Earticle

Deep learning techniques such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) show superior performance in text classification than traditional approaches such as Support Vector Machines (SVMs) and Naïve Bayesian approaches. When using CNNs for text classification tasks, word embedding or character embedding is a step to transform words or characters to fixed size vectors before feeding them into convolutional layers. In this paper, we propose a parallel word-level and character-level embedding approach in CNNs for text classification. The proposed approach can capture word-level and character-level patterns concurrently in CNNs. To show the usefulness of proposed approach, we perform experiments with two English and three Korean text datasets. The experimental results show that character-level embedding works better in Korean and word-level embedding performs well in English. Also the experimental results reveal that the proposed approach provides better performance than traditional CNNs with word-level embedding or character-level embedding in both Korean and English documents. From more detail investigation, we find that the proposed approach tends to perform better when there is relatively small amount of data comparing to the traditional embedding approaches.

Text classification is a challenging task, especially when dealing with a huge amount of text data. The performance of a classification model can be varied depending on what type of words contained in the document corpus and what type of features generated for classification. Aside from proposing a new modified version of the existing algorithm or creating a new algorithm, we attempt to modify the use of data. The classifier performance is usually affected by the quality of learning data as the classifier is built based on these training data. We assume that the data from different domains might have different characteristics of noise, which can be utilized in the process of learning the classifier. Therefore, we attempt to enhance the robustness of the classifier by injecting the heterogeneous data artificially into the learning process in order to improve the classification accuracy. Semi-supervised approach was applied for utilizing the heterogeneous data in the process of learning the document classifier. However, the performance of document classifier might be degraded by the unlabeled data. Therefore, we further proposed an algorithm to extract only the documents that contribute to the accuracy improvement of the classifier.

본 논문에서는 텍스트 분류 작업에서 TF-IDF(Term Frequency-Inverse Document Frequency)와 같은 희소 벡터(sparse vector)와 임베딩 모델을 통해 계산하는 밀집 벡터(dense vector)의 성능 차이를 비교하기 위해 실험을 수행한다. 과거에는 SVM, RandomForest와 같은 전통적인 머신 러닝 알고리즘에 TF-IDF 기반의 희소 벡터를 주로 사용하였으나, 최근에는 임베딩 모델의 발전으로 문장의 의미적 이해의 수준이 높아져 도메인별 문제 해결에 많이 활용되어 오고 있다. 본 논문에서는 임베딩 모델을 활 용하여 밀집 벡터 값을 계산하고, 이를 특징 값으로 하여 전통적인 머신러닝 알고리즘에 적용하여 실험을 수행하였다. 실험 데이 터 셋으로는 이메일 스팸/비스팸 데이터셋을 사용하였고, 5가지의 임베딩 모델과 4가지 분류 알고리즘으로 10 폴드 교차 검증 (10-fold cross validation)으로 검증하였다. 실험 결과에서 SVM에 희소 벡터 값을 사용했을 때 보다, SVM에 밀집 벡터 값을 사 용했을 때, F-Score 기준으로 2.6%의 향상이 있었다. 이로써 텍스트 분류에 있어 임베딩 모델을 활용한 밀집 벡터가 효과가 있 음을 알아냈다.

This paper presents an experiment comparing the performance of sparse vectors, such as those based on Term Frequency-Inverse Document Frequency (TF-IDF), with dense vectors computed through embedding models in text classification tasks. Traditionally, sparse vectors using TF-IDF have been commonly applied in conventional machine learning algorithms such as Support Vector Machines (SVM) and Random Forest. However, with the recent advancements in embedding models, the level of semantic understanding of sentences has significantly improved, and these models have been increasingly employed for solving domain-specific problems. In this study, we calculate dense vector representations using embedding models and apply them as feature values to traditional machine learning algorithms. The experiment was conducted using a spam/non-spam email dataset, and we validated the models through 10-fold cross-validation, utilizing five embedding models and four classification algorithms. The experimental results demonstrate a 2.6% improvement in the F-score when applying dense vectors with SVM, compared to using sparse vectors. This confirms the effectiveness of dense vectors generated by embedding models in text classification tasks.

