5ad742cd15633b26fdce1b80f7b39f7c-AuthorFeedback.pdf

Comment 3: Eq. (10) is slightly different from [18] in the first term as we hope to better control of the lengths for22 apple-to-applecomparison. RNPisgoodatfinding the30 aspects that correlate most closely with the output labels, but lacks class-specific capabilities; whereas POST-EXP31 isgood atfinding class-specific words and phrases, butdoes poorly infinding the right aspect. The quantization operation squeezes out information, and deteriorates the performance of the RNP baseline. For our method on factual rationales, the sentiment agreement is 0.71 on beer reviews, 0.79 on hotel45 reviewsand0.81onAmazon