Earticle

기계번역 품질 예측(Quality Estimation, QE)은 정답 문장(Reference sentence) 없이도 기계번역 결과의 질을 평가할 수 있으며, 활용도가 높다는 점에서 그 필요성이 대두되고 있다. Conference on machine translation(WMT)에서 매년 이와 관련한 shared task가 열리고 있고 최근에는 대용량 데이터 기반 Pretrained language model(PLM)을 적용한 연구들이 주로 진행되고 있다. 본 논문에서는 기계번역 품질 예측 task에 대한 설명 및 연구 동향에 대한 전반적인 survey를 진행했고, 최근 자주 활용되는 PLM의 특징들에 대해 정리하였다. 더불어 아직 활용된 바가 없는 multilingual BART 모델을 이용하여 기존 연구들인 XLM, multilingual BERT, XLM-RoBERTa와 의 비교 실험 및 분석을 진행하였다. 실험 결과 어떤 사전 학습된 다중언어 모델이 QE에 적용했을 때 가장 효과적인지 확인하였을 뿐 아니라 multilingual BART 모델의 QE 태스크 적용 가능성을 확인했다.

Quality estimation (QE) can evaluate the quality of machine translation output even for those who do not know the target language, and its high utilization highlights the need for QE. QE shared task is held every year at Conference on Machine Translation (WMT), and recently, researches applying Pretrained Language Model (PLM) are mainly being conducted. In this paper, we conduct a survey on the QE task and research trends, and we summarize the features of PLM. In addition, we used a multilingual BART model that has not yet been utilized and performed comparative analysis with the existing studies such as XLM, multilingual BERT, and XLM-RoBERTa. As a result of the experiment, we confirmed which PLM was most effective when applied to QE, and saw the possibility of applying the multilingual BART model to the QE task.

기계번역이란 소스언어를 목적언어로 컴퓨터가 번역하는 소프트웨어를 의미하며 규칙기반, 통계기반 기계번역 을 거쳐 최근에는 인공신경망 기반 기계번역에 대한 연구가 활발히 이루어지고 있다. 인공신경망 기계번역에서 중요한 요소 중 하나로 고품질의 병렬 말뭉치를 뽑을 수 있는데 이제까지 한국어 관련 언어쌍의 고품질 병렬 코퍼스를 구하기 쉽지 않은 실정이었다. 최근 한국정보화진흥원의 AI HUB에서 고품질의 160만 문장의 한-영 기계번역 병렬 말뭉치를 공개하였다. 이에 본 논문은 AI HUB에서 공개한 데이터 및 현재까지 가장 많이 쓰인 한-영 병렬 데이터인 OpenSubtitles와 성능 비교를 통해 각각의 데이터의 품질을 검증하고자 한다. 테스트 데이터로 한-영 기계번역 관련 공식 테스트셋인 IWSLT에서 공개한 테스트셋을 이용하여 보다 객관성을 확보하였다. 실험결과 동일한 테스트셋으로 실험한 기존의 한-영 기계번역 관련 논문들보다 좋은 성능을 보임을 알 수 있었으며 이를 통해 고품질 데이터의 중요성 을 알 수 있었다.

Machine translation refers to software that translates a source language into a target language, and has been actively researching Neural Machine Translation through rule-based and statistical-based machine translation. One of the important factors in the Neural Machine Translation is to extract high quality parallel corpus, which has not been easy to find high quality parallel corpus of Korean language pairs. Recently, the AI HUB of the National Information Society Agency(NIA) unveiled a high-quality 1.6 million sentences Korean-English parallel corpus. This paper attempts to verify the quality of each data through performance comparison with the data published by AI Hub and OpenSubtitles, the most popular Korean-English parallel corpus. As test data, objectivity was secured by using test set published by IWSLT, official test set for Korean-English machine translation. Experimental results show better performance than the existing papers tested with the same test set, and this shows the importance of high quality data.

