Technological advancements in recent years have profoundly changed people's lives in the physical world. With technology becoming more advanced and accessible, fundamental developments such as analysis, evaluation, and commentary have been integrated. Given that comments encompass a wide array of issues, including sentiment, emotion, and interaction, analyzing comments has become essential in today's digital age. Sentiment analysis (SA) is intricately tied to the rapid technological advancements in the real world. It stands out as one of the most dynamic research domains within Natural Language Processing (NLP), owing to its signicant potential applications in both business and society [1,2,3]. This calls for the development of new evaluation models and presents considerable challenges. While these models have seen widespread use in recent years, there remains room for improvement in this trend. Unfortunately, these models are trained on various architectures, pre-trained data, and preprocessing steps, leading to inconsistencies and errors in systems. All studies compare and evaluate performance using both the monolingual PhoBERT and the ViT5 model [1]. Furthermore, the study [1] stated that it is the rst to investigate the performance of ne-tuning Transformerbased models on ve datasets with dierent domains and scales for Vietnam's SA task. Considering cultural factors is crucial during application, as misinformation can directly impact the training of the model [4,5,6,7,8].
To eectively analyze Vietnamese comments on social media platforms [4,8], we propose a comprehensive system that integrates natural language processing and machine learning techniques. This system aims to classify the sentiment of Vietnamese comments as either positive or negative, oering valuable insights into user opinions and emotions, as depicted in Figure 1.
Therefore, to analyze the sentiment of comments on Vietnamese social networks, we propose the following listed implementation procedures:
• Using Ontology for sentiment analysis in Vietnamese social media.
• Collecting and preprocessing comment data from popular Vietnamese social media platforms.
• Using Ontology to assign sentiment labels (positive, negative, neutral) to comments.
• Analyzing sentiment patterns and trends in comments.
• Comparing the performance of the Ontology-based method to traditional sentiment analysis. This study's ndings contribute to expanding social data analysis techniques and provide practical insights into user behavior on Vietnamese social media platforms. These insights can be directly applied to enhance social media strategies and user engagement.
The rest of this paper is organized as follows: Section 2 discusses related work, Section 3 describes the proposed model, and Section 4 evaluates performance. Finally, Section 5 discusses our ndings and pending issues.
This research relates to our investigation on chatbots, social media comment analysis, and ontology. The study focuses on the impact of managerial relationships, social media, and business attitudes on the nancial performance of small and medium-sized enterprises (SMEs) in Vietnam [9]. The report outlines some theoretical advances and provides useful recommendations for practitioners in Vietnamese SMEs who want to increase productivity and eciency.
Motivated by GLUE's success, they provide the Social Media Text Classication Evaluation (SMTCE) benchmark, which consists of a set of models and datasets covering a wide range of SMTC tasks. They apply and evaluate a range of BERT-based [10] multilingual (mBERT, XLM-R, and DistilmBERT) and monolingual (PhoBERT, viBERT, vELECTRA, and viBERT4news) models for tasks in the SMTCE benchmark using the suggested benchmark. Monolingual models produce state-of-the-art results on all The study of [11] examines customer perceptions of Vietnamese hotel services in general and aspects of hotel services by combining natural language rules and inferential statistics; they analyze to understand customer experiences with the hotel industry recovering from the pandemic and thereby provide what customers want to enhance the customer experience.
The purpose of [12] is to investigate the factors leading to social commerce adoption in Vietnam. The participants of this study were 447 social networking website users in Vietnam. The results identify important antecedents that inuence Vietnamese consumers' propensity to participate in social commerce. These ndings have implications for research and practical applications in understanding social commerce adoption in emerging economies.
One of the negative side eects of more people using social networking sites is an increase in rude and nasty language aimed at other members. Because of this, reviewing tagged comments that have been ltered by categorization systems may become challenging for human moderators. In an eort to solve this problem, [13] presents the ViHOS (Vietnamese Hate and Oensive Spans) dataset. They also oer comprehensive annotation rules and denitions of hateful and oensive spans in Vietnamese comments.
Around the world, a lot of individuals utilize social media for education and amusement. Furthermore, as several foreign language specialists and academics across the world have demonstrated, using this kind of technology helps students learn foreign languages, including English, Chinese, French, Japanese, and so forth. The purpose of the study [14] is to see how students feel about expanding their vocabulary in English through social media use.
