The article presents a comprehensive system for the automated analysis of comments on Instagram, focusing on multilingual content and the unique characteristics of social networks. The system includes a module for automatic parsing of dynamic content, an algorithm for determining the language of comments, and mood analysis modules built on the basis of modern transformer models, in particular XLM-RoBERTa. Particular attention is paid to supporting Ukrainian, Russian, and English, as well as processing texts with informal elements, including slang, abbreviations, emojis, and symbols. An approach to analysing mood dynamics over time by combining models of time series, moving averages, and clustering is proposed. The system is complemented by interactive visualisation of results, which enables researchers and businesses to gain in-depth insights from large amounts of data. The analysis of existing solutions demonstrates the advantages of the proposed approach, particularly its high accuracy for local languages and its adaptation to social media content. The developed tool is crucial for monitoring public sentiment, gathering business intelligence, and enhancing information security, particularly in the Ukrainian context.
In today's world, social media has become a powerful tool for communication, marketing, opinion
analysis, and research into consumer sentiment [
0000-0001-6417-3689 (V. Vysotska); 0009-0007-4948-4586 (A. Hyriak); 0000-0002-9448-1751 (L. Chyrun);
0009-00026669-0369 (R. Fedchuk); 0000-0002-4909-6723 (O. Lavrut); 0000-0003-4858-4511 (D. Uhryn); 0000-0002-9690-8042
(L. Kolyasa); 0000-0002-8411-3584 (S. Smailova); 0000-0002-1101-7479 (M. Brygadyr)
the distinct characteristics of language groups, such as Ukrainian and Russian [
In the context of Ukraine, there is a growing need to analyse local content, including Ukrainian
and Russian. Existing solutions lack high accuracy due to the absence of specialised models for
these languages [
The creation of specialised models for the Ukrainian language will contribute to the development of local technologies and increase the competitiveness of Ukrainian companies in the global market. In times of war and post-conflict reconstruction, sentiment analysis on social media can be helpful in monitoring public opinion, detecting disinformation, and evaluating the effectiveness of information campaigns. Businesses will be able to receive more accurate analytics on reactions to their products or services, taking into account local language features and nuances. The proposed solution opens up new opportunities for research in the field of natural language processing, mood analysis and sociology. Thus, the proposed project is not only relevant but also critically important for the development of text analysis technologies in social networks within a multilingual environment, particularly in the context of Ukraine. It allows you to address existing limitations by providing accurate, contextual and multilingual analysis of comments, which is essential for both business and society as a whole.
The purpose of the study is to develop a comprehensive system of automated analysis of comments on Instagram, which provides multilingual text processing, sentiment identification, trend building and in-depth analysis of the interaction of comments with posts, taking into account the specifics of language groups and the context of social networks, to increase the efficiency of decision-making by businesses, researchers and organisations.
To achieve this goal, it is necessary to solve the following tasks: 1. Development of a mechanism for automatic parsing of comments from Instagram: ensure efficient retrieval of data from posts, including comments, emojis, and symbols; consider the technical limitations of the Instagram API and optimise for high scraping performance. 2. Definition of comment language: implement an algorithm to automatically detect the language of comments, taking into account multilingualism, abbreviations, slang and symbols, and add labelling of comments (e.g. "ua", "ru", "en", "symbol_only") for further analysis. 3. Development of a mechanism for analysing the sentiments of comments: integrate specialised pre-trained transformer models from the Hugging Face platform to analyse the sentiments of comments in different languages, and to ensure high accuracy of analysis, adapted to the specifics of texts from social networks. 4. Building a comprehensive analysis of the results: develop algorithms to assess the solidarity of comments with posts, determine the general mood, and build trends for maintaining posts over time, and perform an analysis of the sentiment of comments based on the language group. 5. Visualisation of data and analysis results: develop tools for visualising results in the form of graphs, charts, and interactive reports, and to provide the possibility of segmented analysis (by language, time, mood, etc.). 6. Testing and optimisation of the system: test the system on real data from Instagram to check its performance and accuracy, and optimise the algorithm to ensure efficient processing of large datasets.
Thus, the implementation of these tasks will create an innovative solution for the multilingual analysis of comments in social networks, which will be useful for businesses, researchers, and organisations, particularly in Ukraine, and will contribute to the development of natural language processing technologies. The object of this study is the automated processing and analysis of text data generated by users in social networks, specifically Instagram comments that contain multilingual content, symbols, emojis, and other text features. The subject of the study is methods and algorithms for automatic parsing, multilingual text processing, sentiment determination, and analysis of comment trends in social networks, particularly specialised transformer models for analysing texts in different languages, as well as approaches to visualising results to provide a deep understanding of user interaction with content. Within the framework of the study, the following new scientific provisions and solutions were obtained, which differ from the previously known ones and have the subsequent degree of novelty: 1. For the first time, an approach to analysing multilingual comments on Instagram has been developed, taking into account the specifics of the Ukrainian, Russian, and English languages, as well as content consisting only of characters (symbol_only). An algorithm for automatically determining the language of comments is proposed, taking into account the features of texts in social networks, such as slang, abbreviations, emojis and symbols. 2. For the first time, specialised pre-trained transformer models have been integrated to analyse the sentiments of comments in various languages. Adaptation of models from the Hugging Face platform for text analysis in the multilingual environment of Instagram, in particular for the Ukrainian language, which was previously underrepresented in existing solutions. 3. The process of analysing the solidarity of comments with posts has been improved. A new approach has been developed to assess the level of support or criticism of comments regarding the content of the initial post, enabling a more accurate evaluation of user interaction with the content. 4. The method of complex analysis of moods in the dynamics of time was further developed.
An algorithm for identifying trends in comment sentiments based on the time of post publication is proposed, enabling the detection of changes in audience reactions to content over time. 5. Improved the approach to analysing comment sentiment based on language group. For the first time, a comparative analysis of the moods of comments in different languages (Ukrainian, Russian, English, and symbol_only) was conducted, taking into account linguistic and cultural features. 6. A new approach to visualising the results of analysing multilingual comments has been developed. Interactive data presentation tools have been developed that enable businesses, researchers, and organisations to quickly gain insights from large datasets, taking into account both linguistic and emotional characteristics.
Thus, the results of the study have a high degree of novelty, as the proposed solutions provide a more accurate, contextual and multilingual analysis of comments on social networks, which has not been implemented in this form before, especially for the Ukrainian context.
The developed project has significant practical value, since its implementation opens up new
opportunities for data analysis in social networks, in particular on Instagram, and can be applied in
various areas [
The developed project is of particular importance to Ukraine, as it contributes to the development of local IT solutions, supports the Ukrainian language in the technological environment, helps analyse public sentiment in challenging socio-political conditions, and provides tools to combat disinformation. Thus, the practical value of the project lies in its versatility, adaptability to a multilingual environment, ability to generate deep insights from social media data, and support for the development of technologies focused on local needs.
