CN106446195A - News recommending method and device based on artificial intelligence - Google Patents

Abstract

The embodiment of the present invention discloses a news recommendation method and device based on artificial intelligence. The method includes: obtaining the first news feature of the news to be recommended and the second news feature of the browsed news; according to the first news feature and The second news feature determines whether the news to be recommended and the browsed news are similar news; if it is similar news, then refuse to recommend the news to be recommended; if it is non-similar news, then recommend the news to be recommended . In the embodiment of the present invention, by determining whether the news to be recommended and the browsed news are similar news, only non-similar news is recommended to the user, which can effectively avoid repeatedly recommending similar news to the user, and improve the efficiency of news recommendation.

Description

News recommendation method and device based on artificial intelligence

technical field

The embodiments of the present invention relate to the technical field of information processing, and in particular to an artificial intelligence-based news recommendation method and device.

Background technique

Artificial Intelligence (AI) is a new technical science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence. Artificial intelligence is a branch of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a manner similar to human intelligence. Research in this field includes robotics, language recognition, image recognition, natural language processing and expert systems, etc.

In the era of rapid information development, with the development of Internet technology, more and more news information has entered the public eye, and news-related products have also appeared in people's daily life, which are closely related to people's daily life. For example, common news-related products include "Today's Headlines", "Baidu Kankan", "Mobile Baidu Information", etc. They mainly mine news from various news websites, and recommend news to users according to the interests of different users and the popularity and attention of the news itself.

Due to the wide range of news sources, it is inevitable that there will be similar news. For example, website A creates an original piece of news, website B reprints it, website C reprints it after slightly modifying it, and website D creates another piece of news based on the same news event. The four pieces of news are actually the same event. The news recommendation products will dig out the difference, and recommend them to users without knowing the difference. However, users often know the news event after reading a piece of news, and there is no need to see the same news again. Especially for those similar news with different titles, users often only know that they are similar news after reading the news, resulting in a waste of user time.

Contents of the invention

Embodiments of the present invention provide an artificial intelligence-based news recommendation method and device, which can avoid repeated recommendation of similar news to users, so as to improve news recommendation efficiency.

In the first aspect, the embodiment of the present invention provides a news recommendation method based on artificial intelligence, including:

Obtaining the first news feature of the news to be recommended and the second news feature of the browsed news;

determining whether the news to be recommended and the browsed news are identical news according to the first news feature and the second news feature;

If it is the same news, refuse to recommend the news to be recommended; if it is not the same news, recommend the news to be recommended.

In the second aspect, the embodiment of the present invention also provides a news recommendation device based on artificial intelligence, including:

A feature acquisition module, configured to acquire the first news feature of the news to be recommended and the second news feature of the browsed news;

A similarity determination module, configured to determine whether the news to be recommended and the browsed news are similar news according to the first news feature and the second news feature;

The news recommendation module is configured to refuse to recommend the news to be recommended if it is similar news; to recommend the news to be recommended if it is non-similar news.

The embodiment of the present invention provides a method for news recommendation based on artificial intelligence. By determining whether the news to be recommended and the browsed news are similar news, only non-similar news is recommended to the user, which can effectively avoid repeatedly recommending similar news to the user. News to improve the efficiency of news recommendation.

Description of drawings

Fig. 1 is a flow chart of a news recommendation method based on artificial intelligence in Embodiment 1 of the present invention;

FIG. 2 is a flow chart of an artificial intelligence-based news recommendation method in Embodiment 2 of the present invention;

FIG. 3 is a flow chart of an artificial intelligence-based news recommendation method in Embodiment 3 of the present invention;

FIG. 4 is a flow chart of an artificial intelligence-based news recommendation method in Embodiment 4 of the present invention;

FIG. 5 is a flow chart of an artificial intelligence-based news recommendation method in Embodiment 5 of the present invention;

FIG. 6 is a structural diagram of an artificial intelligence-based news recommendation device in Embodiment 6 of the present invention;

FIG. 7 is a structural diagram of an artificial intelligence-based news recommendation device in Embodiment 7 of the present invention.

detailed description

The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, but not to limit the present invention. In addition, it should be noted that, for the convenience of description, only some structures related to the present invention are shown in the drawings but not all structures.

Embodiment one

Figure 1 is a flow chart of an artificial intelligence-based news recommendation method provided by Embodiment 1 of the present invention. This embodiment is applicable to various news recommendation situations, and the method can be executed by the news recommendation device provided by the embodiment of the present invention. , the device can be implemented in the form of software and/or hardware, and the device can be integrated into any device that provides news recommendation functions, such as a typical user terminal device, which can be a computer, or a mobile terminal (such as a mobile phone), Tablet PCs, etc., as shown in Figure 1, specifically include:

S110. Obtain the first news feature of the news to be recommended and the second news feature of the browsed news.

Among them, news refers to reports of recent facts, and is a general term for various news styles such as news, communications, close-ups, sketches, reportage, etc. News in a narrow sense refers to news. The first news feature of the news to be recommended can be the title, text, picture, video, etc. of the news to be recommended, which can report some key information of the news to be recommended in detail; similarly, the second news feature of the browsed news can also be the Browse the headlines, texts, pictures, videos, etc. of the news, and be able to report some key information of the news in detail. When obtaining the first news feature of the news to be recommended is the title of the recommended news, then obtaining the second news feature of the browsed news is also the title of the browsed news; when obtaining the first news feature of the news to be recommended is the text of the recommended news content, then the second news feature of the browsed news is also the text content of the browsed news; when the first news feature of the news to be recommended is the video of the recommended news, the second news feature of the browsed news is also It is the video of the viewed news. The executor for acquiring the above news features may be a search engine in a webpage, or a news application software contained in a mobile terminal.

S120. Determine whether the news to be recommended and the browsed news are identical news according to the first news feature and the second news feature.

Among them, the same news may be news with exactly the same news content, or different reports on the same news event. Although the text expressions are different, the substantive content is the same. The first news feature and the second news feature may be titles, text content, pictures or videos, and the like. It is determined whether the news to be recommended and the browsed news are identical news according to the first news feature and the second news feature. The method for judging whether it is similar news can be Min Hash algorithm, Shingling algorithm, Sim Hash de-duplication algorithm, or constructing a training model (such as a neural network model), etc.

Specifically, if it is determined that the news to be recommended and the news that has been browsed are identical news, then step S130 is performed; if it is determined that the news to be recommended and the news that has been browsed are not identical news, then step S140 is performed.

S130. Refuse to recommend the news to be recommended.

S140. Recommend the news to be recommended.

Specifically, if the news to be recommended and the news that has been browsed are the same news, it means that the news to be recommended and the news that has been browsed are repetitive news or similar news, then the recommended news should be rejected; but if the news to be recommended and the news that has been viewed The browsed news is non-similar news, indicating that the repeatability of the news to be recommended and the news that has been browsed is not high, so the news to be recommended can be recommended.

For example, user A opens the news application software in the mobile phone to browse news, and the homepage of the news application software will recommend the latest and hottest news content closely related to user A's life or related to user A's hobbies for user A to choose to watch. When user A clicks on the news content of "2016 National Day Holiday Arrangement", then the title of "2016 National Day Holiday Arrangement" is used as the second feature information of the browsed news. When the news to be recommended is "How long is the National Day holiday in 2016?", the title of the news to be recommended is obtained as the first news feature. Since the two titles are similar, it is determined that the news to be recommended and the news that has been viewed are the same news, so the news application software will no longer recommend the news of "How long is the National Day holiday in 2016?" for user A.

Since the rise of the Internet has brought great challenges to traditional media such as newspapers, TV, and magazines, Internet news breaks through the traditional concept of news and gives users a brand new experience in terms of sight and feel. The same piece of news not only contains text and pictures, but also video, audio, and online comments, etc., presented in various ways. Internet news satisfies users' needs for information in the information age, but it also prevents users from obtaining the part of information that is really useful to them when faced with a large amount of information, and the efficiency of using information is reduced instead. Usually, the best solution to this problem is The conventional method is news recommendation. However, there are a large number of nearly repeated or completely repeated news on the Internet. Therefore, after a user has browsed a news, whether it is a news browsed a day ago or a news browsed a month ago, when recommending news, it is necessary to The recommended news is detected, and the first news feature of the news to be recommended is compared with the second news feature of the browsed news to avoid repeated recommendation of the news.

In this embodiment, by obtaining the first news feature of the news to be recommended and the second news feature of the news that has been browsed, it is determined whether the news to be recommended and the news that has been browsed are similar news, which can effectively avoid repeatedly recommending similar news to users, so as to improve News recommendation efficiency.

Embodiment two

Fig. 2 is a flow chart of an artificial intelligence-based news recommendation method provided by Embodiment 2 of the present invention. This embodiment is optimized on the basis of the above-mentioned embodiments, and provides optimized information based on the first news feature and the first news feature. The processing method of determining whether the news to be recommended and the browsed news are similar news by the second news feature is specifically: inputting the first news feature and the second news feature into a neural network model for learning; The learning result determines whether the news to be recommended and the browsed news are identical news.

Correspondingly, the method of this embodiment includes:

S210. Obtain the first news feature of the news to be recommended and the second news feature of the browsed news.

S220. Input the first news feature and the second news feature into a neural network model for learning.

