To develop, implement and test algorithms for identifying relevant topics (for a given problem owner), for identifying the stakeholders that entertain a position on those topics in news data and in user generated content, and for signalling significant changes in topics of stakeholders’ positions.
Topic detection and tracking has been an active research area for many years, but restricted to edited news content and to facts rather than entities. Robust mining of associations between topics and entities has mostly been limited to expertise in the case of academics or other knowledge workers, but needs to be extended to “perspectives”—moreover, the associations need to be explained and equipped with brief supporting evidence. First attempts at automatic aggregation of perspectives were recently evaluated at the Text Analytics Conference. Large scale analyses of positions have so far been limited to static positions only.
The core application is issue management, in both edited and user generated content.
Description of work: UvA will develop models, algorithms and prototypes to be professionalized in WP5; Talking Trends will provide labelled data and aid in specifying evaluation criteria.
- Task 2.1: modelling themes and their dynamics using language models and (dynamic) topic models (UvA, Talking Trends).
- Task 2.2: Identifying stakeholders and characterizing their positions using entity recognition, relation extraction and association mining (UvA, Talking Trends).
- Task 2.3: aggregation of static stakeholders’ positions using graph-based methods (UvA).
- Task 2.4: Identifying and predicting changes in stakeholders’ positions (UvA, Talking Trends).