Academia.eduAcademia.edu

Figure from:Learning Classification with Both Labeled and Unlabeled Data

See full PDFdownloadDownload figure
Table 1. Comparison between kupiec et al.’s summarizer system and discriminant and genera- tive CEM algorithms for the cmp_lg collection. All classifiers are trained in a fully supervised way The two CEM classifiers allow approximately 10% increase both in average preci- sion and in accuracy over Kupiec et al’s system. Another interesting result is that both discriminant and generative CEM trained on semi-supervised learning scheme (using 10% of labeled sentences together with 90% of unlabeled sentences in the training set) gave similar performances to the Kupiec et al.’s summarizer system fully supervised.

Table 1 Comparison between kupiec et al.’s summarizer system and discriminant and genera- tive CEM algorithms for the cmp_lg collection. All classifiers are trained in a fully supervised way The two CEM classifiers allow approximately 10% increase both in average preci- sion and in accuracy over Kupiec et al’s system. Another interesting result is that both discriminant and generative CEM trained on semi-supervised learning scheme (using 10% of labeled sentences together with 90% of unlabeled sentences in the training set) gave similar performances to the Kupiec et al.’s summarizer system fully supervised.

Related Figures (2)

Fig. 1. Web Pages classification (a) and e-mail Spam detection (5) - Performance of two clas- sifiers with respect to the ratio x of labeled data in the training set. The classifier is a logistic unit trained with labeled data in a supervised scheme (dashed bottom curves) and using the semi-supervised discriminant-CEM algorithm (solid top curves) For comparison, we have also performed tests with the logistic classifier trained in a supervised way using the same x% labeled data in the training set.
Fig. 1. Web Pages classification (a) and e-mail Spam detection (5) - Performance of two clas- sifiers with respect to the ratio x of labeled data in the training set. The classifier is a logistic unit trained with labeled data in a supervised scheme (dashed bottom curves) and using the semi-supervised discriminant-CEM algorithm (solid top curves) For comparison, we have also performed tests with the logistic classifier trained in a supervised way using the same x% labeled data in the training set.
Fig. 2. Average precision of 3 trainable summarizers with respect to the ratio of labeled sen- tences in the training set for the cmp_lg collection For text summarization, we have also compared in figure 2, discriminant-CEM to the generative-CEM algorithm presented in section 3.2. For the latter, we assume that the conditional density functions {f;,};-1,.... are normal distributions.
Fig. 2. Average precision of 3 trainable summarizers with respect to the ratio of labeled sen- tences in the training set for the cmp_lg collection For text summarization, we have also compared in figure 2, discriminant-CEM to the generative-CEM algorithm presented in section 3.2. For the latter, we assume that the conditional density functions {f;,};-1,.... are normal distributions.