• The Transactions of the Korea Information Processing Society
• /
• v.13 no.4
• /
• pp.174-180
• /
• 2024

In this paper, we explore an emotion classification method through multimodal learning utilizing wav2vec 2.0 and KcELECTRA models. It is known that multimodal learning, which leverages both speech and text data, can significantly enhance emotion classification performance compared to methods that solely rely on speech data. Our study conducts a comparative analysis of BERT and its derivative models, known for their superior performance in the field of natural language processing, to select the optimal model for effective feature extraction from text data for use as the text processing model. The results confirm that the KcELECTRA model exhibits outstanding performance in emotion classification tasks. Furthermore, experiments using datasets made available by AI-Hub demonstrate that the inclusion of text data enables achieving superior performance with less data than when using speech data alone. The experiments show that the use of the KcELECTRA model achieved the highest accuracy of 96.57%. This indicates that multimodal learning can offer meaningful performance improvements in complex natural language processing tasks such as emotion classification.

• Journal of Biomedical Engineering Research
• /
• v.43 no.2
• /
• pp.109-115
• /
• 2022

Purpose: Medical records classification using vectorization techniques plays an important role in natural language processing. The purpose of this study was to investigate proper vectorization techniques for electronic medical records classification. Material and methods: 403 electronic medical documents were extracted retrospectively and classified using the cosine similarity calculated by Scikit-learn (Python module for machine learning) in Jupyter Notebook. Vectors for medical documents were produced by three different vectorization techniques (TF-IDF, latent sematic analysis and Word2Vec) and the classification precisions for three vectorization techniques were evaluated. The Kruskal-Wallis test was used to determine if there was a significant difference among three vectorization techniques. Results: 403 medical documents were relevant to 41 different diseases and the average number of documents per diagnosis was 9.83 (standard deviation=3.46). The classification precisions for three vectorization techniques were 0.78 (TF-IDF), 0.87 (LSA) and 0.79 (Word2Vec). There was a statistically significant difference among three vectorization techniques. Conclusions: The results suggest that removing irrelevant information (LSA) is more efficient vectorization technique than modifying weights of vectorization models (TF-IDF, Word2Vec) for medical documents classification.

• Journal of the Korea Convergence Society
• /
• v.12 no.1
• /
• pp.57-63
• /
• 2021

Spoken language understanding, which aims to understand utterance as naturally as human would, are mostly focused on English language. In this paper, we construct a Korean language dataset for spoken language understanding, which is based on a conversational corpus between reservation system and its user. The domain of conversation is limited to restaurant reservation. There are 7 types of slot tags and 5 types of intent tags in 6857 sentences. When a model proposed in English-based research is trained with our dataset, intent classification accuracy decreased a little, while slot filling F1 score decreased significantly.

• Journal of Intelligence and Information Systems
• /
• v.28 no.3
• /
• pp.45-55
• /
• 2022

Recently, large-scale language models (LPLM) have been shown state-of-the-art performances in various tasks of natural language processing including intent classification. However, fine-tuning LPLM requires much computational cost for training and inference which is not appropriate for dialog system. In this paper, we propose compressed intent classification model for multi-agent in low-resource like CPU. Our method consists of two stages. First, we trained sentence encoder from LPLM then compressed it through knowledge distillation. Second, we trained agent-specific adapter for intent classification. The results of three intent classification datasets show that our method achieved 98% of the accuracy of LPLM with only 21% size of it.

• Journal of the Korean Society for information Management
• /
• v.21 no.4 s.54
• /
• pp.251-263
• /
• 2004

For question-answering system, question analysis module finds the question points from user's natural language questions, classifies the question types, and extracts some useful information for answer. This paper proposes a question type classifying technique based on focus words extracted from questions and word semantic information, instead of complicated rules or huge knowledge resources. It also shows how to find the question type without focus words, and how useful the synonym or postfix information to enhance the performance of classifying module.

