온톨로지 지식 기반 특성치를 활용한Bidirectional LSTM-CRF 모델의시퀀스 태깅 성능 향상에 관한 연구 (Improving Bidirectional LSTM-CRF model Of Sequence Tagging by using Ontology knowledge based feature)

서지정보

· 발행기관 : 한국지능정보시스템학회
· 수록지 정보 : 지능정보연구 / 24권 / 1호 / 253 ~ 267페이지
· 저자명 : 진승희, 장희원, 김우주

초록

본 연구는 질의 응답(QA) 시스템에서 사용하는 개체명 인식(NER)의 성능을 향상시키기 위하여 시퀀스 태깅방법론을 적용한 새로운 방법론을 제안한다. 사용자의 질의를 입력 받아 데이터베이스에 저장된 정답을 추출하기 위해서는 사람의 언어를 컴퓨터가 알아들을 수 있도록 구조화 질의어(SQL)와 같은 데이터베이스의 언어로전환하는 과정이 필요한데, 개체명 인식은 사용자의 질의에서 데이터베이스에 포함된 클래스나 데이터 명을 식별하는 과정이다. 기존의 데이터베이스에서 질의에 포함된 단어를 검색하여 개체명을 인식하는 방식은 동음이의어와 문장성분 구를 문맥을 고려하여 식별하지 못한다. 다수의 검색 결과가 존재하면 그들 모두를 결과로 반환하기 때문에 질의에 대한 해석이 여러 가지가 나올 수 있고, 계산을 위한 시간복잡도가 커진다. 본 연구에서는 이러한 단점을 극복하기 위해 신경망 기반의 방법론을 사용하여 질의가 가지는 문맥적 의미를 반영함으로써이러한 문제를 해결하고자 했고 신경망 기반의 방법론의 문제점인 학습되지 않은 단어에 대해서도 문맥을 통해식별을 하고자 하였다. Sequence Tagging 분야에서 최신 기술인 Bidirectional LSTM-CRF 모델을 도입함으로써신경망 모델이 가진 단점을 해결하였고, 학습되지 않은 단어에 대해서는 온톨로지 기반 특성치를 활용하여 문맥을 반영한 추론을 사용하였다. 음악 도메인의 온톨로지(Ontology) 지식베이스를 대상으로 실험을 진행하고 그성능을 평가하였다. 본 연구에서 제안한 방법론인 L-Bidirectional LSTM-CRF의 성능을 정확하게 평가하기 위하여 학습에 포함된 단어들뿐만 아니라 학습에 포함되지 않은 단어들도 포함한 질의를 평가에 사용하였다. 그 결과 L-Bidirectional LSTM-CRF 모형을 재학습 시키지 않아도 학습에 포함되지 않은 단어를 포함한 질의에 대한개체명 인식이 가능함을 확인하였고, 전체적으로 개체명 인식의 성능이 향상됨을 확인할 수 있었다.

영어초록

This paper proposes a methodology applying sequence tagging methodology to improve the performance of NER(Named Entity Recognition) used in QA system. In order to retrieve the correct answers stored in the database, it is necessary to switch the user's query into a language of the database such as SQL(Structured Query Language). Then, the computer can recognize the language of the user. This is the process of identifying the class or data name contained in the database. The method of retrieving the words contained in the query in the existing database and recognizing the object based on the index does not identify the homophone and the word phrases because it does not consider the context of the user’s query. If there are multiple search results, all of them are returned as a result, so there can be many interpretations on the query and the time complexity for the calculation becomes large. In addition, in the existing QA system, chunking has a higher probability of selecting the longest phrase among the sentence component phrases retrieved after obtaining a subset of all cases that can come from the user query. This chunking process has a problem that the algorithm is complicated, but it is not always correct according to the query. To overcome these, this study aims to solve this problem by reflecting the contextual meaning of the query using neural network-based methodology and to identify the problem of the neural network based methodology. Context-sensitive tagging combined with CRF and Bidirectional LSTM can be used to supplement the LSTM-based model, which can limit long-term memory problems, but can be biased toward recent input because of the data nature of word embedding. Therefore, we have solved the disadvantages of the neural network model by introducing the latest technology Bidirectional LSTM-CRF model in Sequence Tagging field. We used reasoning that reflects context by using ontology-based characteristic values for untrained words. In case that untrained words come in, we store the object name recognition tag information obtained from the ontology knowledge base in the Lucene index DB as the ontology knowledge based feature. and In this paper, we propose a neural network model based on ontology knowledge based feature.
Experiments were conducted on the ontology knowledge base of music domain and the performance was evaluated. In order to accurately evaluate the performance of the L-Bidirectional LSTM-CRF proposed in this study, we experimented with converting the words included in the learned query into untrained words in order to test whether the words were included in the database but correctly identified the untrained words. Through experimentation, it was confirmed that tag recognition based on the context of the homophone and chunking is possible. In addition, we could apply the features that match the data type by using the CRF property which can use user defined function as a variable. As a result, it was possible to recognize objects considering the context and can recognize the untrained words without re-training the L-Bidirectional LSTM-CRF model, and it is confirmed that the performance of the object recognition as a whole is improved. Therefore, the proposed L-Bidirectional LSTM-CRF methodology can be applied to the ontology knowledge base of various fields to solve the object name recognition problem. However, the proposed L-Bidirectional LSTM-CRF method performed better than the conventional LSTM-CRF method, but lacked the ability to process homophones compared to the conventional Bidirectional LSTM-CRF. This is due to the fact that the hypothesis of this study inserts a feature that can infer a pattern of unknown words based on the entity name pattern of trained queries. This is because the weight increases linearly with the inclusion in the training data. In future research, it is necessary to experiment with various methods of applying weights so that homophones can be found well while maintaining overall performance.

참고자료

· 없음