💬 NLP

GUMBridge: a Corpus for Varieties of Bridging Anaphora

This paper introduces GUMBridge, a new corpus featuring 16 diverse English genres with granular annotations for bridging anaphora subtypes, while demonstrating that both bridging resolution and subtype classification remain challenging tasks for contemporary large language models.

Original authors: Lauren Levine, Amir Zeldes

Published 2026-03-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Lauren Levine, Amir Zeldes

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are listening to a friend tell a story about a trip to a new city. They say, "I stayed at a hotel. The room was small, but the view was amazing."

You didn't need them to say "the room of the hotel" or "the view from the hotel." Your brain instantly connected the dots. You knew exactly which room and which view they meant because they belonged to the hotel they just mentioned. In the world of language, this mental "bridge" you built is called Bridging Anaphora.

For a long time, computers (AI) have been terrible at building these bridges. They often get confused, thinking "room" is just a random word, not realizing it's part of the "hotel."

This paper introduces a new tool called GUMBridge to help fix that. Here is the simple breakdown:

1. The Problem: The "Old Maps" Were Too Small

Before this paper, researchers had a few old maps (datasets) to teach computers about bridging. But these maps had big holes:

They were tiny: Like trying to learn about the whole ocean by looking at a single puddle.
They were boring: They mostly used old newspaper articles from the 1990s. They didn't have modern slang, podcasts, or funny internet forums.
They were messy: Some maps counted "bridges" differently than others, making it impossible to compare who was the best at solving the puzzle.

2. The Solution: GUMBridge (The "Super-Map")

The authors, Lauren Levine and Amir Zeldes, built a massive new library called GUMBridge. Think of it as upgrading from a single puddle to a high-definition satellite view of the entire ocean.

Huge Size: It contains nearly 6,000 examples of bridging, which is way more than any previous English resource.
Diverse Genres: It doesn't just use news. It includes 24 different types of writing, from academic textbooks and legal documents to vlogs, podcasts, and even live esports commentary. It's like teaching a computer to understand bridging in a courtroom, a kitchen, and a concert hall.
The "Multi-Label" Trick: This is the coolest part. In the past, a bridge could only be one thing (e.g., "part of a whole"). But GUMBridge realized that sometimes a bridge is two things at once.
- Example: If someone says, "I bought a car. The engine is loud," "engine" is a part of the car (meronomy), but it's also a property of the car (it's loud). GUMBridge lets computers tag it as both. It's like realizing a fruit can be both "red" and "sweet" at the same time.

3. The Test: Can AI Do It?

The authors didn't just build the map; they tested the best AI brains (Large Language Models) on it to see if they could learn to cross the bridges.

The Result: Even the smartest AI (like GPT-5) struggled. It got about 40% of the bridges right.
The Analogy: Imagine teaching a robot to read a mystery novel. The robot is great at spotting names and dates, but when the detective says, "The clue was hidden in the lamp," the robot often forgets that the lamp belongs to the room mentioned earlier.
The Good News: While the AI isn't perfect yet, it did better than older, smaller models. This proves that with better data (like GUMBridge), AI can get much smarter at understanding how our words connect.

Why Should You Care?

If AI gets better at "bridging," it means:

Better Chatbots: They won't get confused when you say, "I love my new phone. The battery dies fast." They'll know you mean that phone's battery.
Smarter Summaries: They can summarize a long story without losing the connections between characters and objects.
Deeper Understanding: It helps computers move from just recognizing words to actually understanding the relationships between things, just like humans do.

In a nutshell: The authors built a giant, diverse, and detailed training manual for computers to learn how to connect the dots in language. They showed that while computers are still learning to walk across these bridges, this new manual is the best map we've ever had to help them get there.

1. Problem Statement

Bridging anaphora is a linguistic phenomenon where a newly introduced entity (the anaphor) is interpreted based on a previous, non-identical entity (the antecedent) through an inferential relationship (e.g., "There is a house. The door is red," where "door" is inferred from "house").

Despite its importance for Natural Language Processing (NLP) tasks like question answering and summarization, bridging resolution remains under-resourced compared to coreference resolution. The paper identifies three critical gaps in existing English resources (ISNotes, BASHI, ARRAU):

Limited Scale and Coverage: Existing corpora are small (ranging from ~40k to ~230k tokens) and contain relatively few bridging instances.
Lack of Genre Diversity: Most resources rely heavily on Wall Street Journal (WSJ) news data from decades ago, lacking representation of spoken language, fiction, social media, and other modern genres.
Inconsistent Annotation Schemes: Resources differ fundamentally in their definitions. Some focus strictly on referential bridging (where the antecedent is strictly required for interpretation, e.g., BASHI), while others include lexical bridging (semantic relations like part-whole that may not be strictly anaphoric, e.g., ARRAU). Furthermore, no existing English resource supports multi-subtype annotation (labeling a single instance with multiple bridging categories).

2. Methodology

The authors constructed GUMBridge, a new corpus built upon the existing GUM v12 corpus (a multi-genre English corpus).

Data Construction

Scale: The corpus contains 291,000 tokens and 5,698 bridging instances (approx. 19.6 instances per 1k tokens), covering 24 diverse genres (including academic, fiction, interviews, vlogs, legal documents, and esports commentary).
Annotation Strategy: The authors adopted an Information Status (IS) informed approach (similar to ISNotes and BASHI) to identify bridging. An entity is marked as bridging if it is "Accessible" (inferable) specifically due to a relationship with a previous, non-identical entity, excluding cases solvable by generic world knowledge or coreference.
Subtype Schema: A new, unified schema was developed to categorize bridging varieties, inspired by ARRAU but expanded. It includes 10 specific subtypes grouped into three main categories:
1. Comparison: relative, sense, time.
2. Entity: associative, meronomy (part-whole), property, resultative.
3. Set: member, subset, span-interval.
4. Other: For unique cases not fitting the above.
Multi-Subtype Capability: Unlike previous resources, GUMBridge allows annotators to assign multiple subtypes to a single bridging instance if the relationship is complex (e.g., an instance could be both meronomy and associative).

Annotation Procedure

Human Annotation: The dev and test sets were manually annotated by the authors and trained graduate students. The train and test2 sets were corrected and expanded from previous GUM annotations.
Quality Control: An Inter-Annotator Agreement (IAA) study was conducted on the dev set.
- Anaphor Recognition: Micro F1 score of 0.62, indicating this is the most difficult subtask.
- Antecedent Resolution: 84% accuracy (when anaphors were agreed upon).
- Subtype Classification: Cohen's $\kappa$ of 0.76 (substantial agreement) among expert annotators.

3. Key Contributions

GUMBridge Corpus: The largest and most diverse English resource for bridging anaphora to date, offering 24 genres and over 5,000 instances.
Unified Annotation Framework: It bridges the gap between "referential" and "lexical" bridging definitions by using an IS-based approach while providing granular subtype labels.
Multi-Subtype Annotation: It is the first English resource to formally recognize and annotate multiple bridging subtypes within a single instance, acknowledging the complexity of natural language inference.
Baseline Evaluation: The paper provides the first comprehensive evaluation of Large Language Models (LLMs) on bridging resolution across diverse genres.

4. Results

The authors evaluated three LLMs (GPT-5, Llama-3.3-70B, and Qwen3-235B) on three subtasks: Anaphor Recognition, Antecedent Resolution, and Subtype Categorization.

Model Performance:
- GPT-5 achieved the best results but still struggled significantly:
  - Anaphor Recognition F1: 0.40
  - Antecedent Resolution Accuracy: 0.49
  - Subtype Categorization Accuracy: 0.69
- Open-source models (Llama/Qwen) performed poorly on recognition (F1 ~0.17–0.20) and antecedent selection (Acc ~0.17–0.28), though they performed moderately on subtype classification.
Genre Difficulty: GPT-5 performed significantly worse on spoken genres compared to written genres, highlighting a gap in current models' ability to handle informal or conversational bridging.
Error Analysis: The most common confusion occurred between entity-meronomy (part-whole) and entity-associative relations, both for humans and models.

5. Significance

Benchmarking: GUMBridge establishes a rigorous, genre-diverse benchmark for bridging resolution, moving beyond the limitations of WSJ-only datasets.
NLP Challenges: The results demonstrate that even state-of-the-art LLMs struggle with bridging resolution (F1 < 0.5), indicating that current models lack the necessary world knowledge or discourse reasoning capabilities to reliably infer non-identical entity relationships.
Linguistic Insight: The multi-subtype annotation reveals the complexity of bridging, showing that a single linguistic phenomenon often spans multiple semantic categories, necessitating more nuanced annotation schemes for future research.
Future Directions: The paper highlights that bridging resolution remains an open, challenging task, suggesting that future NLP systems must improve their handling of discourse structure and implicit inference, particularly in spoken and diverse genre contexts.