WO2017166512A1

WO2017166512A1 - Video classification model training method and video classification method

Info

Publication number: WO2017166512A1
Application number: PCT/CN2016/089246
Authority: WO
Inventors: 张立宁; 余婧
Original assignee: 乐视控股（北京）有限公司; 乐视云计算有限公司
Priority date: 2016-03-31
Filing date: 2016-07-07
Publication date: 2017-10-05
Also published as: CN105913072A

Abstract

A training method of a video classification model and a video classification method on the basis of the trained video classification model, the training method of the video classification model comprising: obtaining text content and existing category labels for each video in a video set of a certain domain (S1); segmenting words of the text content of each video to obtain a set of attribute words for each video (S2); and establishing a Bayesian model, and inputting the set of attribute words and the existing category labels of each video in the video set of the certain domain into the Bayesian model to train the Bayesian model, so as to obtain a video classification model (S3). The video classification method comprises: segmenting words of the text content of a to-be-classified video, to obtain a set of attribution words of the to-be-classified video (S02); inputting each attribute word in the set of attribute words into a video classification model to determine a category label of the to-be-classified video in the category directory. With this method, it is possible to achieve the classification of video efficiently, easily and with high accuracy.

Description

Training method and video classification method for video classification model

The present application claims priority to Chinese Patent Application No. 2016102024974, entitled "Training Method of Video Classification Model and Video Classification Method", filed on March 31, 2016, the entire contents of which are incorporated by reference. In this application.

Technical field

The present disclosure belongs to the field of Internet technologies, and in particular, to a training method and a video classification method for a video classification model.

Background technique

In the context of big data, the classified storage of video plays an important role in the management of video and the recommendation of interest. In the prior art, for some professional category video playing websites (for example, an educational platform for playing teaching videos), it has its own set of video management system to classify and store videos on the website. However, due to the limited capacity of the category video playback website, it does not have long-distance transcoding capability. When it wants to upload a video, it needs to generate the ID of the video by using the long-distance transcoding function provided by the video service provider (such as LeEco Cloud Platform). The ID is distributed to the video service provider's CDN platform. When it needs to play the video, it only needs to obtain the network address of the video from the video service provider's CDN platform to play. Since the ID is generally composed of a series of meaningless letters and numbers (the ID of each video is unique), for the video service provider, the content tag of the video stored in its cloud platform is only a string of Meaning letters and numbers. Therefore, it is very difficult for video service providers to classify this type of video in their cloud platform.

Summary of the invention

The purpose of the present disclosure is to enable an accurate classification of video stored by a video service provider (eg, LeEco Cloud Platform) in a cluster of cloud platform servers that it is built.

In order to achieve the objectives of the present disclosure, the present disclosure provides a training method for a video classification model, including the following steps:

Obtain the text content and existing category labels of each video in a certain area video collection;

Segmenting the text content of each video to obtain a set of attribute words for each video;

A Bayesian model is established, and a set of attribute words and an existing category label of each video in the domain video set are input to a Bayesian model to train the Bayesian model to obtain a video classification model.

Further, the training method of the video classification model, after the step of acquiring the text content of each video in the video collection of a certain domain and the existing category label, further includes:

A category directory of the video collection of the domain is established according to the existing category label.

Further, the training method of the video classification model, wherein the input parameter of the video classification model is an attribute word, and the output parameter is: a plurality of category probability values; wherein each category probability value indicates that the attribute word belongs to the The probability of a category in the category catalog.

Further, the training method of the video classification model, wherein the step of acquiring the text content and the category label of each video in the video collection of a certain domain comprises:

Obtaining a network address of each video in a video collection of a certain domain stored in the cloud server;

Obtaining a play webpage of each video by using a webpage crawling algorithm according to the network address of the video;

The text content and category labels of the current video are extracted from each video play page network.

Further, the training method of the video classification model, wherein the step of segmenting the text content of each video to obtain a set of attribute words of each video comprises:

Segmenting the text content to obtain a word segmentation result;

Performing part-of-speech tagging on each word in the segmentation result according to the part-of-speech tagging algorithm, and filtering the words in the segmentation result according to the tagging result to obtain a first-level keyword set;

According to the stop word table, the first level keyword set is filtered to obtain a set of attribute words.

Further, the training method of the video classification model, wherein the text content includes a title and/or a content introduction of a current video.

