Gi-Taek An

In product search, fine-grained cues within user queries—such as attribute terms and constraint expressions—play a crucial role in determining ranking quality. However, existing dense and late-interaction retrieval models, including ColBERT, assign uniform importance to all query tokens, thereby failing to capture attribute sensitivity. To overcome this limitation, we propose ColBERT-AW (Attribute-aware Query-Token Weighting), a lightweight extension that injects token-specific scalar weights into ColBERT’s late-interaction computation. The module operates exclusively on the query side, preserving document embeddings and the index, and introduces only a minimal computational overhead through a single MLP pass. During training, attribute-token annotations generated by a large language model (LLM) guide the model to learn attribute-aware token weighting through a dual-objective loss that jointly optimizes token- and document-level supervision. Experiments on the TREC 2023 Product Search test set demonstrate consistent and significant improvements, achieving an NDCG@10 of 0.7584, a +0.0496 absolute gain (+7.0%) over the baseline ColBERT. These findings confirm that augmenting ColBERT with attribute-aware query weighting substantially enhances retrieval effectiveness while maintaining efficiency, scalability, and deployability, especially in scenarios requiring fine-grained understanding of attribute and constraint signals.

Recent advances in deep learning-based retrieval have established models such as SPLADE—a prominent variant of Learned Sparse Retrieval (LSR)—as state-of-the-art approaches, particularly in benchmark tasks like TREC Product Search. However, conventional LSR models (e.g., SPLADE) often struggle to effectively capture the semantics of negative Query terms and their lexical variations, leading to misinterpretation of user intent. This limitation leads to retrieval results that do not accurately reflect user intent, especially in queries containing negations like “without”. To overcome this challenge, we propose a novel activation function, Negative Gaussian Linear Unit (NeGLU), designed to assign negative weights that better represent the semantics of negated terms. By integrating NeGLU into an LSR framework, our method improves the model’s ability to identify and suppress negatively influential terms, thereby enhancing ranking performance. Extensive experiments on the TREC2023 product search test collection demonstrate the effectiveness of our approach, achieving an NDCG@100 score of 0.745, significantly outperforming existing methods. To further validate its robustness, we evaluated NeGLU on 186 test queries, including 54 involving negation, where it showed consistent gains—with an average improvement rate of 26.95% overall and 35.

Food research is uniquely intertwined with everyday life and necessitates the utilization of big data. Within this domain, the research data consist of various forms and formats, encompassing biological experiment results, chemical analysis data, nutritional information, microbiological data, sensor data, images, and videos. This diversity stems from the integration of data from various subdomains within the larger field. With recent advancements in deep learning technology, the importance of data has grown significantly, resulting in increased reliance on data-driven research. Although specialized platforms for sharing and utilizing data have been established at the national level, particularly in the bioscience field, food research lacks a dedicated infrastructure and specialized data-sharing platforms. In this study, we develop a platform that leverages Hadoop-based distributed file systems to create a data lake. This platform enables data storage and sharing through a web-based interface. The distributed file system supports scalability by adding data nodes, making it an effective solution for capacity expansion. In addition, the web-based platform ensures high accessibility, allowing users access from anywhere, at any time, using any device. Finally, we introduce the establishment of a 1.8 PB Hadoop-based physical storage system and present an approach for building a highly accessible web platform with substantial utility.

최근 딥러닝을 이용한 자연어처리 분야에서 다양한 모델이 소개되는 가운데 BERT와 GPT 등 트랜스포머(Transformer) 기반의 사전훈련 언어모델(Pre-trained model)이 기본이 되고 있다. 트랜스포머 기반 모델의 파인-튜닝(Fine-tuning)은 전체 모델의 파라미터가 업데이트되어 우수한 성능을 보여주고 있다. 최근에는 적은 양의 파라미터를 업데이트하여 성능을 개선하는 P-tuning 방식이 등장하였다. 본 논문에서는 모델의 파라미터의 학습을 동결하여 적은 양의 파라미터만 업데이트하더라도 기존의 파인-튜닝과 비슷한 성능을 달성할 수 있는 피-튜닝 방식에서 프롬프트 인코더(Prompt-encoder)를 변경한 방법을 제안하였다. 성능 검증을 위하여 GPT-2 모델은 KoGPT2를 사용하였다. NSMC와 KorNLI 데이터셋을 이용한 분류 결과, 제안한 방법은 기존의 피-튜닝 방식과 비교하여 NSMC와 KorNLI 데이터셋으로 각각 4.56%와 11%의 정확도가 향상된 성능 향상 결과를 보였다.

Information systems play an important role in business management, especially in personnel, budget, and financial management. If an anomaly ensues in an information system, all operations are paralyzed until their recovery. In this study, we propose a method for collecting and labeling datasets from actual operating systems in corporate environments for deep learning. The construction of a dataset from actual operating systems in a company’s information system involves constraints. Collecting anomalous data from these systems is challenging because of the need to maintain system stability. Even with data collected over a long period, the training dataset may have an imbalance of normal and anomalous data. We propose a method that utilizes contrastive learning with data augmentation through negative sampling for anomaly detection, which is particularly suitable for small datasets. To evaluate the effectiveness of the proposed method, we compared it with traditional deep learning models, such as the convolutional neural network (CNN) and long short-term memory (LSTM). The proposed method achieved a true positive rate (TPR) of 99.47%, whereas CNN and LSTM achieved TPRs of 98.8% and 98.67%, respectively. The experimental results demonstrate the method’s effectiveness in utilizing contrastive learning and detecting anomalies in small datasets from a company’s information system.