Text Analytics Meetup – KPN Results

The Text Analytics Meetup series is a collaboration in transfer learning for Natural Language Processing (NLP), initiated by Gianluigi Bardelloni and Yury Kashnitsky (KPN). It was intended to develop best practices in using unlabeled data to boost performance in classification tasks. Given that labeling data is particularly cumbersome and expensive in NLP tasks, one of the key motivations for such a collaboration was to find ways of leveraging current State-of-the-Art transfer learning techniques (such as ULMFiT and transformer-based approaches, BERT) to diminish the need for vast amounts of labeled data in applied business tasks.

The key contribution done by the mentioned transfer learning “club” is a DistilBERT classification pipeline with Catalyst, which helps any NLP practitioner quickly test transformers by HuggingFace in their classification tasks, reusing best DL practices through the Catalyst framework.

At the same time several seminars were organized during this collaboration, where participants shared various tips and best practices, e.g. how to train deep learning models with TPUs or how BERT in general works.

The main outcomes from KPN are as follows:

• Intro, ideas for collaboration within Amsterdam Data Science by Yury Kashnitsky
• Common  classification pipeline (plans) by Yury Kashnitsky
• Common classification pipeline (result, with Catalyst) by Yury Kashnitsky & Co.
• “Unsupervised Data Augmentation” – summary by Boris Zubarev
• Overview of the Jigsaw competition by Yury Kashnitsky
• Intro to Transformers by Boris Zubarev
• Intro to BERT by Boris Zubarev
• Intro to Transformers (formerly, PyTorch-transformers) by Boris Zubarev
• Tips on training with TPUs + ipynb by Dmitry Leghikov
• Overview of NLP from RNNs to transformers, ADS collaboration on text analytics by Yury Kashnitsky

Text Analytics – Edition I

Text Analytics – Edition II

If you’re interested collaborating by pitching a challenge, please email info@amsterdamdatascience.nl.