Lexicon-based Non-Compositional Multiword Augmentation Enriching Tweet Sentiment Analysis

One of the benefits of recognizing a slang, an id-iom or an abbreviation in a tweet is the ability to help in finding certain sentiment in a concise and understandable manner. However, a lack of adequate annotated “idiomatic tweets” makes classification challenging. We propose a pliable augmen-tation technique to improve the classification of idiomatic tweets with tiny training samples. For classification, we evaluate the performance of fine-tuning version of a pre-trained embedding model at different flavors. During the augmentation process, we deduce the intrinsic propositional meaning of the idiomatic ex-pression from IBM's SliDE (Sentiment Lexicon of IDiomatic Expressions) and another lexicon we built. The empirical results show that the proposed method is beneficial in concealing the actual intent of the tweet and advantageous to tackle the prob-lem of overfitting caused by smaller training sets. The experi-ment shows that using data augmentation of the idiomatic ex-pressions has reduced the classification error rate with 16%.