Will Deep Learning and Knowledge Graphs outperform Existing Systems?

February 01, 2019 by Stefan Summesberger

SEMANTiCS Poster & Demo Chair Ricardo Usbeck is Postdoc at Paderborn University. The focal areas of his research are Natural Language Processing, Question Answering, Search, Chatbots, and Smart Assistants. Ricardo shares some insights into these areas, his experiences as a participant of prior SEMANTiCS Conferences as well as expectations and advice regarding posters and demos for SEMANTiCS 2019.

You are an expert in question-answering systems over Knowledge Graphs. What are the hot topics in this area? What are the most exciting technology trends?

You may have heard of the Stanford SQuAD dataset where novel deep learning architectures outperformed human performance in 2018. These algorithms get a Wikipedia page and extract the correct answer given a question from the text. We have also seen the rise of property graphs as well as tensors to incorporate different data models into machine learning. Most exciting for me was the Alexa Price. Different university teams try to use an intelligent assistant device to talk for 20 mins with an end user who is able to end the conversation at any point. Here, we saw great strides in conversational AI using Question Answering.

We have seen QA systems for some years now, with lots of ups and downs. Are we finally getting to a stage, where machines can understand complex questions and return reasonable answers? What are typical constraints you are currently facing?

I do not think so. Despite neural network-based approaches reaching new leaderboard positions on different QA datasets weekly, they do not really understand natural language, human reasoning or contextualized information needs. They are however very good at answering factual questions over a limited set of knowledge sources, though. Current text or knowledge graph-based algorithm and their output also lack explainability. Especially with respect to trust, this is important but an unsolved issue. And do not get me started with GDPR compliance in that area.The single most important breakthrough happened recently in the area of representing language as low-level vector representations, also known as embeddings. These embeddings are used to train machine learning algorithms and recently became language- and domain-independent. To be so, they demand large corpora and even larger computational resources. However, the big companies and research labs made these embedding models and their generating algorithms open-source so it can be used by anybody without these resources.

Your research project is ready to be reviewed soon? Calls for SEMANTiCS 2019 are open!

What prospects do you see for QA systems? What will be the most relevant application areas in the future? Please share some insights about your perspective on projects that currently are on your radar.

All the progress in the area of natural language understanding fueled a new discussion on whether Artificial General Intelligence (AGI) is in reach. We are definitely advancing with high speed and QA technology is further spread than ever before. Think of your smartphone, your home assistant or modern in-car voice assistants. I predict that we will see even more adoption of voice-driven technology in the future as this is the easiest means of communication for us humans. I am especially looking forward to non-factual QA algorithms helping us in our daily business routines, e.g. finding the latest research methods, analyzing and understanding customer behaviour but also assisting elderly people via ambient assisted living. Finally, we will see the combination of Deep Learning and Knowledge Graphs, sometimes called informed Machine Learning, outperform neural approaches over text.

At SEMANTiCS 2019 you will be chairing the Posters and Demos Track. What are your expectations in this respect? What do you especially look forward to about the submissions?

I hope we can continue the spirit of last SEMANTiCS conferences, where the interaction between industry and research posters and demos sparked novel ideas, interesting collaborations and a vivid community that lasted way longer than the conference did. I am also looking forward to actively promote bidirectional knowledge and data transfer. Some researchers need to get out of their ivory tower and make their inventions applicable and some companies should share their data and interesting problems with the community more. This way, both sides benefit. I am in particular keen to see working prototypes and demos as well as open source projects and datasets. For me as a fresh PhD student, the SEMANTiCS 2014 in Leipzig was an eye-opening experience!

Any last words to those who haven’t submitted yet, researchers facing constraints during the work on their projects or those who still have a lot on their ToDo-Lists before the calls close in April?

Submitting to SEMANTiCS will definitely help you to spread the word about your work. I want to suggest that you write your submissions understandable also for non-field experts. A goodie will be the excellent organization by this year's local host Harald Sack. Feel free to ask him about where to get the perfect espresso in Karlsruhe!

Ready to share your project with the SEMANTiCS Community? Calls are open NOW!


The annual SEMANTiCS conference is the meeting place for professionals who make semantic computing work, and understand its benefits and know its limitations. Every year, SEMANTiCS attracts information managers, IT-architects, software engineers, and researchers, from organisations ranging from NPOs, universities, public administrations to the largest companies in the world. http://www.semantics.cc