• ETRI Journal
• /
• 제32권1호
• /
• pp.160-162
• /
• 2010

For the acoustic models of embedded speech recognition systems, hidden Markov models (HMMs) are usually quantized and the original full space distributions are represented by combinations of a few quantized distribution prototypes. We propose a maximum likelihood objective function to train the quantized distribution prototypes. The experimental results show that the new training algorithm and the link structure adaptation scheme for the quantized HMMs reduce the word recognition error rate by 20.0%.

• 한국음향학회:학술대회논문집
• /
• 한국음향학회 1994년도 제11회 음성통신 및 신호처리 워크샵 논문집 (SCAS 11권 1호)
• /
• pp.395-398
• /
• 1994

In this paper the word recognition using sub-model based Hidden Markov Model was studied. Phoneme models were composed of 61 phonemes in therms of Korean language pronunciation characteristic. Using this, word model was maded by serial concatenation. But, in case of this phoneme concatenation, the second and the third phoneme of syllable are overlapped in distribution at the same time. So considering this, the method that combines the second and the third phoneme to one model was proposed. And to prevent the increase in number of model, similar phonemes were combined to one, and finially, 57 models were created. In experiment proper model structure of sub-model was searched for, and recognition results were compared. So similar recognition results were maded, and overall recognition rates were increased in case of using parameter tying method.

• 한국소음진동공학회:학술대회논문집
• /
• 한국소음진동공학회 2007년도 춘계학술대회논문집
• /
• pp.1074-1079
• /
• 2007

We will introduce speech/speaker recognition algorithm for the isolated word. In general case of speaker verification, Gaussian Mixture Model (GMM) is used to model the feature vectors of reference speech signals. On the other hand, Dynamic Time Warping (DTW) based template matching technique was proposed for the isolated word recognition in several years ago. We combine these two different concepts in a single method and then implement in a real time speaker/speech recognition system. Using our proposed method, it is guaranteed that a small number of reference speeches (5 or 6 times training) are enough to make reference model to satisfy 90% of recognition performance.

• 전자공학회논문지B
• /
• 제31B권6호
• /
• pp.17-24
• /
• 1994

This paper is a study on Korean word recognition and suggest the method that stabilizes the state-transition in the HMM by applying the `inertia' to the feature vector sequences. In order to reduce the quantized distortion considering probability distribution of input vectors, we used SOFM, an unsupervised learning method, as a vector quantizer, By applying inertia to the feature vector sequences, the overlapping of probability distributions for the response path of each word on the self organizing feature map can be reduced and the state-transition in the Hmm can be Stabilized. In order to evaluate the performance of the method, we carried out experiments for 50 DDD area names. The results showed that applying inertia to the feature vector sequence improved the recognition rate by 7.4% and can make more HMMs available without reducing the recognition rate for the SOFM having the fixed number of neuron.

• 한국음향학회지
• /
• 제13권4호
• /
• pp.40-49
• /
• 1994

본 논문에서는 단어 인식 시스템의 성능 개선을 위하여 다음과 같은 두가지 방법을 제안한다. 첫번째 방법은 VQ 코드간의 천이를 안정화시키기 위하여 음성신호의 특징벡터 시퀀스에 관성을 적용하는 방법이고, 두번째 방법은 이산 HMM 모델에서 인접 프레임 간의 시간 상관성을 고려하기 위하여 VQ 코드의 천이행렬을 출력 심벌의 관측확률에 가중치로 이용하여 새로운 관측확률을 발생하는 방법이다. 특징벡터 시퀀스에 관성을 도입함으로서, SOFM상의 각 단어에 대한 반응경로에서 확률분포가 중첩되는 것을 억제하여 HMM의 상태천이를 안정화 시킬 수 있다. 기존의 이산 HMM에 VQ 코드의 천이행렬을 가중치로 적용함으로써, 특징벡터의 확률분포를 더욱 세분화하고, 특징분포를 적당한 영역으로 제한함으로써 인식시스템의 성능을 개선할 수 있다. 제안한 방법을 평가하기 위하여 50개의 DDD 지역명을 대상으로 인식 실험을 수행하였다. 실험 결과에 의하면, 제안된 방법이 기존의 HMM 모델에 비해 화자종속 실험에서는 $4.2\%$의 인식률 향상과 화자 독립 실험에서는 $12.45\%$의 인식률 향상을 얻을 수 있었다.

