Search

Outlier analysis has been widely studied and has produced many methods. However, there is still rare a method to detect outliers for dynamically streaming batch data (online learning). In the present research, a novel online algorithm to detect outliers in such dataset is proposed. Data points are proceeded by applying a modified recursive PCA to predict sequentially parameters of the model; eigenvalues and eigenvectors of the statistical detection model are recursively updated using approximate values by perturbation methods. More specifically, the recursive eigenstructure is obtained from the derivation of the covariance matrix using the first-order perturbation technique. The Mahalanobis distance is then used as an outlier score. Our algorithm performances are evaluated using some metrics, namely accuration, precision, recall, F1-score, AUC-PR, and the execution time. Results show that the proposed online outlier detection is computationally efficient in time and the algorithm's performance effectiveness is comparable to that of the offline outlier detection algorithm via classical PCA.

Studies in low-resource languages have become more challenging with the increasing volume of texts in today's digital era. Also, the lack of labeled data and text processing libraries in a language further widens the research gap between high and low-resource languages, such as English and Indonesian. This has led to the use of a transfer learning approach, which applies pre-trained models to solve similar problems, even in different languages by using bilingual or cross-lingual word embedding. Therefore, this study aims to investigate two bilingual word embedding methods, namely VecMap and BiVec, for Indonesian – English language and evaluates them for bilingual lexicon induction and text summarization tasks. The generated bilingual embedding was compared with MUSE (Multilingual Unsupervised and Supervised Embeddings) as the existing multilingual word created with the generative adversarial network method. Furthermore, the VecMap was improved by creating shared vocabulary spaces and mapping the unshared ones between languages. The result showed the embedding produced by the joint methods of BiVec, performed better in intrinsic evaluation, especially with CSLS (Cross-Domain Similarity Local Scaling) retrieval. Meanwhile, the improved VecMap outperformed the regular type by 16.6% without surpassing the BiVec evaluation score. These methods enabled model transfer between languages when applied to cross-lingual-based text summarization. Moreover, the ROUGE score outperformed classical text summarization by adding only 10% of the training dataset of the target language.

Dynamic heterogeneous graphs can represent real-world networks. Predicting links in these graphs is more complicated than in static graphs. Until now, research interest of link prediction has focused on static heterogeneous graphs or dynamically homogeneous graphs. A link prediction technique combining temporal RNN and hierarchical attention has recently emerged, called DyHATR. This method is claimed to be able to work on dynamic heterogeneous graphs by testing them on four publicly available data sets (Twitter, Math-Overflow, Ecomm, and Alibaba). However, after further analysis, it turned out that the four data sets did not meet the criteria of dynamic heterogeneous graphs. In the present work, we evaluated the performance of DyHATR on dynamic heterogeneous graphs. We conducted experiments with DyHATR based on the Yelp data set represented as a dynamic heterogeneous graph consisting of homogeneous subgraphs. The results show that DyHATR can be applied to identify link prediction on dynamic heterogeneous graphs by simultaneously capturing heterogeneous information and evolutionary patterns, and then considering them to carry out link predicition. Compared to the baseline method, the accuracy achieved by DyHATR is competitive, although the results can still be improved.

Aspect detection systems for online reviews, especially based on unsupervised models, are considered better strategically to process online reviews, generally a very large collection of unstructured data. Aspect embedding-based deep learning models are designed for this problem however they still rely on redundant word embedding and they are sensitive to initialization which may have a significant impact on model performance. In this research, a pruning approach is used to reduce the redundancy of deep learning model connections and is expected to produce a model with similar or better performance. This research includes several experiments and comparisons of the results of pruning the model network weights based on the general neural network pruning strategy and the lottery ticket hypothesis. The result of this research is that pruning of the unsupervised aspect detection model, in general, can produce smaller submodels with similar performance even with a significant amount of weights pruned. Our sparse model with 80% of its total weight pruned has a similar performance to the original model. Our current pruning implementation, however, has not been able to produce sparse models with better performance.

Many fruit recognition works have applied statistical approaches to make an exact correlation between low-level visual feature information and high-level semantic concepts given by predefined text caption or keywords. Two common fruit recognition models include bag-of-features (BoF) and convolutional neural network (ConvNet), which achieve high-performance results. In most cases, the overfitting problem is unavoidable. This problem makes it difficult to generalize new instances with only a slightly different appearance, although belonging to the same category. This article proposes a new fruit recognition model by associating an object's low-level features in an image with a high-level concept. We define a perceptual color for each fruit species to construct a relationship between fruit color and semantic color name. Furthermore, we develop our model by integrating the perceptual color and semantic template concept to solve the overfitting problem. The semantic template concept as a mapping between the high-level concept and the low-level visual feature is adopted in this model. The experiment was conducted on three different fruit image datasets, with one dataset as train data and the two others as test data. The experimental results demonstrate that the proposed model, called perceptual color on semantic template (PCoST), is significantly better than the BoF and ConvNet models in reducing the overfitting problem.