In today's world, social media, particularly Instagram, has become a powerful source of data for
analysing user sentiment, public opinion, and behaviour [
Social media is a primary source of textual data for businesses, researchers, and organisations
seeking to understand user sentiments, trends, and behaviours [
Let's review the most well-known systems, such as Google Cloud Natural Language, Microsoft Azure Text Analytics, Brandwatch, and Sprout Social, with an emphasis on their functionality, multilingual capabilities, sentiment analysis accuracy, integration, and visualisation features. 1. Google Cloud Natural Language API is a service from Google that allows you to analyse text data using artificial intelligence. It supports features such as language detection, sentiment analysis, entity recognition, and parsing. Determines the polarity (positive, negative, neutral) and mood intensity of the text. Automatically identifies the language of the text. Highlights key objects, such as names, places, and organisations. It supports more than 20 languages, including English and Russian, but does not provide full support for the Ukrainian language. The accuracy of sentiment analysis is high for English-language content; however, it decreases for other languages, particularly for texts from social networks that contain slang, emojis, and abbreviations. The API integrates seamlessly into applications thanks to REST APIs and SDKs for different programming languages. It does not provide built-in visualisation tools, but the results can be integrated with other graphing tools. Limited support for the Ukrainian language. High cost for large amounts of data. Google Cloud Natural Language is a powerful tool for text analysis; however, its limitations in local languages and the specific characteristics of texts from social networks reduce its effectiveness for multilingual analysis. 2. Microsoft Azure Text Analytics is a part of Azure Cognitive Services that provides an API for text analysis. It supports sentiment detection, entity recognition, text classification, and language detection. Determines the emotional tone of the text. Highlights entities and categorises them. Automatically detects the language of the text. Supports more than 120 languages, including Ukrainian, Russian, and English. The accuracy of sentiment analysis is high for major languages; however, it decreases for texts from social networks, particularly those containing symbols and emojis. Easy integration via REST API and SDK for different programming languages. It does not have built-in visualisation tools, but the results can be used in other tools for visual representation. High cost for large amounts of data. Limited adaptation to the specifics of social media texts. Microsoft Azure Text Analytics has broader language support than Google Cloud Natural Language, but its accuracy for texts from social networks remains limited. 3. Brandwatch is a social media monitoring platform that allows you to collect data, analyse sentiment, identify trends, and generate reports. Collects data from various platforms, including Instagram, Twitter, and Facebook. Determining the mood of texts in social networks. Automatic identification of key trends in texts. Supports text analysis in multiple languages, with varying accuracy levels by language. The accuracy of sentiment analysis is high for English-language content, but may be reduced for less popular languages such as Ukrainian. The platform offers APIs for integration with other systems. It has built-in tools for creating graphs, charts, and reports. High cost. Closed access to sentiment analysis algorithms. Brandwatch is an effective tool for monitoring social networks, but its limitations in supporting local languages and high cost make it less accessible for multilingual analysis. 4. Sprout Social is a social media management platform that also includes text analysis features. Data collection from different platforms. Determination of the moods of texts. Automatically generate reports on user interactions. Supports text analysis in multiple languages, but is limited. The accuracy of sentiment analysis is high for English-language content, but limited for other languages. Integration with other platforms via API. It has built-in visualisation tools for generating reports. High cost. Limited support for local languages. Sprout Social is useful for social media management, but its text analysis functionality is basic.
Existing social media text analysis systems offer a wide range of features; however, they have limitations, including support for local languages (such as Ukrainian), the accuracy of text analysis from social networks, and adaptation to the specifics of multilingual content. The system under development aims to solve these problems by integrating specialised models for sentiment analysis, multilingual analysis and adaptation to texts from social networks.
Collecting data from social networks, particularly Instagram, is a crucial step in further analysing
comments, sentiments, and trends [
1. Selenium is a browser automation tool that allows you to interact with web pages the way a real user does. It is ideal for dynamic parsing because it can interact with page elements generated by JavaScript. Selenium allows you to load comments that appear after clicking the "Load more comments" button. It is possible to automate any actions on the page, including scrolling, clicking buttons, and entering data. Using a real-time browser reduces the risk of detecting automated actions – the ability to work with Chrome, Firefox, Edge and other browsers. Selenium consumes a lot of RAM and computing resources due to page rendering in the browser. Less productive compared to tools that work without a browser. Instagram may detect automated activities, so additional measures, such as using proxies, are required. Selenium is a powerful tool for dynamic scraping, particularly for handling complex pages with dynamically loaded elements, such as those found on Instagram comments. 2. Scrapy is a Python web scraping framework that allows you to quickly scrape data from static web pages. Scrapy is much faster than Selenium because it doesn't use a browser. Suitable for collecting large amounts of data from simple static pages. Easily integrates with other libraries for data processing. Scrapy cannot work with JavaScript-generated elements, such as dynamically loaded comments. To work with Instagram, you need additional tools or workarounds. Scrapy is effective for static pages, but its limitations in working with dynamic content make it less suitable for scraping comments from social media platforms like Instagram. 3. BeautifulSoup is a Python library for analysing HTML and XML documents. It allows you to extract data from web pages using a simple API. Easy to set up for basic parsing. It can be used to work with HTML structures of any complexity. It can be combined with other tools such as Scrapy or Selenium. BeautifulSoup cannot work with dynamically uploaded content. Works slower than Scrapy, especially for large amounts of data. BeautifulSoup is suitable for basic HTML page scraping, but it cannot work with dynamic content, such as Instagram comments. 4. Instaloader is a Python tool for downloading data from Instagram, including posts, profiles, and comments. Easily configurable for basic data collection from public profiles. Supports logging in to the account to access private data. Can extract comments on posts. Using Instaloader may violate Instagram's terms of service, resulting in account suspension or ban. The tool is subject to Instagram's restrictions, which may affect the stability of the work. Cannot interact with elements that JavaScript generates. Instaloader is helpful for basic data collection from Instagram, but its limitations in working with dynamic content make it less effective for scraping comments. 5. Instagram Graph API is an official tool for accessing Instagram data that provides features to collect information about posts, comments, and other metadata. The API complies with Instagram's privacy policy. Allows access to structured data with high accuracy. You can obtain additional information, such as the publication date, the author of the comment, and so on. The API only works with business accounts and creator accounts. There are strict limits on the number of requests, which makes it challenging to work with large amounts of data. The data is only available for public profiles. The Instagram Graph API is a reliable and legal tool, but its limitations make it less flexible for comprehensive comment scraping.
Among all the tools reviewed, Selenium is the best choice for scraping comments from Instagram, as it enables working with dynamically loaded elements, such as the "Load more comments" button. Although Selenium has high resource requirements and is slower, its flexibility and ability to mimic user actions make it ideal for tasks that involve dynamic content. Other tools, such as Scrapy, BeautifulSoup, and Instaloader, are less suitable for this task due to limitations in working with JavaScript-generated elements. At the same time, the Instagram Graph API has strict access restrictions.
Low Low 2.3. Comparison of methods for determining the language of texts Determining the language of a text is a key step for multilingual data analysis on social media. In the context of Instagram, comments can be written in different languages, contain mixed language elements, symbols, emojis, and abbreviations, making it difficult to automatically detect the language [18–24]. We will review the most common algorithms and libraries for determining the language of texts, such as LangDetect, FastText, and DeepLang, and their adaptation to the specific needs of texts from social networks. Particular attention is paid to accuracy, performance and the possibilities of integrating these methods into the system being developed.
1. LangDetect is a popular text language detection library based on an algorithm that uses statistical models and Bayesian classification. It supports more than 50 languages. Supports most common languages, including Ukrainian, Russian, and English. Easily integrates into Python projects – high speed of text processing, which allows you to work with large amounts of data. LangDetect has difficulty detecting the language of short texts, such as social media comments. Texts that contain multiple languages or characters may be classified incorrectly. LangDetect is effective for detecting the language of long texts, but its accuracy for short and mixed texts is limited. 2. FastText is a text classification and vectorisation library developed by Facebook AI. It also features a model for detecting the language of the text, which supports over 170 languages. FastText demonstrates high accuracy even for short texts. Supports a greater number of languages than LangDetect. FastText models are designed to be fast, making them wellsuited for processing large amounts of data. The model can be additionally trained on specific data from social networks. To use FastText, you must have a sufficient amount of RAM. The integration can be more complicated compared to LangDetect. FastText is one of the best options for determining the language of short texts such as Instagram comments, due to its accuracy and adaptability. 3. DeepLang is a text-language detection library that uses neural networks for classification.
