Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

XNLP: An Interactive Demonstration System for Universal Structured NLP

Hao Fei, Meishan Zhang, Min Zhang, Tat-Seng Chua

TL;DR

XNLP addresses the fragmentation of structured NLP tasks by proposing a unified, interactive demonstration platform powered by a backbone LLM. It introduces a structure-aware, instruction-tuned framework and a universal structure formatter to predict spans and relations across diverse XNLP tasks, while enabling new task definitions and multi-turn user interactions. The system combines a Vicuna-13B backbone, in-context learning prompts, and brat-based visualizations to deliver end-to-end, interpretable predictions with rationale explanations. Empirical results suggest that with broad-cover instruction tuning, the smaller Vicuna model can outperform larger models like ChatGPT on XNLP tasks, highlighting practical utility for cross-task generalization and community-driven task expansion.

Abstract

Structured Natural Language Processing (XNLP) is an important subset of NLP that entails understanding the underlying semantic or syntactic structure of texts, which serves as a foundational component for many downstream applications. Despite certain recent efforts to explore universal solutions for specific categories of XNLP tasks, a comprehensive and effective approach for unifying all XNLP tasks long remains underdeveloped. In the meanwhile, while XNLP demonstration systems are vital for researchers exploring various XNLP tasks, existing platforms can be limited to, e.g., supporting few XNLP tasks, lacking interactivity and universalness. To this end, we propose an advanced XNLP demonstration platform, where we propose leveraging LLM to achieve universal XNLP, with one model for all with high generalizability. Overall, our system advances in multiple aspects, including universal XNLP modeling, high performance, interpretability, scalability, and interactivity, providing a unified platform for exploring diverse XNLP tasks in the community. XNLP is online: https://xnlp.haofei.vip

XNLP: An Interactive Demonstration System for Universal Structured NLP

TL;DR

XNLP addresses the fragmentation of structured NLP tasks by proposing a unified, interactive demonstration platform powered by a backbone LLM. It introduces a structure-aware, instruction-tuned framework and a universal structure formatter to predict spans and relations across diverse XNLP tasks, while enabling new task definitions and multi-turn user interactions. The system combines a Vicuna-13B backbone, in-context learning prompts, and brat-based visualizations to deliver end-to-end, interpretable predictions with rationale explanations. Empirical results suggest that with broad-cover instruction tuning, the smaller Vicuna model can outperform larger models like ChatGPT on XNLP tasks, highlighting practical utility for cross-task generalization and community-driven task expansion.

Abstract

Structured Natural Language Processing (XNLP) is an important subset of NLP that entails understanding the underlying semantic or syntactic structure of texts, which serves as a foundational component for many downstream applications. Despite certain recent efforts to explore universal solutions for specific categories of XNLP tasks, a comprehensive and effective approach for unifying all XNLP tasks long remains underdeveloped. In the meanwhile, while XNLP demonstration systems are vital for researchers exploring various XNLP tasks, existing platforms can be limited to, e.g., supporting few XNLP tasks, lacking interactivity and universalness. To this end, we propose an advanced XNLP demonstration platform, where we propose leveraging LLM to achieve universal XNLP, with one model for all with high generalizability. Overall, our system advances in multiple aspects, including universal XNLP modeling, high performance, interpretability, scalability, and interactivity, providing a unified platform for exploring diverse XNLP tasks in the community. XNLP is online: https://xnlp.haofei.vip
Paper Structure (36 sections, 11 figures)

This paper contains 36 sections, 11 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Structured NLP
  4. Demonstration for XNLP.
  5. Model Unification
  6. System Design
  7. Architecture overview.
  8. Backend
  9. Backbone LLM.
  10. In-context learning.
  11. Broad-cover structure-aware instruction tuning.
  12. Structure formatting.
  13. Frontend
  14. Pre-defined XNLP tasks.
  15. New task definition.
  16. ...and 21 more sections

Figures (11)

  • Figure 1: Illustration of the Structured NLP (XNLP) tasks, and the unification of XNLP by decomposing into the predictions of spans and relations.
  • Figure 2: The overall architecture of our XNLP system includes the frontend module and the backend module.
  • Figure 3: Screenshot of the XNLP web application, where key functions are annotated.
  • Figure 4: Structure formatter for universal XNLP.
  • Figure 5: Screenshots of the visualizations of 12 representative XNLP tasks. Best viewing with zooming in.
  • ...and 6 more figures