Among them, the neural network is an attempt to imitate the mode of transmitting and processing information between neurons in the brain. The neural network is divided into two processes, including the training phase of the neural network model and the learning phase using the neural network model. Typical neural network models include perceptrons, linear neural networks, BP (Back Propagation) networks, deep learning network models, etc. It is determined that the news to be recommended is similar to the news that has been browsed. The existing model can be used for learning, and the existing model can also be improved on the basis of the existing model for learning. In general, neural network models have two characteristics: including neurons and the strength of information transmission between neurons. The construction of the neural network model is divided into three parts, including the input layer, hidden layer and output layer. On this basis, in the training phase, the neural network model of this embodiment needs to rely on a large number of known similar news data and/or to be trained by known non-identical news data. The first news feature and the data of the browsed news, that is, the second news feature, are input to the trained neural network model for learning, and the learning result is output at the output layer. According to the learning result, it can be determined whether the news to be recommended and the news that has been browsed are similar news.

S230. Determine whether the news to be recommended and the browsed news are identical news according to the learning result.

Wherein, the learning result output by the neural network model may be text, value, score or feature information about news. If the output learning result is text, for example, directly outputting similar news and non-similar news, it can be determined whether the news to be recommended and the browsed news are similar news according to the text;

If the output learning result is a score, then determine whether the news to be recommended and the browsed news are similar news according to the output score, for example, if the score is higher than the preset score, then determine the news to be recommended It is the same news as the browsed news, otherwise, it is not the same news;

If the output learning result is a numerical value, then determine whether the news to be recommended and the news that has been browsed are similar news according to the numerical value of the output, for example, if the value is 1, then determine that the news to be recommended and the news that has been browsed are For similar news, if the value is -1, it is determined that the news to be recommended and the browsed news are non-similar news, and so on.

If it is determined that the news to be recommended and the news that has been browsed are identical news, execute step S240, and if it is determined that the news to be recommended and the news that has been viewed are non-identical news, execute step S250.

S240. Refusing to recommend the news to be recommended.

S250. Recommend the news to be recommended.

For example, the above-mentioned user A opens the news application software in the mobile phone to browse news. When the title of "2016 National Day holiday arrangement" is used as the second characteristic information of the browsed news, and the title of "2016 National Day holiday how many days?" is used as the first news feature to obtain news to be recommended, the two titles " 2016 National Day Holiday Arrangement" and "2016 National Day Holiday How many days?" were input into the neural network model for learning, and if the results are determined to be similar news, then refuse to recommend the title "2016 National Day holiday how many days?" News, if it is determined that the result is not the same news, then recommend the news titled "How many days are off for the National Day in 2016?".

This embodiment determines whether it is the same news by inputting the first news feature and the second news feature into the neural network model for learning. Since the neural network model has the function of simulating the information processing of the human brain, it can improve the judgment of similar news. The accuracy and efficiency of news recommendation.

Embodiment Three

Fig. 3 is a flow chart of an artificial intelligence-based news recommendation method provided by Embodiment 3 of the present invention. This embodiment is optimized on the basis of the above-mentioned embodiments, and provides an optimized method of determining the news to be recommended and the news to be recommended according to the learning results. The processing method of whether the browsed news is the same news is specifically: if the learning result meets the preset condition, then determine that the news to be recommended and the browsed news are the same news; if the learning result does not meet the preset condition conditions, it is determined that the news to be recommended and the browsed news are non-similar news.

Correspondingly, the method of this embodiment includes:

S310. Obtain the first news feature of the news to be recommended and the second news feature of the browsed news.

S320. Input the first news feature and the second news feature into a neural network model for learning.

S330. Determine whether the learning result satisfies a preset condition.

Wherein, the preset condition includes at least one of the following: the score output by the neural network model is higher than the preset score, the title is consistent, the text similarity is higher than the preset threshold, and the source is the same. If the learning result satisfies the preset condition, determine that the news to be recommended and the news that has been viewed are the same news, execute step S340, and if the learning result does not meet the preset condition, determine that the news to be recommended and the news that has been browsed If the news is non-similar news, go to step S350.

Specifically, when the score output by the neural network model is higher than the preset score, it is determined that the news to be recommended and the browsed news are the same news. Dynamic value set by demand;

When the titles of the news to be recommended and the news that have been browsed are consistent, it is determined that the news to be recommended and the news that has been browsed are identical news, otherwise they are non-identical news;

When the similarity of the text of the news to be recommended is higher than the preset threshold, it is determined that the news to be recommended and the news that has been browsed are similar news, otherwise it is non-similar news. Dynamic value set by demand;

When the sources of the news to be recommended and the news that has been browsed are the same, it is determined that the news to be recommended and the news that has been browsed are identical news, otherwise they are non-identical news.

