• Annual Conference of KIPS
• /
• 2019.10a
• /
• pp.1026-1029
• /
• 2019

최근 제안되고 있는 Bert 등의 딥러닝 언어 모델 기반 pre-training 기법은 다양한 NLP 분야에서 활용되고 있다. 텍스트로 작성된 데이터 셋을 딥러닝 언어 모델이 학습하기 위해서는 토크나이징(tokenizing) 기술이 필요하다. 그러나 기존 토크나이징 방식은 한국어 및 한글이 가지는 고유한 특성(교착어적 특성과 모아쓰기 반영)을 반영하기 어렵다는 한계를 가지고 있다. 본 논문에서는 한국어와 한글이 가지는 고유한 특성을 고려하기 위하여 음소 단위의 임베딩 기법을 제안하며, 이를 기반으로 언어 모델을 설계 및 구현한다. 또한 음소 단위 임베딩 기반 한국어 모델이 실제 데이터 집합(구약성서)에서 나타나는 언어적 패턴을 학습할 수 있다는 것을 실험을 통하여 밝힌다.

• Journal of the Korea Society of Computer and Information
• /
• v.24 no.2
• /
• pp.9-16
• /
• 2019

Computer-Assisted Pronunciation Training system is considered to be a useful tool for pronunciation learning for students who received elementary level English pronunciation education, especially for students who have difficulty in correcting their pronunciation in front of others or who are not able to receive face-to-face training. The conventional Computer-Assisted Pronunciation Training system shows the word to the user, the user pronounces the word, and then the system provides phoneme or audio feedback according to the pronunciation of the user. In this paper, we propose a Computer-Assisted Pronunciation Training system that can practice on the varying pronunciation according to positions of adjacent phonemes. To achieve this, the proposed system is implemented by recommending a series of words by focusing on adjacent phonemes for simplicity and clarity. Experimental results showed that word recommendation considering adjacent phonemes leads to improvement of pronunciation accuracy.

• Annual Conference on Human and Language Technology
• /
• 2019.10a
• /
• pp.368-372
• /
• 2019

맞춤법 교정이란 주어진 문장에서 나타나는 철자 및 맞춤법 오류들을 올바르게 교정하는 것을 뜻하며 맞춤법 교정 시스템이란 컴퓨터가 이를 자동으로 수행하는 것을 의미한다. 본 논문에서는 맞춤법 교정을 기계번역의 관점으로 바라보고 문제를 해결하였다. 소스문장에 맞춤법 오류문장, 타겟 문장에 올바른 문장을 넣어 학습시키는 방법을 제안한다. 본 논문에서는 단일 말뭉치로 한국어 맞춤법 병렬 말뭉치를 구성하는 방법을 제안하며 G2P(Grapheme to Phoneme)를 이용한 오류 데이터 생성, 자모 단위 철자 오류데이터 생성, 통번역 데이터 기반 오류 데이터 생성 크게 3가지 방법론을 이용하여 맞춤법 오류데이터를 생성하는 방법론을 제안한다. 실험결과 GLEU 점수 65.98의 성능을 보였으며 44.68, 39.55의 성능을 보인 상용화 시스템보다 우수한 성능을 보였다.

• Korean Journal of Cognitive Science
• /
• v.23 no.1
• /
• pp.1-23
• /
• 2012

This study investigated the lexical access processing of inflected Korean verbs in the mental lexicon. In Korean, verbs can be classified into two main types of inflections, which are regular and irregular inflections, which can be further divided into three types of regular inflections and two types of irregular inflections. A masked priming lexical decision task was used and the priming effects were compared. Experiments were carried out using the five different types of verbal inflections in Korean: (1) No change-regularity (regular verbs with no orthographical or phonological changes), (2) Phonological change-regularity (regular verbs with phonological changes to the stem only), (3) Orthographical change-regularity (regular verbs that only undergo orthographical changes), (4) Stem change-irregularity (the stem is omitted or alternated with the other phoneme of the stem in irregular verbs), (5) Ending change-irregularity (irregular verbs with changes in the endings by phoneme substitution). The first three types are regarded as regular verbal inflections whereas the latter two types are regarded as irregular verbal inflections. The infinitive forms of the verb were presented as target words and three different conditions were presented as prime words. The three conditions included regular verbal inflection, irregular verbal inflection, and a control condition in which morphologically and semantically unrelated primes were presented. In addition, different stimulus onset asynchrony (SOA) were manipulated (43ms, 72ms, 230ms) to examine the time frame of the morphological decomposition process in word recognition. The results revealed that there were significant priming effects in all three SOAs across conditions. Hence, there was no significant differences in priming effects between regular and irregular verbal inflection conditions. This may suggest that Korean verb processing does not adopt different processing routes for regular and irregular inflections, which can also be an indication of earlier morphological information processing for Korean verbs.

