Meme-Sentiment-Analysis

Course project of BITS F312, Neural Networks and Fuzzy Logic

The project attempts to score internet memes on the basis of three measures, offensiveness, sentiment and expressed motivation.

Preprocessing

Images were all cropped to the same size, and standard NLP preprocessing techniques like stemming, lemmatization and hyperlink removal etc. were performed on the text.

Model

A pretrained ResNet is used to extract image features. The text is processed using pretrained GloVe vectors and a couple of stacked Bidirectional LSTM layers. The feature representations of both text and images are then combined. A few more fully connected layers are included to operate upon the combined representations.

Metrics

Since the dataset is imbalanced, an F1 score is more meaningful than accuracy for the classification tasks. The offensiveness task was cast as a regression problem, and the mean square error was used as the loss metric.