It supports more than 100 languages and is one of the most advanced technologies in this field. The use of neural networks ensures high accuracy even for short and mixed texts. Can work with texts that contain symbols, emojis, and abbreviations. You can additionally train the model on specific data. DeepLang requires significant computing resources to operate. Integration into the system requires additional effort due to the complexity of working with neural networks. DeepLang is the best option for accurately determining the language of texts from social networks; however, it requires significant resources to utilise.
Low
For a system that handles comments from Instagram, LangDetect is the best choice. However, FastText offers high accuracy for short texts, broad language support, and the ability to adapt to the specifics of texts from social networks. DeepLang is also a promising option for providing the highest accuracy, but its resource requirements can be a limiting factor. LangDetect can be used for basic tasks; it does not require significant resource expenditures on the part of the system, so it was chosen for the project.
Analysing the sentiments of texts is a crucial task for understanding the emotional reaction of the
audience to specific content on social networks [
1. Traditional dictionary approaches to sentiment analysis are based on the use of pre-created dictionaries that contain words with the meanings of their polarity (positive, negative, neutral). The most common tool in this category is VADER (Valence Aware Dictionary and sEntiment Reasoner). Advantages of VADER: ease of use, high accuracy for Englishlanguage text, particularly short texts such as tweets or comments and taking into account the intensity of the mood through punctuation, capital letters and emojis. Disadvantages of VADER: 1.1. Limited language support, i.e., VADER is primarily focused on the English language. 1.2. Low accuracy for multilingual analysis or texts from social networks that contain slang, abbreviations, and mixed languages. 1.3. Difficulties with context arise when dictionary methods fail to take into account the context, which can lead to errors in determining mood.
VADER is effective for fundamental sentiment analysis of English-language content; however, its limitations in multilingualism and contextuality render it unsuitable for complex multilingual tasks.
2. Transformers are modern models for natural language processing (NLP) that use a selfattention mechanism to take into account the context of each word in the text. They provide high accuracy for sentiment analysis, even for multilingual content. 2.1. BERT (Bidirectional Encoder Representations from Transformers) is a bidirectional model that takes into account the context of words both to the left and right of the current word. Advantages: high accuracy for sentiment analysis due to contextual consideration and the ability to further train on specific data, which increases accuracy for particular tasks. 2.2. RoBERTa (A Robustly Optimised BERT Pretraining Approach) is an advanced version of BERT that uses better pre-training techniques and larger amounts of data – advantages: higher accuracy compared to BERT, as well as the ability to adapt to specific tasks. The disadvantage is that, like BERT, RoBERTa is based on English-language data in the basic version. RoBERTa is effective for English-language sentiment analysis, but its limitations in multilingualism remain. 2.3. DistilBERT is a simplified version of BERT that provides faster operation and lower resource requirements. Advantages: lower requirements for computing resources and speedier word processing. Disadvantages: reduced accuracy compared to BERT and limited language support. DistilBERT is a compromise between accuracy and speed, but its limitations in multilingualism remain. 2.4. XLM-RoBERTa (Cross-lingual RoBERTa) is a multilingual version of RoBERTa that supports more than 100 languages, including Ukrainian, Russian, and English. Advantages:
Low
Low
Low
High accuracy for multilingual sentiment analysis. Adaptation to texts in different languages, including less common ones. The possibility of additional training on specific data.
The disadvantage is the high requirements for computing resources. XLM-RoBERTa is the best choice for multilingual sentiment analysis because it achieves high accuracy across texts in various languages.
Among the sentiment analysis models considered, XLM-RoBERTa is one of the best choices for multilingual text analysis on social networks due to its accuracy, contextuality, and support for multiple languages. Traditional dictionary approaches, such as VADER, are less effective due to their limited multilingual capabilities and lack of contextual consideration. Transformer-based models, such as BERT, RoBERTa, and DistilBERT, demonstrate high accuracy for English-language content; however, their limitations in multilingualism make them less suitable for the system under development. That is why it was decided to use an individual approach, employing different models for various languages, as well as for symbols and emojis. In the future, with the development of models, there will be a complete transition to specialised models for each language.
Analysing mood trends over time is a crucial task for understanding the dynamics of changes in
users' emotional responses to content on social networks [
1. Time series is a method of analysing data that changes over time, using forecasting and trend detection models. In the context of sentiment analysis, time series enable you to estimate the change in sentiment of comments on posts during specific time periods, offering advantages such as forecasting, trend detection, and flexibility. Time series models such as ARIMA (AutoRegressive Integrated Moving Average) allow you to predict future mood changes. Will enable you to identify long-term sentiment trends. Suitable for analysing data with different time intervals (hours, days, weeks). Disadvantages: data requirements and complexity of setup. Practical analysis requires large amounts of data that have a regular time structure. Time series models require careful optimisation of their parameters. Time series are a powerful method for analysing sentiment trends over time; however, their effectiveness depends on the quality and regularity of the data. 2. Moving averages are a method that uses averaging sentiment values over a specific time period to smooth out short-term fluctuations and identify general trends. Advantages: simplicity, data smoothing and visualisation. Easy to implement and customise. Allows you to reduce the impact of noise and short-term fluctuations. Moving averages integrate seamlessly with charting tools to build trends – disadvantages include the loss of parts and limited predictive ability. Anti-aliasing can hide significant short-term mood changes. Moving averages do not allow you to predict future mood changes. Moving averages are a simple and effective method for identifying general trends, but their limitations in forecasting make them less suitable for complex tasks. 3. Mood clustering is a method of grouping comments according to similar emotional characteristics, followed by an analysis of their dynamics over time. Advantages: pattern detection, relationship analysis, and flexibility. Allows you to identify comment groups with similar sentiments and their changes over time. You can identify relationships between sentiment and other factors (for example, when posts are posted). Various clustering algorithms, such as K-Means, DBSCAN, or hierarchical clustering, can be used. Disadvantages are complexity and resource requirements. Clustering requires careful adjustment of parameters, such as the number of clusters. Analysing large amounts of data can be a resource-intensive process. Sentiment clustering is an effective method for analysing sentiment dynamics in depth; however, its complexity can be a limiting factor.
In the system being developed, a combined approach will be used to analyse mood trends over time:
2. Moving averages are used to smooth out short-term fluctuations and identify general trends. 3. Clustering to group comments by similar emotional characteristics and analyse their relationships over time.
This approach allows you to consider the benefits of each method, providing an accurate and indepth analysis of moods over time.
Analysing mood trends over time is a crucial task for understanding the dynamics of the audience's emotional responses. Among the methods considered, time series provide the best predictive ability, while moving averages allow data to be smoothed to identify general trends, and clustering adds the ability to analyse the relationships between sentiment and other factors in depth. The proposed combined approach in the system under development enables the integration of the advantages of all methods, providing an accurate analysis of mood changes over time.
