• Proceedings of the Acoustical Society of Korea Conference
• /
• spring
• /
• pp.33-36
• /
• 2002

We implemented a robust Korean broadcast news transcription system for out-of-vocabulary (OOV), tested its performance. The occurrence of OOV words in the input speech is inevitable in large vocabulary continuous speech recognition (LVCSR). The known vocabulary will never be complete due to the existence of for instance neologisms, proper names, and compounds in some languages. The fixed vocabulary and language model of LVCSR system directly face with these OOV words. Therefore our Broadcast news recognition system has an offline OOV update module of language model and vocabulary to solve OOV problem and selects morpheme-based recognition unit (so called, pseudo-morpheme) for OOV robustness.

• Smart Media Journal
• /
• v.10 no.1
• /
• pp.16-24
• /
• 2021

The exiting word embedding methodology such as word2vec represents words, which only occur in the raw training corpus, as a fixed-length vector into a continuous vector space, so when mapping the words incorporated in the raw training corpus into a fixed-length vector in morphologically rich language, out-of-vocabulary (OOV) problem often happens. Even for sentence embedding, when representing the meaning of a sentence as a fixed-length vector by synthesizing word vectors constituting a sentence, OOV words make it challenging to meaningfully represent a sentence into a fixed-length vector. In particular, since the agglutinative language, the Korean has a morphological characteristic to integrate lexical morpheme and grammatical morpheme, handling OOV words is an important factor in improving performance. In this paper, we propose parallel Tri-LSTM sentence embedding that is robust to the OOV problem by extending utilizing the morphological information of words into sentence-level. As a result of the sentiment analysis task with corpus in Korean, we empirically found that the character unit is better than the morpheme unit as an embedding unit for Korean sentence embedding. We achieved 86.17% accuracy on the sentiment analysis task with the parallel bidirectional Tri-LSTM sentence encoder.

• The Journal of Information Systems
• /
• v.29 no.3
• /
• pp.103-117
• /
• 2020

Purpose Recently, various issues related to toxic comments on web portal sites and SNS are becoming a major social problem. Toxic comments can threaten Internet users in the type of defamation, personal attacks, and invasion of privacy. Over past few years, academia and industry have been conducting research in various ways to solve this problem. The purpose of this study is to develop the deep learning modeling for toxic comments classification. Design/methodology/approach This study analyzed 7,878 internet news comments through CNN classification modeling based on Highway Network and OOV process. Findings The bias and hate expressions of toxic comments were classified into three classes, and achieved 67.49% of the weighted f1 score. In terms of weighted f1 score performance level, this was superior to approximate 50~60% of the previous studies.

• The Journal of the Acoustical Society of Korea
• /
• v.21 no.1E
• /
• pp.3-11
• /
• 2002

Broadcast news transcription is one of the hardest tasks in speech recognition because broadcast speech signals have much variability in speech quality, channel and background conditions. We developed a Korean broadcast news speech recognizer. We used a morpheme-based dictionary and a language model to reduce the out-of·vocabulary (OOV) rate. We concatenated the original morpheme pairs of short length or high frequency in order to reduce insertion and deletion errors due to short morphemes. We used a lexicon with multiple pronunciations to reflect inter-morpheme pronunciation variations without severe modification of the search tree. By using the merged morpheme as recognition units, we achieved the OOV rate of 1.7% comparable to European languages with 64k vocabulary. We implemented a hidden Markov model-based recognizer with vocal tract length normalization and online speaker adaptation by maximum likelihood linear regression. Experimental results showed that the recognizer yielded 21.8% morpheme error rate for anchor speech and 31.6% for mostly noisy reporter speech.

• Proceedings of the KSPS conference
• /
• 2005.04a
• /
• pp.95-98
• /
• 2005

Many utterance verification (UV) algorithms have been studied to reject out-of-vocabulary (OOV) in speech recognition systems. Most of conventional confidence measures for UV algorithms are mainly based on log likelihood ratio test, but these measures take much time to evaluate the alternative hypothesis or anti-model likelihood. We propose a novel confidence measure which makes use of a momentary best scored state sequence during Viterbi search. Our approach is more efficient than conventional LRT-based algorithms because it does not need to build anti-model or to calculate the alternative hypothesis. The proposed confidence measure shows better performance in additive noise-corrupted speech as well as clean speech.

• Journal of Multimedia Information System
• /
• v.9 no.2
• /
• pp.93-102
• /
• 2022

Dependency parsing is a decision problem of the syntactic relation between words in a sentence. Recently, deep learning models are used for dependency parsing based on the word representations in a continuous vector space. However, it causes a mislabeled tagging problem for the proper nouns that rarely appear in the training corpus because it is difficult to express out-of-vocabulary (OOV) words in a continuous vector space. To solve the OOV problem in dependency parsing, we explored the proper noun embedding method according to the embedding unit. Before representing words in a continuous vector space, we replace the proper nouns with a special token and train them for the contextual features by using the multi-layer bidirectional LSTM. Two models of the syllable-based and morpheme-based unit are proposed for proper noun embedding and the performance of the dependency parsing is more improved in the ensemble model than each syllable and morpheme embedding model. The experimental results showed that our ensemble model improved 1.69%p in UAS and 2.17%p in LAS than the same arc-eager approach-based Malt parser.