• International conference on construction engineering and project management
• /
• 2022.06a
• /
• pp.744-751
• /
• 2022

Construction is among the most dangerous industries with numerous accidents occurring at job sites. Following an accident, an investigation report is issued, containing all of the specifics. Analyzing the text information in construction accident reports can help enhance our understanding of historical data and be utilized for accident prevention. However, the conventional method requires a significant amount of time and effort to read and identify crucial information. The previous studies primarily focused on analyzing related objects and causes of accidents rather than the construction activities. This study aims to extract construction activities taken by workers associated with accidents by presenting an automated framework that adopts a deep learning-based approach and natural language processing (NLP) techniques to automatically classify sentences obtained from previous construction accident reports into predefined categories, namely TRADE (i.e., a construction activity before an accident), EVENT (i.e., an accident), and CONSEQUENCE (i.e., the outcome of an accident). The classification model was developed using Convolutional Neural Network (CNN) showed a robust accuracy of 88.7%, indicating that the proposed model is capable of investigating the occurrence of accidents with minimal manual involvement and sophisticated engineering. Also, this study is expected to support safety assessments and build risk management systems.

• Annual Conference on Human and Language Technology
• /
• 2018.10a
• /
• pp.155-159
• /
• 2018

최근 화두가 되고있는 지능형 개인 비서 시스템에서 자연어 이해(NLU) 시스템은 중요한 구성요소이다. 자연어 이해 시스템은 사용자의 발화로부터 대화의 도메인(domain), 의도(intent), 의미적 슬롯(semantic slot)을 분류하는 역할을 한다. 하지만 자연어 이해 시스템을 학습하기 위해서는 많은 양의 라벨링 된 데이터를 필요로 하며 새로운 도메인으로 시스템을 확장할 때, 새롭게 데이터 라벨링을 진행해야 하는 한계점이 존재한다. 이를 해결하기 위해 본 연구는 적대 학습 방법을 이용하여 풍부한 양으로 구성된 기존(source) 도메인의 데이터부터 적은 양으로 라벨링 된 데이터로 구성된 대상(target) 도메인을 위한 슬롯 채우기(slot filling) 모델 학습 방법을 제안한다. 실험 결과 적대 학습을 적용할 경우, 적대 학습을 적용하지 않은 경우 보다 높은 f-1 score를 나타냄을 확인하였다.

• The Journal of Information Systems
• /
• v.25 no.4
• /
• pp.121-139
• /
• 2016

Purpose In this study, we propose the special purposed R package named ""new_Noun()" to process nonstandard texts appeared in various social networks. As the Big data is getting interested, R - analysis tool and open source software is also getting more attention in many fields. Design/methodology/approach With more than 9,000 R packages, R provides a user-friendly functions of a variety of data mining, social network analysis and simulation functions such as statistical analysis, classification, prediction, clustering and association analysis. Especially, "KoNLP" - natural language processing package for Korean language - has reduced the time and effort of many researchers. However, as the social data increases, the informal expressions of Hangeul (Korean character) such as emoticons, informal terms and symbols make the difficulties increase in natural language processing. Findings In this study, to solve the these difficulties, special algorithms that upgrade existing open source natural language processing package have been researched. By utilizing the "KoNLP" package and analyzing the main functions in noun extracting command, we developed a new integrated noun processing package "new_Noun()" function to extract nouns which improves more than 29.1% compared with existing package.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2021.05a
• /
• pp.78-80
• /
• 2021

Currently, research on natural language processing (NLP) is rapidly evolving. Natural language processing is a technology that allows computers to analyze the meanings of languages used in everyday life, and is used in various fields such as speech recognition, spelling tests, and text classification. Currently, the most commonly used natural language processing library is NLTK based on English, which has a disadvantage in Korean language processing. Therefore, after introducing KonLPy and Soynlp, the Korean Tokenizing libraries, we will analyze morphology analysis and processing techniques, compare and analyze modules with Soynlp that complement KonLPy's shortcomings, and use them as natural language processing models.

• Journal of the Korea Society of Computer and Information
• /
• v.13 no.6
• /
• pp.263-271
• /
• 2008

This paper presents a Korean conversational agent system in a mobile environment using natural language processing techniques. The aim of a conversational agent in mobile environment is to provide natural language interface and enable more natural interaction between a human and an agent. Constructing such an agent, it is required to develop various natural language understanding components and effective utterance generation methods. To understand spoken style utterance, we perform morphosyntactic analysis, shallow semantic analysis including modality classification and predicate argument structure analysis, and to generate a system utterance, we perform example based search which considers lexical similarity, syntactic similarity and semantic similarity.