Aspect Based Sentiment Analysis

I want to detect aspect-based sentiment. e.g. He is going to New York instead of Nevada. In this case, I want New York as -ve entity and Nevada as +ve entity.

Hi @abhishekborse, I am currently looking into composite entities (something, for example, mentioned here: Meta entities by amn41 · Pull Request #3889 · RasaHQ/rasa · GitHub). With composite entities you are able to assign specific “roles” to entities. For example, you could annotate New York and Nevada with He is going to [New York](ve@negative) instead of [Nevada](ve@positive). ve would be the entity and negative and positive the roles. The idea is that you can later on access the entities together with their “roles” in custom actions etc. Would that help?

We are still developing the approach and testing it. In the meantime, you could try to annotate your training examples like He is going to [New York](ve-) instead of [Nevada](ve+). Instead of one entity with different roles, you would simply have two different entities. Just try to train the model and see how good the results are. In our experience they are actually not that bad.

In order to probably test our approach, we are still looking for good datasets. Do you have by any chance a good dataset with the above mentioned annotations? If yes, would you be able to share the dataset with me?

1 Like