Title: Revealing Implicit Knowledge in Form of Single- and Multihop Commonsense Knowledge Paths in Texts
Speaker: Maria Becker (ICL)
In this work we leverage commonsense knowledge in form of single and multihop paths for establishing connections between concepts from different sentences, as a form of explicitation of implicit knowledge. The connections can either be direct (car --> IsA --> vehicle) or require intermediate concepts (waste --> ReceivesAction --> recycle --> PartOf --> environmental protection). For constructing paths we first propose a tool which extracts meaningful concepts from natural language texts and maps them to conjunct concept nodes in ConceptNet, utilizing the maximum of relational information stored in the ConceptNet knowledge graph. For finding paths between the extracted concepts, we then present a framework in which we combine two model types: a relation classifier that predicts direct connections between two concepts; and a target prediction model that generates intermediate concepts given a source concept and a relation, which we use for building multi-hop paths. As opposed to models that retrieve knowledge from static knowledge bases, we dynamically combine the power of language models with knowledge stored in ConceptNet. We design manual and automatic evaluation settings for assessing the quality of the paths. Evaluation on two argumentative datasets shows that with our framework we can generate meaningful, high-quality knowledge paths that connect sentences
and reveal implicit knowledge conveyed by the text.