• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.17 no.10
• /
• pp.2768-2787
• /
• 2023

As an important research direction of the application of computer science in the medical field, the automatic generation technology of radiology report has attracted wide attention in the academic community. Because the proportion of normal regions in radiology images is much larger than that of abnormal regions, words describing diseases are often masked by other words, resulting in significant feature loss during the calculation process, which affects the quality of generated reports. In addition, the huge difference between visual features and semantic features causes traditional multi-modal fusion method to fail to generate long narrative structures consisting of multiple sentences, which are required for medical reports. To address these challenges, we propose a multi-feature optimization Transformer (MFOT) for generating radiology reports. In detail, a multi-dimensional mapping attention (MDMA) module is designed to encode the visual grid features from different dimensions to reduce the loss of primary features in the encoding process; a feature pre-fusion (FP) module is constructed to enhance the interaction ability between multi-modal features, so as to generate a reasonably structured radiology report; a detail enhanced attention (DEA) module is proposed to enhance the extraction and utilization of key features and reduce the loss of key features. In conclusion, we evaluate the performance of our proposed model against prevailing mainstream models by utilizing widely-recognized radiology report datasets, namely IU X-Ray and MIMIC-CXR. The experimental outcomes demonstrate that our model achieves SOTA performance on both datasets, compared with the base model, the average improvement of six key indicators is 19.9% and 18.0% respectively. These findings substantiate the efficacy of our model in the domain of automated radiology report generation.

• Smart Media Journal
• /
• v.12 no.9
• /
• pp.81-88
• /
• 2023

Recently, communication through online is increasing due to the spread of non-face-to-face services due to COVID-19. In non-face-to-face situations, the other person's opinions and emotions are recognized through modalities such as text, speech, and images. Currently, research on multimodal emotion recognition that combines various modalities is actively underway. Among them, emotion recognition using speech data is attracting attention as a means of understanding emotions through sound and language information, but most of the time, emotions are recognized using a single speech feature value. However, because a variety of emotions exist in a complex manner in a conversation, a method for recognizing multiple emotions is needed. Therefore, in this paper, we propose a multi-emotion regression model that extracts feature vectors after preprocessing speech data to recognize complex, inherent emotions and takes into account the passage of time.

• Journal of Intelligence and Information Systems
• /
• v.24 no.1
• /
• pp.183-203
• /
• 2018

News articles are the most suitable medium for examining the events occurring at home and abroad. Especially, as the development of information and communication technology has brought various kinds of online news media, the news about the events occurring in society has increased greatly. So automatically summarizing key events from massive amounts of news data will help users to look at many of the events at a glance. In addition, if we build and provide an event network based on the relevance of events, it will be able to greatly help the reader in understanding the current events. In this study, we propose a method for extracting event networks from large news text data. To this end, we first collected Korean political and social articles from March 2016 to March 2017, and integrated the synonyms by leaving only meaningful words through preprocessing using NPMI and Word2Vec. Latent Dirichlet allocation (LDA) topic modeling was used to calculate the subject distribution by date and to find the peak of the subject distribution and to detect the event. A total of 32 topics were extracted from the topic modeling, and the point of occurrence of the event was deduced by looking at the point at which each subject distribution surged. As a result, a total of 85 events were detected, but the final 16 events were filtered and presented using the Gaussian smoothing technique. We also calculated the relevance score between events detected to construct the event network. Using the cosine coefficient between the co-occurred events, we calculated the relevance between the events and connected the events to construct the event network. Finally, we set up the event network by setting each event to each vertex and the relevance score between events to the vertices connecting the vertices. The event network constructed in our methods helped us to sort out major events in the political and social fields in Korea that occurred in the last one year in chronological order and at the same time identify which events are related to certain events. Our approach differs from existing event detection methods in that LDA topic modeling makes it possible to easily analyze large amounts of data and to identify the relevance of events that were difficult to detect in existing event detection. We applied various text mining techniques and Word2vec technique in the text preprocessing to improve the accuracy of the extraction of proper nouns and synthetic nouns, which have been difficult in analyzing existing Korean texts, can be found. In this study, the detection and network configuration techniques of the event have the following advantages in practical application. First, LDA topic modeling, which is unsupervised learning, can easily analyze subject and topic words and distribution from huge amount of data. Also, by using the date information of the collected news articles, it is possible to express the distribution by topic in a time series. Second, we can find out the connection of events in the form of present and summarized form by calculating relevance score and constructing event network by using simultaneous occurrence of topics that are difficult to grasp in existing event detection. It can be seen from the fact that the inter-event relevance-based event network proposed in this study was actually constructed in order of occurrence time. It is also possible to identify what happened as a starting point for a series of events through the event network. The limitation of this study is that the characteristics of LDA topic modeling have different results according to the initial parameters and the number of subjects, and the subject and event name of the analysis result should be given by the subjective judgment of the researcher. Also, since each topic is assumed to be exclusive and independent, it does not take into account the relevance between themes. Subsequent studies need to calculate the relevance between events that are not covered in this study or those that belong to the same subject.

