Mining Argumentative Structure from Natural Language text using Automatically Generated Premise-Conclusion Topic Models

This paper presents a method of extracting argumentative structure from natural language text. The approach presented is based on the way in which we understand an argument being made, not just from the words said, but from existing contextual knowledge and understanding of the broader issues. We leverage highprecision, low-recall techniques in order to automatically build a large corpus of inferential statements related to the text’s topic. These statements are then used to produce a matrix representing the inferential relationship between different aspects of the topic. From this matrix, we are able to determine connectedness and directionality of inference between statements in the original text. By following this approach, we obtain results that compare favourably to those of other similar techniques to classify premise-conclusion pairs (with results 22 points above baseline), but without the requirement of large volumes of annotated, domain specific data.