arXiv subjects

Jaewook Kang

Publications and source records attributed to Jaewook Kang.

16 recordsLinked to original sources

arXiv2024

DSAI: Unbiased and Interpretable Latent Feature Extraction for Data-Centric AI

Large language models (LLMs) often struggle to objectively identify latent characteristics in large datasets due to their reliance on pre-trained knowledge rather than actual data patterns. To address this data grounding issue, we propose Data Scientist AI (DSAI), a framework that enables unbiased and interpretable feature extraction through a multi-stage pipeline with quantifiable prominence metrics for evaluating extracted features. On synthetic datasets with known ground-truth features, DSAI demonstrates high recall in identifying expert-defined features while faithfully reflecting the underlying data. Applications on real-world datasets illustrate the framework's practical utility in uncovering meaningful patterns with minimal expert oversight, supporting use cases such as interpretable classification. The title of our paper is chosen from multiple candidates based on DSAI-generated criteria.

cs.LG↗

arXiv2024

Navigating the Path of Writing: Outline-guided Text Generation with Large Language Models

Large Language Models (LLMs) have impacted the writing process, enhancing productivity by collaborating with humans in content creation platforms. However, generating high-quality, user-aligned text to satisfy real-world content creation needs remains challenging. We propose WritingPath, a framework that uses explicit outlines to guide LLMs in generating goal-oriented, high-quality text. Our approach draws inspiration from structured writing planning and reasoning paths, focusing on reflecting user intentions throughout the writing process. To validate our approach in real-world scenarios, we construct a diverse dataset from unstructured blog posts to benchmark writing performance and introduce a comprehensive evaluation framework assessing the quality of outlines and generated texts. Our evaluations with various LLMs demonstrate that the WritingPath approach significantly enhances text quality according to evaluations by both LLMs and professional writers.

cs.CL↗

arXiv2024

HyperCLOVA X Technical Report

We introduce HyperCLOVA X, a family of large language models (LLMs) tailored to the Korean language and culture, along with competitive capabilities in English, math, and coding. HyperCLOVA X was trained on a balanced mix of Korean, English, and code data, followed by instruction-tuning with high-quality human-annotated datasets while abiding by strict safety guidelines reflecting our commitment to responsible AI. The model is evaluated across various benchmarks, including comprehensive reasoning, knowledge, commonsense, factuality, coding, math, chatting, instruction-following, and harmlessness, in both Korean and English. HyperCLOVA X exhibits strong reasoning capabilities in Korean backed by a deep understanding of the language and cultural nuances. Further analysis of the inherent bilingual nature and its extension to multilingualism highlights the model's cross-lingual proficiency and strong generalization ability to untargeted languages, including machine translation between several language pairs and cross-lingual inference tasks. We believe that HyperCLOVA X can provide helpful guidance for regions or countries in developing their sovereign LLMs.

cs.CL↗

arXiv2024

CheckEval: A reliable LLM-as-a-Judge framework for evaluating text generation using checklists

Existing LLM-as-a-Judge approaches for evaluating text generation suffer from rating inconsistencies, with low agreement and high rating variance across different evaluator models. We attribute this to subjective evaluation criteria combined with Likert scale scoring in existing protocols. To address this issue, we introduce CheckEval, a checklist-based evaluation framework that improves rating reliability via decomposed binary questions. Through experiments with 12 evaluator models across multiple datasets, we first demonstrate that CheckEval strongly correlates with human judgments. More importantly, CheckEval dramatically improves the average agreement across evaluator models by 0.45 and reduces the score variance. CheckEval scores furthermore have the benefit of being more interpretable because it decomposes evaluation criteria into traceable binary decisions, allowing analyses of specific attributes driving quality judgments.

cs.CL↗

arXiv2023

Empowering Sentence Encoders with Prompting and Label Retrieval for Zero-shot Text Classification

With contrastive pre-training, sentence encoders are generally optimized to locate semantically similar samples closer to each other in their embedding spaces. In this work, we focus on the potential of their embedding spaces to be readily adapted to zero-shot text classification, as semantically distinct samples are already well-separated. Our framework, RaLP (Retrieval augmented Label Prompts for sentence encoder), encodes prompted label candidates with a sentence encoder, then assigns the label whose prompt embedding has the highest similarity with the input text embedding. In order to compensate for the potentially poorly descriptive labels in their original format, RaLP retrieves sentences that are semantically similar to the original label prompt from external corpora and use them as additional pseudo-label prompts. RaLP achieves competitive or stronger performance than much larger baselines on various closed-set classification and multiple-choice QA datasets under zero-shot settings. We show that the retrieval component plays a pivotal role in RaLP's success, and its results are robustly attained regardless of verbalizer variations.

cs.CL↗

arXiv2022

Enhancing Out-of-Distribution Detection in Natural Language Understanding via Implicit Layer Ensemble