In today's world, social media, including Instagram, has become a crucial source of data for analysing audience sentiments, trends, and behaviours. User comments contain valuable information that can be used for business intelligence, sociological research, and public opinion monitoring. However, existing solutions for automating text analysis in social networks have significant drawbacks that limit their effectiveness, especially for multilingual content and local languages such as Ukrainian. This section summarises the main problems that arise during the analysis of texts from social networks, substantiates the need for a new system, and determines how the proposed approach is superior to its analogues. Disadvantages of existing solutions: 1. Low accuracy for local languages. Most modern text analysis systems, such as Google Cloud Natural Language and Microsoft Azure Text Analytics, are focused on Englishlanguage content. The Ukrainian language is often underrepresented in educational datasets, resulting in low accuracy in mood analysis and language definition. Many models overlook the slang, abbreviations, and specific grammar of local languages, which compromises the quality of the analysis. 2. Limited support for multilingualism. Existing systems have limited support for multilingual content, especially for texts that contain mixed languages (e.g., Ukrainian and English). Multilingual models, such as mBERT, often exhibit reduced efficiency for less common languages, including English. 3. The complexity of working with the texts of social networks. Texts from social networks have specific features, such as slang, abbreviations, emojis, and symbols, that are not taken into account by many existing models. Dynamic content, such as Instagram comments, uploaded via JavaScript, creates technical challenges for data collection. The lack of adaptation to short texts, which are often found in comments, reduces the accuracy of the analysis. 4. Lack of integration. Most systems do not offer a comprehensive approach that includes parsing, sentiment analysis, language group definition, and trending over time. Existing solutions often require the integration of multiple tools, making them difficult to use.
Based on the analysis of the shortcomings of existing solutions, the following key problems are formulated, which are solved by the developed system: 1. Improving accuracy for local languages. The system under development utilises modern transformer models, such as XLM-RoBERTa, which offer high accuracy for Ukrainian, Russian, and English. Adding models to specific data from social networks enables you to account for the peculiarities of slang, abbreviations, and symbols. 2. Expanding support for multilingualism. The system provides analysis of mixed-language texts, including determining the language group for each comment. The integration of multilingual models enables you to work effectively with texts in multiple languages. 3. Adaptation to texts from social networks. The system takes into account the specific characteristics of texts from social networks, including emojis, symbols, and abbreviated text forms. The use of specialised algorithms to collect data from dynamic content (Selenium) provides access to comments that are loaded via JavaScript. 4. Integrated approach. The system integrates all the key stages of analysis, including parsing comments, determining language, analysing moods, building trends over time, and visualising results. The possibility of interactive data analysis through visualisation (Plotly) is provided.
Justification of the need to develop the system: 1. The uniqueness of the proposed approach. Existing solutions focus on specific aspects of text analysis, such as parsing or sentiment analysis, but do not offer a comprehensive approach. The system under development integrates all stages of analysis, which provides a complete cycle of data processing from social networks. 2. Advantages over analogues: accuracy, flexibility, interactivity and comprehensiveness. The use of XLM-RoBERTa enables you to achieve high accuracy for multilingual content, including Ukrainian. The system is adapted to texts from social networks, taking into account their specifics. Data visualisation through Plotly enables users to interact with the analysis results. The system under development encompasses all stages of data analysis within a single product. 3. Importance for Ukraine. The development of a system that takes into account the specific characteristics of the Ukrainian language contributes to the advancement of local technologies in the field of natural language processing (NLP). The system can be used to monitor public sentiment, assess the effectiveness of information campaigns, and combat disinformation.
Existing social media text analysis solutions have significant drawbacks that limit their effectiveness for multilingual content and local languages. The system under development addresses these problems by integrating modern models for sentiment analysis, multilingual text processing, and adaptation to the specifics of social networks, while providing an integrated approach to data analysis. Its implementation will contribute to improving the accuracy of analysis, expanding support for local languages, and providing interactive data analysis, making it an important tool for businesses, researchers, and organisations, especially in the context of Ukraine.
The analytical review revealed that existing solutions for analysing texts from social networks have significant limitations, making it difficult to utilise them effectively for multilingual content, particularly for local languages such as Ukrainian. The main drawbacks include low accuracy of sentiment analysis for less common languages, limited support for texts from social networks, difficulty working with dynamic content, and a lack of a comprehensive approach to data analysis. The considered tools and methods, such as Google Cloud Natural Language, Microsoft Azure Text Analytics, VADER, XLM-RoBERTa, Selenium, and Plotly, demonstrate strengths in certain aspects of the analysis but do not provide a comprehensive data processing cycle from social networks. They are either focused on English-language content or do not take into account the specifics of texts from social networks, such as slang, emojis, and mixed languages. The developed system has several significant advantages over existing analogues, including multilingualism, adaptation to texts from social networks, comprehensiveness, interactivity, and flexibility. The use of the modern XLM-RoBERTa model yields high accuracy in sentiment analysis for Ukrainian, Russian, and English, taking into account the characteristics of local language groups. The system takes into account the specific characteristics of texts from social networks, such as slang, abbreviations, emojis, and symbols, which enhances the quality of analysis. Integration of all key stages of analysis – parsing comments, determining language, sentiment analysis, building trends over time, and data visualisation – within one product. Using Plotly to visualise results enables interaction with the data, making it easier to analyse and interpret. The system can be further trained on specific data from social networks, which allows it to be adapted to new challenges and tasks.
The developed system addresses the key problems that arise during the analysis of texts from social networks, offering multilingualism, adaptation to local languages, and an integrated approach to data processing. Its implementation will contribute to increasing the efficiency of text analysis, the development of local technologies, and the creation of new opportunities for businesses, researchers, and organisations.
System analysis is a crucial stage in the development of information systems, as it enables a thorough exploration of the subject area, identification of key system requirements, and justification of its architecture and structure. In the context of developing a system for automating the analysis of comments from social networks, system analysis provides an understanding of the complex, multi-level interaction between the system's components, as well as determining the optimal solutions for addressing the tasks. Social networks, such as Instagram, generate vast amounts of text data that contain valuable insights into user sentiment, trends, and behaviour. However, analysing this data is a challenge due to multilingualism, the use of slang, emojis, symbols, and the dynamic nature of the content. Developing a system that automates the analysis of such data requires a systematic approach that takes into account all aspects, from data collection to processing, analysis, and visualisation. The main goals of the system development include: 1. Automation of data collection from social networks – ensuring effective scraping of comments from dynamic Instagram content. 2. Multilingual analysis of texts – determination of the language group of comments, analysis of moods and taking into account the specifics of local languages (Ukrainian, Russian, English). 3. Building sentiment trends over time – identifying changes in user sentiment depending on time and other factors. 4. Interactive visualisation – creating graphs and reports that allow users to interact with the results of the analysis.
System analysis will not only allow us to determine the main components of the system and their interactions, but also to justify the choice of architecture and implementation methods. It will ensure the creation of an efficient, reliable, and scalable system that addresses the current challenges of analysing texts in social networks.
Social media platforms like Instagram are platforms where users actively interact through text, images, videos, and comments. Instagram allows users to publish posts that other users can comment on and react to with likes and emojis. Comment texts on social networks have several essential characteristics that affect their analysis: content dynamism, multilingualism and a large amount of data. Comments are constantly changing, new ones are added, and old ones can be deleted. Instagram is a global platform, allowing comments to be written in a variety of languages, including English, Ukrainian, Russian, and others. Popular posts can garner thousands of comments, which creates challenges in processing large amounts of information. Multilingual content on Instagram often features comments written in different languages within the same post. It creates challenges for analysis, such as speech recognition (automatic detection of the language of the comment), processing of texts with different grammatical structures (each language has its own rules for constructing sentences that must be taken into account), translation and normalisation (for sentiment analysis, comments can be translated or brought to a single format). The dynamism of comments means that the data is constantly changing. It creates challenges such as the timeline and the relevance of the data. Comments are added in real-time, which is essential for analysing mood changes. Old comments may become irrelevant over time, so the system must take this into account. Social media comment texts often feature several characteristics, including slang and abbreviations, emojis, spelling mistakes, and short phrases. Users frequently use informal words, abbreviations, and acronyms, such as "OMG" and "LOL". Emojis are an integral part of communication on Instagram, as they effectively convey emotions and moods. Comments often contain errors due to the speed of writing or the informality of the platform. The comments are typically brief, making it challenging to comprehensively analyse the context.
