Greenpeace Magazine and Shoof. Dataset Description Extractor data Data from the extractors is stored inside Marmotta and is queried using the SPARQL endpoint, using Anno4j, as described in section 3.5. To remove complexity and interdependencies between recommendation module and extractors, it is assumed that all content items to be consid- ered for a recommendation task, have previously been annotated within the platform, using the appropriate pipeline. As described in the WP2 broker chapter of this deliverable, broker v3 will provide improved func- tionalities to define such pipelines, and to perform respective jobs, i.e. to execute a specific pipeline using a specific content set. Furthermore, it is assumed that content items referred to by user activities that are fed to the API are mappable to content items analyzed by the platform, as described in the WP2 high-level broker data model. Using this approach, the recommendation provides a best effort service: If annotations are available, the recommendation will use it. If only a subset of annotations exists, recommendation will still work, but its qual- ity might decrease. For the Greenpeace Magazine and Shoof showcase, any pipeline config- uration that produces named entities which can be processed within the recommendation module can be used. With reference to the extractor overview table in section 2.1, the following extractors are required by this specific recommendation: audiodemux to extract audio from video files for further processing, which is especially useful if the audio stream contains speech that can be further processed to deduce information about the content of the video. For details, see section 2.2.9. kaldi2txt to extract audio transcriptions for further processing. The task of speech-to-text extraction for different language is implemented using the Kaldi speech recognition toolkit. For details, see sec- tion 2.2.6. redlink-analysis provides textual analysis for Greenpeace articles and videos, and for Shoof videos. The NER capabilities of this extrac- tor will be used for subsequent clustering and therefore support recommendation of similar items. For details, see section 2.2.7. User behavior data User behavior is tracked as pseudonymous log data, e.g., from Google Analytics. The minimal set of required fields is as follows • date: date of the logging event • eventType: type of the logged event, see below • pagePath: URL to the respective content item • userId: numeric value for each user For the Greenpeace Magazi...
