求贵人帮忙翻译

A Multi-Classifier Based Guideline SentenceClassification System
A 10-fold cross validation using 340 sentences was done. Althoughthe typical method is to split data into training, testingand validation sets, n-fold cross validation is preferredwhen the available data are highly limited in size. So we splitthe data into 10 sub-sets and tested on every nth data aftertraining each model using the rest of the data, calculatingprecision, recall and f-measure on average. The ratio of thesize of sentences for four classes was identical, the number ofinstances thus being 85 per class. Each classifier was trainedusing a set of 340 feature vectors.
Much to the contrary to the initial intuition, the f-measureof overall classifiers given in Table 4 was over 0.78. Theperformance of Naïve Bayes was relatively lower than otheralgorithms, probably due to its being based on independenceassumption, which does not stand in harmony with the nonlinearnature inherent in textual object generation systems.Since this research presupposed that these base learnerswould choose weak classification model parameters andthat it would be necessary for them to be sequentially combinedby boosting for learning strong classifiers based on aweighted voting scheme, the result above simply eradicatedthe necessity of using a meta-algorithm (Table 5).
Since weak classifiers are generally defined as having slightlybetter accuracy than a random classifier with an accuracyof less than 50%, we concluded that even boosting NaïveBayes classifiers would not cause distinct improvement.Table 6 is the actual test in which we sequentially combinedNaïve Bayes classifiers using the AdaBoost.M1 meta-algorithm.
As given in (Table 6), the improvement of Naïve Bayes classifierscombined by AdaBoost.M1 was not sufficiently conspicuous.This is because the transformation of the originalfeature space by feature extractors into a lower dimensionmay have reduced the complexity and non-linearity of thedata, decreasing the width of scatter and the number of outliers.In actual fact, boosting Neural Network algorithm suchas Multi-layered Perceptron or Radial Basis Function Networksaw no increase of accuracy by AdaBoost.M1.
The aims of this research were to apply transformation totackle the problem of dimensionality and to increase classificationaccuracy by applying a boosting algorithm to learna strong classifier that is robust to outliers and the non-linearcharacteristics of data. The second purpose turns out to bemeaningless when the curse of dimensionality is resolvedand robust classification algorithms such as a multilayer perceptronor a radial basis function network are adopted.
Moreover, we found that transformation has the advantageof exploiting structural and underlying features which gounseen by the BOW model. We also realized that integratinga sentential classifier with a TF-IDF-based search engineenhances a search process by realizing the capabilityof maximizing the probability of automatically presentingrelevant information required in the context generated inthe guideline authoring environment. This, however, has adisadvantage of increasing the total amount of time requiredto parse and classify the set of sentences within a documentat runtime. Therefore, our future study shall be focused onexcluding slow parsing processes while extracting structuralfeatures from a textual object.

全部

圆***

回答