Until the 2000s, the concept of non-statistical methods was used to solve the problem of automatic extraction of causal relationships (CR). These methods used manually constructed linguistic templates. Obviously, the CR that did not fit into the built templates could not be defined. Non-statistical methods required constant manual control by experts, up to the evaluation. Almost all methods were aimed at extracting explicit CR. In some methods, attempts were made to untie the extraction system from a specific subject area. To eliminate the above disadvantages, the methods developed in the future began to shift towards statistical data processing and machine learning. In this article, statistical and machine methods of CR e xtraction are considered. A few valuable papers related to the new paradigm of CR extraction were analyzed. The aim of the research was to evaluate new methods with the ability to identify their advantages and disadvantages. The great advantage of machine and statistical methods is independence from the subject area while maintaining the accuracy of extraction. Such methods are worse in accuracy, but they are not tied to a specific problem area. The methods themselves, unlike non-statistical ones, which used linguistic and syntactic comparison with templates manually, are focused on finding these templates. Even though machine and statistical methods are mostly independent of the subject area and use large corpora oftext for teaching, they are intended mainly for the English language. There is also no standardized data set that would allow methods to be compared with each other. All works devoted to methods ignored the extraction of implicit CR.
Most of the first attempts to extraction of causal relationship were tied with complex and manual linguistic patterns, syntactic rules and small datasets based on domain. This article examines the paradigm of a non-statistical approach to the extraction of causal relationships, its basis, language constructs, patterns, and classification of causal relationships. The aim was to study the methods of this paradigm, to determine their disadvantages, advantages, and the possibility of their application.The article discusses various approaches given by the authors of well-known and highly cited research papers and their impact on the success of the extraction of causal relationships. The analysis of these scientific papers has unequivocally confirmed that the task of extracting CR is an extremely difficult task of natural language processing. The presence of a variety of linguistic constructions of the language, ambiguities of various kinds, as well as language features greatly affect the accuracy of CR extraction. Almost all non-statistical methods have encountered the problem of highly specialized fields of knowledge, where expert description is almost always required. Also, almost all non-statistical methods are manual or semi-automatic, because assume the construction of templates for determining the CR in the text. Even though non-static methods with sufficient accuracy (on average 70-80%) successfully cope with the task under consideration, there is currently no universal method for extracting CR. The proposed method should be universal with respect to languages, universal with respect to subject areas and with the possibility of defining implicit CR.
