Relevance & Research Question

"Something has changed – as globalization has marched on, [political] debate is taking place in a completely new media environment. Opinions aren't formed the way they were 25 years ago" as Angela Merkel stated in her speech to Parliament on November 23rd 2016.[1]

We are in a world where social media and peer-to-peer communication is taking the lead over curated contents. The exponential growth of trolls, fake news and chat bots raises the question how state-of-art Natural Language Processing powered by deep learning algorithms can be applied to curating online contents.

Methods & Data

We build on a wide array of self-developed as well as open source software, originating from research performed at ETH, Zurich.[2] Thereby we make use of state of the art machine learning technologies such as word embedding, convolutional or recurrent neural networks, but also more traditional, heuristic approaches.

We pre-train our model using publicly available German text data and fine tune it with a dataset obtained from a major Swiss-German online-newspaper.

Results

We have been able to establish leading sentiment analysis models on challenging data inputs such as Tweets, winning the 2016 Semeval competition (task 4a).[3] Building on those proven methods, we have since tackled the problem of detecting unwanted user comments in German language online newspapers. In particular, our system aims at detecting “trolls” who try to capture online discussions and manipulate the public opinion.

Added Value

This approach can be used to manage online communication through cost-effective and repeatable tools to complement human judgment on the quality and validity of online data. It can further be used to monitor and act upon customer satisfaction and sentiment at company or product level.

[1] http://www.usnews.com/news/features/news

[2] http://e-collecHon.ethbib.ethz.ch/show?t

[3] https://www.inf.ethz.ch/news-and-events/

https://aclweb.org/anthology/S/S16/S16-1173.pdf