Long Short-Term Memory에 기반한 한국어 언어모델 연구

이선정(Sunjeong Lee) 융복합지식학회 2020 융복합지식학회논문지 Vol.8 No.1

본고에서는 LSTM에 기반을 둔 한국어 언어모델에 관한 연구를 수행하였으며 GloVe를 사용하는 LSTM 기반 언어모델을 제안한다. 먼저 PTB 영어 코퍼스를 이용하여 전통적인 n-gram 방식의 통계적인 언어모델과 LSTM 기반의 언어모델을 비교하였고 그 결과 47.3%의 복잡도가 감소되는 효과를 얻었다. 한국어에 적용 실험을 위해서 기본 토큰 단위로 WPM (word-piece model)을 사용하였으며 통계적인 n-gram 언어모델과 LSTM 언어모델을 비교한다. 또한, LSTM 언어모델을 만들 때 GloVe를 단어 표현 벡터로 사용하는 방법을 제안하여 비교 연구도 수행하였다. 한국어 평가 코퍼스 10만 문장을 이용하여 성능 비교를 한 결과 LSTM 방식을 사용하였을 경우n-gram 방식보다 28.8%의 복잡도가 감소하였고 GloVe와 같이 사용할 경우 43.4%의 복잡도가 감소되었다. 영어와 한국어 코퍼스의 비교 실험으로 GloVe를 사용하는 LSTM 기반 언어모델의 제안이 우수하다는 것을 입증하였다. In this paper, we make a comparative study on the language model based on long short-term memory (LSTM) and propose a language model based on LSTM using GloVe as a word representation vector. For this purpose, traditional n-gram statistical language model is compared with LSTM language model using PTB English corpus. The experimental result yields that LSTM language model get the perplexity (ppl) reduction of 47.3% compared with traditional n-gram model. In order to expand this approach to Korean language, we design a language model of which basic unit is word-piece model (WPM). And we also make a comparative study of statistical language model and neural language model. Especially, we propose a LSTM language model using glove vector (GloVe) as a word representation vector. For our study, 100,000 Korean sentences are used as a test set. Our experimental result yields that LSTM language model get the reduction of 28.8% compared with n-gram language model and LSTM with GloVe get the reduction of 43.3%. In conclusion, we show that the proposed language model is good approach as a language model.

