Enhancing Sentiment Analysis with Multimodal Large Language Models

This study aimed to evaluate the diagnostic accuracy of multimodal large language models in classifying superior labial frenulum attachments from intraoral photographs using expert consensus as the reference standard. Five experts (two periodontists and three orthodontists) established the consensus standard by classifying frenulum attachments in 117 intraoral images as mucosal, gingival, papillary, and papilla penetrating. The same photographs were then presented to three multimodal large language models (ChatGPT 4o, Gemini 2.5 Pro, and Microsoft Copilot GPT-4), and their diagnostic performance was evaluated using accuracy, sensitivity, specificity, and F1 score. Reliability was assessed using Fleiss' and Cohen's Kappa, and diagnostic performances were compared using Cochran's Q test. Human raters demonstrated almost perfect agreement (κ = 0.838, p < 0.001), whereas large language models showed poor inter-model agreement (κ = -0.124, p < 0.001). ChatGPT achieved slight significant agreement with the consensus (κ = 0.114, p = 0.019), although its clinical relevance was negligible. Gemini (κ = 0.099) and Copilot (κ = 0.027) showed no significant agreement (p > 0.05). Copilot yielded the highest overall accuracy (46.2%), followed by Gemini (44.5%) and ChatGPT (35.0%). The performance of the large language models varied across frenulum types. Current multimodal large language models demonstrate inconsistent and clinically insufficient accuracy in the classification of superior labial frenulum attachments from photographs. Domain-specific training is essential before large language models can be considered reliable diagnostic tools in dentistry.

Sentiment Analysis has emerged as one of the most challenging and significant research directions in text mining. The first milestone of Sentiment Analysis involves identifying the polarity of sentiments in social media and classifying them as positive or negative. Most of the previous work on sentiment analysis has focused on designing a classifier, for a specific dataset, using supervised machine learning approaches and feature extraction methods. Notably, choosing the most appropriate classification technique to correctly classify the sentiment polarity of text is crucial. Thus, this study proposes a novel ensemble classifier approach, which utilizes a combination of multiple feature sets with ensemble classification by combining multiple base classifiers, which are weak learners, into an ensemble classifier. These feature sets include Bag of Words, Term Frequency–Inverse Document Frequency, Part of Speech, N-gram, Opinion Lexicon, and Term Frequency. The experiments confirm that the proposed ensemble method outperforms all individual classifiers and significantly improves the overall sentiment classification performance on the most frequently used datasets in Sentiment Analysis.

The utilization of sentiment analysis has gained significant importance as a valuable method for obtaining meaningful insights from textual data. The research progress in languages such as English and Chinese has been notable. However, there is a noticeable dearth of attention towards creating tools for sentiment analysis in the Bangla language. Currently, datasets are limited for Bangla sentiment analysis, especially balanced datasets capturing both binary and multiclass sentiment for e-commerce applications. This paper introduces a new sentiment analysis dataset from the popular Bangladeshi e-commerce site “Daraz”. The dataset contains 1000 reviews across 5 product categories, with both binary (positive/negative) and multiclass (very positive, positive, negative, very negative) sentiment labels manually annotated by native Bangla speakers. Reviews were collected using an organized process, and labels were assigned based on standardized criteria to ensure accuracy. In addition, a benchmark evaluation of the performance achieved by Machine Learning and Deep Learning algorithms on this dataset is also provided. The new dataset can aid research on multiclass and binary Bangla sentiment analysis utilizing both machine learning, deep learning, and Large Language Models. It can aid e-commerce platforms in analysing nuanced user opinions and emotions from online reviews. The utilization of categorized product reviews also facilitates research in the field of text categorization.