• Journal of the Institute of Convergence Signal Processing
• /
• v.4 no.4
• /
• pp.30-39
• /
• 2003

In this paper, we carried out the performance evaluation of HM-Net(Hidden Markov Network) speech recognition system for Korean speech databases. We adopted to construct acoustic models using the HM-Nets modified by HMMs(Hidden Markov Models), which are widely used as the statistical modeling methods. HM-Nets are carried out the state splitting for contextual and temporal domain by PDT-SSS(Phonetic Decision Tree-based Successive State Splitting) algorithm, which is modified the original SSS algorithm. Especially it adopted the phonetic decision tree to effectively express the context information not appear in training speech data on contextual domain state splitting. In case of temporal domain state splitting, to effectively represent information of each phoneme maintenance in the state splitting is carried out, and then the optimal model network of triphone types are constructed by in the parameter. Speech recognition was performed using the one-pass Viterbi beam search algorithm with phone-pair/word-pair grammar for phoneme/word recognition, respectively and using the multi-pass search algorithm with n-gram language models for sentence recognition. The tree-structured lexicon was used in order to decrease the number of nodes by sharing the same prefixes among words. In this paper, the performance evaluation of HM-Net speech recognition system is carried out for various recognition conditions. Through the experiments, we verified that it has very superior recognition performance compared with the previous introduced recognition system.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.5
• /
• pp.530-536
• /
• 2021

Recently, due to the development of deep learning, end-to-end speech recognition, which directly maps graphemes to speech signals, shows good performance. Especially, among the end-to-end models, conformer shows the best performance. However end-to-end models only focuses on the probability of which grapheme will appear at the time. The decoding process uses a greedy search or beam search. This decoding method is easily affected by the final probability output by the model. In addition, the end-to-end models cannot use external pronunciation and language information due to structual problem. Therefore, in this paper conformer with lexicon transducer is proposed. We compare phoneme-based model with lexicon transducer and grapheme-based model with beam search. Test set is consist of words that do not appear in training data. The grapheme-based conformer with beam search shows 3.8 % of CER. The phoneme-based conformer with lexicon transducer shows 3.4 % of CER.

• Proceedings of the KSPS conference
• /
• 1996.10a
• /
• pp.247-251
• /
• 1996

This paper is intended as a preliminary study on phonetic and phonological differences between Polish and Korean languages. In this paper an attempt is made to examine the most conspicious difficulties encountered by Polish learners who begin to speak Korean (and in doing so, 1 would hope that it might be of help to future learners of both languages). Since the phoneme inventory and general phonetic rules for both languages are very different, teaching and learning accurate pronunciation is extremely difficult for both the Poles and Koreans without any previous phonetic training. In the case of Polish and Korean we can see how strong and persistent the influences of the mother-tongue are on the target language. As an example I would like to discuss the basic differences between Polish and Korean consonants. The most important consonantal opposition in Polish is voice-/voicelessness (f. ex.; 〔b〕 / 〔p〕, 〔g〕 / 〔k〕) while in Korean, opposition such as voice-/voicelessness is of secondary importance. Therefore Korean speakers do not perceive the difference between Polish voiced and voiceless consonants. On the other hand, Polish speakers can not distinguish Korean lenis / fortis / aspirated consonants (f. ex.; ㅂ 〔b〕 / ㅃ 〔p〕 / ㅍ〔ph〕, ㄱ 〔g〕 / ㄲ 〔k〕 / ㅋ 〔kh〕)) opposition. The other very important factor is palatalization which is of vital importance in Polish and, because of this, Polish speakers are extremely sensitive to it. In Korean palatalization is not important phonetically and Korean speakers do not distinguish between palatalized and non-palatalized consonants. The transcription used here is based on ' The principles of the International Phonetic Association and the Korean Phonetic Alphabet ' (1981) by Hyun Bok Lee.

