Media of the future: what neural networks can already be used in the media today

The main articles are:

• Neural networks (neural networks)
• Artificial intelligence for writing texts in the media and literature

Generating Images Based on Material Context

Images in news or social media posts attract attention. Text without pictures, no matter how interesting, is less read according to statistics than text with any picture. Therefore, journalists spend minutes and hours daily looking for thematic images in photo banks. Such photos are not directly related to the news and are uploaded only for beauty. Why not instruct the neural network to generate thematic images to news or posts?

Using neural networks to generate images can significantly reduce the time spent on selecting pictures for articles.

However, without experience in working with generative networks, the journalist's time spent will be comparable to the labor spent on searching by photo banks: instead of formulating keywords for a photo bank, you will need to write prompt (or "prompts," approx. TAdviser), and instead of choosing a suitable search result - wait for generation, reformulate the request if necessary, improve the quality and download the result.

In favor of neural networks, it also says that the generated images have high originality, attract attention, often turn out to be quite beautiful, do not violate copyright and are cheaper to buy licenses from authors. And embedding algorithm directly into the administrative panel of the site with the automation of creating a request to the neural network based on the selected context of the material (which can be written by another neural network that distinguishes meanings) - can reduce the journalist's labor costs to values ​ ​ close to zero: up to a couple of clicks. mouse

The table below lists the products and their comparative characteristics. This list is not exhaustive, and there are other products and algorithms for generating images based on text descriptions.

Product Comparison Table for Generating Images Based on Text Descriptions

Let's dwell on the three most popular products: DALL-E 2, Midjourney, Stable Diffusion, which can potentially be used in media work.

The DALL-E algorithm developed by OpenAI can generate images based on complex text descriptions, including concepts that were not presented in the training data. One of the features of DALL-E 2 is that it can not only create photorealistic images in seconds, but also unusual works, such as abstract illustrations and non-existent objects, which will take the artist weeks or months.

DALL-E 2 realistically edits images by removing and adding elements, changing composition, shadows, and texture. Can create variations by being inspired by the original. DALL-E 2 provides a public API that can potentially be used to embed features in a media site engine.

Midjourney uses its own deep learning technology to create images with a high degree of detail and realism. Several off-the-shelf products already exist, such as applications for creating automatic advertising banners and book covers based on textual descriptions.

Midjourney is currently available through Discord-boat on the official Discord server-. The user generates an image using the "/imagine "command and enters the request, as in any other generator. artificial intelligence Then the bot returns 4 images, you can select 1 of 4 and create variations based on it or increase its detail.

The figure shows a screenshot of one of the iterations of the authors' request to generate an image suitable for illustrating this article:

Midjouney developers provide free access to their brainchild with some restrictions.

Stable Diffusion is a neural network architecture developed by researchers at Stability AI that can also be used to generate images based on text descriptions.

The architecture has been trained on a large set of image data and text descriptions, allowing it to generate qualitative and realistic images based on different material contexts.

The main advantage that should interest a developer planning to introduce a neural network into the work of the media editorial office is open source under the Creative ML OpenRail-M license. Stable Diffusion can run on the local computer, not via the website or API. But before using the code for commercial purposes, you should obtain a special permission.

Creative ML OpenRail-M requires that any derived networks (Derivatives) be published under an open license and be available for free use, excluding any patent requirements. Ethical restrictions are imposed on the content of pictures and use cases.

How to embed picture generation in the administrative panel of the site?

In an ideal world, an image generation algorithm is already built into sites and social networks. The generation process itself can proceed in three ways:

• The user is forced to describe in a separate field (window) the image that he wants to receive and generate directly in the admin panel application of the site or personal account of the social network.
• The text of an article or post is automatically processed, the image is generated based on the context of the material or part of it. A text neural network like ChatGPT can participate in this scheme, which will be discussed below.
• A combined option is when the user can use the proposed illustrations or give his own description of what he wants to get.