For example, the preset condition is that the text similarity is higher than the preset threshold as an example for detailed description. If the text similarity is higher than the preset threshold, it is determined that the news to be recommended and the browsed news are the same news. Among them, the numerical value corresponding to the similarity degree of news with completely identical texts can be set to 100, the numerical value corresponding to the similarity degree of news with completely non-identical texts can be set to 0, and the preset threshold can be customized to 70. When the text similarity of "2016 National Day holiday arrangement" is used as the second feature information of the browsed news, and the text similarity of "2016 National Day holiday?" is used as the first news feature of the news to be recommended, the The two texts "2016 National Day holiday arrangement" and "2016 National Day holiday how many days? If you have browsed the same news, then refuse to recommend the news titled "How long is the National Day holiday in 2016?".

S340. Refusing to recommend the news to be recommended.

S350. Recommend the news to be recommended.

In this embodiment, by setting predetermined conditions, it is determined whether the news to be recommended and the browsed news are identical news, so that the news recommendation is more accurate.

Embodiment Four

FIG. 4 is a flow chart of an artificial intelligence-based news recommendation method provided by Embodiment 4 of the present invention. This embodiment is optimized on the basis of the above-mentioned embodiments, and provides a combination of the first news feature and the second news feature. The processing method of inputting the news features into the neural network model for learning is specifically: determining whether the news to be recommended and the browsed news are similar news; Similarity standard, then adopt the first kind of similarity training sub-model in neural network model to learn described first news feature and described second news feature, if described judgment similarity standard is the second kind of similarity standard, then adopt neural network model The second similarity training sub-model in the network model learns the first news feature and the second news feature.

Correspondingly, the method of this embodiment includes:

S410. Obtain the first news feature of the news to be recommended and the second news feature of the browsed news.

S420. Determine a similarity criterion adopted for judging whether the news to be recommended and the browsed news are similar news.

Specifically, determining whether the news to be recommended and the browsed news are identical news may also use a similarity criterion. Among them, the similarity judgment standard can have different judgment standards, including the first type of similarity standard and the second type of similarity standard, according to which judgment similarity standard is more in line with the news to be recommended and the browsed news, and then the recommended news is treated according to the similarity judgment standard determination. Among them, the first type of similarity standard refers to the same news event and the same news source, which is subjected to any of the following operations by other news websites: reprinting, piracy, and adaptation; the second type of similarity standard refers to the same news event, which is used by different media Do any of the following: report, write original, and add commentary from your own media.

Which standard to choose as the judging standard can be customized by the user. Specifically, standard selection options can be set on the news recommendation device provided in the embodiment of the present invention for users to choose.

If the adopted standard for determining similarity is the first type of similarity standard, execute step S430, and if the adopted standard for determining similarity is the second type of identical standard, execute step S440.

S430. Use the first similarity training sub-model in the neural network model to learn the first news feature and the second news feature.

S440. Use the second similarity training sub-model in the neural network model to learn the first news features and the second news features.

Among them, the first type of similarity standard refers to the same news event and the same news source, which are reproduced intact by other news websites, pirated, added supplementary content such as advertisements on the basis of the original text, or slightly processed and adapted the content, etc. . Much of the content of this type of news is exactly the same. We think it is enough for all users to read this kind of news once, and will not show it twice. If it is determined that the similarity standard is the first type of similarity standard, then the first type of similarity training sub-model in the neural network model is used to learn the first news feature and the second news feature. Among them, the first type of similar training sub-model is a training model based on deep learning technology.

The second type of similarity criteria is the same news event but different news sources. Specifically, it refers to reporting on the same news event by different media, with original and additional media’s own comments. The content of this type of news is generally not exactly the same, but it describes one thing. Different users of this kind of news have different needs. Some users do not want to display it again after reading it, and some users want to read reports of the same event in different media. Determining whether the news to be recommended and the browsed news are identical news may also use a similarity criterion. If it is determined that the similarity standard is the second type of similarity standard, then the first news feature and the second news feature are learned by using the second type of similarity training sub-model in the neural network model. Among them, the second type of similar training sub-model is also a training model based on deep learning technology.

S450. Determine whether the news to be recommended and the browsed news are identical news according to the learning result.

If it is determined that the news to be recommended and the news that has been browsed are identical news, execute step S460, and if it is determined that the news to be recommended and the news that has been viewed are non-identical news, execute step S470.

Specifically, if the first type of similarity standard is used, the first type of similarity training sub-model in the neural network model is used to learn the first news feature and the second news feature, if determined according to the learning result If the to-be-recommended news and the browsed news are identical news, step S460 is performed, and if it is determined according to the learning result that the to-be-recommended news and the browsed news are not identical news, then step S470 is performed.

Or, if the second type of similarity standard is adopted, the first news feature and the second news feature are learned through the second type of similarity training sub-model in the neural network model. The news to be recommended and the news that has been browsed are similar news, and it is detected that the recommendation setting item is to reject the recommendation, then perform step S460; if it is determined according to the learning result that the news to be recommended and the news that has been browsed are non-similar news, Or, if it is determined according to the learning result that the news to be recommended and the browsed news are the same news, and it is detected that the recommendation setting item is not set to reject recommendation, then step S470 is performed.

