An interpretable contrastive logical knowledge learning method for sentiment analysis

Yulin Chen, Bo Yuan, Dongheng Chen, Dov Gabbay, Beishui Liao

Keywords: interpretable sentiment analysis, Talmudic public announcement logic, contrastive logical knowledge learning, knowledge reasoning

Abstract: Current interpretable sentiment analysis (ISA) methods frequently underperform state-of-the-art models, and few of them cast light on the inner working of pre-trained models. In this work, we fill the gap by addressing four key research challenges in ISA—knowledge acquisition, knowledge representation, knowledge learning and knowledge reasoning—in one unified framework. Theoretically, we propose a novel contrasitive logical knowledge learning (CLK) framework that can visualize the decisions made through deterministic Talmudic public announcement logic semantics. We apply CLK to current popular sentiment analysis models to obtain CLK based interpretable ones. Empirically, experimental results of both binary sentiment analysis tasks and fine-grained sentiment analysis tasks indicate that CLK can achieve an effective trade-off between accuracy and interpretability. Furthermore, we find that CLK can reduce the uncertainty of logical knowledge for discriminative labels by visualizing the learned feature representations and model output. Besides, we carry out a case study to investigate the fidelity of model interpretability through knowledge reasoning, which demonstrates that the explanations provided by our method are reasonable and consistent for sentiment analysis tasks.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)

TL;DR: We present a novel contrastive logical knowledge learning (CLK) method to learn interpretable TPK models and generate explanations for sentiment analysis tasks.