Text-Based Event Detection: Deciphering Date Information Using Graph Embeddings

Abstract

Event detection is increasingly gaining attention within the fields of natural language processing and social network analysis. Graph models have always been integral to social media analysis literature. Owing to the long processing time and time complexities of graph-based algorithms, these models were initially very difficult to improve upon. Over the past few years, researchers proposed many approaches to create representations such as word2vec and doc2vec [11]. With the emergence of graph embedding techniques in recent years using deep learning techniques such as node2vec, it is possible to extract node embeddings that can be used to embed graph information into machine learning methods. We introduce SnakeGraph, a new model which uses the sequences of words making up each body of text along with key representations such as the user and the date. These representations can help us learn about the main ideas communicated via written language. However, our method not only looks at both the content of text and how it links to other key information, but also factors the relationship between words in our text as they appear in sequence and overlap as they appear across different bodies of text. We believe that date and user embeddings can especially shed light on event detection literature.