Towards a neural extraction framework

In the large majority of cases in DBpedia, it is not clear what kind of relationship exists between the entities. Instead of extracting the triples from semi-structured data only, we want to leverage information found in the entirety of a Wikipedia article, including page text. The goal of this project is to develop a framework for predicate resolution of wiki links among entities, specifically, we focus on the direct cause-effect relations between events. Our task then is to extract the cause-effect entity pairs (e.g., Peaceful_Revolution, German_reunification) from the wikipedia text. We combine the idea of using the seed data (e.g., the known cause-effect entity pairs) with training a classifier (e.g., a discriminative model--- LSTM ), so as to discover more cause-effect entity pairs from wikipedia text, which is known as distant supervised relation extraction. The procedures include the pattern matching, knowledge exploration, entity recognition, entity mapping, etc. Eventually, we aim to acquire more reliable causal relations between entities of DBpedia.