자질집합선택 기반의 기계학습을 통한 한국어 기본구 인식의 성능향상

황영숙,정후중,박소영,곽용재,임해창,Hwang, Young-Sook,Chung, Hoo-jung,Park, So-Young,Kwak, Young-Jae,Rim, Hae-Chang 한국정보과학회 2002 정보과학회논문지 : 소프트웨어 및 응용 Vol.29 No.9

In this paper, we present an empirical study for improving the Korean text chunking based on machine learning and feature set selection approaches. We focus on two issues: the problem of selecting feature set for Korean chunking, and the problem of alleviating the data sparseness. To select a proper feature set, we use a heuristic method of searching through the space of feature sets using the estimated performance from a machine learning algorithm as a measure of "incremental usefulness" of a particular feature set. Besides, for smoothing the data sparseness, we suggest a method of using a general part-of-speech tag set and selective lexical information under the consideration of Korean language characteristics. Experimental results showed that chunk tags and lexical information within a given context window are important features and spacing unit information is less important than others, which are independent on the machine teaming techniques. Furthermore, using the selective lexical information gives not only a smoothing effect but also the reduction of the feature space than using all of lexical information. Korean text chunking based on the memory-based learning and the decision tree learning with the selected feature space showed the performance of precision/recall of 90.99%/92.52%, and 93.39%/93.41% respectively.

특집 : 병렬말뭉치의 구축과 활용 ; 통계기반 기계번역 기술의 소개 및 최신 연구동향 분석

황영숙 ( Young Sook Hwang ) 연세대학교 언어정보연구원(구 연세대학교 언어정보개발원) 2010 언어사실과 관점 Vol.25 No.-

The filed of machine translation has recently been energized by the statistical techniques, which try to discover the rules of translations automatically from a large parallel corpus, by pairing the input and output of the translation process and learning from the statistics over the parallel corpus. Statistical Machine Translation is a machine translation paradigm where machine translations are generated on the basis of statistical models whose parameters are learned from the parallel corpus. Statistical machine translation has been in the spotlight recently in both in the research community and in the commercial fields. This paper introduces the statistical machine translation methods, including the basis of statistical machine translation, automatic evaluation methods of machine translation systems, the benefits of SMT, and the advanced techniques. Moreover. we discuss the limitations of the statistical machine translation methods without using the linguistic features. Then, we examine the methods which has tried to adopt some linguistic features such like the morphological and syntactical features into the SMT models, in order to overcome the shortage of the simple SMT methods. Finally, we will discuss the future direction of the machine translation fields, such as data challenge.

한국어의 음절 결합 특성 및 통사적 어휘 특성을 이용한 문자인식 후처리 시스템

황영숙(Young-Sook Hwang),박봉래(Bong-Rae park),임해창(Hae-Chang Rim) 한국정보과학회 언어공학연구회 1997 한국정보과학회 언어공학연구회 학술발표 논문집 Vol.1997 No.10

개체동결 굴(Crassostrea gigas)을 이용한 레토르트파우치 굴국의 제조 및 품질특성

황영숙 ( Young-sook Hwang ),조준현 ( Jun-hyun Cho ),황석민 ( Seok-min Hwang ),김상현 ( Sang-hyun Kim ),김병균 ( Byeong-gyun Kim ),오광수 ( Kwang-soo Oh ) 한국수산과학회(구 한국수산학회) 2016 한국수산과학회지 Vol.49 No.6

