• Annual Conference on Human and Language Technology
• /
• 2018.10a
• /
• pp.541-545
• /
• 2018

본 논문에서는 Sent2Vec을 이용한 문장 임베딩으로 구현한 유사 문장 판별 시스템을 제안한다. 또한 한국어 특성에 맞게 모델을 개선하여 성능을 향상시키는 방법을 소개한다. 고성능 라이브러리 구현과 제품화 가능한 수준의 완성도 높은 구현을 보였으며, 자체 구축한 평가셋으로 한국어 특성을 반영한 모델에 대한 P@1 평가 결과 Word2Vec CBOW에 비해 9.25%, Sent2Vec에 비해 1.93% 더 높은 성능을 보였다.

• Annual Conference on Human and Language Technology
• /
• 2019.10a
• /
• pp.388-391
• /
• 2019

과편향 뉴스 판별(hyperpartisan news detection)은 뉴스 기사가 특정 인물 또는 정당에 편향되었는지 판단하는 task이다. 이를 위해 feature-based ELMo + CNN 모델이 제안되었으나, 이는 문서 임베딩이 아닌 단어 임베딩의 평균을 사용한다는 한계가 존재한다. 따라서 본 논문에서는 feature-based 접근법을 따르며 Sentence-BERT(SentBERT)의 문서 임베딩을 이용한 feature-based SentBERT 기반의 과편향 뉴스 판별 모델을 제안한다. 제안 모델의 효과를 입증하기 위해 ELMO, BERT, SBERT와 CNN, BiLSTM을 적용한 비교 실험을 진행하였고, 기존 state-of-the-art 모델보다 f1-score 기준 1.3%p 높은 성능을 보였다.

• Smart Media Journal
• /
• v.7 no.4
• /
• pp.24-29
• /
• 2018

Text analysis technique for natural language processing in deep learning represents words in vector form through word embedding. In this paper, we propose a method of constructing a document vector and classifying it into spam and normal text message, using word embedding and deep learning method. Automatic spacing applied in the preprocessing process ensures that words with similar context are adjacently represented in vector space. Additionally, the intentional word formation errors with non-alphabetic or extraordinary characters are designed to avoid being blocked by spam message filter. Two embedding algorithms, CBOW and skip grams, are used to produce the sentence vector and the performance and the accuracy of deep learning based spam filter model are measured by comparing to those of SVM Light.

• Annual Conference on Human and Language Technology
• /
• 2022.10a
• /
• pp.526-530
• /
• 2022

자동 쓰기 평가 연구는 쓰기 답안지를 채점하는데 드는 시간과 비용을 절감할 수 있어, 교육 분야에서 큰 관심을 가지고 있다. 본 연구의 목적은 쓰기 답안지의 문서 구조를 효과적으로 학습하여 평가하고, 문장단위의 피드백을 제공하는데 있다. 그 방법으로는 문장 레벨에서 한국어 Sentence-BERT 모델을 활용하여 각 문장을 임베딩하고, LSTM 어텐션 모델을 활용하여 문서 레벨에서 임베딩 문장을 모델링한다. '한국어 쓰기 텍스트-점수 구간 데이터'를 활용하여 해당 모델의 성능 평가를 진행하였으며, 다양한 KoBERT 기반 모델과 비교 평가를 통해 제안하는 모델의 방법론이 효과적임을 입증하였다.

• Journal of Korea Society of Digital Industry and Information Management
• /
• v.15 no.1
• /
• pp.87-97
• /
• 2019

Reaction of people is importantly considered about specific case as a social network service grows. In the previous research on analysis of social network service, they predicted tendency of interesting topic by giving scores to sentences written by user. Based on previous study we proceeded research of sentiment analysis for social network service's sentences, which predict the result as positive or negative for movie reviews. In this study, we used movie review to get high accuracy. We classify the movie review into positive or negative based on the score for learning. Also, we performed embedding and morpheme analysis on movie review. We could predict learning result as positive or negative with a number 0 and 1 by applying the model based on learning result to social network service. Experimental result show accuracy of about 80% in predicting sentence as positive or negative.

• ETRI Journal
• /
• v.45 no.1
• /
• pp.18-27
• /
• 2023

We present an English-Korean speech translation corpus, named EnKoST-C. End-to-end model training for speech translation tasks often suffers from a lack of parallel data, such as speech data in the source language and equivalent text data in the target language. Most available public speech translation corpora were developed for European languages, and there is currently no public corpus for English-Korean end-to-end speech translation. Thus, we created an EnKoST-C centered on TED Talks. In this process, we enhance the sentence alignment approach using the subtitle time information and bilingual sentence embedding information. As a result, we built a 559-h English-Korean speech translation corpus. The proposed sentence alignment approach showed excellent performance of 0.96 f-measure score. We also show the baseline performance of an English-Korean speech translation model trained with EnKoST-C. The EnKoST-C is freely available on a Korean government open data hub site.

• Smart Media Journal
• /
• v.11 no.3
• /
• pp.31-38
• /
• 2022

In the document extraction summary study, various methods for selecting important sentences based on the relationship between sentences were proposed. In the Korean document summary using the summation similarity of sentences, the summation similarity of the sentences was regarded as the amount of sentence information, and the summary sentences were extracted by selecting important sentences based on this. However, the problem is that it does not take into account the various importance that each sentence contributes to the entire document. Therefore, in this study, we propose a document extraction summary method that provides a summary by selecting important sentences based on the amount of quantitative and semantic information in the sentence. As a result, the extracted sentence agreement was 58.56% and the ROUGE-L score was 34, which was superior to the method using only the combined similarity. Compared to the deep learning-based method, the extraction method is lighter, but the performance is similar. Through this, it was confirmed that the method of compressing information based on semantic similarity between sentences is an important approach in document extraction summary. In addition, based on the quickly extracted summary, the document generation summary step can be effectively performed.