• Journal of Intelligence and Information Systems
• /
• v.27 no.3
• /
• pp.175-197
• /
• 2021

Recently, with the development of deep learning technology, research on unstructured data analysis is being actively conducted, and it is showing remarkable results in various fields such as classification, summary, and generation. Among various text analysis fields, text classification is the most widely used technology in academia and industry. Text classification includes binary class classification with one label among two classes, multi-class classification with one label among several classes, and multi-label classification with multiple labels among several classes. In particular, multi-label classification requires a different training method from binary class classification and multi-class classification because of the characteristic of having multiple labels. In addition, since the number of labels to be predicted increases as the number of labels and classes increases, there is a limitation in that performance improvement is difficult due to an increase in prediction difficulty. To overcome these limitations, (i) compressing the initially given high-dimensional label space into a low-dimensional latent label space, (ii) after performing training to predict the compressed label, (iii) restoring the predicted label to the high-dimensional original label space, research on label embedding is being actively conducted. Typical label embedding techniques include Principal Label Space Transformation (PLST), Multi-Label Classification via Boolean Matrix Decomposition (MLC-BMaD), and Bayesian Multi-Label Compressed Sensing (BML-CS). However, since these techniques consider only the linear relationship between labels or compress the labels by random transformation, it is difficult to understand the non-linear relationship between labels, so there is a limitation in that it is not possible to create a latent label space sufficiently containing the information of the original label. Recently, there have been increasing attempts to improve performance by applying deep learning technology to label embedding. Label embedding using an autoencoder, a deep learning model that is effective for data compression and restoration, is representative. However, the traditional autoencoder-based label embedding has a limitation in that a large amount of information loss occurs when compressing a high-dimensional label space having a myriad of classes into a low-dimensional latent label space. This can be found in the gradient loss problem that occurs in the backpropagation process of learning. To solve this problem, skip connection was devised, and by adding the input of the layer to the output to prevent gradient loss during backpropagation, efficient learning is possible even when the layer is deep. Skip connection is mainly used for image feature extraction in convolutional neural networks, but studies using skip connection in autoencoder or label embedding process are still lacking. Therefore, in this study, we propose an autoencoder-based label embedding methodology in which skip connections are added to each of the encoder and decoder to form a low-dimensional latent label space that reflects the information of the high-dimensional label space well. In addition, the proposed methodology was applied to actual paper keywords to derive the high-dimensional keyword label space and the low-dimensional latent label space. Using this, we conducted an experiment to predict the compressed keyword vector existing in the latent label space from the paper abstract and to evaluate the multi-label classification by restoring the predicted keyword vector back to the original label space. As a result, the accuracy, precision, recall, and F1 score used as performance indicators showed far superior performance in multi-label classification based on the proposed methodology compared to traditional multi-label classification methods. This can be seen that the low-dimensional latent label space derived through the proposed methodology well reflected the information of the high-dimensional label space, which ultimately led to the improvement of the performance of the multi-label classification itself. In addition, the utility of the proposed methodology was identified by comparing the performance of the proposed methodology according to the domain characteristics and the number of dimensions of the latent label space.