외국인 모델의 매력도와 외국어 사용의 상호작용 효과

이지현(Lee Ji-hyun),이동일(Dong Il Lee) 한국마케팅과학회 2007 마케팅과학연구 Vol.17 No.3

현재 한국시장의 광고에는 외국인 모델과 외국어를 사용하는 것이 일반적이다. 그러나 그 효과에 대한 검증은 거의 이루어지지 않고 있다. 기존의 연구에 의하면 마케팅 커뮤니케이션과 소비자의 문화적 가치가 일치하게 될 때 커뮤니케이션의 효과가 높아진다고 한다. 즉, 광고에의 외국문화 사용은 기존 연구에서 주장하고 있는 바와는 반대의 현상인 것이다. 그러나 인터넷 등 글로벌 매체에 의해 글로벌 문화가 출현함에 따라 마케팅 커뮤니케이션과 소비자의 문화적 가치가 일치하는 것이 가장 바람직한 커뮤니케이션 방법이라 볼 수는 없다. 이에 본 연구는 외국인 모델을 사용하는 광고에 있어서 모델의 매력도와 외국어 사용이 소비자의 광고와 제품에 대한 태도, 구매의도 등에 어떤 영향을 미치는가를 살펴보고 효과적인 광고 커뮤니케이션에 대한 제언을 하고자 했다. 연구결과, 광고에 사용된 언어와 광고의 시각적 단서인 모델 매력도는 광고의 효과에 상호작용 효과를 나타내는 것이 밝혀졌다. 광고에 사용된 모델 매력도가 높은 경우, 제품에 대한 태도, 구매의도가 사용된 언어에 대해 유의한 차이를 보이지 않았다. 광고에 사용된 모델의 매력도가 낮은 경우에는 영어를 사용한 경우가 제품에 대한 태도, 구매의도가 높았다. 즉, 외국인 모델을 광고에 사용할 경우, 그 모델의 매력도의 높고 낮음에 따라 광고에 사용되는 언어를 선택함으로써 광고의 효과를 조절할 수 있다는 실무적 시사점을 제공한다. sing to increase the effect of advertising is general. However, recent studies show that targeted women audience tend to compare themselves to the highly attractive models and do experience negative sentiment. Bower (2001) proved the difference between "comparer" and "noncomparer" when women face highly attractive models. The results show that a comparer who has an intention to compare highly attractive model (HAM) with herself has a significantly negative effect on model expertise, product argument, product evaluation and buying intention. Therefore, HAM is not always a good choice and model attractiveness plays a role in the processing other cues or changing the advertising effect from result of processing other cues. The purpose of this study is to investigate the effect of the use of foreign language on the advertising response of the audience with regard of the model attractiveness. For the empirical study, the virtual advertising using foreign models (HAM, NAM), brand names and slogans (Korean, English) were used as stimuli. The respondents of each stimulus were 75("HAM-Korean"), 75("NAM-Korean"), 66("HAM-English") and 66("NAM-English") respectively. To establish the effect of marketing communication, the attitude for media(AM), the attitude for product(AP), targetedness(TD), overall quality(OQ), and purchase intention(PI) with 7 point likert scale were measured. The manipulation was verified to check the difference between HAM attractiveness assessment (m=3.27) and NAM attractiveness assessment (m=5.12). The mean difference was statiscally significant (P<05). As a result, all consequences were significantly changed with model attractiveness, and overall quality evaluation(OQ) were significantly changed with language. The interaction effect from model attractiveness and language was significant on attitude toward the product(AP) and purchase intention(PI). To analyze the difference, the mean values and standard deviation of consequences were compared. The result was more positive when model attractiveness was high for all consequences. For language effect, the assessment was more positive when English was used for OQ. Considering model attractiveness and language simultaneously, HAM-Korean was more positive for AP and PI, and NAM-English was more positive for AP and PI. In other words, the interaction effect was confirmed by model attractiveness and language. As mentioned above, use of foreign models and foreign language in advertising was explained by cultural match up hypothesis (Leclerc et al. 1994) which claimed that culture of origin effect. In other words, in advertising, use of same cultural language with the foreign model could make positive assessment for OQ. But this effect was moderated by model attractiveness. When the model attractiveness was low, the use of English makes PI high because of the effect of foreign language which supported the cultural match up hypothesis. ""When the model attractiveness was low, the use of Korean made AP and PI high because the effect of foreign language was diluted. It was a general notion that the visual cues got processed before (Holbrook and Moore, 1981; Sholl et al, 1995) compared to linguistic cues. Therefore, when consumers were faced HAM, so much perception was already consumed at processing visual cues making their native language of Korean to strongly and positively connected with the advertising concept. On the contrary, when consumers were faced with NAM, less perception was consumed compared to HAM, making English to accompany cultural halo effect which affected more positively. Therefore, when foreign models were employed in advertising, the language must be carefully selected according to the level of model attractiveness.

