Huynh Thai Hoc

The increasing volume of online product reviews presents both opportunities and challenges for e-commerce platforms seeking to leverage customer sentiment for strategic decision-making. This study investigates the optimization of Large Language Models (LLMs) for product review sentiment analysis through systematic fine-tuning and hyperparameter optimization. We evaluated three LLMs (gpt-3.5-turbo-0125, gpt-4o-mini-2024-07-18, and distilbert-base-uncased) using 12,000 product reviews for training and 3,000 for validation, derived from an original corpus of 205,052 reviews, across nine hyperparameter scenarios. Validation results demonstrate that smaller batch sizes improved proprietary model performance, while DistilBERT achieved superior accuracy through moderate-low learning rates but exhibited extreme hyperparameter sensitivity. However, generalization testing on unseen data revealed distinct performance patterns: GPT models demonstrated exceptional accuracy with substantial improvements over their validation results, while DistilBERT maintained strong performance but showed reduced improvement compared to its validation peak. These findings establish fundamental differences in optimization characteristics between proprietary and open-source models. GPT models exhibited robust generalization capabilities with consistent performance improvements across diverse datasets, while DistilBERT required precise hyperparameter tuning to achieve optimal results. The results demonstrate that generalization performance provides critical insights for deployment decisions, with GPT models offering superior robustness for production environments requiring consistent performance across varied review contexts, while DistilBERT remains viable for specialized applications with controlled data distributions.

Due to the ongoing global debate regarding the relationship between fintech and banks, including developing countries, this study aims to investigate this relationship in the case of Vietnam, an emerging nation. The study analyzes the relationship between fintech search and bank stock returns, which are measures of fintech and banks, respectively. The time series data for fintech and bank stock returns were obtained from Google Trends and Vietstock, respectively. Exploratory factor analysis was utilized to derive the fintech variables, while the bank stock return variable was calculated using a basket of eight listed banks from 2017w46 to 2021w46. The results were estimated using the vector autoregression and Granger causality method and validated with the copula method. A key finding of this study is the presence of a simultaneous negative change and bidirectional causality between bank stock returns and fintech lending. Furthermore, several other interesting findings were discovered: (1) the causal relationship from fintech to bank stock returns is weaker compared with the opposite direction; (2) unidirectional causality exists between different types of fintech, such as influence from FinFintech to FinLending, from FinPayment to FinLending and FinWallet, from FinMoney to FinFintech, from FinWallet to FinLending, and from FinProduct to FinFintech; and (3) there is an equal occurrence of simultaneous increase or decrease between bank stock returns and certain types of fintech, specifically between BankReturn and FinPayment, BankReturn and FinLending, as well as BankReturn and FinWallet. These findings shed light on the complex relationship between fintech and banks, offering insights that contribute to our understanding of this dynamic interplay in the context of Vietnam’s emerging fintech landscape.

Objective: The advancement of Internet technology brought up the tourism industry towards new development and opportunities. With application of the Internet technology tourism industry comprises a vast range of virtual communities such as Trip Advisor, Agoda, Travelocity and so on. Existing research concentrated on evaluating the factors influencing virtual communities’ behaviour with limited evaluation of tourist perception. This paper focused on examining the tourists' perception of a virtual tour through the sentimental analysis model based on eWOM for sustainable development. Method: The developed model comprises the group average Bayesian network with the computation of the polarity of the tourist perception. A Bayesian network is a data-driven method involved in estimating dependence among the variable with probabilistic computation. Results: The analysis is based on data collected from sample population in Vietnam with consideration of the 11 variables. Participation intensity, social identity, functional value, emotional value, sharing, interaction, and user satisfaction are among eleven primary variables that have been chosen. Conclusions: The analysis of the results expressed that the user satisfaction level is based on the user's experience and functional value. Additionally, the analysis stated that social value comprises the intermediary role in virtual tourism. This research adds to research methodologies of user engagement methods as well as serves as a reference for theoretical research and management practise in the virtual tourist community.

Over the last few years, more and more people have been using YouTube videos to experience virtual reality travel. Many individuals utilize comments to voice their ideas or criticize a subject on YouTube. The number of replies to 360-degree and unidirectional videos is enormous and might differ between the two kinds of videos. This presents the problem of efficiently evaluating user opinions with respect to which type of video will be more appealing to viewers, positive comments, or interest. This paper aims to study SentiStrength-SE and SenticNet7 techniques for sentiment analysis. The findings demonstrate that the sentiment analysis obtained from SenticNet7 outperforms that from SentiStrength-SE. It is revealed through the sentiment analysis that sentiment disparity among the viewers of 360-degree and unidirectional videos is low and insignificant. Furthermore, the study shows that unidirectional videos garnered the most traffic during COVID-19 induced global travel bans. The study elaborates on the capacity of unidirectional videos on travel and the implications for industry and academia. The second aim of this paper also employs a Convolutional Neural Network and Random Forest for sentiment analysis of YouTube viewers’ comments, where the sentiment analysis output by SenticNet7 is used as actual values. Cross-validation with 10-folds is employed in the proposed models. The findings demonstrate that the max-voting technique outperforms compared with an individual fold.

This study compares the performance of Pytorch-based Deep Learning, Multiple Perceptron Neural Networks with Multiple Linear Regression in terms of software effort estimations based on function point analysis. This study investigates Adjusted Function Points, Function Point Categories, Industry Sector, and Relative Size. The ISBSG dataset (version 2020/R1) is used as the historical dataset. The effort estimation performance is compared among multiple models by evaluating a prediction level of 0.30 and standardized accuracy. According to the findings, the Multiple Perceptron Neural Network based on Adjusted Function Points combined with Industry Sector predictors yielded 53% and 61% in terms of standardized accuracy and a prediction level of 0.30, respectively. The findings of Pytorch-based Deep Learning are similar to Multiple Perceptron Neural Networks, with even better results than that, with standardized accuracy and a prediction level of 0.30, 72% and 72%, respectively. The results reveal that both the Pytorch-based Deep Learning and Multiple Perceptron model outperformed Multiple Linear Regression and baseline models using the experimental dataset. Furthermore, in the studied dataset, Adjusted Function Points may not contribute to higher accuracy than Function Point Categories.