• Annual Conference on Human and Language Technology
• /
• 1997.10a
• /
• pp.175-182
• /
• 1997

지금까지의 한글 문자인식 후처리 연구분야에서 미등록어와 비문맥적 오류 문제는 아직까지 잘 해결하지 못하고 있는 문제이다. 본 논문에서는 단어로서 가능한지를 결정하는 기준으로 확률적 음절 결합 정보를 사용하여 형태소 분석 기법만을 사용했을 때 발생할 수 있는 미등록어 문제를 해결하고, 통사적 기능의 어말 어휘를 고려한 문맥 결합 정보를 이용함으로써 다수의 후보 어절 가운데에서 최적의 후보 어절을 선택하는 방법을 제안한다. 제안된 시스템은 인식기에서 내보낸 후보 음절과 학습된 혼동 음절을 조합하여 하나 이상의 후보 어절을 생성하는 모듈과 통계적 언어 정보를 이용하여 최적의 후보 어절을 선정하는 모듈로 구성되었다. 실험은 1000만 원시 코퍼스에서 추출한 음절 결합 정보와 17만 태깅된 코퍼스에서 추출한 어절 결합 정보를 사용하였으며, 실제 인식 결과에 적용한 결과 문자 단위에서는 94.1%의 인식률을 97.4%로, 어절 단위에서는 87.6%를 96.6%로 향상시켰다. 교정률과 오교정률은 각각 문자 단위에서 56%와 0.6%, 어절 단위에서 83.9%와 1.66%를 보였으며, 전체 실험 어절의 3.4%를 차지한 미등록어 중 87.5%를 올바로 인식하는 한편, 전체 오류의 20.3%인 비문맥 오류에 대해서 91.6%를 올바로 교정하는 후처리 성능을 보였다.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.31 no.9C
• /
• pp.853-858
• /
• 2006

TMHMM(Tied Mixture Hidden Markov Model) is an important approach to reduce the number of free parameters in speech recognition. However, this model suffers from a degradation in recognition accuracy due to its GPDF (Gaussian Probability Density Function) clustering error. This paper proposes a clustering algorithm, called HCNN(Homogeneous Centroid Neural network), to cluster acoustic feature vectors in TMHMM. Moreover, the HCNN uses the heterogeneous distance measure to allocate more code vectors in the heterogeneous areas where probability densities of different states overlap each other. When applied to Korean digit isolated word recognition, the HCNN reduces the error rate by 9.39% over CNN clustering, and 14.63% over the traditional K-means clustering.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.659-663
• /
• 2023

이 논문은 한국어 음성인식 성능을 개선하고자 기존 음성인식 과정을 자모단위 음성인식 모델과 신경망 기반 자모 병합 모델 총 두 단계로 구성하였다. 한국어는 조합어 특성상 음성 인식에 필요한 음절 단위가 약 2900자에 이른다. 이는 학습 데이터셋에 자주 등장하지 않는 음절에 대해서 음성인식 성능을 저하시키고, 학습 비용을 높이는 단점이 있다. 이를 개선하고자 음절 단위의 인식이 아닌 51가지 자모 단위(ㄱ-ㅎ, ㅏ-ㅞ)의 음성인식을 수행한 후 자모 단위 인식 결과를 음절단위의 한글로 병합하는 과정을 수행할 수 있다[1]. 자모단위 인식결과는 초성, 중성, 종성을 고려하면 규칙 기반의 병합이 가능하다. 하지만 음성인식 결과에 잘못인식된 자모가 포함되어 있다면 최종 병합 결과에 오류를 생성하고 만다. 이를 해결하고자 신경망 기반의 자모 병합 모델을 제시한다. 자모 병합 모델은 분리되어 있는 자모단위의 입력을 완성된 한글 문장으로 변환하는 작업을 수행하고, 이 과정에서 음성인식 결과로 잘못인식된 자모에 대해서도 올바른 한글 문장으로 변환하는 오류 수정이 가능하다. 본 연구는 한국어 음성인식 말뭉치 KsponSpeech를 활용하여 실험을 진행하였고, 음성인식 모델로 Wav2Vec2.0 모델을 활용하였다. 기존 규칙 기반의 자모 병합 방법에 비해 제시하는 자모 병합 모델이 상대적 음절단위오류율(Character Error Rate, CER) 17.2% 와 단어단위오류율(Word Error Rate, WER) 13.1% 향상을 확인할 수 있었다.

• Journal of Intelligence and Information Systems
• /
• v.10 no.2
• /
• pp.65-77
• /
• 2004

This paper presents a priori and the local font classification method. The font classification uses ascenders, descenders, and serifs extracted from a word image. The gradient features of those sub-images are extracted, and used as an input to a neural network classifier to produce font classification results. The font classification determines 2 font styles (upright or slant), 3 font groups (serif sans-serif or typewriter), and 7-font names (Postscript fonts such as Avant Garde, Helvetica, Bookman, New Century Schoolbook, Palatine, Times, and Courier). The proposed a priori and local font classification method allows an OCR system consisting of various font-specific character segmentation tools and various mono-font character recognizers. Experiments have shown font classification accuracies reach high performance levels of about 95.4 percent even with severely touching characters. The technique developed for tile selected 7 fonts in this paper can be applied to any other fonts.

