SynerRRL: Synergizing Small and Large Language Models for Rhetorical Role Labeling

ACL ARR 2026 January Submission8307 Authors

Keywords: Sentence-Level Classification, Specialized Domains, Hybrid NLP Models

Abstract: Rhetorical Role Labeling (RRL) assigns a functional role to each sentence in a document and is widely used in legal, medical, and scientific domains. Encoder-based Small Language Models (SLMs) are effective for sentence-level classification but do not capture the broader discourse and domain knowledge encoded by autoregressive Large Language Models (LLMs). We introduce **SynerRRL**, a hybrid framework that leverages the complementary strengths of SLMs and LLMs by aligning their internal representations through a lightweight residual fusion mechanism, without relying on prompting. Experiments with five SLMs, three LLMs, and eight RRL datasets show consistent improvements, yielding average gains of $5.14$ macro-F1 points. An expert-based evaluation shows that SynerRRL improves the classification of rhetorically ambiguous sentences and performs robustly across annotation difficulty levels.

Paper Type: Long

Research Area: Generalizability and Transfer

Research Area Keywords: legal NLP, NLP in resource-constrained settings

Contribution Types: Approaches to low-resource settings

Languages Studied: english

Submission Number: 8307