기계번역이란 소스문장(Source Sentence)을 타겟문장(Target Sentence)으로 컴퓨터가 번역하는 시스템을 의 미한다. 기계번역에는 다양한 하위분야가 존재하며 APE(Automatic Post Editing)이란 기계번역 시스템의 결과물을 교정하여 더 나은 번역문을 만들어내는 기계번역의 하위분야이다. 즉 기계번역 시스템이 생성한 번역문에 포함되어 있 는 오류를 수정하여 교정문을 만드는 과정을 의미한다. 기계번역 모델을 변경하는 것이 아닌 기계번역 시스템의 결과 문장을 교정하여 번역품질을 높이는 연구분야이다. 2015년부터 WMT 공동 캠페인 과제로 선정되었으며 성능 평가는 TER(Translation Error Rate)을 이용한다. 이로 인해 최근 APE에 모델에 대한 다양한 연구들이 발표되고 있으며 이에 본 논문은 APE 분야의 최신 동향에 대해서 다루게 된다.

Machine translation refers to a system where a computer translates a source sentence into a target sentence. There are various subfields of machine translation. APE (Automatic Post Editing) is a subfield of machine translation that produces better translations by editing the output of machine translation systems. In other words, it means the process of correcting errors included in the translations generated by the machine translation system to make proofreading. Rather than changing the machine translation model, this is a research field to improve the translation quality by correcting the result sentence of the machine translation system. Since 2015, APE has been selected for the WMT Shaed Task. and the performance evaluation uses TER (Translation Error Rate). Due to this, various studies on the APE model have been published recently, and this paper deals with the latest research trends in the field of APE.

Neural machine translation has achieved great success in the last decade. It has been widely adopted in multiple commercial products and domains, facilitating communication and trade between countries. In this paper, we present our Chinese to Vietnamese neural machine translation system, which integrates recent advances in deep learning and is enhanced with an external named entity translation model.

Machine translation is one of the classic sub-fields of natural language processing, a field that has been studied for a long time. There are several methods related to machine translation, and neural machine translation is one of them. Neural machine translation is a method that translates the source sentence to the target sentence through a neural network model. Recently, the transformer model using the attention mechanism has become the SOTA technique in the field. In this research, the transformer model using the attention mechanism is trained through 2 parallel corpora: English-German and English-Korean. The results of translations are evaluated with a BLEU score for each. The transformer shows an 8.60 BLEU score when trained by the English-German parallel corpora, and a 0.43 BLEU score in the case of English-Korean. It seems like the reason why the performance of English-German translations is higher than the performance of English-Korean translations is that English and Germany are in the Germanic language family, which is the same. In contrast, unlike English, Korean belongs to Altaic or Isolated languages and has different linguistic characteristics. The performance of the transformer is also affected by the quality of the parallel corpus used for training. It can be interpreted that it is necessary to secure a higher quality English-Korean parallel corpus.

한–중 법률 번역은 법률 용어의 복잡성, 일관성에 대한 엄격한 요구, 그리고 도메인 특화 이중언어 자원의 부족으로 인해 고유한 어려움을 가진다. 최근 신경망 기계번역(NMT)의 발전으로 전반적인 번역 성능은 향상되었으나, 용어의 신뢰성과 오류 분포 문제는 여전히 해결되지 않아 법률 분야에서의 실제 적용 가능성을 제한하고 있다. 본 연구는 mBART 기반 NMT 모델의 디코딩 단계와 사후편집 단계에 이중언어 법률 용어집을 통합하는 용어 인식 번역 프레임워크를 제안한다. 또한 의미 불일치, 문법 오류, 용어 일탈 등을 포함한 주요 법률 번역 오류를 분류하고 정량화하기 위한 종합적인 오류 분석 체계를 제시한다. 한–중 법률 코퍼스를 이용한 실험 결과, 제안된 방법은 BLEU 및 TER과 같은 기존 평가 지표를 향상시킬 뿐만 아니라, 기준 시스템에 비해 더 높은 용어 정확도와 일관성을 달성함을 보여준다. 이러한 결과는 도메인 특화 용어 통제와 체계적인 오류 분석이 기계 기반 법률 번역의 신뢰성을 향상시키는 데 핵심적임을 시사하며, 사법 및 입법 분야에서 보다 신뢰할 수 있는 활용을 가능하게 한다.

