LingMess-Coref
| Property | Value |
|---|---|
| Authors | Shon Otmazgin, Arie Cattan, Yoav Goldberg |
| Paper | arXiv:2205.12644 |
| License | Creative Commons Attribution 4.0 International |
| Best Performance | 81.4 F1 on OntoNotes 5.0 |
What is lingmess-coref?
LingMess is an innovative approach to coreference resolution that employs linguistically motivated categorization of mention-pairs into six distinct types of coreference decisions. The model implements dedicated trainable scoring functions for each category, resulting in significantly improved accuracy in both pairwise scoring and overall coreference resolution tasks.
Implementation Details
The model builds upon the Longformer-large architecture and introduces a novel multi-expert scoring system. It achieves state-of-the-art performance with an impressive 81.4 F1 score on the OntoNotes 5.0 dataset, surpassing previous approaches like SpanBERT-large (79.6 F1) and standard Longformer-large implementations (80.3 F1).
- Utilizes 6 specialized scoring functions for different types of coreference decisions
- Built on Longformer-large architecture
- Implements linguistically informed categorization
- Achieves superior performance on OntoNotes benchmark
Core Capabilities
- Advanced coreference resolution across multiple mention types
- Improved accuracy in pairwise mention scoring
- Robust performance on long-form text analysis
- Linguistically motivated decision making
Frequently Asked Questions
Q: What makes this model unique?
LingMess stands out for its linguistic categorization approach, dividing coreference decisions into six distinct types and employing specialized scoring functions for each category. This multi-expert approach leads to significantly better performance compared to traditional methods.
Q: What are the recommended use cases?
The model is particularly well-suited for tasks requiring high-accuracy coreference resolution in English text, especially in academic and professional contexts where precise reference resolution is crucial. It's ideal for applications in natural language understanding, document analysis, and information extraction systems.