The general goal of the system is to develop an automated system for analysing texts in Instagram comments to determine user sentiments, taking into account multilingualism, the dynamism of content, and the specificity of texts. The primary objectives are data collection, text processing, mood analysis, and visualisation of results. Data collection involves parsing comments, processing dynamic data, and utilising web scraping. Text processing encompasses normalisation, pre-processing, and processing of multilingual content. Sentiment analysis involves pinpointing sentiments, analysing emojis, and utilising Transformer models. Visualising the results consists of distributing sentiments by language, displaying mood changes in graphs, and filtering data. The results of achieving the objectives will define the criteria for the system's functioning, including relevance, accuracy, speed, scalability, and versatility.
To develop a system for analysing texts in Instagram comments, several alternative options are being considered. These options included a selection of models for sentiment analysis, tools for parsing comments, and methods for processing multilingual content. After analysing the advantages and disadvantages of each option, it was decided to use the transformer models for sentiment analysis, since they provide high accuracy and take into account the context of the texts. Alternative options for sentiment analysis: 1. Rule-based models use predefined lexicons (for example, dictionaries of positive and negative words). Advantages: ease of implementation, does not require large amounts of data for training. Disadvantages include low accuracy for texts containing slang, emojis, and abbreviations. 2. Machine learning-based models use algorithms such as Logistic Regression, Random Forest, or SVM. Advantages: high accuracy for texts with well-defined features.
Disadvantages: require large amounts of data for training. 3. Deep Learning-based models use neural networks such as LSTM, GRU, or transformers (e.g., BERT, RoBERTa). Advantages: high accuracy for complex texts, takes into account the context. Disadvantages include high computational complexity and the need for significant resources for training.
For sentiment analysis, transformer models (e.g., BERT, RoBERTa, or XLM-R) were chosen, as they provide the highest accuracy for complex texts, take context into account, and support multilingualism. It is vital for analysing texts on Instagram, where comments may contain slang, emojis, and abbreviations. Alternative options for parsing comments: 1. Instagram API – Using the official API to fetch data. Advantages: reliability, access to upto-date data. Disadvantages include restrictions on the number of requests and the need for authorisation. 2. Web Scraping – using libraries for parsing HTML (for example, BeautifulSoup).
Advantages: Ability to bypass API restrictions. Disadvantages: Risk of violating Instagram's terms of service and privacy policy.
Web Scraping was chosen for parsing comments because it provides reliable access to up-todate data. In the event of API restrictions, alternative approaches can be used as a fallback.
Alternative options for handling multilingual content: 1. Automatic language detection – using libraries such as langdetect or fastText. Advantages: speed and accuracy of language detection. Disadvantages: possible errors for short texts. 2. Multilingual models for sentiment analysis – using transformers such as mBERT or XLM-R.
Advantages: Support for multiple languages, taking context into account. Disadvantages: high computational complexity.
Multilingual transformer models (e.g., mBERT or XLM-R) were chosen to handle multilingual content because they support text analysis in multiple languages and take into account context, which is crucial for multilingual content on Instagram.
To select the best option, the analytical hierarchy process (AHP) method was employed. A comparison of alternative options was carried out according to the following criteria:
Based on the scores obtained using these criteria, it was decided to utilise transformer models for sentiment analysis, Instagram APIs for parsing comments, and multilingual models for processing multilingual content. Thus, the systematic analysis of the object of study and the subject area made it possible to consider all the features of the texts in Instagram comments and select the best options for implementing the system. The system under development is designed to automate the process of analysing comments on social media platforms, particularly on Instagram. Its main functions are:
1. Comment Scraping – Automatically collect comments from social media posts, including dynamic content that is loaded via JavaScript. Parsing is provided using the Selenium tool. Also, collecting metadata such as the time the comment was published, the author, and the related post. 2. Language Definition – Automatically detects the language group of each comment (for example, Ukrainian, Russian, English, or mixed languages). Also, marking comments that consist only of symbols or emojis. 3. Sentiment analysis – the use of modern transformer models, such as XLM-RoBERTa, to determine the polarity of moods (positive, negative, neutral). Additionally, adapting models to the specificities of texts from social networks, including slang, abbreviations, and emojis. 4. Trend building – analysing mood changes over time, creating time series, and identifying long-term trends. Additionally, the construction of graphs that reflect the dynamics of moods over time, language groups, or other factors. 5. Interactive visualisation – using interactive graphs (based on Plotly) to present analysis results that allow users to interact with data, scale graphs, and highlight key segments.
The system is a universal tool for multilingual text analysis, enabling the automation of data collection, processing, and presentation. The system has a wide range of potential users, including: 1. Businesses, in particular, analyse consumer reactions to products, services, or marketing campaigns, as well as identifying trends in customer feedback to make informed decisions. 2. Researchers, such as sociologists, linguists, and analysts, can use the system to study public opinion, audience behaviour, and language characteristics, as well as monitor changes in the population's mood on socially significant topics. 3. Organisations, such as civil society organisations, can utilise the system to monitor public sentiment, evaluate the effectiveness of information campaigns, and detect disinformation. Government agencies can also use the system to analyse public opinion on political/social initiatives.
The system is helpful for any organisation that works with large amounts of text data on social media platforms and requires automated analysis. Social media text analysis faces several challenges that limit the effectiveness of existing solutions, including low accuracy for local languages, limited support for multilingualism, the complexity of working with social media texts, and a lack of integrated solutions. Most commercial systems, such as Google Cloud Natural Language and Microsoft Azure Text Analytics, are primarily focused on English-language content and exhibit lower accuracy for Ukrainian and Russian. Existing solutions do not cope well with texts that contain mixed languages, slang and symbols. Features of texts from social networks, such as abbreviations, emojis, and dynamic content, are often overlooked. Most systems focus on specific aspects of analysis (for example, only parsing or only sentiment analysis), which makes it challenging to use them for complex analysis. The system under development addresses these problems by integrating modern technologies, including transformers (XLM-RoBERTa), multilingual models for language detection (FastText), and tools for interactive visualisation (Plotly). Expected effects from the implementation of the system: improving the accuracy of sentiment analysis, supporting multilingualism, process automation, interactivity, and the development of local technologies. The use of modern models for multilingual analysis provides high accuracy even for local languages (Ukrainian, Russian). The system supports text analysis in different languages, including mixed languages, making it versatile for working in multilingual environments. Automatic parsing of comments and sentiment analysis significantly reduces labour costs for processing large amounts of data. Interactive graphs allow users to easily analyse the results and gain deeper insights. The system contributes to the development of natural language processing (NLP) technologies for the Ukrainian language, which is essential in the context of Ukraine. Conceptual model of the system: 1. Input: Comments (text data collected from Instagram, including mixed languages, symbols, and emojis) and metadata (time the comment was published, author, related post). 2. Initial data: the results of sentiment analysis (the polarity of each comment as positive, negative, or neutral), trends (graphs of mood changes over time) and reports (interactive reports including sentiment analysis, frequency of language groups, time trends). 3. Functions and structure of the system: parsing module (automatic collection of data from Instagram), language detection module (classification of the language group of each comment), mood analysis module (determination of the emotional polarity of texts) and visualisation module (construction of interactive graphs and reports). 4. System requirements: performance (fast processing of large amounts of data), accuracy (high accuracy of sentiment analysis for multilingual content), relevance (correspondence of data information in real time), and scalability (ability to process data from different sources, not only Instagram).