S460. Refusing to recommend the news to be recommended.

S470. Recommend the news to be recommended.

In this embodiment, by selecting a similarity determination criterion, it is determined whether the news to be recommended and the browsed news are similar news, so that the news recommendation is more accurate.

In addition, in order to ensure more accurate results, the first type of similarity standard and the second type of similarity standard can also be used to determine whether the news to be recommended and the browsed news are similar news, that is, by using the neural network model at the same time The first type of similarity training sub-model and the second type of similarity training sub-model respectively learn the first news feature and the second news feature, and determine whether the news to be recommended and the browsed news are similar news according to the learning results.

In this embodiment, by selecting similarity determination criteria, it is determined whether the news to be recommended and the browsed news are similar news according to different similarity determination criteria, so that the news recommendation is more accurate.

Embodiment five

FIG. 5 is a flow chart of an artificial intelligence-based news recommendation method provided by Embodiment 5 of the present invention. On the basis of the above embodiments, this embodiment further includes: obtaining known similar news and/or known non-identical news; Construct a first training sample of known similar news according to the third news feature of the known similar news, and/or construct a second training sample of known non-identical news according to the fourth news feature of the known non-identical news ; using a neural network to train the first training sample and/or the second training sample to obtain the neural network model.

Correspondingly, the method of this embodiment includes:

S510. Obtain known similar news and/or known non-similar news.

Specifically, the neural network model is a training model based on a deep learning algorithm for deduplication of similar news. The neural network model is constructed, and the neural network model for effectively judging whether the news to be recommended is similar news is obtained by analyzing and training known similar news and known non-similar news.

For example, crawler technology can be used to mine news data from the Internet, determine similar news and non-similar news according to the news data, and use the determined similar news and non-similar news as known similar news and known non-similar news respectively.

S520. Construct a first training sample of known similar news according to the third news feature of the known similar news, and/or construct a second training sample of known non-identical news according to the fourth news feature of the known non-identical news. Training samples.

Specifically, for known similar news, extract the news features of known similar news, that is, the third news feature including but not limited to title, text and source, as the first training sample, and mark the training sample, that is, tell the neural network the The training samples are the training samples of the same news. Similarly, for known non-similar news, extract the news features of known non-similar news, that is, the fourth news feature, including but not limited to title, text and source, as the second training sample, and mark the training sample, that is, inform the neural network The training samples of the network are training samples of non-similar news. .

S530. Use a neural network to train the first training sample and/or the second training sample to obtain the neural network model.

Specifically, the final neural network model is obtained by repeatedly training the first training sample and the second training sample. Among them, the first training sample conducts separate and multiple trainings on features such as titles, text content, pictures or videos that have been similar to news; the second training sample performs training on features such as titles, text content, pictures or videos that have not been similar Separate and multiple training sessions.

S540. Obtain the first news feature of the news to be recommended and the second news feature of the browsed news.

S550. Input the first news feature and the second news feature into a neural network model for learning.

S560. Determine whether the news to be recommended and the browsed news are identical news according to the learning result.

If it is determined that the news to be recommended and the news that has been browsed are identical news, execute step S570, and if it is determined that the news to be recommended and the news that has been viewed are not identical news, execute step S580.

S570. Refusing to recommend the news to be recommended.

S580. Recommend the news to be recommended.

In this embodiment, since the joint training of news headlines and texts is carried out in the constructed deep neural network, it does not only rely on the coincidence degree of the headlines or texts, which reduces the repetition rate of news recommendation.

On the basis of the foregoing embodiments, the neural network model may include a first identical training sub-model and/or a second identical training sub-model. Wherein, the training process of the first similar training sub-model includes:

Using crawler technology to dig out news clusters that have at least one relationship among reprinting, piracy, and adaptation, and use the news clusters as the first known similar news;

Use news that does not have at least one relationship among reprinting, piracy, and adaptation as the first known non-identical news;

Using a neural network to train the training samples of the first known similar news and/or the training samples of the first known non-similar news to obtain a first similar training sub-model in the neural network model;

Crawler technology is used to mine news that has a relationship among reprinting, piracy and adaptation, and this news is regarded as the first known similar news. For example, if A is reposted by B, and B is reposted by C, then A, B, and C are all pairs with a reprint relationship. You can find out where a certain news is reprinted by clearly indicating the "source" and its link in the news .

Similarly, news that does not have at least one relationship among reprinting, piracy, and adaptation is taken as the first known non-identical news. From the entire news collection mined by crawler technology, two news samples are randomly selected. If these two news are not the first known identical news, we consider such news pair as the first known non-identical news. Correspondingly, the first similarity training sub-model in the neural network model is trained from the first known training samples of similar news and the first known training samples of non-similar news. For example, the crawler technology mines all the news about "2016 National Day Holiday Arrangements", and groups the news that belongs to the same news content and the same news source as the first known similar news; This news is not the first known identical news as the first known non-identical news. Finally, the first similar training sub-model in the neural network model about the news "2016 National Day holiday arrangements" is obtained through training.

