SynSem-ICL: syntax–semantic fusion retrieval for structured sentiment extraction with in-context learning

Document Type

Article

Publication Date

1-1-2026

Abstract

Retrieval-based ICL for structured sentiment extraction typically selects examples by semantic similarity, which can fail to preserve correct tuple boundaries under long-distance relations or ambiguous syntax. We propose SynSem-ICL, a retrieval module that explicitly fuses dependency syntax and contextual semantics to select structurally aligned examples for a frozen large language model (LLM). SynSem-ICL encodes dependency parses with a structure-aware graph transformer and performs bidirectional cross-attention fusion. We further train a lightweight retriever with a dual-similarity contrastive objective that leverages semantic and syntactic hard negatives. On ASTE-Data-V2, using five retrieved examples, SynSem-ICL narrows the gap to fully supervised systems to 4.98–7.74 F1 points without downstream fine-tuning. On the SemEval-2022 multilingual benchmark, SynSem-ICL improves Sentiment Graph F1 (SF1) by up to 8.1 points using off-the-shelf Universal Dependencies parses. The code is available at https://github.com/xiay9/SynSem-ICL.git.

Keywords

Contrastive learning, Dependency syntax, In-context learning, Intelligent information systems, Retrieval-augmented prompting, Structured sentiment extraction

Publication Title

Journal of Intelligent Information Systems

ISSN

09259902

DOI

10.1007/s10844-026-01038-7

Share

COinS