• Phonetics and Speech Sciences
• /
• v.1 no.2
• /
• pp.129-140
• /
• 2009

The purpose of the present study is to investigate frequency related information and syllable structure constraints on Sino-Korean. Previous studies on Sino-Korean have mostly investigated the historical change of sounds and reviewed archaic features of Chinese language in Sino-Korean. Unfortunately, there is little study on the sounds of contemporary Sino-Korean in terms of syllable structure constraints. For the purpose of the present study, sounds of 7,742 Chinese characters used in Sino-Korean (7,795 syllables) were investigated and syllable matrices made based on the results of frequency related information. As a result, 483 syllable types were observed and the most frequently observed syllables were as follows: /ku/ (103) > /ki/ (100) > /ju/ (87) > /pi/ (86). Only 16 out of 19 consonants are used for Sino-Korean. /$t^{\ast}$/ and /$p^{\ast}$/ are never used in Sino-Korean and /kh, $s^{\ast}$, $k^{\ast}$/ occur only a few times (3, 2, 1 respectively). /k/ (17.5%) shows the highest frequency and /n, ${\eta}$, 1, tc, m/ occupied the next rankings. Among 20 vowel types, /a/ showed the highest frequency and /o, u, i, $j{\Lambda}$, ${\Lambda}$/ occupied the next rankings. Based on the syllable matrices, gaps were observed and classified into accidental or systematic ones. Onset and nucleus, nucleus and coda, onset and coda, and other syllable structure constraints of Sino-Korean were listed.

• Journal of Korean language education
• /
• v.28 no.1
• /
• pp.251-268
• /
• 2017

This study is based on differences of syllable type between Korean and Chinese language pronunciation. For example, Nasal consonant ending 【n】 and 【${\eta}$】 reside in both Korean and Chinese phonetics simultaneously. However, in experiential training, Chinese learners will make errors in pronunciation of the Korean syllable nasal consonant ending like 【n】 and 【${\eta}$】. In the previous research, analysis of pronunciation errors were often based on the perspective of phonological system and combination of the phoneme rules. However, in this study, the analysis is based on the differences between Korean and Chinese syllables category to indicate the cause of pronunciation errors. The main findings of this study indicated that in the process of pronunciation of Chinese, nasal consonant syllable rime and its 【back】 tongue vowel are combined with each other. However, this rule does not apply in Korean pronunciation. Therefore, the Korean syllabic types like "앤, 응, 옹, 앵, 은, 온, 언" also exist in the Chinese language. When theChinese learners pronounce these types of syllables, the combination of the voweland nasal syllable rime rule will be taken, which will result in pronunciationerrors.

• Phonetics and Speech Sciences
• /
• v.7 no.4
• /
• pp.111-121
• /
• 2015

The present study examines the auditory boundary between Korean /o/ and /u/ on a synthesized vowel continuum by Chinese learners of Korean language. Preceding researches reported that the Chinese learners have difficulty pronouncing Korean monophthongs /o/ and /u/. In this experiment, a nine-step continuum was resynthesized using Praat from a vowel token from a recording of a male announcer who produced it in isolated form. F1 and F2 were synchronously shifted in equal steps in qtone (quarter tone), while F3 and F4 values were held constant for the entire stimuli. A forced choice identification task was performed by the advanced learners who speak Mandarin Chinese as their native language. Their experiment data were compared to a Korean native group. ROC (Receiver Operating Characteristic) analysis and logistic regression were performed to estimate the perceptual boundary. The result indicated the learner group has a different auditory criterion on the continuum from the Korean native group. This suggests that more importance should be placed on hearing and listening training in order to acquire the phoneme categories of the two vowels.