Earticle

본 고에서는 멀티스크린 기반의 컨택트 센터를 위한 언어처리 기술에 대한 연구를 수행하였다. 개인이 갖고 있는 디바이스가 다양화 됨에 따라 전통적인 전화망 기반의 콜센터는 다양한 개인용 디바이스를 통한 입력도 처리가 가능한 컨택트 센터로 진화되고 있다. 이러한 입력 방식은 음성과 문자가 주종을 이루고 있으므로 컨텍트 센터도 두 종류의 입력에 대한 구조가 정의되어야 한다. 이를 위해 음성 질의어 처리와 문자 질의어 처리가 순차적으로 이루어지도록 하는 하이브리드 질의 시스템을 제안하였다. 그리고 음성 질의어 시스템을 위한 언어처리 기술 알고리즘에 대해서 사용되는 메모리 양, 인식률 및 인식 속도를 고려한 네 가지 방식이 제안 되었다. 제안된 방식을 평가하기 위하여 121 단어로 구성된 음성 DB와 이것을 전사한 문자 DB를 사용하였다. Bigram 방식에 비해서 bigram + trigram + 핵심어기반 방식이 22.3% 인식률이 향상되었고 look forward 방식을 도입할 경우 동일한 인식률 성능에 검색속도는 1% 정도 향상이 되었다. Look forward 방식의 검색 속도 개선이 작은 것은 단어의 양이 크지 않았기 때문인 것으로 검토가 되었다. 향후 단어의 양을 늘리고 실제 상황의 DB를 활용한 실험으로 확대하여 제안된 알고리즘의 성능을 확대 평가할 필요성이 있다.

In this paper, we present a study on the language processing technology for the contact center based on multi-screen devices. The traditional telephone-based call center is evolving into contact center which can handle information entered by various personal devices. Since those information include voice and text, the contact center should have architecture for those modes. We propose a hybrid system in which text query system is followed by voice query system. And we also suggest four kinds of language processing algorithms considering memory, speed in addition to performance in a voice query system. The experimental DB consists of voice and text of vocabulary size 121. The experimental result shows that bigram + trigram + keyword algorithm yields higher performance of 22.3% than bigram only algorithm. And we get time reduction of 1% when the look forward algorithm is added. We expect that look forward algorithm may give better improvement in case of large vocabulary system. In future, we are going to make various experiments on the large vocabulary system with real DB.

본 연구의 목적은 저자원 아프리카 언어에서 생성형 AI가 독해 지문을 정답 선택의 근거로 활용할 수 있는지 검토하는 데 있다. 이를 위해 Belebele의 영어, 암하라어, 스와힐리어, 요루바어 병렬 독해 문항을 수집하고 문항 등가성 검토를 거쳐 16개 문항을 선정하였다. qwen3:8b 모델을 활용하여 동일한 문항에 대해 각 언어별로 질문·선택지만 제시한 조건과 질문·선택지에 지문을 포함해서 제시한 조건에서의 응답 결과를 분석하였다. 분석 결과, 전체 정답률은 질문·선택지만 제시한 조건에서 45.3%, 지문·질문·선택지를 함께 제시한 조건에서 64.1%로 나타나 지문을 제시했을 때의 정답율이 높아짐을 확인했다. 반면 언어별로는 영어, 암하라어, 스와힐리어의 정답율은 높아졌으나 언어마다 차이를 보였고 요루바어는 변화가 없었다. 이러한 결과는 저자원 언어에서 동일한 정보가 제공되더라도 생성형 AI가 이를 처리하고 응답에 반영하는 방식이 다를 수 있음을 시사한다.

This study examines whether generative AI can use reading passages as evidence for answer selection in low-resource African languages. Using the Belebele benchmark, parallel reading-comprehension items in English, Amharic, Swahili, and Yoruba were collected, and 16 items were selected through item-equivalence review. The qwen3:8b model was tested on the same items under two conditions: one in which only the question and answer options were provided, and another in which the passage, question, and answer options were provided in the same language. The overall accuracy was 45.3% without passages and 64.1% with passages. By language, accuracy increased in English, Amharic, and Swahili, while Yoruba showed no change. These results suggest that, even when the same information is provided, generative AI may process and reflect it differently across low-resource languages.

