VisBERT: Hidden-State Visualizations for Transformers

With the development of Internet technology and the transformation of news media to informatization, news images, texts, and other information have exploded. The news image and text classification can effectively solve the disorder problem of news information. The early news image text classification is to establish artificial classifiers according to specific classification rules, but the classification error rate is high and the classification speed is slow. Later, machine learning technology replaced manual classification of news image texts. Although the classification efficiency has been greatly improved, the classification time is still a bit long. The bidirectional encoder representations from transformers (BERT) model uses transformer and encoder to pretrain news image text to improve classification efficiency. By comparing the differences between machine learning and BERT models in news image text classification, the experiments showed that the average precision, recall, and F1 values of the news image text classification algorithm using the BERT model were 96.6%, 95.7%, and 96.1%, respectively. All three evaluation criteria were about 5% more than the classification algorithm of the machine learning model. The classification speed of the news image text classification algorithm using the BERT model was 1.8 times that of the news image text classification algorithm based on support vector machine. Therefore, the news image text classification algorithm using the BERT model can improve the classification accuracy and efficiency of news image text.

The article examines the use of neural networks in telemedicine for patient self-diagnosis, focusing on custom LSTM (Long Short-Term Memory) models and pre-trained BERT (Bidirectional Encoder Representations from Transformers) models. Telemedicine has become essential in modern healthcare, particularly due to the rise in remote consultations and diagnostics during global health crises. The study reviews research in this area, highlighting contributions from international and Ukrainian experts. It explores the effectiveness of neural networks in interpreting patient symptom descriptions and generating accurate diagnoses. The custom LSTM model processes text-based symptom descriptions but faces challenges with limited datasets, impacting its diagnostic accuracy. On the other hand, the BERT model leverages transfer learning, achieving over 92% accuracy in predictions, thanks to its ability to understand complex language inputs. An enhanced LSTM model is also presented, incorporating Dropout and Batch Normalization layers, which help increase the robustness and accuracy of the predictions while managing the model's complexity. The article also discusses the technical challenges of training these neural networks, including the variability of symptom descriptions and the need for large annotated medical datasets. Proposed solutions include domain-specific enhancements and data augmentation to improve the models' performance. In comparing the three models, the study identifies that while the custom LSTM model is more flexible and simple, it struggles with generalization. The BERT model, by contrast, excels in understanding and accuracy but requires more computational resources. The enhanced LSTM model offers a balance between complexity and performance. In conclusion, neural networks offer promising advances in telemedicine by providing accurate, AI-assisted self-diagnosis. Although challenges remain, such as data requirements and computing power, models like BERT are key to improving healthcare accessibility and accuracy in telemedicine applications. Among the main advantages of using artificial intelligence for the healthcare system are in-depth diagnostic analysis, individual treatment regimens, and optimization of medical processes. However, modern technologies in the medical field are not yet able to completely replace experienced specialists.

