In this paper, we systematically explore feature definition and selection strategies for sentiment polarity classification. We begin by exploring basic questions, such as whether to use stemming, term frequency versus binary weighting, negation-enriched features, n-grams or phrases. We then move onto more complex aspects including feature selection using frequency-based vocabulary trimming, part-of-speech and lexicon selection (three types of lexicons), as well as using expected Mutual Information (MI). Using three product and movie review datasets of various sizes, we show, for example, that some techniques are more beneficial for larger datasets than the smaller. A classifier trained on only few features ranked high by MI outperformed one trained on all features in large datasets, yet in small dataset this did not prove to be true. Finally, we perform a space and computation cost analysis to further understand the merits of various feature types.

In this paper, we investigate the utility of linguistic features for detecting the sentiment of Twitter messages. We evaluate the usefulness of existing lexical resources as well as features that capture information about the informal and creative language used in microblogging. We take a supervied approach to the problem, but leverage existing hashtags in the Twitter data for building training data.

Using social media for political discourse is becoming common practice, especially around election time. One interesting aspect of this trend is the possibility of pulsing the public’s opinion about the elections, and that has attracted the interest of many researchers and the press. Allegedly, predicting electoral outcomes from social media data can be feasible and even simple. Positive results have been reported, but without an analysis on what principle enables them. Our work puts to test the purported predictive power of socialmedia metrics against the 2010 US congressional elections. Here, we applied techniques that had reportedly led to positive election predictions in the past, on the Twitter data collected from the 2010 US congressional elections. Unfortunately, we find no correlation between the analysis results and the electoral outcomes, contradicting previous reports. Observing that 80 years of polling research would support our findings, we argue that one should not be accepting predictions about events using social media data as a black box. Instead, scholarly research should be accompanied by a model explaining the predictive power of social media, when there is one.

In times of mass emergency, vast amounts of data are generated via computer-mediated communication (CMC) that are difficult to manually cull and organize into a coherent picture. Yet valuable information is broadcast, and can provide useful insight into time- and safety-critical situations if captured and analyzed properly and rapidly. We describe an approach for automatically identifying messages communicated via Twitter that contribute to situational awareness, and explain why it is beneficial for those seeking information during mass emergencies. We collected Twitter messages from four different crisis events of varying nature and magnitude and built a classifier to automatically detect messages that may contribute to situational awareness, utilizing a combination of hand-annotated and automatically-extracted linguistic features. Our system was able to achieve over 80% accuracy on categorizing tweets that contribute to situational awareness. Additionally, we show that a classifier developed for a specific emergency event performs well on similar events. The results are promising, and have the potential to aid the general public in culling and analyzing information communicated during times of mass emergency.

Although sentiment analysis has attracted a lot of research, little work has been done on social media data compared to product and movie reviews. This is due to the low accuracy that results from the more informal writing seen in social media data. Currently, most of sentiment analysis tools on social media choose the lexicon-based approach instead of the machine learning approach because the latter requires the huge challenge of obtaining enough human-labeled training data for extremely large-scale and diverse social opinion data. The lexicon-based approach requires a sentiment dictionary to determine opinion polarity. This dictionary can also provide useful features for any supervised learning method of the machine learning approach. However, many benchmark sentiment dictionaries do not cover the many informal and spoken words used in social media. In addition, they are not able to update frequently to include newly generated words online. In this paper, we present an automatic sentiment dictionary generation method, called Constrained Symmetric Nonnegative Matrix Factorization (CSNMF) algorithm, to assign polarity scores to each word in the dictionary, on a large social media corpus — digg.com. Moreover, we will demonstrate our study of Amazon Mechanical Turk (AMT) on social media word polarity, using both the human-labeled dictionaries from AMT and the General Inquirer Lexicon to compare our generated dictionary with. In our experiment, we show that combining links from both WordNet and the corpus to generate sentiment dictionaries does outperform using only one of them, and the words with higher sentiment scores yield better precision. Finally, we conducted a lexicon-based sentiment analysis on human-labeled social comments using our generated sentiment dictionary to show the effectiveness of our method.