Recently, there have been extensive developments in Natural Language Processing (NLP). However, most studies have focused primarily on English, and relatively little research has been conducted on Tamil. More than 80 million people worldwide use Tamil, one of the national languages of Sri Lanka. It has several unique features that make NLP work more difficult, such as different regional dialects, complex morphemes, and different ways of writing non-standardized words. These challenges are further compounded by the lack of digital content, labeled datasets, and pretrained language models available for Tamil. Although some useful resources exist, they remain limited and fall short of expectations. In this study, we review NLP on Tamil language, identify problems with existing resources, and suggest future developments. Additionally, we highlight the need for collaboration to create better datasets, stronger language tools, and new applications designed to fit the special features of Tamil. This study helps researchers to better understand Tamil and supports the inclusion of low-resource languages in NLP research.

The system evaluates the impact of applying NLP to library services on user accessibility, response speed, and interactive capabilities. Information processing enables computers to both process written documents and convert spoken words into written text, which becomes usable content. The system operates with spaCy, Hugging Face, MARC, and Dublin Core metadata platforms to clean data and tokenize information before building semantic models. The combination of GPT with BERT alongside improved language models enables libraries to automate their cataloging system and deliver both contextual search options and personalized suggestions. The platform gives users multilingual capabilities and efficiency tools and protects privacy through security protocols. Through neural translation, libraries eliminate linguistic roadblocks so users obtain speech tools for visual help. Despite strategic and ethical issues with resource limitations, solutions that aim to change library management systems can still be implemented. The framework proved that AI systems have the potential to improve the information organization and search capabilities in libraries. This methodology uses intelligent automation to determine how libraries will develop into space, providing inclusive knowledge access for future generations.

The vast number of biomedical literature is an important source of biomedical interaction information discovery. However, it is complicated to obtain interaction information from them because most of them are not easily readable by machine. In this paper, we present a method for extracting biomedical interaction information assuming that the biomedical Named Entities (NEs) are already identified. The proposed method labels all possible pairs of given biomedical NEs as INTERACTION or NO-INTERACTION by using a Maximum Entropy (ME) classifier. The features used for the classifier are obtained by applying various NLP techniques such as POS tagging, base phrase recognition, parsing and predicate-argument recognition. Especially, specific verb predicates (activate, inhibit, diminish and etc.) and their biomedical NE arguments are very useful features for identifying interactive NE pairs. Based on this, we devised a twostep method: 1) an interaction verb extraction step to find biomedically salient verbs, and 2) an argument relation identification step to generate partial predicate-argument structures between extracted interaction verbs and their NE arguments. In the experiments, we analyzed how much each applied NLP technique improves the performance. The proposed method can be completely improved by more than 2% compared to the baseline method. The use of external contextual features, which are obtained from outside of NEs, is crucial for the performance improvement. We also compare the performance of the proposed method against the co-occurrence-based and the rule-based methods. The result demonstrates that the proposed method considerably improves the performance.

Background: The rapid growth of the aging population has increased the demand for technologies related to muscle health, assessment, prevention, and functional support. Purpose: This study aimed to analyze technology trends in muscle-related innovations using natural language processing and machine learning applied to South Korean patent and utility model data. Study design: Descriptive study using unsupervised machine learning. Methods: A total of 2,836 records were analyzed using combined title and abstract texts. Semantic embeddings were generated using Sentence-BERT, followed by dimensionality reduction with Uniform Manifold Approximation and Projection and clustering with Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN). Keywords for each cluster were extracted using term frequency-inverse document frequency, and annual filing patterns were examined to identify the trends. Results: Two dominant macro-domains were identified, including a Wearable and Sensor Technology domain (n=1,695) and a Biological and Preventive Technology domain (n=903). HDBSCAN yielded substantially better cluster validity, with a silhouette coefficient of 0.610 and Davies-Bouldin index of 0.525. Within these macro-domains, sub-domains related to stem cell and regenerative therapy, functional food and bioactive extracts, antibody and antiviral therapeutics, aging and sarcopenia mechanisms, and massage devices were identified. Conclusions: These findings provide an overview of muscle-related technology development and demonstrate the value of machine learning-based patent text analysis for technology landscape mapping and innovation assessment.

