FAME - A Framework for Argument Mining and Evaluation

Two different perspectives on argumentation have been pursued in computer science research, namely approaches of argument mining from natural language texts on the one hand and formal argument evaluation on the other. So far these research areas are largely independent and unrelated. The overall goal of the FAME project is to link these two perspectives and their respective research agendas. We develop a framework which integrates argument mining and formal argument evaluation. Evaluation results will allow for new types of queries to be answered by argumentation retrieval systems and large-scale content analysis of empirical argument use. Moreover, feeding back evaluation results in the mining process can be utilized to improve the obtained results.

The challenge here is to bridge the gap between processing texts in natural language and evaluating arguments expressed in some formal language with a corresponding formal semantics. The hypothesis underlying this proposal is that controlled natural language (CNL) can provide the necessary link for bridging this gap. CNL can serve as an intermediate representation of argumentative text. They have the look and feel of natural language and are thus close enough to natural language based argument processing. On the other hand, the CNL we will be using is one which also possesses a well-defined formal semantics and is thus amenable to formal evaluation methods. CNLs of the kind mentioned already exist. A prominent example is Attempto Controlled English (ACE). ACE is easily understood by humans. At the same time, its semantics is formally defined based on a translation into discourse representation structures, which can be further translated into predicate logic. Since ACE was not primarily developed with argumentation in mind, we will thoroughly investigate whether the language constructs provided are sufficient for our purposes, or whether we need to extend the language. 

Within the project, we investigate argumentation in public, political discourse represented in newspaper text and related user comments with respect to five selected, controversial issues (e.g. free trade agreements). By modeling specific, empiric issues as CNL statements in a knowledge base, we extend existing approaches of argument mining from identification of generic, functional types of argument structures (e.g. premise, claim) towards semantic argument constituents. At the same time, representing such constituents as formally evaluable statements in a knowledge base allows for combination with approaches of abstract argumentation evaluation. We expect this to be a suitable framework to handle specifics argument structures as they generally occur in empiric communication (e.g. incomplete or inconsistent arguments).