The technical implementation of such an idea is no longer long and expensive. It will not be years, but months, when such an algorithm can be built as simply as creating an adaptive design in the landing designer.

Improve photo quality

Upscale image is a service offered by dozens of sites in search results. The function can also be installed directly in the administrative panel of the site and increase the resolution of photos significantly. This is relevant for the media, which use amateur photos from the scene. In addition, many old articles, years or even decades old, are still ranked in search results and receive traffic. And if the text in them looks decent, then the photo gives the advanced age of publication. Editors of sites with long-term photo archives should think about processing all old photos, which will increase their ranking in search and improve the appearance of last year's news, which is still subject to traffic.

Automatic text generation

There are many text neural networks that are used in various fields, such as natural language processing (NLP), machine translation, text generation, tonality analysis, topic definition, and much more.

Some of the most famous and widely used models are:

• BERT (Bidirectional Encoder Representations from Transformers)
• GPT (Generative Pre-trained Transformer)
• LSTM (Long Short-Term Memory)
• CNN (Convolutional Neural Networks)
• Word2Vec (Word to Vector)
• Seq2Seq (Sequence to Sequence)

Each of these models has its own features and is used to solve certain problems in the field of NLP. For example, BERT is used to classify text, and GPT is used to generate text. LSTM and CNN are used for tonality analysis, and Word2Vec and Seq2Seq are used for machine translation.

ChatGPT from OpenAI in some scenarios compose texts indistinguishable from humans. It has little to do with writing analytical or artistic material. News also arises from the current agenda, they cannot be composed, especially given the increasing severity of the current legislation of Russia. But with the help of neural networks, you can process already written texts and make new ones from them, automating the work of rewriters. And robots here will not replace a person at all, but will only increase his productivity. If earlier 1 rewriter could write from 10 to 30 texts per shift, now he will be able to choose 100-200 information channels, feed their neural networks and check the results, adjusting if necessary.

Also, the neural network delightfully composes headlines and can give them out in dozens, creating an assortment for the editor to choose from.

On March 1, 2023, the GPT-3.5 API was announced, which means that it will be possible to integrate modules into the administrative panel of the site or the posting program in the social network that will help further simplify the rewright function, as well as reduce the amount of manual work on the author's materials.

Joint work of the neural network and the journalist

One of the possible options for using the API and the collaboration of a journalist and a neural network: the neural network helps the journalist and offers options for the next sentence, paragraph or title in a format similar to auto-completion, which everyone saw in the work of search strings or filters.

Text neural networks are already used in daily work. And they allow you to get outstanding traffic results. However, often the neural network shows excessive imagination, writing facts. And it is worth admitting, in Russian she writes not as brilliantly as in her native, English. The neural network has to be monitored by the corrector, which significantly slows down the work. For example, the plot from 15 materials was generated by the forces of a 1 person and one neural network in 30 minutes, but the corrector took almost 2 hours to read and edit. And in ordinary life, writing 15 materials of the same type would take about 3-4 hours, but it would take no more than an hour to correct. At first glance, the result does not seem impressive due to inflated expectations at the start. But in fact, labor costs are reduced by 2 times, and the speed of delivery of materials is doubled. Managers who make management decisions that even lead to 20% growth are awarded medals. And here the result is doubled by connecting just one tool.

If you calculate the labor costs, then you can determine in numbers the effectiveness of using neural networks in the work of journalists. The diagram below shows the result of the most pessimistic calculations when using the browser version of ChatGPT on a paid tariff. The drama of saving time should be evaluated by looking at the numbers, because a logarithmic scale is used to reflect small values ​ ​ together with large ones when building a diagram.

Labor costs for working with typical materials: journalist, neural network and combined method. Work specified in seconds

Preparing a full-fledged response by a neural network, about 3000 characters long in Russian, takes about 1 minutes, depending on the time of day and the workload of the program. It is worth noting that the generation of English texts is much faster. A person can make 5 options for headers to choose from in about 60 seconds. Neural network - for 15. However, joint work involves generating 5 headers by a neural network, choosing and adjusting by a journalist the best of the proposed ones.

