Online AI Meetup - Natural Language Processing

Dear Stuttgart AI Community,

after almost two years, we are happy to announce the restart of this meetup group on December 8th! Due to the current situation, we will hold the meetup online.

Prior to this meetup, we have asked you what topics you would be interested in. According to the results, we have prepared two talks on Natural Language Processing:

PEER TALKS @ Stuttgart AI

1. Christoph Bensch - Data Scientist @ KI performance

"Open-domain Question Answering: ElasticAlbert"

Question answering is one of the oldest tasks of natural language processing (NLP), which is concerned with answering questions asked by users in natural language. While the question answering area already exceeds human performance like the Retro-Reader on the Stanford Question Answering Dataset (SQuAD) 2.0, open-domain question answering is still lacking in performance.
ElasticAlbert is a project that aims to reduce the response time of current question answering models. In order to improve the performance, a question answering model is prepended with a retriever which pre-selects only relevant articles for the question answering model.
In summary, this approach tries to make open domain question answering applicable, since the long computing time of parametric question answering systems is significantly reduced by using a non-parametric retriever.

2. Oliver Bensch - Data Scientist @ KI performance

"Key Information Extraction From Documents: Evaluation And Generator"

Extracting information from documents usually relies on natural language processing methods working on one-dimensional sequences of text. In some cases, e.g. for the extraction of key information from semi-structured documents such as invoice-documents, spatial and formatting information of text are crucial to understand the contextual meaning.
In this research project a template-based document generator was created to compare state-of-the-art models for information extraction. An existing information extraction model "Chargrid" (Katti et al., 2019) was reconstructed and the impact of a bounding box regression decoder, as well as the impact of an NLP pre-processing step was evaluated for information extraction from documents.

3. TBD

We are still looking for a third talk. If you want to give a talk at this or one of the following meetups, please feel free to contact us! :)

Looking forward to seeing you all online on December 8th! :)