Aspect Sentiment Triplet Extraction via Disentangled Attention and Task-specific Graph Neural Networks
MSc Thesis Defense by: Asmita Prabhakar
Date: Friday, June 19th, 2026
Time: 12:00pm to 2:00pm
Location: Essex Hall 122
Abstract:
Aspect-Based Opinion Mining (ABOM) analyzes customer feedback (reviews) to identify sentiments associated with specific product or service features (aspects). Traditional ABOM approaches relying on pre-defined rules struggle when a single review expresses multiple aspects and sentiments. In contrast, Aspect Sentiment Triplet Extraction (ASTE), a subset of ABOM, extracts structured triplets of (aspect, opinion, sentiment) from text. For example, from the review "The food was delicious but the service was slow," ASTE extracts (food, delicious, positive) and (service, slow, negative), preserving fine-grained sentiment structure, which star ratings collapse into a single score. Existing ASTE solutions such as MSFAN22, BMRC-DA-TF24, and ATF-ASTE25 address ASTE through transformer-based (eg. BERT) and graph neural architectures. MSFAN22 relies on multiscale local feature aggregation and span-level modeling but struggles with long-range dependencies; for example, in sentences with distant aspect opinion pairs, it may fail to correctly link them into complete triplets. BMRC-DA-TF24 improves adaptability through GPT-4 augmentation but its BERT extractor fragments multi-word expressions. For example, "solid state drive performance" is reduced to partial aspect "solid," missing tokens that complete the term. ATF-ASTE25 eliminates error propagation and parser dependency by integrating BERT embeddings with Graph Attention Networks built from BIO (Beginning-Inside-Outside) tag predictions, labelling tokens as B, I, or O of aspect or opinion spans, achieving 66.43% F1 on laptop reviews dataset. Since BERT combines token and positional embeddings before attention computation, it lacks an explicit mechanism to distinguish whether adjacent tokens belong together because of semantic coherence or merely because of proximity. This can lead to boundary errors in complex phrases such as "thermal inefficiency" in "The laptop's proclivity for thermal inefficiency undermines its performance," producing incomplete triplets like (laptop, thermal, negative) rather than (laptop, thermal inefficiency, negative). This thesis proposes DeBERTa-GNN-ASTE, extending ATF-ASTE25 system based on the BERT transformer, by replacing BERT with pretrained encoder DeBERTa-v3-base capable of separating content and position through three attention components: content to-content measuring semantic similarity, content-to-position enabling tokens to query relative distances, and position-to-content allowing position to query content. This correctly identifies "thermal inefficiency" as a coherent compound opinion linked to "laptop" seven tokens earlier, producing (laptop, thermal inefficiency, negative). The system generates position-aware embeddings via DeBERTa-v3-base, predicts BIO tags through three classification heads, constructs tripartite graphs with distance-weighted edges, applies 4-head GAT, and trains on GPT-4 augmented datasets. Experimental evaluation on ASTE-Data-V2 using SemEval 2014 datasets shows DeBERTa-GNN-ASTE improves F1 on 14lap (+2.65), 14res (+3.51) and 15res (+1.62), with only a slight drop on 16res (-0.68) over ATF-ASTE25, showing its effectiveness in capturing multi-word triplets compared to BERT encoding.
Keywords: Aspect Sentiment Triplet Extraction, ASTE, DeBERTa, Disentangled Attention, Graph Attention Networks
Thesis Committee:
Internal Reader: Dr. Muhammad Asaduzzaman
External Reader: Dr. Karim Malik (School of Environment)
Advisor: Dr. Christie Ezeife
Chair: Dr. Abdulrauf Gidado (CS Algoma University, adjunct CS UWindsor)