In aspect-based sentiment analysis, tasks are diverse and consist of aspect term extraction, aspect categorization, opinion term extraction, sentiment polarity classification, and relation extractions of aspect and opinion terms. These tasks are generally carried out sequentially using more than one model. However, this approach is inefficient and likely to reduce the model’s performance due to cumulative errors in previous processes. The co-extraction approach with Dual crOss-sharEd RNN (DOER) and span-based multitask acquired better performance than the pipelined approaches in English review data. Therefore, this research focuses on adapting the co-extraction approach where the extraction of aspect terms, opinion terms, and sentiment polarity are conducted simultaneously from review texts. The co-extraction approach was adapted by modifying the original frameworks to perform unhandled subtask to get the opinion triplet. Furthermore, the output layer on these frameworks was modified and trained using a collection of Indonesian-language hotel reviews. The adaptation was conducted by testing the output layer topology for aspect and opinion term extraction as well as variations in the type of recurrent neural network cells and model hyperparameters used, and then analysing the results to obtain a conclusion. The two proposed frameworks were able to carry out opinion triplet extraction and achieve decent performance. The DOER framework achieves better performance than the baselines on aspect and opinion term extraction tasks.

The paraphrase generator for citation sentences is used to produce several sentence alternatives to avoid plagiarism. Furthermore, the generation results need to pay attention to semantic similarity and lexical divergence standards. This study proposed the StoPGEN model as an algorithm for generating citation paraphrase sentences with stochastic output. The generation process is guided by an objective function using a simulated annealing algorithm to maintain the properties of semantic similarity and lexical divergence. The objective function is created by combining the two factors that maintain these properties. This study combined METEOR and PINC Scores in a linear weighting function that can be adjusted for its value tendency in one of the matrix functions. The dataset of citation sentences that had been labeled with paraphrases was used to test StoPGEN and other models for comparison. The StoPGEN model, with the citation sentences dataset, produced a BLEU score of 55.37, outperforming the bidirectional LSTM method with a value of 28.93. StoPGEN was also tested using Quora data by changing the language source in the architecture section resulting in a BLEU score of 22.37, outperforming UPSA 18.21. In addition, the qualitative evaluation results of the citation sentence generation based on respondents obtained an acceptance value of 50.80.

Automatic news categorization is essential to automatically handle the classification of multi-label news articles in online portal. This research employs some potential methods to improve performance of hierarchical multi-label classifier for Indonesian news article. First potential method is using Convolutional Neural Network (CNN) to build the top level classifier. The second method could improve the classification performance by calculating the average of the word vectors obtained from distributed semantic model. The third method combines lexical and semantic method to extract documents features, which multiplied word term frequency (lexical) with word vector average (semantic). Model build using Calibrated Label Ranking as multi-label classification method, and trained using Naïve Bayes algorithm has the best F1-measure of 0.7531. Multiplication of word term frequency and the average of word vectors were also used to build this classifiers. This configuration improved multi-label classification performance by 4.25%, compared to the baseline. The distributed semantic model that gave best performance in this experiment obtained from 300-dimension word2vec of Wikipedia’s articles. The multi-label classification model performance is also influenced by news’ released date. The difference period between training and testing data would also decrease models’ performance.

Captioning is the process of assembling a description for an image. Previous research on captioning has usually focused on foreground objects. In captioning concepts, there are two main objects for discussion: background object and foreground object. In contrast to the previous image-captioning research, generating captions from the geological images of rocks is more focused on the background of the images. This study proposed image captioning using a convolutional neural network, long short-term memory, and word2vec to generate words from the image. The proposed model was constructed by a convolutional neural network (CNN), long short-term memory (LSTM), and word2vec and gave a dense output of 256 units. To make it properly grammatical, a sequence of predicted words was reconstructed into a sentence by the beam search algorithm with K = 3. An evaluation of the pre-trained baseline model VGG16 and our proposed CNN-A, CNN-B, CNN-C, and CNN-D models used BLEU score methods for the N-gram. The BLEU scores achieved for BLEU-1 using these models were 0.5515, 0.6463, 0.7012, 0.7620, and 0.5620, respectively. BLEU-2 showed scores of 0.6048, 0.6507, 0.7083, 0.8756, and 0.6578, respectively. BLEU-3 performed with scores of 0.6414, 0.6892, 0.7312, 0.8861, and 0.7307, respectively. Finally, BLEU-4 had scores of 0.6526, 0.6504, 0.7345, 0.8250, and 0.7537, respectively. Our CNN-C model outperformed the other models, especially the baseline model. Furthermore, there are several future challenges in studying captions, such as geological sentence structure, geological sentence phrase, and constructing words by a geological tagger.

