DiscoExplorer: An Open Interface for the Study of Multilingual Discourse Relations

Technical Summary: DiscoExplorer

Problem Statement
The study of discourse relations (e.g., CAUSE, CONCESSION) across languages is hindered by the fragmentation of data. While recent efforts like the DISRPT Shared Task have successfully unified relation inventories across 38 datasets in 16 languages, researchers lack accessible tools to query, visualize, and compare this data. Existing tools are often limited to sentence-level searches, tied to specific theoretical frameworks (e.g., RST or PDTB), or rely on heavy server-side architectures (like ANNIS) that are slow and complex to deploy. There is a specific gap in providing a lightweight, client-side interface that allows for cross-linguistic comparison and pattern identification without requiring dedicated server infrastructure.

Methodology
The authors present DiscoExplorer, a client-side web interface built entirely in pure JavaScript (React) that runs locally in the browser without a database backend or dedicated indexing libraries.

• Architecture: The system loads datasets entirely into the browser's main memory. It avoids heavy dependencies (e.g., D3.js, Meilisearch) to ensure stability and low compute costs. The data model treats discourse relations as the primary search unit, aligning them to token positions and spanning two argument spans (which may be discontinuous). It preserves context by associating relations with pre-, inter-, and post-context spans.
• Data Integration: The system ingests datasets from the DISRPT 2025 Shared Task, covering 38 datasets from five frameworks (RST, PDTB, SDRT, eRST, and discourse dependencies) across 16 languages. These datasets are mapped to a unified schema where relations carry labels, directionality, and signal tokens (connectives).
• Query Language (DEQL): The interface utilizes a flexible query language (DiscoExplorer Query Language) that supports:
  • Simple text searches and exact sequence matching.
  • Constraints on specific argument spans (source/target) and text order.
  • Integration with Universal Dependencies (UD) annotations (POS tags, dependency labels, lemmas) available in the DISRPT data.
  • Filters for relation labels, signal types, and negation.
• Visualization and Analysis: The interface offers three primary views:
  1. Concordance View: Displays search results with highlighted matches and signals.
  2. Frequencies Tab: Provides quantitative breakdowns (counts, percentages, plots) of categorical variables (labels, signal types) or numerical variables (argument length, token distance). It includes cross-tabulation with chi-squared residual plots to identify significant associations.
  3. Comparison Tab: Allows side-by-side comparison of two datasets, visualizing distribution differences in relation labels or signal frequencies using paired barplots and boxplots.

Key Contributions

1. Open Source Client-Side Interface: A high-performance, pure JavaScript tool that can be run on any PC, eliminating the need for expensive server infrastructure.
2. Public Accessibility: The authors have made the DISRPT 2025 datasets searchable online for the first time via this interface.
3. Unified Query Language: The proposal of DEQL, which balances simplicity for new users with the power to restrict queries by token annotations and argument structure.
4. Cross-Linguistic and Cross-Resource Comparison: The ability to directly compare distributional patterns (e.g., explicitness of signals) across different languages and annotation frameworks (e.g., comparing GUM vs. GENTLE corpora).

Results and Performance

• Data Coverage: The system currently supports over 300,000 relations across 5 million tokens in nearly 10,000 documents.
• Performance Benchmarks: In a timing experiment comparing DiscoExplorer to ANNIS on a consumer laptop, DiscoExplorer demonstrated significantly faster query response times (e.g., 0.022s vs. 3.98s for a simple token search). While ANNIS offers more complex dependency graph queries, DiscoExplorer's in-memory approach provides near-instantaneous responses for relation-based searches after an initial dataset load.
• Resource Usage: Loading large datasets (e.g., the English PDTB with ~1.1M tokens) requires approximately 444 MB of RAM, a manageable footprint for modern browsers.
• Empirical Observations: The authors demonstrate the tool's utility by identifying patterns such as the prevalence of implicit CAUSAL relations versus explicit CONJUNCTION relations, and the specific signal preferences for different subtypes of EXPLANATION relations (e.g., "so" for MOTIVATION vs. "for example" for EVIDENCE).

Significance and Claims
The paper positions DiscoExplorer as a practical solution to the "access gap" in discourse research. Its significance lies in democratizing access to large-scale, multilingual discourse data without requiring specialized server setups. The authors claim that by leveraging the standardized DISRPT format, the tool facilitates the first easy cross-linguistic comparisons of discourse relations.

The paper modestly notes that it does not conduct a formal user study, relying instead on initial feedback from students and plans to gather further input from the CODI and DISRPT communities. It acknowledges that the system is tailored to the DISRPT data model and does not support the complex, multi-layer dependency queries possible in systems like ANNIS. However, the authors argue that the trade-off is justified by the system's speed, ease of deployment, and ability to handle the growing volume of predicted and annotated discourse data generated by multilingual models (such as DeDisCo), which are expected to expand the searchable corpus further.