Research on reform and breakthrough of news, film, and television media based on artificial intelligence

1 Introduction

Artificial intelligence [1] refers to an intelligent machine that can respond in a similar way to human intelligence [2], which has been highly valued in various industries, such as intelligent healthcare [3] and smart grid [4], and has invaded all aspects of people’s lives [5]. With the rapid breakthrough and development of artificial intelligence [6], media technology has also been significantly transformed, increasingly moving in the direction of intelligence and automation. Under the influence of artificial intelligence, the media has undergone tremendous changes in production and communication [7]. Fernandez-Luque et al. [8] studied the application of artificial intelligence technology in dealing with humanitarian health crises. A search and analysis of the literature revealed that artificial intelligence had been successfully applied, for example, for epidemic detection, and no major crises were found, demonstrating the feasibility of artificial intelligence to extract valuable information. Gavai et al. [9] accurately retrieved all needed content about detecting unknown stimulants from scientific literature and articles from media with artificial intelligence and found that the method was reliable. Alattar and Shaalan [10] used an artificial intelligence approach to analyze the causes of mood changes in social media, built a filtered-latent dirichlet allocation framework, and carried out a study on Twitter to automatically extract the causes of mood changes and help decision-makers make appropriate decisions. Stolikj et al. [11] used artificial intelligence techniques to detect media piracy. They searched and identified the source of the content based on the visual information of images to combat piracy. This article studied the application of artificial intelligence in the field of news, film, and television media. Rumors about news, film, and television media were detected using the intelligent algorithm. The back-propagation neural network (BPNN) algorithm was proposed to detect rumors about news, film, and television media, and then it was compared with K-means and support vector machine (SVM) algorithms through simulation experiments. The final experimental results demonstrated that the BPNN algorithm could more accurately identify rumors. The contribution of this article is replacing manpower with artificial intelligence to detect rumors about news, film, and television media and providing an effective reference for refuting rumors.

2 Application of artificial intelligence in media

The 2016 China Internet News Market Research Report released by China Internet Network Information Center (CNNIC) shows that with the iterative update of the network, China has fully entered the mobile era, which provides a broad stage for the application of artificial intelligence. The development of China’s Internet users from June 2012 to June 2016 is shown in Figure 1.

Figure 1

Development of Internet users in China from June 2012 to June 2016.

It is seen from Figure 1 that the number of Internet users in China has been growing and the penetration rate of the Internet has also been increasing. The 48th Statistical Report on the Development Status of China’s Internet shows that as of June 2021, the size of China’s Internet users reached 1.011 billion, the penetration rate of the Internet reached 71.6%, the size of cell phone users reached 1.007 billion, the proportion of Internet users using cell phones to access the Internet was 99.6%, the scale of Internet news users reached 760 million, accounting for 75.2% of all Internet users, and the number of Internet video users was 944 million, accounting for 87.8% of all Internet users [12].

Figure 2 shows some of artificial intelligence applications in news media and film and television. The first is the application in news media, including robot writing, news recommendation, and rumor detection. Robot writing means collecting news materials with crawlers, analyzing the valuable information and completing news writing according to the pre-defined reporting format. It can output lots of news quickly but it is difficult to produce news with human feelings and dig deeper into news clues from surface events to distinguish the authenticity of the materials. Next is news recommendation, which uses intelligent algorithms to collect and analyze user information and make personalized recommendations for news according to user needs. It improves the user’s news receiving experience and helps users get the content they are interested in faster, but it may lead to being trapped in the information cocoon, i.e., receiving only the news they are interested in and losing the opportunity to access heterogeneous information, which aggravates the division of the community. Rumor detection uses intelligent algorithms to detect online opinions formed during news dissemination and checks online rumors to reduce social harm.

Figure 2

Application of artificial intelligence in news media and film and television media.

Then, it is the application of artificial intelligence in film and television media, including script writing, voice synthesis, and spam filtering [13]. Script writing is similar to news writing in news media, in which intelligent algorithms are used to edit the collected material according to a set template. It also has the advantage of fast output and the disadvantage of weak humanistic sentiment. Voice synthesis means generating a near-human voice with intelligent algorithms. The synthesized voice can replace the anchor to broadcast around the clock and improve work efficiency, but similar to robotic writing, it lacks emotion. The digital broadcast form hardly resonates with the audience, and the unchanging emotional tone cannot reflect the deeper connotation of the news. Spam filtering is related to today’s network film and television broadcast mode. Nowadays, the network film and television have the mechanism of bullet screen in the process of playback, i.e., users comment in real-time on the movies and TV shows that are playing, but in the actual use, bullet screen comments may be worthless and even untrue rumors rather than effective comments on the video. Therefore, it is essential to detect and filter these messages.

