한국어 형용사 의미계층의 전산적 추출

송상헌(Song Sanghoun),최재웅(Jae-Woong Choe) 한국정보과학회 언어공학연구회 2006 한국정보과학회 언어공학연구회 학술발표 논문집 Vol.2006 No.10

자연 언어의 각 어휘는 서로 관계를 가지고 계층적ㆍ입체적 모델로 존재한다. 이러한 전제에서 출발한 연구 가운데 대표적인 것이 의미 계층이다. 본고에서는 한국어 형용사의 의미 계층을 추출하는 것을 목표로 하여, 형식적ㆍ객관적 방법론을 정립하고, 결과를 비교적 신속하고 정확하게 이끌어 낼 수 있는 전산적 처리 도입하였다. 우선 전체 구축에 필요한 절차를 세우고 각 단계에서 필요한 방법과 휴리스틱을 정리하였다. 이를 바탕으로 사전 뜻풀이 말을 이용하여 반자동으로 작업하였으며, 일부 코퍼스를 활용하였다. 최종 알고리즘으로는 Top-Down 방식을 택하였다. 이렇게 추출된 한국어 형용사 의미 계층은 226개의 최상위어에서 시작하여 총 3,792개의 표제어를 망라한다. 또한 수직적 계열 관계만을 명시했을 경우 나타날 수 있는 한계를 보완하기 위해, 동의어ㆍ반의어와 같은 수평적 의미 관계와 공기 명사와 같은 결합 관계 등을 함께 기술하였다. 한편 표제항을 뜻풀이말의 공기 명사를 이용하여 의미별로 분류하고, 각 분류마다 별도의 의미 계층을 수립하였다.

딥러닝-워드임베딩을 기반으로 한 말더듬 대상자의 읽기과제 비유창성 분석

송상헌(Sanghoun Song),전희정(HeeCheong Chon),이수복(Soo Bok Lee) 한국언어청각임상학회 2020 Communication Sciences and Disorders Vol.25 No.3

배경 및 목적: 최근의 딥러닝 자연어 처리는 언어단위를 수치 벡터로 변환하여 공간상에서 연산을 도모하는 임베딩 기술을 활용한다. 본 연구는 이 기법을 언어병리학 유창성장애 데이터에 적용하여 비유창성의 위치와 분포 특성을 파악하고자 하였다. 방법: 110명의 중학생 이상 청소년 및 말더듬 성인의 읽기발화(800음절) 데이터를 음소 단위로 분절한 뒤 수치 벡터로 변환하여 거리 연산을 수행하였다. Word2Vec을 활용하여 코사인 유사도를 측정하여 각 비유창성 유형 별 유사성을 도출하고, 또한 전체 데이터를 t-SNE 그림으로 모델을 시각화하여 제시하였다. 또한, 음소 환경을 분석하고자 파라다이스-유창성검사-II의 읽기발화와 세종 코퍼스 데이터를 비교 분석하였다. 결과: 첫 번째, 총 8개의 ND 유형(‘UR’, ‘I’, ‘H’, ‘R1’)과 AD 유형(‘URa’, ‘Ia’, ‘Ha’, ‘R1a’)은 .9 이상의 유사도로 근접하여 출현하였다. 두 번째, ND 유형과 AD 유형 간의 분포적 차이를 분석한 결과, 분포적 차이가 크지 않은 것으로 나타났다. 또한, AD 유형은 상당히 높은 유사도로 AD 유형 간에 발생 위치가 중첩되는 것으로 나타났고 ‘R2’와 ‘DP’는 다른 비유창성 유형과 다른 양상을 보였다. 마지막으로, 음소 환경에 따른 비유창성 출현 빈도는 ‘ㅁ’, ‘ㄷ’, ‘ㅈ’ 음소에서 비유창성 빈도가 비교적 높게 나타났다. 논의 및 결론: 본 연구는 딥러닝 기법을 활용하여 비유창성 유형들 간의 군집화 특성, 정상적 비유창성과 비정상적 비유창성 사이의 분포적 차이가 크지 않다는 것, 그리고 비유창성 출현 음소 등을 확인하였고 이러한 결과는 유창성장애 진단과 치료에 활용될 수 있을 것이다. Objectives: Recent natural language processing systems employ embedding techniques, which convert linguistic expressions into numerical vectors in order to measure the geometric distance between expressions. Using skills and focusing on the reading tasks, the present study aims to reveal the distributional properties of disfluencies. Methods: The current work segmented the reading data of 110 adolescents and adults who stutter, transformed the data into a vector space, and then conducted the embedding calculation. Utilizing Word2Vec, the cosine similarity was measured so as to look at how the types of disfluencies were co-related to each other. Results: The eight ND (Normal disfluencies) and AD (Abnormal disfluencies) types, excluding the R2 (Repetition 2) and DP (Disrhythmic Phonation) types, were close to each other with respect to the cosine similarity (>.9). In particular, the AD types such as Ha (Abnormal hesitation), Ia (Abnormal interjection), URa (Abnormal unfinished/revision word), and R1a (Abnormal Repetition1) largely overlapped with each other. R2 and DP showed different distributional properties from other types of disfluencies. The results also indicated that each ND and AD pair seldom differed in their distributional properties. Finally, this study it found that several consonants tended to appear more often when the speakers produced disfluencies. Conclusion: This study draws the distributional patterns of fluency disorders in an automatic way using deep learning skills. The findings are of use for the diagnosis and treatment of the fluency disorders.

