Acronym:

Event Retrieval Based on semantically Enriched Structures for interactive user Tasks 

From: 2015-12-15 To: 2017-12-14 (Completed)

UKF 

- 

prof. dr. sc. Bojana Dalbelo Bašić

- 

News stories revolve around events, their protagonists, circumstances, backgrounds, and implications. Online news consumption has become the predominant way of news consumption. With the availability of tremendous amount of news content online, the technical challenge now lies in providing event-oriented search and recommendation capabilities that meet the diverse information needs. From an information retrieval (IR) perspective, the key issue is to devise an effective event representation model. Existing research opted for a priori event representations, seemingly intuitive for the application at hand, but not grounded in users’ understanding of an event nor aligned with their event consumption habits. In EVERBEST, we aim to to develop novel techniques and models for event-oriented search and recommendation grounded in event consumption habits. Consumption of online news is essential for shaping one's perspective and public opinion, thus having impact on the society as a whole, and critical for a range of businesses. We will therefore focus on three different aspects: support for authoring news (by journalists), news analysis in a specific sector, and news consumption by public. We anticipate that they differ in the level of analysis that need to be supported and interaction models, but share the common framework for representing and analysing events. We will adopt an empirical approach and conduct a systematic user study to discover how users approach two broad types of event-oriented tasks: search by example and searching via exploratory queries. As the users may be engage in search at any point in the life-cycle of an evolving event, and may be interested in different aspects of it, our primary goal is to devise a method that interactively refines user’s initial query and incrementally synthesizes the search task. To this end, we will model information-seeking tasks via a set of generalization/specialization operators defined over semantic event representations derived from documents, inspired by recent advances in program synthesis research. We will consider event representations of varying complexity, including structured representations obtained using natural processing (NLP) and machine learning techniques, as well as corresponding sets of transformation operators, to determine the optimal configuration with respect to the use cases. Finally, we will devise a prototype event search and recommendation engine as an extension of the EventRegistry, a platform for large-scale collection and indexing of news. By integrating findings from three research areas (NLP, IR, and program synthesis), our research may open the way for uncovering new paradigms in digital content analysis. Objectives and expected results We aim to to develop novel techniques and models for event-oriented search and recommendation grounded in event consumption habits. The first objective of the our research is to identify patterns in the ways the user approach two two broad types event-oriented information seeking tasks: search by example and exploratory queries. The second major objective is to model information-seeking tasks as pairs (S,L) of the Semantic Representation (S) for a given corpus of documents, and the set of operators (L) on that representation S that underpin the query expressions. Our final objective is to deploy and evaluate our method on a specific news platform such as the EventRegistry that collect news in real-time from a broad set of RSS feeds.  We will extend EventRegistry service with a search capability that leverages user browsing (capturing relevant examples) and search specifications through a /desktop and a mobile application. By integrating findings from three research areas (NLP, IR, and program synthesis), our research may open the way for uncovering new paradigms in digital content analysis.