Social Media Cross-Source and Cross-Domain Sentiment Classification

Abstract

Due to the expansion of Internet and Web 2.0 phenomenon, there is a growing interest in sentiment analysis of freely opinionated text. In this paper, we propose a novel cross-source cross-domain sentiment classification, in which cross-domain-labeled Web sources (Amazon and Tripadvisor) are used to train supervised learning models (including two deep learning algorithms) that are tested on typically nonlabeled social media reviews (Facebook and Twitter). We explored a three-step methodology, in which distinct balanced training, text preprocessing and machine learning methods were tested, using two languages: English and Italian. The best results were achieved using undersampling training and a Convolutional Neural Network. Interesting cross-source classification performances were achieved, in particular when using Amazon and Tripadvisor reviews to train a model that is tested on Facebook data for both English and Italian.

Environmental disclosure and stakeholder engagement via social media: State of the art and potential in public utilities

“…In the last few years, a wide range of scientific research has aimed to study the large amount of new data generated by internet users. Some applications have already been made in order to evaluate the sentiments of people related to any topic (Estévez‐Ortiz, García‐Jiménez, & Glösekötter, ; Zola, Cortez, Ragno, & Brentari, ) and to measure the outcomes of different disclosure strategies (Castelló, Etter, & Årup Nielsen, ; Colleoni, ; Etter et al, ). In this paper, we suggest the use of social media data and sentiment analysis to study the interactions on FB between stakeholders and PUs regarding environmental disclosure.…”

Sentimental classification analysis of polarity multi-view textual data using data mining techniques

“…In the related research of user's reviews, a sentiment classification of movie reviews has been handled by Ahuja et al [16] using dual training and dual prediction while addressing polarity shift. Consequently, Zola et al [17] have propose a novel sentiment classification, in which a three-step methodology is explored based on balanced training, text preprocessing and machine learning using two languages: English and Italian.…”

Twitter alloy steel disambiguation and user relevance via one-class and two-class news titles classifiers

“…Moreover, most supervised learning studies used binary labels, while we approach two training setups: one-class (unary), in which only positive financial texts are available; and two-class (binary), which assumes an access to both positive (financial) and negative (non financial) messages. Since Twitter texts are unlabeled, and in order to avoid a laborious manual effort, we use public and freely available news titles to set the positive and negative messages, thus making use of a transfer learning [22,39]. As for the TFD models, we adjust and compare several data preprocessing, statistical measures and ML algorithms, including recent Word2Vec encoding and deep learning methods (e.g., siamese autoencoder, deep multilayer perceptron).…”

“…In order to get labeled data for train the models, we use easy to collect and freely available news titles (as detailed in Section 4.1). Given the samples of P and N , the TFD models use a transfer learning [22,39], where the models are adjusted to one training source (news titles) and tested on a different source (Twitter).…”

“…All tested supervised ML methods (RF, MLP, SVM) and autoencoders require a fixed input size but the analyzed texts include a variable number of tokens. To handle this issue, when using direct text token inputs (TF-IDF or W2V), the truncation technique employed in [50,39] is adopted, which considers only the first M tokens, as they appear in the texts. If the texts have less than M tokens then we use padding, which consists in adding null values (e.g., 0) [39].…”