• Phonetics and Speech Sciences
• /
• v.7 no.4
• /
• pp.3-8
• /
• 2015

In this paper, we investigate an input dimension reduction method using continuous word vector in deep neural network language model. In the proposed method, continuous word vectors were generated by using Google's Word2Vec from a large training corpus to satisfy distributional hypothesis. 1-of-${\left|V\right|}$ coding discrete word vectors were replaced with their corresponding continuous word vectors. In our implementation, the input dimension was successfully reduced from 20,000 to 600 when a tri-gram language model is used with a vocabulary of 20,000 words. The total amount of time in training was reduced from 30 days to 14 days for Wall Street Journal training corpus (corpus length: 37M words).

• The Journal of Bigdata
• /
• v.7 no.2
• /
• pp.11-29
• /
• 2022

Pre-trained language models are the most important and widely used tools in natural language processing tasks. Since those have been pre-trained for a large amount of corpus, high performance can be expected even with fine-tuning learning using a small number of data. Since the elements necessary for implementation, such as a pre-trained tokenizer and a deep learning model including pre-trained weights, are distributed together, the cost and period of natural language processing has been greatly reduced. Transformer variants are the most representative pre-trained language models that provide these advantages. Those are being actively used in other fields such as computer vision and audio applications. In order to make it easier for researchers to understand the pre-trained language model and apply it to natural language processing tasks, this paper describes the definition of the language model and the pre-learning language model, and discusses the development process of the pre-trained language model and especially representative Transformer variants.

• Journal of Intelligence and Information Systems
• /
• v.29 no.1
• /
• pp.27-40
• /
• 2023

With the increase in the spread of smart devices and the impact of COVID-19, the consumption of media contents through smart devices has significantly increased. Along with this trend, the amount of media contents viewed through OTT platforms is increasing, that makes contents recommendations on these platforms more important. Previous contents-based recommendation researches have mostly utilized metadata that describes the characteristics of the contents, with a shortage of researches that utilize the contents' own descriptive metadata. In this paper, various text data including titles and synopses that describe the contents were used to recommend similar contents. KLUE-RoBERTa-large, a Korean language model with excellent performance, was used to train the model on the text data. A dataset of over 20,000 contents metadata including titles, synopses, composite genres, directors, actors, and hash tags information was used as training data. To enter the various text features into the language model, the features were concatenated using special tokens that indicate each feature. The test set was designed to promote the relative and objective nature of the model's similarity classification ability by using the three contents comparison method and applying multiple inspections to label the test set. Genres classification and hash tag classification prediction tasks were used to fine-tune the embeddings for the contents meta text data. As a result, the hash tag classification model showed an accuracy of over 90% based on the similarity test set, which was more than 9% better than the baseline language model. Through hash tag classification training, it was found that the language model's ability to classify similar contents was improved, which demonstrated the value of using a language model for the contents-based filtering.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.19 no.2
• /
• pp.73-81
• /
• 2024

Recently, mobile manipulators have been utilized in agriculture industry for weed removal and harvest automation. This paper proposes a domain adaptive fruit detection method for harvest automation, by utilizing OWL-ViT model which is an open-vocabulary object detection model. The vision-language model can detect objects based on text prompt, and therefore, it can be extended to detect objects of undefined categories. In the development of deep learning models for real-world problems, constructing a large-scale labeled dataset is a time-consuming task and heavily relies on human effort. To reduce the labor-intensive workload, we utilized a large-scale public dataset as a source domain data and employed a domain adaptation method. Adversarial learning was conducted between a domain discriminator and feature extractor to reduce the gap between the distribution of feature vectors from the source domain and our target domain data. We collected a target domain dataset in a real-like environment and conducted experiments to demonstrate the effectiveness of the proposed method. In experiments, the domain adaptation method improved the AP50 metric from 38.88% to 78.59% for detecting objects within the range of 2m, and we achieved 81.7% of manipulation success rate.

• Journal of Korea Society of Industrial Information Systems
• /
• v.28 no.5
• /
• pp.31-39
• /
• 2023

The evolution of Natural Language Processing (NLP) and the rise of large language models (LLM) like ChatGPT have paved the way for specialized question-answering (QA) systems tailored to specific domains. This study outlines a system harnessing the power of LLM in conjunction with document search algorithms to interpret and address user inquiries using documents from the Korea Atomic Energy Research Institute (KAERI). Initially, the system refines multiple documents for optimized search and analysis, breaking the content into managable paragraphs suitable for the language model's processing. Each paragraph's content is converted into a vector via an embedding model and archived in a database. Upon receiving a user query, the system matches the extracted vectors from the question with the stored vectors, pinpointing the most pertinent content. The chosen paragraphs, combined with the user's query, are then processed by the language generation model to formulate a response. Tests encompassing a spectrum of questions verified the system's proficiency in discerning question intent, understanding diverse documents, and delivering rapid and precise answers.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.14 no.3
• /
• pp.357-362
• /
• 2004

