Harvesting Knowledge from Web Data and Text

Hady W. Lauw, Ralf Schenkel, Fabian M. Suchanek, Martin Theobald, Gerhard Weikum2010

TL;DR

This paper proposes a tutorial for CIKM 2010 on "Harvesting Knowledge from Web Data and Text." It outlines how to systematically extract structured knowledge from the vast, unstructured web and compile it into a machine-readable knowledge base, enabling more precise querying and advanced applications like semantic search and question answering.

What Problem Does This Paper Address?

The internet is a treasure trove of information, but most of it is in unstructured text and media, making it difficult to access and utilize effectively. This paper addresses the challenge of transforming this raw web data into a structured, machine-readable knowledge base. Such a knowledge base would contain not only entities (like people, places, and organizations) but also their semantic properties and relationships, akin to a "Structured Wikipedia." This would unlock powerful applications such as highly precise semantic querying, entity disambiguation, and natural language question answering, moving beyond current keyword-based search limitations.

Methods

The tutorial, and by extension the research area it covers, breaks down knowledge harvesting into three main tasks:

  1. Entity and Class Detection/Disambiguation: Identifying individual entities (e.g., "Albert Einstein") and assigning them to semantic classes (e.g., "scientist," "physicist"). This involves techniques like name matching and entity resolution to handle ambiguity.
  2. Relationship Extraction: Detecting binary relationships between entities (e.g., "Albert Einstein was born in Ulm"). This can be achieved through pattern-based extraction, where textual patterns are used to identify facts, or through wrapper induction for structured web data.
  3. Higher-Arity Relations and Temporal Reasoning: Extending beyond simple binary facts to capture more complex relationships involving multiple entities or temporal aspects (e.g., when a fact was true).

The paper also discusses knowledge representation using RDF (Resource Description Framework) and the importance of ranking results based on informativeness and confidence.

Datasets

The paper discusses several existing large-scale knowledge bases that are leveraged or serve as examples, including:

  • Freebase.com
  • TrueKnowledge.com
  • DBpedia
  • KnowItAll
  • TextRunner
  • Kylin/KOG
  • Omnivore
  • ReadTheWeb
  • YAGO (developed by some of the authors)

It also mentions resources like Wikipedia (for its semi-structured content), WordNet (for semantic classes), and various online dictionaries and social tagging sites. No new datasets are introduced in this proposal.

Key Insights and Ideas

  • The Web as a Structured Encyclopedia: The core idea is to treat the web not just as a collection of documents, but as a vast, albeit noisy, source for building a structured, encyclopedic knowledge base.
  • Bridging Databases and Web Search: This work aims to bring database methodologies to web search, enabling precise, structured queries rather than just keyword matching.
  • The Power of RDF: RDF is highlighted as a standard for representing knowledge bases, allowing for graph-like structures and interlinking of diverse data sources.
  • Informativeness and Confidence for Ranking: Simply retrieving facts isn't enough; ranking them by how informative and trustworthy they are is crucial for user satisfaction.
  • The "Machine Reading" Paradigm: Harvesting knowledge from the web is framed as a form of "machine reading," where systems learn to understand and extract information from text.

Key Results

This paper is a tutorial proposal, not a research paper presenting experimental results. Therefore, specific quantitative results are not available. However, it outlines the potential benefits and applications of successful knowledge harvesting, such as:

  • Formalized, queryable encyclopedias.
  • Improved entity disambiguation.
  • Enhanced semantic search capabilities.
  • More accurate question-answering systems.
  • Facilitation of machine translation and dialogue systems.

Limitations and Future Work

The paper acknowledges several challenges and areas for future work:

  • Handling Noise and Low-Quality Data: The web is inherently noisy, and extracting reliable facts requires robust methods.
  • Higher-Arity Relations and Temporal Reasoning: Extending knowledge extraction beyond simple binary facts to capture complex, time-varying relationships is an ongoing challenge.
  • Efficient Querying and Ranking: While models for ranking exist, efficiently evaluating them at query time for large knowledge bases remains an open issue.
  • Maintaining Knowledge Base Quality: Ensuring the accuracy and consistency of harvested knowledge, especially as it grows, requires sophisticated techniques like consistency reasoning.
  • Open Information Extraction: The paper notes that systems like TextRunner aim to extract all meaningful relations, not just predefined ones, which is a broader but less canonical approach.

Summary updated

Abstract

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et a ̀ la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
BibTeX
@misc{lauw2010harvesting,
  author = {Hady W. Lauw and Ralf Schenkel and Fabian M. Suchanek and Martin Theobald and Gerhard Weikum},
  title = {{Harvesting Knowledge from Web Data and Text}},
  year = {2010},
}
Figure