• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.25-30
• /
• 2023

초거대 언어모델은 과연 수수께끼 문제에 재치있는 답변을 할 수 있을까? 최근 초거대 언어모델(Large language model, LLM)은 강력한 성능 및 유저 만족도를 보이며 세계의 이목을 집중시키고 있다. 여러 태스크들에 대한 정량 평가를 넘어서 최근에는 LLM의 창의력 및 고도화된 언어능력을 평가하는 연구들이 등장하고 있다. 본 논문에서는 이러한 연구 흐름에 따라 LLM의 재치에 관해 고찰해본다. 이때 재치를 평가하기 위한 태스크로 이를 요구하는 말놀이인 수수께끼를 활용한다. 본 논문은 LLM이 수수께끼를 잘 수행하는지를 모델 추론을 통해 평가하며, 모델 추론 시 활용되는 프롬프트들의 성격에 따른 성능 변화를 관찰한다. 또한 수수께끼의 종류에 따른 모델의 능력을 비교 분석하며 LLM의 추론 결과에 대한 오류 분석을 수행한다. 본 논문은 실험을 통해 GPT-4가 가장 높은 성능을 보이며, 설명글이나 데이터 예시를 추가할 시 성능을 한층 더 향상시킬 수 있음을 확인한다. 또한 단어 기반보다는 특성 기반의 수수께끼에 더욱 강력한 성능을 보이며, 오류 유형 분석을 통해 LLM이 환각(hallucination) 문제와 창의력을 동시에 가지고 있다고 분석한다.

• The Transactions of the Korea Information Processing Society
• /
• v.13 no.4
• /
• pp.181-188
• /
• 2024

Traditional profanity detection methods have limitations in identifying intentionally altered profanities. This paper introduces a new method based on Named Entity Recognition, a subfield of Natural Language Processing. We developed a profanity detection technique using sequence labeling, for which we constructed a dataset by labeling some profanities in Korean malicious comments and conducted experiments. Additionally, to enhance the model's performance, we augmented the dataset by labeling parts of a Korean hate speech dataset using one of the large language models, ChatGPT, and conducted training. During this process, we confirmed that filtering the dataset created by the large language model by humans alone could improve performance. This suggests that human oversight is still necessary in the dataset augmentation process.

• Language and Information
• /
• v.7 no.2
• /
• pp.103-115
• /
• 2003

In recent years, many organizations have been constructing their own large corpora to achieve corpus representativeness. However, there is no reliable guideline as to how large corpus resources should be compiled, especially for Korean corpora. In this study, we have contrived a new statistical model, ARIMA (Autoregressive Integrated Moving Average), for predicting the relationship between linguistic items (the number of types) and corpus size (the number of tokens), overcoming the major flaws of several previous researches on this issue. Finally, we shall illustrate that the ARIMA model presented is valid, accurate and very reliable. We are confident that this study can contribute to solving some inherent problems of corpus linguistics, such as corpus predictability, corpus representativeness and linguistic comprehensiveness.

• The Journal of the Acoustical Society of Korea
• /
• v.23 no.3
• /
• pp.242-247
• /
• 2004

Conventional keyword spotting systems use the connected word recognition network consisted by keyword models and filler models in keyword spotting. This is why the system can not construct the language models of word appearance effectively for detecting keywords in large vocabulary continuous speech recognition system with large text data. In this paper to solve this problem, we propose a keyword spotting system using pseudo N-gram language model for detecting key-words and investigate the performance of the system upon the changes of the frequencies of appearances of both keywords and filler models. As the results, when the Unigram probability of keywords and filler models were set to 0.2, 0.8, the experimental results showed that CA (Correctly Accept for In-Vocabulary) and CR (Correctly Reject for Out-Of-Vocabulary) were 91.1% and 91.7% respectively, which means that our proposed system can get 14% of improved average CA-CR performance than conventional methods in ERR (Error Reduction Rate).