The use of graph pattern association rules (GPARs) on the Yago knowledge base is proposed. Extending association rules for itemsets, GPARS can help to discover regularities between entities in a knowledge base. A rule-generated graph pattern (RGGP) algorithm was used for extracting rules from the Yago knowledge base and a GPAR algorithm for creating the association rules. Our research resulted in 1114 association rules, with the value of standard confidence at 50.18% better than partial completeness assumption (PCA) confidence at 49.82%. Besides that the computation time for standard confidence was also better than for PCA confidence.

A common task in summarizing scientific articles is employing the rhetorical structure of sentences. Determining rhetorical sentences itself passes through the process of text categorization. In order to get good performance, some works in text categorization have been done by employing word embedding. This paper presents rhetorical sentence categorization of scientific articles by using word embedding to capture semantically similar words. A comparison of employing Word2Vec and GloVe is shown. First, two experiments are evaluated using five classifiers, namely Naïve Bayes, Linear SVM, IBK, J48, and Maximum Entropy. Then, the best classifier from the first two experiments was employed. This research showed that Word2Vec CBOW performed better than Skip-Gram and GloVe. The best experimental result was from Word2Vec CBOW for 20,155 resource papers from ACL-ARC, features from Teufel and the previous label feature. In this experiment, Linear SVM produced the highest F-measure performance at 43.44%.

This paper presents a studyon automatic title generation for scientific articles considering sentence information types known as rhetorical categories. A title can be seenas a high-compression summary of a document. A rhetorical category is an information type conveyed by the author of a text for each textual unit, for example: background, method, or result of the research. The experiment in this studyfocused on extracting the research purpose and research method information for inclusion in a computer-generated title. Sentences are classifiedinto rhetorical categories, after which these sentences are filtered using three methods. Three title candidates whose contents reflect the filtered sentencesare then generated using a template-based or an adaptive K-nearest neighbor approach. The experiment was conducted using two different dataset domains: computational linguistics and chemistry. Our study obtained a 0.109-0.255 F1-measure score on average for computer-generated titles compared to original titles. In a human evaluation the automatically generated titles were deemed 'relatively acceptable' in the computational linguistics domain and 'not acceptable' in the chemistry domain. It can be concluded that rhetorical categories have unexplored potential to improve the performance of summarization tasks in general.

In order to assist researchers in addressing time constraint and low relevance in using scientific articles, an automatic tailored multi-paper summarization (TMPS) is proposed. In this paper, we extend Teufel’s tailored summary to deal with multi-papers and more flexible representation of user information needs. Our TMPS extracts Rhetorical Document Profile (RDP) from each paper and presents a summary based on user information needs. Building Plan Language (BPLAN) is introduced as a formalization of Teufel’s building plan and used to represent summary specification, which is more flexible representation user information needs. Surface repair is embedded within the BPLAN for improving the readability of extractive summary. Our experiment shows that the average performance of RDP extraction module is 94.46%, which promises high quality of extracts for summary composition. Generality evaluation shows that our BPLAN is flexible enough in composing various forms of summary. Subjective evaluation provides evidence that surface repair operators can improve the resulting summary readability.

RESTRICTED CONTENT CLASSIFICATION BASED ON VIDEA METADATA AND COMMENTS (CASE STUDY : YOUTUBE.COM) aStefanus Thobi Sinaga, aMasayu Leylia Khodra a,bSekolah Teknik Elektro dan Informatika, Institut Teknologi Bandung, Jl. Ganesha 10 Bandung E-Mail: s.thobi.sinaga@gmail.com Abstrak Klasifikasi konten terbatas merupakan kegiatan memisahkan konten video yang layak untuk seluruh pengguna dari konten yang tidak layak untuk pengguna di bawah umur (

Since abstract of scientific paper is author biased, readers’ required information may not be included in the abstract. Tailored summary may help them to get a summary based on their information needs. This research is the first one that implements tailored summary system for scientific paper. Tailored summary applies information extraction that transforms a scientific paper into Rhetorical Document Profile, a structured representation of paper content based on rhetorical scheme of fifteen slots. This research adapted building plan that used rhetorical scheme of seven slots. We also implement tailored summary system. After generating initial summary, surface repair is conducted to improve summary readability. Each sentence in initial summary is combined with template phrase based on syntax-tree combination method. There are five groups of template phrases provided in surface repair. We construct evaluation standards by asking five human raters. The best method for sentence selection subsystem that uses Maximal Marginal Importance-Multi Sentence is employing TF.IDF weighting system with precision/recall of 0.61. The surface repair subsystem has acceptance of 0.91.