This study investigates the self-rationalization framework constructed with a cooperative game, where a generator initially extracts the most informative segment from raw input, and a subsequent predictor utilizes the selected subset for its input. The generator and predictor are trained collaboratively to maximize prediction accuracy. In this paper, we first uncover a potential caveat: such a cooperative game could unintentionally introduce a sampling bias during rationale extraction. Specifically, the generator might inadvertently create an incorrect correlation between the selected rationale candidate and the label, even when they are semantically unrelated in the original dataset. Subsequently, we elucidate the origins of this bias using both detailed theoretical analysis and empirical evidence. Our findings suggest a direction for inspecting these correlations through attacks, based on which we further introduce an instruction to prevent the predictor from learning the correlations. Through experiments on six text classification datasets and two graph classification datasets using three network architectures (GRUs, BERT, and GCN), we show that our method not only significantly outperforms recent rationalization methods, but also achieves comparable or even better results than a representative LLM (llama3.1-8b-instruct).
With the advance of large language models (LLMs), LLMs have been utilized for the various tasks. However, the issues of variability and reproducibility of results from each trial of LLMs have been largely overlooked in existing literature while actual human annotation uses majority voting to resolve disagreements among annotators. Therefore, this study introduces the straightforward ensemble strategy to a sentiment analysis using LLMs. As the results, we demonstrate that the ensemble of multiple inference using medium-sized LLMs produces more robust and accurate results than using a large model with a single attempt with reducing RMSE by 18.6%.
Large language models (LLMs) have revolutionized natural language processing by solving a wide range of tasks simply guided by a prompt. Yet their performance is highly sensitive to prompt formulation. While automated prompt optimization addresses this challenge by finding optimal prompts, current methods require a substantial number of LLM calls and input tokens, making prompt optimization expensive. We introduce CAPO (Cost-Aware Prompt Optimization), an algorithm that enhances prompt optimization efficiency by integrating AutoML techniques. CAPO is an evolutionary approach with LLMs as operators, incorporating racing to save evaluations and multi-objective optimization to balance performance with prompt length. It jointly optimizes instructions and few-shot examples while leveraging task descriptions for improved robustness. Our extensive experiments across diverse datasets and LLMs demonstrate that CAPO outperforms state-of-the-art discrete prompt optimization methods in 11/15 cases with improvements up to 21%p. Our algorithm achieves better performances already with smaller budgets, saves evaluations through racing, and decreases average prompt length via a length penalty, making it both cost-efficient and cost-aware. Even without few-shot examples, CAPO outperforms its competitors and generally remains robust to initial prompts. CAPO represents an important step toward making prompt optimization more powerful and accessible by improving cost-efficiency.
In-context learning (ICL) enables Large Vision-Language Models (LVLMs) to adapt to new tasks without parameter updates, using a few demonstrations from a large support set. However, selecting informative demonstrations leads to high computational and memory costs. While some methods explore selecting a small and representative coreset in the text classification, evaluating all support set samples remains costly, and discarded samples lead to unnecessary information loss. These methods may also be less effective for image classification due to differences in feature spaces. Given these limitations, we propose Key-based Coreset Optimization (KeCO), a novel framework that leverages untapped data to construct a compact and informative coreset. We introduce visual features as keys within the coreset, which serve as the anchor for identifying samples to be updated through different selection strategies. By leveraging untapped samples from the support set, we update the keys of selected coreset samples, enabling the randomly initialized coreset to evolve into a more informative coreset under low computational cost. Through extensive experiments on coarse-grained and fine-grained image classification benchmarks, we demonstrate that KeCO effectively enhances ICL performance for image classification task, achieving an average improvement of more than 20\%. Notably, we evaluate KeCO under a simulated online scenario, and the strong performance in this scenario highlights the practical value of our framework for resource-constrained real-world scenarios.
