vectors corresponding to the words.

5. The method according to claim 1, wherein the converting the updated bullet screen text into the corresponding word vector matrix specifically includes:

converting each word contained in the updated bullet screen text into a corresponding word vector, wherein one replacement identifier in the updated bullet screen text corresponds to one word, and obtaining a word vector matrix corresponding to the bullet screen text based on the word vector corresponding to each word; alternatively, the first and second electrodes may be,

6. The method according to claim 3, wherein the obtaining of the drama popup screen sample set specifically comprises:

the training of the pivot recognition model based on the word vector matrix and the pivot identification corresponding to the pivot samples in the pivot bullet screen sample set specifically includes:

7. The method of claim 6, further comprising: determining the plot type of the multimedia resource to which the barrage text belongs;

and inputting the word vector matrix corresponding to the barrage text into a drama identification model corresponding to the scenario type of the multimedia resource to which the barrage text belongs.

8. The method of claim 1, wherein the pivot recognition model comprises an input layer, a first predetermined number of convolutional pooling layers, a second predetermined number of fully-connected layers, and an output layer connected in sequence, each convolutional pooling layer comprising one convolutional layer and one pooling layer connected in sequence;

inputting the word vector matrixes passing through the input layer into a first preset number of convolution pooling layers which are sequentially connected, so as to perform convolution pooling processing on the word vector matrixes for a first preset number of times, and obtain convolution pooling processing results;

9. A bullet screen processing method is characterized by comprising the following steps:

displaying the bullet screen file according to a set display mode, and displaying the non-pivot bullet screen and not displaying the pivot bullet screen when the set display mode is that the pivot bullet screen is not displayed; when the set display mode is that the pivot bullet screen is displayed according to the specified display effect, the non-pivot bullet screen is displayed, and the pivot bullet screen is displayed according to the specified display effect;

the pop-up screen file further comprises a spoiler probability value corresponding to the spoiler pop-up screen, wherein the spoiler probability value is determined based on a spoiler identification model;

the showing of the pivot bullet screen according to the specified showing effect specifically comprises:

according to the appointed display effect, obtaining a weakening display parameter corresponding to the perspective probability value of the perspective bullet screen, wherein the weakening display parameter is negatively related to the perspective probability value; the weakening display parameters comprise fonts, word sizes, bullet screen rolling speed and bullet screen retention time;

10. The method of claim 9, further comprising:

responding to a pivot bullet screen marking operation aiming at the displayed non-pivot bullet screen, and marking the non-pivot bullet screen as a pivot bullet screen;

and sending the non-pivot bullet screen marked as the pivot bullet screen to the server so that the server adds the non-pivot bullet screen marked as the pivot bullet screen to a pivot bullet screen sample set for training the pivot recognition model.

11. A bullet screen processing apparatus, comprising:

the system comprises a preprocessing module, a word vector matrix module and a word vector matrix module, wherein the preprocessing module is used for acquiring a bullet screen text, updating words belonging to a preset type in the bullet screen text into a replacement identifier corresponding to the preset type, and converting the updated bullet screen text into a corresponding word vector matrix;

the judging module is used for determining the bullet screen text with the pivot probability value larger than a set threshold value as a pivot bullet screen so that the client displays the pivot bullet screen according to a set display mode when receiving the pivot bullet screen; the display mode comprises a font, a font size, a bullet screen rolling speed and bullet screen retention time;

the preprocessing module is specifically configured to: acquiring related information corresponding to the bullet screen text, and performing named entity recognition on the bullet screen text by combining a preset named entity according to the named entity extracted from the related information to recognize words belonging to the named entity; the related information at least comprises a plot introduction and a character introduction of the multimedia resource corresponding to the barrage text; and determining a preset type corresponding to the vocabulary belonging to the named entity, and updating the vocabulary belonging to the named entity into a replacement identifier corresponding to the preset type.

12. A bullet screen processing apparatus, comprising:

the bullet screen request module is used for responding bullet screen display operation input through a control interface of the client and sending a bullet screen acquisition request to the server;

the bullet screen display module is used for displaying the bullet screen file according to a set display mode, and when the set display mode is that the full play bullet screen is not displayed, the non-full play bullet screen is displayed, and the full play bullet screen is not displayed; when the set display mode is that the pivot bullet screen is displayed according to the specified display effect, the non-pivot bullet screen is displayed, and the pivot bullet screen is displayed according to the specified display effect;

the bullet screen file further comprises a perspective probability value corresponding to the perspective bullet screen, and the perspective probability value is determined based on a perspective identification model; the bullet screen display module is specifically used for:

13. A computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, implement the steps of the method of any one of claims 1 to 10.