This study examined public perceptions of generative AI through the analysis of user comments on YouTube news videos regrading generative AI from three major South Korean broadcasting networks. Using structural topic modeling, 56,708 comments from 105 news videos were analyzed. Among nine distinct topics that emerged, the top three prevalent topics centered on labor market change, AI control concerns, and service automation benefits. Temporal analysis revealed evolving discourse patterns. Employment-related concerns peaked after the release of ChatGPT's but subsequently declined; three topics gained increasing prominence including human-centered AI development, control concerns, and educational applications. These findings offer theoretical and practical implications for individuals, organizations, and institutions adopting generative AI.

This paper suggests a variation of a well-known probabilistic matrix factorization algorithm which is commonly used in data analysis and scientific computing, and which has been considered recently to serve natural language processing. The proposed variation is meant to take benefit from the fact that matrices processed in natural language processing tasks are normally sparse rectangular matrices with one dimension much larger than the other, and this can be used to ensure adequate accuracy with acceptable computation time. Preliminary experiments on real-world textual corpora show that the proposed algorithm achieves relevant improvements compared to the original one.

Recently, effort to obtain various information from the vast amount of social network services (SNS) big data generated in daily life has expanded. SNS big data comprise sentences classified as unstructured data, which complicates data processing. As the amount of processing increases, a rapid processing technique is required to extract valuable information from SNS big data. We herein propose a system that can extract human sentiment information from vast amounts of SNS unstructured big data using the naïve Bayes algorithm and natural language processing (NLP). Furthermore, we analyze the effectiveness of the proposed method through various experiments. Based on sentiment accuracy analysis, experimental results showed that the machine learning method using the naïve Bayes algorithm afforded a 63.5% accuracy, which was lower than that yielded by the NLP method. However, based on data processing speed analysis, the machine learning method by the naïve Bayes algorithm demonstrated a processing performance that was approximately 5.4 times higher than that by the NLP method.

Objective To investigate the hemispheric contributions to prosody recognitions and interference effects of semantic processing on prosody for stroke patients by using the Korean language.Methods Ten right hemisphere damaged patients (RHD), nine left hemisphere damaged patients (LHD), and eleven healthy controls (HC) participated. In pure prosody recognition task, four semantically neutral sentences were selected and presented in both sad and happy prosodies. In interference task, participants listened to emotionally intoned sentences in which the semantic contents were congruent or incongruent with prosody. Participants were asked to rate the valence of prosody while ignoring the semantic contents, and thus, reaction time and accuracy were estimated.Results In pure prosody recognition task, RHD showed low accuracy as compared to HC (p=0.013), and the tendency of group response showed that RHD performed worse than HC and LHD with regards to accuracy and reaction time. In interference task, analysis of accuracy revealed a significant main effect of groups (p=0.04), and the tendency implied that RHD is less accurate as compared to LHD and HC. The RHD took longer reaction times than HC in congruent and incongruent items (p<0.001).Conclusion Right hemispheric laterality to prosody processing of Korean language in stroke patients was observed. Interference effects of semantic contents to prosody processing were not observed, which suggested unique characteristics of prosody for Korean language. These results could be referred as preliminary data for future researches on Korean languages.