Sentiment analysis is a computational technique employed to extract and interpret subjective information from textual data. It involves the identification and classification of sentiments, opinions, and emotions expressed within the text. By analyzing linguistic cues, such as word choice, syntax, and sentiment lexicons, sentiment analysis can discern a range of emotions, from positive to negative, as well as more nuanced sentiments, such as anger, joy, or surprise. This powerful tool has the potential to unlock valuable insights from vast amounts of unstructured text data, which enables informed decision-making and effective communication in various domains, including education. Recent advances in sentiment analysis have leveraged the power of deep neural networks, particularly general-purpose Large Language Models (LLMs) trained on extensive labeled datasets. However, real-world applications frequently encounter challenges related to the availability of large, high-quality labeled data and the computational resources necessary for training such models. This research addresses these challenges by investigating effective strategies for utilizing LLMs in scenarios with limited data and computational resources. Specifically, this study explores three techniques: zero-shot learning, <i>N</i>-shot learning and fine-tuning. By evaluating these methods, this research aims to demonstrate the feasibility of employing general-purpose LLMs for sentiment analysis within educational contexts even when access to computational resources and labeled data is limited. The findings of this study reveal that different adaptation methods lead to significantly different LLM performance outcomes.

To develop and evaluate the effectiveness of multimodal machine learning approach for the differentiation of NTM from MTB. The clinical data and CT images of 175 patients were retrospectively obtained. We established clinical data-based model, radiomics-based model, and multimodal (clinical plus radiomics) model gradually using 5 machine learning algorithms (Logistic, XGBoost, AdaBoost, RandomForest, and LightGBM). Optimal algorithm in each model was selected after evaluating the differentiation performance both in training and validation sets. The model performance was further verified using external new MTB and NTM patient data. Performance was also compared with the existing approaches and model. The clinical data-based model contained age, gender, and IL-6, and the RandomForest algorithm achieved the optimal learning model. Two key radiomics features of CT images were identified and then used to establish the radiomics model, finding that model from Logistic algorithm was the optimal. The multimodal model contained age, IL-6, and the 2 radiomics features, and the optimal model was from LightGBM algorithm. The optimal multimodal model had the highest AUC value, accuracy, sensitivity, and negative predictive value compared with the optimal clinical or radiomics models, and its' favorable performance was also verified in the external test dataset (accuracy = 0.745, sensitivity = 0.900). Additionally, the performance of multimodal model was better than that of the radiologist, NGS detection, and existing machine learning model, with an increased accuracy of 26, 4, and 6%, respectively. This is the first study to establish multimodal model to distinguish NTM from MTB and it performs well in differentiating them, which has the potential to aid clinical decision-making for experienced radiologists.

Sentiment analysis refers to the automated process of identifying the underlying emotional tone conveyed in a text. This task has become increasingly crucial today due to the exponential growth of opinion-oriented text generation on various online platforms including social media and e-commerce websites. The process of sentiment analysis has evolved through different phases like the use of handcrafted rules, sentiment-oriented dictionaries, machine learning, deep learning, transformer models, and large language models. These different phases have enriched the field of sentiment analysis and enabled it to get better insights into text for sentiment analysis. One of the most impactful developments in sentiment analysis was the creation of embedding’s for words or sentences, which are useful for representing unstructured data as structured data. These embedding are representing words or sentences as numerical vectors in a high-dimensional space and have proven effective in capturing the semantic relationships between words. By leveraging the power of embedding’s, sentiment analysis models can understand the complexities of human language and provide more precise insights into people's emotions and opinions. However, word embedding’s not inherently generated for sentiment analysis, which means that they are not inherently sentiment-oriented. Therefore, many researchers have focused on developing sentiment-oriented word embedding’s for sentiment words, but they have overlooked the importance of intensity words such as “little”, “high”, and “extremely”. These intensity words determine the level of sentiment expressed in text, which is particularly useful for fine-level sentiment analysis. To address this issue, a review proposes an intensity-aware word embedding development model and sentiment analysis with ML.

Sentiment analysis has changed completely in 2024, with developments in deep learning and machine learning, and multimodal approaches. This paper reviews ten recent studies that explore various sentiment analysis techniques, including transformer-based models (GPT-4, LLAMA 3, FinBERT), conventional techniques for machine learning (Nave Bayes, Logistic Regression), and multimodal frameworks integrating text and images. The findings suggest that large language models (LLMs) perform well in Learning situations with zero-shot and few-shot but struggle with complex sentiment understanding. Traditional models like Logistic Regression remain competitive in financial sentiment prediction, while multimodal approaches such as M2SA and topic-oriented models excel in image-text sentiment analysis. This review highlights The advantages and disadvantages of different techniques, supplying information for researchers and practitioners in choosing the most suitable approach for their particular uses.

