Sentiment Polarity Shifters: Creating Lexical Resources through Manual Annotation and Bootstrapped Machine Learning

Abstract

Alleviating pain is good and abandoning hope is bad. We instinctively understand how words like alleviate and abandon affect the polarity of a phrase, inverting or weakening it. When these words are content words, such as verbs, nouns and adjectives, we refer to them as polarity shifters. Shifters are a frequent occurrence in human language and an important part of successfully modeling negation in sentiment analysis; yet research on negation modeling has focussed almost exclusively on a small handful of closed class negation words, such as not, no and without. A major reason for this is that shifters are far more lexically diverse than negation words, but no resources exist to help identify them.

We seek to remedy this lack of shifter resources. Our most central step towards this is the creation of a large lexicon of polarity shifters that covers verbs, nouns and adjectives. To reduce the prohibitive cost of such a large annotation task, we develop a bootstrapping approach that combines automatic classification with human verification. This ensures the high quality of our lexicon while reducing annotation cost by over 70%.

In designing the bootstrap classifier we develop a variety of features which use both existing semantic resources and linguistically informed text patterns. In addition we investigate how knowledge about polarity shifters might be shared across different parts of speech, highlighting both the potential and limitations of such an approach.

The applicability of our bootstrapping approach extends beyond the creation of a single resource. We show how it can further be used to introduce polarity shifter resources for other languages. Through the example case of German we show that all our features are transferable to other languages. Keeping in mind the requirements of under-resourced languages, we also explore how well a classifier would do when relying only on data- but not resource-driven features. We also introduce ways to use cross-lingual information, leveraging the shifter resources we previously created for other languages.

Apart from the general question of which words can be polarity shifters, we also explore a number of other factors. One of these is the matter of shifting directions, which indicates whether a shifter affects positive polarities, negative polarities or whether it can shift in either direction. Using a supervised classifier we add shifting directions information to our bootstrapped lexicon.

For other aspects of polarity shifting, manual annotation is preferable to automatic classification. Not every word that can cause polarity shifting does so for every of its word senses. As \wsdLong{} technology is not robust enough to allow the automatic handling of such nuances, we manually create a complete sense-level annotation of verbal polarity shifters.

To verify the usefulness of the lexica which we create, we provide an extrinsic evaluation in which we apply them to a sentiment analysis task. In this task the different lexica are not only compared amongst each other, but also against a state-of-the-art compositional polarity neural network classifier that has been shown to be able to implicitly learn the negating effect of negation words from a training corpus. However, we find that the same is not true for the far more lexically diverse polarity shifters. Instead, the use of the explicit knowledge provided by our shifter lexica brings clear gains in performance.