Event Knowledge Graph
Event data is an important aspect of the EPOCH prediction models. In the first year of the project, the involved researchers have developed an implementation that can detect and resolve temporal references in German documents (not only absolute dates, but also references relative to the publication date). This temporal reference detection is a first means to predicting future events.
Based on this work, a first Event Knowledge Graph (EKG) has been created with events sourced from WikiData and public iCals, encompassing more than 30.000 events with a daily query updating event instances and descriptions from WikiData. We developed an API that helps us query events according to any of their properties, e.g. label, start, end, type, location etc. The API returns matching events for past or future dates, as feature candidates to train prediction models.
Named Entity Graph
The knowledge graph also stores information about relevant persons, organisations and locations. Initial experiments started to introduce ML algorithms to improve the detection of name variants through our NER/NEL engine Recognyze. Name variants are important as the full entity names often do not appear in a text. Modifications are common due to language, use of abbreviations or position within a text – a person’s role may or may not be included in its name; publicly traded companies are often identified by stock ticker symbols, e.g. TSLA for Tesla. Initial experiments used a fine-tuned Support Vector Machine (SVM) for name generation from Knowledge Bases like DBpedia and Wikidata, improving the results as compared to a baseline using no name generation and a heuristic (rule-based) method.