CoRBS: a dynamic storytelling algorithm using a novel contextualization approach for documents utilizing BERT features

Storytelling is the process of connecting documents one after another representing the evolution of an event. Existing algorithms for storytelling connect events based on content overlaps between consecutive documents ignoring the role of the same term in different documents and the contemporary contexts (e.g., dynamic embeddings) of documents and terms. Due to the lack of role and contemporary contexts in the designs of the existing storytelling methods, the resultant stories frequently jump to documents with the keywords to form a chain but not a meaningful one.

In this paper, we present a novel storytelling algorithm—Contextual Role-Based Storytelling (CoRBS)—that generates a chain of documents explaining the evolution of an event, addressing role and contemporary context issues of existing methods. CoRBS starts with a given document and moves forward temporally, stitching together role and context-driven documents to represent the evolution of the events that appear in the first document. We define the role of a term in a document as a distribution of similarities of the nearest neighbors of the term based on BERT embeddings of all terms of that document. Contemporary contexts are incorporated as a mechanism to discover a coherent next document while the story progresses.

Our experiments demonstrate that CoRBS generates more meaningful stories compared to other baseline storytelling techniques.