Accurate classification of public policy is essential for effective policy analysis, design, comparison, and formulation across countries. However, manual classification by policy experts can lead to inconsistencies and human errors, highlighting the need for a more reliable and efficient approach. This study proposes a deep learning-based model to support policy classification using artificial intelligence. Leveraging Korean policy datasets, comprising administrative data (1988–2018), legislative data (1987–2018), and media data (1988–2020), previously curated by experts, we developed an AI model for automated policy classification based on the KoBERT language model. Designed as a supplementary tool for policy experts, this model enhances classification consistency, reduces decision-making time, and improves overall productivity. Moreover, the model enables the classification, comparison, and evaluation of diverse policies at both local and national levels, offering valuable support for strategic public policy development. The proposed model achieved a Top-1 accuracy of 62.4% and a Top-3 accuracy of 71.6%, outperforming traditional baselines and demonstrating its practical potential for real-world policy analysis.

최근, 대형 언어 모델(Large Language Model)을 대상으로 혐오적, 폭력적, 또는 공격적인 표현을 프롬프트로 입력하여 모 델의 응답을 왜곡하거나 악용하는 사례가 발생하고 있다. 이러한 공격적 사용에 대비하기 위해, 본 논문에서는 이중 필터링 기법을 제안한다. 제안된 기법은 1단계에서 혐오적 문장 여부를 분류할 수 있는 분류기를 적용하고, 2단계에서 기존 지식 내 에서 벡터 검색을 활용하여 문장의 의미를 추가적으로 평가하여 잘못 분류된 문장을 걸러내는 과정을 포함한다. 이를 통해 모 델의 응답이 공격적 콘텐츠를 포함하지 않도록, 보다 정교한 필터링을 구현하여 언어 모델의 안전성을 강화할 수 있도록 한다. 이를 위해, 1단계에서는 텍스트 임베딩 모델과 텍스트 분류기를 결합하여 혐오적 문장 분류 실험을 수행하였다. 그 결과, 0.98 이상의 f-score를 보였다. 2단계에서는 텍스트 임베딩 모델을 활용하여 질문 및 답변을 벡터 값으로 전환하여 질문에 대한 답 을 찾는 실험을 수행하였다. 그 결과, 0.87의 Precision@1을 보였다.

Recently, there have been instances where prompts containing hateful, violent, or offensive expressions are input into large language models (LLMs) to distort or misuse their responses. To counteract such abusive usage, this paper proposes a dual filtering technique. The proposed method applies a classifier in the first stage to detect hateful sentences, followed by a second stage where vector search is employed to further assess the semantics of the sentence within the model's existing knowledge base, filtering out any misclassified sentences. This approach aims to strengthen the safety of language models by implementing more sophisticated filtering that prevents the inclusion of offensive content in model responses. In the first stage, we combined a text embedding model with a text classifier to conduct an experiment on classifying hateful sentences, achieving an f-score of over 0.98. In the second stage, we conducted experiments using a text embedding model to convert questions and answers into vector values, with a focus on finding accurate responses to queries. The results showed a Precision@1 of 0.87.

본 논문은 지식·서비스 산업을 중심으로, 직원 만족도와 기업의 재무 성과 간 상호작용을 정량적으로 분석하고, 이를 바탕으로 조직의 리스크를 진단하며 맞춤형 경영 전략을 수립하기 위한 XAI(설명 가능한 인공지능) 기법의 활용 가능성을 탐구한다. 코스피 및 코스닥 상장 기업의 재무정보와 온라인 기업 리뷰 플랫폼 ‘잡플래닛’의 리뷰 데이터를 결합하여, 재무 성과와 직원 만족도를 각각 예측하는 모형을 구축하였다. 재무 성과 예측 모형은 전통적인 머신러닝 기법을 활용하였으며, 직원 만족도 모형은 딥러닝 기반 텍스트 분류 모델(BERT 미세 조정)을 적용하여 개발되었다. 예측 결과를 기반으로, 각 기업은 재무 성과와 직원 만족도에 따라 성장형, 성과 중시형, 만족 중심형, 개선형의 네 가지 유형으로 분류되었으며, SHAP와 LIME 기법을 활용하여 각 모형의 예측에 기여한 주요 요인을 정량화하고 해석하였다. 특히, 재무 성과 모형은 SHAP를 통해 간략히 시각화하였고, 직원 만족도 모형은 LIME을 중심으로 텍스트 리뷰의 긍·부정 요인을 심층적으로 분석하여, 조직별 강점과 약점을 도출하였다. 본 연구의 결과는 경영진 및 HR 담당자가 조직의 내·외부 리스크를 명확히 인지하고, 맞춤형 전략 수립을 위한 의사결정 지원에 활용될 수 있음을 보여준다.

