Year: 2025
Type: article
Abstract: While deep learning models show impressive performance in natural language processing (NLP) tasks, they raise concerns about reliability and ethical implications due to their black-box nature. Consequently, a burgeoning area of research focuses on elucidating NLP models by clarifying the importance of input features to the models’ output predictions. Among the various levels of NLP predictions, sequence-level predictions are crucial for tasks such as named entity recognition, semantic role labeling, and event extraction. However, current explanation techniques have largely overlooked sequence-level predictions. This paper presents an approach to explaining the importance of word features in sequence-level predictions within the semantic role labeling task. Our method evaluates both the impact and usefulness of each word in a sentence regarding model predictions. Additionally, we propose a novel data perturbation mechanism that strategically selects substitution words to more effectively mask word feature information, addressing the limitations of existing data perturbation techniques. Through experiments on biomedical texts, we demonstrate the effectiveness of our explanation method in providing explanations that are both comprehensible to humans and faithful to the actual processing within NLP models.
Author: Hoai Duc Tuan Nguyen
Institution: Ho Chi Minh City University of Science
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4414744827
FWCI:
Cites:
Cited by:
Related to: