Towards Unified, Dynamic and Annotation-based Visualisations and Exploration of Annotated Big Data Corpora with the Help of Unified Corpus Explorer

Kevin Bönisch, Giuseppe Abrami, Alexander MehlerNAACL 2025

TL;DR

This paper introduces the Unified Corpus Explorer (UCE), a novel, standardized, and dynamic platform for visualizing and exploring large, annotated big data corpora. UCE addresses the challenges of data heterogeneity and the lack of unified tools by leveraging UIMA-based annotations and a microservices architecture, enabling flexible and scalable corpus navigation.

What Problem Does This Paper Address?

The analysis and exploration of large text corpora, common in fields like linguistics, digital humanities, biology, and legal science, are hampered by significant challenges. These include the heterogeneity of processing methods and data formats, making visualization, interaction, and integration difficult. Furthermore, efficiently searching and interactively visualizing results from these diverse corpora, especially when dealing with massive datasets, is computationally demanding and lacks general-purpose tools. This paper tackles the need for a unified, flexible, and scalable solution to navigate and explore annotated big data corpora.

Methods

The core of the solution is the Unified Corpus Explorer (UCE), built upon a microservices architecture deployed using Docker. Key components include:

  • Corpus-Importer: Reads UIMA-annotated documents and configuration files, performing preprocessing steps like text segmentation, database indexing, keyword extraction, and embedding generation.
  • Relational Database (PostgreSQL with pgvector): Stores processed documents and high-dimensional vector embeddings, enabling efficient vector operations.
  • Graph Database (Apache Jena): Incorporates semantic searches using RDF and OWL data, allowing for the integration of domain-specific ontologies.
  • Python Webserver: Provides an interface for machine learning and AI models, facilitating access to embedding generation, dimensionality reduction (t-SNE, PCA), and LLM inference via a REST API.
  • UCE Web Portal: The user interface, implemented in Java, allowing interaction with all features and services.

UCE supports various features:

  • Document Reader: Displays annotated documents with interactive features and highlights.
  • Search Layers: Includes full-text search, Named-Entity (NE) Search, Semantic Role (SR) Search, and Embedding Search.
  • Chatbot: Leverages pre-trained LLMs fine-tuned for instruction-following and Retrieval-Augmented Generation (RAG) to support contextual, multi-turn question-answering. A custom CCC-BERT model helps determine when additional context is needed for the LLM.
  • 3D-TAXONOMY-VIZ: Visualizes annotated geographic entities on a 3D globe.

The entire process is standardized through the DUUI (Docker Unified UIMA Interface) pipeline, ensuring efficient and scalable processing of UIMA-annotated corpora.

Datasets

The paper does not explicitly mention specific datasets used for training the core UCE system. However, it highlights the use of the GERPARCOR corpus (German parliamentary debates) for evaluation. This corpus was extended with Semantic Role Labeling for the evaluation scenarios. The chatbot's training involved "35000 synthetic multi-turn chats generated with OpenAI's GPT-3.5 Turbo."

Key Insights and Ideas

  • Unifying Heterogeneity: The primary insight is the successful unification of diverse NLP tools and annotation formats through a standardized UIMA-based pipeline and a microservices architecture. This breaks down silos and enables cross-domain exploration.
  • Dynamic Adaptability: UCE's ability to dynamically adapt to different corpora and their annotations, without requiring code changes, is a significant advantage for researchers across various disciplines.
  • RAG for Enhanced Chatbots: The integration of Retrieval-Augmented Generation (RAG) with a custom context classification model (CCC-BERT) for the chatbot significantly improves its ability to handle nuanced queries and provide relevant, context-aware responses.
  • Visualizing Complex Relationships: The 3D-TAXONOMY-VIZ feature offers a novel way to explore and understand the spatial and taxonomic relationships within biodiversity data, making complex information more accessible.
  • Importance of User Experience: The evaluation highlights that while powerful features are available, clear explanations and intuitive interfaces are crucial for user adoption, especially for those less familiar with NLP tools.

Key Results

  • The user study with 15 evaluators found UCE to be generally acceptable, easy to use, and with tolerable load times.
  • The CHATBOT was the most positively rated feature, praised for its clarity and ease of use.
  • Group B, which used the CHATBOT and EMBEDDING SEARCH, provided more favorable feedback than Group A (SR-SEARCH and NE-SEARCH).
  • 86% of evaluators indicated they would use UCE again for similar tasks.
  • Despite positive feedback, a significant number of users reported frustration due to insufficient explanations of certain tools (SR-SEARCH, NE-SEARCH, EMBEDDING-SEARCH), leading to lower usage of these features.
  • Users often reverted to full-text search when overwhelmed by other features.

Limitations and Future Work

  • Explanations and Onboarding: The paper acknowledges that insufficient explanations for some tools led to frustration and underutilization. Future work includes developing better explanatory and onboarding materials.
  • Improved Filtering and Reading: Enhancements in filtering capabilities and document structure alignment are suggested for future development.
  • Scalability Testing: Stress testing UCE for very large datasets is identified as a necessary step.
  • Impact Assessment: Systematic evaluations of specific tools like CCC-BERT are needed to assess their impact on research projects.
  • Generative AI Integration: The success of the CHATBOT suggests further opportunities for integrating generative AI.
  • DH Integration: Incorporating Wikidition technologies is proposed as a future direction from a Digital Humanities perspective.
  • Broader Data Types: Extending UCE's corpus concept to include relational, non-primary textual data, as well as video and image data, is a promising avenue for future work.

Summary updated

BibTeX
@inproceedings{bonisch2025towards,
  author = {Kevin Bönisch and Giuseppe Abrami and Alexander Mehler},
  title = {{Towards Unified, Dynamic and Annotation-based Visualisations and Exploration of Annotated Big Data Corpora with the Help of Unified Corpus Explorer}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure