5W1H-Aware Approach for Retrieving Semantically Rich Multidimensional Events in Social Media Ecosystem
No Thumbnail Available
Date
2024-08
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Nowadays, multimedia digital ecosystems (e.g., social media sites) become a great source of user-contributed multimedia documents for many types of real-world events. Social media documents about events are multimedia (e.g., image, video, text, and others), contain multiple features (describing different characteristics in it such as 5W1H), and come from multiple sources. Such events can be used to construct Event Knowledge Graph (EKG) which is basic to retrieve semantically rich multidimensional events. However, social media documents cannot be used directly to construct such a knowledge graph. First, social media documents must be represented in order to detect multimedia events as well as their different types of semantic relationships. To achieve these tasks, it is necessary to carry out preliminary event-related tasks, such as detecting, linking, and representing events. By doing these, we can provide an event search API that presents a concise summary of events focused on temporal, spatial, semantic, and participant aspects. For this, we proposed a novel 5W1H-aware framework consisting of six modules. Each of these modules uses 4W elements. More specifically, we first represent social media documents that have been used to detect real-world events with their 4W elements based on event-only descriptive features. The detected events are used to identify the three main event relations categories (such as, spatial, temporal, and sematic) and many relation types under these categories by comparing dimensions over multimedia events. We then used a graph database to store event detection outputs as nodes and relationship identification outputs as edges to construct an Event Knowledge Graph (EKG). Finally, we integrated the EKG with an event retrieval API to retrieve events.
Each of these event-related tasks were evaluated using various datasets. For instance, the effectiveness of an incremental event detection algorithm was evaluated using the MediaEval 2013 dataset. The algorithm achieved an NMI score of 0.9914 and an F-score value of 0.9928 when the feature weights were assigned as 0.40 for participant, 0.30 for temporal, 0.35 for spatial, and 0.45 for semantic. Manually crafted spatial and temporal datasets are also used for evaluating the effectiveness of event relationship identification algorithm. Finally, the effectiveness of the
developed EKG was evaluated through its downstream tasks, like retrieving similar nodes, whereas the event search API was evaluated via PageRank and search result relevance analysis. Results from the experiments showed that the proposed approach was more effective compared with alternative solutions.
Description
Keywords
Multimedia Event, Event Detection, Event Relationships, Event Relationships Identification, Social Media Platforms, Social Media Documents Representation, Event Representation, and Event Knowledge Graph