Out-of-distribution (OOD) detection aims to discern outliers from the intended data distribution, which is crucial to maintaining high reliability and a good user experience. Most recent studies in OOD detection utilize the information from a single representation that resides in the penultimate layer to determine whether the input is anomalous or not. Although such a method is straightforward, the potential of diverse information in the intermediate layers is overlooked. In this paper, we propose a novel framework based on contrastive learning that encourages intermediate features to learn layer-specialized representations and assembles them implicitly into a single representation to absorb rich information in the pre-trained language model. Extensive experiments in various intent classification and OOD datasets demonstrate that our approach is significantly more effective than other works.

cs.CL↗

arXiv2021

What Changes Can Large-scale Language Models Bring? Intensive Study on HyperCLOVA: Billions-scale Korean Generative Pretrained Transformers

GPT-3 shows remarkable in-context learning ability of large-scale language models (LMs) trained on hundreds of billion scale data. Here we address some remaining issues less reported by the GPT-3 paper, such as a non-English LM, the performances of different sized models, and the effect of recently introduced prompt optimization on in-context learning. To achieve this, we introduce HyperCLOVA, a Korean variant of 82B GPT-3 trained on a Korean-centric corpus of 560B tokens. Enhanced by our Korean-specific tokenization, HyperCLOVA with our training configuration shows state-of-the-art in-context zero-shot and few-shot learning performances on various downstream tasks in Korean. Also, we show the performance benefits of prompt-based learning and demonstrate how it can be integrated into the prompt engineering pipeline. Then we discuss the possibility of materializing the No Code AI paradigm by providing AI prototyping capabilities to non-experts of ML by introducing HyperCLOVA studio, an interactive prompt engineering interface. Lastly, we demonstrate the potential of our methods with three successful in-house applications.

cs.CL↗

arXiv2021

GPT3Mix: Leveraging Large-scale Language Models for Text Augmentation

Large-scale language models such as GPT-3 are excellent few-shot learners, allowing them to be controlled via natural text prompts. Recent studies report that prompt-based direct classification eliminates the need for fine-tuning but lacks data and inference scalability. This paper proposes a novel data augmentation technique that leverages large-scale language models to generate realistic text samples from a mixture of real samples. We also propose utilizing soft-labels predicted by the language models, effectively distilling knowledge from the large-scale language models and creating textual perturbations simultaneously. We perform data augmentation experiments on diverse classification tasks and show that our method hugely outperforms existing text augmentation methods. Ablation studies and a qualitative analysis provide more insights into our approach.

cs.CL↗

arXiv2018

An iALM-ICA-based Anti-Jamming DS-CDMA Receiver for LMS Systems

We consider a land mobile satellite communication system using spread spectrum techniques where the uplink is exposed to MT jamming attacks, and the downlink is corrupted by multi-path fading channels. We proposes an anti-jamming receiver, which exploits inherent low-dimensionality of the received signal model, by formulating a robust principal component analysis (Robust PCA)-based recovery problem. Simulation results verify that the proposed receiver outperforms the conventional receiver for a reasonable rank of the jamming signal.

cs.IT↗

arXiv2015

Bernoulli-Gaussian Approximate Message-Passing Algorithm for Compressed Sensing with 1D-Finite-Difference Sparsity

This paper proposes a fast approximate message-passing (AMP) algorithm for solving compressed sensing (CS) recovery problems with 1D-finite-difference sparsity in term of MMSE estimation. The proposed algorithm, named ssAMP-BGFD, is low-computational with its fast convergence and cheap per-iteration cost, providing phase transition nearly approaching to the state-of-the-art. The proposed algorithm is originated from a sum-product message-passing rule, applying a Bernoulli-Gaussian (BG) prior, seeking an MMSE solution. The algorithm construction includes not only the conventional AMP technique for the measurement fidelity, but also suggests a simplified message-passing method to promote the signal sparsity in finite-difference. Furthermore, we provide an EM-tuning methodology to learn the BG prior parameters, suggesting how to use some practical measurement matrices satisfying the RIP requirement under the ssAMP-BGFD recovery. Extensive empirical results confirms performance of the proposed algorithm, in phase transition, convergence speed, and CPU runtime, compared to the recent algorithms.

cs.IT↗

arXiv2015

Fast Signal Separation of 2D Sparse Mixture via Approximate Message-Passing

Approximate message-passing (AMP) method is a simple and efficient framework for the linear inverse problems. In this letter, we propose a faster AMP to solve the \emph{$L_1$-Split-Analysis} for the 2D sparsity separation, which is referred to as \emph{MixAMP}. We develop the MixAMP based on the factor graphical modeling and the min-sum message-passing. Then, we examine MixAMP for two types of the sparsity separation: separation of the direct-and-group sparsity, and that of the direct-and-finite-difference sparsity. This case study shows that the MixAMP method offers computational advantages over the conventional first-order method, TFOCS.

