Major Entity Identification: A Generalizable Alternative to Coreference Resolution

TL;DR

This work reframes coreference resolution as Major Entity Identification (MEI), a task that assumes a predefined set of major entities and focuses on detecting and labeling mentions that refer to them. By offloading domain adaptation to inference and modeling MEI as a classification problem, the authors demonstrate stronger cross-domain generalization than traditional CR across literary datasets, and show MEI is amenable to robust evaluation with simple metrics. They propose MEIRa, a supervised MEI architecture with document encoding, mention proposals, and a memory-based identification module, plus a fast variant MEIRa-S. They further show that few-shot prompting with LLMs—especially GPT-4—can approach or match supervised MEIRa performance in end-to-end MEI, given a two-stage prompting strategy (word-level MEI followed by head-to-span retrieval). Overall, MEI offers practical benefits for entity-centric retrieval and long-context referential tasks, with scalable, parallelizable inference and potential for broad applicability in literature, film, finance, and legal domains.

Abstract

The limited generalization of coreference resolution (CR) models has been a major bottleneck in the task's broad application. Prior work has identified annotation differences, especially for mention detection, as one of the main reasons for the generalization gap and proposed using additional annotated target domain data. Rather than relying on this additional annotation, we propose an alternative referential task, Major Entity Identification (MEI), where we: (a) assume the target entities to be specified in the input, and (b) limit the task to only the frequent entities. Through extensive experiments, we demonstrate that MEI models generalize well across domains on multiple datasets with supervised models and LLM-based few-shot prompting. Additionally, MEI fits the classification framework, which enables the use of robust and intuitive classification-based metrics. Finally, MEI is also of practical use as it allows a user to search for all mentions of a particular entity or a group of entities of interest.

Figures (6)

• Figure 1: CR vs. MEI. The CR task aims to detect and cluster all mentions into different entities, shown in various colors. MEI takes major entities as additional input and aims to detect and classify the mentions that refer only to these entities.
• Figure 2: Identification module of MEIRa. A mention encoding $\bm_i$ is concatenated with each entity's embedding in $\mathcal{E}^W$ and the metadata $\chi(m_i, e_j)$. Network $f$ scores the likelihood of assigning $m_i$ to each major entity. If the highest score $s_i^*$ is above the threshold $\tau$, $m_i$ is associated with the highest scoring major entity $e_i^*$ or discarded. In MEIRa-S, the entity memory $\mathcal{E}^W$ remains static. For MEIRa-H (blue path), the assigned entity's working memory is updated, and both the static (top half) and updated working memory (bottom half) are utilized to compute a final score.
• Figure 3: Linking speed comparison between MEIRa-S and longdoc for the combined LitBank and FantasyCoref test set. There exists 6 documents with (0, 100] mentions, 19 with (100, 500] mentions, 5 with (500, 1250] mentions and 3 with (1250, 2500] mentions.
• Figure 4: Performance of MEIRa-H across number of target entities for the document Ali Baba and the Forty Thieves.
• Figure 5: Performance of MEIRa-H across number of target entities for Aladdin.