언어영재와 일반 영재를 위한 이원화 언어 교수-학습모형

김주연 국어문학회 2015 국어문학 Vol.60 No.-

The purpose of this study is to develop dual learning models for the gifted in language and the others. It is proposed also a learning model for the gifted in language 'linguistic talent-development model', And the language-learning model for the gifted in general. "linguistic model fulfillment verbal -grounding’. The goal of 'Linguistic talent-development model "is improving special competence of the gifted in language more and more. So they will be provided and supported various ways that appropriate their linguistic aptitude. That is based on individual instruction model including individual language competence and educational environmental factors, cognitive, and emotional characteristics. Next, ‘Linguistic model fulfillment verbal -grounding’ is used frame of language training program for the gifted in general. In the future, Professionalism is completed how to using of language. So the goal of this model is practicing and studying about the high quality expressions. Listening and speaking, reading and writing to their major This model is based on integrated education model. This study proposed two models of language learning program for the gifted in language and the other gifted students. This dual model show the importance of language in gifted education. And this study can prove the identification of language giftedness and propose direction of languge gifted education. 본 연구의 목적은 언어영재와 일반 영재의 특성과 교육 목적에 적합한 이원화 언어교수-학습 모형을 개발하는 데 있다. 언어영재를 위한 언어 교수- 학습모형으로는 ‘언어적 재능 계발 모형’을 제안되었고, 일반 영재를 위한 언어 교수-학습모형으로는 ‘언어적 소양 함양 모형’을 제안되었다. ‘언어적 재능계발 모형’은 영역 특수적으로 언어영역에 탁월한 영재성을 지니고 있는 학생들을 위한 모형으로서 자신의 언어적성에 맞는 분야의 기능을 더욱 향상시킬수 있도록 지원해 주는 모형이다. 개별화 교육모형에 근거하고 있으며, 개개인의 언어적성과 교육적 환경 요인, 인지적・정서적 특성 등을 충분히 고려하여 전략을 세운 뒤 적용하여 산출물을 생산해내는 단계를 거친다. ‘언어적 소양 함양 모형’은 일반 영재를 위한 언어 교육 프로그램 설계 시 활용될 수 있는 모형이다. 듣기와 말하기, 읽기, 쓰기 등 의사소통의 도구로서 활용되는 언어의 특성을 고려하여 어떤 분야의 영재라 할지라도 자신이 연구한 것을 언어적으로 정확하고 효율적으로 소통할 수 있는 언어능력을 함양시켜 주는 데 목적이 있다. 이 모형은 통합교육모형을 기본 바탕으로 하고 있다. 본 연구는 언어영재의 특징에 맞는 교수-학습 모형과 일반 영재를 위한 언어 교수-학습모형의 이원화 언어 교수-학습모형을 제시하여 영재교육에 있어 언어가 차지하고 있는 중요성과 언어영재의 정체성 그리고 언어영재 교육의 방향 등을 제시하였다는 데 큰 의의와 가치가 있다.

