Information Extraction for Social Media

Mena Badieh Habib Morgan, Maurice van Keulen

Introduction: The Challenge of Noisy Data

Traditional Information Extraction (IE) systems were designed for the structured environments of the "Old Web," such as Wikipedia entries or formal news reports. These systems rely on grammatical consistency, standard capitalization, and extensive context to identify entities and their relationships. However, social media platforms like Twitter present a fundamentally different landscape. Posts are characterized by brevity, slang, typos, and a high volume of "pointless babble" (estimated at 40% of content). Furthermore, social media often discusses "non-famous" entities that do not exist in standard knowledge bases, leading traditional tools to either miss connections or "hallucinate" facts.

A Non-Linear Extraction Framework

To address these challenges, this research proposes a specialized IE framework that replaces traditional linear pipelines with iterative feedback loops. Key components include:

  • Noisy Text Filtering: A pre-processing layer designed to prune irrelevant content before it reaches computationally expensive extraction modules.
  • Iterative Feedback Loops: Unlike traditional systems where errors propagate downstream, this framework allows modules to communicate backwards. For instance, if a disambiguation module determines that "Apple" refers to a fruit rather than the tech company in a specific context, it signals the extraction module to correct its initial classification.
  • Open-World Disambiguation: The system looks beyond Wikipedia, attempting to link entities to social media profiles or personal homepages to account for the "long tail" of non-famous users.
  • Probabilistic Knowledge-base (PKB): Instead of binary truth values, the system stores facts with confidence scores. This allows the system to manage uncertainty and update its "understanding" as new, potentially contradictory information arrives.

The Role of Context and Metadata

A central insight of this work is that when text is too sparse to provide internal context, metadata becomes the primary source of truth. The framework leverages the source of a tweet to resolve ambiguity that would otherwise baffle a standard NLP tool.

Examples of social media posts where metadata and source context are crucial for entity disambiguation.

As shown in the examples above, a tweet mentioning a "fire on Belmont" (Example 3) is highly ambiguous in isolation. However, by identifying the source as the "Fort Worth Fire Dept," the system can accurately ground the location and event. Similarly, sports updates (Example 2) or local arguments (Example 1) require specific temporal and source-based context to be correctly structured into a knowledge base.

Performance and Real-World Application

The effectiveness of this modular, iterative approach has been validated in several high-profile benchmarks. The authors' system ranked first in the #msm2013 Concept Extraction Challenge and second in the #Microposts 2014 Named Entity Extraction and Linking Challenge.

Beyond benchmark success, the framework is designed for portability. Individual modules, such as the noisy text filter, can be integrated into external systems like sentiment analysis tools or specialized search engines. This modularity is currently being applied to the TEC4SE project, which aims to assist security forces by enriching real-time social media streams for operational decision-making.

Conclusion and Future Directions

While the iterative nature of this framework introduces higher computational overhead compared to linear pipelines, the gain in accuracy for noisy data is significant. Future work focuses on addressing the challenge of untrustworthy content and misinformation. By providing these modules as open-source tools, the researchers hope to enable more robust data extraction across the increasingly chaotic landscape of social media.

Summary updated

BibTeX
@inproceedings{morgan2014information,
  author = {Mena Badieh Habib Morgan and Maurice van Keulen},
  title = {{Information Extraction for Social Media}},
  booktitle = {SWAIE @ COLING},
  year = {2014},
  doi = {10.3115/v1/w14-6202},
}
Figure