State-of-the-art in Sentiment Analysis accepted to WASSA 2017
What is State-of-the-art in Sentiment Analysis?
State-of-the-art is a tricky concept. In sentiment analysis, which approach works best often depends on the data you have at hand, whether your interested in knowing the general sentiment of a document or sentence, which is dominated by neural networks, or if you want to know what the sentiment is of a specific target entity, where an ensemble of techniques often gives the best results.
Another obstacle is that we often don’t have the time or energy to devote to reimplementing others’ approaches and rely on the results they present in their papers.
We were interested in seeing if any particular model was better on a certain kind of data, or if there was one model that would perform better overall. We decided to choose 7 standard sentiment models and look more at some of the hyperparameters which are often ignored.
We ran 7 different models on 6 different datasets with different characteristics (you can find the details here).
The first major surprise is that the old bag-of-words approach works amazingly well across datasets. In fact, it comes very close to the best performing algorithms on almost all of the datasets, as you can see in the next table.
Rather unsurprisingly given their performance on so many tasks these days, bidirectional LSTMs are the big winners.