Aspect-Based Sentiment Analysis (ABSA) is the task of identifying aspect terms and categories from a given sentence, and to then associate a sentiment polarity to each of them.
It can be also seen as composed by 4 different sub-tasks, namely:
- (A) Aspect term identification
- (B) Aspect term polarity classification
- (C) Aspect category identification
- (D) Aspect category polarity classification
In this project we developed several approaches and carried out experiments to jointly solve tasks A and B (A+B) and then we adapted the same architecture to also solve C and D (C+D) together. The architecture is based on 2 stacked BiLSTMs and Attention layers, leveraging PoS, GloVe and BERT (frozen) embeddings.
For further information, you can read the detailed report or take a look at the presentation slides (pages 10-18).
This project has been developed during the A.Y. 2020-2021 for the Natural Language Processing course @ Sapienza University of Rome.
- Word-in-Context (WiC) disambiguation as a binary classification task, experimenting with a word-level approach (MLP + ReLU) and a sequence encoding one (LSTMs), on top of GloVe embeddings
- Word Sense Disambiguation (WSD) for WiC disambiguation, experimenting with BERT feature-based and fine-tuning approaches (GlossBERT)