• 정보과학회 논문지
• /
• 제43권9호
• /
• pp.1008-1014
• /
• 2016

트위터, 페이스북, 온라인 고객 리뷰 등은 신문기사처럼 정제된 글이 아닌 자유롭게 기술되는 비격식(informal) 텍스트 문서에 속한다. 이러한 비격식 문서에서 일관된 규칙이나 패턴을 찾는 일은 격식(formal) 문서 경우에 비해 용이하지 않기 때문에, 비격식 문서 분석을 위해서는 성능 개선을 위한 추가적인 접근 방법 필요다고 판단된다. 본 연구에서는 대표적 비격식 문서인 트위터 데이터를 열 가지 카테고리로 분류함에 있어 LDA(Latent Dirichlet allocation) 단어 분포를 사용하여 자질(feature)을 교정하고 확장한다. 토픽별로 상위에 랭크된 단어 자질들을 기반으로 다른 단어 자질들을 분해 및 병합하는 방식으로 유용한 자질 집합을 반복적으로 확장시킨다. 이렇게 생성된 자질로 문서 분류를 수행한 결과 자질 확장 이전에 비해 마이크로 평균 F1-score 7.11%p의 성능 개선 효과를 확인할 수 있었다.

• Journal of Information Science Theory and Practice
• /
• 제9권1호
• /
• pp.24-34
• /
• 2021

The popularity of social networking sites (SNS) has facilitated communication between users. The usage of SNS helps users in their daily life in various ways such as sharing of opinions, keeping in touch with old friends, making new friends, and getting information. However, some users misuse SNS to belittle or hurt others using profanities, which is typical in cyberbullying incidents. Thus, in this study, we aim to identify profane words from the ASKfm corpus to analyze the profane word distribution across four different roles involved in cyberbullying based on lexicon dictionary. These four roles are: harasser, victim, bystander that assists the bully, and bystander that defends the victim. Evaluation in this study focused on occurrences of the profane word for each role from the corpus. The top 10 common words used in the corpus are also identified and represented in a graph. Results from the analysis show that these four roles used profane words in their conversation with different weightage and distribution, even though the profane words used are mostly similar. The harasser is the first ranked that used profane words in the conversation compared to other roles. The results can be further explored and considered as a potential feature in a cyberbullying detection model using a machine learning approach. Results in this work will contribute to formulate the suitable representation. It is also useful in modeling a cyberbullying detection model based on the identification of profane word distribution across different cyberbullying roles in social networks for future works.

• 산경연구논집
• /
• 제13권9호
• /
• pp.37-50
• /
• 2022

Purpose: The purpose of this study is to explore and compare research trends in Korea and overseas academic papers on social media marketing, and to present new academic perspectives for the future direction in Korea. Research design, data and methodology: We used English abstract of research paper (Korea's: 1,349, overseas': 5,036) for word frequency analysis, topic modeling, and trend analysis for each topic. Results: The results of word frequency and co-occurrence frequency analysis showed that Korea researches focused on the experiential values of users, and overseas researches focused on platforms and content. Next, 13 topics and 12 topics for Korea and overseas researches were derived from topic modeling. And, trend analysis showed that Korean studies were different from overseas in applying marketing methods to specific industries and they were interested in the short-term performance of social media marketing. Conclusions: We found that the long-term strategies of social media marketing and academic interest in the overall industry will necessary in the future researches. Also, data mining techniques will necessary to generate more general results by quantifying various phenomena in reality. Finally, we expected that continuous and various academic approaches for volatile social media is effective to derive practical implications.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제12권6호
• /
• pp.2470-2491
• /
• 2018

Web-scale open information extraction (Open IE) plays an important role in NLP tasks like acquiring common-sense knowledge, learning selectional preferences and automatic text understanding. A large number of Open IE approaches have been proposed in the last decade, and the majority of these approaches are based on supervised learning or dependency parsing. In this paper, we present a novel method for web scale open information extraction, which employs cosine distance based on Google word vector as the confidence score of the extraction. The proposed method is a purely unsupervised learning algorithm without requiring any hand-labeled training data or dependency parse features. We also present the mathematically rigorous proof for the new method with Bayes Inference and Artificial Neural Network theory. It turns out that the proposed algorithm is equivalent to Maximum Likelihood Estimation of the joint probability distribution over the elements of the candidate extraction. The proof itself also theoretically suggests a typical usage of word vector for other NLP tasks. Experiments show that the distance-based method leads to further improvements over the newly presented Open IE systems on three benchmark datasets, in terms of effectiveness and efficiency.

• Asia pacific journal of information systems
• /
• 제25권3호
• /
• pp.563-578
• /
• 2015

Forecasting daily box-office performance is critical for planning the distribution of marketing resources, and by extension, maximizing profits. For certain movies, the number of viewers increases rapidly at the beginning of their theatrical run, and the increments slow down later. Other movies are not popular in the beginning, but the audience sizes grow rapidly afterward. Thus, the audience attendance of movies grow in different trajectories, which are influenced by various factors including marketing budget, distributors, directors, actors, and word of mouth. In this paper, we propose a method for predicting the daily performance trajectory of running movies based on the hierarchical linear model. More specifically, we focus on the effect of online word of mouth on the shape of the growth curves. We fitted the mean trajectory of the cumulative audience size as a cubic function of time, and allowed the intercept and slope to vary movie-to-movie. Moreover, we fitted the linear slope with a function of online word of mouth predictors to help determine the shape of the trajectories. Finally, we provide performance predictions for individual movies.