• Journal of Intelligence and Information Systems
• /
• v.19 no.3
• /
• pp.1-23
• /
• 2013

To discover significant social issues such as unemployment, economy crisis, social welfare etc. that are urgent issues to be solved in a modern society, in the existing approach, researchers usually collect opinions from professional experts and scholars through either online or offline surveys. However, such a method does not seem to be effective from time to time. As usual, due to the problem of expense, a large number of survey replies are seldom gathered. In some cases, it is also hard to find out professional persons dealing with specific social issues. Thus, the sample set is often small and may have some bias. Furthermore, regarding a social issue, several experts may make totally different conclusions because each expert has his subjective point of view and different background. In this case, it is considerably hard to figure out what current social issues are and which social issues are really important. To surmount the shortcomings of the current approach, in this paper, we develop a prototype system that semi-automatically detects social issue keywords representing social issues and problems from about 1.3 million news articles issued by about 10 major domestic presses in Korea from June 2009 until July 2012. Our proposed system consists of (1) collecting and extracting texts from the collected news articles, (2) identifying only news articles related to social issues, (3) analyzing the lexical items of Korean sentences, (4) finding a set of topics regarding social keywords over time based on probabilistic topic modeling, (5) matching relevant paragraphs to a given topic, and (6) visualizing social keywords for easy understanding. In particular, we propose a novel matching algorithm relying on generative models. The goal of our proposed matching algorithm is to best match paragraphs to each topic. Technically, using a topic model such as Latent Dirichlet Allocation (LDA), we can obtain a set of topics, each of which has relevant terms and their probability values. In our problem, given a set of text documents (e.g., news articles), LDA shows a set of topic clusters, and then each topic cluster is labeled by human annotators, where each topic label stands for a social keyword. For example, suppose there is a topic (e.g., Topic1 = {(unemployment, 0.4), (layoff, 0.3), (business, 0.3)}) and then a human annotator labels "Unemployment Problem" on Topic1. In this example, it is non-trivial to understand what happened to the unemployment problem in our society. In other words, taking a look at only social keywords, we have no idea of the detailed events occurring in our society. To tackle this matter, we develop the matching algorithm that computes the probability value of a paragraph given a topic, relying on (i) topic terms and (ii) their probability values. For instance, given a set of text documents, we segment each text document to paragraphs. In the meantime, using LDA, we can extract a set of topics from the text documents. Based on our matching process, each paragraph is assigned to a topic, indicating that the paragraph best matches the topic. Finally, each topic has several best matched paragraphs. Furthermore, assuming there are a topic (e.g., Unemployment Problem) and the best matched paragraph (e.g., Up to 300 workers lost their jobs in XXX company at Seoul). In this case, we can grasp the detailed information of the social keyword such as "300 workers", "unemployment", "XXX company", and "Seoul". In addition, our system visualizes social keywords over time. Therefore, through our matching process and keyword visualization, most researchers will be able to detect social issues easily and quickly. Through this prototype system, we have detected various social issues appearing in our society and also showed effectiveness of our proposed methods according to our experimental results. Note that you can also use our proof-of-concept system in http://dslab.snu.ac.kr/demo.html.

• Journal of KIISE:Software and Applications
• /
• v.34 no.12
• /
• pp.1045-1055
• /
• 2007

Latent topic models are statistical models which automatically captures salient patterns or correlation among features underlying a data collection in a probabilistic way. They are gaining an increased popularity as an effective tool in the application of automatic semantic feature extraction from text corpus, multimedia data analysis including image data, and bioinformatics. Among the important issues for the effectiveness in the application of latent topic models to the massive data set is the efficient learning of the model. The paper proposes an accelerated learning technique for PLSA model, one of the popular latent topic models, by an incremental EM algorithm instead of conventional EM algorithm. The incremental EM algorithm can be characterized by the employment of a series of partial E-steps that are performed on the corresponding subsets of the entire data collection, unlike in the conventional EM algorithm where one batch E-step is done for the whole data set. By the replacement of a single batch E-M step with a series of partial E-steps and M-steps, the inference result for the previous data subset can be directly reflected to the next inference process, which can enhance the learning speed for the entire data set. The algorithm is advantageous also in that it is guaranteed to converge to a local maximum solution and can be easily implemented just with slight modification of the existing algorithm based on the conventional EM. We present the basic application of the incremental EM algorithm to the learning of PLSA and empirically evaluate the acceleration performance with several possible data partitioning methods for the practical application. The experimental results on a real-world news data set show that the proposed approach can accomplish a meaningful enhancement of the convergence rate in the learning of latent topic model. Additionally, we present an interesting result which supports a possible synergistic effect of the combination of incremental EM algorithm with parallel computing.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.3
• /
• pp.24-33
• /
• 2001