Pneumothorax requires rapid recognition and accurate interpretation of chest X-rays (CXRs), particularly in acute settings where delays can have serious consequences. With the emergence of advanced image interpretation models capable of visual analysis, their diagnostic reliability in radiology practice remains to be determined. This study aimed to assess the diagnostic performance of three state-of-the-art systems in detecting pneumothorax using a large, well-annotated dataset. A total of 10,675 CXRs from the publicly available SIIM-ACR Pneumothorax Segmentation dataset were analyzed. Three multimodal models (GPT-4o, Gemini 2 Pro, and Claude 4 Sonnet) were evaluated using a uniform, image-based approach. Each model's binary outputs (presence: 1, absence: 0) were compared with reference results to determine accuracy, sensitivity, specificity, precision, and F1 scores. Additional subgroup analyses were conducted across pneumothorax size categories: small, medium, and large. Pairwise statistical comparisons were performed using McNemar's test. Sensitivity, specificity, and overall accuracy are reported with corresponding 95% confidence intervals. The prevalence of pneumothorax in the dataset was 22.3% (n = 2,379). All models demonstrated high specificity (above 0.90) but consistently low sensitivity (0.16-0.36). The best overall performance was observed with Gemini 2, which achieved an accuracy of 0.79 and specificity of 0.95, whereas Claude 4 showed greater sensitivity (0.20-0.34) across lesion-size categories. Diagnostic performance improved with increasing pneumothorax size, but smaller lesions remained difficult to identify. Pairwise comparisons confirmed statistically significant differences among all evaluated systems (P < 0.050). In this large-scale evaluation, the tested models exhibited strong reliability in identifying normal examinations but limited ability to detect subtle or small pneumothoraxes. Despite high specificity, low sensitivity limits the use of current Multimodal large language models as rule-out tools for pneumothorax. With continued refinement, these models may eventually support radiologists by improving workflow efficiency and diagnostic confidence. Automated systems capable of high specificity but low sensitivity should not be relied upon to exclude pneumothorax. However, they may serve as valuable assistants for confirming positive findings and prioritizing urgent cases in busy clinical workflows.

The integration of visual encoders and large language models (LLMs) has driven recent progress in multimodal large language models (MLLMs). However, the scarcity of high-quality instruction-tuning data for vision-language tasks remains a challenge. The current leading paradigm, such as LLaVA, relies on language-only GPT-4 to generate data, which requires pre-annotated image captions and detection bounding boxes, suffering from understanding image details. A practical solution to this problem would be to utilize the available multimodal large language models to generate instruction data for vision-language tasks. However, it's worth noting that the currently accessible MLLMs are not as powerful as their LLM counterparts, as they tend to produce inadequate responses and generate false information. As a solution for addressing the current issue, this paper proposes the Visual Instruction Generation and Correction (VIGC) framework that enables multimodal large language models to generate instruction-tuning data and progressively enhance its quality on-the-fly. Specifically, Visual Instruction Generation (VIG) guides the vision-language model to generate diverse instruction-tuning data. To ensure generation quality, Visual Instruction Correction (VIC) adopts an iterative update mechanism to correct any inaccuracies in data produced by VIG, effectively reducing the risk of hallucination. Leveraging the diverse, high-quality data generated by VIGC, we finetune mainstream models and validate data quality based on various evaluations. Experimental results demonstrate that VIGC not only compensates for the shortcomings of language-only data generation methods, but also effectively enhances the benchmark performance. The models, datasets, and code are available at https://opendatalab.github.io/VIGC

AI-enabled language models (LMs) to large language models (LLMs) and multimodal large language models (MLLMs) in drug discovery and development

