APPLYING DEEP LEARNING TO EXTRACT CAUSALITY FROM TEXT USING SYNTHETIC DATA

Abstract

This article addresses the problem of developing a causal full-tuples extraction model from unstructured texts to represent decision-making situations in complex social and humanitarian environments. We present a causal full-tuples extraction model using a pre-trained BERT with additional feature-based special fine-tuning. To refine the causal classification, the model uses two types of features (verb causality and cause-and-effect quality metrics) to recognize a causal tuple, automatically extracts semantic features from sentences, increasing the accuracy of extraction. Text preprocessing is performed using the open source SpaCy library. The extracted cause-and-effect tuples in the format <cause phrase, verb phrase, effect phrase, polarity> are easily transformed into the corresponding elements of the graph <outgoing graph node, graph arc direction, incoming graph node, graph connection weight sign> and can then be used to construct a directed weighted signed graph with deterministic causality on arcs. In order to reduce dependence on external knowledge, synthetic generated annotated datasets are used to fine-tune and test the BERT model. Experimental results show that the accuracy of extracting cause-and-effect relationships on synthetic data reaches 94%, and the F1 value is 95%. The advantages of the presented technological solution are that the model does not require high operating costs, is implemented on a computer with standard characteristics, uses free software, which makes it accessible to a wide variety of users. It is expected that the proposed model can be used to automate text analysis and support decision-making in conditions of high uncertainty, which is especially important for social and humanitarian environments.