Citation Context

...erate ranking functions which predict the relevance of a document. One class of algorithms minimizes the errors resulting from ordering document pairs incorrectly. Examples include (Joachims, 2002), (=-=Freund et al., 2003-=-) and (Qin et al., 2007). In particular, the training phase of the Joachims’ Ranking SVM (Joachims, 2002) is formulated as the following SVM optimization problem: min w, ξ ∑ 1 w T ξ i , j , k 2 , subj...

Citation Context

... the examples (Ω(di), fi(ej)) for all documents di in the training corpus. The regression framework we adopt is Support Vector Regression (SVR), which is a regression analysis technique based on SVM (=-=Schölkopf et al., 2000-=-). We require hj to have the form w T Ω(di) + b. Finding hj is equivalent to solving the following optimization problem: 1 min T w , b, ξ , C ( , 1 , 2) i , 1 ξ w w + ξ i , 2 2 ∑ i + ξi subject to: ∀d...

Citation Context

...hereas Aman and Szpakowicz (2007) manually label phrases with emotional categories. In 2007, the SemEval-2007 workshop organized a task on the unsupervised annotation of news headlines with emotions (=-=Strapparava and Mihalcea, 2007-=-). As for the task of ranking, many machinelearning algorithms have been proposed in information retrieval. These techniques generate ranking functions which predict the relevance of a document. One c...

Citation Context

...ich predict the relevance of a document. One class of algorithms minimizes the errors resulting from ordering document pairs incorrectly. Examples include (Joachims, 2002), (Freund et al., 2003) and (=-=Qin et al., 2007-=-). In particular, the training phase of the Joachims’ Ranking SVM (Joachims, 2002) is formulated as the following SVM optimization problem: min w, ξ ∑ 1 w T ξ i , j , k 2 , subject to: ∀( q , d ), ( q...

Citation Context

...s. To do this, we use Kendall’s τb correlation coefficient, which is a statistical measure for determining the correlation between two ranked lists when there may be ties between two items in a list (=-=Liebetrau, 1983-=-). The value of τb is determined based on the number of concordant pairwise orders and the number of discordant pairwise orders between two ranked lists. Therefore, this measure is appropriate for eva...

Citation Context

...ment. They discover that using Support Vector Machines (SVM) with word unigram features results in the best performance. Since then, more work has been done to find features better than unigrams. In (=-=Hu et al., 2005-=-), word sentiment information is exploited to achieve better classification accuracy. Experiments have been done to extract emotional information from texts at granularities finer than documents. Wieb...

Citation Context

...ss of correct Chinese words. This problem is particularly severe if there are a lot of out-of-vocabulary words in a dataset. In Chinese, around 70% of all Chinese words are Chinese character bigrams (=-=Chen et al., 1997-=-). Thus, using Chinese character bigrams as features will allow us to identify a lot of Chinese words, which when combined with the words extracted by the word segmenter, will give us a wider coverage...