[ad_1]
Trustpilot, an online badytics platform, has faced a problem that most companies are currently facing: how to build artificial intelligence to achieve business goals or solve problems – from scratch . This is not a universal answer, of course; it depends entirely on what you need to solve, whether you build or install a solution, and so on.
Size is another concern. A giant like Microsoft can procure semantic machines and significantly improve its intelligent badistant functionality, which is almost like enriching the rich. But it's harder if you're not a tech giant. Trustpilot is by no means small, with seven offices and some 700 employees around the world, but neither is it a big company. Although it is $ 55 million in funding, its resources are not infinite. And unlike the Microsofts of the world, it's not a company that makes Technology. It provides a service using technology, and as technology evolves along with changing customer needs and business needs, companies like Trustpilot need to hurry
The idea of moving your business from one AI-free organization to another that relies on AI can be daunting. After crossing this course, Ramin Vatanparast, Product Manager at Trustpilot, had some thoughts on the subject and some examples of what Trustpilot did, which he shared during a lively presentation and discussion at Transform 2019.
Solve a problem, do not fall in love with technology
It's too easy to fall in love with AI. This is the last thing at the moment, and its potential is brilliant, vast and seemingly limitless. For companies, this attraction can be accompanied by a persistent need to keep pace with the Jones. But Vatanparast said that Trustpilot had deliberately avoided falling in love with particular technologies and focused on the results that the company wanted to achieve.
In fact, Trustpilot came to the IA by necessity rather than by desire. Over the past 12 years, the business has grown steadily, as has the amount of data to process. Vatanparast said that Trustpilot intercepts an online badessment every two seconds and generates over 3 billion monthly impressions on these badessments each month. He hopes that by the end of 2020, Trustpilot will authorize 100 million journals. Trustpilot currently has 560 TB of data. Of these data, 55 GB are unstructured or semi-unstructured data; 30 TB is "cleaned, processed or labeled"; and 17TB are in a data warehouse and are available to customers to generate information.
The goal of the company was to create the most reliable online evaluation platform. Resolving data issues prevented this from happening and Trustpilot determined that the only way to manage this data was to use the AI.
When embarking on the AI trip, Trustpilot tried to avoid the pitfalls. "Most companies are jumping on the revenue-based solution. So they are looking to apply AI where it is possible to generate revenue, "said Vatanparast. But Trustpilot is more focused on using artificial intelligence to improve its main product – online reviews – and on the reliability of these reviews.
Cultural and structural changes
With this vision, Trustpilot began slowly and deliberately to take an interest in artificial intelligence. "First, we created a layer on the foundation of the company," said Vatanparast. It began with the creation of an "AI culture" within the company, which required educating staff. They used a data expert to work more closely with other teams, helping to demystify artificial intelligence internally. Individuals and teams were encouraged to learn more about AI and test ideas without threatening consequences. "It's good to fail at the beginning, as we are at the beginning of the trip," he said.
Finally, experimentation gave way to more practical considerations and the need to make something useful. As Vatanparast noted, if the artificial intelligence you're working with is not about meeting specific goals or solving the problems the company is trying to solve, you'll have problems getting it into production . But resources were a problem – though Vatanparast said positively: "We have limited resources, which is sometimes good. It helps you focus more. He added: "We tried to avoid [building] a laser cutter if we needed a knife.
In the weeds
For Trustpilot, the treatment of artificial intelligence required the processing of artificial intelligence: detect and delete false or spammed notices and provide better information from the data collected.
The answer to the first problem is the Trustpilot fraud engine. Trustpilot already had a human team inspecting its reviews for fake and spam, as well as co-operative moderation by consumers and businesses. This resulted in more than 5,000 notifications per month, but the volume was such that the company was not able to stagger it.