KoBERT와 KoGPT2 기반의 대형언어모델과 딥러닝을 통합한 리뷰 유용성 예측모형

김은미,남승진,김태이,홍태호 한국지능정보시스템학회 2024 지능정보연구 Vol.30 No.2

AI 기술이 산업 전반에서 광범위하게 적용되면서 텍스트 데이터 기반의 대형언어모델이 높은 관심을 받고 있다. 대형 언어모델은 번역, 챗봇, 콘텐츠 생성 등과 같은 자연어 처리 분야에서 활발히 연구되고 있으며, 이커머스 분야에서도 고객 데이터 분석을 위해 사용되고 있다. 제품 및 서비스에 대한 사용 경험을 기반으로 사용자가 직접 작성하는 온라인 리뷰는 고객분석을 위한 중요한 자료이며, 대형언어모델의 활용은 텍스트로 작성되어 있는 리뷰 데이터의 의미 파악을 보다 정확 하게 할 수 있도록 한다. 본 연구는 SVM, 1D-CNN, 2D-CNN, CNN-LSTM의 딥러닝 모델과 대형 언어 모델인 KoBERT 와 KoGPT2로 리뷰 유용성 예측모형을 구축하고, 이를 통합하여 텍스트 내의 복잡한 의미가 반영된 리뷰 유용성 예측모 형을 제안한다. 구글 지도의 리뷰 데이터를 활용하였으며, 딥러닝 기법에서는 CNN-LSTM의 예측 성과가 72.74%로 가장 우수한 것으로 나타났다. KoBERT와 KoGPT2의 대형언어모델은 73.22%와 75.74%로 기존의 머신러닝 기법의 예측 모델 보다 대형언어모델을 기반으로 한 예측 모형이 우수한 성능을 보였다. 본 연구에서 제안한 딥러닝 기법과 대형 언어 모델을 통합한 통합모형에서는 76.37%의 정확도로 예측성과를 향상시켰으며, 통합모형은 텍스트의 의미를 보다 정확하게 반영하고, 예측성과를 향상시키며 예측모형의 안정성을 높일 수 있다. As AI technology is widely applied across industries, text data-based large language models are gaining significant attention. Large language models are actively researched in natural language processing fields such as translation, chatbots, and content creation, and are also used for customer data analysis in e-commerce. Online reviews, which are written directly by customers based on their experiences with products and services, are crucial for customer analysis, and leveraging large language models can help better understand the meanings embedded in these text reviews. This study proposes a review helpfulness prediction model by integrating deep learning models such as SVM, 1D-CNN, 2D-CNN, CNN-LSTM, and large language models KoBERT and KoGPT2, thereby reflecting the complex semantics within the text. Experiment results indicate that the CNN-LSTM model showed the best prediction performance at 72.74%. The large language models KoBERT and KoGPT2 achieved 73.22% and 75.74%, respectively, showing that prediction models based on large language models performed better than traditional machine learning models. The integrated model, combining the deep learning techniques and large language models proposed in this study, improved prediction performance with an accuracy of 76.37%, indicating that the integrated model can more accurately reflect the text’s meaning, enhance prediction performance, and improve model stability.

한국어 음소 단위 LSTM 언어모델을 이용한 문장 생성

안성만(SungMahn Ahn),정여진(Yeojin Chung),이재준(Jaejoon Lee),양지헌(Jiheon Yang) 한국지능정보시스템학회 2017 지능정보연구 Vol.23 No.2

Language models were originally developed for speech recognition and language processing. Using a set of example sentences, a language model predicts the next word or character based on sequential input data. N-gram models have been widely used but this model cannot model the correlation between the input units efficiently since it is a probabilistic model which are based on the frequency of each unit in the training set. Recently, as the deep learning algorithm has been developed, a recurrent neural network (RNN) model and a long short-term memory (LSTM) model have been widely used for the neural language model (Ahn, 2016; Kim et al., 2016; Lee et al., 2016). These models can reflect dependency between the objects that are entered sequentially into the model (Gers and Schmidhuber, 2001; Mikolov et al., 2010; Sundermeyer et al., 2012). In order to learning the neural language model, texts need to be decomposed into words or morphemes. Since, however, a training set of sentences includes a huge number of words or morphemes in general, the size of dictionary is very large and so it increases model complexity. In addition, word-level or morpheme-level models are able to generate vocabularies only which are contained in the training set. Furthermore, with highly morphological languages such as Turkish, Hungarian, Russian, Finnish or Korean, morpheme analyzers have more chance to cause errors in decomposition process (Lankinen et al., 2016). Therefore, this paper proposes a phoneme-level language model for Korean language based on LSTM models. A phoneme such as a vowel or a consonant is the smallest unit that comprises Korean texts. We construct the language model using three or four LSTM layers. Each model was trained using Stochastic Gradient Algorithm and more advanced optimization algorithms such as Adagrad, RMSprop, Adadelta, Adam, Adamax, and Nadam. Simulation study was done with Old Testament texts using a deep learning package Keras based the Theano. After pre-processing the texts, the dataset included 74 of unique characters including vowels, consonants, and punctuation marks. Then we constructed an input vector with 20 consecutive characters and an output with a following 21st character. Finally, total 1,023,411 sets of input-output vectors were included in the dataset and we divided them into training, validation, testsets with proportion 70:15:15. All the simulation were conducted on a system equipped with an Intel Xeon CPU (16 cores) and a NVIDIA GeForce GTX 1080 GPU. We compared the loss function evaluated for the validation set, the perplexity evaluated for the test set, and the time to be taken for training each model. As a result, all the optimization algorithms but the stochastic gradient algorithm showed similar validation loss and perplexity, which are clearly superior to those of the stochastic gradient algorithm. The stochastic gradient algorithm took the longest time to be trained for both 3- and 4-LSTM models. On average, the 4-LSTM layer model took 69% longer training time than the 3-LSTM layer model. However, the validation loss and perplexity were not improved significantly or became even worse for specific conditions. On the other hand, when comparing the automatically generated sentences, the 4-LSTM layer model tended to generate the sentences which are closer to the natural language than the 3-LSTM model. Although there were slight differences in the completeness of the generated sentences between the models, the sentence generation performance was quite satisfactory in any simulation conditions: they generated only legitimate Korean letters and the use of postposition and the conjugation of verbs were almost perfect in the sense of grammar. The results of this study are expected to be widely used for the processing of Korean language in the field of language processing and speech recognition, which are the basis of artificial intelligence systems.

