• Journal of the Korean Society for information Management
• /
• v.22 no.3 s.57
• /
• pp.261-287
• /
• 2005

The purpose of this paper is to explore the ways to improve the performance of SVM (Support Vector Machines) text classifier using inter-document similarities. SVMs are powerful machine learning systems, which are considered as the state-of-the-art technique for automatic document classification. In this paper text categorization via SVMs approach based on feature representation with document vectors is suggested. In this approach, document vectors instead of index terms are used as features, and vector similarities instead of term weights are used as feature values. Experiments show that SVM classifier with document vector features can improve the document classification performance. For the sake of run-time efficiency, two methods are developed: One is to select document vector features, and the other is to use category centroid vector features instead. Experiments on these two methods show that we can get improved performance with small vector feature set than the performance of conventional methods with index term features.

• Journal of the Korean Society for information Management
• /
• v.24 no.4
• /
• pp.267-283
• /
• 2007

In this study, measures of distributional similarity such as KL-divergence are applied to cluster documents instead of traditional cosine measure, which is the most prevalent vector similarity measure for document clustering. Three variations of KL-divergence are investigated; Jansen-Shannon divergence, symmetric skew divergence, and minimum skew divergence. In order to verify the contribution of distributional similarities to document clustering, two experiments are designed and carried out on three test collections. In the first experiment the clustering performances of the three divergence measures are compared to that of cosine measure. The result showed that minimum skew divergence outperformed the other divergence measures as well as cosine measure. In the second experiment second-order distributional similarities are calculated with Pearson correlation coefficient from the first-order similarity matrixes. From the result of the second experiment, secondorder distributional similarities were found to improve the overall performance of document clustering. These results suggest that minimum skew divergence must be selected as document vector similarity measure when considering both time and accuracy, and second-order similarity is a good choice for considering clustering accuracy only.

• The Journal of the Korea Contents Association
• /
• v.9 no.2
• /
• pp.142-151
• /
• 2009

Previous document clustering method, NSTC measures similarities between two document pairs using TF-IDF during web document clustering. In this paper, we propose new similarity measure using common phrase-based relational graph, not TF-IDF. This method suggests that weighting common phrases by relational graph presenting relationship among common phrases in document collection. And experimental results indicate that proposed method is more effective in clustering document collection than NSTC.

• Journal of Information Processing Systems
• /
• v.16 no.2
• /
• pp.277-300
• /
• 2020

This paper presents a method for clustering short text documents, such as news headlines, social media statuses, or instant messages. Due to the characteristics of these documents, which are usually short and sparse, an appropriate technique is required to discover hidden knowledge. The objective of this paper is to identify the combination of document representation, document distance, and document clustering that yields the best clustering quality. Document representations are expanded by external knowledge sources represented by a Distributed Representation. To cluster documents, a K-means partitioning-based clustering technique is applied, where the similarities of documents are measured by word mover's distance. To validate the effectiveness of the proposed method, experiments were conducted to compare the clustering quality against several leading methods. The proposed method produced clusters of documents that resulted in higher precision, recall, F1-score, and adjusted Rand index for both real-world and standard data sets. Furthermore, manual inspection of the clustering results was conducted to observe the efficacy of the proposed method. The topics of each document cluster are undoubtedly reflected by members in the cluster.

• Journal of KIISE
• /
• v.44 no.2
• /
• pp.179-185
• /
• 2017

Document summarization aims to generate a summary that is consistent and contains the highly related sentences in a document. In this study, we implemented for document summarization that extracts highly related sentences from a whole document by considering both similarities and entailment relations between sentences. Accordingly, we proposed a new algorithm, TextRank-NLI, which combines a Recurrent Neural Network based Natural Language Inference model and a Graph-based ranking algorithm used in single document extraction-based summarization task. In order to evaluate the performance of the new algorithm, we conducted experiments using the same datasets as used in TextRank algorithm. The results indicated that TextRank-NLI showed 2.3% improvement in performance, as compared to TextRank.