Knowledge distillation (KD) is a technique for transferring knowledge from complex teacher models to simpler student models, significantly enhancing model efficiency and accuracy. It has demonstrated substantial advancements in various applications including image classification, object detection, language modeling, text classification, and sentiment analysis. Recent innovations in KD methods, such as attention-based approaches, block-wise logit distillation, and decoupling distillation, have notably improved student model performance. These techniques focus on stimulus complexity, attention mechanisms, and global information capture to optimize knowledge transfer. In addition, KD has proven effective in compressing large language models while preserving accuracy, reducing computational overhead, and improving inference speed. This survey synthesizes the latest literature, highlighting key findings, contributions, and future directions in knowledge distillation to provide insights for researchers and practitioners on its evolving role in artificial intelligence and machine learning.
Telecom services are at the core of today's societies' everyday needs. The availability of numerous online forums and discussion platforms enables telecom providers to improve their services by exploring the views of their customers to learn about common issues that the customers face. Natural Language Processing (NLP) tools can be used to process the free text collected. One way of working with such data is to represent text as numerical vectors using one of many word embedding models based on neural networks. This research uses a novel dataset of telecom customers' reviews to perform an extensive study showing how different word embedding algorithms can affect the text classification process. Several state-of-the-art word embedding techniques are considered, including BERT, Word2Vec and Doc2Vec, coupled with several classification algorithms. The important issue of feature engineering and dimensionality reduction is addressed and several PCA-based approaches are explored. Moreover, the energy consumption used by the different word embeddings is investigated. The findings show that some word embedding models can lead to consistently better text classifiers in terms of precision, recall and F1-Score. In particular, for the more challenging classification tasks, BERT combined with PCA stood out with the highest performance metrics. Moreover, our proposed PCA approach of combining word vectors using the first principal component shows clear advantages in performance over the traditional approach of taking the average.
With the introduction of the PSD2 regulation in the EU which established the Open Banking framework, a new window of opportunities has opened for banks and fintechs to explore and enrich Bank transaction descriptions with the aim of building a better understanding of customer behavior, while using this understanding to prevent fraud, reduce risks and offer more competitive and tailored services. And although the usage of natural language processing models and techniques has seen an incredible progress in various applications and domains over the past few years, custom applications based on domain-specific text corpus remain unaddressed especially in the banking sector. In this paper, we introduce a language-based Open Banking transaction classification system with a focus on the french market and french language text. The system encompasses data collection, labeling, preprocessing, modeling, and evaluation stages. Unlike previous studies that focus on general classification approaches, this system is specifically tailored to address the challenges posed by training a language model with a specialized text corpus (Banking data in the French context). By incorporating language-specific techniques and domain knowledge, the proposed system demonstrates enhanced performance and efficiency compared to generic approaches.
Federated Learning (FL) faces major challenges regarding communication overhead and model privacy when training large language models (LLMs), especially in healthcare applications. To address these, we introduce Selective Attention Federated Learning (SAFL), a novel approach that dynamically fine-tunes only those transformer layers identified as attention-critical. By employing attention patterns to determine layer importance, SAFL significantly reduces communication bandwidth and enhances differential privacy resilience. Evaluations on clinical NLP benchmarks (i2b2 Clinical Concept Extraction and MIMIC-III discharge summaries) demonstrate that SAFL achieves competitive performance with centralized models while substantially improving communication efficiency and privacy preservation.
Traditionally, authorship attribution (AA) tasks relied on statistical data analysis and classification based on stylistic features extracted from texts. In recent years, pre-trained language models (PLMs) have attracted significant attention in text classification tasks. However, although they demonstrate excellent performance on large-scale short-text datasets, their effectiveness remains under-explored for small samples, particularly in AA tasks. Additionally, a key challenge is how to effectively leverage PLMs in conjunction with traditional feature-based methods to advance AA research. In this study, we aimed to significantly improve performance using an integrated integrative ensemble of traditional feature-based and modern PLM-based methods on an AA task in a small sample. For the experiment, we used two corpora of literary works to classify 10 authors each. The results indicate that BERT is effective, even for small-sample AA tasks. Both BERT-based and classifier ensembles outperformed their respective stand-alone models, and the integrated ensemble approach further improved the scores significantly. For the corpus that was not included in the pre-training data, the integrated ensemble improved the F1 score by approximately 14 points, compared to the best-performing single model. Our methodology provides a viable solution for the efficient use of the ever-expanding array of data processing tools in the foreseeable future.