Information extraction systems have been recently proposed for organizing and exploring content in large online text corpora as information networks . In such networks, the nodes are named entities (e.g., people, organizations) while the edges correspond to statements indicating relations among such entities. To date, such systems extract rather primitive networks, capturing only those relations which are expressed by direct statements. In many applications, it is useful to also extract more subtle relations which are often expressed as meta statements in the text. These can, for instance provide the context for a statement (e.g., “Google acquired YouTube on October 2006”), or repercussion about a statement (e.g., “The US condemned Russia’s invasion of Georgia”). In this work, we report on a system for extracting relations expressed in both direct statements as well as in meta statements. We propose a method based on Conditional Random Fields that explores syntactic features to extract both kinds of statements seamlessly. We follow the Open Information Extraction paradigm, where a classifier is trained to recognize any type of relation instead of specific ones. Finally, our results show substantial improvements over a state-of-the-art information extraction system, both in terms of accuracy and, especially, recall.

The amount of information available on the Web grows at an incredible high rate. Systems and procedures devised to extract these data from Web sources already exist, and different approaches and techniques have been investigated during the last years. On the one hand, reliable solutions should provide robust algorithms of Web data mining which could automatically face possible malfunctioning or failures. On the other, in literature there is a lack of solutions about the maintenance of these systems. Procedures that extract Web data may be strictly interconnected with the structure of the data source itself; thus, malfunctioning or acquisition of corrupted data could be caused, for example, by structural modifications of data sources brought by their owners. Nowadays, verification of data integrity and maintenance are mostly manually managed, in order to ensure that these systems work correctly and reliably. In this paper we propose a novel approach to create procedures able to extract data from Web sources -- the so called Web wrappers -- which can face possible malfunctioning caused by modifications of the structure of the data source, and can automatically repair themselves.

Inferring the sentiment of social media content, for instance blog postings or online product reviews, is both of great interest to businesses and technically challenging to accomplish. This paper presents two computational methods for estimating social media sentiment which address the challenges associated with Web-based analysis. Each method formulates the task as one of text classification, models the data as a bipartite graph of documents and words, and assumes that only limited prior information is available regarding the sentiment orientation of any of the documents or words of interest. The first algorithm is a semi-supervised sentiment classifier which combines knowledge of the sentiment labels for a few documents and words with information present in unlabeled data, which is abundant online. The second algorithm assumes existence of a set of labeled documents in a domain related to the domain of interest, and leverages these data to estimate sentiment in the target domain. We demonstrate the utility of the proposed methods by showing they outperform several standard methods for the task of inferring the sentiment of online reviews of movies, electronics products, and kitchen appliances. Additionally, we illustrate the potential of the methods for multilingual business informatics through a case study involving estimation of Indonesian public opinion regarding the July 2009 Jakarta hotel bombings.

Nowadays, the huge amount of information distributed through the Web motivates studying techniques to be adopted in order to extract relevant data in an efficient and reliable way. Both academia and enterprises developed several approaches of Web data extraction, for example using techniques of artificial intelligence or machine learning. Some commonly adopted procedures, namely wrappers, ensure a high degree of precision of information extracted from Web pages, and, at the same time, have to prove robustness in order not to compromise quality and reliability of data themselves. In this paper we focus on some experimental aspects related to the robustness of the data extraction process and the possibility of automatically adapting wrappers. We discuss the implementation of algorithms for finding similarities between two different version of a Web page, in order to handle modifications, avoiding the failure of data extraction tasks and ensuring reliability of information extracted. Our purpose is to evaluate performances, advantages and draw-backs of our novel system of automatic wrapper adaptation.

This paper presents a co-regularization based approach to semi-supervised domain adaptation. Our proposed approach (EA) builds on the notion of augmented space (introduced in EASYADAPT (EA) [1]) and harnesses unlabeled data in target domain to further assist the transfer of information from source to target. This semi-supervised approach to domain adaptation is extremely simple to implement and can be applied as a pre-processing step to any supervised learner. Our theoretical analysis (in terms of Rademacher complexity) of EA and EA show that the hypothesis class of EA has lower complexity (compared to EA) and hence results in tighter generalization bounds. Experimental results on sentiment analysis tasks reinforce our theoretical findings and demonstrate the efficacy of the proposed method when compared to EA as well as few other representative baseline approaches.