Now is the time for IS scholars to demonstrate the added value of academic theory through its integration with text mining, clearly outline how to implement this for text mining experts outside of the academic field, and move towards establishing this integration as a standard practice. Therefore, in this study we develop a systematic theory-based text-mining framework (TTMF), and illustrate the use and benefits of TTMF by conducting a text-mining project in an actual business case evaluating and improving hotel service quality using a large volume of actual user-generated reviews. A total of 61,304 sentences extracted from actual customer reviews were successfully allocated to SERVQUAL dimensions, and the pragmatic validity of our model was tested by the OLS regression analysis results between the sentiment scores of each SERVQUAL dimension and customer satisfaction (star rates), and showed significant relationships. As a post-hoc analysis, the results of the co-occurrence analysis to define the root causes of positive and negative service quality perceptions and provide action plans to implement improvements were reported.

한국 문화의 역사, 전통과 관련된 디지털 웹 문서가 증가하게 되었다. 하지만 창작자 또는 전통 문화와 관련된 소재를 찾는 사용자들은 정보를 검색해도 결과가 충분하지 않았으며 원하는 정보를 얻지 못하는 경우가 나타나고 있다. 이런 효과적인 정보를 접하기 위해서는 문서 분류가 필요하다. 과거에 문서 분류는 작업자가 수작업으로 문서 분류하여 시간과 비용이 많이 소비하는 어려움이 있었지만, 최근 기계학습 기반으로 한 자동 문서 분류를 통해 효율적인 문서 분류가 이루어진다. 이에 본 논문은 전통문화 콘텐츠를 체계적인 분류체계로 구성한 한민족정보문화마당 데이터를 기반으로 전통문화 콘텐츠 자동 텍스트 분류 모델을 개발한다. 본 연구는 한민족정보문화마당 텍스트 데이터에 대해 단어 빈도수를 추출하기 위해 TF-IDF모델, Bag-of-Words 모델, TF-IDF/Bag-of-Words를 결합한 모델을 적용하여 각각 SVM 분류 알고리즘을 사용하여 전통문화 콘텐츠 자동 텍스트 분류 모델을 개발하여 성능평가를 확인하였다.

The Internet have increased the number of digital web documents related to the history and traditions of Korean Culture. However, users who search for creators or materials related to traditional cultures are not able to get the information they want and the results are not enough. Document classification is required to access this effective information. In the past, document classification has been difficult to manually and manually classify documents, but it has recently been difficult to spend a lot of time and money. Therefore, this paper develops an automatic text classification model of traditional cultural contents based on the data of the Korean information culture field composed of systematic classifications of traditional cultural contents. This study applied TF-IDF model, Bag-of-Words model, and TF-IDF/Bag-of-Words combined model to extract word frequencies for 'Korea Traditional Culture' data. And we developed the automatic text classification model of traditional cultural contents using Support Vector Machine classification algorithm.

Text categorization systems generally use single words (unigrams) as features. A deceptively simple algorithm for improving text categorization is investigated here, an idea previously shown not to work. It is to identify useful word pairs (bigrams) made up of adjacent unigrams. The bigrams it found, while small in numbers, can substantially raise the quality of feature sets. The algorithm was tested on two pre-classified datasets, Reuters-21578 for English and Korea-web for Korean. The results show that the algorithm was successful in extracting high quality bigrams and increased the quality of overall features. To find out the role of bigrams, we trained the Naive Bayes classifiers using both unigrams and bigrams as features. The results show that recall values were higher than those of unigrams alone. Break-even points and F1 values improved in most documents, especially when documents were classified along the large classes. In Reuters-21578 break-even points increased by 2.1%, with the highest at 18.8%, and F1 improved by 1.5%, with the highest at 3.2%. In Korea-web break-even points increased by 1.0%, with the highest at 4.5%, and F1 improved by 0.4%, with the highest at 4.2%. We can conclude that text classification using unigrams and bigrams together is more efficient than using only unigrams.