The above content briefly describes some of artificial intelligence applications in news media and film and television media. Whether it is news or film, the ultimate role of both media is the transfer of information. News is inherently informative, while film and television media transfer information through comments and bullet screens. Worthless duplicate information and untrue information may exist in transferred information; the former will affect users’ browsing experience, while the latter will affect social stability. Therefore, rumor information in news, film, and television media need to be detected and filtered. In the face of today’s big data environment, it is difficult to identify rumor information by human beings alone effectively, so the focus of this article is to reform the detection of rumor information in news media and film and television media with intelligent algorithms.

3 Detection of rumor information in news media and film and television media

In this article, artificial intelligence was applied to detect rumor information in media spreading news (e.g., Weibo) and bullet screen comments of film and TV works so as to reduce the damage of rumors to social stability, strengthen the authenticity of news media, and enhance the feeling when watching film and television works. Figure 3 shows the process of detecting rumor information in news media and film and television media. The specific steps are as follows.

Figure 3

Detection process of rumor information in news media and film and television media.

① The text data to be detected are pre-processed to remove noise and segment words [14].

② Features are extracted from the text data. As rumors should be detected and excluded as soon as possible, features that can be obtained as soon as the text is published are selected, including punctuation, part of speech (POS), emotion, emoji, common social media characters, length, and hyperlinks. In addition to the text features that can be directly counted, this article also adds two features, implication degree and pseudo-feedback, to improve the accuracy of detection, and the implication degree is calculated as follows:

where d ₁ is the information of social media, i.e., the information to be detected, d ₁ is the set of all sub-documents of a real news report divided by text granularity sliding, n is the text granularity, i.e., the number of segmented words of the information to be detected, i is the serial number of segmented words, d _1,i is the term frequency-inverse document frequency (TF-IDF) of no. i segmented word in the information to be detected, d _2,i is the TF-IDF of no. i segmented word in the sub-document set of the news report, and E(d ₁,d ₂) is the implication degree of the information to be detected.

③ The pseudo-feedback eigenvalue is calculated by the following formula:

where d ₁ is the text to be detected, d ₃ is the text in the rumor candidate set, t is the segmented word owned by d ₁ and d ₃, and tf·idf(t,d ₁) and tf·idf(t,d ₃) are the TF-IDF value of t in d ₁ and d ₃. After comparing all the texts in the rumor candidate set, if the maximum CS(d ₁,d ₃) value exceeds the set threshold, the pseudo-feedback eigenvalue PF of the text to be detected is recorded as 1; otherwise, it is recorded as 0.

④ After the features of the text to be detected are calculated, they are input to the rumor evaluation model to calculate the rumor score. There are various algorithms available for the rumor evaluation model, including SVM, neural network, clustering algorithm, etc. All of these algorithms calculate the corresponding scores based on text features. This article chose a BPNN to build the rumor evaluation model. The basic structure of the BPNN is shown in Figure 4, including the input layer, hidden layer, and output layer. Text features are input into the node of the input layer. The number of nodes in the input layer depends on the dimension of the input text features. The hidden layer is the core structure of the BPNN, and its number is set according to needs. The input text features will be processed by forward calculation in the hidden layer according to the formula, and the results can be obtained in the output layer. The formula of forward calculation in the hidden layer is:

where x _i is the text feature, a is the output of every layer, β is the adjustment term of every layer, f(·) is the activation function [13], and ω is the weight between the layers. If it is in the training phase of the evaluation model [15], the results calculated layer-by-layer are compared with the actual results, and the weight of the hidden layer is adjusted reversely according to the difference for layer-by-layer calculation again until the difference converges to stability. If it is in the application phase, the layer-by-layer calculation is performed directly.

Figure 4

The basic structure of BPNN.

⑤ Whether the score calculated by the rumor evaluation model exceeds the set threshold value is determined. If not, no treatment is done, and if it does, the text is assigned to the rumor candidate set.

4 Simulation experiments

4.5 Experimental results

Figure 5 shows the statistical results obtained after testing three different rumor evaluation models with the test set. When facing Weibo texts, the precision rate, recall rate, and F-measure of the K-means-based rumor evaluation model were 74.3, 65.8, and 69.8%, respectively; the precision rate, recall rate, and F-measure of the SVM-based rumor evaluation model were 88.4, 80.2, and 84.1%, respectively; the precision rate, recall rate, and F-measure of the BPNN-based rumor evaluation model were 98.3, 95.7, and 97.0%, respectively. When facing bullet screen texts, the precision rate, recall rate, and F-measure of the K-means-based rumor evaluation model were 64.3, 55.8, and 59.7%, respectively; the precision rate, recall rate, and F-measure of the SVM-based rumor evaluation model were 78.4, 70.2, and 74.1%, respectively; the precision rate, recall rate, and F-measure of the BPNN-based rumor evaluation model were 88.3, 85.7, and 87.0%, respectively. It was concluded that the BPNN-based rumor evaluation model had the highest recognition performance, followed by the SVM-based rumor evaluation model, and the K-means-based rumor evaluation model, both for Weibo texts and bullet screen texts. In the same rumor evaluation model, the recognition performance for Weibo texts was higher than that for bullet screen texts.

Figure 5

Recognition performance of three rumor evaluation models for Weibo texts and bullet screen texts.

Figure 6 shows the time consumption of the three rumor evaluation models for identifying Weibo texts and bullet screen texts. The time consumption of the K-means-based rumor evaluation model, the SVM-based rumor evaluation model, and the BPNN-based rumor evaluation model was 1,350 ms, 985 ms, and 573 ms, respectively, for Weibo texts. The BPNN-based rumor evaluation model took the least time, and the K-means-based rumor evaluation model took the most time for Weibo texts and bullet screen texts. The same rumor evaluation model took more time to identify Weibo texts.

Figure 6

Recognition time consumption of three rumor evaluation models for Weibo texts and bullet screen texts.

Figure 7 shows changes in TCR 5 days after applying the three rumor detection models in actual Weibo platform and bullet screen website. It is seen from Figure 7 that as time went on, the TCR of detecting Weibo texts and bullet screen texts with the three models showed a decreasing tendency, the decrease amplitude in the TCR of the BPNN-based model was the smallest, and the decrease amplitude in the TCR of the K-means-based model was the largest. After the same detection time, the TCR of the BPNN-based model was the highest, followed by the SVM-based model, and the K-means-based model. According to the rule that the larger the value of TCR is, the lower the cost of rumor detection is. The above results indicated that the longer the time of using the rumor detection model was, the larger the cost consumed by rumor detection was, but the cost increase of the BPNN-based model was slow; after the same detection time, the BPNN-based rumor detection model had a lower detection cost.

Figure 7

False alarm rates of three rumor models applied in the actual Weibo platform and bullet screen video websites within 5 days.

5 Discussion

With the development of technology, forms of news media and film and television media have also changed. For example, Weibo and WeChat have good advantages in spreading short information, which can simply be regarded as a kind of short news [16], and in film and television media, the emergence of bullet screens makes the traditional film and television works have the possibility of real-time interaction. The changes in news media and film and television video media have made the dissemination of information more rapidly, but they have also made the dissemination of false information more convenient. In the face of fast-spreading rumors in news media and film and television media, the traditional manual disinformation mode cannot respond rapidly, and a rumor often has been spread through news and film and television media platforms before it is detected and refuted. In order to refute rumors timely, artificial intelligence is introduced into the news and film and television media for reform [17]. Some relevant studies are as follows. Bi et al. [18] modeled the information transmission network of Weibo as a heterogeneous graph containing various semantic information, constructed a graph-based rumor detection model, and captured and aggregated semantic information using an attention layer. The experimental results showed that the accuracy of the detection model exceeded 92%. Wu et al. [19] studied the full convolutional neural network for image recognition and improved its detection performance by multi-scale feature fusion so that it could be applied to the recognition of rumor images more effectively. They also verified the effectiveness of the recognition algorithm through experiments. Cheng et al. [20] proposed a hierarchical rumor detection model based on generative adversarial networks and verified the model's effectiveness through experiments. Artificial intelligence algorithms are mostly imitations of human thinking. When applying intelligent algorithms to news and film and television media for rumor identification reform, the basic principle is to analyze the text features to be detected and categorize the text as rumor or non-rumor according to the features. Intelligent algorithms need sufficient training before identifying rumors, where training is either supervised or unsupervised. The training of the K-means algorithm is unsupervised, and the training of SVM and BPNN is supervised. The application of intelligent algorithms to the reform of media rumor identification not only frees up human resources but also allows them to continuously expand their identification performance with its learning characteristic during the process of use, thus reducing human interference and making the discrimination of media rumors more objective.

In this article, the rumor recognition model was established using the BPNN algorithm and compared with K-means and SVM algorithms, and the final results have shown above. The BPNN-based rumor recognition model not only had the best recognition performance but also took the least time when dealing with Weibo texts and bullet screen texts. The reason is that the unsupervised K-means algorithm only roughly divided the feature vector distance in the training process and did not dig deeper into the laws; the SVM algorithm further utilized the recognition features, but the division was linear, which was difficult to fit the non-linear laws; the BPNN was fitted to the non-linear laws based on the activation function in the hidden layer. When recognizing Weibo texts and bullet screen texts, the three rumor recognition models performed better in recognizing Weibo texts. The reason why the recognition of Weibo texts was more accurate than bullet screen texts is that the content contained by Weibo texts were completer than bullet screen texts.

6 Conclusion

This article briefly describes the application of artificial intelligence in news, film and television media, used the BPNN algorithm to reform disinformation in the news media and film and television media, and compared the BPNN algorithm with K-means and SVM algorithms. The results are as follows: (1) the BPNN-based rumor evaluation model had better recognition performance and was more advantageous in recognizing Weibo texts, (2) the BPNN-based rumor evaluation model consumed less time in recognition and was more advantageous in recognizing bullet screen texts, and (3) the BPNN-based rumor evaluation model had the lowest detection cost and remained stable when being applied in the actual Weibo platform and bullet screen video websites.