Machine learned annotation of tweets about politicians' reputation during presidential elections: the cases of Mexico and France.

Abstract

With regular elections challenges, opinion mining on Twitter recently attracted research interest in politics using Information Retrieval (IR) and Nat- ural Language Processing (NLP). However, get- ting language and domain-specific annotated data still remains a costly manual step. In addition, the amount and quality of these annotations may be critical regarding the performance of NLP-based Machine Learning (ML) techniques. An alterna- tive solution is to use cross-language and cross- domain sets to simulate training data. This paper describes ML approaches to automatically annotate Spanish tweets dealing with the online reputation of politicians. Our main finding is that a simple sta- tistical NLP classifier without in-domain training can provide as reliable annotation as humans an- notators can. It also outperforms more specific re- sources such as polarity lexicon or in-domain man- ually translated data.