Large language models (LLMs) have emerged as a milestone in the realm of machine learning, offering groundbreaking capabilities in understanding human language with remarkable coherence and versatility. Their application spans a multitude of domains, ranging from text classification to information extraction. Such models are first pre-trained on vast data corpora to learn general language patterns and then are fine-tuned on specific downstream tasks. Despite their impressive capabilities, LLMs encounter several limitations, including sub-optimal performance in understanding the domain-specific language of drug labels, limited performance in analyzing class-imbalanced data, and high computational complexity of training. Human prescription drug labeling contains a summary of the essential scientific information needed for the safe and effective use of the drug and includes critical information about drug products, such as pharmacokinetics and adverse events. Automatic information extraction from drug labels can facilitate processes such as finding the adverse reaction of the drugs or finding the interaction of one drug with another drug. Although LLMs are an interesting solution to perform this task, because of their training on generic data, they may not possess the depth of knowledge or the ability required to understand the language in such a specialized field. To address this issue, this work presents the developed PharmBERT, which is a BERT model specifically pretrained on the drug labels, publicly available online for the NLP community. We demonstrate that our model outperforms models such as vanilla BERT, ClinicalBERT, and BioBERT in multiple NLP tasks in the drug label domain. Another shortcoming of LLMs lies in their handling of class-imbalanced data, leading to a subpar performance in identifying and accurately responding to minority classes. This class-imbalance can skew the model's learning process, favoring majority classes, and thus diminishing its effectiveness in recognizing and appropriately addressing less represented categories. To cope with these challenges, we present a simple modification of standard fine-tuning. Specifically, we propose a two-stage fine-tuning. In Stage 1, we fine-tune the final layer of the pretrained model with class-balanced augmented data, generated using ChatGPT. As a large generative language model, ChatGPT is capable of generating novel yet contextually similar responses to a given text sample, which makes it an excellent candidate for data augmentation. In Stage 2, we perform the standard fine-tuning. Our modification has several benefits: (1) it leverages pretrained representations by only fine-tuning a small portion of the model parameters while keeping the rest untouched; (2) it allows the model to learn an initial representation of the specific task; and importantly (3) it protects the learning of tail classes from being at a disadvantage during the model updating. The experimental results show that the proposed two-stage fine-tuning outperforms vanilla fine-tuning and state-of-the-art methods on the above datasets. The third issue of LLMs is the high computational complexity of training. Due to the large number of parameters in many state-of-the-art LLMs, such as Bidirectional Encoder Representations from Transformers (BERT), the process of fine-tuning exhibits a high level of intricacy. One attractive solution to this issue is parameter-efficient fine-tuning, which means only fine-tuning a small portion of the model and not changing the rest. However, the question that which portion of the BERT model is the most important one for fine-tuning remains unanswered. Here, we first analyze different components in the BERT model to see which one undergoes the most change after fine-tuning. We show that output LayerNorm changes more than other components when finetuned for different General Language Understanding Evaluation (GLUE) tasks. Then we show that only fine-tuning this component reaches a comparable, or in some cases better, performance to full fine-tuning and other parameter-efficient fine-tuning methods. Moreover, we use Fisher information to find out which parameters in LayerNorm are the most important ones and show that with a small performance degradation, many NLP tasks in the GLUE benchmark can be solved by fine-tuning only a small, selected portion of LayerNorm. In summary, LLMs hold immense potential to redefine and enhance various fields, including drug-label analysis, but they also suffer from limitations like sub-optimal performance in understanding domain-specific texts of drug labels, not learning minority classes in the class-imbalanced data, and high computational complexity. In this thesis, for each one of these shortcomings, first, the problem is analyzed and then a solution is provided.

BackgroundDeaths related to physical trauma impose a heavy burden on society, and the Abbreviated Injury Scale (AIS) is an important tool for injury research. AIS covers injuries to various parts of the human body and scores them based on the severity of the injury. In practical applications, the complex AIS coding rules require experts to encode by consulting patient medical records, which inevitably increases the difficulty, time, and cost of evaluation of patient and also puts higher demands on the workload of information collection and processing. In some cases, the sheer number of patients or the inability to access detailed medical records necessary for coding further complicates independent AIS codes.ObjectiveThis study aims to use advanced deep learning techniques to predict AIS codes based on easily accessible diagnostic information of patients to improve the accuracy of trauma assessment.MethodsWe used a dataset of patients with trauma (n=26,810) collected by the Chongqing Daping Hospital between October 2013 and June 2024. We mainly selected the patient’s diagnostic information, injury description, cause of injury, injury region, injury types, and present illness history as the key feature inputs. We used a robust optimization Bidirectional Encoder Representations from Transformers (BERT) pretraining method to embed these features and constructed a prediction model based on BERT. This model aims to predict AIS codes and comprehensively evaluate its performance through a 5-fold cross-validation. We compared the BERT model with previous research results and current mainstream machine learning methods to verify its advantages in prediction tasks. In addition, we also conducted external validation of the model using 244 external data points from the Chongqing Emergency Center.ResultsThe BERT model proposed in this paper performs significantly better than the comparison model on independent test datasets with an accuracy of 0.8971, which surpassed the previous study by 10 % points. In addition, the area under the curve (AUC value of the BERT model is 0.9970, and the F1-score is 0.8434. In the external dataset, the accuracy, AUC, and F1-score results of the model are 0.7131, 0.8586, and 0.6801, respectively. These results indicate that our model has high generalization ability and prediction accuracy.ConclusionsThe BERT model we proposed is mainly based on diagnostic information to predict AIS codes, and its prediction accuracy is superior to previous investigations and current mainstream machine learning methods. It has a high generalization ability in external datasets.

Bidirectional Encoder Representations from Transformers (BERT) has gained increasing attention from researchers and practitioners as it has proven to be an invaluable technique in natural languages processing. This is mainly due to its unique features, including its ability to predict words conditioned on both the left and the right context, and its ability to be pretrained using the plain text corpus that is enormously available on the web. As BERT gained more interest, more BERT models were introduced to support different languages, including Arabic. The current state of knowledge and practice in applying BERT models to Arabic text classification is limited. In an attempt to begin remedying this gap, this review synthesizes the different Arabic BERT models that have been applied to text classification. It investigates the differences between them and compares their performance. It also examines how effective they are compared to the original English BERT models. It concludes by offering insight into aspects that need further improvements and future work.