The system being developed is designed to automate the analysis of comments on social media platforms, particularly Instagram. Its implementation will enhance the accuracy of sentiment analysis, support multilingualism, and interactivity, making it a valuable tool for businesses, researchers, and organisations. The conceptual model of the system defines its key components and requirements, ensuring the effective implementation of tasks.
The system analysis made it possible to thoroughly investigate the problems associated with automating text analysis in social networks and to identify the key aspects of system development. The subject area was studied in detail, and the general goal of the system was formulated. A tree of goals was built, which structures the tasks that must be performed to achieve this goal. The analysis of alternative approaches to building a system made it possible to justify the choice of the optimal architecture, which provides efficiency, flexibility, and scalability. The key aspects of system development were: justification of the chosen architecture, modelling the system's structure and functions, and determining the system's requirements. The system is based on modern technologies, including transformer models (XLM-RoBERTa) for multilingual sentiment analysis, Selenium for the automatic collection of dynamic content, and seaborn for visualisation. This choice enables you to ensure high accuracy of analysis, adapt to the specifics of texts from social networks, and achieve ease of use. The requirements for performance, accuracy, interactivity and scalability of the system were formulated. It ensures its ability to process large amounts of data, work with multilingual content, and adapt to changes in the subject area. The creation of such a system is a crucial step in addressing the current challenges of analysing texts in social networks. It enables you to automate data collection, processing, and analysis, increase the accuracy of sentiment analysis, maintain multilingual support, and provide interactive visualisation of results. It makes the system useful for businesses, researchers, NGOs, and government agencies that work with large amounts of text data. Thus, the conducted system analysis laid the groundwork for the further implementation of a system that addresses the challenges of analysing texts in social networks and meets the modern requirements for information systems.
The choice of methods and means is a key stage in the development of information systems, as the efficiency, accuracy, and productivity of the final product depend on it. For the development of a system to automate the analysis of comments on social networks, particularly Instagram, it is essential to consider the specific characteristics of texts, which are often multilingual, unstructured, and contain slang, emojis, abbreviations, and symbols. It creates additional challenges for data collection, processing, and analysis. The main tasks that the system solves are: 1. Comment parsing – automatic collection of text data from social networks, including dynamic content generated by JavaScript. 2. Language Detection– Automatically detect the language group of texts, including support for mixed languages and texts with symbols. 3. Sentiment analysis – determining the emotional polarity of comments (positive, negative, neutral), taking into account the context. 4. Trend building – analysing mood changes over time and identifying long-term trends. 5. Interactive visualisation – presenting results in the form of graphs and reports that allow users to interact with data.
To address these issues, several key requirements for the methods and technologies employed in the system have been identified: accuracy, performance, multilingualism, adaptability, and interactivity. The methods should provide high accuracy of sentiment analysis, especially for local languages (Ukrainian, Russian) and multilingual content. The system must efficiently process large amounts of data, including thousands of comments from social networks. The methods should support the analysis of texts in various languages, including those that are mixed. The technologies used should take into account the specific characteristics of texts from social networks, such as slang, emojis, and abbreviations. Visualisation tools should provide ease of use and allow for interactive graphing. To achieve the set goals, modern technologies and libraries that meet these requirements were chosen. In particular, Selenium is used for parsing, while transformer models from the Transformers library are utilised for sentiment analysis. The langdetect library is employed for language detection, and the Seaborn and Matplotlib libraries are utilised for data visualisation. Each of these tools has been selected based on its benefits in its respective field, ensuring the system's efficiency and reliability. Thus, the choice of methods and means is a critical stage that determines the success of the system implementation. In the following sections, the choice of each technique and tool will be substantiated in detail, and a comparative analysis with analogous tools will be conducted to confirm their effectiveness.
Sentiment analysis is a key task for the system being developed, as it enables the determination of users' emotional reactions to content on social networks. For this, modern transformer models are used, which provide high accuracy of analysis due to the context. Advantages of transformer models include high accuracy, multilingualism, and adaptability. Transformers consider the context of each word, enabling you to accurately determine the emotional polarity of the text. Models such as XLM-RoBERTa support text analysis in different languages, which is critical for a multilingual social media environment. Models can be further trained on specific data from social networks, which allows slang, abbreviations, and emojis to be taken into account. Comparison with traditional methods: 1. Dictionary Approaches (VADER) is a simple sentiment analysis tool that is based on word polarity dictionaries. Erevaga: Ease of use, high accuracy for English-language content. Disadvantages: low accuracy for multilingual content and a lack of consideration for context. 2. Transformers offer significantly higher accuracy due to their contextual understanding, but require more resources to operate.
1. A multilingual version of XLM-RoBERTa that supports more than 100 languages, including Ukrainian, Russian, and English. It is used to analyse the sentiments of comments in Ukrainian. Model is cardiffnlp/twitter-xlm-roberta-base-sentiment. 2. RuBERT is a model designed for analysing texts in Russian, specifically adapted to the language's unique characteristics. It is used to analyse the sentiment of comments in Russian. Model is blanchefort/rubert-base-cased-sentiment. 3. RoBERTa is an English-language model that demonstrates high accuracy for texts from social networks. It is used to analyse the sentiment of comments in the English language.
Model is cardiffnlp/twitter-roberta-base-sentiment. 4. RoBERTa for symbols is used to analyse texts consisting only of symbols or emojis. Model is cardiffnlp/twitter-roberta-base-sentiment.
Determining the language of texts is an essential task for multilingual analysis. The system under development uses the langdetect library, which provides automatic detection of the language group of texts. The library uses statistical models to classify texts by language. Advantages: ease of integration into Python projects, high speed and support for more than 50 languages, including Ukrainian, Russian and English. Disadvantages include low accuracy for short texts, which are often found in comments, as well as limited support for mixed languages (for example, when Ukrainian and English are used in the exact text).
Comparison with other methods: 1. FastText is a library that demonstrates high accuracy for short texts and supports more than 170 languages. Advantages: high accuracy, adaptability to texts from social networks.
Disadvantages: resource requirements, complexity of integration. 2. Langdetect is a straightforward solution for basic language detection, but it is less effective for complex texts.
Pre-processing of texts is a critical step to ensure the accuracy of the analysis. The following approaches are used in the system being developed:
I. Pre-treatment: 1. Normalisation of texts: Remove unnecessary characters, spaces, and punctuation; Convert text to lower case; Using the unicodedata library to remove accents and special characters. 2. Cleaning Texts: Remove URLs, emojis, and other irrelevant items; Using regular expressions (re) to find and remove specific patterns; Tokenisation is the process of splitting text into separate words or tokens for further analysis.
1. Simplicity – regular expressions allow you to process text data efficiently. 3. Flexibility – the ability to adapt to the specifics of texts from social networks.
Data visualisation enables you to present the results of the analysis clearly and understandably. The following libraries are used in the system under development: 1. Seaborn is a tool for creating aesthetic charts. It is used to build bar graphs, heat maps, and line graphs. 2. Matplotlib is a basic library for creating graphs. It is used to adjust the details of graphs (colours, labels, axes). 3. Statsmodels is used to smooth data using the LOWESS (Locally Weighted Scatterplot
Smoothing) method. Allows you to identify trends in time series.
