What is the Tunable Neural Sentiment Detection Model?
Up to this point there were two general options for sentiment analysis inside the platform: the Lexicon-based Sentiment Analysis and the Neural Network based Analysis. Out-of-box lexicon-based sentiment analysis has an accuracy of about 70%, however this analysis can be tuned by a custom sentiment dictionary to achieve higher accuracy up to 85%. NN (Neural Network) based sentiment models typically have a higher out-of-box accuracy of about 85%+. The tradeoffs are that this type of analysis is not tunable in the platform, and, in some cases, users can spot individual, yet obvious errors, which discredit the overall model results.
The goal of the Tunable Neural Sentiment Detection Model is to take advantage of our NN sentiment model’s high accuracy and add in the capability to fine tune the analysis by leveraging a custom sentiment dictionary. This feature allows a way to achieve about 90% sentiment accuracy.
When should I use Tunable Neural Sentiment Detection Model?
There are several reasons to consider leveraging this feature:
- You are currently not satisfied with the out-of-box sentiment accuracy from either the Lexicon or the NN approach.
- Your use-case requires exceptional sentiment accuracy.
- Sentiment scoring within your industry / use-case is unique.
How to apply this model
There are 2 parts to leveraging this feature: base-line neural network model, and a sentiment list.
- Choose a base-line neural network model from the available options in the drop-down or contact Stratifyd to train one for your specific use-case.
2. Create and attach a sentiment list to the analysis in the ‘advanced settings’:
- Sentiment words: custom words
- Negation words: words that invert the value of next sentiment word
- Booster words: words that increase the value of the next sentiment word
- The feature works best when a customized neural network model has been trained from use case specific data, and a well-crafted customized sentiment list has been attached.
- An iterative approach when crafting the custom sentiment list is recommended.
- This feature is a little more complicated and might require a bit more collaboration with Straitfyd Solutions Team for exceptional results. Don’t hesitate to reach out!
Please sign in to leave a comment.