Sentence Boundary Disambiguation (SBD) is crucial for building datasets for tasks such as machine translation, syntactic analysis, and semantic analysis. Currently, most automatic sentence segmentation in Tibetan adopts the methods of rule-based and statistical learning, as well as the combination of the two, which have high requirements on the corpus and the linguistic foundation of the researchers and are more costly to annotate manually. In this study, we explore Tibetan SBD using deep learning technology. Initially, we analyze Tibetan characteristics and various subword techniques, selecting Byte Pair Encoding (BPE) and Sentencepiece (SP) for text segmentation and training the Bidirectional Encoder Representations from Transformers (BERT) pre-trained language model. Secondly, we studied the Tibetan SBD based on different BERT pre-trained language models, which mainly learns the ambiguity of the shad (“།”) in different positions in modern Tibetan texts and determines through the model whether the shad (“།”) in the texts has the function of segmenting sentences. Meanwhile, this study introduces four models, BERT-CNN, BERT-RNN, BERT-RCNN, and BERT-DPCNN, based on the BERT model for performance comparison. Finally, to verify the performance of the pre-trained language models on the SBD task, this study conducts SBD experiments on both the publicly available Tibetan pre-trained language model TiBERT and the multilingual pre-trained language model (Multi-BERT). The experimental results show that the F1 score of the BERT (BPE) model trained in this study reaches 95.32% on 465,669 Tibetan sentences, nearly five percentage points higher than BERT (SP) and Multi-BERT. The SBD method based on pre-trained language models in this study lays the foundation for establishing datasets for the later tasks of Tibetan pre-training, summary extraction, and machine translation.

BackgroundThe accumulation of vast electronic medical records (EMRs) through medical informatization creates significant research value, particularly in obstetrics. Diagnostic standardization across different health care institutions and regions is vital for medical data analysis. Large language models (LLMs) have been extensively used for various medical tasks. Prompt engineering is key to use LLMs effectively.ObjectiveThis study aims to evaluate and compare the performance of LLMs with various prompt engineering techniques on the task of standardizing obstetric diagnostic terminology using real-world obstetric data.MethodsThe paper describes a 4-step approach used for mapping diagnoses in electronic medical records to the International Classification of Diseases, 10th revision, observation domain. First, similarity measures were used for mapping the diagnoses. Second, candidate mapping terms were collected based on similarity scores above a threshold, to be used as the training data set. For generating optimal mapping terms, we used two LLMs (ChatGLM2 and Qwen-14B-Chat [QWEN]) for zero-shot learning in step 3. Finally, a performance comparison was conducted by using 3 pretrained bidirectional encoder representations from transformers (BERTs), including BERT, whole word masking BERT, and momentum contrastive learning with BERT (MC-BERT), for unsupervised optimal mapping term generation in the fourth step.ResultsLLMs and BERT demonstrated comparable performance at their respective optimal levels. LLMs showed clear advantages in terms of performance and efficiency in unsupervised settings. Interestingly, the performance of the LLMs varied significantly across different prompt engineering setups. For instance, when applying the self-consistency approach in QWEN, the F1-score improved by 5%, with precision increasing by 7.9%, outperforming the zero-shot method. Likewise, ChatGLM2 delivered similar rates of accurately generated responses. During the analysis, the BERT series served as a comparative model with comparable results. Among the 3 models, MC-BERT demonstrated the highest level of performance. However, the differences among the versions of BERT in this study were relatively insignificant.ConclusionsAfter applying LLMs to standardize diagnoses and designing 4 different prompts, we compared the results to those generated by the BERT model. Our findings indicate that QWEN prompts largely outperformed the other prompts, with precision comparable to that of the BERT model. These results demonstrate the potential of unsupervised approaches in improving the efficiency of aligning diagnostic terms in daily research and uncovering hidden information values in patient data.

