Aspect-based sentiment analysis with graph convolution over syntactic dependencies.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Aspect-based sentiment analysis (ABSA) aims to identify sentiment towards specific entities or attributes in text.
• The health and well-being domain presents unique challenges, often exhibiting a bias towards negative sentiment in user reviews.
• Existing deep learning models for ABSA require enhancement to handle the nuances of specialized domains like healthcare.

Purpose of the Study:

• To propose a novel model for aspect-based sentiment analysis using graph convolution.
• To evaluate the model's performance on a corpus of drug reviews within the health and well-being domain.
• To compare the effectiveness of graph convolution over dependency parse trees versus flat sequence representations.

Main Methods:

• Developed a novel aspect-based sentiment analysis model utilizing graph convolution.
• Employed dependency parse trees to represent sentence structure for the graph convolution model.
• Utilized a corpus of drug reviews, with aspects grounded in the Unified Medical Language System (UMLS).

Main Results:

• The proposed graph convolution model significantly outperformed standard deep learning architectures for ABSA.
• Graph convolution applied to dependency parse trees achieved a higher F-score (0.8179) compared to a flat sequence representation (0.7332).
• The model successfully addressed the negative sentiment bias common in health and well-being text.

Conclusions:

• Graph convolution over dependency parse trees is a highly effective approach for aspect-based sentiment analysis.
• This method enhances sentiment analysis performance in the health and well-being domain, aligning it with other fields.
• The study demonstrates the potential of graph-based methods for nuanced sentiment analysis in specialized textual data.