AR-BERT: Aspect-relation enhanced Aspect-level Sentiment Classification with Multi-modal Explanations

TL;DR

AR-BERT enhances aspect-level sentiment classification by integrating aspect-relation knowledge graphs with a novel embedding scheme and disambiguation techniques, significantly improving accuracy on benchmark datasets.

Contribution

It introduces a two-level entity embedding scheme and disambiguation method for joint training of KG-based aspect representations with ALSC models.

Findings

Achieved a 2.5-4.1% improvement over BERT-based baselines.

Effectively incorporated aspect-relation information from knowledge graphs.

Proposed a novel disambiguation detection technique for aspect entities.

Abstract

Aspect level sentiment classification (ALSC) is a difficult problem with state-of-the-art models showing less than 80% macro-F1 score on benchmark datasets. Existing models do not incorporate information on aspect-aspect relations in knowledge graphs (KGs), e.g. DBpedia. Two main challenges stem from inaccurate disambiguation of aspects to KG entities, and the inability to learn aspect representations from the large KGs in joint training with ALSC models. We propose AR-BERT, a novel two-level global-local entity embedding scheme that allows efficient joint training of KG-based aspect embeddings and ALSC models. A novel incorrect disambiguation detection technique addresses the problem of inaccuracy in aspect disambiguation. We also introduce the problem of determining mode significance in multi-modal explanation generation, and propose a two step solution. The proposed methods show a consistent improvement of 2.5 - 4.1 percentage points, over the recent BERT-based baselines on benchmark datasets. The code is available at https://github.com/mainuliitkgp/AR-BERT.git.