인공지능 다국어 번역을 위한 언어학적 전 처리와 공학적 처리 방법론 연구-역 방향 번역과 합성 코퍼스 생성을 중심으로 = A Study on Linguistic Preprocessing and Engineering Processing Methodology for Artificial Intelligence Multilingual Translation

다국어 초록 (Multilingual Abstract)

In this paper, we studied the basic humanities work that linguists can perform in the preprocessing process for natural language processing. Embedding an atypical and infinite human language into a structured and finite computer resource is a very important task that has not yet been solved, but must be solved in order to complete the natural language processing by artificial intelligence. In order to solve these tasks, humanities and linguistic knowledge must be mobilized, and engineering computational skills must be supported. In this paper, we introduced the underlying technology, focusing on the creation of the artificial intelligence synthesis corpus and the back translation, and introducing various issues for accurate meaning extraction in natural language. In the text, in particular, the problems of meaning accompaniment and categorization, morphological negation, direct visibility and implicitity, and existence were dealt with, and the introduction of Word2Vec, the concept of Subword, and the use of big data were suggested as engineering solutions. In connection with this, we proposed back translation and artificial intelligence synthesis corpus construction technologies, and explained that these technologies can play a particularly large role in natural language meaning extraction and multilingual translation. In fact, this research team conducted a Korean-English artificial intelligence translation test using the above technology, supported by NIPA high-performance computing resources. Based on a total of 4.6 million Korean-English parallel data, the data were trained by repeating 120 times for 60 days. It was found from the experimental results that the translation performance improved by about 5% when the back translation was repeated 40 times for 20 days (about 0.3 → 0.312).

참고문헌 (Reference)

1 이기창, "한국어 임베딩" 에이콘 2019

2 김재훈, "통합국어정보베이스를 위한 한국어 형태 통사 태그 설정" 한국과학기술원 1996

3 Matteo Negri, "eSCAPE: a Large Synthetic Corpus for Automatic Post-Editing"

4 Jun-Yan Zhu, "Unfaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks"

5 Sergey Edunov, "Understanding Back-Translation at Scale"

6 Sutskever, I., "Sequence to sequence learning with neural networks" 2014

7 Richard Socher, "Recursive Deep Models for Semantic Compositionality over a Sentiment Treebank"

8 Marek Rei, "Jointly Learning to Label Sentences and Tokens"

9 Rico Sennrich, "Improving neural machine translation models with monolingual data"

10 R. Bowman, "GLUE: A multi-task benchmark and analysis platform for natural language understanding" 1-20, 2019

11 Tomas Mikolov, "Distributed Representations of Words and Phrases and their Compositionality" 3111-3119, 2013

12 K. Papineni, "BLEU: a method for automatic evaluation of machine translation" 2002

13 Marek Rei, "Auxiliary Objectives for Neural Error Detection Models"