An Infrastructure for Building Semantic Web Portals

Yuangui Lei, Vanessa López, Enrico MottaWISM 2006 · OpenReview

TL;DR

This paper introduces the KMi semantic web portal infrastructure, designed to tackle the challenges of extracting high-quality semantic metadata and providing intuitive data querying. It offers a novel combination of automated metadata extraction, ontology-driven question answering, and enhanced semantic search to make semantic web data more accessible and useful.

What Problem Does This Paper Address?

The Semantic Web aims to make information understandable not just to humans but also to machines, enabling more intelligent applications. Semantic web portals are key to realizing this vision by aggregating and providing access to information within specific communities. However, existing portals often struggle with two critical aspects:

  1. Metadata Extraction Quality: Automatically extracting high-quality semantic metadata from diverse, heterogeneous sources is difficult. Ensuring this metadata is accurate, free from errors, and consistently updated is a significant challenge.
  2. Intuitive Data Querying: Users often face a steep learning curve when querying semantic web data. They need to understand complex ontologies and query languages, which hinders accessibility. The goal is to allow users to ask questions in natural language, as they would to a human expert.

The KMi infrastructure aims to bridge this gap by providing a robust solution for both high-quality metadata acquisition and user-friendly data access.

Methods

The KMi semantic web portal infrastructure is built around three core components:

  • ASDI (Automated Semantic Data Acquisition): This tool focuses on extracting high-quality metadata. It comprises:
  • Information Extraction: Uses ESpotter, an adaptive Named Entity Recognition (NER) system, to identify entities in textual sources. Its adaptability allows it to correctly identify entities in different contexts (e.g., "Magpie" as a project vs. a bird).
  • Semantic Transformation: A semantic transformation engine converts extracted data into a specified domain ontology using a mapping ontology.
  • Information Verification: A crucial component that checks the quality of extracted data through three steps:
  • Internal Lexicon Library: Uses domain-specific lexicons to identify and map abbreviations to entities.
  • Repository Querying: Queries the existing semantic web data repository to resolve typos and minor errors, employing string matching algorithms and disambiguation mechanisms.
  • External Resource Investigation: Utilizes tools like PANKOW and WordNet to classify entities and determine their correct type when internal methods fail.
  • AquaLog (Ontology-Driven Question Answering): This component enables users to ask questions in natural language. It works by:
  • Linguistic Parsing: Translating natural language queries into linguistic triples.
  • Relation Similarity Service (RSS): Interpreting these triples as ontological triples by leveraging domain ontologies, string metrics, and lexical resources like WordNet.
  • Inference Engine: Deriving answers from the ontological triples by querying the semantic data repository. AquaLog also incorporates a learning mechanism to improve its performance over time by adapting to user feedback and ontology-specific jargon.
  • Semantic Search Engine: This engine enhances traditional keyword search by incorporating semantic relationships. It aims to:
  • Understand Keyword Meaning: Differentiate between concepts, instances, and literals represented by a keyword.
  • Identify Related Instances: Find instances relevant to the keyword, whether they are direct matches or indirectly linked.
  • Leverage Semantic Relations: Utilize both explicit and implicit semantic relations between entities to retrieve more relevant results, going beyond simple keyword matching.
  • Ranking: Ranks search results based on string similarity, domain context, and semantic relation closeness.

Datasets

The paper describes an experimental evaluation using data from the Knowledge Media Institute (KMi) at the Open University. This includes:

  • Departmental databases: Used for extracting metadata about PhD students.
  • Electronic newsletter archive: Contains news items describing events significant to KMi members.
  • Domain-specific ontologies: Used by ASDI and AquaLog to structure and interpret the extracted data.

The paper does not introduce new benchmark datasets but rather uses existing data within the KMi context for evaluation.

Key Insights and Ideas

  • Quality is Paramount: The paper strongly emphasizes that for semantic web portals to be truly useful, the quality of the acquired semantic metadata is as crucial as the ability to access it. This means ensuring data is accurate, error-free, and reflects reality.
  • Verification is Key to Quality: The ASDI's verification engine significantly improves the quality of extracted data, boosting precision even if recall slightly decreases. This highlights the importance of robust error-checking mechanisms in automated data extraction.
  • Natural Language Querying is Achievable: AquaLog demonstrates that users can indeed ask questions in natural language and receive precise answers, significantly lowering the barrier to accessing semantic web information.
  • Beyond Keyword Matching: Semantic search, by leveraging ontologies and semantic relations, can provide much richer and more relevant results than traditional keyword search, uncovering information that might otherwise be missed.
  • Portability is a Goal: While not fully achieved, the design of AquaLog emphasizes portability with respect to ontologies, aiming for broader applicability.

Key Results

  • ASDI Performance: The ASDI component, with its verification engine, achieved a precision of 0.911 and a recall of 0.775 for metadata extraction, outperforming the baseline ESpotter (precision 0.787, recall 0.798) in terms of overall quality (F-measure of 0.837 vs. 0.792).
  • AquaLog Performance: An initial study with 69 questions showed that AquaLog handled 40 questions correctly and 19 more could be handled with minor reformatting by the user. Failures were primarily due to linguistic handling limitations or lack of appropriate services/metadata.
  • Semantic Search: While still in its early stages, the semantic search engine showed encouraging results, successfully ranking news stories about "phd students" based on explicit and implicit semantic relations.

Limitations and Future Work

  • Manual Mapping: The current setup for metadata extraction requires manual specification of mappings, making it "heavy to launch." Future work involves investigating automatic or semi-automatic mapping algorithms to improve portability across domains.
  • Trust Management: The infrastructure lacks robust support for trust management, which is crucial when data is contributed from various sources. Future research will focus on associating trust factors with metadata and how processing tools handle trust.
  • Ontology Limitations: AquaLog's performance can be limited by the coverage and structure of the underlying ontology. Future work includes using multiple ontologies to overcome individual limitations.
  • Recall Improvement: The authors plan to improve ASDI's recall by introducing additional information extraction engines to work in parallel.
  • Semantic Search Refinement: Further investigation into a more fine-grained ranking algorithm and the impact of implicit relations on search results is planned.

Summary updated

BibTeX
@inproceedings{lei2006infrastructure,
  author = {Yuangui Lei and Vanessa López and Enrico Motta},
  title = {{An Infrastructure for Building Semantic Web Portals}},
  booktitle = {WISM},
  year = {2006},
  url = {https://openreview.net/forum?id=ctKvozOzhY4},
}
Figure