Integrating User Rules into Neural Text Inference

Wei Zheng, YUQING SUN, Yexin Zhang, Bin Gong

Abstract:

Incorporating user-defined heuristic rules into neu-
ral text inference methods has the potential to align models
with user intentions and domain knowledge, thereby improv-
ing interpretability. In this study, we introduce a novel rule
pattern that includes both domain-specific keywords and the
logical relationships between keywords, which can be defined
by users. We propose an approach to integrate explicit rule-
based reasoning with the semantic modeling capabilities of neural
networks. Specifically, our method employs a parallel framework
wherein a neural classifier is trained on labeled text data for
prediction, while a Semantic-Logic Network (SLN) forms rule
inference as a satisfiability problem. We use a Jensen-Shannon
(JS) loss to ensure consistent predictions on both sides for mutual
regularization. The experiment results show that our approach
outperforms baseline methods. We also did ablation analysis on
our method, it shows that the performance of both the SLN and
the classifier contribute to the final results. Additionally, for the
case that lacks explicit user rules, we propose a boosting method
to automatically generate rules from labeled texts which is
beneficial for text inference and improve the model performance