MEIC-DT: Memory-Efficient Incremental Clustering for Long-Text Coreference Resolution with Dual-Threshold Constraints

TL;DR

MEIC-DT tackles the memory bottleneck in long-text coreference resolution by introducing a dual-threshold, memory-efficient incremental clustering framework based on a lightweight Transformer. It combines a statistics-aware eviction strategy (SAES) that adapts cache management between training and inference with an internal regularization policy (IRP) that condenses clusters to representative mentions. The approach achieves competitive CoNLL-style F1 scores across OntoNotes, LitBank, and WikiCoref under strict memory budgets, while also accelerating convergence and maintaining efficiency. This memory-conscious design enhances practicality for long documents and demonstrates robust cross-domain generalization in CR tasks.

Abstract

In the era of large language models (LLMs), supervised neural methods remain the state-of-the-art (SOTA) for Coreference Resolution. Yet, their full potential is underexplored, particularly in incremental clustering, which faces the critical challenge of balancing efficiency with performance for long texts. To address the limitation, we propose \textbf{MEIC-DT}, a novel dual-threshold, memory-efficient incremental clustering approach based on a lightweight Transformer. MEIC-DT features a dual-threshold constraint mechanism designed to precisely control the Transformer's input scale within a predefined memory budget. This mechanism incorporates a Statistics-Aware Eviction Strategy (\textbf{SAES}), which utilizes distinct statistical profiles from the training and inference phases for intelligent cache management. Furthermore, we introduce an Internal Regularization Policy (\textbf{IRP}) that strategically condenses clusters by selecting the most representative mentions, thereby preserving semantic integrity. Extensive experiments on common benchmarks demonstrate that MEIC-DT achieves highly competitive coreference performance under stringent memory constraints.

Figures (4)

• Figure 1: Analysis of motivations om the LitBank training corpus. Notably, the "unbound" condition failed due to out-of-memory (OOM) errors under our memory constraints.
• Figure 2: The MEIC-DT Coreference Resolution pipeline. The core innovation is a dual-threshold constraint mechanism---comprising a statistics-aware eviction strategy and an internal regularization policy---that enables high-performance, memory-efficient incremental clustering.
• Figure 3: Learning curves and total training time. Results are shown for configurations with $\tau_1/\tau_2=50/-$ and $50/30$, across different datasets (LitBank, OntoNotes) and classifier backbones.
• Figure 4: An example of semantic space distributions for different sampling strategies on OntoNotes.