According to the same logic, a journalist can draw up a thesis plan for a ready-made text or for a new idea, outlining the neural networks the essence of what he plans to write.

When it is necessary to lengthen the text ("pour water") - the neural network is indispensable, the journalist will have to adjust the finished material, since ChatGPT's literary Russian is not as good as English.

Reducing the text will also save half the time when a journalist with a neural network works in a team, in comparison with the individual labor of a representative of natural intelligence.

The greatest time savings are visible when preparing a rewright. It is most often enough for a journalist to make changes in style. But sometimes the neural network is updated with non-existent facts, which can lead to dire consequences for the media. Therefore, it is still risky to allow a neural network into a media admin with the right to publish without human control.

The reader can independently translate time savings into money savings in relation to his project and make one of two possible decisions:

• save on the salaries of journalists and riders, delegating half of the work of the neural network,
• keep the staff, but increase at least twice the amount of content produced.

In addition to writing reites, you can automatically generate digests and summaries, like the results of the week or the results of the day. The neural network is quite capable of coping with a brief retelling of the main events noted by the editor.

Thus, the use of neural networks significantly speeds up the process of writing materials and saves time for journalists, increasing the amount of content produced, or reduces the cost of paying riders and journalists. The neural network can be especially useful when writing reites and editing finished materials. However, so far the use of neural networks requires human control in order to avoid the publication of incorrect or false information.

Extracting meanings from text

Extracting meanings for a neural network is a simpler task. However, this also simplifies the work of people working with large amounts of data.

The neural network can highlight tags, write subheadings for the material, draw up an annotation and conclusion, and form a table of contents.

As for tags, categorization, or in other words automatic tagging, is a task that the IT departments of many major news agencies are beating right now. Tags should accumulate the basic meaning of the material. This is necessary for communication with other materials, with a similar meaning. For many years, journalists put tags manually from under the stick. Now, from 2023, it seems that this work was not initially human. But there has been such an opportunity for several years.

Automatic tagging, in addition to saving journalists time, has many other advantages. Firstly, so you can put a lot of tags. For the sake of aesthetics, some of them can be hidden. They will be needed for various official purposes, such as the withdrawal of similar material, the assembly of headings, the formation of plots, a dossier for persons. You can create new complex types of materials that will be hidden at least from readers, and maybe from journalists. You can build interesting matching algorithms on top of them. For example, define the subject, object, actions, and tonality of the material.

What else can you do?

Briefly list other possible ways to use neural networks in the work of online media.

• Video generation. Combining several neural networks in series, you can generate clips with video sequence, voice acting and credits. The existing technologies will be primitive, but such short videos can be put in stories, shorts or just social media feeds. This is more attractive than static pictures and carries a minimum of human labor.
• Processing time series. Time series of metrics, different graphs, dashboards... Now the graphs are viewed through the eyes, processed using data analytics, then the deviations that occurred in the past and the general trend are determined. But most anomalies at parameter intersections are still noticed by chance. Neural networks could be entrusted with predicting anomalies. Neuronka learns to determine how the schedule behaves before a sharp rise or a sharp fall and warns of anomalies. Back in 2016-2017, large companies spoke in reports on such practices. Since then, this has become easier to implement. Such a network will not be expensive in resources.

The graph below shows an example of finding anomalies in the number of visits to the regional media site. The yellow trend line shows the averages, and the red and green show the allowable corridor. Finding the graph within the corridor is considered the norm. The curve going beyond the corridor is an anomaly that requires increased attention from analysts.

