Abstract
Although interpretable methods for deep learning models have become popular in sentiment analysis domains in recent years, existing methods still face the challenge of providing predictions with both high accuracy and user-friendly explanations. To address this problem, we propose a novel framework called Contrasting Logical Knowledge Learning (CLK) that utilizes contrastive learning, label knowledge, and logical rule learning. Logical rule learning is used to provide human-understandable explanations while label knowledge and contrastive learning are used to achieve high performance on both pre-trained models and ordinary DNNs. To ensure model interpretability, we design a novel knowledge reasoning strategy based on learned logical rules and trained models. Empirical results from binary sentiment analysis tasks and fine-grained sentiment analysis tasks show that CLK can effectively balance accuracy and interpretability. Additionally, we conduct two case studies to demonstrate the process of explanation generation and knowledge reasoning, which shows that our method's explanations are causally consistent with the implicit model decision logic.
Original language | English |
---|---|
Article number | 110863 |
Journal | Knowledge-Based Systems |
Volume | 278 |
DOIs | |
State | Published - 25 Oct 2023 |
Externally published | Yes |
Bibliographical note
Publisher Copyright:© 2023 Elsevier B.V.
Funding
The research reported in this paper was partially supported by the National Key Research and Development Program of China ( 2022YFC3340900 ), the “2030 Megaproject” - New Generation Artificial Intelligence of China under Grant No. 2018AAA0100904 , and the National Social Science Fund of China ( 20 & ZD047 ).
Funders | Funder number |
---|---|
New Generation Artificial Intelligence of China | 2018AAA0100904 |
National Key Research and Development Program of China | 2022YFC3340900 |
National Office for Philosophy and Social Sciences | ZD047 |
Keywords
- Contrastive learning
- First order logic
- Interpretable sentiment analysis
- Knowledge reasoning