Commonsense knowledge relations as represented in ConceptNet are  
crucial for advanced NLU tasks, including argument analysis and the  
reconstruction of implicit premises. In preceding studies we found  
that missing knowledge between adjacent sentences in argumentative  
texts can in most cases be encoded as commonsense knowledge relations  
between concepts that appear in these sentences, such as in the  
following example of two adjacent sentences from the Microtext Corpus:  
(S1) We need independent media. (S2) Everyone should pay for public  
broadcasters. (Relation) public broadcasters HAS_PROPERTY independent.
However, commonsense knowledge relations as represented in ConceptNet  
are difficult to learn due to very specific properties of this  
resource, such as the complexity of argument types and relation  
ambiguity. We examine the learnability of such relations with a neural  
open world multi-label classification approach. Based on an in-depth  
study of the specific properties of ConceptNet, we investigate the  
impact of different relation representations and model variations. In  
contrast to current link prediction systems based on ConceptNet, our  
model assesses classification accuracy for individual relation types  
and achieves overall F1 scores of up to 68 in an open world and 71 in  
a closed world setting.
We then investigate the usefulness of commonsense knowledge for  
argumentation analysis in two different settings: for (1) classifying  
argumentative relations and for (2) reconstructing implicit premises  
in argumentative texts. Argumentative relation classification aims at  
determining the type of relation (e.g., support or attack) that holds  
between two argument units. We find that adding links between premises  
in the form of commonsense knowledge relations as contained in  
ConceptNet and as predicted by our classifier helps the   
attention-based argumentative relation classification model proposed  
by Paul et al. (to appear). Establishing links between concepts in  
argumentative unit pairs is related to the long-standing problem of  
reconstructing implicit premises, so-called enthymemes. In future  
work, we want to tackle this problem by combining the learning of  
inference paths for commonsense knowledge relations with targeted  
information extraction for filling knowledge gaps in argumentative  
texts.