Abstract:A complete sentiment analysis of product and service reviews has attracted growing concerns from merchants to enhance personalized marketing activities. Aspect sentiment quadruple prediction (ASQP) is a demanding and challenging task with the objective to predict four sentiment elements from given reviews. Existing methods for ASQP face certain issues, with pipeline-based non-generative approaches prone to error propagation and generative models at the potential risk of producing unexpected outputs or longer inference times. To avoid these shortcomings, we develop a novel end-to-end non-generative model for ASQP involving multi-task decomposition within machine reading comprehension (MRC) framework. Specifically, the ASQP task is decomposed into six query-induced subtasks by introducing task-specific question templates. The proposed model, named MRC-CLRI, is trained with multi-task joint learning. It also incorporates contrastive learning for category identification and sentiment classification to enhance the correlation of the six subtasks. To further promote the quadruple prediction, we present a refined inference algorithm in a bidirectional multi-turn inference procedure to effectively match aspect and opinion terms and optimize two inference hyperparameters: distance threshold and probability threshold. Experimental results exhibit superior performance compared to existing two non-generative and seven generative baselines. Our proposed MRC-CLRI, as a novel non-generative model, outperforms the best existing generative method by an average F1 score improvement of 1.69% and the best previous non-generative method by an average F1 score improvement of 15.77% across four review datasets. Ablation experiments further validate the efficacy of the designed contrastive learning and the refined inference algorithm.

MvP: Multi-view Prompting Improves Aspect Sentiment Tuple Prediction

Prior-Bert and Multi-Task Learning for Target-Aspect-Sentiment Joint Detection

Making Flexible Use of Sub-tasks: A Multiplex Interaction Network for Unified Aspect-based Sentiment Analysis

Few-shot Joint Multimodal Aspect-Sentiment Analysis Based on Generative Multimodal Prompt

SentiPrompt: Sentiment Knowledge Enhanced Prompt-Tuning for Aspect-Based Sentiment Analysis

E2TP: Element to Tuple Prompting Improves Aspect Sentiment Tuple Prediction

Exploiting Duality in Aspect Sentiment Triplet Extraction with Sequential Prompting

Improving Aspect-Based Sentiment Analysis Via Tuple-Order Learning

MCPL: Multi-model Co-Guided Progressive Learning for Multimodal Aspect-Based Sentiment Analysis

Query-induced multi-task decomposition and enhanced learning for aspect-based sentiment quadruple prediction

BvSP: Broad-view Soft Prompting for Few-Shot Aspect Sentiment Quad Prediction

Enhancing Generative Aspect-Based Sentiment Analysis with Relation-Level Supervision and Prompt

A Novel Energy Based Model Mechanism for Multi-modal Aspect-Based Sentiment Analysis

3M: Multi-Task Multi-Prompt Learning Model for Aspect Based Sentiment Analysis

Improve Chinese Aspect Sentiment Quadruplet Prediction via Instruction Learning Based on Large Generate Models

Syntax-aware Hybrid prompt model for Few-shot multi-modal sentiment analysis

Improving Aspect Sentiment Quad Prediction Via Template-Order Data Augmentation.

Mixture of Hybrid Prompts for Cross-Domain Aspect Sentiment Triplet Extraction

Aspect-based Sentiment Analysis Via Multitask Learning for Online Reviews

PGSO: Prompt-based Generative Sequence Optimization Network for Aspect-based Sentiment Analysis

Prompted and Integrated Textual Information Enhancing Aspect-Based Sentiment Analysis