Wherein, the training process of the second identical training sub-model includes:

Using crawler technology to dig out news combinations whose matching degree of news titles, content and release time exceeds the preset level, and use said news combination as the second known similar news;

Taking the news whose matching degree is less than the preset degree as the second known non-identical news;

A neural network is used to train the training samples of the second known similar news and/or the training samples of the second known non-similar news to obtain a second similar training sub-model in the neural network model.

Use web crawler technology to mine some news, use the title and content of this part of the news to search the search engine, and find similar news through the results returned by the search engine. If the matching degree of two news release times exceeds the preset degree, the content of the two news contents has a high degree of literal overlap, or the news combination of news titles, we consider the news combination as the second known similar news. Wherein, the preset degree in news matching degree exceeding the preset degree may be a default static value, or may be a dynamic value set according to individual needs.

The second similarity training sub-model in the neural network model is trained from the second known training samples of similar news and the second known training samples of non-similar news. For example, the crawler technology mines all the news about "2016 National Day holiday arrangements", and combines the news belonging to the same news event, similar release time, and different news sources as the second known similar news; The two news that came out were not the second known identical news as the second known non-identical news. Finally, the training obtains the second identical training sub-model in the neural network model about the news "2016 National Day holiday arrangement".

This embodiment has described in detail the construction of the first similarity training sub-model and the second similarity training sub-model in the neural network model. This kind of targeted and supervised training for the special application of news similarity can improve News recommendation efficiency.

Embodiment six

FIG. 6 is a schematic structural diagram of an artificial intelligence-based news recommendation device provided by Embodiment 6 of the present invention. This embodiment is applicable to various news recommendation situations. The method can be executed by the news recommendation device provided in the embodiment of the present invention. The device can be implemented in the form of software and/or hardware. The device can be integrated in any news recommendation Functional devices, such as typical user terminal devices, can be computers, mobile terminals (such as mobile phones), tablet computers, etc., as shown in Figure 6, specifically include: feature acquisition module 61, similarity determination module 62 and News recommendation module63.

The feature acquisition module 61 is used to acquire the first news feature of the news to be recommended and the second news feature of the browsed news;

The similarity determination module 62 is used to determine whether the news to be recommended and the browsed news are similar news according to the first news feature and the second news feature;

The news recommendation module 63 is configured to refuse to recommend the news to be recommended if it is similar news; to recommend the news to be recommended if it is non-similar news.

The news recommendation device in this embodiment is used to implement the news recommendation methods described in the above-mentioned embodiments, and its technical principles and technical effects are similar, and will not be repeated here.

Embodiment seven

FIG. 7 is a schematic structural diagram of an artificial intelligence-based news recommendation device provided by Embodiment 7 of the present invention. As shown in Figure 7:

On the basis of the above embodiments, the similarity determining module 62 preferably includes a learning unit 71 and a similarity determining unit 72 .

The learning unit 71 is used for inputting the first news feature and the second news feature into the neural network model for learning;

The similarity determining unit 72 is configured to determine whether the news to be recommended and the browsed news are similar news according to the learning result.

On the basis of the above-mentioned embodiments, the similarity determining unit 72 is specifically configured to: if the learning result satisfies the preset condition, then determine that the news to be recommended and the browsed news are similar news; if the learning result does not meet the preset condition conditions, then determine that the news to be recommended and browsed news are non-similar news

On the basis of the above embodiments, the preset conditions of the device include at least one of the following: the score output by the neural network model is higher than the preset score, the title is consistent, the similarity of the text is higher than the preset threshold and the source same.

On the basis of the above-mentioned embodiments, the learning unit 71 is specifically configured to: determine whether the news to be recommended and the browsed news are identical news; standard, then the first type of similarity training sub-model in the neural network model is used to learn the first news feature and the second news feature; or, if the judgment similarity standard is the second type of similarity standard, use The second similarity training sub-model in the neural network model learns the first news features and the second news features.

On the basis of the above embodiments, the first type of similarity standard is the same news event and the same news source, and any of the following operations are performed by other news websites: reprinting, piracy, and adaptation; and/or, the second type The same standard refers to the same news event, which is performed by any of the following operations by different media: reporting, originality, and adding comments from its own media.

On the basis of the above-mentioned embodiments, the news recommendation module 63 is specifically configured to: if it is determined according to the first similarity standard that the news to be recommended and the news that has been browsed are similar news, then directly refuse to recommend the news to be recommended. Recommending news; or, if it is determined according to the second similarity standard that the news to be recommended and the news that has been browsed are similar news, and it is detected that the recommendation setting item is rejection of recommendation, then refuse to recommend the news to be recommended.

