A Transformer-Based Sentimental Analysis System for Cultural Heritage Sites

Abstract

User-generated positive and negative sentiment content of reviews and comparisons between heritage sites are fast-growing, resulting in a massive amount of data that plays a vital part in consumer decision-making. However, consumers may find it challenging to pick which heritage sites to visit or which activities to participate in without the assistance of tourism focused websites. The sentiment analysis technique is used to assess offered text, to determine its explicit, embedded, or underpinning emotions and to determine the ground truth of the collected reviews. Although several models are employed to improve the outcome, certain factors based on the text size and quality pave the way for better results. This chapter explains how transformer models fare when compared with other neural net models and puts forward the importance and implications of digitalization with regards to cultural heritage sites. The sentiment analysis approach employed in this study seeks to measure the strength (both positive and negative) of a sentiment conveyed, in this example, in reviews submitted on four popular tourism websites. The sentiment analysis technique is used to evaluate the provided text in 260order to identify the emotions in it, in order to acquire the ground truth of the gathered reviews. In sentiment analysis, Bidirectional Encoder Representations from Transformers (BERT), Long Short-Term Memory (LSTM) models, and Bi-directional LSTM are used. To give high accuracy and polarity in contrast to common machine learning classifiers, the reviews were analyzed using BERT, LSTM, Bi-directional LSTM (Long short term memory) models. The experimental results comparing Bidirectional Encoder Representations from Transformers model to LSTM, Bi-LSTM models indicate the clear influence of the self-attention mechanism and positional embeddings by BERT in understanding the text and performing any text-based tasks.