• Journal of Intelligence and Information Systems
• /
• v.29 no.3
• /
• pp.125-144
• /
• 2023

Recently, the use of Large Language Models (LLMs) such as ChatGPT has been increasing in various fields such as interactive commerce and mobile financial services. However, LMMs, which are mainly created by learning existing documents, can also learn various human biases inherent in documents. Nevertheless, there have been few comparative studies on the aspects of bias and discrimination in LLMs. The purpose of this study is to examine the existence and extent of nine types of discrimination (Age, Disability status, Gender identity, Nationality, Physical appearance, Race ethnicity, Religion, Socio-economic status, Sexual orientation) in LLMs and suggest ways to improve them. For this purpose, we utilized BBQ (Bias Benchmark for QA), a tool for identifying discrimination, to compare three large-scale language models including ChatGPT, GPT-3, and Bing Chat. As a result of the evaluation, a large number of discriminatory responses were observed in the mega-language models, and the patterns differed depending on the mega-language model. In particular, problems were exposed in elder discrimination and disability discrimination, which are not traditional AI ethics issues such as sexism, racism, and economic inequality, and a new perspective on AI ethics was found. Based on the results of the comparison, this paper describes how to improve and develop large-scale language models in the future.

• Proceedings of the Acoustical Society of Korea Conference
• /
• spring
• /
• pp.33-36
• /
• 2002

We implemented a robust Korean broadcast news transcription system for out-of-vocabulary (OOV), tested its performance. The occurrence of OOV words in the input speech is inevitable in large vocabulary continuous speech recognition (LVCSR). The known vocabulary will never be complete due to the existence of for instance neologisms, proper names, and compounds in some languages. The fixed vocabulary and language model of LVCSR system directly face with these OOV words. Therefore our Broadcast news recognition system has an offline OOV update module of language model and vocabulary to solve OOV problem and selects morpheme-based recognition unit (so called, pseudo-morpheme) for OOV robustness.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.399-403
• /
• 2023

본 논문에서는 다중 문서 기반의 대화 시스템을 통한 효율적인 정보 검색과 응답 생성에 중점을 둡니다. 대규모 데이터 집합에서 정확한 문서를 선택하는 데 필요한 검색의 중요성을 강조하며, 현재 검색 방법의 한계와 문제점을 지적합니다. 또한 더 자연스러운 답변을 생성하기 위해 대규모 언어 모델을 사용하게 되면서 fine-tuning 시에 발생하는 제약과 낭비를 모델의 제로샷 생성 능력을 활용하여 개선하려는 방안을 제안하며, 모델의 크기와 자원의 효율성에 대한 고려사항을 논의합니다. 우리의 접근 방식은 대규모 언어 모델을 프롬프트와 함께 다중 문서로 학습 없이 정보를 검색하고 응답을 생성하는 방향으로 접근하여 대화 시스템의 효율성과 유용성을 향상시킬 수 있음을 제시합니다.

• Phonetics and Speech Sciences
• /
• v.15 no.3
• /
• pp.83-88
• /
• 2023

This study investigates the fine-tuning of large-scale Automatic Speech Recognition (ASR) models, specifically OpenAI's Whisper model, for domain-specific applications using the KsponSpeech dataset. The primary research questions address the effectiveness of targeted lexical item emphasis during fine-tuning, its impact on domain-specific performance, and whether the fine-tuned model can maintain generalization capabilities across different languages and environments. Experiments were conducted using two fine-tuning datasets: Set A, a small subset emphasizing specific lexical items, and Set B, consisting of the entire KsponSpeech dataset. Results showed that fine-tuning with targeted lexical items increased recognition accuracy and improved domain-specific performance, with generalization capabilities maintained when fine-tuned with a smaller dataset. For noisier environments, a trade-off between specificity and generalization capabilities was observed. This study highlights the potential of fine-tuning using minimal domain-specific data to achieve satisfactory results, emphasizing the importance of balancing specialization and generalization for ASR models. Future research could explore different fine-tuning strategies and novel technologies such as prompting to further enhance large-scale ASR models' domain-specific performance.

• Journal of Intelligence and Information Systems
• /
• v.29 no.4
• /
• pp.165-184
• /
• 2023

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.

• Journal of Sensor Science and Technology
• /
• v.33 no.2
• /
• pp.78-85
• /
• 2024

We proposed a method for visual place recognition that represents images using objects as visual words. Visual words represent the various objects present in urban environments. To detect various objects within the images, we implemented and used a zero-shot detector based on a large-scale image language model. This zero-shot detector enables the detection of various objects in urban environments without additional training. In the process of creating histograms using the proposed method, frequency-based weighting was applied to consider the importance of each object. Through experiments with open datasets, the potential of the proposed method was demonstrated by comparing it with another method, even in situations involving environmental or viewpoint changes.