Aspect Sentiment Triplet Extraction via Disentangled Attention and task-specific graph neural networks
MSc Thesis Proposal by: Asmita Prabhakar
Date: Friday, March 13th, 2026
Time: 12:00 pm – 2:00 pm
Location: EH 186
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 that star ratings collapse into a single score. Existing ASTE solutions such as ASTE-RL21, BMRC-DA-TF24, and ATF-ASTE25 address ASTE through transformer-based and graph neural architectures. ASTE-RL21 uses sequential subtasks causing irreversible error propagation; if "terrible" is missed in "Great food but terrible service," the triplet (service, terrible, negative) is permanently lost. 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. However, BERT combines token meaning and position before attention computation, causing boundary errors. When processing "thermal inefficiency" in "The laptop's proclivity for thermal inefficiency undermines its performance," BERT cannot distinguish whether "thermal" (token embedding) and "inefficiency" are related semantically or by proximity, fragmenting the compound opinion into the incorrect partial triplet (laptop, thermal, negative) instead of (laptop, thermal inefficiency, negative). This thesis proposes DeBERTa-GNN-ASTE, extending ATF-ASTE25 by replacing BERT with pretrained encoder DeBERTa-v3-base, 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 ATF-ASTE25's 4-head GAT, and trains on GPT-4 augmented datasets. Experimental evaluation on ASTE-Data-V2 using SemEval 2014 datasets shows DeBERTa-GNN-ASTE achieves F1 69.08% on 14lap and F1 79.98% on 14res, improvements of +2.65 and +3.51 F1 points over ATF-ASTE25, with recall gains of +5.02 and +5.43 confirming disentangled attention identifies complete multi-word triplets that BERT-based encoding fragments.
Keywords: Aspect Sentiment Triplet Extraction, ASTE, DeBERTa, Disentangled Attention, Graph Attention Networks
