By mapping messages into a large context, we can compute the distances between them, and then classify them. We test this conjecture on Twitter messages: Messages are mapped onto their most similar Wikipedia pages, and the distances between pages are used as a proxy for the distances between messages. This technique yields more accurate classification of a set of Twitter messages than alternative techniques using string edit distance and latent semantic analysis.
Yegin Genc, Yasuaki Sakamoto, Jeffrey V. Nickerson