Advantages: flexibility (the ability to create both static and interactive graphs), aesthetics (graphs look modern and clear), and interactivity (integration with Plotly to build interactive graphs). The methods for solving the problem were chosen taking into account the specifics of texts from social networks, as well as the requirements for accuracy, productivity, and multilingualism. The use of transformer models for sentiment analysis, language detection libraries such as langdetect, regular expressions for text processing, and modern libraries for visualisation ensures the efficiency and reliability of the system. Each method was chosen based on its advantages in the relevant field, allowing you to solve tasks with high quality.
Python was chosen as the primary programming language to automate the analysis of comments from social networks. Python is a versatile, user-friendly, and powerful tool for working with text data, analysing large datasets, and developing complex systems. Its popularity in the field of natural language processing (NLP), the presence of numerous libraries, and an active community make Python the optimal choice for this project. The rationale for choosing Python is versatility, simplicity, community, and scalability. Python supports a wide range of libraries for working with texts, data analysis, and visualisation. Easy to digest and allows you to quickly develop complex systems – a large user base and an active community that provides access to documentation and support. Python is well-suited for working with both small datasets and large amounts of information. Overview of the libraries used: 1. Selenium is used to automatically scrape comments from dynamic Instagram content that JavaScript generates. Selenium enables you to interact with web pages in a manner similar to a real user, including clicking buttons, scrolling through pages, and filling out forms. Advantages: the ability to work with dynamic content and flexibility in setting up parsing for specific tasks. Disadvantages: high resource requirements compared to other tools. 2. Pandas is used to process and analyse structured data. Pandas allows you to work conveniently with data tables (DataFrames). Support for filtering, aggregation, ru, HEROES GLORY symbols_only,symbols_only,symbols_only,en,No ❤️ one will care as much as He is true en, Glory to Ukraine symbols_only,en,A ❤️❤️ uk,Win en,is with symbols_only, symbols_only, UK,take ❤️❤️❤️❤️❤️❤️❤️❤️❤️❤️ care of all and symbols_only,symbols_only,UK,MPs ' salaries and then there will be
Main_Language,Filtered_Comment en, I presented state awards Crosses of Military Merit and Orders of the Golden Star to our and handed the orders to the family members of the Heroes who were posthumously awarded this Everyone who Everyone who works for defense and truly for the state not just for Everyone who I thank I thank the families of our warriors for such for such We will undoubtedly endure this And we will certainly ensure a dignified life for great President to be respected and thank you to the Warriors of %❤️
The Analyze_data module uses transformer models that add a sentiment column to the database, where one of three values (positive, neutral, or negative) is possible (Fig. 8).
Database view after executing the module Process_data: Main_Language,Filtered_Comment,Sentiment en,I presented state awards Crosses of Military Merit and Orders of the Golden Star to our and handed the orders to the family members of the Heroes who were posthumously awarded this Everyone who Everyone who works for defense and truly for the state not just for Everyone who I thank I thank the families of our warriors for such for such We will undoubtedly endure this And we will certainly ensure a dignified life for,positive ru,HEROES GLORY,neutral symbols_only, symbols_only, symbols_only, ,positive ,positive ,positive en,No one will care as much as He is true,neutral en,Glory to Ukraine,positive en,A great President to be respected and thankyou to the Warriors of,positive uk,Win,positive symbols_only,
,neutral
In the Aggregate_data module, all files with posts are numbered and combined into one file (Fig. 9), where the id column indicates the post number, with 1 being the newest and each subsequent number representing an older post. And the sub_id column means the order in which the comment is displayed, which is formed by Instagram algorithms based on likes and other indicators.
id,sub_id,Main_Language,Sentiment,Filtered_Comment1,1,en,positive,I presented state awards Crosses of Military Merit and Orders of the Golden Star to our and handed the orders to the family members of the Heroes who were posthumously awarded this Everyone who Everyone who works for defense and truly for the state not just for Everyone who I thank I thank the families of our warriors for such for such We will undoubtedly endure this And we will certainly ensure a life for1,2,ru,neutral,HEROES dignified ,No
️❤ GLORY1,3,symbols_only,positive,1,5,symbols_only,positive,1,6,symbols_only,positive,1,7,en,neutral
will care as much as He is true1,8,en,positive,Glory to Ukraine1,9,symbols_only,positive,1 ❤️❤️, 10,en,positive,A great President to be respected and thankyou to the
Warriors of1,11,uk,positive,Win1,12,en,neutral,is with1,13,symbols_only,neutral,1,14,symbols_only,positive,1,15,uk,neutral,save ❤️❤️❤️❤️❤️❤️❤️❤️❤️❤️all %❤️ salaries and then it will be1, 19,symbols_only,positive,1,20,en,positive,I am continuously so impressed by
He seems to work so tirelessly for his country ' and God1,21,uk,neutral,Eternal1,22,uk,negative,Eternal and Light1,23,uk,neutral,how
The visualize_data module is the final module, in which the dataset is evaluated, and all the data within it is visualised.
In the following visualisation, you can see the number of comments for each post, where the first is the most recent, the twentieth is the oldest (Fig. 11a). Next, you can see how many comments were written by each of the language categories (Fig. 11b).
In the following images, you can see the tonality distribution for each language, as well as for each post (Fig. 12). Fig. 13 shows the distribution of comment lengths. Fig. 14 shows how the length of a comment depends on the key, as well as on the language in which it was written.
Fig. 15a analysed the longest comments. Fig. 15b presents the solidarity index in tabular form. Fig. 16a and Fig. 17a show the distribution of moods within the comments of each language. Next, the solidarity index for each language was analysed, as well as the sentiments expressed in the comments for each language across all posts (Fig. 16b and Fig. 17b).
In the following Figs. In Figures 18–20, you can see the characteristics of the distribution of the solidarity index for each of the languages.
Graphs in Figs. 21–25 show a change in the positivity of sentiments in the comments for each language over time. You need to read them from right to left, because 1 is the most recent post and 20 is the oldest. You can see that comments in Russian are very rarely positive, for apparent reasons (Fig. 22). Fig. 23 shows that comments from symbols are most often positive. The remarks in Ukrainian have the most stable positivity index, which fluctuates approximately within the range of 0.2-0.4 (Fig. 24). The positivity graph in Fig. 25, which combines comments from all languages, shows a rather volatile trend towards change.
So, following this control example, with the help of existing modules, you can download and analyse comments on posts from any Instagram page.
In modern information systems, the efficiency of software is a key factor in determining its practical value. Performance analysis enables you to assess how efficiently the system performs tasks, identify potential bottlenecks, and determine ways to optimise. This section presents a statistical analysis of the system's implementation for automating comment analysis in social networks. The main attention is paid to the following indicators: 1. The execution time of the main modules is a measurement of the time required to complete each stage of data processing (collection, processing, sentiment analysis, aggregation, and visualisation). 2. Data processing – analysis of the number of processed comments at each stage. 3. Resources – An estimate of the use of computing resources such as RAM and CPU.
It is necessary to evaluate the system's efficiency, identify its strengths and weaknesses, and provide recommendations for enhancing performance. The statistics will be presented in the form of tables, graphs, and charts, allowing you to visualise the results of the analysis and make them more understandable to users. The results obtained will not only confirm the effectiveness of the developed system, but also provide valuable information for its further improvement. Let's proceed to the analysis of the program's main modules. Since selenium and an automatic browser client are used to load (Load_data) all comments as the basis of the dataset, the downloading process takes a significant amount of time. After all, to make the process as secure as possible and the script faulttolerant, various approaches, such as WebDriverWait and Time.sleep(), are integrated into it (Fig. 26). The combination of these approaches avoids blocking the IP address due to excessive requests. Also, it ensures the fault tolerance of the script (Fig. 27a). Additionally, in guaranteeing the script's speed, storing browser cookies after the first authorisation plays a crucial role. It avoids a delay in logging into your account. Also, the performance of the script is affected by the number of posts uploaded to the target page, as well as the number of comments in each of them (Fig. 27b). In our case, when uploading 20 posts with an average of 180 comments in each of them, the execution time reached 36 minutes.