프라이버시 보호를 위한 오프사이트 튜닝 기반 언어모델 미세 조정 방법론

정진명,김남규 한국지능정보시스템학회 2023 지능정보연구 Vol.29 No.4

최근 구글의 BERT, OpenAI의 GPT 등, 언어모델(Language Model)을 사용한 비정형 텍스트 데이터에 대한 딥러닝 (Deep Learning) 분석이 다양한 응용에서 괄목할 성과를 나타내고 있다. 대부분의 언어모델은 사전학습 데이터로부터 범용적인 언어정보를 학습하고, 이후 미세 조정(Fine-Tuning) 과정을 통해 다운스트림 태스크(Downstream Task)에 맞추어 갱신되는 방식으로 사용되고 있다. 하지만 최근 이러한 언어모델을 사용하는 과정에서 프라이버시가 침해될 수 있다는 우려가 제기되고 있다. 즉 데이터 소유자가 언어모델의 미세 조정을 수행하기 위해 다량의 데이터를 모델 소유자에게 제공 하는 과정에서 데이터의 프라이버시가 침해될 수 있으며, 반대로 모델 소유자가 모델 전체를 데이터 소유자에게 공개하면 모델의 구조 및 가중치가 공개되어 모델의 프라이버시가 침해될 수 있다는 것이다. 이러한 상황에서 프라이버시를 보호하며 언어모델의 미세 조정을 수행하기 위해 최근 오프사이트 튜닝(Offsite Tuning)의 개념이 제안되었으나, 해당 연구는 제안 방법론을 텍스트 분류 모델에 적용하는 구체적인 방안을 제시하지 못했다는 한계를 갖는다. 이에 본 연구에서는 한글 문서에 대한 다중 분류 미세 조정 수행 시, 모델과 데이터의 프라이버시를 보호하기 위해 분류기를 추가한 오프사이트 튜닝을 적용하는 구체적인 방법을 제시한다. 제안 방법론의 성능을 평가하기 위해 AIHub에서 제공하는 ICT, 전기, 전자, 기계, 그리고 의학 총 5개의 대분야로 구성된 약 20만건의 한글 데이터에 대해 실험을 수행한 결과, 제안하는 플러그인 모델이 제로 샷 모델 및 오프사이트 모델에 비해 분류 정확도 측면에서 우수한 성능을 나타냄을 확인하였다. Recently, Deep learning analysis of unstructured text data using language models, such as Google’s BERT and OpenAI’s GPT has shown remarkable results in various applications. Most language models are used to learn generalized linguistic information from pre-training data and then update their weights for downstream tasks through a fine-tuning process. However, some concerns have been raised that privacy may be violated in the process of using these language models, i.e., data privacy may be violated when data owner provides large amounts of data to the model owner to perform fine-tuning of the language model. Conversely, when the model owner discloses the entire model to the data owner, the structure and weights of the model are disclosed, which may violate the privacy of the model. The concept of offsite tuning has been recently proposed to perform fine-tuning of language models while protecting privacy in such situations. But the study has a limitation that it does not provide a concrete way to apply the proposed methodology to text classification models. In this study, we propose a concrete method to apply offsite tuning with an additional classifier to protect the privacy of the model and data when performing multi-classification fine-tuning on Korean documents. To evaluate the performance of the proposed methodology, we conducted experiments on about 200,000 Korean documents from five major fields, ICT, electrical, electronic, mechanical, and medical, provided by AIHub, and found that the proposed plug-in model outperforms the zero-shot model and the offsite model in terms of classification accuracy.

