• Annual Conference on Human and Language Technology
• /
• 1996.10a
• /
• pp.158-162
• /
• 1996

본 논문에서는 문자출력이 가능한 무선호출기에 음성인식 기술을 접목한, 특성화된 한 음성인식 시스템에 대하여 설명하고자 한다. 시스템 동작 과정은, 일단 호출자가 음성인식 서버와 접속하게 되면 서버는 호출자의 자연스런 입력음성을 인식, 그 결과를 문장 형태로 피호출자의 호출기 단말기에 출력시키는 방식으로 되어 있다. 본 시스템에서는 통계적 음성인식 기법을 도입하여, 각 단어를 연속 HMM으로 모델링하였다. 가우시안 혼합 확률밀도함수를 사용하는 각 모델은 전통적인 HMM 학습법들 중의 하나인 Baum-Welch 알고리듬에 의해 학습되고 인식시에는 이들에 비터비 빔 탐색을 적용하여 최선의 결과를 얻도록 한다. MFCC와 파워를 혼용한 26 차원 특징벡터를 각 프레임으로부터 추출하여, 최종적으로, 83 개의 도메인 어휘들 및 무음과 같은 특수어휘들에 대한 모델링을 완성하게 된다. 여기에 구문론적 기능과 의미론적 기능을 함께 수행하는 FSN을 결합시켜 자연발화음성에 대한 연속음성인식 시스템을 구성한다. 본문에서는 이상의 사항들 외에도 음성 데이터베이스, 레이블링 등과 갈이 시스템 성능과 직결되는 시스템의 외적 요소들에 대해 고찰하고, 시스템에 구현되어 있는 다양한 특성들에 대해 밝히며, 실험 결과 및 앞으로의 개선 방향 등에 대해 논의하기로 한다.

• Smart Media Journal
• /
• v.9 no.3
• /
• pp.90-97
• /
• 2020

With the development of the Internet and smartphones, digital sound sources are easily accessible, and accordingly, interest in music search and recommendation is increasing. As a method of recommending music, research using melodies such as pitch, tempo, and beat to classify genres or emotions is being conducted. However, since lyrics are becoming one of the means of expressing human emotions in music, the role of the lyrics is increasing, so a study of emotion classification based on lyrics is needed. Therefore, in this thesis, we analyze the emotions of the farewell lyrics in order to subdivide the farewell emotions based on the lyrics. After constructing an emotion dictionary by vectoriziong the similarity between words appearing in the parting lyrics through Word2Vec learning, we propose a method of classifying parting lyrics emotions using Word2Vec and LSTM, which classify lyrics by similar emotions by learning lyrics using LSTM.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.5
• /
• pp.483-489
• /
• 2020

This paper proposes acoustic model training using self-attention for low-resource speech recognition. In low-resource speech recognition, it is difficult for acoustic model to distinguish certain phones. For example, plosive /d/ and /t/, plosive /g/ and /k/ and affricate /z/ and /ch/. In acoustic model training, the self-attention generates attention weights from the deep neural network model. In this study, these weights handle the similar pronunciation error for low-resource speech recognition. When the proposed method was applied to Time Delay Neural Network-Output gate Projected Gated Recurrent Unit (TNDD-OPGRU)-based acoustic model, the proposed model showed a 5.98 % word error rate. It shows absolute improvement of 0.74 % compared with TDNN-OPGRU model.

• Journal of Digital Convergence
• /
• v.15 no.12
• /
• pp.341-347
• /
• 2017

As the amount of document data increases, the user needs summarized information to understand the document. However, existing document summary research methods rely on overly simple statistics, so there is insufficient research on multiple document summaries for ambiguity of sentences and meaningful sentence generation. In this paper, we investigate semantic connection and preprocessing process to process unnecessary information. Based on the vocabulary semantic pattern information, we propose a multi-document summarization method that enhances semantic connectivity between sentences using VAE. Using sentence word vectors, we reconstruct sentences after learning from compressed information and attribute discriminators generated as latent variables, and semantic connection processing generates a natural summary sentence. Comparing the proposed method with other document summarization methods showed a fine but improved performance, which proved that semantic sentence generation and connectivity can be increased. In the future, we will study how to extend semantic connections by experimenting with various attribute settings.

