Causality mining by candidate events extension and multi-level background knowledge

Mining implicit and ambiguous causality in natural language texts has been a core topic for decades and plays a fundamental role in information retrieval, question answering, decision making, and future event prediction. Usually, in the past several traditional, machine learning, and deep learning approaches for causality mining plays a significant role, but the performance is not very satisfactory due to informal, implicit, and ambiguous expression of causality without any explicit signal in the natural language source corpora, which kept it hard. Focusing implicit and ambiguous causality requires deep neural models with multi-level features. In this study, we combine two modules for effective causality mining, named Causality Mining by Candidate Event-Extension and prior Background Knowledge (CMCE+BK). The CMCE module is used to focus on extended segment and connective level features. Extended segments and connective describe the extended nature of candidate segments and connective through causal keywords extracted from the context word bank. The background knowledge (BK) module is used to enhance the model ability to perceive causally related BK in sentences, which further strengthens the key features of causality at the segment and connective levels. The experiments and ablation studies on the AltLexes corpus indicate that the extended nature of contextual words and multi-level BK enhanced precision by a maximum of 21.81%, F1-score 57.42%, accuracy 22.15%, and recall 66.73% on the training dataset. Similarly, on the Bootstrapped training dataset, low precision of 15.16% is achieved, but the maximum F1-score of 68.04%, accuracy of 22.67%, and recall of 55.68% are achieved over state-of-the-art implicit causality and text mining techniques.