fani-lab / LADy

LADy 💃: A Benchmark Toolkit for Latent Aspect Detection Enriched with Backtranslation Augmentation
Other
5 stars 6 forks source link

2023, Information, Semi-Supervised Model for Aspect Sentiment Detection #91

Open Sepideh-Ahmadian opened 1 month ago

Sepideh-Ahmadian commented 1 month ago

Paper Semi-Supervised Model for Aspect Sentiment Detection

Introduction This paper presents a semi-supervised model for Aspect-Based Sentiment Analysis (ABSA), aiming to detect both explicit and implicit sentiments from online reviews in SemEval datasets and Amazon.

Main Problem The main problem is the inability of traditional supervised models to effectively detect sentiments, especially implicit sentiments, in different domains due to the lack of sufficient labeled data. The authors propose a semi-supervised approach to overcome these challenges by combining unsupervised pre-training and minimal labeled data.

Illustrative Example Sentence: "The vacuum cleaner is very quiet." Aspect: "device noise" Sentiment: positive (depending on the context this feature would be negative for a speakerphone).

Input A sentence and an aspect category.

Output The detected sentiment (positive, negative, or neutral)

Motivation The authors were motivated by the challenge of detecting implicit sentiments in multi-domain reviews with limited labeled data. Traditional supervised models require large datasets, and the authors aimed to develop a more efficient method that can work with small amounts of labeled data.

Related works and their gaps The paper addresses the gap where previous models, especially, supervised ones, require large amounts of labeled data to function effectively. It also fills the need for models capable of implicit sentiment detection and domain adaptation.

Contribution of this paper The main contributions include: Proposing a semi-supervised aspect sentiment detection model based on the aspect-embedded attentional encoder-decoder (AE-AED) architecture. Demonstrating how the model can detect implicit sentiments with limited labeled data across multiple domains and outperforms the existing baselines.

Proposed methods Not included

Experiments Datasets: Semeval datasets (14, 15) and Amazon. Two phases of 1) unsupervised (on Semeval datasets and Amazon) and 2) semi-supervised models

Implementation Not mentioned

Gaps this work The reliance on Word2Vec embeddings and pre-trained models may limit the performance of this research. These embeddings may struggle to capture new trends within a domain or different ways of expressing sentiment across diverse domains. Additionally, the dataset used is restricted to a limited number of domains, which could impact the generalizability of the model. It is unclear how effectively this model would perform on low-resource datasets, where data scarcity poses a significant challenge.