영어 "형용사+전치사구" 구문의 의미적 제약 - ICE-GB와 WordNet을 활용한 통계적 검증

송상헌 ( Sanghoun Song ),전지은 ( Jieun Jeon ),최재웅 ( Jae Woong Choe ) 한국외국어대학교 언어연구소 2008 언어와 언어학 Vol.0 No.41

In this paper, we test a working hypothesis that there is a semantic co-occurrence restriction between an adjective and the preposition in its complement in English. In the process, we propose a statistical method to evaluate the semantic constraint. The hypothesis of this study is that if a group of adjectives form a cluster sharing a common meaning, then they tend to co-occur with the same preposition. In order to test the validity of the hypothesis, we make use of two language resources, namely, ICE-GB and WordNet. We conclude that the proposed hypothesis holds true to the degrees of 45~63%, based on our statistical analysis of a list of 57 adjective clusters gathered from the 549 adjectives extracted from the one million word corpus of ICE-GB.

챗GPT가 유발한 교육 환경의 변화 1년 - 회고와 전망

송상헌 ( Sanghoun Song ) 중앙대학교 인문콘텐츠연구소 2023 인공지능인문학연구 Vol.15 No.-

챗GPT 출시 이후 지난 1년간 교육 현장에서 챗GPT에 대한 무분별한 기대와 우려가 공존하였지만, 실제로 교육이 크게 퇴보하거나 개선된 것은 아니다. 이러한 관찰을 토대로 본고는 챗GPT를 둘러싼 과장된 기대와 우려가 교육 현장에서 AI를 둘러싼 잘못된 오해를 낳고 있음을 지적한다. 오히려 챗GPT를 비롯한 AI 기술을 유연하게 활용할 경우 교육 방법론을 향상시킬 수 있는 가능성이 커질 수 있다. 나아가 본고는 교육 현장에서 챗GPT를 포함한 AI 기술을 통합하는 방법에 대한 실질적인 대안을 제시한다. AI 리터러시의 핵심 교과로 코딩 교육의 중요성을 강조하며, AI를 비판적으로 이해하고 책임감 있게 사용할 수 있도록 AI 윤리 교육이 필수적으로 병행되어야 함을 지적한다. 본고가 AI 기술의 교육적 적용에 대한 깊이 있는 논의를 촉진할 수 있기를 기대한다. Since the launch of ChatGPT, there have been a variety of unwanted expectations. However, its effect on education is neither notably favorable nor detrimental. Misconceptions regarding the use of AI in education stem from inflated expectations and anxieties about ChatGPT. Instead, the previous year‘s experience convinced us that teaching approaches could be enhanced by the intelligent application of AI tools such as ChatGPT. This study also offers a feasible solution for incorporating ChatGPT and other AI tools into instructional strategies. This study emphasizes the importance of programming skills as an essential component of AI literacy instruction. In addition, this study contends that educating students on AI ethics will foster critical thinking and widespread AI application. Thus, the purpose of this study is to initiate a detailed discussion on the application of AI in educational environments.

Measuring the Degree of Convergence between Informal and Formal Acceptability Judgments for Korean

Sanghoun Song(송상헌),Sang-Geun Lee(이상근),Jae-Woong Choe(최재웅),Eunjeong Oh(오은정) 한국언어학회 2017 언어 Vol.42 No.4

The present study attempts to quantify the extent to which the acceptability judgments of naïve native speakers coincide with those of linguists in the domain of Korean syntax. We excerpted 1,554 pairs from articles published in a theoretical syntax journal on Korean, Studies in Generative Grammar, from 1991 to 2014, and then randomly chose 287 pairs. Employing PsychoPy as the toolkit and a five-point Likert scale as the task, we collected 68,388 data points from 328 naïve participants. Two statistical analyses, a one-tailed T-test and a κ-test, were then run on the Z-score transformed data points. The T-test, which determines the directionality of difference, yielded a convergence rate of 88.15% between the two types of judgments. The κ-test, which estimates both the size and directionality of difference, obtained a convergence rate of 82.93%. 16 pairs that passed the T-test in a measure of directionality failed to pass the κ-test when testing both directionality and size of difference by degree of match/mismatch between the two types of judgments. Such consideration leads us to believe that the κ-test is more appropriate for detecting the size of the difference between the two types of judgments, as well as the directionality of the difference. We also classified convergence results into 4 patterns: (i) convergence for both sentences in the pair, (ii) divergence for only the bad sentences in the pair, (iii) divergence for only the good sentences in the pair, and (iv) divergence for both sentences in the pair. These patterns play a role in interpreting different convergence rates obtained in the study. We discuss the implications of the κ-test convergence rate for our research question and the use of κ-tests and patterns in experimental syntax studies.