• Annual Conference on Human and Language Technology
• /
• 2020.10a
• /
• pp.343-346
• /
• 2020

개체명 인식은 주어진 문장 내에서 OOV(Out of Vocaburary)로 자주 등장하는 고유한 의미가 있는 단어들을 미리 정의된 개체의 범주로 분류하는 작업이다. 최근 개체명이 문장 내에서 OOV로 등장하는 문제를 해결하기 위해 외부 리소스를 활용하는 연구들이 많이 진행되었다. 본 논문은 의미역, 의존관계 분석에 한국어 어휘지도를 이용한 자질을 추가하여 성능 향상을 보인 연구들을 바탕으로 이를 한국어 개체명 인식에 적용하고 평가하였다. 실험 결과, 한국어 어휘지도를 활용한 자질을 추가로 학습한 모델이 기존 모델에 비해 평균 1.83% 포인트 향상하였다. 또한, CRF 단일 모델만을 사용했음에도 87.25% 포인트라는 높은 성능을 보였다.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.2
• /
• pp.27-34
• /
• 2001

Utterance verification is used in variable vocabulary word recognition to reject the word that does not belong to in-vocabulary word or does not belong to correctly recognized word. Utterance verification is an important technology to design a user-friendly speech recognition system. We propose a new utterance verification algorithm for no-training utterance verification system based on the minimum verification error. First, using PBW (Phonetically Balanced Words) DB (445 words), we create no-training anti-phoneme models which include many PLUs(Phoneme Like Units), so anti-phoneme models have the minimum verification error. Then, for OOV (Out-Of-Vocabulary) rejection, the phoneme-based confidence measure which uses the likelihood between phoneme model (null hypothesis) and anti-phoneme model (alternative hypothesis) is normalized by null hypothesis, so the phoneme-based confidence measure tends to be more robust to OOV rejection. And, the word-based confidence measure which uses the phoneme-based confidence measure has been shown to provide improved detection of near-misses in speech recognition as well as better discrimination between in-vocabularys and OOVs. Using our proposed anti-model and confidence measure, we achieve significant performance improvement; CA (Correctly Accept for In-Vocabulary) is about 89％, and CR (Correctly Reject for OOV) is about 90％, improving about 15-21％ in ERR (Error Reduction Rate).

• Smart Media Journal
• /
• v.8 no.2
• /
• pp.58-65
• /
• 2019

Automatic text summarization is a process of shortening a text document by either extraction or abstraction. The abstraction approach inspired by deep learning methods scaling to a large amount of document is applied in recent work. Abstractive text summarization involves utilizing pre-generated word embedding information. Low-frequent but salient words such as terminologies are seldom included to dictionaries, that are so called, out-of-vocabulary(OOV) problems. OOV deteriorates the performance of Encoder-Decoder model in neural network. In order to address OOV words in abstractive text summarization, we propose a copy mechanism to facilitate copying new words in the target document and generating summary sentences. Different from the previous studies, the proposed approach combines accurate pointing information and selective copy mechanism based on bidirectional RNN and bidirectional LSTM. In addition, neural network gate model to estimate the generation probability and the loss function to optimize the entire abstraction model has been applied. The dataset has been constructed from the collection of abstractions and titles of journal articles. Experimental results demonstrate that both ROUGE-1 (based on word recall) and ROUGE-L (employed longest common subsequence) of the proposed Encoding-Decoding model have been improved to 47.01 and 29.55, respectively.

• Annual Conference on Human and Language Technology
• /
• 2018.10a
• /
• pp.434-439
• /
• 2018

문체 변환 시스템을 학습하는 데 있어서 가장 큰 어려움 중 하나는 병렬 말뭉치가 부족하다는 것이다. 최근 대량의 비병렬 말뭉치만으로 문체 변환 문제를 해결하려는 많은 연구들이 발표되었지만, 아직까지도 원 문장의 정보 보존(Content preservation)과 문체 변환(Style transfer) 모두를 이루는 것이 쉽지 않은 상태이다. 특히 비교사 학습의 특성상 문체 변환과 동시에 정보를 보존하는 것이 매우 어렵다. Attention 기반의 Seq2seq 네트워크를 이용할 경우에는 과도하게 원문의 정보가 보존되어 문체 변환 능력이 떨어지기도 한다. 그리고 OOV(Out-Of-Vocabulary) 문제 또한 존재한다. 본 논문에서는 Attention 기반의 Seq2seq 네트워크를 이용하여 어절 단위의 정보 보존력을 최대한 높이면서도, 입력 문장에 효과적으로 Noise를 넣어 문체 변환 성능을 저해하는 과도한 정보 보존 현상을 막고 문체의 특성을 나타내는 어절들이 잘 변환되도록 할 뿐 아니라 OOV 문제도 줄일 수 있는 방법을 제안한다. 우리는 비교 실험을 통해 본 논문에서 제안한 방법들이 한국어 문장뿐 아니라 영어 문장에 대해서도 state-of-the-art 시스템들에 비해 향상된 성능을 보여준다는 사실을 확인하였다.