Ioannis Kopanakis

Social media generates massive amounts of unstructured, multilingual textual data that contain many sentiments, emotions, and opinions from users that provide insights for enterprises undergoing digital transformation. In this paper, we evaluate a variety of commercially available, lightweight large language models (LLMs) for sentiment, emotion, aspect-based sentiment, and toxic speech detection on noisy, real-world social media datasets. These lightweight models are compared (against traditional RoBERTa-based classifiers and heavyweight “pro” LLMs) in terms of performance, latency, and cost trade-offs that are key for scalable deployment across the enterprise. Our results show that lightweight LLMs achieve good accuracy with considerably lower response times and costs, which makes them suitable as the next step in digital transformation for real-time social media analytics. Additionally, we investigate the importance of prompt engineering and show that preprocessing can be very limited for LLMs. This study provides operational guidelines regarding model selection based on latency, accuracy, and cost trade-offs as well as optimal prompt engineering techniques. Finally, it evaluates the intersection of operational performance and resource efficiency to help researchers and developers integrate LLM-based social media analytics into digitized business practices.

Social media generate massive amounts of unstructured, multilingual textual data that contain many sentiments, emotions and opinions from users that provide insights for enterprises undergoing digital transformation. In this paper, we evaluate a variety of commercially available, lightweight Large Language Models (LLMs) for sentiment, emotion, aspect-based sentiment, and toxic speech detection on noisy, real-world social media datasets. These lightweight models are compared (against traditional RoBERTa-based classifiers and heavyweight “pro” LLMs) in terms of performance, latency, and cost trade-offs that are key for scalable deployment across the enterprise. Our results show that lightweight LLMs achieve good accuracy with considerably lower response times and costs, which makes them suitable as the next step in digital transformation for real-time social media analytics. Additionally, we investigate the importance of prompt engineering and show that preprocessing can be very limited for LLMs. This study provides tactical knowledge for researchers and developers seeking to integrate LLM based social media analytics into digitized business practices and evaluates the intersection of operational performance and resource efficiency.

The vast amount of social media and web data offers valuable insights for purposes such as brand reputation management, topic research, competitive analysis, product development, and public opinion surveys. However, analysing these data to identify patterns and extract valuable insights is challenging due to the vast number of posts, which can number in the thousands within a single day. One practical approach is topic clustering, which creates clusters of mentions that refer to a specific topic. Following this process will create several manageable clusters, each containing hundreds or thousands of posts. These clusters offer a more meaningful overview of the discussed topics, eliminating the need to categorise each post manually. Several topic detection algorithms can achieve clustering of posts, such as LDA, NMF, BERTopic, etc. The existing algorithms, however, have several important drawbacks, including language constraints and slow or resource-intensive data processing. Moreover, the labels for the clusters typically consist of a few keywords that may not make sense unless one explores the mentions within the cluster. Recently, with the introduction of AI large language models, such as GPT-4, new techniques can be realised for topic clustering to address the aforementioned issues. Our novel approach (AI Mention Clustering) employs LLMs at its core to produce an algorithm for efficient and accurate topic clustering of web and social data. Our solution was tested on social and web data and compared to the popular existing algorithm of BERTopic, demonstrating superior resource efficiency and absolute accuracy of clustered documents. Furthermore, it produces summaries of the clusters that are easily understood by humans instead of just representative keywords. This approach enhances the productivity of social and web data researchers by providing more meaningful and interpretable results.

The wealth of data available in social media and the web can be valuable for many purposes, including brand reputation management, topic research, competitive analysis, product development, public opinion surveys and many more. However, analyzing this data to identify patterns and extract useful conclusions can be challenging due to the volume of collected posts, which can be several thousands, even for a single day. One useful approach is the use of topic clustering, to create clusters of mentions that refer to a specific topic. Doing so, will result in a number of manageable clusters that can contain hundreds or thousands of posts. With these clusters, one can now have a much more meaningful overview of the topics being discussed, instead of having to go through each post and manually categorize them into topics. There are several topic detections algorithms that can achieve clustering of posts, such as LDA, NMF, BERTopic etc. The existing algorithms however have several important drawbacks, including language constraints and slow or resource-intensive processing of data. More importantly, the label for the clusters is usually comprised of a few keywords, that do not make much sense unless someone browses through the mentions inside the cluster. Recently, with the introduction of AI Large Language Models, such as GPT-4, new techniques can be realized for topic clustering, to address the aforementioned issues. Our novel approach (AI Mention Clustering) employs LLMs at its core, to produce an algorithm for efficient and accurate topic clustering of web and social data. Our solution was tested on actual social and web data and was compared to the popular existing algorithm of BERTopic, demonstrating superior resource efficiency and absolute accuracy of clustered documents. Furthermore, it produces summaries of the clusters and are easily understood by humans, instead of just representative keywords. This way, it makes the work of a researcher of social & web data, faster, easier and more productive.

It is estimated that European households are responsible for 55% of food waste generation. Key factors contributing to household food waste generation include food spoilage, confusion over expiration dates, overbuying, and inadequate shopping planning. Thus, food waste prevention at the household level depends heavily on food supplies monitoring and management. To this end, during the last decade, several consumer-oriented digital tools have been designed and launched. A literature review showed that currently accessible digital tools are scarce and cover a narrow range of functionalities. Here, we address these issues by designing and launching a decision support tool implemented in a smart mobile phone application (app), the FoodSaveShare Mobile App. The application development followed a traditional client–server architecture using state-of-the-art software and hardware technologies. Additionally, a survey of 340 individuals was conducted to better understand end-user motivation for and barriers against adopting this and similar apps. The developed application combines user-provided data with a retailer loyalty program to leverage the integrated features for tracking shopping activities. The app features a household shopping list populated by product barcode scanning and manual entry. Based on food and packaging type, food products are assigned approximate expiration dates to issue product expiration reminders. For products about to expire, suggestions for their utilization are provided, drawing from a list of over 7000 recipes. Additional functionality allows users to identify products that have either been consumed in time or that need to be discarded. Analytical tools, such as past purchase and resources discarded versus resources saved statistics, offer comprehensive insight and encourage improved shopping and consumption practices. The FoodSaveShare App was launched during the A2UFood Project, which allowed an organised campaign for its use. The app was tested under real customer data and conditions, and selected features have been adopted by the largest supermarket chain on the Island of Crete, Greece. The potential end-user survey results suggest that, provided personal data use issues are addressed, such apps can have a significant impact on reducing household food waste. Future work will focus on analysing the datasets produced by the application to assess its impact on household food waste management.