Studies on the development of L2 negation suggest that preverbal negation is learned earlier than postverbal negation regardless of a learner’s first language background. Korean provides avenues of L2 negation research because it allows both preverbal and postverbal negative constructions. The current paper aims to propose a rank order of processing difficulty involved in different types of errors committed by learners acquiring Korean as a foreign language. Twenty-four second language learners of Korean enrolled in two intermediate courses at a university participated in this investigation. An Elicited Imitation (EI) task was administered to each learner. EI accuracy scoring showed that a higher level of processing difficulty was required in repeating post-verbally negated stimuli than in repeating pre-verbally negated ones. Five different types of errors were identified, and an implicational pattern of these error types was found. Among the five types of errors, the conversion of post-verbal into grammatical pre-verbal negation was the most frequently committed error. This indicates learners’ simplification strategy for ease of production in preverbal negation while fully processing negative meaning in a postverbal construction. The results are also consistent with the prediction made by the typological markedness theory (Dahl 1979) and Processibility Theory (Pienemann 1999, 2007) claiming that preverbal negation appears earlier in the development than its postverbal counterpart.

Natural language processing, as an integral part of artificial intelligence technology, has foundations in a variety of disciplines, including linguistics, computer science, and mathematics. Rapid advances in natural language processing provide solid backing for machine translation research. This document first sets out the key concepts and key points of computational linguistics, followed by a brief review of the history and progress of NLP research in the United States and abroad. The document then summarizes the three stages of machine translation as well as the current state of research. Historically, the advancement curves of natural language processing and machine translation have almost coincided, as well as the two complement each other. On this premise, the paper examines NLP applications in machine translation and highlights problems and trends in the fields of artificial intelligence. Finally, the authors examine the link between machine translation and human interpretation in the era of artificial intelligence and speculate on machine translations long term prospects.

In general, for the explanation of language processing of morphologically complex words, there are two competing models; the continuous (or left-to-right) processing approach and the discontinuous (or decompositional) processing approach. One of the typical arguments for the discontinuous processing approach is the Prefix-Stripping Hypothesis(cf. Taft and Foster 1975). The main claim of the Prefix-Stripping Hypothesis is that prefixes should be stripped off before the lexical access, since the lexical access can be done via only stem. In this paper, we provided an account of some asymmetrical behaviors of English prefixes and suffixes on the basis of the language parsing processes. Moreover, we attempted to provide further evidence supporting the Prefix-Stripping Hypothesis. As discussed in Melinger(2001) and Wurm(1996, 1997), prefixes and suffixes in English exhibit several asymmetrical behaviors. For instance, when suffixes are attached to the stem, they cause changes to the consonants of the stem. However, prefixes do not cause such change. In addition, we can find more suffixes across languages than prefixes. Through explaining and analyzing such prefix/suffix asymmetries in English, we provided further evidence for the Prefix-Stripping Hypothesis and argued that the Prefix-Stripping Hypothesis is more advantageous for the explanation of the prefix/suffix asymmetries in English.

본 연구는 영어 원어민과 비원어민 대학생의 주장문 쓰기에서 나타나는 어휘 묶음(lexical bundle) there are so many의 의미적 운율(semantic prosody)을 분석한다. 기존 연구들이 주로 구조와 기능적 특성에 집중해온 반면, 본 연구는 담화 맥락에서 드러나는 긍·부정·중립적 의미 함의에 초점을 맞춘다. 자연어처리 기반 감성 분석을 활용하여, 두 학생 집단이 해당 어휘 묶음을 통해 어떤 입장과 태도를 드러내는지를 두 개의 비교 가능한 말뭉치를 바탕으로 분석하였다. 두 말뭉치는 각각 미국, 한국 대학생들이 작성한 영어 주장문으로 구성되어 있으며, 두 집단 모두 동일한 주제에 대해 제한된 시간 내에 글을 작성하였다. 연구 결과, 원어민 학생들은 해당 어휘 묶음을 주로 중립적이거나 긍정적인 맥락에서 사용하였으며, 이는 주로 우회적 표현이나 예시 제시에 활용되었다. 반면, 비원어민 학습자들은 이 어휘 묶음을 부정적인 맥락에서 더 자주 사용하였고, 그 결과 의미적으로 어색하거나 담화 상황에 부적절하게 사용되는 경우도 나타났다. 이러한 결과는 두 집단이 동일한 어휘 묶음을 문맥에 따라 다르게 활용한다는 점을 보여주며, 이는 담화 맥락에 적절한 표현 선택을 강조하는 글쓰기 교육이 필요함을 시사한다.