Legal translation between Korean and Chinese poses unique challenges due to the complexity of legal terminology, strict requirements for consistency, and the scarcity of domain-specific bilingual resources. While recent advances in neural machine translation (NMT) have improved overall translation performance, issues of terminology reliability and error distribution remain unresolved, limiting the practical applicability of such systems in legal contexts. This study proposes a terminology-aware translation framework that integrates a bilingual legal glossary into both the decoding and post-editing stages of an mBART-based NMT model. In addition, a comprehensive error analysis framework is introduced to classify and quantify common legal translation errors, including semantic inconsistency, grammatical inaccuracies, and terminological deviations. Experiments on a Korean–Chinese legal corpus demonstrate that the proposed approach not only enhances conventional evaluation metrics such as BLEU and TER, but also achieves higher terminology accuracy and consistency compared with baseline systems. These findings highlight the importance of domain-specific terminology control and systematic error analysis in improving the reliability of machine-assisted legal translation, thereby contributing to more trustworthy applications in judicial and legislative contexts.

This study aims to explore the limitations of Neural Machine Translation (NMT) in processing spoken text. Following literature review, an analysis of government press briefing transcriptions that have been processed by Google Translate will be analyzed based on the framework of the features of spoken language. The concept of cognitive complements will be employed to attempt to explain any identified pattern of errors made by the NMT. Specifically, the role of cognitive complements that exist outside of the utterances will be discussed, drawing on the theories of Lederer(1989) and Gutt(2000) among other scholars. A sample text analysis of spoken text translation will be conducted to investigate whether this inherently human mechanism works in the latest neural based machine translation engine, ‘Google Translate’. Analysis will show that while Google Translate demonstrated quite impressively on some of the well-crafted segments that resemble written language, it performed poorly on segments laden with features of spoken language. The findings of the study suggest that the need for human interpreters’ in mediating highly professional inter-lingual communication will likely persist in the future. As long as artificial intelligence does not acquire the ability to infer meaning from the cognitive complements as human beings to, the application of automated inter-lingual interpreting devices may need to be limited to certain market segments such as for tourism, entertainment and socializing. leisure.

This paper addresses the quality of Korean-English legal translation outputs generated by three neural machine translation engines, namely NMT1, NMT2 and NMT3. NMT1 and NMT2 are widely available multilingual generic machine translation engines, whereas NMT3 is a custom machine translation engine for legal translation. In both automatic and human evaluations of the three engines’ English translations of Korean statutes, the custom engine outperformed the other two NMTs. BLEU and METEOR scores revealed that NMT3’s output was more similar to human reference translation. Human evaluation results confirmed NMT3’s strong performance, outperforming the generic engines in accuracy and fluency while underperforming in terminology. However, further analysis revealed that NMT3 was superior in handling terms comprising compound nouns and sentences over 25 words. Although the custom engine’s performance still requires substantial human involvement for quality translation, it showed a potential for the legal domain, demonstrating human parity in short sentences. The growing demand for speedy translation under budget constraints calls for further research on technology-based translation tools and human-computer interaction for legal translation.

At the present time, when the development of neural machine translation presents a significant turning point, this paper attempts to compare NMT with existing statistically-based machine translation, and discuss implications for future translator education. Multiple types of MT errors and mistranslations are identified by analysis. This study intends to propose a codified table of those error types as a tool for objective assessment and ongoing monitoring of the quality of MT. Based on a translation editing scheme for learners, error types at lexical, phrasal, syntactic and textual levels were identified, which were then re-sorted for MT. The classification was applied to the analysis of translation of informative texts and news articles, for different error types, MT platforms (Naver, Google) and language directions, to draw graphs to demonstrate the results. Considering that the work of translation in the future will likely evolve into a combination of MT with human post-editing, continuous MT quality assessment would be necessary. The outcome of this study would help post-editors figure out where to focus working with different MT platforms, which is potentially useful for translator training.