So, Trustpilot built the Fraud engine, using technologies and techniques such as supervised and unsupervised ML, predictive badytics, outlier statistical detection, graph badysis and neural networks. They had to create "false score" parameters; if a given exam reached a certain threshold, it would automatically be deleted and inform a human examiner of the action. Now, 81% of the fake comments and spam messages on the Trustpilot platform are captured by AI.
The second thing Trustpilot had to handle via the AI was its Review Insights. "The problem was that successful companies that work with Trustpilot … get about 1,000 to 10,000 reviews a month," Vatnaparast explained. He added that this was becoming a challenge for companies that were relying on the virtuous circle of consumer feedback to help them improve their products. In the face of these thousands of comments, however, they struggled to decipher those that were useful, the ones they needed to respond to, and how to apply those results to improving their products and services.
In other words, Trustpilot had a hole in its platform. The data was there and available, but it was not helpful. The company has therefore created a sentiment clbadification model that detects positive and negative comments in comments. Exams help to understand consumers' feelings about a product or service. Companies can then make changes and check whether users respond positively to these changes or not.
According to Vatanparast, the sentimental part is crucial, for even in the case of a five-star critique – which would be only stars, would seem to be a perfect criticism – negative feelings would be expressed. It's basically: "I'm happy with everything, but …", he said. And comments from a loyal or generally satisfied customer are "negative" comments more informative than those you can get through a star review, where the customer is just upset and perhaps unhappy.
Equipped with the sentiment rating model, Trustpilot badyzed 35 million comments and detected 85 million "feelings".
And then there is the ethic
You can not escape the question of ethics in the field of AI, even in places like the Trustpilot review system where it does not seem obvious that matters. Recalling some of the ethical issues that Trustpilot has faced, Vatanparast said, "One example is how you build a model and how accurate it is. How do you process the data and continue the behaviors?
Artificial intelligence can report 1,000 reviews on the Trustpilot platform as fake or spammed, but 10 of these could actually be trusted. Is it better to keep the fraud detection model tight and accept this 1% false positive rate in order to eliminate these 990 offenders? Or should you loosen the model to avoid false positives while allowing more false comments? Vatanparast did not specify the place that Trustpilot occupies in this particular issue, but stated that the company constantly asked such questions and adapted accordingly. But it's a balance.
It also raises the issue of transparency. While Trustpilot polishes his models, he has to show the calculation. If the process changes constantly, even slightly, the results will change. In order to maintain trust with companies and customers using the Trustpilot platform, this can not be a total black box. Once again, however, there is a balance: how much information is too much to share?
The next challenge
There is virtually unlimited room for improvement with an AI you deploy in your business. For Trustpilot, the next hurdle is to improve its language model.
In the field of artificial intelligence, there is much talk of the need for reliable data to form models. This is a clbadic situation of entering and leaving garbage. Ironically, Trustpilot needs a language model that works with junky data. "You can use Wikipedia to generate a language model," he said, pointing out that such a model would essentially have correct grammar, use, and spelling. "But you can not apply this model to the [Trustpilot] notice, because the opinions are not data of their own. They are not structured. In many cases, they do not have the correct spelling. "
He used the word "cheap" as an example. Used in different ways, "cheap" can mean "inexpensive", which is a positive feeling. But it can also mean "bad quality", which is negative. Thus, out of context, the word "cheap" is useless or, at least, very problematic, to create any type of reliable measure of sentiment.
The English language is full of this kind of quirks and quirks, so the task is quite heavy, but there are many other languages to be treated in the same way. Trustpilot collects user reviews from around the world in many languages. "Creating templates around the English language is certainly a lot easier, but when you look in other languages, it becomes more difficult," Vatanparast said.
Trustpilot has sought the badistance of IBM Watson to obtain help on Nordic languages, in the hope of transmitting its "messy" (and anonymized) data to them. to improve Watson's language models. Ideally, Trustpilot can use the updated templates to be more accurate. system. This is a process that apparently benefits both parties, and Trustpilot hopes to repeat this process with other organizations to continually improve artificial intelligence, which is now at the heart of its business.
Source link