Transformer-based architectures have significantly advanced Natural Language Processing (NLP), with Bidirectional Encoder Representations from Transformers (BERT) serving as a strong baseline for extractive Question Answering (QA). This study aims to evaluate the performance of Indonesian BERT models on extractive QA tasks and to identify the most effective model for low-resource language settings. This research employed a comparative experimental method using two Indonesian BERT variants: indobert-base- ncased (IndoLEM) and indobert-base-p1 (IndoNLU/IndoBenchmark). Both models were fine-tuned on an Indonesian version of SQuAD 2.0, automatically translated via the Google Translate API. Answer-span alignment errors caused by translation were corrected using fuzzy string matching. Evaluation was conducted under identical hyperparameter settings and training schemes, using Exact Match (EM) and F1-score as performance metrics. The results indicate that IndoLEM achieved superior performance, with better loss convergence and a higher F1-score (71.58) than IndoNLU (63.59), and the difference was statistically significant (p < 0.001). In conclusion, IndoLEM is a more effective baseline model for Indonesian extractive QA systems. The findings also demonstrate that the composition and scale of pre-trained corpora substantially influence model performance in low-resource language contexts and highlight the importance of transfer learning for advancing NLP in underrepresented languages.

Given the foundational role of system requirements in design projects, designers can benefit from classifying, comparing, and observing connections between requirements. Manually undertaking these processes, however, can be laborious and time-consuming. Previous studies have employed Bidirectional Encoder Representations from Transformers (BERT), a state-of-the-art natural language processing (NLP) deep neural network model, to automatically analyze written requirements. Yet, it remains unclear whether BERT can sufficiently capture the nuances that differentiate requirements between and within design documents. This work evaluates BERT’s performance on two requirement classification tasks (one inter- document and one intra-document) executed on a corpus of 1,303 requirements sourced from five system design projects. First, in the “parent document classification” task, a BERT model is fine-tuned to classify requirements according to their originating project. A separate BERT model is then fine-tuned on a “functional classification” task where each requirement is classified as either functional or nonfunctional. Our results also include a comparison with a baseline model, Word2Vec, and demonstrate that our model achieves higher classification accuracy. When evaluated on test sets, the former model receives a Matthews correlation coefficient (MCC) of 0.95, while the latter receives an MCC of 0.82, indicating BERT’s ability to reliably distinguish requirements. This work then explores the application of BERT’s representations, known as embeddings, to identify similar requirements and predict requirement change.

It is essential to understand how environmental factors and cultural backgrounds affect emotional responses to build Human-Computer Interaction (HCI) technologies. This study uses the BERT (Bidirectional Encoder Representations from Transformers) model to investigate the emotional landscape of interactions on the X platform across various cultural contexts. The aims of the study are as follows: (1) Gather and prepare X platform data from COVID-19 sources in India, Pakistan, Malaysia, the US, the UK, Australia, and Canada. (2) Use a BERT model to qualitatively analyze tweet sentiment. (3) Assess the accuracy of the model and look for sentiment trends in tweets from nations throughout the pandemic. The BERT model successfully classified sentiment, as seen by its 80.29% accuracy rate on test data. Sentiment research showed that positive sentiment was far more prevalent in the US, Canada, and Australia, indicating that these countries were better able to adjust to the COVID-19 situation. Stability was seen in the balanced sentiment distributions displayed by Pakistan, India, and the United Kingdom. Despite having fewer data points, Pakistan and Malaysia continued to have largely positive attitudes. This study provides the basis for a comparative analysis of emotional responses by taking contextual and cultural aspects into account.

Twitter has emerged as a critical platform for capturing public sentiment, offering a valuable source for sentiment analysis. This study presents a comparative evaluation of two BERT (Bidirectional Encoder Representations from Transformers) models—baseline and fine-tuned—targeted at analyzing Indonesian-language tweets. Employing the CRISP-DM framework, the methodology encompasses automated data crawling, comprehensive text pre-processing (including case folding, cleaning, tokenisation, normalisation, and data augmentation), and model development using the IndoBERT-base-p1 architecture. The experimental results reveal that the fine-tuned BERT model achieves significantly improved performance over the non-optimized model, with accuracy, precision, recall, and F1-score values reaching 91%, 0.91, 0.90, and 0.91, respectively. These findings indicate the fine-tuned model's superior ability to capture linguistic subtleties and contextual sentiment features within informal social media text. Furthermore, the model is deployed in a web-based application for real-time sentiment classification, demonstrating its practical applicability. This study underscores the effectiveness of Fine-tuning in enhancing BERT-based sentiment analysis for low-resource languages. It highlights its potential for informing decisionmaking in digital communication, marketing, and policy research contexts.