Terminology translation plays a key role in technical machine translation. This paper conducts quantitative and qualitative evaluation and presents a comprehensive error analysis on terminology translation from Korean to English. For this study, 637 terms are selected from a patent domain corpus. Among them, 425 terms are translated correctly while other 212 terms have one or more errors. Those identified errors are classified into seven error categories which include addition, omission, lexical errors, order errors, morphological errors, partial errors, and miscellanies. Among them, partial errors and morphological errors are the two top error categories, and lexical errors and omission cases are also high. The study also presents n-gram wise error rates. The calculation of n-gram wise terminology translation error rates shows that the curve for the Google Translation results goes upwards with the increasing size of n. This shows that neural machine translation can be vulnerable in translating a higher-order of n-gram terms. This study has implications for Korean-English machine translation research, in that it classifies error cases spotted in machine translation of terminology which is critical in securing accuracy of translations.

There was a big technical progress in the research field of machine translation: the main approach has switched from statistical machine translation (SMT) to neural machine translation (NMT), leading to dramatic improvements in translation quality. Recently, another progress has been taking place from recurrent neural network(RNN)-based NMT to transformer-based NMT (T-NMT). As the performance of NMT has evolved, a lot of research papers for machine translation have been published in the field of interpretation and translation. Their main focus is on whether machine translation can replace human translation, and analyzing the quality of translation results. In this paper, we briefly explain the history of the machine translation research and review the mechanism of NMT. NMT is basically composed of three parts: encoder, attention mechanism, and decoder. Further we discuss the new transformer structure based on the encoder-decoder model. We also discuss the challenges in NMT and explain the research direction or solutions to the problems. Particular attention is given to the mistranslation of NMT, quality of the translation, and robustness against the noises in the training dataset as well as in the testing sentences. In order to test the performance of transformer-based NMT, we used the Google NMT (GNMT) service for 4 languages – Korean, English, German, and Japanese. We confirmed the robustness against sentences with noises. However, we found unexpected volatility of NMT models where the input sentence is semantically and syntactically correct, resulting in critical degradation of translation quality.

최근 세계보건기구(WHO)의 Coronavirus Disease-19(COVID-19)에 대한 팬데믹 선언으로 COVID-19는 세계적인 관심사이며 많은 사망자가 속출하고 있다. 이를 극복하기 위하여 국가 간 정보 교환과 COVID-19 관련 대응 방안 등의 공유에 대한 필요성이 증대되고 있다. 하지만 언어적 경계로 인해 원활한 정보 교환 및 공유가 이루어지지 못하고 있는 실정이다. 이에 본 논문은 COVID-19 도메인에 특화 된 인공신경망 기반 기계번역(Neural Machine Translation(NMT)) 모델을 제안한다. 제안한 모델은 영어를 중심으로 프랑스어, 스페인어, 독일어, 이탈리아어, 러시아 어, 중국어 지원이 가능한 Transformer 기반 양방향 모델이다. 실험결과 BLEU 점수를 기준으로 상용화 시스템과 비교 하여 모든 언어 쌍에서 유의미한 높은 성능을 보였다.

With the recent World Health Organization (WHO) Declaration of Pandemic for Coronavirus Disease-19 (COVID-19), COVID-19 is a global concern and many deaths continue. To overcome this, there is an increasing need for sharing information between countries and countermeasures related to COVID-19. However, due to linguistic boundaries, smooth exchange and sharing of information has not been achieved. In this paper, we propose a Neural Machine Translation (NMT) model specialized for the COVID-19 domain. Centering on English, a Transformer based bidirectional model was produced for French, Spanish, German, Italian, Russian, and Chinese. Based on the BLEU score, the experimental results showed significant high performance in all language pairs compared to the commercialization system.

최신 기계번역 연구 동향을 살펴보면 대용량의 단일말뭉치를 통해 모델의 사전학습을 거친 후 병렬 말뭉치로 미세조정을 진행한다. 많은 연구에서 사전학습 단계에 이용되는 데이터의 양을 늘리는 추세이나, 기계번역 성능 향상을 위해 반드시 데이터의 양을 늘려야 한다고는 보기 어렵다. 본 연구에서는 병렬 말뭉치 필터링을 활용한 mBART 모델 기반의 실험을 통해, 더 적은 양의 데이터라도 고품질의 데이터라면 더 좋은 기계번역 성능을 낼 수 있음을 보인다. 실험결과 병렬 말뭉치 필터링을 거친 사전학습모델이 그렇지 않은 모델보다 더 좋은 성능을 보였다. 본 실험결과를 통 해 데이터의 양보다 데이터의 질을 고려하는 것이 중요함을 보이고, 해당 프로세스를 통해 추후 말뭉치 구축에 있어 하나의 가이드라인으로 활용될 수 있음을 보였다.

