Effective sentiment stream analysis with self-augmenting training and demand-driven projection

Abstract

How do we analyze sentiments over a set of opinionated Twitter messages? This issue has been widely studied in recent years, with a prominent approach being based on the application of classification techniques. Basically, messages are classified according to the implicit attitude of the writer with respect to a query term. A major concern, however, is that Twitter (and other media channels) follows the data stream model, and thus the classifier must operate with limited resources, including labeled data for training classification models. This imposes serious challenges for current classification techniques, since they need to be constantly fed with fresh training messages, in order to track sentiment drift and to provide up-to-date sentiment analysis.We propose solutions to this problem. The heart of our approach is a training augmentation procedure which takes as input a small training seed, and then it automatically incorporates new relevant messages to the training data. Classification models are produced on-the-fly using association rules, which are kept up-to-date in an incremental fashion, so that at any given time the model properly reflects the sentiments in the event being analyzed. In order to track sentiment drift, training messages are projected on a demanddriven basis, according to the content of the message being classified. Projecting the training data offers a series of advantages, including the ability to quickly detect trending information emerging in the stream. We performed the analysis of major events in 2010, and we show that the prediction performance remains about the same, or even increases, as the stream passes and new training messages are acquired. This result holds for different languages, even in cases where sentiment distribution changes over time, or in cases where the initial training seed is rather small. We derive lower-bounds for prediction performance, and we show that our approach is extremely effective under diverse learning scenarios, providing gains that range from 7% to 58%.

Sentiment Analysis in Social Streams

“…Research is currently being conducted to adapt to new domains, by automatically assigning sentiment to terms not previously covered by the lexicons, and by providing dynamic re-training of existing classifiers [9][50][61] [65]. There are also recent approaches aimed to generate domain-dependent lexicons, such as that presented in [25].…”

Discovering and monitoring product features and the opinions on them with OPINSTREAM

“…For feature polarity learning, we compare with the non-adaptive semi-supervised classifier of [2], denoted as PolarityBaseline hereafter, and also with the method of Silva et al [21], denoted as Silva hereafter.…”

“…Semi-supervised stream learning is used by Silva et al [21]: they require only a small number of labeled documents, on which they train a classifier based on association rules; then, they update this initial training dataset incrementally with new documents, the label of which is derived by the classifier. Drury et al [22] also use self-training to extend the initial training dataset, but they assume a static setting.…”

“…Drury et al [22] also use self-training to extend the initial training dataset, but they assume a static setting. In [3], we propose a semi-supervised stream classifier for opinionated documents under concept drift: similarly to the methods of [21,22], it expands the training dataset with new documents, but it also forgets old documents, so that the training set grows and shrinks with time; this allows for a better response to concept drift. In OPINSTREAM we invoke the semi-supervised classifier [3] within each cluster.…”

Learning to Rank Similar Apparel Styles with Economically-Efficient Rule-Based Active Learning

“…Active learning algorithms have been proposed to help dealing with the labeling effort problem in L2R [27,8,26]. The motivation behind active learning is that it may be possible to learn effective ranking models by labeling only data instances (i.e., images) that are "informative" to the learning algorithm.…”