The primary function of the Process_data module is to filter comments and determine the primary language in each of them (Fig. 28). In order to save time and resources of the system during scaling, a highly efficient Langdetect library was chosen, which made it possible to reduce the module's execution time to about 8 seconds on a dataset of almost 4000 entries. The primary function of the Analyze_data module is to utilise transformer models specific to each language to assess the mood of each comment and add the corresponding designation to the dataset (Fig. 29). This analysis, which involves processing comments through models, is the most time-consuming step. The longer the comment, the longer it will take to process. Since all the models were preloaded and most comments on social media are pretty brief, the execution time was only 4 minutes.
The Aggregate_data module aggregates all received files into a single dataset using the pandas library, so it does not require much time. It took less than a second to execute (Fig. 30).
The Visualize_data module does not manipulate data; its primary function is to visualise an already created dataset (Fig. 31–32). Those graphs and statistics, which are built for the entire dataset, do not depend significantly on its size, and in general, each took less than a second. The only exception is the visualisation of mood trends, as we used the lowess function from the statsmodels library, which is picky about resources. However, the section for analysing individual posts took significantly longer – 6.2 seconds - and the time of its execution is directly proportional to the number of posts and comments within them.
The analysis of the system's implementation statistics enabled an assessment of its effectiveness, performance, and compliance with the assigned tasks. The results confirmed that the system successfully handles the functions of collecting, processing, analysing, and visualising large amounts of text data. Key takeaways: 1. Module execution time. The stage of mood analysis using transformer models takes the most time, which is due to the high computational requirements of these models. Other modules, such as data collection, word processing, and aggregation, are relatively fast due to the use of efficient algorithms and streamlined approaches. 2. System performance. The system is capable of processing large amounts of data, maintaining high accuracy at every stage. The use of multilingual transformer models provides a qualitative analysis of moods for Ukrainian, Russian and English. 3. Use of resources. The main load falls on the stage of mood analysis, where transformer models are used. It requires significant computing resources, including RAM and CPU time. Other stages, such as text scraping, filtering, and data aggregation, have low resource requirements. 4. Analysis of processed data. The system demonstrates high efficiency in working with texts of varying lengths and complexities, including comments with symbols, slang, and mixed languages. Visualising the results makes it easy to interpret the data obtained, which is essential for end users.
Recommendations for improvement: 1. Optimisation of transformer models – use of less resource-intensive models (e.g.
DistilBERT) for texts with low complexity, as well as integration of methods of preliminary classification of texts to determine which comments require detailed analysis. 2. Data collection optimisation – using Instagram's official API (subject to availability) to reduce comment collection time and avoid possible restrictions from the platform. 3. System scaling – the implementation of multithreading or distributed computing for parallel processing of large amounts of data, as well as the use of cloud services for the analysis of large data sets.
The developed system demonstrates high efficiency and productivity in solving the problems of analysing comments in social networks. It meets modern requirements for automating the processing of large amounts of text data, providing accurate sentiment analysis, multilingual capabilities, and convenient visualisation of results. However, further optimisation of computing resources and integration with other platforms will make the system even more versatile and productive.
accounts analysis of different categories of Instagram As an additional task, a new dataset was created, which included more than 20000 comments. Let's conduct a comprehensive analysis of various categories of Instagram accounts. The first category we will consider is the business page (Fig. 33). This dataset includes 1600 comments and three different pages (Fig. 34).
Therefore, even Ukrainian business accounts primarily focus on the English language. As expected, according to Fig. 35, symbolic comments contain the most positive dynamics, while in other languages, neutral comments prevail. On average, comments for business accounts are not very long, but they are nevertheless longer than those for personal pages (Fig. 36 –37). In these images, you can see that the comments in business accounts for the most part coincide in tone with the description from the author of the page.
Let's proceed to the analysis of the positivity of comments (Figs. 38-41). You can see that most comments in this topic are neutral or positive, with a clear trend towards improving the sentiment of comments over time.
The entertainment category contains 1766 comments from 3 different pages (Fig. 42). In category discussions, most of the comments turned out to be neutral (Fig. 43a), even in the symbol_only category. Comments turned out to be significantly longer than those in the category of business pages (Fig. 43b). The solidarity index was, on average, the same as that of business pages, but with a different distribution by language (Fig. 44–45).
Let's proceed to the analysis of moods (Figs. 46–50). You can see that pages from the entertainment category have a significantly lower number of positive comments compared to business accounts. However, a positive trend is also evident over time.
The last and largest analysed category is political accounts (Fig. 51). Three accounts, with a total of 17,000, were analysed. As in the category of business accounts, the largest categories of mood in languages are symbolic and English positive comments (Fig. 52a). It can be seen that in the sample of political commentaries there were significantly longer comments than in the previous categories (Fig. 52b). The solidarity index for political posts was considerably lower than for other categories (Fig. 53). The graph also shows that most of the solidarity comments are symbolic (Fig. 54).
Let's move on to sentiment analysis (Fig. 55–59). You can see that here, unlike other categories, there are clear peaks and troughs on the chart, indicating a high correlation between political events and the mood of comments. It is also worth noting that for comment schedule in Ukrainian and all languages, there is a strong trend towards improving the comments mood over time.
As a result of the study, an integrated approach to the automated analysis of Instagram comments was developed, which considers the multilingual nature of content, the dynamism of social networks, and the characteristics of the informal online communication environment. The proposed system provides a comprehensive cycle of data processing, encompassing automatic parsing of comments, language detection, sentiment analysis, time trend analysis, and interactive visualisation of results. Based on a critical analysis of existing solutions, it was found that most commercial and open source tools demonstrate low accuracy for local languages, in particular Ukrainian, work with short and mixed texts to a limited extent, and do not take into account the specifics of social networks (emojis, slang, symbols). It confirmed the need to create a specialised system capable of adapting to the real-world conditions of data processing on social platforms.
As part of the study, a method for multilingual analysis of Instagram comments using specialised transformer models, specifically XLM-RoBERTa and individual models for specific language groups, was proposed for the first time. It enabled the achievement of high accuracy in determining moods for Ukrainian, Russian, and English, as well as in processing content consisting only of symbols or emojis. The developed algorithms for identifying the language group and assessing the solidarity of comments with posts provide a deeper contextual analysis of user reactions. An additional scientific result is a technique for analysing mood dynamics over time, which combines time series models, moving averages, and clustering. It enables you to identify changes in the audience's emotional response, predict trends, and assess long-term engagement with content. Visualising results using interactive tools enhances the practical value of the system for businesses, researchers, analysts, and organisations. The created system is relevant and significant for Ukraine, as it contributes to the development of local NLP solutions, supports the analysis of Ukrainian-language content, enables the monitoring of public sentiment, counters disinformation, and provides new opportunities for business intelligence.
Thus, the developed system demonstrates high efficiency in the multilingual analysis of texts from social networks, opening up prospects for further improvement, including expanding language support, integrating new transformer models, and enhancing the accuracy of analysing emotional and semantic characteristics of comments.