In the latest trend of machine translation research, the model is pretrained through a large mono lingual corpus and then finetuned with a parallel corpus. Although many studies tend to increase the amount of data used in the pretraining stage, it is hard to say that the amount of data must be increased to improve machine translation performance. In this study, through an experiment based on the mBART model using parallel corpus filtering, we propose that high quality data can yield better machine translation performance, even utilizing smaller amount of data. We propose that it is important to consider the quality of data rather than the amount of data, and it can be used as a guideline for building a training corpus.

딥러닝 모델을 활용한 인공신경망 기계번역 (Neural machine translation)이 주류 분야로 떠오르면서 최고의 성능을 위해 모델과 데이터 언어 쌍에 대한 많은 투자와 연구가 활발하게 진행되고 있다. 그러나, 최근 대부 분의 인공신경망 기계번역 연구들은 번역 문장의 품질을 극대화하는 자연어 생성을 위한 디코딩 전략 (Decoding strategy)에 대해서는 미래 연구 과제로 남겨둔 채 다양한 실험과 구체적인 분석이 부족한 상황이다. 기계번역에서 디코딩 전략은 번역 문장을 생성하는 과정에서 탐색 경로를 최적화 하고, 모델 변경 및 데이터 확장 없이도 성능 개선이 가능하다. 본 논문은 시퀀스 투 시퀀스 (Sequence to Sequence) 모델을 활용한 신경망 기반의 기계번역 에서 고전적인 그리디 디코딩 (Greedy decoding)부터 최신의 방법론인 Dynamic Beam Allocation (DBA)까지 비교 분석하여 디코딩 전략의 효과와 그 의의를 밝힌다.

Neural machine translation using deep neural network has emerged as a mainstream research, and an abundance of investment and studies on model structure and parallel language pair have been actively undertaken for the best performance. However, most recent neural machine translation studies pass along decoding strategy to future work, and have insufficient a variety of experiments and specific analysis on it for generating language to maximize quality in the decoding process. In machine translation, decoding strategies optimize navigation paths in the process of generating translation sentences and performance improvement is possible without model modifications or data expansion. This paper compares and analyzes the significant effects of the decoding strategy from classical greedy decoding to the latest Dynamic Beam Allocation (DBA) in neural machine translation using a sequence to sequence model.

As a case study, this paper presents an investigation on gender bias in neural machine translation with Naver Papago and Google Translate. In spite of the remarkable progress in machine translation system, there are remaining challenging tasks, especially the gender bias in machine translation and processing itself. This kind of phenomenon threatens the fairness of a translation system and runs the risk of amplifying biases in machine translation. Under the circumstances, this study examines gender bias in neural machine translation system with Google Translate and Naver Papago. As for the gender bias evaluation in two machine translation, this paper makes English-Korean and Korean-English language sets using a lexicon set of adjectives. Also, this study purposes to compare Naver Papago and Google Translate. Upon the qualitative examination, the findings revealed that Naver Papago and Google Translate are significantly prone to gender-biased translation and biased results against women are much more frequent. Also, Naver Papagoseems to be more abusive against women. On the other hand, Google Translate seems to be more balanced than Naver Papago, using gender-neutral vocabulary and structure. Moreover, Naver papago and Google Translate exhibit a tendency toward male defaults typically associated with unbalanced gender distribution or stereotypes.

