3rd Belgium NLP Meetup

On Thursday May 18th, we'll continue our tour through Belgium and meet at the offices of The Language Industry in Ghent for three exciting NLP talks. By popular request, we'll finish the evening with drinks, an opportunity to network and a chance to discuss the most pressing current questions in Natural Language Processing: how much linguistic information do we need to achieve natural language understanding, how far will deep learning take us, are GRUs better than LSTMs, and how can we overcome the eternal confusion with Neuro-Linguistic Programming?

Quality Estimation for Machine TranslationJoachim Van den Bogaert, Crosslang

Traditional MT evaluation metrics (such as BLEU, TER and METEOR) are static – they assess overall MT quality over a pre-defined sample of translated sentences. They can only be used during engine development, because in production, no reference translations are available and quality must be predicted for each sentence separately. Quality Estimation (QE) provides technology that only requires a source sentence and its MT output to produce a reliable quality assessment per sentence. QE can be used to increase the effectiveness of MT workflows (by marking high-quality output as "review-only") and to remove frustration at the post-editor side (by hiding low-quality output). Quality Estimation can furthermore provide an objective instrument to conduct negotiations using a metric that is similar to the widely used "TM fuzzy match level".

"Lekker, yummy, délicieux”. Fine-grained sentiment analysis of customer reviews.
Orphée De Clercq, LT3 Ghent University

Every day large amounts of opinionated content are being created online. In NLP, the task of automatically deriving these opinions is known as sentiment analysis and in the past decade this task has yielded a lot of attention both in academia and in commerce. Recently, the attention in sentiment analysis research has shifted from the coarse-grained detection of the polarity of a given piece of text to the more fine-grained detection of not only polarity, but also the target of the expressed sentiment. This is known as aspect-based sentiment analysis. During this talk, I will zoom in on the different subtasks and discuss the state of the art when applying supervised machine learning techniques to each of these. This will be done by focusing on the analysis of customer reviews, more specifically restaurant reviews written in three languages: Dutch, English and French. Working with multilingual data raises questions such as: do sentiment expressions differ from one language to another and is it possible to apply the same sentiment analysis methodology to all languages? Besides multilinguality, I will also discuss other challenges the field is still facing.

Information extraction from unstructured medical dataGeorges de Feu & Charlotte Hansart, LynxCare

The volume of medical data is multiplying at a yearly rate of >148%, expected to reach 1200 exabytes in 2017. >85% of that data volume is unstructured. This unstructured medical data is currently not used for patient follow-up, optimization of healthcare, medical and pharmaceutical research. LynxCare is a start/scale-up company that builds clinical text-mining technology and helps hospitals translate their stack of medical data into granular, structured, analyzable data. We are currently active in seven Flemish hospitals, where we unlock medical data, create insights for thousands of patients, enable better follow-up by their physicians and drive analytics to promote medical and pharmaceutical research. We'll give you a brief insight in LynxCare's approach, technology and real-world applications of clinical NLP.