A few words with Andreas Vlachos
If you have checked out the list of speakers on our website, you must know Andreas Vlachos will be speaking at Lisbon.ai. If you’re really excited about the event, then you must know his talk is going to be about the intersection between Imitation Learning and Natural Language Processing (more on this later).
Andreas is a lecturer and NLP/ML researcher from the University of Sheffield. He holds a PhD from the University of Cambridge and has been involved in various interesting ML/NLP projects.
We decided to ask Andreas a few questions about his experience and his talk at Lisbon.ai. Our goal is to share more information about the event’s speakers and their talks so as to get everyone even more hyped.
What got you into AI and then, more specifically, NLP?
When I was an undergraduate student in Greece, I was employed as an RA in a research lab. The project I was working on was mostly web development, which didn’t excite me. The head of the group suggested to me and and a friend to work on sentence splitting in Greek financial texts. After spending sometime devising and coding rules for this task, we thought we should get the computer to do this instead. We came up with an algorithm, wrote a paper for the Greek AI conference, and I became fascinated with learning from data and text in particular, and went on to study it in Edinburgh and then Cambridge.
You have done a lot of work in the field of Automated Fact Checking with NLP. Can you tell us a little bit more about that work and let us know how accurate automated fact checking is becoming?
We started this line of research with Sebastian Riedel at UCL when I joined his group as a postdoc in 2014. We were motivated by our interest in politics as we thought that the lack of grounding in facts was a major problem in public discourse. At the same time we thought that fact checking would be a great challenge to advance NLP and more broadly AI. It involves a number of well-studied NLP tasks such as semantic parsing and textual entailment and at the same time is an application that is very much needed. We detailed this in our 2014 paper describing the task and its challenges, and then in 2015 we presented an algorithm to learn how to fact check simple statistical claims such as “Greece has 10 million residents” (answer: it 10.75). Following this, with William Ferreira in 2016 we automated part of the fact checking process followed in emergent.info, a task which inspired the Fake News Challenge organized by Dean Pomerleau and Delip Rao. More recently, in Sheffield I have been working with James Thorne and developed a fact checking demo presented at the recent EACL in Valencia.
How accurate is automated fact checking? On the rather simple statistical claims we worked on, or on the subtask tackled in the Fake News Challenge, results are promising, e.g. in the latter the top-performing systems achieved >80% accuracy. However these are rather simplified (sub-)tasks, rather far from the task that human fact checkers do in organizations such as Full Fact and Politifact. Thus we have a long way to go!
Your talk is going to be applying Imitation Learning for a variety of NLP tasks such as semantic parsing. What makes Imitation Learning powerful for NLP?
Imitation learning is a learning paradigm originally developed to learn robotic controllers from demonstrations by humans, e.g. autonomous flight from pilot demonstrations. You might wonder what does this have to do with NLP? The answer is that the challenges that a model faces in taking a sequence of actions to navigate through traffic are similar to the challenges a model faces in constructing the syntactic parse for a sentence. A key strength of imitation learning is that one can use any classifier and learn with non-decomposable loss functions which are typically used in NLP. Come to Lisbon.AI and my talk for more!
What excites you about Lisbon?
I have been to Lisbon a number of times in the last couple of years, including EMNLP 2015 and LxMLS 2016, and I love it! It is a great city to have fun, and I found many parallels with my home country Greece. In addition, Lisbon has a thriving AI and NLP scene, thus making it a great place to develop research ideas! Last but not least, I have many friends now here, that I am looking forward to meeting up with!
There is a lot to learn and to absorb from this short interview. My main takeaways are that getting into AI is best and most commonly done via a small, well scoped project where one can make machines solve a well-defined problem for us.
Moreover, Andreas is, as we already knew, very experienced and his projects show immense talent. We can’t wait to host his talk and based on his answer to the last question, it seems we’re not the only ones who are excited.