Identification of Tweets that Mention Books: An Experimental Comparison of Machine Learning Methods

摘要

In this paper, we address the task of the identification of tweets on Twitter that mention books (TMB) among tweets that contain the same strings as full book titles. Although this task can be treated as a kind of Named Entity Recognition, the fact that book titles consist of ordinary expressions (such as "The Girl on the Train") makes the task harder. Furthermore, if tweets are gathered through a dictionary-based search, the tweets that contain the same strings as full book titles are often spam. However, assuming a complete list of book titles (i.e. from a union catalogue from a library or commercial bibliographic data from a book store), this task can be solved by text classification. Thus, we proposed a two-step pipeline consisting of spam filtering and TMB classification based on supervised learning with a small amount of labelled data. We constructed optimal classifiers by comparing combinations of four proven supervised learning methods with different features. Given the difficulty of the task, our pipeline performed highly (about 0.7 in terms of F-score).