To develop a value-added product from individually quick-frozen oysters Crassostrea gigas (IQFO), we prepared a retort pouched oyster soup (RPOS) from IQFOs and characterized its processing conditions and quality metrics. We found that the most appropriate manufacturing process for the RPOS consisted of half-thawing and washing raw IQF oysters, blanching, adding them to the retort pouch along with other ingredients (base soup stock, IQF oyster extract, radish, bean sprouts, garlic, and red pepper), sealing, retort sterilization (120°, F0-value 10 min.), cooling, and packaging inspection. The moisture, crude protein, pH and salinity of the RPOS were 91.0%, 2.8%, 6.20 and 0.9%, respectively. The total amino acid content of the RPOS was 2,163.8 mg/100 g, and the main amino acids were glutamic acid, aspartic acid, leucine, proline, lysine and arginine. The primary inorganic ions were Na, K, S and Zn. In taste compounds, total free amino acid content was 313.4 mg/100 g, and the main free amino acids were glutamic acid, taurine, proline, hydroxyproline, aspartic acid, glycine, alanine, valine, lysine and arginine. This RPOS has good storage stability and organoleptic qualities compared with commercial retort pouched shellfish soup, and is suitable for commercialization as a value-added instant seafood soup.

위수탁 검사의뢰 국산 및 수입화장품의 비교고찰

황영숙 ( Young Sook Hwang ),최채만 ( Chae Man Choi ),정삼주 ( Sam Ju Chung ),박애숙 ( Ae Sook Park ),김현정 ( Hyun Jung Kim ),김정헌 ( Jung Hun Kim ),정권 ( Kwon Jung ) 대한화장품학회 2014 대한화장품학회지 Vol.40 No.4

2010년 1월에서 2012년 12월까지 서울특별시 보건환경연구원에 품질검사가 의뢰된 화장품 9,879건에 대해, 국산과 수입화장품에 대한 집계자료를 바탕으로 제조국가별, 연도별, 적용부위별로 검사의뢰 유형을 비교하여 품질 및 안전성관리에 대한 기초자료를 제공하고자 본 조사를 시행하였다. 전체 화장품 중 국산은 645건 (6.5%), 수입산은 9,234건(93.5%)이며, 제조국가별로는 프랑스 4,342건(44.0%), 독일 1,637건(16.6%), 미국 1,476건(14.9%), 한국 645건(6.5%), 이태리 557건(5.6%), 기타 1,222건(12.4%)이었다. 또한 연도별 위수탁 화장품 검사의뢰건수는 2010년 3,784건, 2011년 3,394건, 2012년 2,701건으로 나타나 일반 화장품은 감소하고 기능성 화장품이나 염모제에 대한 품질 검사가 증가되었다. 화장품 유형별로는 기초제품 5,470건 (55.4%), 색조 1,908건(19.3%), 손발관리 1,026건(10.4%), 두발관리 616건(6.2%), 목욕용이 361건 (3.7%), 기타 498건(5.0%)이며 국산화장품의 유형별 분포는 기초 > 두발관리 > 색조 > 손발관리 > 목욕용 순이나 수입화장품에서는 기초 > 색조 > 손발관리 > 두발관리 > 목욕용 제품의 순서로 나타났다. 국제적 품질관리기준의 우선순위를 설정하기 위해 국내 소비자들의 경향부터 직간접적으로 파악하는 일이 필요하다. 위수탁검사의뢰된 국산 및 수입화장품의 제품 유형과 인체 적용 부위별 비율 등을 비교 활용하여 미래지향적인 화장품 안전관리의 기초자료로 활용하고자 한다. This study is aimed to provide the primary data about safety of cosmetics products using indirect preference of korean cosmetics customer and numerical comparison of applied area. For this study, we collected 9,879 cosmetics products which were inspected in cosmetics research team from January, 2010 to December, 2012. The domestic cosmetics was 645 cases (6.5%) and Imported cosmetics was 9,234 cases (93.5%). As manufacturing country, the France has 4,342 cases (44.0%) and the next ranking were like those, Germany 1,637 cases (16.6%), U.S.A 1,476 cases (14.9%), Republic of Korea 645 cases (6.5%), Italy 557 cases (5.6%), and etc 1,222 cases (12.4%). By the year, the cases of test cosmetics have decreased from 3,784 cases (2010), 3,394 cases (2011) to 2,701 cases (2012), the relative ratio of common cosmetics part was drop in but the other group (functional cosmetics and hair dye related products) was increased. The largest market share product was Skin care 5,470 cases (55.4%) and the next order was like those, Make up 1,908 cases (19.3%), Hand & Foot 1,026 cases (10.4%), Hair Care 616 cases (6.2%), Bath 361 cases (3.7%), and etc 498 cases (5.0%). In domestic cosmetics, the greatest proportion was Skin care and the others were Hair Care > Makeup > Hand & Foot > Bath, but the proportion was evidently changed in imported cosmetics, Skin care > Makeup > Hand & Foot > Hair Care > Bath. It is necessary to set the priority of the international quality standards to identify trends from domestic consumers directly or indirectly. Compare the ratio of category and human application parts from domestic and imported cosmetics, we utilize leverage as the basis for future-oriented cosmetic safety.