Schedule for finding anomalies in the number of visits to the regional media site

• Identify traffic trends. Traffic trends can already be analyzed in real time, and not after the fact. When some news or plot just begins to gain momentum, this may not be noticed. Journalists do not track dashboards in real time and make them useless. And the material that began to go viral in the first minutes or hours of life may not stand out against the background of older materials until it becomes one of the leaders in viewing. With the help of trend analysis, it will be possible to determine the leaders of the agenda much earlier than people and immediately start pumping the topic, collecting all traffic. Now the editors determine the potential virality of the material intuitively.
• Search for cyclicals. Cyclicity can have a recurrence period of a minute to a day, month, or year. What would cycling give in attendance analysis? You can use it as noise cancellation, according to the principle of smart headphones. They record the ambient sound into the microphone and subtract it from what is transmitted in the speaker. If you subtract cycling from attendance, you can see a real change in global metrics. At the same time, the analysis of time series is not particularly tied to the specifics of what is analyzed. Percentage of load on the processor or percentage of transitions from VK - the network looks at the graph as a whole, analyzes it in time, finds cyclicity, and then subtracts it. After that, you can look for anomalies.
• Working with sound. There is already a ready-made project to work with sound. It allows you to analyze sound and turn it into text. The project works on neural networks, journalists have been using it for two years. He originally appeared as an interview transcript assistant voice for journalists. Journalists get time savings, and it can be scientifically proved that the development cost an increase in the economic efficiency of their work. The reverse conversion, the creation of an audio message from the material is already done by everyone. On the voice assistants one hand, for a long time you can ask the voice assistant to read the news. But the media can add a creative component. For example, generate a podcast from the results of the day, a picture of the day at the moment or a summary of a certain plot.

How much is it

For March 2023, the paid browser version of ChatGPT costs $20 a month, MidJourney - $10. The paid ChatGPT API is $0.002 for 1000 tokens, or about 750 words. The average number of characters in a word in English is 5 letters, and in Russian - 6 letters. Taking into account this, the approximate number of characters in 750 words in English is 3,750 characters, and in Russian - 4,500 characters. However, this is only an approximate estimate, and the actual number of characters may differ depending on the text and its content. Thus, an article written on the first attempt will cost no more than 15 kopecks, which can be considered an error and not taken into account in calculating savings on labor costs.

Launching a neural network on your own servers will cost much more and not the fact that it is better. Text neural networks are heavy. To just start such a network, you need a server with a powerful video card. Its cost exceeds a hundred thousand rubles a month as of February 2023.

At the same time, a powerful server is not needed to process pictures. In addition, it is possible to use free algorithms c. open source Now most smartphones of the functions of processing in real time are already built into the gallery of cameras. Therefore, over time, the generation of texts should also evolve and be technically cheaper.

A neural network that gets meaning is cheaper than a text generator. She will need a separate server for at least 20 thousand rubles a month.

The rest of the functions listed in the article are mostly available for use right now: some do not require anything but registration, others - several weeks of clean work on the implementation and rental of expensive servers. Everything else is also feasible in the near future. All technologies exist and these applications will sound naive in a couple of years.

Instead of output

As a conclusion based on the results of the article, without any edits, we present the ChatGPT response to the request "What are the prospects for using neural networks in the media?":

The use of neural networks in the work of the media has wide prospects. One of the main problems in the work of journalists is the huge amount of information that needs to be processed and analyzed in order to select the most significant news. In this case, neural networks can help automate the process of collecting, filtering, processing and analyzing information, which will significantly save the time and efforts of journalists.

For example, neural networks can be used to automatically write news articles based on data collected from various sources. This can help reduce the time spent writing articles, free up journalists for more creative tasks, and improve the quality and accuracy of texts.

In addition, neural networks can be used to analyze data and predict trends, which allows the media to be more informed and accurate in their predictions. For example, neural networks can help predict election results, economic trends, weather and other important events.

Neural networks can also be used to improve interactivity and personalize content for readers. For example, based on the preferences and interests of readers of neural networks, they can recommend the most suitable materials to them, which increases the involvement and retention of the audience.

Thus, the use of neural networks in the work of the media has great prospects and can help make the work of journalists more efficient and accurate, as well as improve the quality of content for readers.

Authors:

• Irina Golmgrain,
• Stanislav Popov

Drawings: Midjourney