this paper describes a baseline for an implementation of a corpus-based Korean TTS system. The conventional TTS systems using small-sized speech still generate machine-like synthetic speech. To overcome this problem we introduce the corpus-based TTS system which enables to generate natural synthetic speech without prosodic modifications. The corpus should be composed of a natural prosody of source speech and multiple instances of synthesis units. To make a phone level synthesis unit, we train a speech recognizer with the target speech, and then perform an automatic phoneme segmentation. We also detect the fine pitch period using Laryngo graph signals, which is used for prosodic feature extraction. For break strength allocation, 4 levels of break indices are decided as pause length and also attached to phones to reflect prosodic variations in phrase boundaries. To predict the break strength on texts, we utilize the statistical information of POS (Part-of-Speech) sequences. The best triphone sequences are selected by Viterbi search considering the minimization of accumulative Euclidean distance of concatenating distortion. To get high quality synthesis speech applicable to commercial purpose, we introduce a domain specific database. By adding domain specific database to general domain database, we can greatly improve the quality of synthetic speech on specific domain. From the subjective evaluation, the new Korean corpus-based TTS system shows better naturalness than the conventional demisyllable-based one.

• The Journal of Society for e-Business Studies
• /
• v.23 no.2
• /
• pp.83-96
• /
• 2018

Extracting keywords representing documents is very important because it can be used for automated services such as document search, classification, recommendation system as well as quickly transmitting document information. However, when extracting keywords based on the frequency of words appearing in a web site documents and graph algorithms based on the co-occurrence of words, the problem of containing various words that are not related to the topic potentially in the web page structure, There is a difficulty in extracting the semantic keyword due to the limit of the performance of the Korean tokenizer. In this paper, we propose a method to select candidate keywords based on semantic similarity, and solve the problem that semantic keyword can not be extracted and the accuracy of Korean tokenizer analysis is poor. Finally, we use the technique of extracting final semantic keywords through filtering process to remove inconsistent keywords. Experimental results through real web pages of small business show that the performance of the proposed method is improved by 34.52% over the statistical similarity based keyword selection technique. Therefore, it is confirmed that the performance of extracting keywords from documents is improved by considering semantic similarity between words and removing inconsistent keywords.

• The KIPS Transactions:PartD
• /
• v.9D no.6
• /
• pp.1145-1156
• /
• 2002

We propose a web document transcoding technique that translates existing web pages designed for desktop computers into an appropriate form for hand-held devices connected to the wireless internet. By defining a content block based on a visual separation and using it as a minimum unit for analyzing and converting processes, we can get web pages converted more exactly. We also apply the reallocation of the content block and the generation of new index in order to provide convenient interface without left-right scrolling in small screen devices. These methods, compared with existing ways such as text level summary or partial extraction method, can provide efficient navigation and a full recognition of web documents. To gain those transcoding benefits, we propose the Layout-Forming Tag Analysis Algorithm that analyzes structural tags, which motivate visual separation and the Component Grouping Algorithm that extracts the content block. We also classify and rearrange the content block and generate the new index to produce an appropriate form of web pages for small display devices. We have designed and implemented our transcoding system in a proxy server and evaluated the methods and the algorithms through an analysis of transcoded results. Our transcoding system showed a good result on most of popular web pages that have complicated structures.

• The KIPS Transactions:PartB
• /
• v.14B no.5
• /
• pp.383-390
• /
• 2007

Interactions between proteins and genes are often considered essential in the description of biomolecular phenomena and networks of interactions are considered as an entre for a Systems Biology approach. Recently, many works try to extract information by analyzing biomolecular text using natural language processing technology. Previous researches insist that linguistic information is useful to improve the performance in detecting gene interactions. However, previous systems do not show reasonable performance because of low recall. To improve recall without sacrificing precision, this paper proposes a new method for detection of gene interactions based on syntactic relations. Without biomolecular knowledge, our method shows reasonable performance using only small size of training data. Using the format of LLL05(ICML05 Workshop on Learning Language in Logic) data we detect the agent gene and its target gene that interact with each other. In the 1st phase, we detect encapsulation types for each agent and target candidate. In the 2nd phase, we construct verb lists that indicate the interaction information between two genes. In the last phase, to detect which of two genes is an agent or a target, we learn direction information. In the experimental results using LLL05 data, our proposed method showed F-measure of 88% for training data, and 70.4% for test data. This performance significantly outperformed previous methods. We also describe the contribution rate of each phase to the performance, and demonstrate that the first phase contributes to the improvement of recall and the second and last phases contribute to the improvement of precision.