Analysis has strange things

[alessio-greco]

I didn't expect for AfDKompakt to have a negative sentiment on AfD, for example, and to see FDP more positively.
It's either true, or something strange happened with the sentiment classification ( e.g. assigning a negative sentiment to AfD itself, without context )

[FabianFalck]

Hi @Arakkun, thanks for your interest in our work!

Yes, we agree that these results are surprising. And it would definitely be an interesting effort to further check your hypothesis on the data (note that we published our data here). We used an off-the-shelve sentiment classification tool for these sentiment predictions. Of course, this tool might be biased, but we would have no influence on it. The key idea of our work was to make sentimental and opinion-based judgements more objective, using the objective tool of sentiment models not trained on the data we apply them to.

If you are interested in performing further or related research on this, please do reach out!

[alessio-greco]

The tool used for sentiment analysis is IBM watson, right?

As far as I know, the actual theme by itself is really important for sentiment analysis. Both thematic (sports, videogames, guns, ecc.) and time( "terrific" is one of the best example for that. "This is terrific!" has a positive connotation) have a great deal on whether a word is to be considered positive or negative ( there should be sentiment lexicons from various subreddits, https://nlp.stanford.edu/projects/socialsent/ )
For example, in the case of instant messaging using punctuation and dot has a big impact on knowing whether the person you're chatting with is angry or not, despite punctuation actually being disregarded in normal sentiment analysis.
So there might be some distortion on the classifier being too general in it's scope