NLATool: an Application for Enhanced Deep Text Understanding

Markus Gärtner, Sven Mayer, Valentin Schwind, Eric Hämmerle, Emine Turcan, Florin Rheinwald, Gustav Murawski, Lars Lischke, Jonas KuhnCOLING 2018

TL;DR

This paper introduces NLATool, a web application designed to bridge the gap between powerful NLP tools and users who need deep text understanding without extensive technical expertise. By integrating named entity recognition, entity linking, and coreference resolution with knowledge base enrichment, NLATool offers an intuitive interface for analysts to gain richer insights from text.

What Problem Does This Paper Address?

The field of Natural Language Processing (NLP) has seen remarkable advancements in tasks like named entity recognition (NER), entity linking, and coreference resolution. However, many existing tools are highly specialized, catering to expert users and often lacking in usability or interoperability. This leaves a significant gap for users in fields like digital humanities or general analysts who require deep text understanding but lack the technical skills to leverage these advanced NLP capabilities. The authors aim to democratize access to these powerful tools by creating a user-friendly application that simplifies complex text analysis and enriches it with external knowledge.

Methods

NLATool is a web application built with a human-centered design process. The core of the system leverages Stanford's CoreNLP for text analysis, including dependency parsing and named entity recognition. To enrich the text, it integrates with the Google Knowledge Graph (GKG) to retrieve additional information about identified named entities.

Key features include:

  • Named Entity Recognition and Linking: CoreNLP identifies entities, and these are then linked to entries in the Google Knowledge Graph.
  • Knowledge Base Enrichment: For identified entities, NLATool retrieves comprehensive text descriptions, images, and even maps (for locations and organizations) from the GKG.
  • Interactive Visualization: The interface uses a split-screen metaphor, with the text on the left and the enriched information (research component) on the right. Hovering over an entity in the text highlights related words and draws a connection line to its corresponding research result, facilitating spatial understanding.
  • User-Friendly Interface: The tool supports an "edit mode" for correcting or adding entities and a "comment mode" for annotating text. It also offers a "multiscreen feature" allowing users to view multiple text/research component pairs side-by-side.
  • Modularity: The system is designed to be modular, allowing for the integration of other NLP analysis infrastructure.

Datasets

The paper does not explicitly list datasets used for training or evaluation. However, it mentions the use of the Google Knowledge Graph and Wikipedia as sources for enriching text with external information. The system's underlying NLP pipeline, CoreNLP, is trained on extensive linguistic corpora.

Key Insights and Ideas

  • Bridging the Usability Gap: The most significant insight is the successful creation of a tool that makes advanced NLP capabilities accessible to non-experts. The human-centered design process, including user studies with computational linguists, was crucial in identifying desired features and ensuring usability.
  • Synergy of NLP and Knowledge Bases: Combining robust NLP pipelines (like CoreNLP) with vast knowledge bases (like GKG) is a powerful approach for deep text understanding. This integration allows for richer context and more comprehensive insights than either component could provide alone.
  • Visualizing Relationships: The highlighting and connection line feature is a clever way to visually represent the link between textual mentions and external knowledge, significantly aiding comprehension and spatial relation.
  • Multiscreen Support: Recognizing the prevalence of modern multi-monitor setups, the multiscreen feature is a practical design choice that enhances productivity for users dealing with large amounts of text and information.

Key Results

The paper focuses on the design and functionality of NLATool rather than presenting quantitative performance metrics. The key results are qualitative:

  • Successful Integration: The tool successfully integrates CoreNLP and the Google Knowledge Graph to provide enhanced text understanding.
  • User-Centric Design: The design process, informed by user studies, resulted in a tool that addresses user needs for information extraction and analysis.
  • Intuitive Interface: The split-screen layout, highlighting features, and editing/commenting modes contribute to an easy-to-use web interface.
  • Accessibility: NLATool makes deep text understanding tasks more accessible to a broader audience, including those without specialized NLP expertise.

Limitations and Future Work

  • User Management: User management is not fully implemented, as the current approach favors a "host your own server" model, which is suitable for privacy-sensitive data.
  • CoreNLP Dependency: The system relies on CoreNLP for its underlying NLP analysis. While powerful, this introduces a dependency on that specific pipeline.
  • Third-Party Server Approach: The current "host your own server" approach, while good for privacy, might present a barrier for users who prefer a fully managed cloud solution.

Future work could involve expanding user management capabilities, exploring alternative NLP pipelines, and potentially offering a more managed deployment option. The authors also suggest that the modular design allows for further integration of other analysis infrastructure.

Summary updated

BibTeX
@inproceedings{gartner2018nlatool,
  author = {Markus Gärtner and Sven Mayer and Valentin Schwind and Eric Hämmerle and Emine Turcan and Florin Rheinwald and Gustav Murawski and Lars Lischke and Jonas Kuhn},
  title = {{NLATool: an Application for Enhanced Deep Text Understanding}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2018},
}
Figure