Aspect Extraction from Bangla Reviews Through Stacked Auto-Encoders

Abstract

Interactions between online users are growing more and more in recent years, due to the latest developments of the web. People share online comments, opinions, and reviews about many topics. Aspect extraction is the automatic process of understanding the topic (the aspect) of such comments, which has obtained huge interest from commercial and academic points of view. For instance, reviews available in webshops (like eBay, Amazon, Aliexpress, etc.) can help the customers in purchasing products and automatic analysis of reviews would be useful, as sometimes it is almost impossible to read all the available ones. In recent years, aspect extraction in the Bangla language has been regarded more and more as a task of growing importance. In the previous literature, a few methods have been introduced to classify Bangla texts according to the aspect they were focused on. This kind of research is limited mainly due to the lack of publicly available datasets for aspect extraction in the Bangla language. We take into account the only two publicly available datasets, recently published, collected for the task of aspect extraction in the Bangla language. Then, we introduce several classification methods based on stacked auto-encoders, as far as we know never exploited in the task of aspect extraction in Bangla, and we achieve better aspect classification performance with respect to the state-of-the-art: the experiments show an average improvement of 0.17 , 0.31 and 0.30 (across the two datasets), respectively in precision, recall and F1-score, reported in the state-of-the-art works that tackled the problem.

Deep-ABSA: A Multichannel Deep Learning Framework For Aspect-Based Bangla Sentiment Analysis

“…They claimed the autoencoder had never been used in Bangla for aspect extraction. So these are new things for experiments in Bangla [7]. In [8] the authors found the best accuracy for Bangla Cricket and Restaurant dataset.…”

“…They got 51% and 64% f1-score for the cricket dataset and restaurant dataset respectively. Bodini and Matteo [7] implement three autoencoders for a Bangla Dataset. These are standard AEs, contractive AEs (CAEs), and sparse AEs (SAEs).…”

Fine-Tuning BERT for Aspect Extraction in Multi-domain ABSA

“…Moreover, the same year [6] witnessed a study on aspect extraction from financial microblogs, comparing supervised and unsupervised methods for explicit and implicit aspect identification, showing promising results. Additionally, 2019 [7] saw a study introducing classification methods based on stacked auto-encoders for aspect extraction in Bangla reviews, outperforming the state-of-the-art methods.…”

Aspect-based Sentiment Analysis for Bengali Text using Bidirectional Encoder Representations from Transformers (BERT)

“…For the 'Restaurant' dataset, KNN, SVM, and RF classification achieved 42%, 38%, and 38% F1-Score, respectively, while CNN outperformed them with 64% F1-Score. In [16], Boidini implemented three stacked Auto-encoders (AEs) models based on the datasets from [15] to categorize aspects in the Bengali text. The stacking network's layers were individually trained to comprehend the encoded data of the layer that came before them.…”