Aim/Purpose: This paper focuses on app review analysis techniques, driven by the rapid advancement of the mobile app market and NLP techniques in optimizing mobile app user experiences. Background: Owing to technological advancements, app review analysis has rapidly evolved. This study examines both conventional and emerging techniques, including current advancements such as large language models (LLMs) in app review analysis. It provides an overview of the various methods used across different categories of app review analysis, comparing effective strategies for identifying user concerns and enhancing app functionality. Methodology: A systematic review was utilized based on two major standard guidelines, PRISMA and Kitchenham’s guidelines, for the period of 2014 to 2024. After defining the review protocol, papers were identified through keyword-based searches on six major online databases: Scopus, Web of Science, IEEE Xplore, ACM Digital Library, Science Direct, and Springer. Following screening and excluding papers based on defined quality criteria, 53 papers were considered for this study. The use of PRISMA ensures a transparent and reproducible review process, while Kitchenham’s guidelines provide a structured and rigorous approach for evaluating and synthesizing the literature. Contribution: This review study aims to evaluate the current state of knowledge on app review analysis techniques to improve mobile app user experiences. This study categorized the existing state-of-the-art papers into eight different categories, such as sentiment analysis, review classification, summarization, and prioritization, and examined challenges related to app review analysis. Furthermore, the study emphasizes the potential of LLMs for optimizing and automating app review analysis and provides future directions to address gaps in user-centric app development. Findings: Among the eight main categories defined in app review analysis, sentiment analysis is the most prevalent, followed by review classification and information extraction. Most studies use a combination of these categories to achieve a comprehensive goal. Prioritization techniques such as risk matrices, thumbs-up count-based approach, and anomaly detection are widely used to identify emerging issues. Extracting meaningful information and evaluating the proposed approach are the most common challenges identified. Novel LLMs, like Chat-GPT, significantly enhance review analysis by automating the process, improving feature extraction, and enabling context-aware review classification. Recommendations for Practitioners: The combination of conventional approaches and novel LLM-based methods can enhance both the efficiency and accuracy in identifying and addressing critical issues raised through mobile app user reviews. It effectively prioritizes user concerns by leveraging the strengths of both traditional preprocessing techniques and advanced LLMs. Recommendation for Researchers: Researchers are encouraged to explore the integration of emerging technologies like LLMs to enhance the of app review analysis, particularly in feature-specific sentiment analysis. Impact on Society: The results of this study contribute to enhancing the mobile app user experience through effective app review analysis, which improves user satisfaction and supports user-centered app development. This ultimately leads to a better mobile app ecosystem, benefiting both users and developers. Future Research: In the future, this research can be extended in multiple directions. Researchers can address the existing research gaps that LLMs have yet to address, particularly in prioritizing user concerns. Additionally, there is potential for further research on tool implementations focusing on identifying persistent issues through time series analysis by considering the app version and date of the app reviews. Moreover, there is a need to develop comprehensive frameworks that are more generalizable across different apps and categories, with a focus on identifying user concerns related to specific features.

Mental health diagnoses possess unique challenges that often lead to nuanced difficulties in managing an individual’s well-being and daily functioning. Self-report questionnaires are a common practice in clinical settings to help mitigate the challenges involved in mental health disorder screening. However, these questionnaires rely on an individual’s subjective response which can be influenced by various factors. Despite the advancements of Large Language Models (LLMs), quantifying self-reported experiences with natural language processing has resulted in imperfect accuracy. This project aims to demonstrate the effectiveness of zero-shot learning LLMs for screening and assessing item scales for depression using LLMs. The DAIC-WOZ is a publicly available mental health dataset that contains textual data from clinical interviews and self-report questionnaires with relevant mental health disorder labels. The RISEN prompt engineering framework was utilized to evaluate LLMs’ effectiveness in predicting depression symptoms based on individual PHQ-8 items. Various LLMs, including GPT models, Llama3_8B, Cohere, and Gemini were assessed based on performance. The GPT models, especially GPT-4o, were consistently better than other LLMs (Llama3_8B, Cohere, Gemini) across all eight items of the PHQ-8 scale in accuracy (M = 75.9%), and F1 score (0.74). GPT models were able to predict PHQ-8 items related to emotional and cognitive states. Llama 3_8B demonstrated superior detection of anhedonia-related symptoms and the Cohere LLM’s strength was identifying and predicting psychomotor activity symptoms. This study provides a novel outlook on the potential of LLMs for predicting self-reported questionnaire scores from textual interview data. The promising preliminary performance of the various models indicates there is potential that these models could effectively assist in the screening of depression. Further research is needed to establish a framework for which LLM can be used for specific mental health symptoms and other disorders. As well, analysis of additional datasets while fine-tuning models should be explored.

