Open Source SUMMA Platform https://github.com/summa-platform/summa-oss Guntis Barzdins (LETA) User Group Meeting 3 20 November 2018 The SUMMA project is funded by the EU H2020 ICT Programme under Grant Agreement 688139
Code and Installation https://github.com/summa-platform/summa-oss SUMMA User Interface Scales to 400 live TV channels: AWS m5.24xlarge instance per 25 live TV channels Code and Installation https://github.com/summa-platform/summa-oss
SUMMA Platform TRL 5-8 TRL 3-7 MediaItem Ingestion 250 TV/radio chanels Text & Social media Speech Recognition (ASR) 9 languages Machine Translation to English Segmentation and Punctuation Natural Language Understanding (NLU) Clustering in Storylines Summarisation Storylines MediaItems Topic Detection Named Entity Recognition (NER) and Linking (NEL) Persons Organizations GPE Events a Knowledge Base (KB) population (Facts about Named Entities) User eXperience (UX) Interface Trending view 24h Named Entities view (KB) Dynamic Storylines FreeText search Scalable to 400 live channels original original EN DE AR SP PT RU IR UA LV EN EN text annotations + text
SUMMA Platform Advantages Completely self-contained No dependence on external (cloud) services All components developed within SUMMA Scalable for BigData All NLP modules are Docker containers Scalable to 400 live streams on 800 servers (e.g. AWS) No external licencing
Multilingual technologies Open Source SUMMA Platform supports EN, DE, LV
Integration Architecture & Scalability for Big Data All components are Docker containers Scalability is achieved by launching as many Docker container instances per task as required Scales to 400 live TV channels
Final Scalability Test Sources and Resources
Final Scalability Test Conclusions Shallow stream processing for live video streams (ASR, punctuation, MT, topic detection) is useful for video content monitoring Natural Language Understanding components (storyline clustering, summarization, NER, NEL, relation extraction, geo-location) are useful only for written text input, but are mostly useless for live video input Language understanding needs to be grounded in video. LETA submited an ERC grant application: «High Dimensional Representation and Computing: Pixels, Objects, Language»
DEMO