Language models based on the Transformer architecture achieve excellent results in many language-related tasks, such as text classification or sentiment analysis. However, despite the architecture of these models being well-defined, little is known about how their internal computations help them achieve their results. This renders these models, as of today, a type of 'black box' systems. There is, however, a line of research -- 'interpretability' -- aiming to learn how information is encoded inside these models. More specifically, there is work dedicated to studying whether Transformer-based models possess knowledge of linguistic phenomena similar to human speakers -- an area we call 'linguistic interpretability' of these models. In this survey we present a comprehensive analysis of 160 research works, spread across multiple languages and models -- including multilingual ones -- that attempt to discover linguistic information from the perspective of several traditional Linguistics disciplines: Syntax, Morphology, Lexico-Semantics and Discourse. Our survey fills a gap in the existing interpretability literature, which either not focus on linguistic knowledge in these models or present some limitations -- e.g. only studying English-based models. Our survey also focuses on Pre-trained Language Models not further specialized for a downstream task, with an emphasis on works that use interpretability techniques that explore models' internal representations.
For a long time, the authorship of the Federalist Papers had been a subject of inquiry and debate, not only by linguists and historians but also by statisticians. In what was arguably the first Bayesian case study, Mosteller and Wallace (1963) provided the first statistical evidence for attributing all disputed papers to Madison. Our paper revisits this historical dataset but from a lens of modern language models, both small and large. We review some of the more popular Large Language Model (LLM) tools and examine them from a statistical point of view in the context of text classification. We investigate whether, without any attempt to fine-tune, the general embedding constructs can be useful for stylometry and attribution. We explain differences between various word/phrase embeddings and discuss how to aggregate them in a document. Contrary to our expectations, we exemplify that dimension expansion with word embeddings may not always be beneficial for attribution relative to dimension reduction with topic embeddings. Our experiments demonstrate that default LLM embeddings (even after manual fine-tuning) may not consistently improve authorship attribution accuracy. Instead, Bayesian analysis with topic embeddings trained on ``function words" yields superior out-of-sample classification performance. This suggests that traditional (small) statistical language models, with their interpretability and solid theoretical foundation, can offer significant advantages in authorship attribution tasks. The code used in this analysis is available at github.com/sowonjeong/slm-to-llm
This study presents a multi-stage approach to mental health classification by leveraging traditional machine learning algorithms, deep learning architectures, and transformer-based models. A novel data set was curated and utilized to evaluate the performance of various methods, starting with conventional classifiers and advancing through neural networks. To broaden the architectural scope, recurrent neural networks (RNNs) such as LSTM and GRU were also evaluated to explore their effectiveness in modeling sequential patterns in the data. Subsequently, transformer models such as BERT were fine-tuned to assess the impact of contextual embeddings in this domain. Beyond these baseline evaluations, the core contribution of this study lies in a novel training strategy involving a dual-model architecture composed of a teacher and a student network. Unlike standard distillation techniques, this method does not rely on soft label transfer; instead, it facilitates information flow through both the teacher model's output and its latent representations by modifying the loss function. The experimental results highlight the effectiveness of each modeling stage and demonstrate that the proposed loss function and teacher-student interaction significantly enhance the model's learning capacity in mental health prediction tasks.