발효주정 첨가 저염 미더덕(Styela clava) 양념젓갈의 제조 및 품질

황영숙 ( Young-sook Hwang ),이현진 ( Hyun-jin Lee ),황석민 ( Seok-min Hwang ),오광수 ( Kwang-soo Oh ) 한국수산과학회(구 한국수산학회) 2021 한국수산과학회지 Vol.54 No.1

In order to develop value-added low-salt fermented seafood with a long shelf-life, we prepared seasoned low-salt fermented Mideoduck (Styela clava) supplemented with fermentation alcohol (SME). The SME was produced by washing and dewatering shelled Mideoduck, followed by cutting and salting for 24 h at 0°C. The salted Mideoduck was seasoned and fermented with ingredients, including garlic, ginger, monosodium glutamate, red pepper, sesame, sorbitol and sugar, for 7-8 days at 0°C. After adding 3-5% fermentation alcohol, the Mideoduck was packed in a polyester container. The salinity, volatile basic nitrogen, and amino nitrogen content of the SME was 4.5%, 20.9 mg/100 g and 92.0 mg/100 g, respectively. In comparison with the control, the addition of 3-5% fermentation alcohol showed inhibitory effects of decreased freshness, texture degradation, and growth of residual bacteria. Additionally, the SME had good storage stability and organoleptic qualities when stored at 4±1°C for 40 days. Therefore, it is suitable for commercialization as a seasoned low-salt fermented product with a long shelf-life. The total amino acid content of the SME was 11,774.5 mg/100 g, majorly comprising glutamic acid, aspartic acid, lysine, arginine, and leucine, and the free amino acid content was 506.4 mg/100 g, majorly comprising hydroxyproline, taurine, and glutamic acid.

자질집합선택 기반의 기계학습을 통한 한국어 기본구 인식의 성능향상

황영숙(Young-Sook Hwang),정후중(Hoojung Chung),박소영(So-Young Park),곽용재(Young-Jae Kwak),임해창(Hae-Chang Rim) 한국정보과학회 2002 정보과학회논문지 : 소프트웨어 및 응용 Vol.29 No.9·10