Cross-Lingual Post-Training (XPT)을 위한 한국어 및 다국어 언어모델 연구

손수현,박찬준,이정섭,심미단,이찬희,박기남,임희석 한국융합학회 2022 한국융합학회논문지 Vol.13 No.3

It has been proven through many previous researches that the pretrained language model with a large corpus helps improve performance in various natural language processing tasks. However, there is a limit to building a large-capacity corpus for training in a language environment where resources are scarce. Using the Cross-lingual Post-Training (XPT) method, we analyze the method's efficiency in Korean, which is a low resource language. XPT selectively reuses the English pretrained language model parameters, which is a high resource and uses an adaptation layer to learn the relationship between the two languages. This confirmed that only a small amount of the target language dataset in the relationship extraction shows better performance than the target pretrained language model. In addition, we analyze the characteristics of each model on the Korean language model and the Korean multilingual model disclosed by domestic and foreign researchers and companies. 대용량의 코퍼스로 학습한 사전학습 언어모델이 다양한 자연어처리 태스크에서 성능 향상에 도움을 주는 것은 많은 연구를 통해 증명되었다. 하지만 자원이 부족한 언어 환경에서 사전학습 언어모델 학습을 위한 대용량의 코퍼스를 구축하는데는 한계가 있다. 이러한 한계를 극복할 수 있는 Cross-lingual Post-Training (XPT) 방법론을 사용하여 비교적 자원이 부족한 한국어에서 해당 방법론의 효율성을 분석한다. XPT 방법론은 자원이 풍부한 영어의 사전학습 언어모델의 파라미터를 필요에 따라 선택적으로 재활용하여 사용하며 두 언어 사이의 관계를 학습하기 위해 적응계층을 사용한다. 이를 통해 관계추출 태스크에서 적은 양의 목표 언어 데이터셋만으로도 원시언어의 사전학습 모델보다 우수한 성능을 보이는 것을 확인한다. 더불어, 국내외 학계와 기업에서 공개한 한국어 사전학습 언어모델 및 한국어 multilingual 사전학습 모델에 대한 조사를 통해 각 모델의 특징을 분석한다

DeNERT: DQN과 BERT를 이용한 개체명 인식 모델

양성민(Sung-Min Yang),정옥란(Ok-Ran Jeong) 한국컴퓨터정보학회 2020 한국컴퓨터정보학회논문지 Vol.25 No.4

In this paper, we propose a new structured entity recognition DeNERT model. Recently, the field of natural language processing has been actively researched using pre-trained language representation models with a large amount of corpus. In particular, the named entity recognition, which is one of the fields of natural language processing, uses a supervised learning method, which requires a large amount of training dataset and computation. Reinforcement learning is a method that learns through trial and error experience without initial data and is closer to the process of human learning than other machine learning methodologies and is not much applied to the field of natural language processing yet. It is often used in simulation environments such as Atari games and AlphaGo. BERT is a general-purpose language model developed by Google that is pre-trained on large corpus and computational quantities. Recently, it is a language model that shows high performance in the field of natural language processing research and shows high accuracy in many downstream tasks of natural language processing. In this paper, we propose a new named entity recognition DeNERT model using two deep learning models, DQN and BERT. The proposed model is trained by creating a learning environment of reinforcement learning model based on language expression which is the advantage of the general language model. The DeNERT model trained in this way is a faster inference time and higher performance model with a small amount of training dataset. Also, we validate the performance of our model’s named entity recognition performance through experiments.

