• Annual Conference on Human and Language Technology
• /
• 2022.10a
• /
• pp.600-603
• /
• 2022

분류체계를 결정하는 일은 매우 중요하지만 어려운 일이다. 우리는 수집된 용어 목록에 클러스터링을 적용하여 상위 범주의 범위를 자동으로 설정하고자 하였다. 용어 클러스터링은 용어를 나타내는 벡터에 큰 의존성을 갖는다. 이에 클러스터링의 성능 향상을 위해 다양한 용어 임베딩 방법을 비교하였으며 용어에 대한 정의문의 벡터를 용어 벡터로 사용하여 가장 우수한 클러스터링 결과를 얻었다. 또한 실험을 통해 클러스터링 알고리즘 중 k-means clustering이 고차원의 벡터에 대해 좋은 성능의 군집을 생성함을 확인하였다.

• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.91-96
• /
• 2021

조기 위험 검출은 실시간으로 들어오는 텍스트를 순차적으로 처리하면서 해당 대화에 위험이 있는지 조기에 분류하는 작업으로, 정확도 저하를 최소화하는 동시에 가능한 한 빨리 대화를 분류하는 것을 목적으로 한다. 이러한, 조기 위험 검출은 온라인 그루밍 검출, 보이스 피싱 검출과 같은 다양한 영역에 활용될 수 있다. 이에, 본 논문에서는 조기 위험 검출 문제를 정의하고, 이를 평가할 수 있는 데이터 셋과 Latency F1 평가 지표를 소개한다. 또한, 점진적 문장 분류 모듈과 위험 검출 결정 모듈로 구성된 점진적 조기 텍스트 분류 시스템을 제안한다. 점진적 문장 분류 모듈은 이전 문장들에 대한 메모리 벡터와 현재 문장 벡터를 통해 현재까지의 대화를 분류한다. 위험 검출 결정 모듈은 softmax 분류 점수와 강화학습을 기반으로 하여 Read 또는 Stop 판단을 내린다. 결정 모듈이 Stop 판단을 내리면, 현재까지의 대화에 대한 분류 결과를 전체 대화의 분류 결과로 간주하고 작업을 종료한다. 해당 시스템은 micro F1과 Latency F1 지표 각각에서 0.9684와 0.8918로 높은 검출 정확성 및 검출 신속성을 달성하였다.

• Annual Conference on Human and Language Technology
• /
• 2022.10a
• /
• pp.473-476
• /
• 2022

문장 임베딩은 문장의 의미를 고려하여 모델이 적절하게 의미적인 벡터 공간에 표상하는 것이다. 문장 임베딩을 위해 다양한 방법들이 제안되었지만, 최근 가장 높은 성능을 보이는 방법은 대조적 학습 방법이다. 대조적 학습을 이용한 문장 임베딩은 문장의 의미가 의미적으로 유사하면 가까운 공간에 배치하고, 그렇지 않으면 멀게 배치하도록 학습하는 방법이다. 이러한 대조적 학습은 비지도와 지도 학습 방법이 존재하는데, 본 논문에서는 효과적인 비지도 학습방법을 제안한다. 기존의 비지도 학습 방법은 문장 표현을 학습하는 언어모델이 자체적인 정보를 활용하여 문장의 의미를 구별한다. 그러나, 하나의 모델이 판단하는 정보로만 문장 표현을 학습하는 것은 편향적으로 학습될 수 있기 때문에 한계가 존재한다. 따라서 본 논문에서는 Cross-Encoder의 Re-Ranker를 통한 의미 검색으로부터 추천된 문장 쌍을 학습하여 기존 모델의 성능을 개선한다. 결과적으로, STS 테스크에서 베이스라인보다 2% 정도 더 높은 성능을 보여준다.

• Proceedings of the Acoustical Society of Korea Conference
• /
• 1998.08a
• /
• pp.396-400
• /
• 1998

문장 독립형 화자인식 시스템에 운율정보 사용을 제안한다. 스펙트럴 특징패턴만을 주로 사용하고 있는 기존의 화자인식 시스템은 채널왜곡이나 기타 잡음환경에서 성능이 크게 저하된다. 그러나 화자의 speaking style을 반영하는 운율정보는 주위환경에 강인한 특성을 갖는다. 적합한 코드북 크기와 피치 컨투어 특징 벡터의 길이를 실험 치로 구하여 자동차 소음과 백색 가우시안 소음이 섞인 음성에 대하여 화자인식 실험을 하였다. 실험 결과 소음 환경에서 운율 정보를 이용한 화자 dsltlr 시스템이 스펙트럴 모델보다 인식율이 높음을 보였다.