Investor attitude is an important factor in driving the behavior of the stock market, usually dwarfing conventional financial markets. Although previous study has investigated the mental side of investing, this has been concentrated mostly on the developed world and has had no real-time behavioral incorporation. This study fulfills this requirement by investigating the roles played by investor mentality, media reporting, and macroeconomic triggers in determining stock market volatility during bear markets with a special emphasis on Kathmandu. The goal is to measure the effect of sentiment on market performance based on both structured survey data and unstructured social media and news text. A hybrid methodological design is used, combining statistical modeling (Chi-Square test) and a machine learning-based sentiment classification approach utilizing a fine-tuned BERT (Bidirectional Encoder Representations from Transformers) model. Social media and financial news information are both drawn from a 2022–2025 Kaggle Stock Market Sentiment dataset, and NVivo provides thematic analysis of investor interviews. The following BERT model, coded through Python's Hugging Face Transformers package, resulted in a 9.2% improvement on accuracy compared to standard sentiment analysis models such as logistic regression and SVM, providing richer contextual knowledge of investor emotion. Empirical findings indicate a statistically significant relationship between sentiment and investment behavior (χ² = 23.250, p = 0.006), validating market change behavior. This study not only emphasizes sentiment as an important predictive variable but also proposes a scalable, real-time predictive framework for risk-sensitive investment strategy. The application of this model can improve the early identification of volatility and guide sentiment-based trading systems for investors, analysts, and policymakers in developing markets.

Semantic matching is one of the critical technologies for intelligent customer service. Since Bidirectional Encoder Representations from Transformers (BERT) is proposed, fine-tuning on a large-scale pre-training language model becomes a general method to implement text semantic matching. However, in practical application, the accuracy of the BERT model is limited by the quantity of pre-training corpus and proper nouns in the target domain. An enhancement method for knowledge based on domain dictionary to mask input is proposed to solve the problem. Firstly, for modul input, we use keyword matching to recognize and mask the word in domain. Secondly, using self-supervised learning to inject knowledge of the target domain into the BERT model. Thirdly, we fine-tune the BERT model with public datasets LCQMC and BQboost. Finally, we test the model’s performance with a financial company’s user data. The experimental results show that after using our method and BQboost, accuracy increases by 12.12% on average in practical applications.

Bidirectional Encoder Representations from Transformers (BERT) models achieve state-of-the-art performance on a number of Natural Language Processing tasks. However, their model size on disk often exceeds 1 GB and the process of fine-tuning them and using them to run inference consumes significant hardware resources and runtime. This makes them hard to deploy to production environments. This paper fine-tunes DistilBERT, a lightweight deep learning model, on medical text for the named entity recognition task of Protected Health Information (PHI) and medical concepts. This work provides a full assessment of the performance of DistilBERT in comparison with BERT models that were pre-trained on medical text. For Named Entity Recognition task of PHI, DistilBERT achieved almost the same results as medical versions of BERT in terms of F1 score at almost half the runtime and consuming approximately half the disk space. On the other hand, for the detection of medical concepts, DistilBERT's F1 score was lower by 4 points on average than medical BERT variants.

PurposeSystems equipped with natural language (NLP) processing can reduce missed radiological findings by physicians, but the annotation costs are burden in the development. This study aimed to compare the effects of active learning (AL) algorithms in NLP for estimating the significance of head computed tomography (CT) reports using bidirectional encoder representations from transformers (BERT).MethodsA total of 3728 head CT reports annotated with five categories of importance were used and UTH-BERT was adopted as the pre-trained BERT model. We assumed that 64% (2385 reports) of the data were initially in the unlabeled data pool (UDP), while the labeled data set (LD) used to train the model was empty. Twenty-five reports were repeatedly selected from the UDP and added to the LD, based on seven metrices: random sampling (RS: control), four uncertainty sampling (US) methods (least confidence (LC), margin sampling (MS), ratio of confidence (RC), and entropy sampling (ES)), and two distance-based sampling (DS) methods (cosine distance (CD) and Euclidian distance (ED)). The transition of accuracy of the model was evaluated using the test dataset.ResultsThe accuracy of the models with US was significantly higher than RS when reports in LD were < 1800, whereas DS methods were significantly lower than RS. Among the US methods, MS and RC were even better than the others. With the US methods, the required labeled data decreased by 15.4–40.5%, and most efficient in RC. In addition, in the US methods, data for minor categories tended to be added to LD earlier than RS and DS.ConclusionsIn the classification task for the importance of head CT reports, US methods, especially RC and MS can lead to the effective fine-tuning of BERT models and reduce the imbalance of categories. AL can contribute to other studies on larger datasets by providing effective annotation.