Better emotion analysis #6
Description
The current approach for using is through lexical matching. With things like LSTMs and Deep RL out there this approach's accuracy pales in comparison.
1- Of particular concern is it's inability to handle negations - i.e. "Not very good" and "very good" are given the same joy score
2- The other big issue is not taking into account the context of a text message. i.e. the final text in this conversation:
- Developer: Is stackoverflow down?
- Octocat: No
- Developer: Fantastic
Will be given the same joy score as the final sarcastic text in this conversation:
- Developer: Is stackoverflow down?
- Octocat: Yes
- Developer: Fantastic
After a quick search, it seems there indeed isn't any pre-trained model out there that has what I'm looking for (the closest is Deepmoji, but it suffers from problem 2), and so this means building a model from scratch - practically a research undertaking in and of itself.
One approach that immediately comes to mind is modeling texts as an Markov Decision Process, with state being the texts so far in a texting session, the action predicting an emotion score, and a reward corresponding to how accurate the emotion prediction was.
Of course there's probably a lot of other approaches (feel free to suggest and discuss in depth, would love to learn more), but regardless the first step is to gather enough data and label it. Perhaps some kind of "human in the loop" training like the one Deepmoji uses