Globally, a large number of individuals utilize social media for learning and amusement. Additionally, the usage of this kind of technology aids with children's foreign language acquisition, as noted by several academics and specialists in foreign languages from across the globe. The study's [15] goal was to nd out how students felt about utilizing social media to expand their vocabulary in English. Van Lang University (VLU), in Vietnam, used a blend of quantitative and qualitative methodologies. Fiffieen of the 154 students who participated in semi-structured interviews had surveys completed on them by dierent majors. Their research showed a connection between negative rejection-related stresses and negative FOMO ratings, as well as a relationship between FOMO scores and worse overall quality of life and increased depression symptoms.
They present a technique in this article for gauging social tension in various Russian regions by examining user posts on the social network Vkontakte (VK) [16,17]. They created a tool to gather postings from VK members that expressed unfavorable opinions regarding prevalent societal problems like ination, corruption, and unemployment. Using this tool, they were able to compile data on the quantity of these postings made over specic timeframes and examine user proles in general.
In this study, we aim to explore Vietnamese comments on social media. We identify and analyze elements of these platforms to develop an eective self-learning support system. The process involves two main stages: data collection and processing, followed by the analysis and evaluation of the comments in Figure 2.
First, we set up the working environment and necessary tools. Next, Vietnamese document datasets, including articles, comments, and other documents related to students' learning on social networks, will be transferred to our database system. This process requires meticulous preparation to collect the data thoroughly and accurately using programming techniques.
Data collection and processing: we gather data from various sources, including popular social media platforms. The collected data is then cleaned and normalized to meet system requirements. This processing involves removing noise, reformatting the data, and organizing it into appropriate categories. We use ontologies to train the model affier transferring the data to the system. Ontologies help identify concepts and the relationships between these concepts in the learning and education domain. Incorporating ontologies during model training allows the system to understand and analyze the data more deeply and accurately.
We begin the data analysis once the data is prepared and the system model is trained. Our system employs natural language processing (NLP) techniques and machine learning algorithms to analyze the sentiment of the comments in the dataset. This includes semantic analysis, keyword identication, and sentiment ratings for each comment.
Affier the system completes the analysis process, the ndings will be displayed as sentiment clas-sications and used for further steps, including understanding the context of the comments. Finally, the program will store the comments and analysis results in the system. This ensures that all data is securely stored and provides a foundation for future updates and analysis.
The sentiment analysis module is at the core of the proposed system [18], which integrates natural language processing techniques and machine learning algorithms. It plays a crucial role in enabling the system to understand and analyze linguistic data, thus providing users with accurate and valuable assessments.
Many studies have demonstrated that self-study on social networking platforms is highly practical. We collected comments from YouTube to assess users' attitudes and emotions regarding self-study using social networks and analyzed them using sentiment analysis (SA). Our automated comment data collection system is illustrated in detailed in Algorithm 1 as follows:
• System Initialization: This stage involves starting and preparing the system to execute the subsequent tasks. • File Transmission: The system receives a specic le containing a list of video URLs.
• Link Retrieval: The system will retrieve each video URL from the provided le in a predetermined order to ensure a sequential process. end if 14: end for
• Video Title Extraction: The system will extract the title of each video from the provided links.
• Video Content Verication: The system will assess the content of each video to determine its reliability. • Comment Section Identication: If the video contains content, the system will locate and retrieve the section containing video comments. • Comment Extraction: The system will extract all comments from the designated comment section. • Comment Count Storage: In the nal and crucial step, upon successful extraction of comments, the system will accurately record the number of comments in the Vietnam Document dataset. • Program Termination: The program will conclude affier completing the aforementioned steps.
In summary, this process involves collecting and analyzing information from Vietnamese-related videos and their comments using data retrieval, natural language processing, and data storage techniques.
