Abstract
This paper presents a demonstration of work to enhance a Web intelligence platform that crawls and analyzes online news and social media content about climate change. By incorporating the content of non-textual media, specifically YouTube videos, the system uncovers sentiment and opinions that were previously inaccessible. We describe a semantic approach to extract and include video content, leading to improved Web intelligence capabilities.
Introduction
Web intelligence platforms are designed to track how public opinion and sentiment evolve over time by crawling news feeds, blogs, and social media. However, these platforms have historically been "text-blind" to audiovisual content. While a platform might know a YouTube video exists because of its title or tags, it typically cannot "see" what is actually being said inside the video.
This creates a massive data gap, especially in fields like climate change communication, where organizations and individuals frequently use video to share complex opinions. To bridge this gap, researchers developed videoLyzard, an extension to the webLyzard platform that treats video fragments as first-class citizens in the world of semantic search and sentiment analysis.
Methodology: The videoLyzard Pipeline
The system transforms raw YouTube URLs into searchable, semantically enriched fragments through a multi-step server-side pipeline:
- Transcription: The system retrieves text transcripts of crawled YouTube videos.
- Named Entity Recognition (NER): Using the NERD service, the system identifies key entities (people, places, concepts) within the transcript.
- Semantic Normalization: Entities are mapped to unique DBpedia URIs. This ensures the system understands that "hydropower" and "hydroelectric" refer to the same concept, leveraging Linked Data to disambiguate terms.
- Temporal Fragmentation: The video is split into distinct segments, usually at the sentence level, using the Media Fragment URI specification.
- RDF Annotation: The system generates machine-readable annotations using the LinkedTV Ontology, stored in a triple store for complex querying.
The Media Watch Demonstrator
The Media Watch on Climate Change dashboard, showing the integration of video fragments alongside text-based news and social media analysis. Source, p. 4
The prototype was implemented using a dataset of 297 annotated YouTube videos integrated into the "Media Watch on Climate Change" demonstrator. As shown in the dashboard above, video content is presented alongside news articles, social media streams, and Fortune 1000 PR feeds. This integration allows users to see how video discourse aligns with or diverges from traditional news.
Key Results and Semantic Recall
The semantic approach significantly increases the recall of relevant content. In a case study, a search for "hydroelectricity" on a standard live site using traditional indexing returned only 3 videos. In contrast, the videoLyzard prototype found 6 video fragments. This was possible because the system recognized semantically similar terms like "hydropower," "hydro geothermal," and "hydro-electric" and correctly associated them with the primary search term through its Linked Data mapping.
Limitations and Future Work
Several hurdles remain before this becomes a production-standard tool. Current automatic transcriptions are often "noisy," requiring domain-specific cleanup methods to improve NER accuracy. Additionally, the pipeline must be optimized for near real-time scalability. Future research will explore "less granular" fragmentation to provide better context for human viewers and apply sentiment analysis algorithms directly to video fragments to compare the tone of audiovisual media against text-based news.
Summary updated