Deep sentiment analysis across language boundaries

Description

As more and more users come online across the globe, increasing numbers of people are voicing their opinion on social media, review sites, and more. Given such valuable data, modern deep learning-based sentiment analysis methods excel at determining the sentiment of what is being said about companies and products. Unfortunately, such deep methods require substantial amounts of training data. This is a significant problem for many of the world’s languages, for which resources may be too costly to obtain and training data is scarce, especially when one considers that new training data is needed for each domain and genre. A model trained on movie reviews, for instance, will fare very poorly on the task of assessing digital camera reviews, let alone social media postings such as tweets. To make things worse, many languages have peculiarities that make them significantly more challenging than English.

Gerard de Melo shares approaches to overcome these challenges, demonstrating how to exploit large amounts of surrogate data to learn advanced word representations that are custom-tailored for sentiment and outlining a special deep neural architecture to use them. For instance, the word “hot” is often positive when referring to music but tends to be negative when referring to the temperature in a hotel room. These sorts of representations can be fed into a bespoke deep convolutional network that is particularly adept at operating on them. Gerard concludes by demonstrating how the resulting freely available resources can be used by developers to effortlessly perform sentiment and emotion analysis in a number of different natural languages.