Local interpretation of deep learning models for Aspect-Based Sentiment Analysis
Stefan Lam, Ying Liu, Max Broers, Jasper van der Vos, Flavius Frăsincar, David Boekestijn, F. van der Knaap
Abstract
Currently, deep learning models are commonly used for Aspect-Based Sentiment Analysis (ABSA). These deep learning models are often seen as black boxes, meaning that they are inherently difficult to interpret. To improve deep learning models, it is crucial to understand their inner workings. We aim to interpret black box models by implementing model-agnostic local interpretation methods. Inspired by Local Interpretable Model-agnostic Explanations (LIME) and Local Rule-based Explanations (LORE) and combined with a Similarity-based Sampling (SS) method, we propose SS-LIME and SS-LORE, and use Anchor to explain two state-of-the-art ABSA deep learning models. The deep learning models build upon the Left-Center-Right separated neural network with Rotatory attention (LCR-Rot) model, extended by iterating multiple times over the rotatory attention mechanism (LCR-Rot-hop) and hierarchical attention and context-dependent word embeddings (LCR-Rot-hop++). We evaluate the proposed models in terms of fidelity, hit rate, and user interpretability using the SemEval 2016 dataset consisting of restaurant reviews for ternary sentiment classification. Results show that the LCR-Rot-hop and LCR-Rot-hop++ models are best explained by SS-LIME and SS-LORE, respectively. Furthermore, we conclude that the LCR-Rot-hop++ model can be better interpreted than the LCR-Rot-hop model. • We explain the predictions of two deep learning models for sentiment analysis. • We provide interpretation models for ternary classification. • We use post-hoc classifiers with a homogeneous sampling method. • We extend the Submodular Pick algorithm by considering local importance.