Further, the training method of the video classification model, wherein the Bayesian model is a naive Bayesian model.

According to another aspect of the present disclosure, there is also provided a video classification method comprising the following steps:

Get the text content of the video to be classified;

Performing word segmentation on the text content of the classified video to obtain a set of attribute words of the video to be classified;

Entering each of the attribute words in the set of attribute words of the video to be classified into the video classification model according to any one of claims 1 to 4, and obtaining a category probability value of each attribute word of the video to be classified;

Determining, according to the category probability value of each attribute word, a category label of the to-be-categorized video in the category directory.

Further, in the video classification method, in the step of obtaining a category probability value of each attribute word of the video to be classified, the attribute word includes at least one category probability value.

Further, in the video classification method, the step of classifying the video to be classified according to the category probability value of each attribute word includes the following steps:

From the plurality of category probability values of each attribute word, the one with the largest value is selected as the optimal category probability value of the attribute word;

Performing a product operation on the optimal class probability value of each attribute word in the attribute word set of the classified video to obtain a class probability of the video to be classified;

Determining, according to a category probability of the video to be classified, a category label of the to-be-categorized video in the category directory.

According to still another aspect of the present disclosure, a computer storage medium is further provided, wherein the computer storage medium can store a program, and when the program is executed, each implementation manner of a training method of a video classification model provided by the present invention can be implemented. Part or all of the steps.

According to still another aspect of the present disclosure, there is also provided a computer storage medium, wherein the computer storage medium may store a program that, when executed, may implement portions of various implementations of a video classification method provided by the present invention or All steps.

The present disclosure enables efficient classification of video by high efficiency, simplicity, and high accuracy.

The above general description and the following detailed description are intended to be illustrative and not restrictive.

DRAWINGS

1 is a flow chart showing the steps of a training method of the video classification model of the present disclosure;

2 is a flow chart showing steps of acquiring text content and category tags of a video in a training method of the video classification model of the present disclosure;

3 is a flow chart showing the steps of segmenting the text content of each video in the training method of the video classification model of the present disclosure;

4 is a flow chart showing the steps of the video classification method of the present disclosure;

FIG. 5 is a flow chart showing the steps of classifying a video to be classified according to a category probability value of each attribute word in the video classification method of the present disclosure.

detailed description

The present disclosure will be further described in detail below with reference to the specific embodiments thereof and the accompanying drawings. It is to be understood that the description is not intended to limit the scope of the disclosure. In addition, descriptions of well-known structures and techniques are omitted in the following description in order to avoid unnecessarily obscuring the concept of the present disclosure.

1 is a flow chart showing the steps of a training method of the video classification model of the present disclosure.

As shown in FIG. 1, a training method for a video classification model includes the following steps:

In step S1, the text content and the existing category label of each video in the video collection in a certain domain are obtained.

In some professional category video playing websites (such as an educational platform for playing instructional videos), the video playing page on the website includes text content that is edited in natural language and describes the content of the video, the text content including the current video. Title and / or content introduction. A field can be in the fields of education, news, entertainment, and so on. In addition, in order to facilitate the management of video, these professional category video playing websites generally establish their own set of category directories, wherein the set of category directories includes multiple category names, and each video is divided into corresponding categories. Under the category name, use the category name as the category label for the video. The existing category label described in the present disclosure refers to the category label of the video in the professional category video playing website.

After the step of acquiring the text content of each video in the video collection of the domain and the existing category label, the method further includes: step of establishing a category directory of the video collection of the domain according to the existing category label. It should be noted that, because the video source of a certain field stored in the video service provider's cloud platform (for example, LeEco Cloud Platform), the video source is not only a video playing website, but may be derived from a large amount of video. Playing the website, therefore, since the existing category directory of each video playing website may not be comprehensive and cannot cover all the videos in the video collection of a certain domain, the present disclosure needs to re-create the video of the field based on the existing category label. The category directory of the collection.

Specifically, the present disclosure takes the field of education as an example, and the category names in the catalogue of the re-established video collections of the educational field mainly include: pre-school, elementary school, junior high school, junior high school, senior high school entrance examination, high school, college entrance examination, university, study abroad, civil servant, and judicial , IT, finance and finance, international study tours, management, life skills, sports, summer camps, interests, arts, language training, pregnancy and baby counseling, vocational skills, and others.

