Abstract
BreXearch is a semantic search system designed to navigate the massive, multilingual data trail generated by the 2016 Brexit referendum. By linking news, social media, and TV broadcasts to structured knowledge in DBpedia, it enables complex, cross-lingual queries that transcend traditional keyword-based search. This paper describes the system's four-stage pipeline—data collection, semantic annotation, data modeling, and query interpretation—which allows researchers to perform sophisticated cross-media analysis and identify patterns across disparate data silos.
Introduction
The 2016 Brexit referendum produced an overwhelming volume of data across global news sites, social media platforms, and television broadcasts. Analyzing this data is historically difficult because it is siloed by language (English, German, Spanish) and media format (text versus video). Most existing search tools rely on simple keyword matching or basic sentiment analysis, which struggle with semantic nuances—such as recognizing that "Prime Minister of the UK" and "David Cameron" refer to the same entity—or connecting a German news article to an English tweet about the same event. BreXearch addresses the need for a unified system that uses explicit knowledge bases to bridge these gaps, allowing for sophisticated cross-media and cross-lingual analysis.
System Architecture and Methodology
The authors developed a four-stage pipeline to process and search Brexit-related data:
- Data Collection: Data was gathered from three distinct sources: global news streams, social media (microposts, blogs, forums), and live-TV streams (audio and video frames).
- Semantic Annotation: Using the X-LiSA tool, the system performs cross-lingual entity linking. It identifies mentions of people, places, and concepts in multiple languages and maps them to unique identifiers in DBpedia. It also categorizes content based on DBpedia’s hierarchy (e.g., linking "Brexit" to the broader category of "Euroscepticism").
- Data Modeling: A specialized schema treats media items as "items" with properties like language, creator, and modality. Crucially, it uses "start" and "end" positions to link specific segments of text or timestamps in a video to specific entities.
- Query Interpretation: To enhance user experience, the system implements Query Entity Graphs (QEGs). When a user types a keyword, the system interprets it as a graph of related entities, helping disambiguate search intent before retrieving results.
Data Modeling and Semantic Enrichment
The core of the BreXearch system is its ability to represent diverse media types within a unified semantic framework. By anchoring every piece of content—whether a sentence in a tweet or a frame in a TV show—to the same semantic entities, the system creates a comparable unit of information across modalities.
As illustrated in the data model, the system utilizes xlime:Item as a base class, connecting it to xlime:EntityAnnotation and xlime:Position. This allows for high granularity, where specific positions within a media item are linked to DBpedia classes and categories.
Experimental Datasets
The researchers compiled a substantial dataset specifically for the Brexit use case:
- News: Approximately 240,000 articles from global sources.
- Social Media: Approximately 12 million microposts.
- Television: 900 TV programs processed for both audio and video content.
- Knowledge Base: DBpedia served as the external "ground truth" for entities and categories.
Search and Analytical Capabilities
The BreXearch prototype demonstrates the power of moving beyond keywords. Because the data is stored as RDF triples, users can execute highly specific, multi-variable SPARQL queries. For example, a researcher can ask: "Which Conservative Party politicians were most present in English microblogs regarding Euroscepticism in the two weeks before the vote?"
The user interface allows for intuitive navigation of these complex relationships. Users can search via keywords, which are automatically mapped to entities, or by selecting entities directly from a graph. The integration with DBpedia allows for advanced "Content Analytics," where results can be filtered by political party, nationality, or complex category hierarchies that were not present in the original raw text.
Conclusion and Future Work
BreXearch successfully integrates disparate data types into a functional semantic search platform. While currently a specialized tool for the Brexit use case, the authors identify several areas for future development. Primary goals include adapting the pipeline to other major global events, such as the 2016 US Presidential election, and exploring the challenges of applying semantic enrichment to real-time, high-velocity data streams. Additionally, further quantitative benchmarking of search precision and recall compared to traditional methods remains a priority for future evaluation.
Summary updated