• Journal of Korea Multimedia Society
• /
• v.8 no.1
• /
• pp.111-118
• /
• 2005

This paper describes an implementation of the spam mail prevention system using reply message with secrete words. When user receives a new e-mail, the e-mail address is compared with the white e-mail addresses in database by the system. If user receives a new e-mail which does not exist in a white e-mail addresses database, a reply e-mail attached with secrete words is delivered automatically. And the system is compared with the white domains first for intranet environment. It speeds up processing time. proposed algorithm is required a small database and faster than the black e-mail addresses comparison. This system is implemented using procmail, PHP and IMAP on Linux and the user can manage the databases on the web.

• Journal of KIISE:Software and Applications
• /
• v.34 no.9
• /
• pp.804-816
• /
• 2007

Sample training data for machine learning often contain irrelevant information or redundant concept. It is also the case that the original data may include noise. If the information collected for constructing learning model is not reliable, it is difficult to obtain accurate information. So the system attempts to find relations or regulations between features and categories in the teaming phase. The feature selection is to remove irrelevant or redundant information before constructing teaming model. for improving its performance. Existing feature selection methods assume that the distribution of documents is balanced in terms of the number of documents for each class and the length of each document. In practice, however, it is difficult not only to prepare a set of documents with almost equal length, but also to define a number of classes with fixed number of document elements. In this paper, we propose a new feature selection method that considers the impurities among the words and unbalanced distribution of documents in categories. We could obtain feature candidates using the word impurity and eventually select the features through unbalanced distribution of documents. We demonstrate that our method performs better than other existing methods via some experiments.

• Journal of Convergence for Information Technology
• /
• v.10 no.10
• /
• pp.40-46
• /
• 2020

With the development of the Internet, users share their experiences and opinions. Since related keywords are used witho0ut considering information such as the general emotion or genre of an unstructured document such as a movie review, the sensitivity accuracy according to the appropriate emotional situation is impaired. Therefore, we propose a system that predicts emotions based on information such as the genre to which the unstructured document created by users belongs or overall emotions. First, representative keyword related to emotion sets such as Joy, Anger, Fear, and Sadness are extracted from the unstructured document, and the normalized weights of the emotional feature words and information of the unstructured document are trained in a system that combines CNN and LSTM as a training set. Finally, by testing the refined words extracted through movie information, morpheme analyzer and n-gram, emoticons, and emojis, it was shown that the accuracy of emotion prediction using emotions and F-measure were improved. The proposed prediction system can predict sentiment appropriately according to the situation by avoiding the error of judging negative due to the use of sad words in sad movies and scary words in horror movies.

• KIPS Transactions on Software and Data Engineering
• /
• v.6 no.5
• /
• pp.267-278
• /
• 2017

As the same letter can be pronounced differently depending on word contexts, one should refer to a lexicon in order to pronounce a word correctly. Phonetic alphabets that lexicons adopt as well as pronunciations that lexicons describe for the same word can be different from lexicon to lexicon. In this paper, we use a sequence-to-sequence model that is widely used in deep learning research area in order to convert automatically from one pronunciation to another. The 12 seq2seq models are implemented based on pronunciation training data collected from 4 different lexicons. The exact accuracy of the models ranges from 74.5% to 89.6%. The aim of this study is the following two things. One is to comprehend a property of phonetic alphabets and pronunciations used in various lexicons. The other is to understand characteristics of seq2seq models by analyzing an error.

• Journal of the Korean Society for Computer Game
• /
• v.31 no.4
• /
• pp.157-166
• /
• 2018

Communication between customers and service providers is an important issue at sites where various businesses and transactions take place. In particular, the ability to solve problems quickly and accurately when a problem arises and when an inquiry is received is directly linked to trust in the site. In this paper, we propose a method of handling complaints and inquiries of site users by using chatbot technology on talent market platform site. First, we implemented chatbot that can communicate with the inquirers in real time, so that users can use the site usage and word search functions. For various errors and problems of the site which can not be defined by a few words or sentences, I have specified an error code and database it. Users of the site were able to contact chatbot with the error code that was output when an error occurred and get the corresponding response in real time. The chatbot implemented in this study provided a satisfactory experience because that was able to provide quick and accurate answers to users who experienced errors or inquiries when using the site. This will have a positive impact on the credibility and favorability of the site over the long term, and will help reduce manpower and time costs for error inquiries.

• The KIPS Transactions:PartB
• /
• v.18B no.6
• /
• pp.405-414
• /
• 2011

Korean grammar checkers typically detect context-dependent errors by employing heuristic rules that are manually formulated by a language expert. These rules are appended each time a new error pattern is detected. However, such grammar checkers are not consistent. In order to resolve this shortcoming, we propose new method for generalizing error decision rules to detect the above errors. For this purpose, we use an existing thesaurus KorLex, which is the Korean version of Princeton WordNet. KorLex has hierarchical word senses for nouns, but does not contain any information about the relationships between cases in a sentence. Through the Tree Cut Model and the MDL(minimum description length) model based on information theory, we extract noun classes from KorLex and generalize error decision rules from these noun classes. In order to verify the accuracy of the new method in an experiment, we extracted nouns used as an object of the four predicates usually confused from a large corpus, and subsequently extracted noun classes from these nouns. We found that the number of error decision rules generalized from these noun classes has decreased to about 64.8%. In conclusion, the precision of our grammar checker exceeds that of conventional ones by 6.2%.