Members of the Applica research team spent last week in beautiful Lausanne, Switzerland to take part in the ICDAR 2021 (16th International Conference on Document Analysis and Recognition). Aside from enjoying the lovely weather and tasting a variety of Gruyères cheeses, the week was packed full of exciting events related to our proprietary work in deep learning and neural language modeling. Some highlights include:
- We presented Going Full-TILT Boogie on Document Understanding with Text-Image-Layout Transformer about our generative language model TILT, which is at the top of the leaderboard for Task 1 – Single Document VQA and Task 3 – Infographics VQA.
- Applica’s Chief Data Scientist, Filip Graliński, took part in the DocVQA Discussion Panel, and it was discussed, among other things, what type of DocVQA challenges would be helpful from the perspective of the industry. Filip highlighted multilinguality and measuring the quality of probabilities returned by the information extraction systems, not just the accuracy on bare pieces of information.
- During the final day of the conference, we presented LAMBERT: Layout-Aware Language Modeling for Information Extraction, which won the IAPR/ICDAR 2021 Best Industry Related Paper Award.
- And last but not least, we presented Kleister: Key Information Extraction Datasets Involving Long Documents with Complex Layouts.
It was incredibly exciting for Applica’s work to be recognized as state of the art by our revered peers in the industry and we will continue pushing the boundaries on what’s possible with document automation.