• Smart Media Journal
• /
• v.10 no.1
• /
• pp.39-46
• /
• 2021

This study is about a method of extracting a summary from a news article in consideration of the importance of each sentence constituting the article. We propose a method of calculating sentence importance by extracting the probabilities of topic sentence, similarity with article title and other sentences, and sentence position as characteristics that affect sentence importance. At this time, a hypothesis is established that the Topic Sentence will have a characteristic distinct from the general sentence, and a deep learning-based classification model is trained to obtain a topic sentence probability value for the input sentence. Also, using the pre-learned ELMo language model, the similarity between sentences is calculated based on the sentence vector value reflecting the context information and extracted as sentence characteristics. The topic sentence classification performance of the LSTM and BERT models was 93% accurate, 96.22% recall, and 89.5% precision, resulting in high analysis results. As a result of calculating the importance of each sentence by combining the extracted sentence characteristics, it was confirmed that the performance of extracting the topic sentence was improved by about 10% compared to the existing TextRank algorithm.

• Journal of Advanced Marine Engineering and Technology
• /
• v.39 no.3
• /
• pp.274-279
• /
• 2015

The main goal of this research is to build a sentiment analysis system which automatically determines user opinions of the Stanford Sentiment Treebank in terms of three sentiments such as positive, negative, and neutral. Firstly, sentiment sentences are POS tagged and parsed to dependency structures. All nodes of the Treebank and their polarities are automatically extracted from the Treebank. We train two Support Vector Machines models. One is for a node level classification and the other is for a sentence level. We have tried various type of features such as word lexicons, POS tags, Sentiment lexicons, head-modifier relations, and sibling relations. Though we acquired 74.2% in accuracy on the test set for 3 class node level classification and 67.0% for 3 class sentence level classification, our experimental results for 2 class classification are comparable to those of the state of art system using the same corpus.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.22 no.5
• /
• pp.656-662
• /
• 2012

In this paper, we describe a new summarizing method based on a graph-based and a sense-based analysis. In the graph-based analysis, we convert sentences in a document into word vectors and calculate the similarity between each sentence using LSA. We reflect this similarity of sentences and the rarity scores of words in sentences to define weights of edges in the graph. Meanwhile, in the sense-based analysis, in order to determine the sense of words, subjectivity or objectivity, we built a database which is extended from the golden standards using Wordnet. We calculate the subjectivity of sentences from the sense of words, and select more subjective sentences. Lastly, we combine the results of these two methods. We evaluate the performance of the proposed method using classification games, which are usually used to measure the performances of summarization methods. We compare our method with the MS-Word auto-summarization, and verify the effectiveness of ours.

• Journal of the Korean Society for information Management
• /
• v.18 no.3
• /
• pp.159-178
• /
• 2001

This paper presents an automatic text summarization model which selects representative sentences from sentence clusters to create a summary. Summary generation experiments were performed on two sets of test documents after learning the optimum environment from a training set. Centroid clustering method turned out to be the most effective in clustering sentences, and sentence weight was found more effective than the similarity value between sentence and cluster centroid vectors in selecting a representative sentence from each cluster. The result of experiments also proves that inverse sentence weight as well as title word weight for terms and location weight for sentences are effective in improving the performance of summarization.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.5
• /
• pp.496-502
• /
• 2021

Recently, deep speaker embedding approach has been widely used in text-independent speaker verification, which shows better performance than the traditional i-vector approach. In this work, to improve the deep speaker embedding approach, we propose a novel method called group-based speaker embedding which incorporates group information. We cluster all speakers of the training data into a predefined number of groups in an unsupervised manner, so that a fixed-length group embedding represents the corresponding group. A Group Decision Network (GDN) produces a group weight, and an aggregated group embedding is generated from the weighted sum of the group embeddings and the group weights. Finally, we generate a group-based embedding by adding the aggregated group embedding to the deep speaker embedding. In this way, a speaker embedding can reduce the search space of the speaker identity by incorporating group information, and thereby can flexibly represent a significant number of speakers. We conducted experiments using the VoxCeleb1 database to show that our proposed approach can improve the previous approaches.

• KIPS Transactions on Software and Data Engineering
• /
• v.8 no.11
• /
• pp.441-448
• /
• 2019

Previous researches on automatic spacing of Korean sentences has been researched to correct spacing errors by using n-gram based statistical techniques or morpheme analyzer to insert blanks in the word boundary. In this paper, we propose an end-to-end automatic word spacing by using deep neural network. Automatic word spacing problem could be defined as a tag classification problem in unit of syllable other than word. For contextual representation between syllables, Bi-LSTM encodes the dependency relationship between syllables into a fixed-length vector of continuous vector space using forward and backward LSTM cell. In order to conduct automatic word spacing of Korean sentences, after a fixed-length contextual vector by Bi-LSTM is classified into auto-spacing tag(B or I), the blank is inserted in the front of B tag. For tag classification method, we compose three types of classification neural networks. One is feedforward neural network, another is neural network language model and the other is linear-chain CRF. To compare our models, we measure the performance of automatic word spacing depending on the three of classification networks. linear-chain CRF of them used as classification neural network shows better performance than other models. We used KCC150 corpus as a training and testing data.