A Comprehensive Neural and Behavioral Task Taxonomy Method for Transfer Learning in NLP
Yunhao Zhang, Chong Li, Xiaohan Zhang, Xinyi Dong, Shaonan Wang · 2023
Transfer learning is frequently utilized in scenarios with limited labeled examples, where a crucial step is to identify a related task to the target task.CogTaskonomy (Luo et al., 2022) was proposed to acquire a taxonomy of NLP tasks, specifically focusing on assessing the similarities between tasks.This method, inspired by cognitive processes, exhibits notable time efficiency.Nevertheless, it does not fully exploit the task-related information present in cognitive data and lacks a comprehensive evaluation of various types of cognitive data.To address these limitations, this paper proposes a comprehensive neural and behavioral method to investigate the relationship among NLP tasks.Our approach utilizes cognitive data, encompassing both neural data such as fMRI and EEG, as well as behavioral data including eye-tracking and semantic feature ratings.Each data modality is employed to establish a common representation space with Representation Similarity Analysis for projecting task-related representations.To fully leverage the cognitive information, we effectively extract the task-relevant information extracted from neural data through feature ranking.Experimental results on 12 NLP tasks demonstrate that our proposed method outperforms state-of-the-art methods on evaluating task similarity.