Cardiovascular disease remains one of the leading causes of mortality worldwide, underscoring the need for accurate as well as interpretable diagnostic machine learning tools. In this work, we investigate heart disease classification using electrocardiogram (ECG) data from two widely-utilized datasets: The MIT-BIH Arrhythmia and the PTB-XL datasets. We adapt a hierarchical attention network (HAN), originally developed for text classification, into an ECG-based heart-disease classification task. Our adapted HAN incorporates two attention layers that focus on ECG data segments of varying sizes. We conduct a comparative analysis between our adapted HAN and a more sophisticated state-of-the-art architecture, featuring a network with convolution, attention, and transformer layers (CAT-Net). Our empirical evaluation encompasses multiple aspects including test accuracy (quantified by 0-1 loss); model complexity (measured by the number of model parameters); and interpretability (through attention map visualization). Our adapted HAN demonstrates comparable test accuracy with significant reductions in model complexity and enhanced interpretability analysis: For the MIT-BIH dataset, our adapted HAN achieves 98.55\% test accuracy compared to 99.14\% for CAT-Net, while reducing the number of model parameters by a factor of 15.6. For the PTB-XL dataset, our adapted HAN achieves a 19.3-fold reduction in model complexity compared to CAT-Net, with only a 5\% lower test accuracy. From an interpretability perspective, the significantly simpler architecture and the hierarchical nature of our adapted HAN model facilitate a more straightforward interpretability analysis based on visualizing attention weights. Building on this advantage, we conduct an interpretability analysis of our HAN that highlights the regions of the ECG signal most relevant to the model's decisions.
In this paper, we introduce Traversal Learning (TL), a novel approach designed to address the problem of decreased quality encountered in popular distributed learning (DL) paradigms such as Federated Learning (FL), Split Learning (SL), and SplitFed Learning (SFL). Traditional FL experiences from an accuracy drop during aggregation due to its averaging function, while SL and SFL face increased loss due to the independent gradient updates on each split network. TL adopts a unique strategy where the model traverses the nodes during forward propagation (FP) and performs backward propagation (BP) on the orchestrator, effectively implementing centralized learning (CL) principles within a distributed environment. The orchestrator is tasked with generating virtual batches and planning the sequential node visits of the model during FP, aligning them with the ordered index of the data within these batches. We conducted experiments on six datasets representing diverse characteristics across various domains. Our evaluation demonstrates that TL is on par with classic CL approaches in terms of accurate inference, thereby offering a viable and robust solution for DL tasks. TL outperformed other DL methods and improved accuracy by 7.85% for independent and identically distributed (IID) datasets, macro F1-score by 1.06% for non-IID datasets, accuracy by 2.60% for text classification, and AUC by 3.88% and 4.54% for medical and financial datasets, respectively. By effectively preserving data privacy while maintaining performance, TL represents a significant advancement in DL methodologies.
Natural language processing models often face challenges due to limited labeled data, especially in domain specific areas, e.g., clinical trials. To overcome this, text augmentation techniques are commonly used to increases sample size by transforming the original input data into artificial ones with the label preserved. However, traditional text classification methods ignores the relationship between augmented texts and treats them as independent samples which may introduce classification error. Therefore, we propose a novel approach called 'Batch Aggregation' (BAGG) which explicitly models the dependence of text inputs generated through augmentation by incorporating an additional layer that aggregates results from correlated texts. Through studying multiple benchmark data sets across different domains, we found that BAGG can improve classification accuracy. We also found that the increase of performance with BAGG is more obvious in domain specific data sets, with accuracy improvements of up to 10-29%. Through the analysis of benchmark data, the proposed method addresses limitations of traditional techniques and improves robustness in text classification tasks. Our result demonstrates that BAGG offers more robust results and outperforms traditional approaches when training data is limited.
Zero-shot text classification typically relies on prompt engineering, but the inherent prompt brittleness of large language models undermines its reliability. Minor changes in prompt can cause significant discrepancies in model performance. We attribute this prompt brittleness largely to the narrow focus on nexttoken probabilities in existing methods. To address this, we propose Placeholding Parallel Prediction (P3), a novel approach that predicts token probabilities across multiple positions and simulates comprehensive sampling of generation paths in a single run of a language model. Experiments show improved accuracy and up to 98% reduction in the standard deviation across prompts, boosting robustness. Even without a prompt, P3 maintains comparable performance, reducing the need for prompt engineering.
