Is sentiment analysis in Twitter old or active research?
May 25, 2018 / Ask Slater
There was a particular surge in Twitter sentiment analysis in the early days of Twitter (pre-2015). It was a somewhat narrow window where this kind of research was extremely exciting because all of the traditional approaches to sentiment analysis relied on more formal corpora and these approaches broke down dramatically when applied to Twitter (as in: our predictions were no longer any better than flipping a coin).
However, things have changed. One of the biggest things that have changed is our approach to sentiment analysis. It has become clear that sentiment analysis is a less consistent and more tractable problem than it used to be with the advent of deep learning techniques. These techniques are very capable of recognizing sentiment with comparable accuracy to a human, meaning that the strict sentiment analysis benchmark of old is less cutting edge and exciting than it once was.
That’s not to say that this is a solved problem, only that the traditional formulation of sentiment analysis (predict a 1 or 0 for positive or negative) is no longer particularly interesting because we’ve collectively gotten too good at the task. When we’re at accuracies close to human accuracy, we start looking for other problems because we can no longer objectively advance.
That’s why when we look at modern problems we both focus on more nuanced understanding and more nuanced tasks. We’re either focusing on things like more nuanced emotional analysis, or more commonly more sophisticated parsing of sentiment involving extracting structured sentiment information from text.
The short is that there is a lot of very interesting research happening in the space of natural language processing, but epsilon research on improving Twitter sentiment analysis is not particularly interesting compared to the problems that we’re trying to solve today. It will be hard to orient a thesis or dissertation around Twitter sentiment analysis unless you’ve got a very interesting new view on it, but generally speaking you have to pick a harder problem to tackle if you want to show serious improvement over state of the art methods.