Language Modeling of English Texts Written by Non-native Speakers

Sanghoun Song(송상헌) 한국사전학회 2017 한국사전학회 학술대회 발표논문집 Vol.2017 No.2

Embedding Calculus with Nonword Properties Improves Word Sense Disambiguation

Seongtae Kim(김성태),Sanghoun Song(송상헌) 한국언어학회 2021 언어 Vol.46 No.2

The present study concerns word sense disambiguation in neural language models using the diagnostic classifiers and hierarchical lexical network. First, we conducted an experiment to see whether the neural models are capable of detecting ambiguous nouns and how they do so. Secondly, we carried out an experiment to verify whether the neural models can identify a specific sense of a lexeme and how they do so. For these experiments, we made use of Word2Vec and FastText as the fixed embedding models and BERT as the contextualized model. In addition, we examined the uniformed and weighted sum method by adding nonword properties (senses). In the case of ambiguity detection, BERT with the general embedding showed better performance than the other models. In regards to sense class detection, BERT with nonword properties showed the best performance on lexemes with numerous senses.

코퍼스 – 실험 – 딥러닝 연구방법론 비교분석: ‘-도록’ 통제 구문을 중심으로

강다은(Daeun Kang),송상헌(Sanghoun Song) 한국중원언어학회 2022 언어학연구 Vol.- No.62

This study aims to examine how convergent results are showing on specific language phenomenon, by using methodological pluralism. Focusing on the ‘-tolok’ control construction, we compared the results of three research methodologies: corpus, experiment, deep learning. Previous studies used corpus exploration and language experiment separately or deep learning based on English data. However, it was not sufficiently implemented that comprehensively examining the three methodologies and deep learning analysis using large amount of data based on specific Korean language phenomenon. Accordingly, we demonstrated whether the results of quantitative analysis agree with each other for the ‘-tolok’ control construction using methodological pluralism. Furthermore, the types of Korean ‘control verb’ are classified into two types. This study is significant in showing that different types of methodology can be complement to each other by adding deep learning to the corpus and experimental methods. Additionally, we empirically revealed the necessity of revisiting the using ‘seltukha-’ as a control verb in Korean and presented four verbs that require further study to be classified as control verb, including ‘seltukha-’.

영어자원문법을 활용한 신경망 기계번역의 데이터 증강과 성능 평가

왕규현 ( Wang Guehyun ),송상헌 ( Song Sanghoun ) 서강대학교 언어정보연구소 2021 언어와 정보 사회 Vol.42 No.-

Machine translation commonly involves both analysis and generation across different human languages. This implies that parallel corpora of a large size are essential to create a theoretically reliable and practically robust translation model. However, as is well known, the parallel corpora between Korean and English are insufficient. In this respect, this study expands the data by means of English Resource Grammar (Flickinger 2000) to improve the translation model between the languages. Then, it looks at whether the neural machine translation model performs better with the augmented data. Unfortunately, it turns out the translation models based on augmented data exhibit rather lower BLEU scores. This study further discusses the reason for the unsatisfactory scores and raises the necessity of human evaluation as a next step.

의문 보문소로서의 ‘ㄹ지’: 상위 술어의 (비)진실성과 (비)사실성

강아름 ( Arum Kang ),송상헌 ( Sanghoun Song ) 한국외국어대학교 언어연구소 2021 언어와 언어학 Vol.- No.92

This article concerns the semantico-pragmatic role of the interrogative complementizer (u)lci in Korean. Thus far, compared to (u)nci, which has a variant of interrogative complementizer, (u)lci has received less attention. The main question we address is why (u)lci gives rise to epistemic modal reading (i.e., speaker’s/subject’s uncertainty) in the embedded clause. Given that the core properties of a complementizer can be characterized by the types of main predicates that it takes, the main focus will be the realization of matrix predicates in Korean. For the detailed discussion, we conduct a corpus study and empirically verify that the predicates taking (u)lci should be non-factive and non-veridical. Such a connection is naturally predicated because historically, (u)lci must have been grammaticalized from the combination of the irrealis mood marker (u)l and the bound noun do which leads to the idea that the lexicon of (u)lci in modern Korean has come to have an irrealis mood meaning. Accordingly, there exists a strong tendency to show a gradience between (u)lci and non-veridical predicates.