본 연구에서는 기계학습을 이용하여 한국어 기본구(base phrase)인식의 성능을 향상시키고자 할 때, 학습집합으로부터 획득 가능한 자질집합들 중 최적의 자질집합이 무엇이며, 자료부족 문제를 어떻게 완화할 것인가에 대해 논한다. 먼저 최적의 자질집합 선택은 “점증적 유용성“이란 관점에서 자질의 적합성을 정의하고 이러한 정의에 따라 자질집합을 선택한다. 그리고, 자료부족 문제 완화의 해결점을 찾기 위해 한국어의 통사적 특성을 고려한 형태소 품사체계 사용 및 선택적 어휘자질의 사용이 성능에 미치는 영향을 분석하고 결과를 제시한다. 다양한 크기의 문맥 및 속성, 품사체계에 따라 자질 집합을 구성하고, 서로 다른 특성을 갖는 학습기법 결정트리와 메모리기반 학습기법을 적용한 결과, 한국어 기본구 인식에 유용한 자질은 품사, 어휘, 그리고 기본구 태그로, 두 학습 알고리즘 모두 동일하였다. 또한 한국어의 특성을 고려한 일반화된 품사체계 및 선택적 어휘자질의 사용이 자료부족 문제를 완화시켜주면서 안정된 성능을 보여주었다. 선택된 최적의 자질집합을 사용하여 결정트리와 메모리 기반 학습을 수행한 결과, 전체 기본구에 대해 각각 93.39%/93.41%, 90.99%/92.52%의 정확률/재현율을 얻었다. In this paper, we present an empirical study for improving the Korean text chunking based on machine learning and feature set selection approaches. We focus on two issues: the problem of selecting feature set for Korean chunking, and the problem of alleviating the data sparseness. To select a proper feature set, we use a heuristic method of searching through the space of feature sets using the estimated performance from a machine learning algorithm as a measure of "incremental usefulness" of a particular feature set. Besides, for smoothing the data sparseness, we suggest a method of using a general part-of-speech tag set and selective lexical information under the consideration of Korean language characteristics. Experimental results showed that chunk tags and lexical information within a given context window are important features and spacing unit information is less important than others, which are independent on the machine learning techniques. Furthermore, using the selective lexical information gives not only a smoothing effect but also the reduction of the feature space than using all of lexical information. Korean text chunking based on the memory-based learning and the decision tree learning with the selected feature space showed the performance of precision/recall of 90.99%/92.52%, and 93.39%/93.41% respectively.

회원 학술보고 : 포스터 발표 / 21. 이온 교환 수지를 이용한 피라세탐의 방출조절

황영숙(Young Sook Hwang),서성훈(Sung Hoon Seo) 한국약제학회 1994 한국약제학회 총회 및 학술대회 요지집 Vol.- No.24

질의 응답 시스템을 위한 가변 길이 단락 검색

이영신(Young-Shin Lee),황영숙(Young-Sook Hwang),임해창(Hae-Chang Rim) 한국정보과학회 언어공학연구회 2002 한국정보과학회 언어공학연구회 학술발표 논문집 Vol.2002 No.10

한국어 구문분석의 효율성을 개선하기 위한 구문제약규칙의 학습

박소영(So-Young Park),곽용재(Yong-Jae Kwak),정후중(Hoo-Jung Chung),황영숙(Young-Sook Hwang),임해창(Hae-Chang Rim) 한국정보과학회 2002 정보과학회논문지 : 소프트웨어 및 응용 Vol.29 No.9·10

본 논문에서는 한국어 구문분석에 적합한 다양한 구문정보에 대해 살펴보고, 이를 바탕으로 학습한 제약규칙을 이용하여 구문분석모델의 효율성을 개선시키는 방법을 제안한다. 제안하는 방법의 특징은 다음과 같다. 첫째, 제약규칙을 이용하여 불필요한 중간결과물의 생성을 제약하므로, 구문분석의 효율성이 향상된다. 둘째, 제약규칙의 학습에 이용되는 구문정보가 한국어의 특성을 적절히 반영하고 있으므로, 한국어 문장에 대해 비교적 견고하게 분석할 수 있다. 셋째, 제약규칙은 결정트리 학습알고리즘에 의해 말뭉치에서 자동으로 학습되므로, 제약규칙의 획득이 용이하다. 제약규칙을 이용하여 실험한 결과 구문분석모델의 과생성이 1/2~1/3로 줄고 처리속도가 2~3배 빨라졌다. In this paper, we observe various syntactic information for Korean parsing and propose a method to learn constraints and improve the efficiency of a parsing model by using the constraints. The proposed method has the following three characteristics. First, it improves the parsing efficiency since we use constraints that can prevent the parser from generating unsuitable candidates. Second, it is robust on a given Korean sentence because the attributes for the constraints are selected based on the syntactic and lexical idiosyncrasy of Korean. Third, it is easy to acquire constraints automatically from a treebank by using a decision tree learning algorithm. The experimental results show that the parser using acquired constraints can reduce the number of overgenerated candidates up to 1/2~1/3 of candidates and it runs 2~3 times faster than the one without any constraints.