How to fine grain neutral sentiment as positive or negative

I'm working on multimodal sentiment analysis with visual and textual cues.

My input dataset is containing neutral sentiment in ground truth but I require to do a binary classification to categorize my input samples as either positive/negative

Is there any possibility to use this neutral class in aiding to remove non-opinion key terms thereby increasing the accuracy of binary categorization?

Is it advised only to adopt a multi-class classification algorithm to categorize as positive, negative or neutral?

P.S: My requirement is to do a binary classification

Thanks in advance

Solution

If your requirement is to do obligatory binary classification, maybe it is worth to perform a hierarchical analysis in two binary classification steps. First you classify the documents into objectives(neutral) or subjectives(positives and negative). Then, for the subjective ones, you classify into positive or negative.

Otherwise, a better way is to simply work on multi-class classification and classify into three classes.