Automatic text classification (ATC) has experienced remarkable advancements in the past decade, best exemplified by recent small and large language models (SLMs and LLMs), leveraged by Transformer architectures. Despite recent effectiveness improvements, a comprehensive cost-benefit analysis investigating whether the effectiveness gains of these recent approaches compensate their much higher costs when compared to more traditional text classification approaches such as SVMs and Logistic Regression is still missing in the literature. In this context, this work's main contributions are twofold: (i) we provide a scientifically sound comparative analysis of the cost-benefit of twelve traditional and recent ATC solutions including five open LLMs, and (ii) a large benchmark comprising {22 datasets}, including sentiment analysis and topic classification, with their (train-validation-test) partitions based on folded cross-validation procedures, along with documentation, and code. The release of code, data, and documentation enables the community to replicate experiments and advance the field in a more scientifically sound manner. Our comparative experimental results indicate that LLMs outperform traditional approaches (up to 26%-7.1% on average) and SLMs (up to 4.9%-1.9% on average) in terms of effectiveness. However, LLMs incur significantly higher computational costs due to fine-tuning, being, on average 590x and 8.5x slower than traditional methods and SLMs, respectively. Results suggests the following recommendations: (1) LLMs for applications that require the best possible effectiveness and can afford the costs; (2) traditional methods such as Logistic Regression and SVM for resource-limited applications or those that cannot afford the cost of tuning large LLMs; and (3) SLMs like Roberta for near-optimal effectiveness-efficiency trade-off.
The Tsetlin Machine (TM) is a propositional logic based model that uses conjunctive clauses to learn patterns from data. As with typical neural networks, the performance of a Tsetlin Machine is largely dependent on its parameter count, with a larger number of parameters producing higher accuracy but slower execution. Knowledge distillation in neural networks transfers information from an already-trained teacher model to a smaller student model to increase accuracy in the student without increasing execution time. We propose a novel approach to implementing knowledge distillation in Tsetlin Machines by utilizing the probability distributions of each output sample in the teacher to provide additional context to the student. Additionally, we propose a novel clause-transfer algorithm that weighs the importance of each clause in the teacher and initializes the student with only the most essential data. We find that our algorithm can significantly improve performance in the student model without negatively impacting latency in the tested domains of image recognition and text classification.
In Biomedical Natural Language Processing (BioNLP) tasks, such as Relation Extraction, Named Entity Recognition, and Text Classification, the scarcity of high-quality data remains a significant challenge. This limitation poisons large language models to correctly understand relationships between biological entities, such as molecules and diseases, or drug interactions, and further results in potential misinterpretation of biomedical documents. To address this issue, current approaches generally adopt the Synthetic Data Augmentation method which involves similarity computation followed by word replacement, but counterfactual data are usually generated. As a result, these methods disrupt meaningful word sets or produce sentences with meanings that deviate substantially from the original context, rendering them ineffective in improving model performance. To this end, this paper proposes a biomedical-dedicated rationale-based synthetic data augmentation method. Beyond the naive lexicon similarity, specific bio-relation similarity is measured to hold the augmented instance having a strong correlation with bio-relation instead of simply increasing the diversity of augmented data. Moreover, a multi-agents-involved reflection mechanism helps the model iteratively distinguish different usage of similar entities to escape falling into the mis-replace trap. We evaluate our method on the BLURB and BigBIO benchmark, which includes 9 common datasets spanning four major BioNLP tasks. Our experimental results demonstrate consistent performance improvements across all tasks, highlighting the effectiveness of our approach in addressing the challenges associated with data scarcity and enhancing the overall performance of biomedical NLP models.
We investigate the effectiveness of fine-tuning large language models (LLMs) on small medical datasets for text classification and named entity recognition tasks. Using a German cardiology report dataset and the i2b2 Smoking Challenge dataset, we demonstrate that fine-tuning small LLMs locally on limited training data can improve performance achieving comparable results to larger models. Our experiments show that fine-tuning improves performance on both tasks, with notable gains observed with as few as 200-300 training examples. Overall, the study highlights the potential of task-specific fine-tuning of LLMs for automating clinical workflows and efficiently extracting structured data from unstructured medical text.