Argument Extraction Based on the Indicator Approach

The article considers an indicator approach to the extraction of arguments found in popular science literature. The types of argumentation indicators and their relationship to the set of discourse markers are presented, and a method of compiling an indicator dictionary is given. An approach to the use of argumentation indicators in deep learning methods based on the analysis of indicator contexts is proposed. To build a training sample for each indicator, the main statement is extracted, as well as the left and right contexts usually represented by neighboring sentences. For each set the presence of argumentation is marked up according to the corpus annotation. To build the classifier, a list of 143 indicators of argumentation and a marked-up corpus including 162 articles of the popular science genre hosted on the ArgNetBank Studio web platform were used. In total, about 4600 training contexts were obtained on the basis of the corpus. The results of the experiments on argument mining showed the best performance of the classifier based on indicators.

Automated summary

The article presents an indicator approach to extracting arguments in popular science literature and proposes a deep learning method based on the analysis of indicator contexts. By constructing training samples and using a classifier, the experiment on argument mining shows the best performance of the classifier based on indicators.