An robust N-gram causal inference approach based on multi-model fusion

Abstract

Causal inference and interpretation generation are currently hot topics in various fields of research. For instance, by describing the causes of accidents and revealing potential accident rules, we can effectively prevent the recurrence of accidents. However, most models can only capture empirical causal patterns from specific datasets, meaning that their performance deteriorates when they are confronted with new samples or applied to other datasets. In this paper, we propose a robust N-Gram causal inference algorithm based on multi-model fusion, which we refer to as MCRN-Gram. Firstly, we innovatively use the Longest Common Subsequence (LCS) algorithm to generate misleading causal pseudo-samples that we randomly inject during model training. This approach encourages the model to develop a more abstract and nuanced understanding of causal relationships while also improving its robustness. Secondly, we use several effective algorithms for text feature extraction to extract the features of the causal relationship dataset respectively. We then use binary cross-entropy loss and N-Gram loss to optimize the loss function. Finally, we compared the evaluation metrics of the proposed algorithm under different N-Gram window sizes and selected the best window size for conducting experiments on three datasets. Through our experiments, we demonstrate that the proposed method outperforms the comparison algorithms in terms of the precision of causal prediction.