Building a Twitter opinion lexicon from automatically-annotated tweets

A method that combines information from automatically annotated tweets and existing hand-made opinion lexicons to expand an opinion lexicon in a supervised fashion that outperforms the three-dimensional word-level polarity classification performance obtained by using PMI-SO alone.

Abstract

Opinion lexicons, which are lists of terms labelled by sentiment, are widely used resources to support automatic sentiment analysis of textual passages. However, existing resources of this type exhibit some limitations when applied to social media messages such as tweets (posts in Twitter), because they are unable to capture the diversity of informal expressions commonly found in this type of media. 
\nIn this article, we present a method that combines information from automatically annotated tweets and existing hand-made opinion lexicons to expand an opinion lexicon in a supervised fashion. The expanded lexicon contains part-of-speech (POS) disambiguated entries with a probability distribution for positive, negative, and neutral polarity classes, similarly to SentiWordNet. 
\nTo obtain this distribution using machine learning, we propose word-level attributes based on (a) the morphological information conveyed by POS tags and (b) associations between words and the sentiment expressed in the tweets that contain them. We consider tweets with both hard and soft sentiment labels. The sentiment associations are modelled in two different ways: using point-wise-mutual-information semantic orientation (PMI-SO), and using stochastic gradient descent semantic orientation (SGD-SO), which learns a linear relationship between words and sentiment. The training dataset is labelled by a seed lexicon formed by combining multiple hand-annotated lexicons. 
\nOur experimental results show that our method outperforms the three-dimensional word-level polarity classification performance obtained by using PMI-SO alone. This is significant because PMI-SO is a state-of-the-art measure for establishing world-level sentiment. Additionally, we show that lexicons created with our method achieve signifi- cant improvements over SentiWordNet for classifying tweets into polarity classes, and also outperform SentiStrength in the majority of the experiments.