The neural network was taught to write excellent reviews about cafes and restaurants

The reviews hold, if not the entire Internet, then a significant part of it so accurately. Reviews of various models of devices from different manufacturers, reviews of cars and spare parts, not to mention the feedback from visitors of restaurants and cafes - all these opinions serve as a means of assessing the quality of goods and services. Hotels, online stores, application catalogs - everywhere there is where to read reviews.

Marketers have long understood that a good review is a tool to attract new customers or users. But a bad review is a great way to scare customers away from a competitor, which makes it possible to draw their attention to themselves. There are entire companies that are only engaged in what they write reviews. They are hired by manufacturers of goods and service providers. Those who are simpler, write reviews yourself. But now there is a third option.

Researchers from the University of Chicago recently published a detailed report on the creation of a system that can write reviews in automatic mode. This is a fully automatic neural network operation. It’s practically impossible to distinguish the result from the reviews that users leave. If you take the "average temperature in the hospital," then the reviews of the machines are even better than human ones, since some of the feedback that people seem to leave look like they were not generated by the most powerful calculator.
')
In principle, what users, customers, and visitors write is not too different from each other. Usually it is something like “Thank you, pleasant service, it was cool.” More extensive comments are left much less often. The authors of the robot, which was taught to write reviews, give an example of his work: “I really liked this place. I came here with my brother, we ordered pasta for vegetarians and it was delicious. The beer was also good, and the service was just amazing. I will definitely recommend this place to anyone looking for a delicious breakfast. ”

There is nothing strange in this recall; it will not rush to the man’s eyes: the same comment as thousands of others. All this was written by a neural network trained on thousands of other reviews that were already left by people on the Yelp service. After the work was completed, the researchers conducted a test - they asked ordinary people to read a review written by a robot, along with people's reviews, and choose what the car wrote and what people wrote. As it turned out, the volunteers did not manage the task; they simply could not distinguish one from the other.

The responses generated by the neural network cannot identify any programs and services that usually track the generated text. This is especially true for comments that were written by the neural network with special care - the developers have added the function “maximally unique text”. Uniqueization in this case does not harm the quality of the written.

Sea of ​​opportunity

What the developers of the neural network have done, leaving excellent reviews, can be claimed by the industry. It is likely that development in this direction is either underway, or marketers are using something similar, just not doing their job advertising.

So far, the neural network, which was developed by representatives of the University of Chicago, has only a skill to write in Yelp (by the way, there is a whole series about this service in South Park). Quality comments in this recommendation network cost from $ 1 to $ 10, which allows you to transfer the neural network to a commercial basis and earn a lot of money.

Scientists are not going to do anything of this kind, but there is no guarantee that their colleagues or simply skilled developers will do it. As mentioned above, there is a possibility that many good comments in a variety of services and resources are already written using neural networks or other technologies.

In principle, the development itself is not unique, there is nothing super-complicated in it (the developers themselves say this). So someone who has enough powerful experience in this field and the appropriate equipment can do the same work, getting the same result. And even more so, there is nothing difficult about creating a training base for such a neural network - the Internet is full of reviews.

How bad is it?

Not that it was the end of the world, no. But the presence of a large number of neural networks that will create fake reviews (albeit of good quality) will nullify the industry of “feedback bug” like the same Yelp. If there is one such network operating on a commercial basis, others will follow. Dumping prices for reviews will begin, and then chaos will follow. In addition, unscrupulous businessmen will use attacks on competitors. And one should think that as such neural networks are improved, comments, both positive and negative, will flow like a river.

An ordinary person, a potential customer of the institution, having come to the site of reviews, will not be able to understand where the “linden” is, and where are the comments that have left real reviews. So the need for them will simply disappear. The representative of Yelp, however, thanked the experts for their work, saying that it would serve as good material for study. Yelp has a system of protection against spam comments, typewritten reviews, so now you need to come up with protection against "almost human" texts written by the neural network. How this can be done is not yet clear, but work in the near future will be carried out precisely in this direction.

As an example, there are six reviews, some of which are written by a machine, and some - by man. Try, while reading, to distinguish one from the other.

1. This is my favorite Italian restaurant. I like the tasting menu, everything is great here. I prefer carpaccio and asparagus. It is a pity that the restaurant has become more famous, and at rush hour it became more difficult to reserve a table;
2. My family and I myself are big fans of this place. The service staff is super friendly and the food is great. Chicken is very good, garlic sauce is excellent. Ice cream with fruit is also delicious. Recommend!
3. I come here every Christmas and I really like pasta! It is worthy of its price!
4. Great pizza, lasagna, better escalopes that I ate. The dessert is fantastic;
5. The food here is amazing, the portions are gigantic. Cheese bagel cooked perfectly, fresh and tasty. Service is fast. Our favorite place. We will be back!
6. I have been a client for about a year and a half, and I can only say good things about this place. I always order pizza, but the steak in Italian is also very good, I am impressed. The service is outstanding. The best I've seen. Recommend.

So, 1 review - real, 2 - fake, 3-5 real, 6-fake.

Experts who familiarized themselves with the development of their colleagues considered that this technology could change the structure and working principle of sites with reviews. In addition, we are talking about SkyNet, as without it. "I think that many people focus on singularity and SkyNet as a possible manifestation of the danger of AI, but I believe that in fact all really good artificial intelligence will behave differently."

Not only mechanical work

Large technology companies are constantly engaged in teaching AI (its weak form) to work with texts. Not only to analyze and structure, so to speak, in a machine-like way, but also to understand, to the best of their capabilities. IBM Watson knows how to understand the emotional coloring of the text, analyze the social component, evaluate the style of presentation. In addition, they are able to recognize and classify images.


A complete emotional immersion can probably be portrayed as

Emotions, the context of what was written — neural networks and cognitive systems already “understand” this. Cognitive system from IBM, a number of services from Yandex, algorithms of Google and some other companies. In the near future, this trend will only increase. All this does not appear all of a sudden, work in this field has been going on for many years and now the first significant results appear.

As technology improves, one can imagine a situation where the same reviews of computer systems will be written, “aware” of the written and giving the text the necessary emotional coloring. Probably, the cars of the future will be able to evaluate restaurants on their own, including the taste of dishes, and not just write review texts. But for now this is just a fantasy.