Abstract—This paper presents a hybrid sentiment analysis system that integrates lexicon-based, traditional machine learning, and deep learning techniques to classify textual data into positive, negative, or neutral sentiments. The system leverages Python-based libraries such as Scikit-learn, NLTK, and Transformers to pre- process text, extract features, and apply models in- cluding Support Vector Machine (SVM), Bidirectional Encoder Representations from Transformers (BERT), Linear Regression, and Valence Aware Dictionary and sEntiment Reasoner (VADER). The framework pro- cesses customer reviews from e-commerce platforms (Amazon, Flipkart) and social media (Instagram) using web scraping and provides actionable insights through sentiment summarization and visualizations (bar and pie charts). Experimental results demonstrate BERT’s superior performance with 92.3% accuracy, followed by SVM (85.6%), Linear Regression (81.2%), and VADER (76.8%). The system addresses challenges like sarcasm, class imbalance, and scalability, offering a scalable, user-friendly solution for real-world applications in e- commerce, social media analytics, and brand reputation management. Index Terms—Sentiment Analysis, Machine Learning, Deep Learning, Natural Language Processing, BERT, SVM, VADER, Text Classification, Data Visualization.

BackgroundMultimodal large language models (LLMs) have the capability to process and integrate both text and image data, offering promising applications in the medical field. This study aimed to evaluate the performance of representative multimodal LLMs in the 2023 Japanese Surgical Specialist Examination, with a focus on image-based questions across various surgical subspecialties. MethodsA total of 98 examination questions, including 43 image-based questions, from the 2023 Japanese Surgical Specialist Examination were administered to three multimodal LLMs: GPT-4 Omni, Claude 3.5 Sonnet, and Gemini Pro 1.5. Each model’s performance was assessed under two conditions: with and without images. Statistical analysis was conducted using McNemar’s test to evaluate the significance of accuracy differences between the two conditions. ResultsAmong the three LLMs, Claude 3.5 Sonnet achieved the highest overall accuracy at 84.69%, exceeding the passing threshold of 80%, which is consistent with the standard set by the Japan Surgical Society for board certification. GPT-4 Omni closely approached the threshold with an accuracy of 79.59%, while Gemini Pro 1.5 scored 61.22%. Claude 3.5 Sonnet demonstrated the highest accuracy in four of six subspecialties for image-based questions and was the only model to show a statistically significant improvement with image inclusion (76.74% with images vs. 62.79% without images, p = 0.041). By contrast, GPT-4 Omni and Gemini Pro 1.5 did not exhibit significant performance changes with image inclusion.ConclusionClaude 3.5 Sonnet outperformed the other models in most surgical subspecialties for image-based questions and was the only model to benefit significantly from image inclusion. These findings suggest that multimodal LLMs, particularly Claude 3.5 Sonnet, hold promise as diagnostic and educational support tools in surgical domains, and that variation in visual reasoning capabilities may account for model-level differences in image-based performance.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12909-025-07938-6.

Sentiment analysis, in the context of digital banking reviews, aims to assess customer satisfaction and support service enhancement. Despite increasing attention to sentiment analysis across domains, Arabic banking reviews remain underexplored. To bridge this gap, we introduce a dataset of 4922 Arabic reviews from three major Saudi digital banks with three sentiment categories positive, negative, or conflict—providing actionable insights for banks. We evaluate the dataset using several machine learning models and four large language models (LLMs)—GPT 3.5, GPT 4, Llama-3-8B-Instruct, and SILMA—using zero-shot (no labeled examples) and few-shot (a few labeled examples) learning strategies. Our results show that GPT 4 performs best among LLMs in few-shot settings, while traditional models still outperform LLMs, with a Voting Classifier achieving 90.24% accuracy. This study contributes a domain-specific dataset and comparative analysis to support research and practical improvements in Arabic digital banking services.