Summary and Comparison of Five Communicative Competence Models

Gilja Byun 한국언어연구학회 2015 언어학연구 Vol.20 No.1

Byun, Gilja. 2015. Summary and comparison of five communicative competence models. Journal of Linguistics Studies, 20(1), 119-132. On the basis of communicative competence models suggested by Canale & Swain (1980), Canale (1983), Bachman (1990), Bachman and Palmer (1996) and Chapell, Grabe and Berns (1997), this paper deals with not only their summaries and comparisons but also their influence on second language teaching and testing. As the result of analyzing the main changes of the models in a chronological order, the following theoretical points are emphasized: 1) language use ability as well as language knowledge, 2) the integrative or holistic model through interactions among components in contrast to the discrete model, 3) the hierarchical and multi-level model instead of the linear model, 4) appropriate language use in a certain context by introducing sociolinguistic competence, 5) function in language use by presenting pragmatic competence, 6) cognitive skills by introducing and elaborating strategic competence, 7) affective factors and personal characteristics in performance. In accordance with these theoretical changes of the models, second language teaching and testing methods should be innovated. Language testing should include language performance as well as language ability and it should be an integrative test reflecting four skills. Likewise, teaching methods should be developed so that learners can use their strategic, functional, sociolinguistic and background competence, but not the previous memorization method in de-contextualized situations. This paper will help to design language testing and teaching methods more effectively. (Macquarie University)

군사용어 Word2Vec과 sLLM 구축 방안 연구

권주영,문호석 한국국방경영분석학회 2025 한국국방경영분석학회지 Vol.51 No.2

최근 급격히 발전한 대규모 언어모델 기술은 국방 분야의 문제 해결과 과업 수행에 큰 잠재력을 지닌다. 그러나 군사용어의 특수성과 작전 보안의 민감성으로 인해 범용 언어모델을 직접 활용하기에는 한계가 존재한다. 군사용어는 일상 언어와는 다른 문맥적 의미를 지니며 국방 분야의 고유한 특수성을 반영해야 한다. 또한, 군 내부망에서도 안전하게 운용할 수 있는 국방분야에 특화된 경량화 언어모델의 구축 필요하다. 본 연구는 국방 분야에 특화된 경량화 언어모델을 성공적으로 구축하는 방법론을 제시하고, 또 잘 활용하기 위한 정책적 방안을 제시하는 데 목적이 있다. 이를 위해 군사용어의 언어적 특성을 분석하고 경량화언어모델에 군사용어를 학습하여 성능을 확인해봄으로써 군사용어 데이터의 확보 및 정제방법, 언어모델 구축 전반에 필요한 요소들을 체계적으로 제시한다. 아울러 우리 군의 환경에 적합한 정책 방향을 제시하였다. 이를 통해 본 연구는 경량화 언어모델 구축을 위한 실질적 정책 수립에 기여하며 언어모델을 기반으로 한국방 역량 강화를 위한 기초를 마련하는 데 기여하고자 한다. Recent rapid advancements in large language model (LLM) technology hold significant potential for problem-solving and task performance in the defense sector. However, limitations exist in directly utilizing gen- eral-purpose language models due to the specificity of military language and the sensitivity of operational security. Military language carries different contextual meanings compared to everyday language and must reflect the unique characteristics of the defense domain. Furthermore, there is a need for the construction of lightweight language models that can be safely operated even within military internal networks. This study aims to propose policy measures for successfully constructing and utilizing lightweight lan- guage models specialized for the defense sector. To this end, by analyzing the linguistic characteristics of military language and validating the performance of lightweight language models trained on military language, the study systematically presents methods for acquiring and refining military language data, and the overall necessary elements for language model construction, and further derives policy directions suit- able for the environment of our military. This study contributes to the practical policy formulation for lightweight language model construction and holds policy significance in laying the foundation for strengthening defense capabilities based on language models.