On the basis of the above embodiments, the device further includes a news acquisition module 73 , a sample construction module 74 and a neural network training module 75 .

The news acquisition module 73 is used to acquire known similar news and/or known non-identical news;

The sample construction module 74 is used to construct the first training sample of known similar news according to the third news feature of the known similar news, and/or construct the known non-identical news according to the fourth news feature of the known non-identical news. A second training sample of news;

The neural network training module 75 is configured to use a neural network to train the first training sample and/or the second training sample to obtain the neural network model.

On the basis of the above embodiments, the output result of the neural network model satisfies the following condition: the score of the news to be recommended and the browsed news is higher than that of non-similar news.

On the basis of the above-mentioned embodiments, the news acquisition module 61 is specifically configured to: use crawler technology to mine news clusters that have at least one relationship among reprinting, piracy, and adaptation, and use the news clusters as the first known Identical news; news that does not have at least one of reprinting, piracy and adaptation as the first known non-identical news;

Correspondingly, the neural network training module 75 is specifically configured to: use a neural network to train the training samples of the first known similar news and/or the training samples of the first known non-identical news to obtain the the first identical training sub-model in the neural network model;

And/or, the news acquisition module 61 is specifically used to: use crawler technology to dig out news combinations whose matching degree of news titles, content and release time exceeds a preset level, and use the news combination as the second known similar news; News whose matching degree does not exceed the preset level is regarded as the second known non-similar news;

Correspondingly, the neural network training module 75 is specifically configured to: use a neural network to train the training samples of the second known similar news and/or the training samples of the second known non-identical news to obtain the The second identically trained submodel in the neural network model.

The news recommendation device described in this embodiment is used to implement the news recommendation methods described in the above embodiments, and its technical principles and technical effects are similar, and will not be repeated here.

Note that the above are only preferred embodiments of the present invention and applied technical principles. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the protection scope of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and can also include more other equivalent embodiments without departing from the concept of the present invention, and the present invention The scope is determined by the scope of the appended claims.

Claims (20)

