No abstract

In this paper, we consider the problem of building models that have high subjectivity classification accuracy across domains. For that purpose, we present and evaluate new methods based on multi-view learning using both high-level (i.e. linguistic features for subjectivity detection) and lowlevel features (i.e. unigrams and bigrams). In particular, we show that multi-view learning, combining high-level and low-level features with adapted classifiers, can lead to improved results compared to one of the state-of-the-art algorithms called Stochastic Agreement Regularization. In particular, the experiments show that dividing the set of characteristics into three views returns the best results overall with accuracy across domains of 91.3% for the Class-Guided Multi-View Learning Algorithm, which combines both Linear Discriminant Analysis and Support Vector Machines.

In this paper we consider the problem of building models that have high sentiment classification accuracy without the aid of a labeled dataset from the target domain. For that purpose, we present and evaluate a novel method based on level of abstraction of nouns. By comparing high-level features (e.g. level of affective words, level of abstraction of nouns) and low-level features (e.g. unigrams, bigrams), we show that, high-level features are better to learn subjective language across domains. Our experimental results present accuracy levels across domains of 71.2% using SVMs learning models.

In this paper, we propose to study the characteristics for analyzing subjective content in documents. For that purpose, we present and evaluate a novel method based on level of abstraction of nouns. By comparing state-of-the-art features and the level of abstraction of nouns between three annotated corpora and texts downloaded from Wikipedia and Web Blogs, we show that, building data sets for the classification of opinionated texts can be done automatically from the web, at the document level. Moreover, we present accuracy levels within domains of 96.5% and across domains of 74.5%.