As expected, the structural methods on either skewed or flattened hierarchies are not significantly better than the flat SVM .

For the flattened hierarchy of 15 leaf genres the maximal accuracy is 54.2% vs. 52.4% for the flat SVM (Figure 3), a nonsignificant improvement.

As a baseline we use the accuracy achieved by a standard "flat" SVM.

A standard flat SVM achieves an accuracy of 64.4% whereas the best structural SVM based on Lin’s information content distance measure (IC-lin-word-bnc) achieves 68.8% accuracy, significantly better at the 1% level.

Table 1 summarizes the best performing measures that all outperform the flat SVM at the 1% level.

The structural SVM (Section 2) requires a distance measure h between two genres.

To strengthen the constraints, the zero value on the right hand side of the inequality for the flat SVM can be replaced by a positive value, corresponding to a distance measure h(yi, m) between two genre classes, leading to the following constraint:

We compared the Gibbs sampling compressor (GS) against a version of maximum a posteriori EM (with Dirichlet parameter greater than 1) and a discriminative STSG based on SVM training (Cohn and Lapata, 2008) ( SVM ).

EM is a natural benchmark, while SVM is also appropriate since it can be taken as the state of the art for our task.4

Nonetheless, because the comparison system is a generalization of the extractive SVM compressor of Cohn and Lapata (2007), we do not expect that the results would differ qualitatively.

We achieve substantial improvements against a number of baselines including EM, support vector machine ( SVM ) based discriminative training, and variational Bayes (VB).

We apply both support vector machine ( SVM ) and Maximum Entropy (ME) algorithms with the help of the SVM-light4 and Mallet5 tools.

We find that ME performs slightly better than SVM on the average.

Transductive SVM , which seeks the largest separation between labeled and unlabeled data through regularization (Joachims, 1999).

While we tested several classifiers, we chose to use two different classifiers based on SVM and Logistic Regression for the final experimental results because they showed the best performance.

Logistic Feature Regression SVM

Logistic Feature Regression SVM

ME SVM Prec.

We employ an SVM coreference resolver trained and tested on ACE 2005 with 79.5% Precision, 66.7% Recall and 72.5% F1 to label coreference mentions of the same named entity in an article.

And thus an SVM classifier can be learned and then used for recognition.

Section 4 introduces the frame work for discourse recognition, as well as the baseline feature space and the SVM classifier.

The classifier learned by SVM is:

One advantage of SVM is that we can use tree kernel approach to capture syntactic parse tree information in a particular high-dimension space.

We set the degree of the kernels to 3 since cubic kernels with SVM have proved effective for Japanese dependency parsing (Kudo and Matsumoto, 2000; Kudo and Matsumoto, 2002).

Stopping Criteria It is known that increment rate of the number of support vectors in SVM indicates saturation of accuracy improvement during iterations of active learning (Schohn and Cohn, 2000).

It is interesting to examine whether the observation for SVM is also useful for support vectors7 of the averaged perceptron.

In the 1st phase, a kernel based classifier, SVM in our study, is employed to classify each candidate subtree pair as aligned or unaligned.

Since SVM is a large margin based discriminative classifier rather than a probabilistic model, we introduce a sigmoid function to convert the distance against the hyperplane to a posterior alignment probability as follows:

We use SVM with binary classes as the classifier.

We learned the feature weights with a linear SVM , using the software SVM-OOPS (Woodsend and Gondzio, 2009).

For each phrase, features were extracted and salience scores calculated from the feature weights determined through SVM training.

The distance from the SVM hyperplane represents the salience score.

SMO is an implementation of Support Vector Machines ( SVM ), rules.JRip is the RIPPER algorithm, and bayes .NaiveBayes is a Naive Bayes classifier.

In task C, the SVM algorithm was also the best performing algorithm among those that were also tried on the English data, but decision trees produced even better results here.

The results are: in task A the lazy.KStar classifier scored 58.6%, and the SVM classifier scored 75.5% in task B and 59.4% in task C, with trees .

SVM ) should be separately created with regards to distinct features.

We utilised SVanl‘;8 as an implementation of the Ranking SVM algorithm, in which the parameter c was set as 1.0 and the remaining parameters were set to their defaults.

Although the work by Denis and Baldridge (2008) uses Maximum Entropy to create their ranking-based model, we adopt the Ranking SVM algorithm (J oachims, 2002), which learns a weight vector to rank candidates for a given partial ranking of each referent.

Previous studies have found that, among several ML-based approaches, the SVM classifier generally performs well in many subjectivity analysis tasks (Pang et al., 2002; Banea et al., 2008).

An SVM score (a margin or the distance from a learned decision boundary) with a positive value predicts the input as being subjective, and negative value as objective.

The second and the third approaches are carried out as follows: Corpus-based (T-CB): We translate the MPQA corpus into the target languages sentence by sentence using a web-based service.6 Using the same method for S-CB, we train an SVM model for each language with the translated training corpora.

We use a Ranking SVM (Si/Mug“ (Joachims, 2002)) to score summaries using our features.

The Ranking SVM seeks to minimize the number of discordant pairs (pairs in which the gold standard has :31 ranked strictly higher than :52, but the learner ranks x2 strictly higher than :01).

For system-level evaluation, we treat the real-valued output of the SVM ranker for each summary as the linguistic quality score.