• Journal of the Korea Society of Computer and Information
• /
• v.27 no.3
• /
• pp.13-23
• /
• 2022

In this study we leveraged a probing method to evaluate how a pre-trained L2 LSTM language model represents sentences with relative and coordinate clauses. The probing experiment employed adapted models based on the pre-trained L2 language models to trace the syntactic properties of sentence embedding vector representations. The dataset for probing was automatically generated using several templates related to different sentence structures. To classify the syntactic properties of sentences for each probing task, we measured the adaptation effects of the language models using syntactic priming. We performed linear mixed-effects model analyses to analyze the relation between adaptation effects in a complex statistical manner and reveal how the L2 language models represent syntactic features for English sentences. When the L2 language models were compared with the baseline L1 Gulordava language models, the analogous results were found for each probing task. In addition, it was confirmed that the L2 language models contain syntactic features of relative and coordinate clauses hierarchically in the sentence embedding representations.

• Journal of Intelligence and Information Systems
• /
• v.25 no.3
• /
• pp.161-177
• /
• 2019

In this paper, we study the performance improvement of the answer extraction in Question-Answering system by using sentence dependency parsing result. The Question-Answering (QA) system consists of query analysis, which is a method of analyzing the user's query, and answer extraction, which is a method to extract appropriate answers in the document. And various studies have been conducted on two methods. In order to improve the performance of answer extraction, it is necessary to accurately reflect the grammatical information of sentences. In Korean, because word order structure is free and omission of sentence components is frequent, dependency parsing is a good way to analyze Korean syntax. Therefore, in this study, we improved the performance of the answer extraction by adding the features generated by dependency parsing analysis to the inputs of the answer extraction model (Bidirectional LSTM-CRF). The process of generating the dependency graph embedding consists of the steps of generating the dependency graph from the dependency parsing result and learning the embedding of the graph. In this study, we compared the performance of the answer extraction model when inputting basic word features generated without the dependency parsing and the performance of the model when inputting the addition of the Eojeol tag feature and dependency graph embedding feature. Since dependency parsing is performed on a basic unit of an Eojeol, which is a component of sentences separated by a space, the tag information of the Eojeol can be obtained as a result of the dependency parsing. The Eojeol tag feature means the tag information of the Eojeol. The process of generating the dependency graph embedding consists of the steps of generating the dependency graph from the dependency parsing result and learning the embedding of the graph. From the dependency parsing result, a graph is generated from the Eojeol to the node, the dependency between the Eojeol to the edge, and the Eojeol tag to the node label. In this process, an undirected graph is generated or a directed graph is generated according to whether or not the dependency relation direction is considered. To obtain the embedding of the graph, we used Graph2Vec, which is a method of finding the embedding of the graph by the subgraphs constituting a graph. We can specify the maximum path length between nodes in the process of finding subgraphs of a graph. If the maximum path length between nodes is 1, graph embedding is generated only by direct dependency between Eojeol, and graph embedding is generated including indirect dependencies as the maximum path length between nodes becomes larger. In the experiment, the maximum path length between nodes is adjusted differently from 1 to 3 depending on whether direction of dependency is considered or not, and the performance of answer extraction is measured. Experimental results show that both Eojeol tag feature and dependency graph embedding feature improve the performance of answer extraction. In particular, considering the direction of the dependency relation and extracting the dependency graph generated with the maximum path length of 1 in the subgraph extraction process in Graph2Vec as the input of the model, the highest answer extraction performance was shown. As a result of these experiments, we concluded that it is better to take into account the direction of dependence and to consider only the direct connection rather than the indirect dependence between the words. The significance of this study is as follows. First, we improved the performance of answer extraction by adding features using dependency parsing results, taking into account the characteristics of Korean, which is free of word order structure and omission of sentence components. Second, we generated feature of dependency parsing result by learning - based graph embedding method without defining the pattern of dependency between Eojeol. Future research directions are as follows. In this study, the features generated as a result of the dependency parsing are applied only to the answer extraction model in order to grasp the meaning. However, in the future, if the performance is confirmed by applying the features to various natural language processing models such as sentiment analysis or name entity recognition, the validity of the features can be verified more accurately.

• English Language & Literature Teaching
• /
• no.6
• /
• pp.143-157
• /
• 2000

It has been one of the puzzles in the English syntax that so called the rule of subject-auxiliary inversion (SAI) is not allowed in subject wh-movement while it is not obligatory in non-subject wh-movement in a root sentence. This asymmetry has been a puzzle since SAI itself was thought to be a part of question construction as we can observe from yes/no questions. The asymmetry gets more complicated in terms of sentence embedding, i.e no SAI is permitted in the embedded context in question. The goal of this paper is to suggest an unified analysis for this unsolved grammatical phenomena on the basis of Rizzi (1997)'s recent work. The main idea is that SAI is not a I-to-C movement but one of I-to-Focus where Focus is a functional category and its phrase is located between CP and IP. The other proposal is that Wh-movement is no more homogeneous in terms of landing site between a root and an embedded sentence: the target for a wh-phrase in the former is the Spec of FocP (Focus Phrase) but the one in the latter is the Spec of CP as the standard theory assumes. Pesetsky (l999)'s analysis is discussed and its theoretical and empirical shortages are pointed out. Its rather radical proposals such as the one that the nominative case is just an uninterpretable tense feature of DP and the other that 'that' is no longer a complementizer but an element of I(nflection) make it less acceptable in spite of the possibility that it can get rid of Case theory entirely, which would be ideal in the spirit of minimalism.