• Journal of Information Processing Systems
• /
• v.9 no.4
• /
• pp.602-620
• /
• 2013

Many previous research studies on extractive text summarization consider a subset of words in a document as keywords and use a sentence ranking function that ranks sentences based on their similarities with the list of extracted keywords. But the use of key concepts in automatic text summarization task has received less attention in literature on summarization. The proposed work uses key concepts identified from a document for creating a summary of the document. We view single-word or multi-word keyphrases of a document as the important concepts that a document elaborates on. Our work is based on the hypothesis that an extract is an elaboration of the important concepts to some permissible extent and it is controlled by the given summary length restriction. In other words, our method of text summarization chooses a subset of sentences from a document that maximizes the important concepts in the final summary. To allow diverse information in the summary, for each important concept, we select one sentence that is the best possible elaboration of the concept. Accordingly, the most important concept will contribute first to the summary, then to the second best concept, and so on. To prove the effectiveness of our proposed summarization method, we have compared it to some state-of-the art summarization systems and the results show that the proposed method outperforms the existing systems to which it is compared.

• Proceedings of the Korean Institute of Intelligent Systems Conference
• /
• 2001.05a
• /
• pp.187-190
• /
• 2001

Due to the recent proliferation of the internet, it is broadly granted that the necessity of the automatic document categorization has been on the rise. Since it is a heavy time-consuming work and takes too much manpower to process and classify manually, we need a system that categorizes them automatically as their contents. In this paper, we propose the automatic E-mail response system that is based on 2 hierarchical document clustering methods. One is to get the final result from the classifier trained seperatly within each class, after clustering the whole documents into 3 groups so that the first classifier categorize the input documents as the corresponding group. The other method is that the system classifies the most distinct classes first as their similarity, successively. Neural networks have been adopted as classifiers, we have used dendrograms to show the hierarchical aspect of similarities between classes. The comparison among the performances of hierarchical and non-hierarchical classifiers tells us clustering methods have provided the classification efficiency.

• ETRI Journal
• /
• v.38 no.3
• /
• pp.518-528
• /
• 2016

The searching of an extensive document database for documents that are locally similar to a given query document, and the subsequent detection of similar regions between such documents, is considered as an essential task in the fields of information retrieval and data management. In this paper, we present a framework for such a task. The proposed framework employs the method of short-read mapping, which is used in bioinformatics to reveal similarities between genomic sequences. In this paper, documents are considered biological objects; consequently, edit operations between locally similar documents are viewed as an evolutionary process. Accordingly, we are able to apply the method of evolution tracing in the detection of similar regions between documents. In addition, we propose heuristic methods to address issues associated with the different stages of the proposed framework, for example, a frequency-based fragment ordering method and a locality-aware interval aggregation method. Extensive experiments covering various scenarios related to the search of an extensive document database for documents that are locally similar to a given query document are considered, and the results indicate that the proposed framework outperforms existing methods.

• The KIPS Transactions:PartB
• /
• v.16B no.3
• /
• pp.233-238
• /
• 2009

This paper presents an approach to the classifications of small size document using the instance-based feature filtering Relief-F algorithm. In the document classifications, we have not always good classification performances of small size document included a few features. Because total number of feature in the document set is large, but feature count of each document is very small relatively, so the similarities between documents are very low when we use general assessment of similarity and classifiers. Specially, in the cases of the classification of web document in the directory service and the classification of the sectors that cannot connect with the original file after recovery hard-disk, we have not good classification performances. Thus, we propose the Extended Relief-F(ERelief-F) algorithm using instance-based feature filtering algorithm Relief-F to solve problems of Relief-F as preprocess of classification. For the performance comparison, we tested information gain, odds ratio and Relief-F for feature filtering and getting those feature values, and used kNN and SVM classifiers. In the experimental results, the Extended Relief-F(ERelief-F) algorithm, compared with the others, performed best for all of the datasets and reduced many irrelevant features from document sets.

• Journal of KIISE:Computing Practices and Letters
• /
• v.16 no.6
• /
• pp.663-667
• /
• 2010

A simple approach to semantic document-retrieval is to measure document similarity based on the bag-of-words representation, e.g., cosine similarity between two document vectors. However, such a syntactic method hardly considers the semantic similarity between documents, often producing semantically-unsound search results. We circumvent such a problem by combining supervised machine learning techniques with ontology information based on Markov logic. Specifically, Markov logic networks are learned from similarity-tagged documents with an ontology representing the diverse relationship among words. The learned Markov logic networks, the ontology, and the training documents are applied to the semantic document-retrieval task by inferring similarities between a query document and the training documents. Through experimental evaluation on real world question-answering data, the proposed method has been shown to outperform the simple cosine similarity-based approach in terms of retrieval accuracy.