Causal Maps for Multi-Document Summarization

Concept maps are concise graphical representations of text data which have been shown to be applicable as a tool for text summarization and downstream tasks. Most prior work has either focused on the generation of concept maps for small corpora or require expensive training data to implement. In this work, we focus on generating causal maps, a subset of concept maps in which only semantically causal relationships are considered. We propose a map generation framework which utilizes a novel mixture model to simultaneously derive concepts and links. This method is computationally efficient and therefore scalable to large datasets, and is fully unsupervised, which makes it suitable for practical applications. We show that our method performs better than a commonly used unsupervised text summarization algorithm, and has results which are comparable to the state-of-the-art supervised method.