Semantic Role Labeling Guided Out-of-distribution Detection

• DOI: 10.48550/arxiv.2305.18026
• Publication Date: 2023-01-01
• Authors: Jinan Zou; Maihao Guo; Yu Tian; Yuhao Lin; Haiyao Cao; Lingqiao Liu; Ehsan Abbasnejad; Qinfeng Shi
• Keywords: FOS: Computer and information sciences; Computer Science - Computation and Language; Computation and Language (cs.CL)

Identifying unexpected domain-shifted instances in natural language processing is crucial real-world applications. Previous works identify the out-of-distribution (OOD) instance by leveraging a single global feature embedding to represent sentence, which cannot characterize subtle OOD patterns well. Another major challenge current methods face learning effective low-dimensional sentence representations hard that are semantically similar in-distribution (ID) data. In this paper, we propose new unsupervised detection method, namely Semantic Role Labeling Guided Out-of-distribution Detection (SRLOOD), separates, extracts, and learns semantic role labeling (SRL) guided fine-grained local from different arguments of full using margin-based contrastive loss. A novel self-supervised approach also introduced enhance such global-local predicting SRL extracted role. The resulting model achieves SOTA performance on four benchmarks, indicating effectiveness our approach. code publicly accessible via \url{https://github.com/cytai/SRLOOD}.