In step S2, the text content of each video is segmented to obtain a set of attribute words for each video.

In this step, the text content of each video can be segmented by using a word segmentation algorithm in the prior art to obtain a set of attribute words for each video. Wherein, the set of attribute words of each video includes at least one attribute word.

Step S3, a Bayesian model is established, and the attribute word set of each video in the domain video set and the existing category label are input to the Bayesian model to train the Bayesian model to obtain a video classification model.

The Bayesian model is a naive Bayesian model. The input parameter of the video classification model is an attribute word, and the output parameter is: a plurality of category probability values. Wherein, each category probability value indicates a probability that the attribute word belongs to a category in the category catalog.

2 is a flow chart showing the steps of acquiring text content and category tags of a video in a training method of the video classification model of the present disclosure.

As shown in FIG. 2, the step of acquiring the text content and the category label of each video in the video collection of a certain domain includes:

Step S11: Obtain a network address of each video in a video collection of a certain domain stored in the cloud server.

Prior to step S1, the professional category video playing website using the long-range transcoding service provided by the cloud platform server cluster generates the long-range transcoding function provided by the video service provider (for example, LeEco Cloud Platform) on the video on the website. The ID of the video is then distributed to one or more servers (ie, cloud servers) in the CDN platform of the video service provider, and the cloud server stores the video. It should be noted that since the video service provider usually provides long-distance transcoding services for a large number of video playing websites, the video service provider's cloud server stores a large amount of video, an ID of each video, and a network of each video. address. Therefore, in step S11, only the network address of the video needs to be acquired.

Step S12: Obtain a broadcast of each video by using a webpage crawling algorithm according to the network address of the video. Put the page.

The web crawling algorithm refers to an algorithm based on the prior art web crawler. The web crawler is a program for automatically extracting web pages, which is a search engine for downloading web pages from the World Wide Web, and is an important component of the search engine. The traditional crawler starts from the URL of one or several initial webpages and obtains the URL on the initial webpage. During the process of crawling the webpage, the new URL is continuously extracted from the current page into the queue until a certain stop condition of the system is satisfied.

In step S13, the text content and the category label of the current video are extracted from each video playpage network.

In some professional category video playing websites (such as an educational platform for playing instructional videos), the video playing page on the website includes text content that is edited in natural language and describes the content of the video, the text content including the current video. Title and / or content introduction. In addition, in order to facilitate the management of video, these professional category video playing websites generally establish their own set of category directories, wherein the set of category directories includes multiple category names, and each video is divided into corresponding categories. Under the category name, use the category name as the category label for the video. The existing category label described in the present disclosure refers to the category label of the video in the professional category video playing website.

3 is a flow chart showing the steps of word segmentation of the text content of each video in the training method of the video classification model of the present disclosure.

As shown in FIG. 3, in step S2, the text content of each video is segmented, and the step of obtaining the attribute word set of each video includes:

Step S21, performing segmentation on the text content, obtaining a word segmentation result, performing part-of-speech tagging on each word in the word segmentation result according to the part-of-speech tagging algorithm, and screening the words in the segmentation result according to the tagging result to obtain a Level keyword collection. The first level keyword set includes multiple first level keywords.

Since the text content is described in natural language, including many words, some of which may be unnecessary words, a predetermined algorithm is needed for the text content to extract keywords to filter out some unnecessary words. In this step, only the part of the text is segmented according to the part of speech of the word segmentation part of the word segmentation. On the one hand, the words are segmented, on the other hand, some structural words, modal particles and other words are filtered out, such as, ah, ah. In addition, before the step, the method further includes storing the participle part-of-speech table in the cloud server, and updating the participle part-of-speech table from time to time.

Step S22, filtering the first-level keyword set according to the stop word table to obtain an attribute word set. Hehe. The attribute word set includes a plurality of attribute words.

Before this step, the method further includes storing the stop word table in the cloud server, and updating the stop word table from time to time. Among them, the stop word list uses the stop word list in the prior art. Filtering the first-level keyword set refers to filtering out the stop words in the primary keyword set.

4 is a flow chart of the steps of the video classification method of the present disclosure.

As shown in FIG. 4, a video classification method includes the following steps:

In step S01, the text content of the video to be classified is obtained.

Obtaining a video to be classified is a new video, which is a new video uploaded to the cloud server.