In recent years, sentiment analysis research has become popular. The research results of sentiment analysis have achieved remarkable results in practical applications, such as in Amazon's book recommendation system and the North American movie box office evaluation system. Analyzing big data based on user preferences and evaluations and recommending hot-selling books and hot-rated movies to users in a targeted manner greatly improve book sales and attendance rate in movies [1, 2]. However, traditional machine learning-based sentiment analysis methods such as the Classification and Regression Tree (CART), Support Vector Machine (SVM), and k-nearest neighbor classification (kNN) had performed poorly in accuracy. In this paper, an improved kNN classification method is proposed. Through the improved method and normalizing of data, the purpose of improving accuracy is achieved. Subsequently, the three classification algorithms and the improved algorithm were compared based on experimental data. Experiments show that the improved method performs best in the kNN classification method, with an accuracy rate of 11.5% and a precision rate of 20.3%.

To improve the performance of text classification, we propose text augmentation based on attention score (TABAS). We recognized that a criterion for selecting a replacement word rather than a random selection was necessary. Therefore, TABAS utilizes attention scores for text modification, processing only words with the same entity and part-of-speech tags to consider informational aspects. To verify this approach, we used two benchmark tasks. As a result, TABAS can significantly improve performance, both recurrent and convolutional neural networks. Furthermore, we confirm that it provides a practical way to develop deep-learning models by saving costs on making additional datasets.

The natural language on social network platforms has a certain front-to-back dependency in structure, and the direct conversion of Chinese text into a vector makes the dimensionality very high, thereby resulting in the low accuracy of existing text classification methods. To this end, this study establishes a deep learning model that combines a big data ultra-deep convolutional neural network (UDCNN) and long short-term memory network (LSTM). The deep structure of UDCNN is used to extract the features of text vector classification. The LSTM stores historical information to extract the context dependency of long texts, and word embedding is introduced to convert the text into low-dimensional vectors. Experiments are conducted on the social network platforms Sogou corpus and the University HowNet Chinese corpus. The research results show that compared with CNN + rand, LSTM, and other models, the neural network deep learning hybrid model can effectively improve the accuracy of text classification.

Patent classification is becoming more critical as patent filings have been increasing over the years. Despite comprehensive studies in the area, there remain several issues in classifying patents on IPC hierarchical levels. Not only structural complexity but also shortage of patents in the lower level of the hierarchy causes the decline in classification performance. Therefore, we propose a new method of classification based on different criteria that are categories defined by the domain's experts mentioned in trend analysis reports, i.e. Patent Landscape Report (PLR). Several experiments were conducted with the purpose of identifying type of features and weighting methods that lead to the best classification performance using Support Vector Machine (SVM). Two types of features (noun and noun phrases) and five different weighting schemes (TF-idf, TF-rf, TF-icf, TF-icf-based, and TF-idcef-based) were experimented on.

In the process of Chinese long-text classification, due to the large amount of text data and complex features, methods suitable for ordinary text classification often lack sufficient accuracy, which directly leads to frequent classification failures in long-text environments. To solve this problem, the research designed a bi-directional long short-term memory (Bi-LSTM) model that combines forward and backward operations and utilized attention mechanisms to improve fusion. At the same time, the bi-directional encoder representations from transformers (BERT) model was introduced into the text processing to form a long-text classification model. Finally, different datasets were tested to verify the actual classification effect of the model. The research results showed that under different dataset environments, the classification accuracy rates of the designed models were 92.93% and 93.77%, respectively, which are the models with the highest classification accuracy rates among the same type of models. The calculation time was 85.42 seconds and 117.51 seconds, respectively, which are the models with the shortest calculation time among the same type of models. It can be seen that the research designed long-text classification model innovatively combined the BERT model, convolutional neural network model, Bi-LSTM model, and attention mechanism structure based on the data characteristics of long-text classification, enabling the model to achieve higher classification accuracy in a shorter computational time. Moreover, it has better classification results in actual long-text classification, overcomes the classification failure problem caused by complex text features in the long-text classification environment, and provides a possibility for long-text specific classification paths.

Nowadays, Islamic content is widely used in research, including Hadith and the Al-Quran. Both are mostly used in the field of natural language processing, especially in text classification research. One of the difficulties in learning the Al-Quran is ambiguity, while the Al-Quran is used as the main source of Islamic law and the life guidance of a Muslim in the world. This research was proposed to relieve people in learning the Al-Quran. We proposed a word embedding feature-based on Tensor Space Model as feature extraction, which is used to reduce the ambiguity. Based on the experiment results and the analysis, we prove that the proposed method yields the best performance with the Hamming loss 0.10317.

