Abstract
Linked Data management faces significant challenges regarding data quality and completeness that automated algorithms often fail to address. This paper proposes a semantically enabled architecture that integrates human intelligence directly into the SPARQL query lifecycle. By extending standard Semantic Web protocols, the system treats crowdsourcing as a first-class query operator, enabling automated task decomposition, semantic optimization, and user interface generation to resolve complex data management problems with minimal human effort.
Introduction
The Web of Data provides a massive, decentralized repository of structured information. However, this ecosystem is frequently plagued by "messiness"—inconsistencies, missing links, and ambiguous classifications—that traditional automated reasoners struggle to navigate. Specifically, tasks such as identity resolution (determining if two URIs refer to the same entity), subjective ranking, and filling information gaps require human cognitive abilities. While crowdsourcing platforms like Amazon Mechanical Turk offer a solution, they typically require manual, ad-hoc task design. This research outlines a vision for a "crowd-aware" data management system that automates the integration of human workers into the Linked Data lifecycle.
A Hybrid Query Architecture
The core of the proposal is a hybrid query engine that bridges the gap between machine processing and human intelligence through formal semantic representations. Unlike traditional systems that treat crowdsourcing as an external post-processing step, this architecture embeds human tasks within the query execution plan.
As shown in the architecture, the system decomposes a single SPARQL query into machine-executable parts and Human Intelligence Tasks (HITs). This allows the engine to apply traditional database optimization techniques, such as predicate push-down and join-ordering, to human-centric operations.
Methods and Semantic Extensions
To facilitate this integration, the authors propose several extensions to existing Semantic Web standards:
- Vocabulary Extensions: The VoID (Vocabulary of Interlinked Datasets) protocol is extended with terms like
void:crowdClassandvoid:crowdProperty. These flags notify the engine which data types or predicates require human intervention. - SPARQL Extensions: A new
ORDER BY CROWDmodifier allows developers to declaratively request subjective ranking based on human preference. - Optimization via SPIN: The system utilizes SPIN (SPARQL Inferencing Notation) to apply logical rules like transitivity. For instance, if a human identifies that Resource A is the same as Resource B, and Resource B is the same as Resource C, the system automatically infers that A equals C. This semantic reduction can theoretically decrease the human workload for identity resolution by up to 33%.
- Automated UI Generation: Because the underlying data is structured, the system can automatically generate HTML interfaces for workers based on the predicates involved in the query. For example, a query involving geographical coordinates can automatically trigger a map-based interface for identity verification.
Key Insights and Results
The primary contribution of this work is the shift toward declarative crowdsourcing. By using SPARQL as the interface, developers can focus on what data they need rather than how to design and deploy crowdsourcing tasks.
Qualitative analysis of the architecture suggests significant efficiency gains. By moving away from the $O(n^2)$ complexity of traditional pairwise comparisons in data cleaning and leveraging semantic pruning, the system reduces the total number of paid human tasks. The authors demonstrated this through a motivational scenario involving the interlinking of METAR weather data, DBpedia, and Geonames, showing how complex queries (e.g., "Find German-labeled commercial airports") can be successfully parsed into a mix of machine and human sub-queries.
Limitations and Future Work
While the theoretical framework is robust, several challenges remain. Crowdsourcing introduces significant latency compared to millisecond machine queries, necessitating new "online" optimization strategies to predict and manage completion times. Furthermore, the current model does not yet account for worker reputation or the economic costs of tasks. Future iterations of the framework aim to semantically model the workers themselves, allowing the system to match specific tasks to individuals with relevant domain expertise.
Summary updated