A multi-modal and multi-scale emotion-enhanced inference model based on fuzzy recognition

Abstract Only the label corresponding to maximum value of fully connected layer is used as output category when a neural network performs classification tasks. When close sub-maximum value, obtained by considering only and ignoring not completely accurate. To reduce noise improve accuracy, combining principles fuzzy reasoning, this paper integrates all results with emotional tendency text based on dictionary establish multi-modal recognition emotion enhancement model. The provided model considers effect negative words, degree adverbs, exclamation marks, question marks smallest subtree defines global membership function emojis corpus. Through comparing CNN, LSTM, BiLSTM GRU Weibo Douyin, it shown that can effectively result clear, especially for long texts.