Step S02: Perform word segmentation on the text content of the classified video to obtain a set of attribute words of the video to be classified.

Step S03: Enter each attribute word in the attribute word set of the video to be classified into the video classification model according to any one of claims 1-4 to obtain a category probability value of each attribute word of the video to be classified.

Determining, according to the category probability value of each attribute word, a category label of the to-be-categorized video in the category directory. In the step of obtaining a category probability value of each attribute word of the video to be classified, the attribute word includes at least one category probability value.

As shown in FIG. 5, the step of classifying the video to be classified according to the category probability value of each attribute word includes the following steps:

Step S031: Select one of the plurality of category probability values of each attribute word as the optimal category probability value of the attribute word.

Step S032: Perform a product operation on the optimal category probability value of each attribute word in the attribute word set of the classified video to obtain a class probability of the to-be-classified video.

Step S033: Determine, according to the category probability of the video to be classified, a category label of the to-be-categorized video in the category directory.

The embodiment of the invention further provides a computer storage medium, wherein the computer storage medium can be stored There is a program, which can implement some or all of the implementation steps of the training method of the video classification model provided by the embodiment shown in FIG. 1 to FIG.

The embodiment of the present invention further provides a computer storage medium, wherein the computer storage medium can store a program, and the program can be implemented in each implementation manner of a video classification method provided by the embodiment shown in FIG. 4-5. Part or all of the steps.

The above-described embodiments of the present disclosure are to be construed as merely illustrative or illustrative of the invention. Therefore, any modifications, equivalent substitutions, improvements, etc., which are made without departing from the spirit and scope of the disclosure, are intended to be included within the scope of the disclosure. Rather, the scope of the appended claims is intended to cover all such modifications and

Claims

A training method for a video classification model, comprising:

Obtain the text content and existing category labels of each video in a certain area video collection;

Segmenting the text content of each video to obtain a set of attribute words for each video;

A Bayesian model is established, and a set of attribute words and an existing category label of each video in the domain video set are input to a Bayesian model to train the Bayesian model to obtain a video classification model.
The method according to claim 1, wherein after acquiring the text content of each video in the video collection of the domain and the existing category label, the method further comprises:

A category directory of the video collection of the domain is established according to the existing category label.
The method according to claim 2, wherein the input parameter of the video classification model is an attribute word, and the output parameter is: a plurality of category probability values; wherein each category probability value indicates that the attribute word belongs to the category directory The probability of a category.
The method according to any one of claims 1-3, wherein the obtaining the text content and the category label of each video in the video collection of a certain domain comprises:

Obtaining a network address of each video in a video collection of a certain domain stored in the cloud server;

Obtaining a play webpage of each video according to the network address of the video;

The text content and category labels of the current video are extracted from each video play page network.
The method according to any one of claims 1-3, wherein the segmentation of the text content of each video, the set of attribute words for each video comprises:

Segmenting the text content to obtain a word segmentation result;

Performing part-of-speech tagging on each word in the segmentation result according to the part-of-speech tagging algorithm, and filtering the words in the segmentation result according to the tagging result to obtain a first-level keyword set;

According to the stop word table, the first level keyword set is filtered to obtain a set of attribute words.
A method according to any of claims 1-3, wherein the textual content comprises a title and/or a content profile of the current video.
The method according to any one of claims 1 to 3, wherein said Bayesian mode The type is a naive Bayesian model.
A video classification method, including:

Get the text content of the video to be classified;

Performing word segmentation on the text content of the classified video to obtain a set of attribute words of the video to be classified;

Entering each of the attribute words in the set of attribute words of the video to be classified into the video classification model according to any one of claims 1 to 4, and obtaining a category probability value of each attribute word of the video to be classified;

Determining, according to the category probability value of each attribute word, a category label of the to-be-categorized video in the category directory.
The video classification method of claim 8, wherein each of the attribute words includes at least one category probability value.
The video classification method according to claim 9, wherein the classifying the video to be classified according to the category probability value of each attribute word comprises:

From the plurality of category probability values of each attribute word, the one with the largest value is selected as the optimal category probability value of the attribute word;

Performing a product operation on the optimal class probability value of each attribute word in the attribute word set of the classified video to obtain a class probability of the video to be classified;

Determining, according to a category probability of the video to be classified, a category label of the to-be-categorized video in the category directory.