Language models are essential in predicting the next word in a spoken sentence, thereby enhancing the speech recognition accuracy, among other things. However, spoken language domains are too numerous, and therefore developers suffer from the lack of corpora with sufficient sizes. This paper proposes a method of combining two n-gram language models, one constructed from a very small corpus of the right domain of interest, the other constructed from a large but less adequate corpus, resulting in a significantly enhanced language model. This method is based on the observation that a small corpus from the right domain has high quality n-grams but has serious sparseness problem, while a large corpus from a different domain has more n-gram statistics but incorrectly biased. With our approach, two n-gram statistics are combined by extending the idea of Katz's backoff and therefore is called a dual-source backoff. We ran experiments with 3-gram language models constructed from newspaper corpora of several million to tens of million words together with models from smaller broadcast news corpora. The target domain was broadcast news. We obtained significant improvement (30%) by incorporating a small corpus around one thirtieth size of the newspaper corpus.

• Journal of the Korea Society of Computer and Information
• /
• v.26 no.10
• /
• pp.157-165
• /
• 2021

Recently, studies using deep learning to analyze a large amount of text are being actively conducted. In particular, a pre-trained language model that applies the learning results of a large amount of text to the analysis of a specific domain text is attracting attention. Among various pre-trained language models, BERT(Bidirectional Encoder Representations from Transformers)-based model is the most widely used. Recently, research to improve the performance of analysis is being conducted through further pre-training using BERT's MLM(Masked Language Model). However, the traditional MLM has difficulties in clearly understands the meaning of sentences containing new words such as newly coined words. Therefore, in this study, we newly propose NTM(Newly coined words Target Masking), which performs masking only on new words. As a result of analyzing about 700,000 movie reviews of portal 'N' by applying the proposed methodology, it was confirmed that the proposed NTM showed superior performance in terms of accuracy of sensitivity analysis compared to the existing random masking.

• Journal of Intelligence and Information Systems
• /
• v.28 no.3
• /
• pp.45-55
• /
• 2022

Recently, large-scale language models (LPLM) have been shown state-of-the-art performances in various tasks of natural language processing including intent classification. However, fine-tuning LPLM requires much computational cost for training and inference which is not appropriate for dialog system. In this paper, we propose compressed intent classification model for multi-agent in low-resource like CPU. Our method consists of two stages. First, we trained sentence encoder from LPLM then compressed it through knowledge distillation. Second, we trained agent-specific adapter for intent classification. The results of three intent classification datasets show that our method achieved 98% of the accuracy of LPLM with only 21% size of it.

• MALSORI
• /
• no.51
• /
• pp.151-165
• /
• 2004

For educational and research purposes, a Korean speech recognition platform is designed. It is based on an object-oriented architecture and can be easily modified so that researchers can readily evaluate the performance of a recognition algorithm of interest. This platform will save development time for many who are interested in speech recognition. The platform includes the following modules: Noise reduction, end-point detection, met-frequency cepstral coefficient (MFCC) and perceptually linear prediction (PLP)-based feature extraction, hidden Markov model (HMM)-based acoustic modeling, n-gram language modeling, n-best search, and Korean language processing. The decoder of the platform can handle both lexical search trees for large vocabulary speech recognition and finite-state networks for small-to-medium vocabulary speech recognition. It performs word-dependent n-best search algorithm with a bigram language model in the first forward search stage and then extracts a word lattice and restores each lattice path with a trigram language model in the second stage.

• Annual Conference on Human and Language Technology
• /
• 2020.10a
• /
• pp.124-127
• /
• 2020

관계 추출은 문장에서 두 개의 엔티티가 주어졌을 때 두 개의 엔티티에 대한 의미적 이해를 통해 관계를 분류하는 작업이다. 이와 같이 관계 추출에서 관계를 분류하기 위해서는 두 개의 엔티티에 대한 정보가 필요하다. 본 연구에서는 관계 추출을 하기 위해 문장에서 엔티티들의 표현을 다르게하여 관계 추출의 성능을 비교 실험하였다. 첫번째로는 문장에서 [CLS] 토큰(Token)으로 관계를 분류하는 Standard 엔티티 정보 표현과 두번째로는 엔티티의 앞과 뒤에 Special Token을 추가하여 관계를 분류하는 Entity-Markers 엔티티 정보 표현했다. 이를 기반으로 문장의 문맥 정보를 학습한 사전 학습(Pre-trained)모델인 BERT-Large와 ALBERT-Large를 적용하여 실험을 진행하였다. 실험 결과 Special Token을 추가한 Entity-Markers의 성능이 높았으며, BERT-Large에서 더 높은 성능 결과를 확인하였다.