• Annual Conference on Human and Language Technology
• /
• 2019.10a
• /
• pp.213-218
• /
• 2019

기계독해는 자연어로 표현된 질문과 단락이 주어졌을 때, 해당 단락 내에 표현된 정답을 찾는 태스크이다. 최근 기계독해 태스크도 다른 자연어처리 태스크와 유사하게 BERT, XLNet, RoBERTa와 같이 사전에 학습한 언어모델을 이용하고 질문과 단락이 입력되었을 경우 정답의 경계를 추가 학습(fine-tuning)하는 방법이 우수한 성능을 보이고 있으며, 특히 KorQuAD v1.0 데이터셋에서 학습 및 평가하였을 경우 94% F1 이상의 높은 성능을 보이고 있다. 본 논문에서는 현재 최고 수준의 기계독해 기술이 학습셋과 유사한 평가셋이 아닌 일반적인 질문과 단락 쌍에 대해서 가지는 일반화 능력을 평가하고자 한다. 이를 위하여 첫번째로 한국어에 대해서 공개된 KorQuAD v1.0 데이터셋과 NIA v2017 데이터셋, 그리고 엑소브레인 과제에서 구축한 엑소브레인 v2018 데이터셋을 이용하여 데이터셋 간의 교차 평가를 수행하였다. 교차 평가결과, 각 데이터셋의 정답의 길이, 질문과 단락 사이의 오버랩 비율과 같은 데이터셋 통계와 일반화 성능이 서로 관련이 있음을 확인하였다. 다음으로 KorBERT 사전 학습 언어모델과 학습 가능한 기계독해 데이터 셋 21만 건 전체를 이용하여 학습한 기계독해 모델에 대해 블라인드 평가셋 평가를 수행하였다. 블라인드 평가로 일반분야에서 학습한 기계독해 모델의 법률분야 평가셋에서의 일반화 성능을 평가하고, 정답 단락을 읽고 질문을 생성하지 않고 질문을 먼저 생성한 후 정답 단락을 검색한 평가셋에서의 기계독해 성능을 평가하였다. 블라인드 평가 결과, 사전 학습 언어 모델을 사용하지 않은 기계독해 모델 대비 사전 학습 언어 모델을 사용하는 모델이 큰 폭의 일반화 성능을 보였으나, 정답의 길이가 길고 질문과 단락 사이 어휘 오버랩 비율이 낮은 평가셋에서는 아직 80%이하의 성능을 보임을 확인하였다. 본 논문의 실험 결과 기계 독해 태스크는 특성 상 질문과 정답 사이의 어휘 오버랩 및 정답의 길이에 따라 난이도 및 일반화 성능 차이가 발생함을 확인하였고, 일반적인 질문과 단락을 대상으로 하는 기계독해 모델 개발을 위해서는 다양한 유형의 평가셋에서 일반화 평가가 필요함을 확인하였다.

• Journal of Intelligence and Information Systems
• /
• v.24 no.4
• /
• pp.111-136
• /
• 2018

In this paper, we propose a methodology to extract answer information about queries from various types of unstructured documents collected from multi-sources existing on web in order to expand knowledge base. The proposed methodology is divided into the following steps. 1) Collect relevant documents from Wikipedia, Naver encyclopedia, and Naver news sources for "subject-predicate" separated queries and classify the proper documents. 2) Determine whether the sentence is suitable for extracting information and derive the confidence. 3) Based on the predicate feature, extract the information in the proper sentence and derive the overall confidence of the information extraction result. In order to evaluate the performance of the information extraction system, we selected 400 queries from the artificial intelligence speaker of SK-Telecom. Compared with the baseline model, it is confirmed that it shows higher performance index than the existing model. The contribution of this study is that we develop a sequence tagging model based on bi-directional LSTM-CRF using the predicate feature of the query, with this we developed a robust model that can maintain high recall performance even in various types of unstructured documents collected from multiple sources. The problem of information extraction for knowledge base extension should take into account heterogeneous characteristics of source-specific document types. The proposed methodology proved to extract information effectively from various types of unstructured documents compared to the baseline model. There is a limitation in previous research that the performance is poor when extracting information about the document type that is different from the training data. In addition, this study can prevent unnecessary information extraction attempts from the documents that do not include the answer information through the process for predicting the suitability of information extraction of documents and sentences before the information extraction step. It is meaningful that we provided a method that precision performance can be maintained even in actual web environment. The information extraction problem for the knowledge base expansion has the characteristic that it can not guarantee whether the document includes the correct answer because it is aimed at the unstructured document existing in the real web. When the question answering is performed on a real web, previous machine reading comprehension studies has a limitation that it shows a low level of precision because it frequently attempts to extract an answer even in a document in which there is no correct answer. The policy that predicts the suitability of document and sentence information extraction is meaningful in that it contributes to maintaining the performance of information extraction even in real web environment. The limitations of this study and future research directions are as follows. First, it is a problem related to data preprocessing. In this study, the unit of knowledge extraction is classified through the morphological analysis based on the open source Konlpy python package, and the information extraction result can be improperly performed because morphological analysis is not performed properly. To enhance the performance of information extraction results, it is necessary to develop an advanced morpheme analyzer. Second, it is a problem of entity ambiguity. The information extraction system of this study can not distinguish the same name that has different intention. If several people with the same name appear in the news, the system may not extract information about the intended query. In future research, it is necessary to take measures to identify the person with the same name. Third, it is a problem of evaluation query data. In this study, we selected 400 of user queries collected from SK Telecom 's interactive artificial intelligent speaker to evaluate the performance of the information extraction system. n this study, we developed evaluation data set using 800 documents (400 questions * 7 articles per question (1 Wikipedia, 3 Naver encyclopedia, 3 Naver news) by judging whether a correct answer is included or not. To ensure the external validity of the study, it is desirable to use more queries to determine the performance of the system. This is a costly activity that must be done manually. Future research needs to evaluate the system for more queries. It is also necessary to develop a Korean benchmark data set of information extraction system for queries from multi-source web documents to build an environment that can evaluate the results more objectively.