Affier the data undergoes raw processing and the removal of any irregular characters, it is input into the word analysis system. Below is an outline of how our proposed system operates:
• Step 1: When words are input into the system, it rst identies the sentence structure and then segments the sentence in Figure 3. These segments are subsequently examined and ltered to eliminate special characters. This ltering step is crucial in text processing to ensure that data is cleansed before analysis, retaining only alphabetic, numeric, and whitespace characters in the input string. Affier cleaning the string, the process proceeds by breaking down the sanitized sentence into individual words, a process known as tokenization. Tokenization converts a text string into a list of words, with each word representing a distinct unit of meaning in the text. This step prepares the data for more advanced semantic analysis and text-processing tasks. • Step 2: The system evaluates the sentiment of sentences by analyzing their context to determine whether they convey a positive or negative meaning. It employs word scanning techniques to identify and lter out sentences with negative sentiment. This process enhances the accuracy of comments and predictions. Next, the system extracts key words from sentences and analyzes them by cross-referencing with its ontology. The ontology is structured as a binary tree, which helps classify text into two primary branches: positive and negative. Each branch represents a dierent type of intent. The leffi branch corresponds to negative features, while the right branch pertains to positive elements. Finaly, this ontology provides a detailed and structured description of terms and their relationships. It aids the system in understanding the semantics of the text more thoroughly and applying logical inference rules eectively. • Step 3: Once the ontology is constructed and trained, the system employs a binary search algorithm to query words within the user's sentence against the ontology. This search method eciently retrieves terms from binary tree data structures. Based on the results of this ontology search, the system can determine whether the user's sentence carries a positive, negative, or neutral sentiment. Corpus comparison and processing algorithms are utilized to conduct this assessment, evaluating the sentence's structure and content in relation to key terms dened in the ontology in Figure 4.
In this paper, we assess the eectiveness of the proposed method by analyzing 3.739 Vietnamese comments. Additionally, we compare our method with three existing methods that analyze English comments. These reference methods are a chatbot system designed to analyze English opinions (referred to as BCSAO), a chatbot for changing lifestyles in education (referred to as ACCLE), and an interactive transport inquiry AI chatbot (referred to as ITEAI). Below is a summary of these reference methods:
• ITEAI [19]: Similar to the ACCLE approach, this method develops a chatbot system that queries users about their current location and nal destination. The design analyzes the user's query and fetches relevant data from the database. It provides comprehensive information, ensuring individuals can safely reach their desired destination.
• ACCLE [20]: The author suggests a Chatbot system to enhance teacher and student collaboration.
In this system, students submit text-based questions to the Chatbot, which uses natural language processing and deep learning technologies to process the data and respond to the students. However, this system is limited to use within schools and does not analyze the respondents' emotions.
• BCSAO [18]: This method resembles the one we propose. However, while their data processing mainly relies on programming techniques, our approach goes further in applying them. Creating an ontology categorizes sentences and performs a more detailed analysis for each specic topic, followed by automated separation using individual models and optimizing comment sentences.
In conclusion, to assess the proposed method's performance against the evaluation methods, we utilize the following formulas according to [21] to calculate Accuracy and F1-score.
Our study aims to assess the accuracy of refer models. Formula 1 determines the ratio of correct predictions to the total number of predictions made. The formula for calculating accuracy is presented below:
Additionally, we recognize that accuracy is only sometimes the best metric for all scenarios. It can be particularly limited in cases of data imbalance, where one class signicantly outweighs the other. Therefore, we also compute other metrics, such as precision (formula 2), recall (formula 3), and F1-score (formula 4). The formulas are as follows:
Where:
• TP: The model predicts 1, and the actual value is also 1.
• TN: The model predicts 0, and the actual value is also 0.
• FN: The model predicts 0, but the actual value is 1.
• FP: The model predicts 1, but the actual value is 0.
Our system outperforms the three methods listed in Table 1. The proposed method consistently achieves a higher percentage than the other methods. The ACCLE method shows a low average result of 0.5027, while the ITEAI and BCSAO methods achieve 0.7622 and 0.7853, respectively. The proposed method reaches the lowest level at 0.8657.
In this study, we developed a social media comment evaluation model using ontology techniques to classify Vietnamese comments as positive or negative. We collected Vietnamese comments from video platforms using Python programming, enhancing our ability to analyze user sentiment. This approach is particularly benecial for businesses in Vietnam that deal with a high volume of customer comments. Our method signicantly outperforms existing ones, achieving an accuracy improvement of up to 0.0804 compared to others.
On the other hand, we also evaluate educational teaching videos by assessing their quality and eectiveness in conveying content through video lectures. Although comments have been collected from these videos, future research will aim to gather a broader and more diverse range of comments to enhance the study's scope.
For future work, it is essential to explore how the proposed framework can be applied across dierent languages, cultures, and age groups. Additionally, understanding how the system identies and interprets irony or sarcasm in comments will be a key area of focus.
This research is funded by the East Asia University of Technology (EAUT).