The Center for Data Innovation spoke with Parsa Ghaffari, chief executive officer and co-founder of AYLIEN, a company based in Dublin, Ireland, that uses artificial intelligence (AI) and natural language programming (NLP) to analyze news content. Ghaffari discussed how AYLIEN’s software helps news and media companies to better understand their online content and audience.
Eline Chivot: What is your story as a tech entrepreneur, and how has it led you to set up AYLIEN?
Parsa Ghaffari: I’m originally from Tehran, Iran. I started coding around the age of 10, and by the time I was in high school I had figured out you can make money by building software!
When I was in college around 2006, I built and marketed various pieces of software: First an ad-hoc content management system (CMS) in the pre-Wordpress/Drupal/Joomla days which I sold to clients in the e-commerce space along with bespoke services, and subsequently a personal newsreader app for pre-smartphone mobile phones that gave you your daily news, finance data (such as FX rates), horoscopes, and so on.
I had always loved video games, so in 2008-2009 I started a game development studio that made casual games for the first smartphones. This was in the early days of the App Store when there wasn’t as much competition, so our four or five games brought us some 60,000 downloads in a short time window—each selling for $1 or $2.
The small success of building games in Iran and selling them to people from all around the world encouraged me to think about building a global business. So in 2010, I started looking at various startup accelerator programs and got admitted to China Accelerator—an accelerator program for web-based startups, based in China—to start what eventually became AYLIEN. I moved to Ireland in 2012 and officially started AYLIEN as a company in 2013.
Our initial product was a CMS plugin for bloggers and journalists that helped them enrich their content using third-party data. For instance, if you were writing about a public company it would suggest a stock price widget for this company, which you could easily drag & drop into your article. Or if you were writing about a celebrity, it would perhaps suggest photos of that celebrity for you to add. This saved the journalists plenty of time and helped create a richer experience for the end user.
What building this plugin revealed to us was that identifying which entities and topics are mentioned in a piece of text is a hard problem that is (or at least was, back in 2013) widely unsolved. That got us to think about the bigger opportunity of building tools and application program interfaces (APIs) that helped developers and data scientists to analyze natural language and text using natural language processing (NLP) techniques. This became our first commercially successful product, on top of which we built the products AYLIEN is offering today.
Chivot: How does your software work? What may be some of the current challenges associated with its development?
Ghaffari: AYLIEN’s product suite is focused on extracting actionable business insights from vast amounts of unstructured data, mainly text. We aggregate news and social media content from thousands of sources, and apply our proprietary NLP technology to it in order to discover and consolidate information that helps businesses operate more efficiently and generate more revenue.
Many businesses have exhausted structured sources of data, and they’re now looking for solutions that help them tap into unstructured data as a source of information. This applies to many use cases in risk management, insurance, banking, credit rating, publishing, media monitoring, and so on.
Our NLP technology is primarily based on deep neural networks that have been trained or tuned on large amounts of high quality data, and it currently operates in 16 languages.
There are three challenges which we—and the NLP industry in general—face from a technological perspective. The first one is accuracy and biases: NLP models are essentially mimicking the human ability to understand and comprehend language. This means that they are not always accurate, and also that they are prone to picking up biases that exist in the data they’re trained on, which is partly introduced by the humans who have provided this data.
A second challenge relates to explainability and transparency: Our NLP models are constantly making predictions about the world—things such as “this article is about a terrorist attack” and “this tweet is about a financial crime.” With such predictions come accountability, and therefore the models must be able to “explain” how they made a certain prediction—perhaps in terms of what words or patterns triggered that prediction.
Finally, a third challenge is multilingual analysis: While there are similarities within certain families of languages, each human language is different to the next one. Our customers target a wide range of markets and therefore require support in our technology for many different languages.
Chivot: How does your solution help news and media companies deliver better analysis of news content—and why is that important?
Ghaffari: We help our clients in the news and media industry to better understand their content, and as a result their audience, using our NLP technology.
On average we enrich each news article with about 25 additional data points. These include things such as sentiment analysis to determine whether the article is positive, neutral, or negative, stance detection for determining whether an article is for or against a topic or entity, and even clickbait headline detection—which is self explanatory.
These data points are then fed into both human-led (that is, editorial) and machine-led (meaning personalization) processes to increase engagement between the user and the content, and help editors operate more efficiently.
But it doesn’t stop here. Our technology also helps reduce bias by providing an objective view of the content regardless of the source or outlet that has published it. This has shown to help in reducing the biases that editors or end-users might have. The simple idea is that you represent all perspectives on a topic and let the user decide what they want and don’t want to believe.
Chivot: NLP can be applied to tackle a wide variety of issues and in various areas. How can AYLIEN provide solutions to address today’s online content challenges such as disinformation?
Ghaffari: As mentioned above, technology helps to reduce many of the issues and biases that exist in today’s media landscape. At the core of this is the idea that if you have technology that can strip layers of bias and provide an objective view of each article essentially as data, you can empower the end-user to navigate different perspectives and form a less biased view of the world.
For instance, for every “left wing” article that bashes a certain person we can surface a “right wing” article that’s praising that person. Or for every English article that is published about an event, we can surface a Russian or Chinese article about the same event which might be portraying it differently.
Similarly, NLP, and more generally AI, have been deployed in many social and commenting platforms to detect obvious signs of hate speech, harassment, explicit content, and so on with relatively good results. While these technologies are not yet perfect, they are showing us the path forward.
Chivot: What kind of future do you think we can expect for AI and NLP in particular and their applications?
Ghaffari: NLP and more generally AI are still in their infancy stage. We’ve built algorithms that can loosely mimic the perception abilities of a young human being. But given our limited understanding of intelligence, the human brain, and our surrounding world, we are nowhere near cracking hard problems such as consciousness and thinking with the algorithms we have built to date.
In a shorter-term time scale however, AI and NLP have proven to be able to reliably automate certain tasks across various industries and use cases, and this creates great opportunities for helping businesses operate more efficiently by leveraging the best of AI and NLP. So the future is bright!