Text data has always accounted for a major portion of the world's information. As the volume of information increases exponentially, the portion of text data also increases significantly. Text classification is therefore still an important area of research. LDA is an updated, probabilistic model which has been used in many applications in many other fields. As regards text data, LDA also has many applications, which has been applied various enhancements. However, it seems that no applications take care of the input for LDA. In this paper, we suggest a way to map the input space to a reduced space, which may avoid the unreliability, ambiguity and redundancy of individual terms as descriptors. The purpose of this paper is to show that LDA can be perfectly performed in a "clean and clear" space. Experiments are conducted on 20 News Groups data sets. The results show that the proposed method can boost the classification results when the appropriate choice of rank of the reduced space is determined.

As the demand for cross-national policy comparisons intensifies, automating the classification of vast volumes of policy texts through deep learning presents a viable solution to alleviate manual labor and enhance policy matching efficiency. This study proposes an automated multi-topic policy classification system grounded in the policy perceptron framework, augmented by the RoBERTa-whole word masking model pretrained on domain-specific corpora. The study curated a dataset of policy documents, applied the model for multi-label classification, and rigorously evaluated its performance via comparative experiments against baseline models. The results present performance for our model in multi-topic scenarios, with robustness to imbalanced classes and lengthy texts. Furthermore, the authors benchmarked the approach against general-purpose large language models, revealing that this domain-tailored model achieves higher classification efficacy. These findings highlight the advantages of specialized deep learning architectures for policy text analysis, examining the conceptual and methodological implications of automating policy classification and offering practical insights for scalable policy organization and future research in automated governance tools.

The narrow-down approach, separately composed of search and classification stages, is an effective way of dealing with large-scale hierarchical text classification. Recent approaches introduce methods of incorporating global, local, and path information extracted from web taxonomies in the classification stage. Meanwhile, in the case of utilizing path information, there have been few efforts to address existing limitations and develop more sophisticated methods. In this paper, we propose an expansion method to effectively exploit category path information based on the observation that the existing method is exposed to a term mismatch problem and low discrimination power due to insufficient path information. The key idea of our method is to utilize relevant information not presented on category paths by adding more useful words. We evaluate the effectiveness of our method on state-of-the art narrow-down methods and report the results with in-depth analysis.

To address the problems of low precision rate, insufficient feature extraction, and poor contextual ability in existing text sentiment analysis methods, a mixed model account of a CNN-BiLSTM-TE (convolutional neural network, bidirectional long short-term memory, and topic extraction) model was proposed. First, Chinese text data was converted into vectors through the method of transfer learning by Word2Vec. Second, local features were extracted by the CNN model. Then, contextual information was extracted by the BiLSTM neural network and the emotional tendency was obtained using softmax. Finally, topics were extracted by the term frequency-inverse document frequency and K-means. Compared with the CNN, BiLSTM, and gate recurrent unit (GRU) models, the CNN-BiLSTM-TE model's F1-score was higher than other models by 0.0147, 0.006, and 0.0052, respectively. Then compared with CNN-LSTM, LSTM-CNN, and BiLSTM-CNN models, the F1-score was higher by 0.0071, 0.0038, and 0.0049, respectively. Experimental results showed that the CNN-BiLSTM-TE model can effectively improve various indicators in application. Lastly, performed scalability verification through a takeaway dataset, which has great value in practical applications.

With the rapid development of Internet of Things (IoT) and big data technology, a large amount of data will be generated during the operation of related industries. How to classify the generated data accurately has become the core of research on data mining and processing in IoT industry chain. This study constructs a classification model of IoT industry chain based on improved random forest algorithm and text analysis, aiming to achieve efficient and accurate classification of IoT industry chain big data by improving traditional algorithms. The accuracy, precision, recall, and AUC value size of the traditional Random Forest algorithm and the algorithm used in the paper are compared on different datasets. The experimental results show that the algorithm model used in this paper has better performance on different datasets, and the accuracy and recall performance on four datasets are better than the traditional algorithm, and the accuracy performance on two datasets, P-I Diabetes and Loan Default, is better than the random forest model, and its final data classification results are better. Through the construction of this model, we can accurately classify the massive data generated in the IoT industry chain, thus providing more research value for the data mining and processing technology of the IoT industry chain.