Class normalization in centroid-based text categorization

Centroid-based categorization is one of the most popular algorithms in text classification. In this approach, normalization is an important factor to improve performance of a centroid-based classifier when documents in text collection have quite different sizes and/or the numbers of documents in classes are unbalanced. In the past, most researchers applied document normalization, e.g., document-length normalization, while some consider a simple kind of class normalization, so-called class-length normalization, to solve the unbalancedness problem. However, there is no intensive work that clarifies how these normalizations affect classification performance and whether there are any other useful normalizations. The purpose of this paper is three folds; (1) to investigate the effectiveness of document- and class-length normalizations on several data sets, (2) to evaluate a number of commonly used normalization functions and (3) to introduce a new type of class normalization, called term-length normalization, which exploits term distribution among documents in the class. The experimental results show that a classifier with weight-merge-normalize approach (class-length normalization) performs better than one with weight-normalize-merge approach (document-length normalization) for the data sets with unbalanced numbers of documents in classes, and is quite competitive for those with balanced numbers of documents. For normalization functions, the normalization based on term weighting performs better than the others on average. For term-length normalization, it is useful for improving classification accuracy. The combination of term- and class-length normalizations outperforms pure class-length normalization and pure term-length normalization as well as unnormalization with the gaps of 4.29%, 11.50%, 30.09%, respectively.     

Pronouncibility index (Π): a distance-based and confusion-based speech quality measure for dysarthric speakers

Recently, there have been many modern speech technologies, including those of speech synthesis and recognition, developed to help people with disabilities. While most of such technologies have successfully been applied to process speech of normal speakers, they may not be effective for speakers with speech disorder, depending on their severity. This paper proposes an automated method to preliminarily assess the ability of a speaker in pronouncing a word. Based on signal features, an indicator called pronouncibility index (Π) is introduced to express speech quality with two complementary measures, called distance-based and confusion-based factors. In the distance-based factor, the 1-norm, 2-norm and 3-norm distance are investigated while boundary-based and Gaussian-based approaches are introduced for confusion-based factors. The Π is used to estimate performance of speech recognition when it is applied to recognize speech of a dysarthric speaker. Three measures are applied to evaluate the effectiveness of Π, rank-order inconsistency, correlation coefficient, and root-mean-square of difference. The evaluations had been done by comparing its predicted recognition rates with ones predicted by the standard methods called the articulatory and intelligibility tests based on the two recognition systems (HMM and ANN). For the phoneme-test set (the training set), Π outperforms the articulatory and intelligibility tests in all three evaluations. The performance of Π decreases for the device-control set (the test set), and the intelligibility test becomes the best method followed by Π and the articulatory test. In general, Π is a promising indicator for predicting recognition rate with comparison to the standard assessments.     

A Learning-Based Approach for Web Cache Management

Web caching has been widely used to alleviate Internet traffic congestion in World Wide Web (WWW) services. To reduce download throughput, an effective strategy on web cache management is needed to exploit web usage information in order to make a decision on evicting the document stored in case of cache saturation. This paper presents a so-called Learning Based Replacement algorithm (LBR), a hybrid approach towards an efficient replacement model for web caching by incorporating a machine learning technique (naive Bayes) into the LRU replacement method to improve prediction of possibility that an existing page will be revised by a succeeding request, from access history in a web log. The learned knowledge includes information on which URL objects in cache should be kept or evicted. The learning-based model is acquired to represent the hidden aspect of user request pattern for predicting the re-reference possibility. By a number of experiments, the LBR gains potential improvement of prediction on revisit probability, hit rate and byte hit rate overtraditional methods; LRU, LFU, and GDSF, respectively.     

Thai spelling analysis for automatic spelling speech recognition

Spelling speech recognition can be applied for several purposes including enhancement of speech recognition systems and implementation of name retrieval systems. This paper presents a Thai spelling analysis to develop a Thai spelling speech recognizer. The Thai phonetic characteristics, alphabet system and spelling methods have been analyzed. As a training resource, two alternative corpora, a small spelling speech corpus and an existing large continuous speech corpus, are used to train hidden Markov models (HMMs). Then their recognition results are compared to each other. To solve the problem of utterance speed difference between spelling utterances and continuous speech utterances, the adjustment of utterance speed has been taken into account. Two alternative language models, bigram and trigram, are used for investigating performance of spelling speech recognition. Our approach achieves up to 98.0% letter correction rate, 97.9% letter accuracy and 82.8% utterance correction rate when the language model is trained based on trigram and the acoustic model is trained from the small spelling speech corpus with eight Gaussian mixtures.     

Multidimensional text classification for drug information

This paper proposes a multidimensional model for classifying drug information text documents. The concept of multidimensional category model is introduced for representing classes. In contrast with traditional flat and hierarchical category models, the multidimensional category model classifies each document using multiple predefined sets of categories, where each set corresponds to a dimension. Since a multidimensional model can be converted to flat and hierarchical models, three classification approaches are possible, i.e., classifying directly based on the multidimensional model and classifying with the equivalent flat or hierarchical models. The efficiency of these three approaches is investigated using drug information collection with two different dimensions: 1) drug topics and 2) primary therapeutic classes. In the experiments, k-nearest neighbor, naive Bayes, and two centroid-based methods are selected as classifiers. The comparisons among three approaches of classification are done using two-way analysis of variance, followed by the Scheffé's test for post hoc comparison. The experimental results show that multidimensional-based classification performs better than the others, especially in the presence of a relatively small training set. As one application, a category-based search engine using the multidimensional category concept was developed to help users retrieve drug information.     

Boosting-based ensemble learning with penalty profiles for automatic Thai unknown word recognition

A boosting-based ensemble learning can be used to improve classification accuracy by using multiple classification models constructed to cope with errors obtained from their preceding steps. This paper proposes a method to improve boosting-based ensemble learning with penalty profiles via an application of automatic unknown word recognition in Thai language. Treating a sequential problem as a non-sequential problem, the unknown word recognition is required to include a process to rank a set of generated candidates for a potential unknown word position. To strengthen the recognition process with ensemble classification, the penalty profiles are defined to make it more efficient to construct a succeeding classification model which tends to re-rank a set of ranked candidates into a suitable order. As an evaluation, a number of alternative penalty profiles are introduced and their performances are compared for the task of extracting unknown words from a large Thai medical text. Using the Naïve Bayes as the base classifier for ensemble learning, the proposed method with the best setting achieves an accuracy of 90.19%, which is an accuracy gap of 12.88, 10.59, and 6.05 over conventional Naïve Bayes, non-ensemble version, and the flat-penalty profile.