인공신경망 기계번역(Neural Machine Translation, NMT)은 한정된 개수의 단어만을 번역에 이용하기 때문 에 사전에 등록되지 않은 단어들이 입력으로 들어올 가능성이 있다. 이러한 Out of Vocabulary(OOV) 문제를 완화하고자 고안된 방법이 서브 워드 분절(Subword Tokenization)이며, 이는 문장을 단어보다 더 작은 서브 워드 단위로 분할하여 단어를 구성하는 방법론이다. 본 논문에서는 일반적인 서브 워드 분절 알고리즘들을 다루며, 나아가 한국어의 무한한 용언 활용을 잘 다룰 수 있는 사전을 만들기 위해 한국어의 음절 중 종성을 분리하여 서브 워드 분절을 학습하는 새로운 방법론을 제안한다. 실험결과 본 논문에서 제안하는 방법론이 기존의 서브 워드 분리 방법론보다 높은 성능을 거두었다.

Since Neural Machine Translation (NMT) uses only a limited number of words, there is a possibility that words that are not registered in the dictionary will be entered as input. The proposed method to alleviate this Out of Vocabulary (OOV) problem is Subword Tokenization, which is a methodology for constructing words by dividing sentences into subword units smaller than words. In this paper, we deal with general subword tokenization algorithms. Furthermore, in order to create a vocabulary that can handle the infinite conjugation of Korean adjectives and verbs, we propose a new methodology for subword tokenization training by separating the Jongsung(coda) from Korean syllables (consisting of Chosung-onset, Jungsung-neucleus and Jongsung-coda). As a result of the experiment, the methodology proposed in this paper outperforms the existing subword tokenization methodology.

Neural Machine Translation (NMT) is known to produce better translation than Statistic Machine Translation(SMT). It can be natural that scholars and media show more interest in the emerging translation technology. However, it is difficult to define what makes a translation better than others. Furthermore, the claims about the NMT quality and wild guess about the future of translation are sometimes misleading and even worrisome. This means more objective and reliable diagnosis about the current state of NMT is necessary so that potential users and readers of machine translation can make informed decisions. Hence, this research aims to investigate if the media's approach, mainly driven by automatic evaluation, is valid and to review what has been considered as important aspects for translation evaluation in the previous translation studies. As a next step toward a more objective diagnosis, this research tries to compare three Neural Machine Translation (NMT) services available in Korea and assess their translation quality from various perspectives. In total, eight translation experts conducted a series of quality assessment on Korean to English and English and Korean translation. The analysis of the experts input will provide a few meaningful findings that can help build more objective yet prudent views on neural machine translation.

최근 딥러닝 기반 자연언어처리 연구들은 다양한 출처의 대용량 데이터들을 함께 학습하여 성능을 올리고자 하는 연구들을 진행하고 있다. 그러나 다양한 출처의 데이터를 하나로 합쳐서 학습시키는 방법론은 성능 향상을 막게 될 가능성이 존재한다. 기계번역의 경우 병렬말뭉치 간의 번역투(의역, 직역), 어체(구어체, 문어체, 격식체 등), 도메인 등의 차이로 인하여 데이터 편차가 발생하게 되는데 이러한 말뭉치들을 하나로 합쳐서 학습을 시키게 되면 성능의 악영 향을 미칠 수 있다. 이에 본 논문은 기계번역에서 병렬말뭉치 간의 균형성을 고려한 Corpus Weight Balance (CWB) 학습 방법론을 제안한다. 실험결과 말뭉치 간의 균형성을 고려한 모델이 그렇지 않은 모델보다 더 좋은 성능을 보였다. 더불어 단일 말뭉치로도 고품질의 병렬 말뭉치를 구축할 수 있는 휴먼번역 시장과의 상생이 가능한 말뭉치 구축 프로세 스를 추가로 제안한다.

Recent deep learning-based natural language processing studies are conducting research to improve performance by training large amounts of data from various sources together. However, there is a possibility that the methodology of learning by combining data from various sources into one may prevent performance improvement. In the case of machine translation, data deviation occurs due to differences in translation(liberal, literal), style(colloquial, written, formal, etc.), domains, etc. Combining these corpora into one for learning can adversely affect performance. In this paper, we propose a new Corpus Weight Balance(CWB) method that considers the balance between parallel corpora in machine translation. As a result of the experiment, the model trained with balanced corpus showed better performance than the existing model. In addition, we propose an additional corpus construction process that enables coexistence with the human translation market, which can build high-quality parallel corpus even with a monolingual corpus.