cs.IT↗

arXiv2015

Bayesian Hypothesis Test using Nonparametric Belief Propagation for Noisy Sparse Recovery

This paper proposes a low-computational Bayesian algorithm for noisy sparse recovery (NSR), called BHT-BP. In this framework, we consider an LDPC-like measurement matrices which has a tree-structured property, and additive white Gaussian noise. BHT-BP has a joint detection-and-estimation structure consisting of a sparse support detector and a nonzero estimator. The support detector is designed under the criterion of the minimum detection error probability using a nonparametric belief propagation (nBP) and composite binary hypothesis tests. The nonzeros are estimated in the sense of linear MMSE, where the support detection result is utilized. BHT-BP has its strength in noise robust support detection, effectively removing quantization errors caused by the uniform sampling-based nBP. Therefore, in the NSR problems, BHT-BP has advantages over CS-BP which is an existing nBP algorithm, being comparable to other recent CS solvers, in several aspects. In addition, we examine impact of the minimum nonzero value of sparse signals via BHT-BP, on the basis of the results of the recent literature. Our empirical result shows that variation of x_min is reflected to recovery performance in the form of SNR shift.

cs.IT↗

arXiv2013

Phase Transition Analysis of Sparse Support Detection from Noisy Measurements

This paper investigates the problem of sparse support detection (SSD) via a detection-oriented algorithm named Bayesian hypothesis test via belief propagation (BHT-BP). Our main focus is to compare BHT-BP to an estimation-based algorithm, called CS-BP, and show its superiority in the SSD problem. For this investigation, we perform a phase transition (PT) analysis over the plain of the noise level and signal magnitude on the signal support. This PT analysis sharply specifies the required signal magnitude for the detection under a certain noise level. In addition, we provide an experimental validation to assure the PT analysis. Our analytical and experimental results show the fact that BHT-BP detects the signal support against additive noise more robustly than CS-BP does.

cs.IT↗

arXiv2013

Bayesian Hypothesis Test for Sparse Support Recovery using Belief Propagation

In this paper, we introduce a new support recovery algorithm from noisy measurements called Bayesian hypothesis test via belief propagation (BHT-BP). BHT-BP focuses on sparse support recovery rather than sparse signal estimation. The key idea behind BHT-BP is to detect the support set of a sparse vector using hypothesis test where the posterior densities used in the test are obtained by aid of belief propagation (BP). Since BP provides precise posterior information using the noise statistic, BHT-BP can recover the support with robustness against the measurement noise. In addition, BHT-BP has low computational cost compared to the other algorithms by the use of BP. We show the support recovery performance of BHT-BP on the parameters (N; M; K; SNR) and compare the performance of BHT-BP to OMP and Lasso via numerical results.

cs.IT↗

arXiv2012

Detection-Directed Sparse Estimation using Bayesian Hypothesis Test and Belief Propagation

In this paper, we propose a sparse recovery algorithm called detection-directed (DD) sparse estimation using Bayesian hypothesis test (BHT) and belief propagation (BP). In this framework, we consider the use of sparse-binary sensing matrices which has the tree-like property and the sampled-message approach for the implementation of BP. The key idea behind the proposed algorithm is that the recovery takes DD-estimation structure consisting of two parts: support detection and signal value estimation. BP and BHT perform the support detection, and an MMSE estimator finds the signal values using the detected support set. The proposed algorithm provides noise-robustness against measurement noise beyond the conventional MAP approach, as well as a solution to remove quantization effect by the sampled-message based BP independently of memory size for the message sampling. We explain how the proposed algorithm can have the aforementioned characteristics via exemplary discussion. In addition, our experiments validate such superiority of the proposed algorithm, compared to recent algorithms under noisy setup. Interestingly the experimental results show that performance of the proposed algorithm approaches that of the oracle estimator as SNR becomes higher.

cs.IT↗

arXiv2012

On Detection-Directed Estimation Approach for Noisy Compressive Sensing

In this paper, we investigate a Bayesian sparse reconstruction algorithm called compressive sensing via Bayesian support detection (CS-BSD). This algorithm is quite robust against measurement noise and achieves the performance of a minimum mean square error (MMSE) estimator that has support knowledge beyond a certain SNR threshold. The key idea behind CS-BSD is that reconstruction takes a detection-directed estimation structure consisting of two parts: support detection and signal value estimation. Belief propagation (BP) and a Bayesian hypothesis test perform support detection, and an MMSE estimator finds the signal values belonging to the support set. CS-BSD converges faster than other BP-based algorithms, and it can be converted to a parallel architecture to become much faster. Numerical results are provided to verify the superiority of CS-BSD compared to recent algorithms.

cs.IT↗