This study explores the semantic prosody of the lexical bundle (LB) there are so many in native and non-native novice English argumentative writing. While previous research on LBs has largely focused on their structural and functional patterns, this study shifts attention to their evaluative meanings in context. Using NLP-based sentiment analysis, it examines how this bundle conveys evaluative stance across two comparable corpora: one consisting of argumentative essays written by native English-speaking students and the other by Korean EFL learners. Both groups were incoming college freshmen who responded to the same prompts under identical time constraints. The results exhibited a clear difference in semantic prosody: native writers tended to use the bundle in neutral or positive contexts, often to hedge or provide examples, whereas non-native writers used it more frequently in negative contexts, which could lead to semantic incongruity or pragmatic misalignment. These findings move beyond existing accounts of LBs by highlighting the role of evaluative meaning. They suggest that L2 learners may benefit from explicit instruction on how stance is expressed in frequently occurring word sequences—an area particularly challenging for novice academic writers. This study highlights the value of sentiment-aware writing pedagogy, as informed by corpus-based insights into learner language.

Introducing the concept of construction safety in the design/planning phase can improve the efficiency and effectiveness of safety management on construction sites. In this sense, further improvements for safety can be made by designers’ involvement in the elimination or management of design-related hazards in addition to contractors’ own efforts, as known as Prevention through Design (PtD). However, this idea has not been well materialized in construction practices due in part to the lack of a PtD tool that allows designers to automatically check and evaluate designs in terms of worker safety. The paper addresses this issue by designing and testing an automated compliance checking framework that extracts information relevant to design-related hazards from regulatory documents and maps it to each of related design components using Natural Language Processing (NLP) and Building Information Modeling (BIM). The proposed framework has two key functions: 1) automatically extracting machine-readable regulation information relevant to design-related hazards using NLP techniques; and 2) automatically mapping such information to corresponding design components in a building information model, enabling the visualization of hazards. How all these components work to automatically display design-related hazards and relevant regulatory information is illustrated with a three-story building, focusing on fall protection. The proposed framework successfully extracted regulation information from textual documents with about 90% of precision and recall rates. It also correctly identified all the hazards in the case study model. The automated framework is expected to facilitate safety design review by improving designers and contractors’ proactive responses against design-related hazards

Service organizations increasingly adopt data-based intelligent engines called chatbots in support of the interaction between customers and the companies. Two different types of chatbots have been suggested and introduced by companies leading the adoption of this emerging technology: rule-based chatbots and natural language processing-based chatbots. While the differences between these two types of technologies look relatively clear, the organizational and practical impacts of the differences have not been systematically explored. This study performed an experiment to compare the use of the two different types of chatbots used in practice by two comparable organizations. These two types of actual chatbots were used by Korean on-line shopping malls with similar business models (mobile shopping), length of history, size and reputation. The comparison was made based on such dimensions as usability, searchability, reliability and attractiveness. Contraty to conventional expectation that the superiority in technology will produce superior usability, the results show mixed superiority. The discussion on the reasons is presented.

With social media growing fast, user-generated content (UGC) has become a key factor in influencing how consumers decide what to buy, especially in the travel and hospitality sector. But there are a lot of fake and extreme reviews, which are making it hard for customers to make the right choice and making competition in the market unfair. The present study focuses on YouTube, a platform with high global activity, and proposes a systematic solution that combines Natural Language Processing (NLP) and machine learning methods (e.g., VADER Sentiment Analysis, Support Vector Machine SVM, and LDA Topic Modelling) for identifying and filtering fake and extreme remarks in hotel reviews. This approach has been shown to enhance the automation and precision of review screening processes. Furthermore, it provides a theoretical foundation and practical methodologies to improve the online information ecology, thereby enhancing the quality of user decision-making.