• Journal of the Korean Data and Information Science Society
• /
• v.27 no.4
• /
• pp.845-854
• /
• 2016

Sentiment analysis or opinion mining is a technique of text mining employed to identify subjective information or opinions of an individual from documents in blogs, reviews, articles, or social networks. In the literature, only a problem of binary classification of ratings based on review texts in an online review. However, because there can be positive or negative reviews as well as neutral reviews, a multi-class classification will be more appropriate than the binary classification. To this end, we consider the multi-class classification of ratings based on review texts. In the preprocessing stage, we extract words related with ratings using chi-square statistic. Then the extracted words are used as input variables to multi-class classifiers such as support vector machines and proportional odds model to compare their predictive performances.

• Journal of Digital Convergence
• /
• v.19 no.9
• /
• pp.181-188
• /
• 2021

Research and importance of legal services applied with AI so that it can be easily understood and predictable in difficult legal fields is increasing. In this study, based on the decision of the Tax Tribunal in the field of tax law, a model was built through self-learning through information collection and data processing, and the prediction results were answered to the user's query and the accuracy was verified. The proposed model collects information on tax decisions and extracts useful data through web crawling, and generates word vectors by applying Word2Vec's Fast Text algorithm to the optimized output through NLP. 11,103 cases of information were collected and classified from 2017 to 2019, and verified with 70% accuracy. It can be useful in various legal systems and prior research to be more efficient application.

• Journal of the Korea Convergence Society
• /
• v.9 no.12
• /
• pp.47-52
• /
• 2018

Named Entity Recognition is a system that extracts entity names such as Persons(PS), Locations(LC), and Organizations(OG) that can have a unique meaning from a document and determines the categories of extracted entity names. Recently, Bi-LSTM-CRF, which is a combination of CRF using the transition probability between output data from LSTM-based Bi-LSTM model considering forward and backward directions of input data, showed excellent performance in the study of object name recognition using deep-learning, and it has a good performance on the efficient embedding vector creation by character and word unit and the model using CNN and LSTM. In this research, we describe the Bi-LSTM-CNN-CRF model that enhances the features of the Korean named entity recognition system and propose a method for constructing the traditional culture corpus. We also present the results of learning the constructed corpus with the feature augmentation model for the recognition of Korean object names.

• Journal of the Korea Convergence Society
• /
• v.10 no.6
• /
• pp.41-47
• /
• 2019

It is very important to search for and obtain an example of a similar judgment in case of court judgment. The existing judge's document search uses a method of searching through key-words entered by the user. However, if it is necessary to input an accurate keyword and the keyword is unknown, it is impossible to search for the necessary document. In addition, the detected document may have different contents. In this paper, we want to improve the effectiveness of the method of vectorizing a document into a three-dimensional space, calculating cosine similarity, and searching close documents in order to search an accurate judge's example. Therefore, after analyzing the similarity of words used in the judge's example, a method is provided for extracting the mode and inserting it into the text of the text, thereby providing a method for improving the cosine similarity of the document to be retrieved. It is hoped that users will be able to provide a fast, accurate search trying to find an example of a tax-related judge through the proposed model.

• Phonetics and Speech Sciences
• /
• v.11 no.3
• /
• pp.39-47
• /
• 2019

We propose a method to improve the performance of spontaneous speech recognizers by extending their phone set using speech data. In the proposed method, we first extract variable-length phoneme-level segments from broadcast speech signals, and convert them to fixed-length latent vectors using an long short-term memory (LSTM) classifier. We then cluster acoustically similar latent vectors and build a new phone set by choosing the number of clusters with the lowest Davies-Bouldin index. We also update the lexicon of the speech recognizer by choosing the pronunciation sequence of each word with the highest conditional probability. In order to analyze the acoustic characteristics of the new phone set, we visualize its spectral patterns and segment duration. Through speech recognition experiments using a larger training data set than our own previous work, we confirm that the new phone set yields better performance than the conventional phoneme-based and grapheme-based units in both spontaneous speech recognition and read speech recognition.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.23 no.3
• /
• pp.237-246
• /
• 2019

In order to improve the accuracy of the rating prediction of the recommendation model, not only user-item rating data are used but also consider auxiliary information of item such as comments, tags, or descriptions. The traditional approaches use a word-level model of the bag-of-words for the auxiliary information. This model, however, cannot utilize the auxiliary information effectively, which leads to shallow understanding of auxiliary information. Convolution neural network (CNN) can capture and extract feature vector from auxiliary information effectively. Thus, this paper proposes character-level deep-Convolution Neural Network based matrix factorization (Char-DCNN-MF) that integrates deep CNN into matrix factorization for a novel recommendation model. Char-DCNN-MF can deeper understand auxiliary information and further enhance recommendation performance. Experiments are performed on three different real data sets, and the results show that Char-DCNN-MF performs significantly better than other comparative models.