• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2018.06a
• /
• pp.336-337
• /
• 2018

최근 이미지의 Visual 정보를 추출하고 Multi label 분류를 통해 나온 결과의 상관관계를 modeling하여 문장으로 출력하는 CNN-RNN 아키텍처가 많은 발전을 이뤘다. 이 아키텍처의 출력은 이미지의 정보가 요약되어 문장으로 표현되기 때문에 Semantic정보가 풍부하여 유사 콘텐츠 검색에도 사용 가능하다. 하지만 결과 문장에 사람이 포함 되면 광범위한 검색 결과를 얻게 되고 부정확한 결과를 초래하게 된다. 이에 본 논문에서는 문장에서 사람을 인식하여 Identity를 부여함으로써 검색어를 좀 더 구체적으로 생성하고자 한다. 이 문제를 해결하기 위해 자연어 처리의 분야 중 하나인 개체명 인식(Named Entity Recognition) 문제로 다루며, 가장 많이 사용되고 있는 모델인 Bidirectional-LSTM-CRF와 CoNLL2003 dataset을 사용하여 수행 한다.

• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.617-621
• /
• 2021

의미역 결정은 문장 속에서 서술어와 그 논항의 관계를 파악하며, '누가, 무엇을, 어떻게, 왜' 등과 같은 의미역 관계를 찾아내는 자연어 처리 기법이다. 최근 수행되고 있는 의미역 결정 연구는 주로 말뭉치를 활용하여 딥러닝 학습을 하는 방식으로 연구가 이루어지고 있다. 최근 구글에서 개발한 사전 훈련된 Bidirectional Encoder Representations from Transformers (BERT) 모델이 다양한 자연어 처리 분야에서 상당히 높은 성능을 보이고 있다. 본 논문에서는 한국어 의미역 결정 성능 향상을 위해 한국어의 언어적 특징을 고려하며 사전 학습된 SNU KR-BERT를 사용하면서 한국어 의미역 결정 모델의 성능을 살펴보였다. 또한, 본 논문에서는 BERT 모델에서 과연 어떤 히든 레이어(hidden layer)에서 한국어 의미역 결정을 더 잘 수행하는지 알아보고자 하였다. 실험 결과 마지막 히든 레이어 임베딩을 활용하였을 때, 언어 모델의 성능은 66.4% 였다. 히든 레이어 별 언어 모델 성능을 비교한 결과, 마지막 4개의 히든 레이어를 이었을 때(concatenated), 언어 모델의 성능은 67.9% 이였으며, 11번째 히든 레이어를 사용했을 때는 68.1% 이였다. 즉, 마지막 히든 레이어를 선택했을 때보다 더 성능이 좋았다는 것을 알 수 있었다. 하지만 각 언어 모델 별 히트맵을 그려보았을 때는 마지막 히든 레이어 임베딩을 활용한 언어 모델이 더 정확히 의미역 판단을 한다는 것을 알 수 있었다.

• 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.