목적: 본 연구는 “학술연구정보서비스”에서 “아시안뷰티화장품학술지”와 국내 학술지 논문 제목에서 “화장품”을 키워드로 사용 하는 학술지의 연구 논문 초록을 크롤링하여, 이들 두 집단은 과연 차별화된 정체성으로 포지셔닝되어 있으며, 어떠한 유사도와 차 이점을 가지고 있을까? 에 대한 의문으로 시작하였다. 방법: 본 연구를 수행하기 위해 Python version 3.10.9 프로그램을 이용하여 학술연구정보서비스에서 아시안뷰티화장품학술지 433편의 연구 논문과 ‘화장품’이라는 키워드로 5,232편의 연구 논문을 크롤링하 여 전처리 과정을 거쳐서 군집 분석, 공저자 네트워크 분석, 토픽 모델링 분석, 유사도 분석을 시행하였다. 결과: 두 학술지 집단은 모두 공통적으로 화장품 산업의 핵심 요소인 피부, 소비자 행동, 브랜드 전략 등 다양한 주제를 다루고 있는 것으로 나타났으나, 세 부적으로는 연구 초점이 다르게 나타났으며, 정체성이 차별화되어 있는 것으로 나타났다. 결론: 공저자 네트워크 분석에서 나타난 각 중심성 분석 결과는 연구의 시너지 효과 극대화를 위한 각 공저자의 역할과 영향력을 알 수 있었으며, 아시안뷰티화장품학술지는 비교 대상 군과는 다른 고유한 차별화된 정체성을 가지고 있는 것으로 나타났다.

Purpose: The aim of this study was to determine whether the “Asian Journal of Beauty and Cosmetology” and domestic journals that use “cosmetics” as a keyword in the abstract of their paper, are positioned with differentiated identities and to identify the similarities and differences between them. All articles were procured from the “Research Information Sharing Service.” Methods: Python version 3.10.9 was used in this study to identify 433 research papers from the “Asian Journal of Beauty and Cosmetology” and 5,232 research papers with the keyword “cosmetic” from the “Research Information Sharing Service.” After preprocessing the articles, a co-author network analysis was performed, followed by cluster analysis and topic modeling analysis. Four types of similarity analyses were performed based on the results obtained. Results: The two journal groups were found to commonly cover a variety of topics such as skin, consumer behavior, and brand strategy, that are central to the cosmetics industry. However, the research topics had different central focuses, indicating distinct identities. Conclusion: The centrality analysis results from the co-authorship network analysis revealed the roles and influence of each co-author in maximizing the synergy of research. The identity of the Asian Journal of Beauty and Cosmetology was found be unique and differentiated, compared with the comparison group.

目的: 这篇研究始于一个关于以下问题：在“学术研究信息服务”中《亚洲美容学术杂志》与国内学术期刊的论文题 目中以“化妆品”为关键词的期刊的论文摘要进行网络爬行，对这两个群体是否真正具有差异化的本质定位，他 们有哪些相同点和不同点？ 方法: 本研究使用Python 3.10.9版本识别出来自《亚洲美容与美容杂志》的433篇研 究论文以及来自“学术研究信息服务”的5,232篇关键词为“化妆品”的研究论文。对文章进行预处理后，进行合著 者网络分析，然后进行聚类分析和主题建模分析。根据获得的结果进行了四种类型的相似性分析。结果: 发现这 两个期刊组普遍涵盖了化妆品行业核心的各种主题，例如皮肤、消费者行为和品牌战略。然而，研究主题的中 心点不同，表现出不同的本性。结论: 合着网络分析的中心性分析结果揭示了每位合著者在最大化研究协同作用 中的作用和影响。与对照组相比，《亚洲美容与美容杂志》的本质是独特且有区别的。