1. A news recommendation method based on artificial intelligence, characterized in that, comprising: Obtaining the first news feature of the news to be recommended and the second news feature of the browsed news; determining whether the news to be recommended and the browsed news are identical news according to the first news feature and the second news feature; If it is the same news, refuse to recommend the news to be recommended; if it is not the same news, recommend the news to be recommended. 2. The method according to claim 1, wherein determining whether the news to be recommended and the browsed news are similar news according to the first news feature and the second news feature includes: Inputting the first news feature and the second news feature into a neural network model for learning; It is determined whether the news to be recommended and the browsed news are identical news according to the learning result. 3. The method according to claim 2, wherein determining whether the news to be recommended and the browsed news are identical news according to the learning result comprises: If the learning result satisfies the preset condition, it is determined that the news to be recommended and the browsed news are identical news; if the learning result does not meet the preset condition, then it is determined that the news to be recommended and the browsed news are non-identical news. 4. The method according to claim 3, wherein the preset conditions include at least one of the following: The score output by the neural network model is higher than a preset score, the title is consistent, the text similarity is higher than a preset threshold, and the source is the same. 5. method according to claim 2, is characterized in that, said first news feature and described second news feature input in neural network model and learning comprises: determining the same criteria for judging whether the news to be recommended and the news viewed are similar news; If the determination similarity standard is the first type of similarity standard, then use the first type of similarity training sub-model in the neural network model to learn the first news feature and the second news feature; and/or, if If the similarity criterion is determined as the second type similarity standard, then the second type similarity training sub-model in the neural network model is used to learn the first news feature and the second news feature. 6. The method according to claim 5, characterized in that, the first type of similarity standard is the same news event and the same news source, and any one of the following operations is performed by other news websites: reprinting, piracy and adaptation; And/or, the second type of similarity standard is the same news event, which is performed by different media in any of the following operations: report, create original content, and add comments from its own media. 7. The method according to claim 5 or 6, wherein if it is the same news, refusing to recommend the news to be recommended comprises: If it is determined according to the first type of similarity standard that the news to be recommended and the news that has been browsed are similar news, then directly refuse to recommend the news to be recommended; Or, if it is determined according to the second similarity standard that the news to be recommended and the news that has been browsed are similar news, and it is detected that the recommendation setting item is rejection of recommendation, then refuse to recommend the news to be recommended. 8. The method according to any one of claims 2-6, further comprising: Obtain known identical news and/or known non-identical news; Construct a first training sample of known similar news according to the third news feature of the known similar news, and/or construct a second training sample of known non-identical news according to the fourth news feature of the known non-identical news ; Using a neural network to train the first training sample and/or the second training sample to obtain the neural network model. 9. method according to claim 8, is characterized in that, the output result of described neural network model satisfies the following conditions: The news to be recommended and the browsed news have higher scores for similar news than non-similar news. 10. The method according to claim 8 or 9, wherein obtaining known similar news and/or known non-identical news comprises: Using crawler technology to dig out news clusters that have at least one relationship among reprinting, piracy, and adaptation, and use the news clusters as the first known similar news; Use news that does not have at least one relationship among reprinting, piracy, and adaptation as the first known non-identical news; Correspondingly, using a neural network to train the first training sample and/or the second training sample, and obtaining the neural network model includes: Using a neural network to train the training samples of the first known similar news and/or the training samples of the first known non-similar news to obtain a first similar training sub-model in the neural network model; and / or, Using crawler technology to dig out news combinations whose matching degree of news titles, content and release time exceeds the preset level, and use said news combination as the second known similar news; Taking the news whose matching degree is less than the preset degree as the second known non-identical news; Correspondingly, using a neural network to train the first training sample and/or the second training sample, and obtaining the neural network model includes: A neural network is used to train the training samples of the second known similar news and/or the training samples of the second known non-similar news to obtain a second similar training sub-model in the neural network model. 11. A news recommendation device based on artificial intelligence, characterized in that it comprises: A feature acquisition module, configured to acquire the first news feature of the news to be recommended and the second news feature of the browsed news; A similarity determination module, configured to determine whether the news to be recommended and the browsed news are similar news according to the first news feature and the second news feature; The news recommendation module is configured to refuse to recommend the news to be recommended if it is similar news; to recommend the news to be recommended if it is non-similar news. 12. The device according to claim 11, wherein the identity determination module specifically comprises: a learning unit, configured to input the first news feature and the second news feature into a neural network model for learning; A similarity determination unit is configured to determine whether the news to be recommended and the browsed news are similar news according to a learning result. 13. The device according to claim 12, wherein the similarity determining unit is specifically configured to: if the learning result satisfies a preset condition, then determine that the news to be recommended and the browsed news are similar news; if If the learning result does not satisfy the preset condition, then it is determined that the news to be recommended and the browsed news are not identical news. 14. The device according to claim 13, wherein the preset conditions include at least one of the following: the score output by the neural network model is higher than the preset score, the titles are consistent, and the text similarity is higher than The preset thresholds are the same as the source. 15. The device according to claim 12, wherein the learning unit is specifically used for: Determining whether the news to be recommended and the browsed news are the similarity criteria adopted for similar news; if the similarity criterion is the first type of similarity standard, the first type of similarity training subclass in the neural network model is used The model learns the first news feature and the second news feature; and/or, if the judgment similarity standard is the second type of similarity standard, the second type of similarity training sub-model in the neural network model is used to pair The first news feature and the second news feature are learned. 16. The device according to claim 15, wherein the first type of similarity standard is the same news event and the same news source, and any of the following operations are performed by other news websites: reprinting, piracy and adaptation; And/or, the second type of similarity standard is the same news event, which is performed by different media in any of the following operations: report, create original content, and add comments from its own media. 17. The device according to claim 15 or 16, wherein the news recommendation module is specifically configured to: if the news to be recommended and the browsed news are determined to be similar news according to the first similarity standard , then directly refuse to recommend the news to be recommended; or, if it is determined according to the second similarity standard that the news to be recommended and the browsed news are similar news, and it is detected that the recommendation setting item is rejection recommendation, then reject Recommend the news to be recommended. 18. The device according to any one of claims 12-16, further comprising: A news acquisition module for acquiring known similar news and/or known non-identical news; The sample construction module is used to construct the first training sample of known similar news according to the third news feature of the known similar news, and/or construct the known non-identical news according to the fourth news feature of the known non-identical news A second training sample of news; A neural network training module, configured to use a neural network to train the first training sample and/or the second training sample to obtain the neural network model. 19. The device according to claim 18, wherein the output result of the neural network model satisfies the following condition: the news to be recommended and the browsed news have higher scores for similar news than for non-similar news score. 20. The device according to claim 18 or 19, wherein the news acquisition module is specifically configured to: use crawler technology to mine news clusters that have at least one relationship among reprinting, piracy and adaptation, and collect the news News gathering is the first known identical news; news that does not have at least one relationship among reprinting, piracy and adaptation is the first known non-identical news; Correspondingly, the neural network training module is specifically configured to: use a neural network to train the training samples of the first known similar news and/or the training samples of the first known non-identical news to obtain the neural network. The first identical training sub-model in the network model; and / or, The news acquisition module is specifically used to: use crawler technology to dig out news combinations whose matching degree of news titles, content and release time exceeds a preset degree, and use the news combination as the second known similar news; News of a preset degree as the second known non-similar news; Correspondingly, the neural network training module is specifically configured to: use a neural network to train the training samples of the second known similar news and/or the training samples of the second known non-identical news to obtain the neural network The second identically trained sub-model in the network model.